content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
According to the current paradigm (see Urry $\&$ Padovani 1995) Active
Galactic Nuclei (AGNs) are galaxies whose emission is dominated by a bright
central core, including a super massive black hole as central engine,
surrounded by an accretion disk and by fast-moving clouds under the influence
of the strong gravitational field, emitting Doppler broadened lines. Absorbing
material in a flattened configuration, idealized as a dust torus, obscures the
central parts so that for transverse line of sight only the more distant
narrow-line emitting clouds are seen directly. In radio-loud objects we have
the additional presence of a relativistic jet, roughly perpendicular to the disk. Within this scheme, blazars represent the fraction of AGNs with their
jet at smaller angles with respect to our line of sight, which causes relativistic aberration and emission amplification (Blandford $\&$ Rees 1978).
Blazars are the most enigmatic subclass of AGNs, characterized by the emission of strong non-thermal radiation across the entire electromagnetic
spectrum, from radio to very high $\gamma$-ray energies. Because of their
peculiar properties and amplified emission, blazars offer the unique
possibility to probe the central region of AGNs, and shed light on the
mechanism responsible for the extraction of energy from the central black hole
and for the acceleration and collimation of relativistic electrons into
jets. With the detection of several blazars in the $\gamma$-rays by EGRET
(Hartman et al.~1999) the study of this class of objects has made significant
progress but, despite the big efforts devoted to the investigation of the
mechanisms responsible for the emission in blazars, the definitive answer is
still missing. The new generation of high-energy space missions, the {\it
Astrorivelatore Gamma a Immagini LEggero} (AGILE) and {\it Fermi} Gamma-ray Space Telescope (GST) satellites, have to address some of the fundamental issues that were left unresolved by EGRET.
PKS 1510$-$089 is a nearby (z=0.361) blazar belonging to the class of the flat
spectrum radio quasars (FSRQs) with radiative output dominated by the
$\gamma$-ray emission, while the synchrotron emission peaks around IR
frequencies, below a pronunced UV bump, likely due to the thermal emission
from the accretion disk (Malkan $\&$ Moore 1986; Pian $\&$ Treves 1993; D'Ammando et al. 2009a).
Its radio emission exhibits very rapid, large amplitude variations in both
total and polarized flux (Aller, Aller, $\&$ Hughes 1996). Moreover, the radio
jet shows superluminal motion with an apparent velocity higher than 20$c$, among the fastest of all blazars observed thus far (Jorstad et al.~2005), with
the parsec and kiloparsec scale jets apparently misaligned by $\sim$180 degrees (Homan et al.~2002, Wardle et al.~2005).
PKS 1510$-$089 has been extensively observed in the X-rays during the last
three decades, since the observations by the {\it Einstein} satellite in 1980s
(Canizares $\&$ White 1989). The X-ray spectrum as observed by ASCA in the
2--10 keV band was very hard with photon index of $\Gamma$ $\simeq$ 1.3
(Singh, Shrader $\&$ George 1997), but steepened ($\Gamma$ $\simeq$ 1.9) in
the ROSAT bandpass (0.1--2.4 keV), suggesting the presence of a spectral break
around 1--2 keV, associated with the existence of a soft X-ray excess (Siebert
et al.~1998). Subsequent observations by {\it Beppo}SAX (Tavecchio et
al.~2000) and {\it Chandra} (Gambill et al.~2003) confirmed the presence of a
soft X-ray excess below 1 keV. Evidence of a similar soft X-ray excess has
been detected in other blazars such as 3C 273, 3C 279, AO 0235+164, and 3C
454.3. The origin of this excess is still an open issue, not only for blazars
but for all AGNs (see e.g.~D'Ammando et al.~2008, and the references therein).
The interest in the peculiar X-ray spectrum of the source led up to a
monitoring campaign organized during August 2006 by {\it Suzaku} and {\it
Swift}. The {\it Suzaku} observations confirmed the presence of a soft X-ray
excess, suggesting that it could be a feature of the bulk Comptonization,
whereas the {\it Swift}/XRT observations revealed significant spectral
evolution of the X-ray emission on timescales of a week: the X-ray spectrum
becomes harder as the source gets brighter (Kataoka et
al.~2008). Unfortunately no $\gamma$-ray satellites were operating at that
time, therefore it was not possible to investigate the correlation between
X-ray and $\gamma$-ray behaviour. However, $\gamma$-ray emission from PKS
1510$-$089 had already been detected in the past by EGRET during
low/intermediate states and was found to be only slightly variable, with an
integrated flux above 100 MeV varying between (13 $\pm$ 5) and (49 $\pm$ 18) $\times$ 10$^{-8}$ photons cm$^{-2}$ s$^{-1}$ (Hartman et al.~1999).
Instead, in the last three years PKS 1510$-$089 showed extreme variability
over the whole electromagnetic spectrum and in particular very intense
$\gamma$-ray activity was detected, intermittently, by the AGILE and {\it
Fermi} satellites. The monitoring of the source between September 2008 and June 2009 with the Large Area Telescope (LAT) onboard {\it Fermi}
is summarized in Abdo et al.~(2010a). The Gamma-Ray Imaging Detector (GRID) onboard AGILE detected flaring episodes in August 2007 (Pucella
et al.~2008) and March 2008 (D'Ammando et al.~2009a). Subsequently, an extraordinary $\gamma$-ray activity was detected in March 2009, with several
flaring episodes and a flux that reached $\sim$ 700 $\times$ 10$^{-8}$ photons
cm$^{-2}$ s$^{-1}$. Preliminary results were presented in D'Ammando et
al.~(2009b), Pucella et al.~(2009), Vercellone et al.~(2009). Recently,
Marscher et al.~(2010) studying the flux behaviour at different frequencies,
linear optical polarization, and parsec-scale structure of PKS 1510$-$089
found a correspondence between the rotation of the optical polarization angle and the $\gamma$-ray activity during the first six months of 2009.
\noindent In this paper we discuss the results of the analysis of the multiwavelength data of PKS 1510$-$089 collected by GLAST-AGILE Support
Program (GASP) of the Whole Earth Blazar Telescope (WEBT), Rapid Eye Mount
(REM), {\it Swift} and AGILE during March 2009 and their implications for the
emission mechanisms at work in this source. The paper is organized as
follows. Section 2 briefly introduces the multiwavelength coverage on PKS
1510$-$089. In Sections 3 through 7 we present the analysis and results of
AGILE, {\it Swift}, GASP-WEBT, REM, and Very Long Baseline Array (VLBA) data,
respectively. In Section 8 we discuss the contribution of the thermal emission
in the low-energy part of the spectrum and the correlation between the
emission at different energy bands. Finally, in Section 9 we draw our conclusions.
Throughout this paper the quoted uncertainties are given at 1-$\sigma$ level,
unless otherwise stated, and the photon indices are parameterized as $N(E)
\propto E^{-\Gamma}$ (ph cm$^{-2}$ s$^{-1}$ keV$^{-1}$ or MeV$^{-1}$ ) with
$\Gamma = \alpha +1$ ($\alpha$ being the spectral index).
\section{The multiwavelength coverage}
The high $\gamma$-ray activity observed by AGILE during March 2009 triggered
15 {\it Swift} ToO observations, starting from 11 March 2009, for a total of
46 ks. Moreover, the monitoring by GASP--WEBT in February--March 2009 provided
important information from radio-to-optical band and, together with the data
collected in near-IR and optical bands by REM during March 2009, allowed us to
also obtain an excellent coverage in the low-energy part of the broad band
spectrum of the source. Finally, high resolution radio VLBI data at 15 GHz
were obtained by monitoring the source within the Monitoring of Jets in Active galactic nuclei with VLBA Experiment (MOJAVE) project.
A summary of the complete multiwavelength data set on PKS 1510$-$089 presented in this paper can be found in Table 1.
\begin{table}[!hhh]
\begin{center}
\begin{tabular}{ccc}
\hline \hline
\label{table:Observatories}
Waveband & Observatory & Frequency/Band \\
\hline
Radio & SMA & 230 GHz \\
& Noto & 43 GHz \\
& Mets\"ahovi & 37 GHz \\
& VLBA & 15, 43 GHz \\
& UMRAO & 8.0, 14.5 GHz \\
& Medicina & 5, 22 GHz \\
\hline
Near-IR & REM & $J$, $H$, $K$ \\
& Campo Imperatore & $J$, $H$, $K$ \\
& Roque (Liverpool) & $H$ \\
\hline
Optical & Abastumani & $R$ \\
& Calar Alto & $R$ \\
& Castelgrande & $B$, $V$, $R$, $I$ \\
& L'Ampolla & $R$ \\
& La Silla (MPG/ESO) & $B$, $V$, $R$, $I$ \\
& Lulin (SLT) & $R$ \\
& Roque (KVA and Liverpool) & $R$ \\
& San Pedro Martir & $R$ \\
& St. Petersburg & $B$, $V$, $R$, $I$ \\
& Valle d'Aosta & $R$ \\
& REM & $V$, $R$, $I$ \\
& {\it Swift}/UVOT & $v$, $b$, $u$ \\
\hline
UV & {\it Swift}/UVOT & $uvw1$, $uvm2$, $uvw2$ \\
\hline
X-ray & {\it Swift}/XRT & 0.3--10 keV \\
& {\it Swift}/BAT & 15--50 keV \\
\hline
Gamma-ray & AGILE GRID & 100 MeV -- 30 GeV \\
\hline \hline
\end{tabular}
\caption{Observatories contributing to the presented data set of PKS 1510$-$089 at different frequencies.}
\end{center}
\end{table}
\section{AGILE observations: data analysis and results}
The AGILE scientific instrument (Tavani et al.~2009) is very compact and combines four detectors that provide broad band coverage from hard X-rays to
$\gamma$-rays: a Silicon Tracker (ST) optimized for $\gamma$-ray imaging in
the 30 MeV -- 30 GeV energy range (Prest et al.~2003; Barbiellini et
al.~2001); a co-aligned coded-mask X-ray imager sensitive in the 18--60 keV
energy range (SuperAGILE; Feroci et al.~2007); a non-imaging CsI(Tl)
Mini-Calorimiter sensitive in the 0.3--100 MeV energy range (MCAL; Labanti et
al.~2009); and a segmented anticoincidence system (ACS; Perotti et
al.~2006). The combination of ST, MCAL and ACS forms the GRID that assures the
$\gamma$-ray detection.
The AGILE satellite observed PKS 1510$-$089 between 1 March 2009 00:01 UT and
31 March 2009 11:41 UT (JD 2454891.5-2454921.0), for a total of 325 hours of
effective exposure time, during which the source moved from $\sim 25^\circ$ to
$\sim 55^\circ$ off the AGILE pointing direction.
AGILE-GRID data were analyzed, starting from the Level--1 data, using the AGILE Standard Analysis Pipeline and the AGILE Scientific Analysis
Package. Counts, exposure, and Galactic background $\gamma$-ray maps, the latter based on the diffuse emission model developed for AGILE (Giuliani
et al.~2004), were generated with a bin size of $0.25^{\circ} \times 0.25^{\circ}$ for photons with energies E $>$ 100 MeV. We selected only the events
flagged as confirmed $\gamma$-ray events, and not collected during the South Atlantic Anomaly or whose reconstructed directions form angles with
the satellite-Earth vector smaller than 80$^{\circ}$, in order to reduce the $\gamma$-ray Earth albedo contamination.
PKS 1510$-$089 was detected over the period 1--30 March 2009 at a significance
level of 19.9-$\sigma$, with an average $\gamma$-ray flux of $F_{\rm
E\,>\,100\, \rm MeV}$ = (162 $\pm$ 12) $\times$ 10$^{-8}$ photons cm$^{-2}$
s$^{-1}$. Figure~\ref{AGILE2009} shows the $\gamma$-ray light curve of March
2009 with 1-day resolution for photons with energy higher than 100 MeV. The
average $\gamma$-ray flux as well as the daily values were derived by the
maximum likelihood analysis, according to the procedure described in Mattox
et al.~(1993): first, the entire period was analyzed to determine the
diffuse emission parameters, then the source flux density was estimated
independently for each of the 1-day periods with the diffuse parameters
fixed at the value obtained in the first step. At the beginning of the
observation period the source was not so active in $\gamma$-rays as
later. AGILE-GRID does not detect PKS 1510$-$089 at a significance level
higher than 3-$\sigma$ between 1 and 8 March 2009 and only upper limits at
95$\%$ confidence level are obtained\footnote{When the significance level of
the detection obtained by the maximum likelihood analysis is $<$ 3-$\sigma$
an upper limit at 95$\%$ confidence level is calculated (see Mattox et
al.~1996).}. Instead, in the period 9--30 March 2009 (JD
2454899.5--2454921.5) different flaring episodes were detected and
considering only this time interval the significance of detection slightly
increases to 21.5-$\sigma$, with an average flux of (311 $\pm$ 21) $\times$
10$^{-8}$ photons cm$^{-2}$ s$^{-1}$.
In the period 9--30 March 2009 the AGILE 95$\%$ maximum likelihood contour
level barycenter of the source is l = 351.278$^{\circ}$ b = 40.058$^{\circ}$,
with a distance between the $\gamma$-ray position and the radio position (l =
351.289$^{\circ}$, b = 40.139$^{\circ}$) of 0.08$^\circ$ and an overall AGILE
error circle of radius r = 0.15$^{\circ}$.
The $\gamma$-ray light curve shows a complex structure with outbursts having a
duration of 4--5 days and approximately symmetrical profile that could be an
indication that the relevant timescale is the light crossing time of the
emitting region. The peak level of activity with daily integration was $F_{\rm
E\,>\,100\, \rm MeV}$ = (702 $\pm$ 131) $\times$ 10$^{-8}$ photons cm$^{-2}$
s$^{-1}$ on 25 March 2009 (JD 2454916.0)\footnote{Integrating between 0:00 UT
and 23:59 UT for each day.}. This is the highest $\gamma$-ray flux observed
from this source on daily timescale and one of the highest fluxes detected
from a blazar. We note that the increasing of the error on the flux estimation
throughout the whole observation period is related to the fact that with the
increase of the off-axis angle between the center of the field of view of the
GRID and the position of the source, the possible background contamination and
the GRID calibration uncertainties become slightly higher. Despite the
different energy range and T$_{start}$ of the daily light curves, the AGILE results are in agreement with the {\it Fermi}-LAT results presented in Abdo et al.~(2010a).
\begin{figure*}[!hhht]
\centering
\includegraphics[width=14cm]{1510_lc_Mar09_new.eps}
\caption{AGILE $\gamma$-ray light curve of PKS 1510--089 between 1 and 30
March 2009 (JD 2454891.5--2454921.5) for E $>$ 100 MeV at 1-day
resolution. The downward arrows represent 2-$\sigma$ upper limits.}
\label{AGILE2009}
\end{figure*}
Unfortunately, the source was located substantially off-axis in the SuperAGILE
field of view during the whole observation period, thus preventing us to
extract a constraining upper limit in the 18--60 keV energy band with SuperAGILE.
The $\gamma$-ray spectrum during the first period of high activity detected by
AGILE (9--16 March 2009; JD 2454899.5--2454907.5) can be fitted with a power
law of photon index $\Gamma$ = 1.95 $\pm$ 0.15. The photon index is calculated
with the weighted least squares method by considering four energy bins:
100--200 MeV, 200--400 MeV, 400 MeV--1 GeV, and 1--3 GeV. A similar value of
the photon index ($\Gamma$ = 1.96 $\pm$ 0.24) is estimated also in the period
of the highest activity (23--27 March 2009; JD
2454913.5--2454918.5). Considering that during this second period the source
was far off-axis with respect to the center of the GRID field of view, the
energy bin 1--3 GeV is not considered in order to reduce the possible
background contamination for high off-axis angles at these energies. Comparing
with the average photon index measured by EGRET ($\Gamma$ = 2.47 $\pm$ 0.21,
Hartman et al.~1999) and that measured by {\it Fermi}-LAT over the first
eleven months of observations ($\Gamma$ = 2.41 $\pm$ 0.01; Abdo et al.~2010b),
the values observed by AGILE confirm a hardening of the $\gamma$-ray spectrum
during the very high activity detected in March 2009. A similar hard
$\gamma$-ray spectrum of PKS 1510$-$089 was observed by AGILE during the flaring episode of March 2008 (D'Ammando et al.~2009a).
The correlation between the flux level and the spectral slope in the
$\gamma$-ray energy band was extensively studied by means of the analysis of
all the EGRET data, but a firm result for the blazars is not found
(Nandikotkur et al.~2007). A `harder-when-brighter' behaviour in $\gamma$ rays
of PKS 1510$-$089 was detected by {\it Fermi}-LAT over a long timescale
(August 2008--June 2009) considering only flux levels of F (E $>$ 200 MeV)
$\gtrsim$ 2.4 $\times$ 10$^{-7}$ photons cm$^{-2}$ s$^{-1}$ (Abdo et
al. 2010a). A similar trend was observed in the EGRET {\it era} only for 3C
279 (Hartman et al.~2001) and marginally for PKS 0528$+$134 (Mukherjee et
al.~1996), but this could be due to the fact that these were the only objects
for which a long-term monitoring in $\gamma$-rays was performed by EGRET. At
present, with AGILE and {\it Fermi} we are able to follow a large number of
blazars on very long timescales and investigate this behaviour on a larger
sample of objects. In this context, considering the long period of high
activity recently observed in $\gamma$-rays, PKS 1510$-$089 is one of the best test cases, together with 3C 454.3 (see e.g. Vercellone et al.~2010).
\section{{\it Swift} observations: data analysis and results}
The {\it Swift} satellite (Gehrels et al.~2004) performed 15 ToOs on PKS
1510$-$089 between 11 and 30 March 2009, triggered by the high $\gamma$-ray
activity of the source. The observations were performed with all the three
onboard instruments: the X-ray Telescope (XRT; Burrows et al.~2005, 0.2--10.0
keV), the UltraViolet Optical Telescope (UVOT; Roming et al.~2005, 170-600
nm), and the Burst Alert Telescope (BAT; Barthelmy et al.~2005, 15--150 keV).
\subsection{{\it Swift}/BAT data}
As part of its normal operations, BAT collects data over a wide area of the sky in its survey mode. The survey data in the 15--50 keV band is
used to produce sky images in which hard X-ray sources can be detected using
the standard {\it Swift} analysis software, following the procedures described
in Krimm et al.~(2008, and the reference therein; see also\footnote{http://swift.gsfc.nasa.gov/docs/swift/results/transients}.).
During March 2009, we detected two short flaring episodes from PKS 1510$-$089. The first covered approximately 2 days beginning on 8 March 2009
(JD 2454899.0), with an average count rate of (0.006 $\pm$ 0.002) counts
s$^{-1}$ cm$^{-2}$ (15--50 keV), which corresponds to 28 mCrab and peaking on
9 March 2009 at 40 mCrab (Krimm et al.~2009). A second weaker episode occurred
on 29 March 2009 (JD 2454920.0), where the average count rate was (0.003 $\pm$
0.001) counts s$^{-1}$ cm$^{-2}$, corresponding to 15 mCrab.
As a comparison we report that during the period before the March 2008
$\gamma$-ray flare detected by AGILE (D'Ammando et al.~2009a), {\it Swift}/BAT
recorded flaring episodes from PKS 1510$-$089 on 16 February 2008 (JD
2454513.0), 21 February 2008 (JD 2454518.0) and 24 February 2008 (JD
2454521.0), with count rates of approximately (0.005 $\pm$ 0.002) counts s$^{-1}$ cm$^{-2}$ on each of those observations.
\subsection{{\it Swift}/XRT data}
The XRT data were processed with standard procedures ({\tt xrtpipeline}
v0.12.4), filtering, and screening criteria by using the {\tt Heasoft} package
(v.6.8). The source count rate was low during the whole campaign (count rate
$<0.5$ counts s$^{-1}$), so we only considered photon counting (PC) data and
further selected XRT event grades 0--12. Pile-up correction was not
required. Source events were extracted from a circular region with a radius
between 15 and 20 pixels (1 pixel $\sim 2.36\arcsec$), depending on the source
count rate, while background events were extracted from a circular region with
radius 40 pixels away from background sources. Ancillary response files were
generated with {\tt xrtmkarf}, and account for different extraction regions,
vignetting and PSF corrections. We used the spectral redistribution matrices
v011 in the Calibration Database maintained by HEASARC. Considering the very
low number of photons collected on 26 and 27 March we summed the data of these
two consecutive days in order to have enough statistics to obtain a good spectral fit.
All spectra were rebinned with a minimum of 20 counts per energy bin to allow
$\chi^2$ fitting within {\sc XSPEC} (v12.5.1; Arnaud 1996). We fit the
individual spectra with a simple absorbed power law, with a neutral hydrogen
column fixed to its Galactic value ($6.89\times10^{20}$ cm$^{-2}$; Kalberla et al.~2005).
The fit results are reported in Table~\ref{XRT_March2009}. The comparison between our results and those presented by Abdo et al.~(2010a) reveals some
differences, which we ascribe to their using a lower number of counts per bin than generally adopted to allow $\chi^{2}$ fitting ($\geq$ 20
counts/bin). Indeed, these differences increase when the number of counts decreases.
\begin{table*}[th!]
\caption{Log and fitting results of {\it Swift}/XRT observations of PKS 1510$-$089 during March 2009. Power law model with $N_{\rm
H}$ fixed to Galactic absorption (6.89 $\times$ 10$^{20}$ cm$^{-3}$) is used. $^{a}$ Unabsorbed flux.}
\centering
\begin{tabular}{cccccc}
\hline
\hline
\noalign{\smallskip}
\multicolumn{1}{c}{Start Time} &
\multicolumn{1}{c}{Stop Time} &
\multicolumn{1}{c}{Exp. Time} &
\multicolumn{1}{c}{Photon Index} &
\multicolumn{1}{c}{Flux 0.3--10.0 keV$^{a}$} &
\multicolumn{1}{c}{$ \chi^{2}_{\rm red}$ (d.o.f.)} \\
\multicolumn{1}{c}{(yyyy-mm-dd hh:mm:ss)} &
\multicolumn{1}{c}{(yyyy-mm-dd hh:mm:ss)} &
\multicolumn{1}{c}{ (sec) } &
\multicolumn{1}{c}{$\Gamma$}&
\multicolumn{1}{c}{($\times$ 10$^{-12}$ erg cm$^{-2}$ s$^{-1}$}) &
\multicolumn{1}{c}{} \\
\hline
\noalign{\smallskip}
2009-03-11 15:07:49 & 2009-03-11 23:25:56 & 4891 & $1.54 \pm 0.10$ & $9.28 \pm 0.77$ & 0.919 (35) \\
2009-03-12 15:13:44 & 2009-03-12 20:29:50 & 4842 & $1.45 \pm 0.09$ & $10.86 \pm 0.84$ & 0.867 (38)\\
2009-03-17 04:29:17 & 2009-03-17 11:07:56 & 4869 & $1.51 \pm 0.11$ & $8.67 \pm 0.77$ & 1.041 (31) \\
2009-03-18 00:02:10 & 2009-03-18 08:03:56 & 4777 & $1.37 \pm 0.10$ & $9.37 \pm 0.88$ & 1.009 (27) \\
2009-03-19 19:14:17 & 2009-03-19 22:39:56 & 2501 & $1.61 \pm 0.14$ & $8.82 \pm 0.99$ & 1.074 (15) \\
2009-03-20 22:34:29 & 2009-03-21 02:00:58 & 2010 & $1.28 \pm 0.18$ & $10.03 \pm 1.33$ & 0.884 (11) \\
2009-03-22 01:30:19 & 2009-03-22 06:38:57 & 2242 & $1.43 \pm 0.18$ & $9.51 \pm 1.38$ & 0.843 (12) \\
2009-03-22 03:05:15 & 2009-03-22 03:48:27 & 2580 & $1.50 \pm 0.17$ & $8.59 \pm 1.14$ & 0.913 (13) \\
2009-03-23 11:25:24 & 2009-03-23 16:32:57 & 2640 & $1.64 \pm 0.15$ & $7.77 \pm 0.91$ & 0.705 (15) \\
2009-03-24 03:27:22 & 2009-03-24 22:57:57 & 1972 & $1.60 \pm 0.21$ & $7.97 \pm 1.26$ & 0.768 (10) \\
2009-03-25 13:23:14 & 2009-03-25 22:58:57 & 2447 & $1.53 \pm 0.14$ & $9.50 \pm 1.13$ & 1.102 (15) \\
2009-03-26 16:27:40 & 2009-03-27 08:43:56 & 5003 & $1.50 \pm 0.13$ & $8.39 \pm 0.89$ & 0.871 (22) \\
2009-03-28 05:18:45 & 2009-03-28 08:48:58 & 2657 & $1.31 \pm 0.14$ & $10.16 \pm 1.30$ & 0.997 (17) \\
2009-03-30 10:18:41 & 2009-03-30 12:12:57 & 2544 & $1.40 \pm 0.14$ & $9.93 \pm 1.25$ & 1.230 (16) \\
\noalign{\smallskip}
\hline
\hline
\noalign{\smallskip}
\end{tabular}
\\
\label{XRT_March2009}
\end{table*}
During the 15 ToOs performed in March 2009, {\it Swift}/XRT observed the
source with a 0.3--10.0 keV flux in the range (0.8 -- 1.1) $\times$ 10$^{-11}$
erg cm$^{-2}$ s$^{-1}$, a somewhat fainter state with respect to the high flux
level observed in August 2006 (Flux$_{\rm 0.5-10\,keV}$ = (1.1 -- 1.8)
$\times$ 10$^{-11}$ erg cm$^{-2}$ s$^{-1}$). Instead, a hint of spectral trend
seems to be present in the X-ray data. Figure~\ref{XRT:hwb} shows the XRT
photon indexes as a function of the X-ray fluxes in the 0.3--10 keV: a possible hardening of the spectrum with the increase of the flux is observed, in
spite of the large errors. A similar harder-when-brighter behaviour in X-rays
was observed for PKS 1510$-$089 in August 2006 (Kataoka et al.~2008){\bf ,}
March 2008 (D'Ammando et al.~2009a), and over the period January--June 2009 (Abdo et al.~2010a).
The significant spectral evolution observed by XRT, with the photon index
changing from 1.3 to 1.6, is not usual for FSRQs, for which only little X-ray
variability is detected both on short and long timescales. Moreover, the
average X-ray photon index observed over March 2009 for this object
($<\Gamma>$ = 1.48 $\pm$ 0.03) is lower than that of the radio-loud quasars ($<\Gamma>$ = 1.66 $\pm$ 0.07; Lawson et al.~1992, Cappi et al.~1997) and is
more similar to that of the high-redshift (z $>$ 2) quasars (e.g. Page et
al.~2005). Since usually in FSRQs the X-ray energy range samples the
low-energy tail of the external Compton (EC) component, as already discussed
in Kataoka et al.~(2008) and Abdo et al.~(2010a), such hard spectral indexes
in X-rays should imply a very flat energy distribution, challenging the
standard shock models of particle acceleration. In a standard shock model
these hard X-ray spectra produce relativistic electrons with an energy
spectrum harder than the canonical power law distribution N($\gamma$)
$\propto$ $\gamma$$^{-2}$, or alternatively, as discussed in Sikora et
al.~(2002), another mechanism (e.g.~instabilities driven by shock-reflected
ions, Hoshino et al.~2002; or magnetic reconnection, Romanova $\&$ Lovelace
1992) should energize the low energies electrons that typically produce the X-ray emission in the EC model.
\begin{figure}[!t]
\epsfig{file=1510_hwb_v2.eps, width=90mm}
\caption{{\it Swift}/XRT photon index as a function of the 0.3--10 keV flux.}
\label{XRT:hwb}
\end{figure}
\subsection{{\it Swift}/UVOT data}
UVOT observed PKS 1510$-$089 in all its optical ($v$, $b$, and $u$) and UV
($uvw1$, $uvm2$, and $uvw2$) photometric bands. Data were reduced with the
{\tt uvotmaghist} task of the HEASOFT package. Source counts were extracted
from a circular region of 5 arcsec radius, centred on the source, while the
background was estimated from a surrounding annulus with 8 and 18 arcsec
radii. In Fig.~\ref{WEBT2009} UVOT magnitudes are displayed with blue circles.
\begin{figure*}[!hhht]
\sidecaption
\includegraphics[width=11.0cm]{uvot.eps}
\caption{Light curves collected in near-IR, optical and UV bands, between 5
March and 2 April 2009 (JD 2454895.5--2454924.5). Blue circles represent the
UVOT data in $v$, $b$, $u$, $uvw1$, $uvm2$, $uvw2$ filters. Red diamonds
represent GASP data in $R$, $J$, $H$, and $K$ bands. Magenta triangles
represent WEBT data in $I$, $V$, and $B$ \,bands. REM data in $V$, $R$, $I$,
$J$, $H$, and $K$ bands are represented with black squares. Yellow regions
in $R$-band light curve indicate the peaks of $\gamma$-ray activity observed by AGILE.}
\label{WEBT2009}
\end{figure*}
\section{GASP observations in radio, near-IR and optical bands: data analysis and results}
The GLAST-AGILE Support Program (Villata et al.~2008, 2009a) is a project born
from the Whole Earth Blazar Telescope\footnote{\tt
http://www.oato.inaf.it/blazars/webt.} in 2007. It is aimed at providing long-term monitoring in the optical ($R$-band), near-IR, and radio bands of 28
$\gamma$-ray-loud blazars during the lifetime of the two $\gamma$-ray satellites.
The $R$-band GASP observations of PKS 1510$-$089 in the period considered in
this paper were performed by the following observatories: Abastumani, Calar
Alto\footnote{Calar Alto data were acquired as part of the MAPCAT project
http://www.iaa.es/~iagudo/research/MAPCAT.}, Castelgrande, L'Ampolla, La
Silla (MPG/ESO), Lulin (SLT), Roque de los Muchachos (KVA and Liverpool), San
Pedro Martir, St. Petersburg, and Valle d'Aosta. These data were calibrated
with respect to stars 2--6 by Raiteri et al.~(1998). The GASP observation in
$R$-band showed that after a low intensity period in February 2009, with the
source observed constantly at $\sim$ 16.5 mag, the optical activity of the
source greatly increased in March 2009 (Villata et al.~2009b) with an intense
flare on 27 March (JD $\sim$ 2454917.6), after a brightening by almost 1 mag
in 3 days, reaching $R$ = 14.35 $\pm$ 0.03 mag (Larionov et al.~2009; see
Fig.~3 and Fig.~4). Near-IR data in $J, H$, and $K$ \,bands are from Campo
Imperatore and Roque de los Muchachos (Liverpool), whereas WEBT data in $B$,
$V$, and $I$ \,bands were taken at Castelgrande, La Silla and
St.~Petersburg. Data collected by GASP and WEBT observatories during March
2009 are reported in Fig.~\ref{WEBT2009} with red diamonds and magenta triangles, respectively.
Radio fluxes were measured at: Submillimeter Array\footnote{These data were obtained as part of the normal monitoring program initiated by the SMA (see
Gurwell et al.~2007).} (SMA, 230 GHz), Medicina (22 and 5 GHz), Mets\"ahovi
(37 GHz), Noto (43 GHz), and UMRAO (14.5 and 8.0 GHz). The radio data are
shown in Fig.~\ref{radio}, together with the optical ($R$-band) and near-IR ($H$-band) light curve of PKS 1510$-$089 in February--March 2009.
We point out that in those cases where different datasets were present in the
same band, we performed a careful data analysis to determine possible
offsets and corrected for them to get homogeneous light curves, as usually
done for GASP--WEBT data (see e.g.~Villata et al.~2002, Raiteri et
al.~2005).
\begin{figure}[!hhh]
\centering
\includegraphics[width=9.0cm]{radiop.eps}
\caption{$R$-band light curve of PKS 1510$-$089 obtained by GASP and REM
during February--March 2009, together with the $H$ band data by GASP (top panel), compared with the radio flux densities at different frequencies (bottom panel).}
\label{radio}
\end{figure}
\section{REM observations: data analysis and results}
During March 2009 photometric near-IR and optical observations were carried
out with REM (Zerbi et al.~2004), a robotic telescope located at the ESO Cerro
La Silla observatory (Chile). The REM telescope has a Ritchey-Chretien
configuration with a 60 cm f/2.2 primary and an overall f/8 focal ratio in a
fast moving alt-azimuth mount providing two stable Nasmyth focal stations. At
one of the two foci, the telescope simultaneously feeds, by means of a
dichroic, two cameras: REMIR for the near-IR (Conconi et al.~2004) and ROSS
(Tosti et al.~2004) for the optical. Both cameras have a field of view of 10
$\times$ 10 arcmin and imaging capabilities with the usual near-IR (z, $J$,
$H$, and $K$) and Johnson-Cousins $VRI$ filters. The REM software system
(Covino et al.~2004) is able to manage complex observational strategies in a
fully autonomous way. All raw near-IR/optical frames obtained with the REM
telescope were then corrected for dark, bias, and flat field following
standard recipes. Instrumental magnitudes were obtained via aperture
photometry and absolute calibration has been performed by means of secondary
standard stars in the field (see Raiteri et al.~1998; Smith and Balonek
1998). The data presented here were obtained by GO program for AOT19 (PI:
F.~D'Ammando). REM data in $V$, $R$, $I$, $J$, $H$, and $K$ \,bands are
reported in Fig.~\ref{WEBT2009} with black squares.
\section{VLBA observations: data analysis and results}
When observed with the high spatial resolution provided by the Very Long
Baseline Interferometer (VLBI) technique, PKS 1510$-$089 shows a radio
structure dominated by the core region from which the pc-scale jet emerges
forming an angle of $-$28$^{\circ}$, i.e. in the north-west
direction. Monitoring campaigns pointed out the ejection of new superluminal components, and followed their evolution throughout several observing epochs
(see Marscher et al.~2010). \\
As part of the MOJAVE program (Lister et al.~2009), PKS 1510$-$089 is
frequently observed with the VLBA at 15 GHz (see Fig.~\ref{VLBA}). To study
variations in the source structure, we retrived the {\it uv}-datasets and we
imported them into NRAO AIPS package. Final images both in total intensity and
polarization have been produced after a few phase-only self-calibration
iterations (for more details on calibration and image analysis see Orienti et
al.~2010). From the analysis of the multi-epoch observations obtained between
2007 and 2010 it was possible to follow the separation between the core
region, considered stationary, and three emitted knots. From linear regression
fits we obtain highly superluminal apparent separation velocities between
16$c$ and 20$c$, consistent with the separation speed derived for other knots
(e.g.~Homan et al.~2002), and at different frequency (Marscher et al.~2010),
indicating an intrinsic separation velocity $\beta_{\rm intr} >0.9989$ and an orientation with the line of sight $\theta < 5^{\circ}$.\\
From the pc-scale resolution images we could separate the core flux density
from the jet emission, in order to study the lightcurve and polarization
properties of each of them. Between 2007 and 2010, three main episodes of
enhanced radio luminosity have been detected, successively followed by marked
flux density decreases. By contrast, no obvious trend of the polarization
percentage has been found.\\
After March 2009, the flux density of the core region, that was in a minimum,
started to increase, while the flux density of the jet component did not show
any significant luminosity changes. A remarkable aspect observed just after
the March 2009 flare is the abrupt change of about 75$^{\circ}$ in the
polarization angle of the core, whereas the fractional polarization remains
constant. More details on the long-term radio monitoring of this source are presented in Orienti et al.~(2010).\\
\begin{figure}[!hhht]
\center
\includegraphics[width=9cm]{1510mar09_VLBA.eps}
\caption{VLBA image at 15 GHz of PKS 1510$-$089 on 25 March 2009. On the image
we provide the restoring beam, plotted in the bottom left corner, the peak flux density in mJy beam$^{-1}$, and the first contour (f.c.) intensity in mJy beam$^{-1}$, which is 3 times the off-source noise level. Contour levels increase by a factor of 2.}
\label{VLBA}
\end{figure}
\section{Discussion}
\subsection{Thermal signatures in the radio-to-UV spectrum}
Evidence of thermal emission in blazars has been detected in a limited number
of objects until now, and usually observed during low activity states (e.g.~3C
279, Pian et al.~1999; 3C 273, Grandi $\&$ Palumbo 2004; 3C 454.3, Raiteri et
al.~2007; AO 0235$+$164, Raiteri et al.~2008; NRAO 150, Acosta-Pulido et
al.~2010). This is because, contrary to the other AGNs, in blazars the jet
emission at small angles with respect to the line of sight of the observer is strongly amplified due to the relativistic beaming effect, overwhelming all
the thermal contributions. However, these thermal features are less prominent
in blazars but not absent, at least for FSRQs, which present a radiatively
efficient accretion disk and broad emission lines.
The knowledge of these thermal features could have important consequences not
only for the low-energy part of the spectrum, in which the thermal emission
could be directly observed, but also for the high-energy part of the spectrum,
because the photons produced by the accretion disk, directly or through the
reprocessing by the broad line region (BLR) or the dusty torus, can become the
main source of seed photons for the EC mechanism that usually dominates the $\gamma$-ray emission of FSRQs (Ghisellini et al.~1998).
In this context PKS 1510$-$089 seems to be peculiar. Since the synchrotron
peak of this source is usually observed in mid-infrared band (see Bach et
al.~2007; Nieppola et al.~2008), the optical/UV part of the spectrum is not
dominated by the non-thermal jet emission, allowing us to observe directly the thermal manifestation of both the accretion disk (the so-called ``big blue
bump'', e.g.~Laor 1990) and of the BLR (the so-called ``little blue bump'',
e.g.~Wills et al.~1985) also during high activity states, like in mid-March
2008 (see D'Ammando et al.~2009a). Moreover, an excess at far-IR wavelengths
was observed in PKS 1510$-$089 by IRAS (Tanner et al.~1996), likely due to
dust radiation from the torus. Recently, {\it Spitzer} observations of PKS
1510$-$089 produced only an upper limit of 2 $\times$ 10$^{45}$ erg s$^{-1}$
to the thermal emission from dust torus (Malmrose et al.~2011).
In order to distinguish between non-thermal and thermal emission
contributions, we built spectral energy distributions (SEDs) from radio to UV
with GASP-WEBT, REM and {\it Swift}/UVOT data. Conversion of magnitudes into
de-reddened flux densities was obtained by adopting the Galactic absorption
value $A_{B}$ = 0.416 from Schlegel et al.~(1998), the extinction laws by
Cardelli et al.~(1989) and the magnitude-flux calibrations by Bessell et al.~(1998).
As for the {\it Swift}/UVOT data, we noticed that PKS 1510$-$089 has a {\it
b-v} $\sim$ 0.3 that is out of the validity range indicated by Poole et
al.~(2008) for their flux calibrations in the UV bands. Hence, following
Raiteri et al.~(2010, 2011) we calculated effective wavelengths
$\lambda_{\rm eff}$, count rate to flux density conversion factors $\rm CF$,
and amount of Galactic extinction $A_\Lambda$ for each UVOT band, by folding
the quantities of interest with the source spectrum and effective areas of UVOT filters. The results are shown in Table 3.
\begin{table}
\caption{Results of the UVOT calibration procedure for PKS 1510-089.}
\label{caluvot} \centering \begin{tabular}{l c c c}
\hline\hline Filter & $\lambda_{\rm eff}$ & $\rm CF$ & $A_\Lambda$ \\
& \AA\ & $10^{-16} \rm \, erg \, cm^{-2} \, s^{-1} \, \AA^{-1}$ & mag \\
\hline $v$ & 5422 & 2.61 & 0.33\\ $b$ & 4346 & 1.47
& 0.43\\
$u$ & 3466 & 1.65 & 0.52\\
$uvw1$ & 2633 & 4.21 & 0.73\\
$uvm2$ & 2251 & 8.45 & 0.90\\
$uvw2$ & 2059 & 6.31 & 0.86\\
\hline\hline \end{tabular}
\end{table}
\begin{figure}[!hhhb]
\centering
\includegraphics[width=9.1cm]{sed_1510_calib.eps}
\caption{SEDs of the low-energy part of the spectrum constructed with data
collected by GASP-WEBT (empty circle), KVA (plus sign), Abastumani (cross),
St.~Petersburg (asteriks), Campo Imperatore (diamonds), REM (squares), and
{\it Swift}/UVOT (filled circles) during March 2008 and March 2009. A cubic
polynomial fit was applied to the SEDs for locating the position of the synchrotron peak.}
\label{SEDoptical}
\end{figure}
UVOT count rates were thus converted into flux densities and then corrected
for Galactic extinction adopting the $\rm CF$ and $A_\Lambda$ reported in
Table 3. In Figure 6 we display three SEDs of March 2009 as well as the March
2008 SED already published by D'Ammando et al.~(2009a). The comparison between
the March 2008 SED obtained with standard calibrations (Figure 3 of D'Ammando
et al.~2009a) and the same SED obtained by applying the procedure described above reveals that the dip in the $uvw1$ band remains, though it is now less
pronounced. This means that it is not due to calibration problems, but likely
reflects a true intrinsic feature. Indeed, we expect the little blue bump due
to BLR emission (MgII, FeII, and Balmer lines) to peak in the $u$--$b$
frequency range. Near-IR to UV spectra of PKS 1510$-$089 in March 2009 have
also been presented by Abdo et al.~(2010a), but for a different choice of
epochs, which prevents a detailed comparison with our results. However, as
expected, we notice a deeper $w1$ dip than that we obtained with our more accurate calibrations.
In other FSRQs, such as 3C 454.3, the presence of the little and big blue
bumps was detected only during low activity state of the source (see Raiteri
et al.~2008, Vercellone et al.~2010). Instead, the SED of PKS 1510$-$089
collected on 18 March 2009 confirmed the evidence of thermal signatures in the
optical/UV spectrum of the source also during high $\gamma$-ray states, with a
contribution in the optical part likely due to the little blue bump and a
significant rise of the spectrum at UV due to the accretion disk emission, as
already observed on 20--22 March 2008. On the other hand, the broad band
spectrum of PKS 1510$-$089 from radio-to-UV during 25--26 March 2009 shows a
flat spectrum in the optical/UV energy band, suggesting an important
contribution of the synchrotron emission in this part of the spectrum during
the brightest $\gamma$-ray flaring episode and therefore a significant shift
of the synchrotron peak, usually observed in this source in the infrared
band. We derived an estimate of the frequency of the synchrotron peak in the
three SEDs applying a cubic polynomial fit (see e.g.~Kubo et al.~1998) to the
radio and IR data, which are likely due to pure synchrotron, as shown in
Fig.~\ref{SEDoptical}. The synchrotron peak shifted from $\nu$ = 1.5 $\times$
10$^{13}$ Hz to $\nu$ = 6.5 $\times$ 10$^{13}$ Hz between 18 and 26 March
2009. Similarly to the harder-when-brighter behaviour observed in X-rays, this
is a typical behaviour of high-frequency-peaked BL Lacs (HBLs) and not so
commonly observed in FSRQs such as PKS 1510$-$089, even if this could be
partially due to the fact that it is more difficult to obtain a long term
monitoring of the synchrotron peak of FSRQs in the IR band with respect to the optical/UV and X-ray band.
Abdo et al.~(2010a) analyzing the UVOT data of PKS 1510$-$089 with standard
calibrations (and thus with possible biases in the results, see above) found
an anticorrelation between UV flux and UV hardness ratio in the period
January-June 2009, suggesting a different level of contamination in UV by
the high-energy branch of synchrotron emission depending on the activity
level. This is in agreement with our results, even if the importance of the
increase of the synchrotron emission observed at the end of March 2009 seems
to be much higher with respect to the general trend.
The UVOT data reported in the SED of 25--26 March 2009 are collected at JD =
2454916.46, instead the BVRI data are collected by St.~Petersburg at JD =
2454916.51-53, thus the separation in time between UVOT and BVRI data is about
1 hour. This rules out the possible bias related to the optical/UV variability
of the object, confirming the flat optical/UV spectrum. The NIR data for the
same SED are collected by REM at JD = 2454916.66, about 3.5 hrs after the
optical data and 4.5 hrs after the UVOT data. In this case it is not possible
to completely rule out a mismatch due to the possible rapid optical/UV
variability, but by comparison with the NIR data collected during 20-22 March
2008 and 18 March 2009, it is evident the change of the NIR spectrum, in
agreement with what we observed also in optical/UV. Moreover the comparison
between the optical/UV data collected by UVOT on 25 March 2009, during the
first UVOT exposure ($\rm JD \sim 2454916.1$, orange squares in
Fig.~\ref{SEDoptical}) and those taken about 9--10 hours after ($\rm JD \sim
2454916.5$, green triangles), reveals a noticeable spectral and flux
variation. Considering that the accretion disk is slowly variable on such
short timescales, this is another proof of the fact that the rapid increase of
the optical emission started on 25 March and peaking on 27 March is mainly due
to the synchrotron mechanism. This is also in full agreement with the
simultaneous rapid increase of the degree of optical polarization shown in
Marscher et al.~2010 (see in particular their Fig.~4). Thermal emission is
unpolarized as it reflects the random walk of atoms and ions within the
emitting region, therefore such increase of the degree of polarization is a
clear signature of a rapid and strong increase of the contribution in the optical band of a non-thermal mechanism such as the synchrotron emission.
\subsection{Light curve behaviour and correlations}
By comparing the source behaviour observed in $\gamma$-rays by AGILE with the
ones observed from optical-to-UV by GASP-WEBT, REM and {\it Swift}/UVOT during
March 2009 we noted that while the $\gamma$-ray light curve shows evidence of
different outbursts of increasing entity, the optical and UV light curves seem
to show a gradual increase of the flux in time with a rapid flux enhancement after JD $\sim$ 2454915.0 and a single major outburst occurred between JD
2454917.25 and JD 2454917.60. A variation of 0.48 mag in 9 days (JD
2454906-2454915) was detected in the $R$-band light curve, followed by a more
rapid variation of 0.95 mag in $\sim$ 2 days. The $uvw1$ light curve shows a
similar behaviour, with a variation of 0.43 mag in 7 days and 0.78 mag in the
following 2 days. The larger increase observed in $R$-band with respect to
$uvw1$ band is in agreement with a larger contribution of the synchrotron
emission in the optical than in the UV band. A linear fit applied to the
$R$-band and $uvw1$-band light curves in Fig.~3 shows a change of slope after
JD = 2454915. The optical/UV peak seems to be delayed with respect to the
brightest $\gamma$-ray peak by 1--2 days, although we cannot exclude a very rapid optical/UV flare occurred during the gap in the optical and UV light curves before JD 2454917.25.
Another possibility is that the delay between optical and $\gamma$-ray
emission is due to our choice of the $T_{0}$ used for building the
$\gamma$-ray daily light curve. To investigate the influence of the $T_{0}$ on
the determination of the $\gamma$-ray peak we constructed three different
light curves, considering an integration interval of 1 day and shifting the
$T_{0}$ of $\Delta$T$_1$ = 8 hours and $\Delta$T$_2$ =16 hours with respect to
our initial choice ($T_{0}$ = 0:00 UT). In Fig.~\ref{corr} the optical
($R$-band), UV ($w1$ filter) and the three $\gamma$-ray light curves built
with the different $T_{0}$ in the period 21--31 March 2009 (labelled as I, II
and III) are shown. For each $\gamma$-ray light curve we assign a weight ``1''
to the three fractional 8-hours bins that constitute the total 1-day bin with
the highest flux and a weight ``0'' to the other fractional bins. Combining
the information derived from the three light curves we can build a probability
function and estimate the fractional bin that corresponds with the highest
probability to the emission peak. In this way we estimated the $\gamma$-ray
emission peak at JD = 2454916.33 $\pm$ 1.00, leading to a reduced delay
between $\gamma$-ray and optical/UV emission, in accordance also with the
delay between the {\it Fermi}-LAT and optical light curve peaks showed in Abdo
et al.~(2010a) for the same period. Moreover, we note that a recent investigation of the ${\it Fermi}$-LAT data of PKS 1510$-$089 collected in
the first half of 2009 by Tavecchio et al.~(2010) showed significant $\gamma$-ray variability on timescales of 6 hours and a flux peak on JD = 2454917.25, compatible with our result.
\begin{figure}[!hhht]
\centering
\includegraphics[width=9.0cm]{1510_mwl.eps}
\caption{Comparison of the $R$-band light curve of PKS 1510$-$089 (panel (a);
GASP-WEBT: red squares, REM: black dots) with the $uvw1$ band data by {\it
Swift}/UVOT (panel (b)), the X-ray data by {\it Swift}/XRT (panel (c)), and
the AGILE $\gamma$-ray light curves in the three different configurations
(panel (d), (e), (f), see Sec. 8.2 for details) during 21--31 March 2009 (JD 2454913.5--2454922.0).}
\label{corr}
\end{figure}
However, a significant increase of the optical/UV emission is detected only for the brightest $\gamma$-ray flare, whereas the optical/UV activity before
JD $\sim$ 2454915 was increasing but only slowly, according to a dominant
contribution of the thermal emission, over which a subsequent rapid
variability of the synchrotron emission is superimposed (see Sec.~8.1). A progressive increase of the source activity, similar to the behaviour of the
optical light curve collected by GASP-WEBT, was observed also in the near-IR
and optical bands by the REM telescope (Fig.~\ref{WEBT2009}), suggesting that
the synchrotron is the dominant mechanism responsible for the flux enhancement
observed from near-IR to UV at the end of March 2009. In this context, the lag
of 13$\pm$1 days between the $\gamma$-ray and the optical peaks reported in
Abdo et al.~(2010a) could suggest a different origin of the dominant component
in optical band during the January and April 2009 flares with respect to the
flare occurred at the end of March 2009. This is another clue in favour of a more complex correlation between the optical and $\gamma$-ray emission in
the FSRQs, in which the optical emission could be due to the contribution of
both the synchrotron emission and the thermal accretion disk emission. Despite
the different dominant contribution in the optical part of the spectrum, a
similar photon index was observed by AGILE in the $\gamma$-ray spectrum during
the first activity period (9--16 March) and the highest activity period (23-27
March), suggesting that the same radiation mechanism is responsible for the
high energy emission of the source over the whole observing period.
\noindent Moreover, it is worth noting that the variability amplitude observed
at the end of March 2009 is similar in optical (a factor of $\sim$3) and
$\gamma$ rays (a factor of $\sim$3--4). If the $\gamma$ rays were produced by
inverse Compton scattering of photons reprocessed by the BLR, as suggested by
Abdo et al.~(2010a), the $\gamma$-ray emission would be proportional to the
number of the emitting electrons, like the synchrotron flux. Therefore, the
nearly linear correlation observed between optical and $\gamma$ rays may be
due to the variation of the electron number in the jet. In this context, the
lag of 1 day of the optical peak with respect to the $\gamma$-ray one could be
only an upper limit due to the lack of data before the observed optical peak.
Compared to the optical and $\gamma$-ray activity, the X-ray flux is not very variable and does not seem to be correlated with the high $\gamma$-ray
activity (see Fig.~7, {\it panel c}). The lack of evident X-ray/$\gamma$-ray
correlation is not so surprising. In fact, Marscher et al.~(2010) already
reported that the long-term X-ray light curve of PKS 1510$-$089 collected
during the 2006--2009 period by {\it Rossi X-ray Timing Explorer} is
significantly correlated with the 14.5 GHz one rather than with the
$\gamma$-ray emission. This evidence could be justified by the fact that the
X-ray photons are likely to originate mostly in the low-energy part of the EC
emission. Therefore the X-ray spectral index seems to reflect the change in
the low-energy tail of the electron distribution, whereas a change in the
high-energy part of the electron distribution should be the driver of the
variation of the $\gamma$-ray flux, as already pointed out by Abdo et
al.~(2010a). Moreover, as already discussed in Kataoka et al.~(2008) and
D'Ammando et al.~(2009a), the spectral evolution of PKS 1510$-$089 could be
justified by the contamination at low energies of a second emission component
with respect to the EC emission. In this case the X-ray spectral shape of the
EC component should remain almost constant, regardless of the activity level,
but its flux should vary. Therefore the amount of contamination of the second
component should change, being more important when the source is fainter and
almost hidden by the EC component when the source is brighter. Possible
origins proposed for this second component are the soft X-ray excess, the bulk
Comptonization or a more significant contribution of the SSC component in any
activity states. A possible hint of increase of activity in X-rays is detected
by XRT on 28 March and this could be an indication of an enhancement of the
contribution of the synchrotron self Compton (SSC) emission, strictly related
to the optical/UV outburst observed on 27 March. The strength of the SSC with
respect to EC emission depends on the ratio between the synchrotron and the
external radiation energy density, as measured in the comoving frame, $U'_{\rm
syn}/U'_{\rm ext}$. Taking into account that, under the assumption that the
dissipation region is within the BLR, $U'_{\rm syn}$ depends on the injected
power, the size of the emission and the magnetic field, a larger magnetic
field can lead to a larger SSC contribution together with an increase and a shift of the synchrotron peak. By contrast, an increase of the SSC
emission should lead to a not negligible contribution of this component, which
usually shows a photon index of $\Gamma$ $\sim$ 2 in X-rays for the FSRQs, and
therefore a softening of the total X-ray spectrum. Instead a hard X-ray
spectrum was observed by XRT ($\Gamma_{x}$ = 1.31 $\pm$ 0.14 on 28 March),
confirming that also during a contemporaneous synchrotron flare the SSC
contribution is not significant in the X-ray spectrum of PKS 1510$-$089, as
already observed by {\it Suzaku} during a low state of the source in 2009 (Abdo et al.~2010c).
An alternative scenario is that the electrons radiating at 14.5 GHz produce
X-rays via IC scattering of the IR photons from a dust torus. This
interpretation favours the ``far dissipation'' scenario recently proposed
again for the emission of FSRQs by Sikora et.~(2008) and Marscher et
al.~(2010), in which the high energy peak is mainly due to the EC scattering
of the IR seed photons produced by the torus. In particular, Marscher et
al.~(2010) identified the dissipation zone with a moving superluminal knot
before and after passing the core, which is interpreted as the first
recollimation shock visible at millimeter wavelengths in the jet.
It is crucial to discriminate between these two scenarios in order to
understand where the plasma blob dissipates in the jet. In fact, the energy
density of the different sources of seed photons varies with the distance from
the central black hole, therefore the dominant population of target photons
depends on the site of the dissipation region that produces the flaring episode. The determination of the location of the emitting region in the jet of
blazars is an open issue, in particular for FSRQs. However, the rapid
variability observed in the $\gamma$-ray light curve of PKS 1510$-$089
collected by AGILE in March 2009 (even more remarkable in the {\it Fermi}-LAT
light curves on timescales of hours, Abdo et al.~(2010a), Tavecchio et
al.~2010) puts severe constraints on the size of the emitting region, leading
to a very compact region difficult to be explained if the $\gamma$-ray
emission is produced far away from the central black hole. This seems to
disfavour the ``far dissipation'' scenario. On the other hand, the recent
detection of this source at TeV energies by H.E.S.S. (Wagner and Behera 2010)
suggests possible complex scenarios. We can envisage that more than one
emission region is simultaneously active and located at different distance
from the central black hole in a spine-sheath (Tavecchio and Ghisellini 2008) or jet-in-jet structure (Giannios et al.~2009).
It is worth noting that the outburst detected by {\it Swift}/BAT on 8--9 March
2009 occurred just at the beginning of the $\gamma$-ray activity observed by
AGILE. On the other hand, no significant activity is detected by BAT simultaneously with the $\gamma$-ray flares detected by AGILE.
Finally, as shown in Figure~\ref{radio}, no clear sign of strong activity in
radio bands is observed during March 2009, although the radio coverage is
sparse. However, the SMA data show an increase of about a factor of two
between February and March 2009 at millimeter wavelengths, similarly to the
optical behaviour even if with a minor variation amplitude. Moreover, the
radio data acquired at Mets\"ahovi at 37 GHz on a longer timescale show that
the increase of activity marginally observed also in Fig.~\ref{radio}, which
starts at the beginning of March 2009 (JD $\sim$ 2454910), has continued
gradually in April--May 2009, reaching a peak flux density of 4.02 Jy on 15
May 2009 (JD 2454969.4; see Abdo et al.~2010a). The increase of flux density
observed at 37 and 230 GHz confirms that the mechanism producing the
$\gamma$-ray flaring events also interested the mm/cm emitting region with
difference in time likely related to opacity effects, as already observed in
March--April 2008 during the period soon after the previous $\gamma$-ray
flare of PKS 1510$-$089 detected by AGILE (D'Ammando et al.~2009a).
In addition, the VLBA data show a minimum of the flux density of the core in
March 2009 and since the rotation of the polarization angle of the core was
very mild in March relative to next 50 days, it is unlikely that the March
$\gamma$-ray flaring episodes are related to the the blob associated with
later flares as in the Marscher et al.~(2010) model.
\section{Summary and concluding remarks}
We reported on the detection by AGILE-GRID of $\gamma$-ray activity from the
FSRQ PKS 1510$-$089 during March 2009. AGILE data showed an amazing
$\gamma$-ray activity, with several flaring episodes and a flux that reached
on daily integration $\sim$ 700 $\times$ 10$^{-8}$ photons cm$^{-2}$ s$^{-1}$,
the highest flux detected from this source until now and one of the highest detected from a blazar.
No significant spectral changes are observed between the different
$\gamma$-ray flares in March 2009, indicating that the same mechanism is
dominant for the high energy emission of the source over all the period.
This $\gamma$-ray activity triggered simultaneous multiwavelength
observations, which provided us a wide dataset for studying the correlation
between the emission properties from radio to $\gamma$ rays. An increasing
activity in near-IR and optical was observed, with a strong flux enhancement
after 24 March and a flaring episode on 26-27 March 2009 observed by GASP-WEBT
and REM, almost simultaneous to the brightest $\gamma$-ray flare. By contrast,
no clear signs of simultaneous high activity were detected from near-IR to UV
during the other $\gamma$-ray flares observed by AGILE.
The broad band spectrum from radio-to-UV collected on 18 March confirmed the evidence of thermal signatures, the little and big blue bumps, in the
optical/UV spectrum of PKS 1510$-$089 also during high $\gamma$-ray states,
and not only during low activity states as observed in other blazars. The
radio-to-UV spectrum on 25-26 March showed a flat spectrum in the optical/UV
energy band, suggesting an important contribution of the synchrotron emission
in this part of the spectrum during the brightest $\gamma$-ray flare and
therefore a significant shift of the synchrotron peak, in agreement with the
rapid variation detected by UVOT and the drastic change of polarization angle observed in VLBA data simultaneously with the optical flare. The optical/UV
spectra presented here clearly show the different contribution of the thermal
and non-thermal components in the various phases of the $\gamma$-ray activity
of the source, with consequent theoretical implications on the modeling of the
broad band spectrum to investigate. It will be important to continue the
monitoring of the source from IR to UV to know if the shift of the synchrotron peak observed in March 2009 is a rare event or not.
The {\it Swift}/XRT observations show no clear correlation between the X-ray and $\gamma$-ray emission. This could be due to the fact
that X-rays and $\gamma$-rays are originated by EC emission of different parts
of the electron distribution with different variability. On the other hand, a
hint of harder-when-brighter behaviour was detected from the XRT observations
suggesting the presence of a second component in the soft X-ray part of the
spectrum that could be associated with the soft X-ray excess rather than to
the variation of the SSC contribution in different activity states of the source.
Moreover, the radio data show that an increase of the activity related to the $\gamma$-ray activity started in March and peaked in April-May 2009,
confirming that the mechanism producing the $\gamma$-ray flare interested also
the mm/cm emitting zone, with some delay likely due to opacity effects.
In conclusion, the $\gamma$-ray emission of PKS 1510$-$089 shows a complex
correlation with the other wavelengths, leading to different scenarios when
modeling its broad band spectrum at different epochs. The multifrequency
observations presented here give new clues, but also offer new questions on
the astrophysical mechanisms at work in this object. Further IR observations
can be important to test the ``far dissipation'' scenario in which the IR
photons are the dominant seed photons for the IC mechanism that produces the
$\gamma$-rays, and in this context the launch of Herschel (Pillbratt et
al.~2010) opened a new window on the infrared Universe and it will be very
important also for the study of blazars (Gonzalez-Nuevo et al.~2010). Finally,
the study of the correlation between the $\gamma$-ray activity and the
variability of optical polarization provided new information on the jet
structure, the location and the causes of the high-energy emission in PKS
1510$-$089. A further piece of the puzzle can be added with the launch of a X-ray polarimeter onboard any future X-ray missions (e.g.~the New Hard
X-ray Mission, Pareschi et al.~2010) giving us a final answer to the real
nature of the X-ray spectrum of this puzzling object.
\begin{acknowledgements}
The AGILE Mission is funded by the Italian Space Agency (ASI) with scientific
participation by the Italian Institute of Astrophysics (INAF) and the Italian
Institute of Nuclear Physics (INFN). We thank the Swift Team for making these
observations possible, particularly the duty scientists and science planners. The Submillimeter Array is a joint project between the Smithsonian
Astrophysical Observatory and the Academia Sinica Institute of Astronomy and Astrophysics and is funded by the Smithsonian Institution and the Academia
Sinica. UMRAO is funded by a series of grants from the NSF and NASA and by the
University of Michigan. This research has made use of data from the MOJAVE
database that is maintained by the MOJAVE team (Lister et al.~2009, AJ, 137, 3718). Acquisition of the MAPCAT data is supported in part by MICIIN (Spain)
grants AYA2007-67627-C03-03 and AYA2010-14844, and by CEIC (Andaluc\'{i}a)
grant P09-FQM-4784. Abastumani Observatory team acknowledges financial support
by the Georgian National Science Foundation through grant GNSF/ST08/4-404. The
Liverpool Telescope is operated on the island of La Palma by Liverpool John
Moores University in the Spanish Observatorio del Roque de los Muchachos of
the Instituto de Astrofisica de Canarias, with funding from the UK Science and
Technology Facilities Council. The Mets\"ahovi team acknowledges the support
from the Academy of Finland to our observing projects (numbers 212656, 210338,
and others). St. Petersburg University team acknowledges support from Russian
RFBR foundation via grant 09-02-00092. This paper is partly based on
observations carried out at the German-Spanish Calar Alto Observatory, which
is jointly operated by the MPIA and the IAA-CSIC. Some of the authors
acknowledge financial support by the Italian Space Agency through contract ASI-INAF I/088/06/0 and I/009/10/0 for the Study of High-Energy Astrophysics.
\noindent {\it Facilities}: AGILE, {\it Swift}, WEBT, REM, UMRAO, and VLBA.
\end{acknowledgements}
|
\section{Introduction}
The effective manipulations by quantum correlations results in significant advantages of the quantum devices (in particular, quantum computers) in comparison with their classical counterparts. However, the problem of correct measure of quantum correlations has not been resolved up to now.
Till recently, the entanglement has been used
as a measure of quantum correlations \cite{W,HW,P,AFOV,DPF}.
However,
it was shown both theoretically \cite{BCJLPS,M}
and experimentally
\cite{LBAW} that some mixed separable states (i.e. states with zero entanglement)
allow one to realize the advantages of quantum algorithms in comparison with their classical analogies and quantum nonlocality
has been observed in the systems without entanglement \cite{BDFMRSSW}.
Such observations suggest us to conclude that the entanglement does not involve all quantum correlations responsible for the advantages of quantum algorithms in comparison with the classical ones.
The concept of
quantum discord, which is intensively developing during last years \cite{OZ,L,ARA}, is based on a separation of a quantum part out of the total mutual information encoded into a bipartite system.
The discord is completely defined by the quantum properties of the system and becomes zero for the classical systems.
The evaluation of the quantum discord is a cumbersome computational problem so that the methods of its
analytical calculations have been developed only for two-qubit systems \cite{L,ARA}. Calculation of the discord
consists of two steps: (a) calculation of the mutual information encoded into two subsystems $A$ and $B$ of the bipartite quantum system and (b) calculation of the classical component of this mutual information.
The second part is the most cumbersome one because it is based on the multiparameter optimization over the von Neumann type of measurements performed on one of the subsystems (say, $B$).
It is seemed out that the quantum discord depends on which subsystem is taken for
measurements { \cite{OZ}.}
{ It will be shown below that the difference between the quantum discords calculated using the projective measurements performed over the different subsystems of a bipartite system (the asymmetry) can be large (about 25$\%$). This means that the quantum discord as a measure of quantum correlations does not provide an unambiguous quantitative characteristic of these correlations. As a result, the quantum discord is insufficient for an estimation of the utility of different materials in quantum devices and requires modification. A possible modification of the quantum discord is suggested in this letter. }
The conditions needed to provide the symmetry of the discord with respect to subsystems $A$ and $B$
have been found only for the particular case of the two-qubit system
\cite{ARA}. These conditions are based on the analysis of the structure of the two-qubit density matrix and have not been related with the physical parameters of the system (such as a temperature and Larmour frequencies).
{ A connection of the quantum discord with the real physical parameters opens a direct way for the experimental investigations of the discord and its asymmetry and allows one to control the quantum discord by means of physical parameters.}
This letter is devoted to the problem of asymmetry of the quantum descord of the two-qubit system with respect to the subsystems $A$ and $B$ chosen for the projective measurements. Considering the simple physical model of two spin-1/2 particles with the dipole-dipole interaction governed by the XY Hamiltonian in the inhomogeneous magnetic field we study the dependence of the above asymmetry on the temperature and Larmour frequencies.
\section{Model}
\label{Section:model}
We consider the quantum system of two spin-1/2 particles governed by the XY Hamiltonian in the external inhomogeneous magnetic field. The Hamiltonian of the system reads:
\begin{multline}
\label{XY}
H={D} (I_{1,x} I_{2,x}+I_{1,y} I_{2,y} ) +D\Omega (1+ \Delta) I_{1,z} +\\
D\Omega(1- \Delta) I_{2,z},
\end{multline}
where $D$ is the constant of the dipole-dipole interaction, $I_{n,\alpha}$ ($n=1,2$, $\alpha=x,y,z$)- $n$th spin projection operator on the axis $\alpha$, $D\Omega(1\pm\Delta)$ are the Larmour frequencies.
We assume that $D>0$, $0\le \Delta\le 1$ and $\Omega\ge 0$.
The system is
in the thermal equilibrium state with the Gibbs density matrix:
\begin{eqnarray}\label{rho}
\rho = e^{-\beta H} /Z,\;\;Z={\mbox{Tr}}\; e^{-\beta H},\;\;\beta = \frac{\hbar}{kT},
\end{eqnarray}
where $T$ is the temperature, $k$ is the Boltsman constant.
It is simple to check that this density matrix has the following form:
\begin{eqnarray}\label{rhog}
\rho=\left(\begin{array}{cccc}
\rho_{11}&0&0&0\cr
0&\rho_{22}&\rho_{23}&0\cr
0&\rho_{23}&\rho_{33}&0\cr
0&0&0&\rho_{44}
\end{array}
\right),
\end{eqnarray}
where all entries are real numbers.
The density matrix (\ref{rhog}) can be considered as
a particular case of so-called two-qubit X-matrix \cite{ARA}. The quantum discord for such matrices can be calculated by the method developed in \cite{ARA}, { where optimization over three continues parameters is reduced to the calculation of the minimum of six values}. This method
becomes much simpler in our case. In particular, the optimization must be performed over the single parameter (instead of three parameters in ref.\cite{ARA}). { As a result, the formulas for the discord calculation becomes simpler and the optimization is reduced to the calculation of the minimum of two values}.
Assuming that the von Neumann type measurements are performed over the subsystem $B$, one can define quantum discord $Q^B$ as follows \cite{OZ}:
\begin{eqnarray}\label{Q}
Q^B={\cal{I}}(\rho) -{\cal{C}}^B (\rho).
\end{eqnarray}
Here ${\cal{I}}(\rho)$ is the total mutual information \cite{OZ} which may be written as follows:
\begin{eqnarray}\label{I}
&&
{\cal{I}}(\rho) =S(\rho^A) + S(\rho^B) + \sum_{j=0}^3 \lambda_j \log_2 \lambda_j,\\\nonumber
\end{eqnarray}
where $\lambda_j$ ($j=0,1,2,3$) are eigenvalues of the density matrix $\rho$, $\rho^A={\mbox{Tr}}_B \rho$ and $\rho^B={\mbox{Tr}}_A \rho$ are the reduced density matrices
and the appropriate entropies $S(\rho^A)$ and $S(\rho^B)$ are given by the following formulas:
\begin{align}\label{SAB}\nonumber
S(\rho^A){}&=-(\rho_{11} + \rho_{22} ) \log_2(\rho_{11}+\rho_{22}) -\\\nonumber
&(\rho_{33} + \rho_{44} ) \log_2(\rho_{33}+\rho_{44}) ,\\\nonumber
S(\rho^B){}&=-(\rho_{11} + \rho_{33} ) \log_2(\rho_{11}+\rho_{33}) -\\
&(\rho_{22} + \rho_{44} ) \log_2(\rho_{22}+\rho_{44}) .
\end{align}
The so-called classical counterpart ${\cal{C}}^B (\rho)$ of the mutual information
can be found considering the minimization over projective measurements on the subsystem $B$ as follows \cite{ARA}:
\begin{eqnarray}\label{CB2}
&&
{\cal{C}}^B (\rho)=S(\rho^A) -\min\limits_{\eta=(0,1)}(p_0 S_0 + p_1 S_1),
\end{eqnarray}
where
\begin{align}\label{S}
&S_i = -\frac{1-\theta^{(i)}}{2}\log_2\frac{1-\theta^{(i)}}{2}-
\\\nonumber
&\hspace{1cm}
\frac{1+\theta^{(i)}}{2}\log_2\frac{1+\theta^{(i)}}{2},
\\\label{p}
&p_i=\frac{1}{2} \Big(1+(-1)^i\eta(2(\rho_{11}+\rho_{33})-1) \Big),\\\label{theta}
&\theta^{(i)}=\frac{1}{p_i}\Big[(1-\eta^2) \rho_{23}^2+\\\nonumber
&
\frac{1}{2}
\Big(
2(\rho_{11}+\rho_{22})-1 +(-1)^i \eta(1-2(\rho_{22}+\rho_{33}))\Big)^2
\Big]^{1/2},
\\\nonumber
&
i=0,1.
\end{align}
It is simple to show that the quantum discord $Q^A$ obtained performing
the von Neumann type measurements on the subsystem $A$ can be calculated
as follows:
\begin{eqnarray}\label{QA}
Q^A=Q^B|_{\rho^{(22)}\leftrightarrow \rho^{(33)}}
\end{eqnarray}
for the system with the density matrix $\rho$ given by eq.(\ref{rhog}).
\iffalse
Below, we
compare quantum discord with the entanglement $E$ defined by the Wootters criterion \cite{} as follows:
\begin{eqnarray}
&&E=-X\log_2 X -(1-X) \log_2 (1-X),\;\;X=(1+\sqrt{1-C^2})/2, \\\nonumber
&&C=\max(0,2 \sqrt{\lambda}-\sum_{i=1}^4 \sqrt{\lambda_i}),\;\;
\lambda=\max_{i=1,2,3,4}{\lambda_i}
\end{eqnarray}
where $C$ is the concurrence and $\lambda_i$ are eigenvalues of the matrix
$\rho (\sigma_y\otimes \sigma_y)\rho(\sigma_y\otimes \sigma_y)$, $\sigma_y$
is the Pauli matrix.
\fi
\section{The quantum discord of the ground state}
\label{Section:gr}
In the considered model, the ground state is defined by the minimal eigenvalue of the Hamiltonian.
Since the eigenvalues read
\begin{align}
\lambda^{(1)}&=-D\Omega,\;\;\;
\lambda^{(2)}=D\Omega ,\\\nonumber
\lambda^{(3)}&=-\frac{D}{2}\sqrt{1+4\Omega^2\Delta^2},\\\nonumber
\lambda^{(4)}&=\frac{D}{2}\sqrt{1+4\Omega^2\Delta^2},
\end{align}
one can conclude that the minimal eigenvalue $\lambda_{min}$ is either $\lambda^{(1)}$ or $\lambda^{(3)}$ depending on the values $\Delta$ and $\Omega$. For the fixed $\Delta$, we introduce the critical value $\Omega_c$ such that $\lambda^{(1)}=
\lambda^{(3)}$ at $\Omega=\Omega_c$, so that $\lambda^{min}=\lambda^{(3)}$ if $\Omega<\Omega_c$ and $\lambda^{min}=\lambda^{(1)}$ if $\Omega>\Omega_c$.
The critical value $\Omega_c$ is following:
\begin{eqnarray}\label{Omegac}
\Omega_c=
\frac{1}{2\sqrt{1- \Delta^2}}
\end{eqnarray}
Both the calculation of the discord $Q^{B}$ using formulas (\ref{Q}-\ref{theta}) and the calculation of the discord $Q^{A}$ using eq.(\ref{QA}) demonstrate that $Q^A=Q^B$ if $\Omega<\Omega_c$ and $\Omega>\Omega_c$, i.e. there is no asymmetry. The discord asymmetry $\delta=Q^A-Q^B$ appears at $\Omega=\Omega_c$ and equals to
\begin{align}\nonumber
\delta(\Delta)&=\frac{1}{4} \left(
3 \log_2\frac{3-\Delta}{3+\Delta}-\log_2\frac{1+\Delta}{1-\Delta}-\Delta
\log_2\frac{9-\Delta^2}{1-\Delta^2} +\right.
\\\nonumber
&
\left.\sqrt{2(1+\Delta)}\log_2\frac{2+\sqrt{2(1+\Delta)}}{2-\sqrt{2(1+\Delta)}} +\right.
\\\label{delta}
&\left.
\sqrt{2(1-\Delta)}\log_2\frac{2-\sqrt{2(1-\Delta)}}{2+\sqrt{2(1-\Delta)}}
\right),
\end{align}
whereas
\begin{align}\label{AB}
Q^{B,A}_c(\Delta)&=
\frac{1}{4} \Big( 12 -(3\mp\Delta)\log_2(3\mp\Delta) -\\\nonumber
&(1\pm\Delta)\log_2(1\pm\Delta)-\\\nonumber
&
(2-\sqrt{2(1\pm\Delta)}) \log_2(2-\sqrt{2(1\pm\Delta)}) -\\\nonumber
&
(2+\sqrt{2(1\pm\Delta)}) \log_2(2+\sqrt{2(1\pm\Delta)})\Big).
\end{align}
Thus, if $\delta\neq 0$, then it seems to be reasonable to define the discord of a bipartite system by the following formula:
\begin{eqnarray}
Q=\min(Q^A,Q^B).
\end{eqnarray}
It follows from eq.(\ref{delta}) that the maximal asymmetry $\delta\approx 0.052$
is achieved at $\Delta_{max}\approx 0.683$ when $\Omega_c\approx 0.684$. Appropriate values of the discords are $Q^{B}_c\approx 0.230$, $Q^A_c\approx 0.282$, { i.e. the asymmetry is about 23$\%$}.
The dependence of both the discord asymmetry $\delta$ and the discord $Q$ on the parameter $\Delta$
is represented in Fig.\ref{Fig:gr}.
{ The absence of the asymmetry for both $\Omega<\Omega_c$ and $\Omega>\Omega_c$ is readily explained by the fact that the ground state is pure unless $\Omega=\Omega_c$. It is known that the discord of a pure state equals to the entanglement and does not depend on which subsystem is taken for the projective measurements \cite{DSC}. On the contrary, the ground state is degenerated if $\Omega=\Omega_c$, so that the state becomes mixed which (together with the asymmetry of the quantum system) results in the discord asymmetry. If $\Delta= 0$, then the system becomes symmetrical so that $\delta=0$, see Fig.\ref{Fig:gr}. If $\Delta=1$, then $\lambda_1>\lambda_3$ so that the ground state is pure for all $\Omega$ and, consequently, $\delta=0$. One has to emphasize also that the ground state can be obtained as the limit of density matrix (\ref{rho}) when the temperature $T$ tends to zero, which uniquely yields either the pure (nondegenerate) or mixed (degenerate) ground state depending on whether $\Omega\neq \Omega_c$ or $\Omega=\Omega_c$.
}
\section{The thermal quantum discord}
The model represented in Sec.\ref{Section:model}~~
allows one to investigate the asymmetry of the thermal discord at $T>0$. We study it for the case $\Delta=\Delta_{max}\approx 0.683$ with the different values of the parameter $\Omega$: $\Omega=\Omega_c(\Delta_{max})-0.1;\;\Omega_c(\Delta_{max});\;\Omega_c(\Delta_{max})+0.1$.
Results of such calculations are represented in Fig.\ref{Fig:Heis}.
The asymmetry of the thermal discord is negligible at high temperatures when quantum correlations disappear. However, the asymmetry increases with the decrease in
the temperature and becomes considerable over a wide temperature range. Asymmetry disappears with $T\to 0$ for all $\Omega$ except $\Omega=\Omega_c$ when $\delta\to 0.052$ and $Q\to 0.230$, { i.e. the asymmetry is about $ 23 \%$ which agrees with the results of Sec.\ref{Section:gr}.}
We also analyze the discord asymmetry $\delta$ as a function of $\Omega$ for the different values of the dimensionless temperature $\bar T=kT/(\hbar D)$ at $\Delta\approx 0.683$: $\bar T=0.02,0.1,0.25,0.5$, see Fig.\ref{Fig:Om}. It was mentioned above that the discord asymmetry of the ground state ($T=0$) appears only for $\Omega=\Omega_c$. On the contrary, if $T>0$ then the asymmetry exists for any $\Omega$ and it becomes essential inside of some interval around the critical value $\Omega_c$. This interval increases with increase in the temperature, which is demonstrated in Fig.\ref{Fig:Om}. { Conversely, if $T\to 0$, then the asymmetry becomes localized in the neighborhood of $\Omega_c$. The asymmetry reduces to the single point $\Omega=\Omega_c$ at $T=0$ (the ground state)}.
The maximum of $\delta$ for $T=0.02$ is found at $\Omega\approx 0.497$, $\delta_{max}\approx 0.098$. It corresponds to $Q\approx 0.399$ so that the asymmetry is about $25 \%$. Similarly, the maximal asymmetry at $\bar T=0.1,\;0.25, \;0.5$ is about $19 \%$, $20 \%$ and $24 \%$ respectively. { We see that the relative asymmetry $\frac{\delta}{Q}$ remains significant inside of the whole temperature interval considered here although both $\delta$ and $Q$ are vanishing with the increase in $T$.}
\begin{figure*}
\epsfig{file=fig1
, scale=0.5,angle=270
}
\caption{Fig.1. The discord asymmetry $\delta=Q^A-Q^B$ and the discord $Q$ as functions of $\Delta$ at $\Omega=\Omega_c$. }
\label{Fig:gr}
\end{figure*}
\begin{figure*}
\epsfig{file=fig2
, scale=0.5,angle=270
}
\caption{Fig.2. The discord asymmetry $\delta=Q^A-Q^B$ and the discord $Q$ (inset) as functions of the dimensionless temperature $\bar T=kT/(\hbar D)$ for $\Delta=\Delta_{max}\approx 0.683$ and different $\Omega$:
$\Omega=\Omega_c(\Delta_{max})-0.1;\;\Omega_c(\Delta_{max});\;\Omega_c(\Delta_{max})+0.1$.
}
\label{Fig:Heis}
\end{figure*}
\begin{figure*}
\epsfig{file=fig3
, scale=0.5,angle=270
}
\caption{Fig.3. The discord asymmetry $\delta=Q^A-Q^B$ and the discord $Q$ (inset) as functions of $\Omega$ for $\Delta=\Delta_{max}\approx0.683$ and different values of the dimensionless temperature $\bar T$: $\bar T=0.02,0.1,0.25,0.5$. The vertical dotted line corresponds to $\Omega=\Omega_c(\Delta_{max})=0.684$, see eq.(\ref{Omegac}).}
\label{Fig:Om}
\end{figure*}
\section{Conclusions}
\label{Section:conclusions}
We have studied the asymmetry of the quantum discord in the bipartite system for the both $T=0$ (the ground state) and $T>0$ with different values of parameters in the Hamiltonian.
The dependence of the discord on the choice of the subsystem for the projective measurements directly follows from the dependence of the classical part of the mutual information on this choice.
Our study demonstrates that, although the description of quantum correlations by the
quantum discord is a significant progress in comparison with such description by the entanglement, it requires the further development.
This work is supported
by the Program of the Presidium of RAS
No.18 "Development of methods of obtaining chemical compounds and creation of new
materials".
|
\section{Introduction}
\label{sec-Int}
The lack of a pre-planned
strategy for splitting a power-grid system into separate parts with
self-sufficient power generation is one of the reasons for large-scale blackouts
that have devastating effects on the economy and welfare of any modern
society~\cite{LILI05,PEIR09}. This defensive strategy (intentional islanding)
is effective in preventing cascading outages~\cite{PEIR09,YANG06}.
Multiple approaches to intentional islanding
(see, e.g., \cite{LILI05,PEIR09,YANG06,KARY98,CHOW08,WANG04,LIU06})
have been suggested for optimizing the selection of the lines to be cut.
These studies can be based on the analysis
of the system topology based on a representation of the network
as a graph \cite{SEAR03,FORT10,NEWM04C,HAMAD2010}.
Some topologies are easier to split into islands than others.
The identification of ``weak'' links and their removal can split
a given topology into independent islands.
While many of the above approaches are very good at the identification
of ``weak'' links, the resulting clusters or islands are usually not
optimized for other qualities such as generating capacity.
Here we present a study utilizing a matrix method for intentional
islanding of a utility power grid. The method uses a
Monte Carlo (MC) simulated annealing~\cite{SIMUL83} technique for optimizing
the resulting islands' internal connectivity as well as balancing their
generating capacity. The concept is illustrated by application to the Floridian
high-voltage power grid.
\section{Methods}
\label{sec:met}
The quality of a particular partitioning of a graph into $M$ communities,
${\mathcal C} = \{C_1, ..., C_M\}$ can be estimated by Newman's
{\it modularity\/} \cite{NEWM04C}. It compares the proportion of edges that are
internal to a community in the particular graph with the same proportion in an
average null-model. It is defined as follows:
\begin{equation}
Q = \frac{1}{w} \sum_{ij} \left( w_{ij} - \frac{w_i w_j}{w} \right)
\delta \left( C(i),C(j) \right)
\;,
\label{eq:Q}
\end{equation}
where $\delta \left( C(i),C(j) \right) = 1$ if nodes $i$ and $j$ belong to the
same community, and $0$ otherwise. Ideally, one would like to maximize $Q$ while
partitioning a power-grid network. This will ensure that the different
communities are well connected internally. Moreover, one would like to minimize
the generating power surplus or deficiency over all the clusters. Here we will
use a partitioning scheme consistent with a power-grid network and try to
optimize the resulting clusters for internal connectivity and power
self-sufficiency using a Monte Carlo simulated annealing approach. The resulting
set of clusters form a new network (in a renormalization-group sense) where each
cluster is represented by a node on the new network. The islanding procedure and
MC optimization are repeated until some required criteria are met.
\subsection{Partitioning}
We use a simplified representation of the
power grid as an undirected graph \cite{SEAR03,NEWM04C} defined by the
$N \times N$ symmetric {\it conductivity matrix\/} $\bf W$, whose
elements $w_{ij} \ge 0$ represent the ``conductivities'' of
the edges (transmission lines) between vertices (generators or loads)
$i$ and $j$,
\begin{equation}
\label{eq:conductivities}
w_{ij}=\frac{ {\rm number\, of\, lines\, between\, vertices\,}i\,{\rm and}\,j}{\rm{\, normalized
\, geographical \, distance}},
\end{equation}
where the ``geographical distance'' is the length of the edge connecting nodes
$i$ and $j$, and is normalized by the minimum geographical distance between two
nodes over the whole network.
The row sums of $\bf W$ define the diagonal matrix $\bf D$. Graph analysis can
be performed using one of several matrices derived from $\bf W$. The Laplacian,
${\bf L} = {\bf D} - {\bf W}$, is a symmetric matrix with vanishing row sums. It
embodies Kirchhoff's laws and thus represents a simple resistor network with
conductances $w_{ij}$. Multiplied by a column vector $| \phi \rangle$ of vertex
potentials, it yields the vector of currents entering the circuit at each
vertex, ${\bf L} | \phi \rangle = | I \rangle$. This equation can be rewritten
as ${\bf L}^{(-1)} | I \rangle = | \phi \rangle$, where ${\bf L}^{(-1)}$ is
the pseudo-inverse of the Laplacian. In other words, given a current vector,
defined as being positive at generator nodes and negative at load nodes, one can
calculate the potential vector $ | \phi \rangle $. Using this potential vector
and the matrix $\bf W$, we calculate the network-current matrix $\bf K$, whose
elements are the currents between the corresponding nodes:
${k}_{ij}=(\phi_j - \phi_i ){w}_{ij}$. Additionally using the matrix
${\bf L}^{(-1)}$, one can calculate an effective distance or equivalent
resistance ${\bf R}$ between any two nodes~\cite{KLEIN93}:
$({\bf R})_{ij}=({\bf L}^{(-1)})_{ii}+({\bf L}^{(-1)})_{jj}-2({\bf L}^{(-1)})_{ij}$.
Consequently, given a current vector and a conductivity matrix ${\bf W}$, the
network-current matrix ${\bf K}$ and the equivalent resistance matrix ${\bf R}$
can be evaluated. In this paper we use these two last matrices to achieve an
initial partitioning of the Floridian high-voltage grid.
The goal is to partition the power grid into communities of vertices that are
highly connected internally, but only sparsely connected to the rest
of the network. For the islanding to be useful, each island should contain at
least one generating plant. To accomplish this, we use a clustering algorithm
where each load $i$ is connected to the ``nearest'' generator $j$. The nearest
generator $j$ is the one located ``upstream'' from load $i$, i.e.
$\phi_j > \phi_i$, and for which $({\bf R})_{ij}$ is minimum. The Floridian
high-voltage grid~\cite{FLAMAP} at this first level of islanding is shown in
figure~\ref{fig:FL}(a).
Kirchhoff's junction law, when applied at each node, tells us that the sum of all
network currents going in and out of node $i$ is equal to $ |I \rangle_i$.
Thus we can think of $ |I \rangle_i$ as the current provided by a generator
at node $i$ if it is positive, or consumed by a load if it is negative.
Moreover, given the constant voltage rating (138, 230, 345, ... kV),
$ |I \rangle_i$ becomes proportional to the power being generated or consumed at
node $i$. This means that we can choose our initial current vector proportional
to the generating power of each power plant. Since the actual power rating for
power plants on the grid was not available to us, we here assume that each
generator's power is proportional to the number of edges linking to it, or its
{\it degree}. The current vector component at each node $i$ is then defined as
$|I \rangle_i = \frac{{\rm degree}_ i}{\sum_{\rm generators}{\rm degree}_j}\,
{\rm for\, generators}$ and
$|I \rangle_i =-\frac{{\rm degree}_ i}{\sum_{\rm loads}{\rm degree}_j}\,
{\rm for\,loads}$.
\begin{figure}
\begin{center}
\includegraphics[angle=-90,width=0.48\textwidth]{beforeMC_arxiv.pdf}
\hspace{0.1truecm}
\includegraphics[angle=-90,width=0.48\textwidth]{afterMC_arxiv.pdf}
\end{center}
\caption[]{
(a) Florida high-voltage power grid partitioned by connecting loads (ovals) to
their ``nearest'' generators (squares). (b) Grid clustering after the MC
simulated annealing algorithm is used for optimizing modularity and power
sufficiency of the partitioned network shown in (a). Nodes of the same color and
label belong to the same cluster. See text for discussion.
}
\label{fig:FL}
\end{figure}
\subsection{Monte Carlo}
Since the power generation or consumption rate of any generator or load is
directly proportional to its current-vector component $ | I \rangle_i$,
a community's total generating surplus or deficiency is proportional to
the sum of its members' current-vector components. Thus, to optimize our
partitioning for well-balanced communities, we try to minimize the
variance of the new current vector $ | \tilde{I} \rangle $, whose components
are defined as $ | \tilde{I} \rangle_i = \sum_{i\ge j} |I \rangle_j \delta
(C(i),C(j))$ after each iteration of the islanding procedure.
We need to maximize the modularity $Q$ for better
internal connectivity at the same time. For this purpose, we define an
optimization parameter
\begin{equation}
E=\frac{Q}{Q_{\rm init}} - \sqrt{ \frac{VAR(| \tilde{I} \rangle )}{VAR(| \tilde{I} \rangle_{\rm init})}},
\end{equation}
where the subscript ``init'' designates the initial value {\it after}
recombination, but {\it before} any MC steps. This form gives equal emphasis on
optimizing modularity and load balance. More emphasis could be given to the
optimization of one quantity versus the other by multiplying the term
corresponding to it by some weighting factor.
The MC process proceeds as follows. First, a load node $i$ is selected at
random. Then, if $i$ is at the edge of the cluster it belongs to, i.e,
if it is connected to a neighboring cluster, we randomly select one of the
neighboring clusters connected to $i$ and attempt to move this load to that
neighboring cluster. If this move does not break the first cluster into two
disconnected parts, the move is accepted with a Metropolis
acceptance rate~\cite{METROP00} $R=\min(1,e^{(- \beta \Delta E)})$,
where $\Delta E = E_a - E_i$ is the difference between the attempted state
and the initial state for that move, and $\beta$ is an ``inverse temperature.''
In a fashion similar to simulated annealing, we start at
a high temperature and gradually decrease it to zero while saving the
configuration for which $E$ is maximum. This process is repeated to look
for the global maximum of $E$.
\subsection{Recombination}
After the initial partitioning and MC, the number of clusters produced is equal
to the number of the generators in the circuit (Fig.~\ref{fig:FL}(a)), as
expected from our clustering scheme. A new network can be constructed from this
group of clusters by regarding each cluster as a new node. The connections
between the new nodes are the same as the connections between the previous
clusters. This defines a new network with new connections and a new conductivity
matrix. The current vector defined above, $| \tilde{I} \rangle $, is the new
current vector because its components represent the generating surplus or
deficiency of each of the old clusters or new ``super-generators'' or
``super-loads,'' respectively. Given the new network and the new current vector,
we repeat the above partitioning and MC schemes on the new network. The number
of clusters at this stage is equal to the number of ``super-generators.'' This
process of recombination is repeated to look for the optimum configuration
until all the original nodes belong to one cluster. The optimization parameter
${\it vs.}$ MC step and the corresponding modularity are shown in
figure~\ref{fig:EandQ}, where the red circles are the values of $Q$ for maximum
$E$ at each level of recombination.
\section{Results and Conclusion}
\label{sec:res}
The map of the Floridian high-voltage grid \cite{FLAMAP} is a
network with 84 vertices, 31 of which are generating plants.
We have modeled it as an undirected graph with
137 edges. The conductivities were calculated according to
equation~(\ref{eq:conductivities}).
While figure~\ref{fig:FL}(a) shows the clusters resulting from the first
partitioning scheme, figure~\ref{fig:FL}(b) shows the same network after
the MC annealing procedure is performed. The current vector and the
corresponding modularity before and after the MC annealing are shown in
figure~\ref{fig:CUR}. As can be seen, the MC process narrows the spread of
the current values or in other words, the average power surplus or deficiency
for the clusters is smaller. Moreover, while the modularity starts at a value of
$0.33$, it ends at a value of $0.47$ after MC optimization and before the first
recombination. The maximum optimization parameter, $E_{\rm max}=1.33$ with a
corresponding $Q=0.63$ was achieved shortly after the first recombination.
Comparable values ($E=1.31$ and $Q=0.62$) are obtained in the second iteration.
While many methods can be used to partition a network into smaller clusters,
there remains the need for further optimization of the resulting cluster
properties. Here we have used a clustering procedure to partition the
Floridian power-grid network that takes into account the generating power of
each of the power plants. Moreover, we have used MC simulated annealing to
optimize the resulting clusters for better internal connectivity and power
self-sufficiency.
\begin{figure}
\begin{center}
\includegraphics[angle=-90,width=0.9\textwidth]{EandQ_paper_arxiv.pdf}
\caption[]{
Modularity Q and the optimization parameter $E$ ${\it vs.}$
MC step.
}
\label{fig:EandQ}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[angle=-90,width=0.9\textwidth]{current_paper_arxiv.pdf}
\caption[]{Components of the current vector $| \tilde{I} \rangle$ {\it vs.} the
cluster index. It can be seen that the MC procedure narrows the spread of the
current values, meaning that, on average, the individual clusters are closer to self-sufficiency than before MC. The legend shows the significant
improvement in modularity.}
\label{fig:CUR}
\end{center}
\end{figure}
\section*{Acknowledgments}
\label{sec:ack}
This work was supported in part by
U.S.\ National Science Foundation Grant No.\ DMR-0802288,
U.S.\ Office of Naval Research Grant No.\
N00014-08-1-0080, and the Institute for Energy Systems, Economics,
and Sustainability at Florida State University.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
Recently superconductivity has been observed in several classes of
iron-pnictide materials. \cite
{kamihara, Wang08, Hsu} Despite that different classes of pnictides have
somewhat different crystal structures, their electronic structures behave
to be quite similar, as confirmed by ARPES and magneto-oscillation measurements.
\cite{Liu177005,Evtushinsky,Coldea}
Band structure calculations indicate that iron pnictides exhibit a
quasi-two-dimensional electronic structure. Their parent compounds are
metals which display an antiferromagnetic long-range order. Superconductivity
is induced either by hole doping or electron doping when part of $\mathrm{Fe}^{2+}$
ions are replaced by $\mathrm{Fe}^{+}$, or
solely by the application of high pressure.
Overall speaking, iron-pnictide shows
a good resemblance with the high-$T_c$ cuprate superconductor,
and hence is an ideal candidate for studying the superconducting (SC)
mechanism of high $T_c$.
One crucial issue towards identifying the interaction that drives superconductivity
is the symmetry structure of Cooper pairs.
A conclusive observation of the pairing symmetry remains unsettled for iron pnictides,
however, Both nodal and nodeless order parameters were reported in
experimental observations. ARPES measurements
clearly indicated a nodeless gap at all points of the Fermi surface.
\cite{Ding_EPL,zhao} Moreover, magnetic penetration depth measurement has
revealed a $T^2$ behavior down to
$0.02T_{c}$ -- a possible signature for the unconventional
$s_{\pm}$ state.\cite{mazin:057003}
The $s_{\pm}$ state is currently a promising pairing candidate for iron pnictides,
which has a sign reversal between $\alpha$ and $\beta$ bands
and can be naturally explained by the spin fluctuation mechanism.
\cite{mazin:057003, wang:047005, Tsuei2010}
On the other hand the neutron-scattering spectrum seems to
suggest the order parameter being fully gapped but without a sign reversal
(called $s_{++}$-wave in contrast to $s_{\pm}$-wave).
\cite{PhysRevB.81.060504, PhysRevB.81.180505}
On the contrary, both scanning SQUID microscopy\cite{hicks-2008} and
NMR measurements \cite{Matano,Grafe08} are
in favor of nodal SC order parameters.
For phase sensitive experiments however, one
point-contact spectroscopy reported was in favor of a nodal gap
\cite{Shan}, while the others reported were in favor of a
nodeless gap. \cite{chen}
Compared to other spectroscopy experiments,
Raman scattering has a unique feature, namely, by manipulating
the polarization directions of incident and scattered photons with
respect to the crystallographic directions,
it can selectively excite quasiparticles (QPs)
on different parts of the Brillouin zone (BZ) . Therefore Raman scattering is
considered a very useful probe for studying pairing symmetry of the SC order
parameter. More explicitly, among various symmetry channels
($A_{1g}$, $A_{2g}$, $B_{1g}$, and $B_{2g}$ within a $D_{\rm 4h}$ symmetry group),
powers laws of low-energy Raman
intensity as well as peak positions can often give
useful information on the pairing symmetry of a superconductor.
Raman scattering had been very
successful on the studies of the high-$T_c$ cuprate superconductors. \cite{RevModPhys.79.175}
It is also anticipated that Raman scattering will give very useful
signals on the iron-pnictide superconductors.
Electronic Raman scattering experiment
has recently been carried out
on high-quality single-crystal
Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$. \cite{PhysRevB.80.180510}
Both normal- and superconducting-state data were taken and for frequency
below 300 cm$^{-1}$,
it showed that a significant change of the Raman response between
the normal and superconducting state occurs only in
the $B_{2g}$ channel to which a strong peak develops at $\omega=69$cm$^{-1}$ in the
superconducting state. Moreover,
the low-energy spectral weight in the SC state suggests
a gap having nodes for certain doping levels of Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$.
It is of importance to perform a more quantitative theoretical fitting on
the Raman data reported in Ref.~[\onlinecite{PhysRevB.80.180510}] and
this is indeed the goal of the present paper.
On the theoretical side, Boyd {\em et al.} \cite{PhysRevB.79.174521} have
calculated the Raman response for iron-pnictides taking into account
multiple gaps on different Fermi sheets. In their calculations, the
Raman vertices are obtained by the expansion of
harmonic functions for a cylindrical FS and the screening effect
due to the long-range Coulomb interaction is considered but leaving out vertex corrections
capturing effects of final state interactions.
Their results give a criterion to distinguish the momentum and frequency dependence of
the SC order parameters by Raman scattering.
In another study, Chubukov and coworkers \cite{PhysRevB.79.220501, Physics.2.46}
have also calculated the Raman response for iron-pnictide superconductors.
By analyzing the vertex corrections for the extended $s$-wave gap symmetry, they
have revealed a collective mode below $2\Delta_0$ in the $A_{1g}$ channel
that may be crucial to unravel the pairing symmetry.
It provides an alternative way to distinguish between various suggested gap
symmetries of the iron-pnictide superconductors.
The electronic structure of iron-pnictide materials is considered to
be more complex than that of the high-T$_c$ cuprates.
The band structure calculations showed that superconductivity
is associated with the Fe-pnictide layer, and the FS
consists of two hole pockets and two electron pockets. \cite{singh:237003,
PhysRevB.77.220506}
The maximum contribution to the density of states near the FS is due to
the Fe-3$d$ orbitals.
Several tight-binding models are thus proposed to construct
the band structure and several competing orders are suggested to exist
in these materials.
\cite{PhysRevB.77.220506, PhysRevLett.101.087004, PhysRevLett.101.057008, 0953-8984-20-42-425203}
In this paper, we shall study the Raman response based on a two-band model.\cite{raghu_prb_08}
This minimal model is considered to be a promising one which captures
the major features of the electronic structures of several classes of
iron-pnictide materials. As mentioned before,
our aim is to give a more quantitative fitting to the current available
Raman data and uncover useful fitting parameters.
It is hoped that our study,
together with other theoretical Raman
works (Refs.~\onlinecite{PhysRevB.79.220501,PhysRevB.79.174521})
can give a more complete picture on the pairing symmetry and electronic structure
of iron pnictides.
The paper is organized as the following. In Sec.~\ref{Model Hamiltonian},
we present the two-band model Hamiltonian for iron-pnictide
superconductors. In Sec.~\ref{Raman Scattering},
theoretical derivations are given for the Raman scattering both in the normal
and in the SC states
with respect to the two-band model given in Sec.~\ref{Model Hamiltonian}.
Calculations of Raman response are presented for both normal and SC states.
Some predictions are made.
Especially fitting to the $B_{2g}$ Raman peak in the SC state are performed
with useful fitting parameters given.
Sec.~\ref{Summary} is a summary of this paper.
\section{Model Hamiltonian} \label{Model Hamiltonian}
For iron-pnictides, we start from the minimal two-band model introduced in
Ref.~[\onlinecite{raghu_prb_08}]. The normal-state Hamiltonian reads as
\begin{equation}
H_{0}=\sum_{\mathbf{k}\sigma }\left(
\begin{array}{cc}
c_{\mathbf{k}\sigma }^{\dag } & d_{\mathbf{k}\sigma }^{\dag
\end{array
\right) \left(
\begin{array}{cc}
\varepsilon _{x}-\mu & \varepsilon _{xy} \\
\varepsilon _{xy} & \varepsilon _{y}-\m
\end{array
\right) \left(
\begin{array}{c}
c_{\mathbf{k}\sigma } \\
d_{\mathbf{k}\sigma
\end{array
\right), \label{normal state Hamiltonian}
\end{equation
where $\mu$ is the chemical potential and
$(c_{\mathbf{k}\sigma}^{(\dagger)},d_{\mathbf{k}\sigma}^{(\dagger)})$
are annihilation (creation) operators that annihilate (create) an electron in
the $d_{xz}$- and $d_{yz}$-orbital with spin $\sigma $ and wavevector $
\mathbf{k}$, respectively. Energy dispersions of orbital $d_{xz}$ and $d_{yz}$
are given by $\varepsilon_{x}=-2t_{1}\cos k_x-2t_{2}\cos k_y-4t_{3}\cos k_x \cos k_y$
and $\varepsilon_{y}=-2t_{2}\cos k_x-2t_{1}\cos k_y-4t_{3}\cos k_x \cos k_y$ respectively,
which are coupled by the $d_{xy}$ orbital with dispersion
$\varepsilon _{xy}=-4t_{4}\sin k_x\sin k_y$.
By the unitary transformation
\begin{equation}
\left(
\begin{array}{c}
c_{\mathbf{k}\sigma } \\
d_{\mathbf{k}\sigma
\end{array
\right) =\left(
\begin{array}{cc}
\cos \theta _{\mathbf{k}} & -\sin \theta _{\mathbf{k}} \\
\sin \theta _{\mathbf{k}} & \cos \theta _{\mathbf{k}
\end{array
\right) \left(
\begin{array}{c}
\alpha _{\mathbf{k}\sigma } \\
\beta _{\mathbf{k}\sigma
\end{array
\right) , \label{unitary transformation}
\end{equation
Hamiltonian (\ref{normal state Hamiltonian}) can be transformed to
\begin{equation}
H_{0}=\sum_{\mathbf{k}\sigma }\left(
\begin{array}{cc}
\alpha _{\mathbf{k}\sigma }^{\dag } & \beta _{\mathbf{k}\sigma }^{\dag
\end{array
\right) \left(
\begin{array}{cc}
E_{-} & 0 \\
0 & E_{+
\end{array
\right) \left(
\begin{array}{c}
\alpha _{\mathbf{k}\sigma } \\
\beta _{\mathbf{k}\sigma
\end{array
\right) , \label{diagnalization Hamiltonian}
\end{equation
where $E_{\pm}=\varepsilon _{+}\pm \sqrt{\varepsilon _{-}^{2}+\varepsilon
_{xy}^{2}}-\mu $ with $\varepsilon _{\pm }\equiv\left( \varepsilon _{x}\pm
\varepsilon _{y}\right) /2$.
The new basis $(\alpha _{\mathbf{k}\sigma },\beta _{\mathbf{k}\sigma})$
consists of new fermionic QP operators in the $E_\pm$ bands which are hybrids
of the $d_{xz}$- and $d_{yz}$-orbital. In the literature, $E_-$ ($E_+$) is usually
labeled as the $\alpha$ ($\beta$) band.
The coherence factors in (\ref{unitary transformation}) can be solved to be
\begin{eqnarray}
\cos ^{2}\theta _{\mathbf{k}} &=&\frac{1}{2}\left( 1+\frac{\varepsilon _
\mathbf{-}}}{\sqrt{\varepsilon _{xy}^{2}+\varepsilon _{\mathbf{-}}^{2}}
\right) \notag\\
\sin ^{2}\theta _{\mathbf{k}} &=&\frac{1}{2}\left( 1-\frac{\varepsilon _
\mathbf{-}}}{\sqrt{\varepsilon _{xy}^{2}+\varepsilon _{\mathbf{-}}^{2}}
\right).
\end{eqnarray
Moreover, the single-particle Matsubara Green's functions in the normal-state
are
\begin{eqnarray}
\mathcal{G}^0_{cc}(\mathbf{k},ip_{n})&=&-\int_0^\beta d\tau~ e^{ip_n\tau}\langle
c_{{\bf k}\sigma}(\tau)c^\dagger_{{\bf k}\sigma}(0)\rangle_0\nonumber\\
&=&\frac{\cos ^{2}\theta _{\mathbf{k}}}
ip_{n}-E_{-}}+\frac{\sin ^{2}\theta _{\mathbf{k}}}{ip_{n}-E_{+}}, \notag\\
\mathcal{G}^0_{dd}(\mathbf{k},ip_{n}) &=&-\int_0^\beta d\tau~ e^{ip_n\tau}\langle
d_{{\bf k}\sigma}(\tau)d^\dagger_{{\bf k}\sigma}(0)\rangle_0\nonumber\\
&=&\frac{\sin ^{2}\theta _{\mathbf{k}}}
ip_{n}-E_{-}}+\frac{\cos ^{2}\theta _{\mathbf{k}}}{ip_{n}-E_{+}}, \\
\mathcal{G}^0_{cd}(\mathbf{k},ip_{n})&=&\mathcal{G}^0_{dc}(\mathbf{k},ip_{n})
=-\int_0^\beta d\tau~ e^{ip_n\tau}\langle
c_{{\bf k}\sigma}(\tau)d^\dagger_{{\bf k}\sigma}(0)\rangle_0\nonumber\\
&=&-\frac{\sin 2\theta _{\mathbf{k}}}{2}\left( \frac{1}{ip_{n}-E_{-}}-\frac{
}{ip_{n}-E_{+}}\right). \notag
\end{eqnarray
Throughout this paper, we shall choose $t_{1}=-1$, $t_{2}=1.3$,
t_{3}=t_{4}=-0.85$, and $\mu =1.54$, all measured in units of $|t_{1}|$.
These represent a good fit to the first-principle calculations
for the band structures of LaOFeAs. \cite{Xu67002}
\section{Raman Scattering} \label{Raman Scattering}
\subsection{Normal state} \label{sec31}
Raman scattering intensity is proportional to the imaginary part of the
effective density-density correlation function $\chi (\mathbf{q},\tau
)=\langle T_{\tau }[\tilde{\rho}(\mathbf{q},\tau ),\tilde{\rho}(-\mathbf{q
,0)]\rangle $ in the $\mathbf{q}\rightarrow 0$ limit. For the current
two-band system, effective density operator associated with Raman scattering
can be given by
\begin{eqnarray}
\tilde{\rho}(\mathbf{q},\tau )& \equiv& {\sum_{\mathbf{k}\sigma }}
[\gamma _{\mathbf{k}}^{c}c_{\mathbf{k}+\mathbf{q}\sigma }^{\dag }(\tau )c_
\mathbf{k}\sigma }(\tau ) \nonumber\\
& +&\gamma _{\mathbf{k}}^{d}d_{\mathbf{k}+\mathbf{q}\sigma }^{\dag }(\tau
)d_{\mathbf{k}\sigma }(\tau )],
\label{rho_tilde}
\end{eqnarray
where $\gamma _{\mathbf{k}}^{c}$ ($\gamma_{\mathbf{k}}^{d}$) is the Raman vertex
associated with the electrons on $d_{xz}$ ($d_{yz}$) orbital.
In the Matsubara frequency space, symmetry-channel dependent
irreducible Raman response function in the normal state has been solved to be
\begin{align}
& \chi_N({\bf q}\rightarrow 0,i\omega_n)={1\over 4}\sum_{\mathbf{k}}\Gamma _{\mathbf{k}}^{2}
\label{Raman response in normal state} \\
& \times \left[ \frac{f(E_-)-f(E_+)}{i\omega
_{n}-(E_+ - E_-)}+\frac{f(E_+)-f(E_-)}{i\omega _{n}-(E_- - E_+)}\right], \notag
\end{align
where $f(E)=1/[\exp(E/{k_B T})+1]$ is the Fermi distribution function and
\begin{equation}
\Gamma _{\mathbf{k}}=\sin 2\theta _{\mathbf{k}}\left( \gamma _{\mathbf{k
}^{c}+\gamma _{\mathbf{k}}^{d}\right)
\label{effect vertex of normal state}
\end{equation}
is a weighting factor for the symmetry-dependent
Raman scattering of a two-band system. $\Gamma_{\mathbf{k}}$ is comprised
of two parts where $\sin 2\theta _{\mathbf{k}}$ corresponds to the
effect of orbital hybridization, while $( \gamma _{\mathbf{k
}^{c}+\gamma _{\mathbf{k}}^{d})$ is the sum of the Raman vertexes associated
with two individual orbitals. In the case of zero orbital hybridization
($\varepsilon_{xy}\rightarrow 0$),
$\sin 2\theta _{\mathbf{k}}\rightarrow 0$ and hence $\chi_N\rightarrow 0$.
\begin{figure}[ptb]
\begin{center}
\includegraphics[
width=9cm ]{fig1.eps}\vspace{-0.5cm}
\end{center}
\caption{(Color online) Panel (a)--(e): To see the properties of the
channel-dependent weighting factor $\Gamma _{\mathbf{k}}$,
$\sin 2\theta _{\mathbf{k}}$, $(\gamma_{\mathbf{k}}^{c}+\gamma _{\mathbf{k}}^{d})_{B_{2g}}$,
$(\gamma_{\mathbf{k}}^{c}+\gamma _{\mathbf{k}}^{d})_{B_{1g}}$,
$\Gamma _{\mathbf{k},B_{2g}}$, and $\Gamma _{\mathbf{k},B_{1g}}$ are plotted respectively
in the first quadrant of the BZ.
Panel (f) is the calculated normal-state Raman intensity.} \label{fig1}
\end{figure}
When the energy of incident light is much smaller than the optical band gap
of the system, the
contribution from the resonant channel is negligible. Consequently, Raman vertex can
be obtained in terms of the curvature of the band dispersion, known as the
inverse effective mass approximation.\cite{PhysRevB.51.16336} That is,
depending on the symmetry of the Raman modes,
$(\gamma _{\mathbf{k}}^{c})_{ij}\sim\partial^2\varepsilon_x/\partial {\bf k}_i\partial {\bf k}_f$
and $(\gamma _{\mathbf{k}}^{d})_{ij}\sim\partial^2\varepsilon_y/\partial {\bf k}_i\partial {\bf k}_f$
with ${\bf k}_i$ and ${\bf k}_f$ being the wavevectors of incident and scattered lights respectively.
In the current two-band model of iron-pnictides, it is obtained that
\begin{eqnarray}
(\gamma _\mathbf{k}^c)_{B_{1g}} &\sim&\frac{1}{2}(\varepsilon_{x}^{xx}
-\varepsilon_{x}^{yy}) =t_{1}\cos{k}_{x}-t_{2}\cos{k}_{y},\nonumber\\
(\gamma _\mathbf{k}^d)_{B_{1g}} &\sim&\frac{1}{2}(\varepsilon_{y}^{xx}
-\varepsilon_{y}^{yy})=t_{2}\cos{k}_{x}-t_{1}\cos{k}_{y},\label{vertex}\\
(\gamma _\mathbf{k}^c)_{B_{2g}}&=&(\gamma _\mathbf{k}^d)_{B_{2g}}\sim\varepsilon_{x}^{xy}
=\varepsilon_{y}^{xy}=4t_{3}\sin{k}_{x}\sin{k}_{y},\nonumber
\end{eqnarray}
where $\varepsilon_{x}^{xx}\equiv\partial^2\varepsilon_x/\partial {\bf k}_x\partial {\bf k}_x$, etc.
In Fig.~\ref{fig1}, we first examine the behaviors of the weighting factor
$\Gamma _{\mathbf{k}}$. Respectively in Fig.~\ref{fig1}(a)--(e), we plot
$\sin 2\theta _{\mathbf{k}}$, $(\gamma_{\mathbf{k}}^{c}+\gamma _{\mathbf{k}}^{d})_{B_{2g}}$,
$(\gamma_{\mathbf{k}}^{c}+\gamma _{\mathbf{k}}^{d})_{B_{1g}}$,
$\Gamma _{\mathbf{k},B_{2g}}$, and $\Gamma _{\mathbf{k},B_{1g}}$
in the first quadrant of the BZ. As shown, both $\sin 2\theta _{\mathbf{k}}$ and
$(\gamma_{\mathbf{k}}^{c}+\gamma _{\mathbf{k}}^{d})_{B_{2g}}$ are peaked at
$(\pi/2,\pi/2)$, while $(\gamma _{\mathbf{k}}^{c}+\gamma
_{\mathbf{k}}^{d})_{B_{1g}}$ is peaked at $(\pi,0)$ and $(0,\pi)$.
Consequently, $\Gamma _{\mathbf{k},B_{2g}}$ is strongly peaked at
$(\pi/2,\pi/2)$, while due to the mismatch of $\sin 2\theta _{\mathbf{k}}$ and
$(\gamma_{\mathbf{k}}^{c}+\gamma_{\mathbf{k}}^{d})_{B_{1g}}$,
$\Gamma_{\mathbf{k},B_{1g}}$ turns out to have two peaks located at
$(k_x,k_y)=(0.26\pi,0.74\pi)$ and $(0.74\pi,0.26\pi)$.
The symmetry mismatch between $\sin 2\theta _{\mathbf{k}}$ and
$(\gamma_{\mathbf{k}}^{c}+\gamma_{\mathbf{k}}^{d})_{B_{1g}}$
implies that the overall normal-state $B_{1g}$ Raman intensity should be {\em weaker}
than that of the $B_{2g}$ channel. The reported normal-state Raman intensities of
Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ seem to be consistent with the prediction.
\cite{PhysRevB.80.180510}
Fig.~\ref{fig1}(f) shows the calculated normal-state Raman spectra in both
$B_{1g}$ and $B_{2g}$ channels. Apart from the feature that
$B_{1g}$ Raman intensity is much weaker than that of the $B_{2g}$ channel,
$B_{2g}$-channel Raman intensity shows a peak at $\omega\simeq 6.8 t_1$
while $B_{1g}$-channel Raman intensity shows a peak at $\omega\simeq 7.3 t_1$.
According to Eq.~(\ref{Raman response in normal state}), the frequencies
where the peaks appear are actually predictable.
When temperature $T\rightarrow 0$, the Raman peak for each channel can be
estimated to be at $\omega=E_+(k_x,k_y)-E_-(k_x,k_y)$ with
$(k_x,k_y)$ corresponding to the maximum of the weighting factor $\Gamma_{\bf k}$.
Therefore for $B_{2g}$-channel Raman intensity, the peak is predicted to be
at $\omega=E_+(\pi/2,\pi/2)-E_-(\pi/2,\pi/2)=6.8 t_1$,
while the $B_{1g}$-channel Raman peak is predicted to be at
$\omega=E_+(0.74\pi,0.26\pi)-E_-(0.74\pi,0.26\pi)=7.3 t_1$.
The above predicted normal-state Raman peaks may provide an
alternative route to the measurement of the band energy scale for
iron-pnictide superconductors.
In iron-pnictide superconductors, the energy scale of the nearest hopping
$t_1$ is estimated to be about $0.05 - 0.3$eV.
Thus the predicted normal-state Raman peaks could have energy $\omega\sim 0.35 - 2.1$eV.
The lower-bound energy (0.35 eV) should well be in the non-resonant regime,
while the upper-bound energy (2.1 eV) could be in the resonant regime and poses a
question mark for the above results to be valid.
Current reported normal-state Raman data on
BaFe$_{1-x}$Co$_x$)$_2$As$_2$ have only be measured up to 300cm$^{-1}$ however.
\cite{PhysRevB.80.180510}
\subsection{Superconducting state} \label{sec32}
We next study the Raman spectra in the SC state. To do so, in a mean-field level
one can add a SC pairing Hamiltonian:
\begin{equation}
H_{\text{SC}}\ =\sum_{\mathbf{k}}\left( \Delta_{\mathbf{k}}^{\alpha}\alpha_
\mathbf{k}\uparrow}^{\dag}\alpha_{-\mathbf{k}\downarrow}^{\dag}+\Delta_
\mathbf{k}}^{\beta}\beta_{\mathbf{k}\uparrow}^{\dag}\beta _{-\mathbf{k
\downarrow}^{\dag}+\text{h.c.}\right)
\label{Hsc}
\end{equation}
to the diagonalized Hamiltonian $H_0$ in
(\ref{diagnalization Hamiltonian}).
In Eq.~(\ref{Hsc}), the pairing is considered between the long-lived
$\alpha_{\mathbf{k}\sigma}$ QPs and between $\beta_{\mathbf{k}\sigma}$ QPs only.
That is, interband pairing is neglected.
Since decoupled $\alpha$ and $\beta$ bands are originated from the
coupled $d_{xz}$ and $d_{yz}$-orbitals, the kind of SC pairing Hamiltonian (\ref{Hsc})
automatically includes both intra- and
inter-orbital pairings in the original fermion basis $\left( c_
\mathbf{k}\sigma},d_{\mathbf{k}\sigma}\right)$. \cite{PhysRevB.79.020501,
Nazario:144513} The QP excitation energy is then given by
$\tilde{E}_{\mathbf{k}l}=(E^2_{l}+|\Delta_{\mathbf{k}}^l|^2)^{1/2}$ ($l=\alpha,\beta$)
for the two bands respectively. In this section, for convenience,
$E_-\rightarrow E_\alpha$ and $E_+\rightarrow E_\beta$.
If the pairing originates from the same mechanism, most likely $\alpha_1$
and $\alpha_2$ bands will have the same pairing symmetry.
Similarly $\beta_1$ and $\beta_2$ bands will also likely have the same
pairing symmetry. The Raman spectra reported in
Ref.~[\onlinecite{PhysRevB.80.180510}] do not show a clear activation threshold
rather exhibit a finite intensity down to an arbitrarily small Raman shift.
It gives a strong evidence that the SC pairing
favors an anisotropic nodal gap rather than a full isotropic gap
such as the $s_\pm$ state or the $s_{++}$ state.
Furthermore, the scanning SQUID microscopy measurements seemed to
exclude the spin-triplet pairing states and suggested that the order parameter
has well-developed nodes [\onlinecite{hicks-2008}].
Within the {\em anisotropic} and {\em nodal} scenarios,
the possible candidates are the extended $s$-wave and $d$-wave states.
\footnote{When the pairing symmetry in the unfolded BZ is extended $s$-wave with $\Delta_{\textbf{k}}=\Delta_0\cos\frac{k_x}{2}\cos\frac{k_y}{2}$ [see Fig.~\ref{fig2}(a)],
it will transform to be $\Delta_{\textbf{k}}=\Delta_0[\cos(k_x)+\cos(k_y)]$
in the folded BZ. Similarly, when the pairing symmetry in the unfolded BZ is
extended $d_{xy}$-wave with $\Delta_{\textbf{k}}=\Delta_0\sin\frac{k_x}{2}\sin\frac{k_y}{2}$
[see Fig.~\ref{fig2}(b)], it will transform to be
$\Delta_{\textbf{k}}=\Delta_0[\cos(k_x)-\cos(k_y)]$ in the folded BZ.}
\begin{figure}[ptb]
\begin{center}
\includegraphics[
width=9cm ]{fig2.eps}\vspace{-0.5cm}
\end{center}
\caption{(Color online) Panel (a)\&(b): Schematic plot of the $\alpha$- and $\beta$-band
Fermi surfaces and the $\mathbf{k}$-dependent amplitude
of (a) extended $s$-wave gap and (b) $d_{xy}$-wave gap. Panel (c)\&(d):
Momentum dependence of the effective Raman vertices
of (c) $(\gamma_{\bf k}^{\alpha\alpha})_{B_{2g}}$ and (d)
$(\gamma_{\bf k}^{\alpha\alpha})_{B_{1g}}$. Here we only show
$(\gamma_{\bf k}^{\alpha\alpha})_{B_{2g}}$ and $(\gamma_{\bf k}^{\alpha\alpha})_{B_{1g}}$ because
$(\gamma_{\bf k}^{\beta\beta})_{B_{2g}}=(\gamma_{\bf k}^{\alpha\alpha})_{B_{2g}}$ and
$(\gamma_{\bf k}^{\beta\beta})_{B_{1g}}\simeq (\gamma_{\bf k}^{\alpha\alpha})_{B_{1g}}$
(see text).}
\label{fig2}
\end{figure}
After a lengthy derivation,
the irreducible Raman response function in the SC state is solved to be
\begin{eqnarray}
\chi_S(\mathbf{q}\rightarrow0,\tau)&=&-\sum_{\mathbf{k},ll^{\prime}=\alpha,\beta}
(\gamma_{\mathbf{k}}^{ll^{\prime}})^{2}[\mathcal{G}_{l}(\mathbf{k},\tau)\mathcal{G
_{l^{^{\prime}}}(\mathbf{k},-\tau) \notag \\
& +&\mathcal{F}_{l}(\mathbf{k},\tau)\mathcal{F}_{l^{^{\prime}}}(\mathbf{k
,-\tau)], \label{response
function}
\end{eqnarray}
where $\mathcal{G}_l$ and $\mathcal{F}_l$ are the usual normal and anomalous Green
functions for a superconductor associated with band $l$.
The intra- and interband vertex functions are solved to be
\begin{align}
\gamma_{\mathbf{k}}^{\alpha\alpha} & =\gamma_{\mathbf{k}}^{c}\cos^{2}\theta_
\mathbf{k}} +\gamma_{\mathbf{k}}^{d}\sin^{2}\theta_{\mathbf{k}}, \notag \\
\gamma_{\mathbf{k}}^{\beta\beta} & =\gamma_{\mathbf{k}}^{c}\sin^{2}\theta_
\mathbf{k}} +\gamma_{\mathbf{k}}^{d}\cos^{2}\theta_{\mathbf{k}},
\label{vertex functions} \\
\gamma_{\mathbf{k}}^{\alpha\beta} & =\gamma_{\mathbf{k}}^{\beta\alpha}=
\sin2\theta_{\mathbf{k}}(\gamma_
\mathbf{k}}^{c}+\gamma_{\mathbf{k}}^{d})/2, \notag
\end{align}
where $\gamma_{\mathbf{k}}^{c}$ and $\gamma_{\mathbf{k}}^{d}$ were defined in (\ref{vertex})
for both $B_{1g}$ and $B_{2g}$ channels.
Frequency and channel-dependent Raman intensity is proportional to the imaginary part
of the effective density-density correlation function (\ref{response
function}) transformed to the Matsubara space.
As shown in Eq.~(\ref{response
function}), Raman spectra are contributed by both intraband and interband transitions.
Nevertheless, due to the little nesting effect occurring across different bands,
interband transitions are negligibly small for the Raman intensity.
Thus one can safely ignore the interband Raman scattering
in the present case. Consequently, at $T\rightarrow 0$,
Raman intensity is proportional to
\begin{equation}\label{Raman intensity}
I(\omega) = \sum_{\mathbf{k},l=\alpha,\beta}(\gamma_{\bf k}^{ll})^2
\left({|\Delta_{\mathbf{k}}^l|\over 2\tilde{E}_{\mathbf{k}l}}\right)^2
\frac{\Gamma}{(\omega-2\tilde{E}_{\mathbf{k}l})^2+\Gamma^2},
\end{equation}
where $\Gamma$ is the broadening which is
set to be 0.08 in our calculations.
As is well-known, Raman scattering is a directional probe for SC
QP excitations. In the present two-band iron-pnictide superconductors,
the directional selectivity is dependent of two factors [see (\ref{Raman intensity})].
One is due to the
Raman vertex $\gamma_{\bf k}^{ll}$ and the other is due to
the symmetry of the pairing gap $\Delta_{\mathbf{k}}^l$.
The overall Raman will also depend on the detailed
locations and topology of the Fermi surfaces of the system.
It is worth noting that, as shown explicitly in (\ref{Raman intensity}),
Raman intensity is directly proportional to the gap maximum of the pairing gap.
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=9cm ]{fig3.eps}\vspace{-0.5cm}
\end{center}
\caption{Frame (a) and (b): Comparison of theoretical fitting and experimental data of the
$B_{2g}$ Raman spectra. The SC gaps tested are extended $s$-wave
$\Delta_{\bf k}=\Delta_{0}\cos k_{x}/2\cos k_{y}/2$ in (a)
and extended $d$-wave $\Delta_{\bf k}=\Delta_{0} \sin k_{x} \sin k_{y}$ in (b). Frame (c) and (d)
are the calculated $B_{1g}$ Raman shift using the same parameters
in frame (a) and (b).}
\label{fig3}
\end{figure}
It is important to first observe how the directional selection occurs
for the iron-pnictide superconductors. We consider
the unfolded BZ for the case of one Fe/cell.
The $\alpha$-band FSs of the 2-orbital model are hole Fermi pockets given by $
E_{-}(\mathbf{k}_{f}) = 0$ which are around the $\Gamma$ point and the
corner, $(\pm\pi,\pm\pi)$. The $\beta$-band FSs are electron Fermi
pockets given by $E_{+} (\mathbf{k}_{f}) = 0$ which are around the $M$
point. In the SC state. the coupling of the two orbitals results in complex Raman vertices
given in (\ref{vertex functions}).
Shown in Fig.~\ref{fig2} (c) and (d) are the momentum dependence of the Raman
vertices $(\gamma_{\bf k}^{\alpha\alpha})_{B_{2g}}$ and $(\gamma_{\bf k}^{\alpha\alpha})_{B_{1g}}$.
Note that for the current two-band model,
$(\gamma_{\bf k}^{\beta\beta})_{B_{2g}}=(\gamma_{\bf k}^{\alpha\alpha})_{B_{2g}}$ and
$(\gamma_{\bf k}^{\beta\beta})_{B_{1g}}\simeq (\gamma_{\bf k}^{\alpha\alpha})_{B_{1g}}$.
It is because
$(\gamma_{\bf k}^{c})_{B_{2g}}=(\gamma_{\bf k}^{d})_{B_{2g}}$ and
$(\gamma_{\bf k}^{c})_{B_{1g}} \simeq (\gamma_{\bf k}^{d})_{B_{1g}}$
[see Eqs.~(\ref{vertex}) and (\ref{vertex functions})] for the current parameters.
As shown in Fig.~\ref{fig2} (c) and (d), $B_{2g}$ vertex peaks at $(\pm \pi/2, \pm\pi/2)$ in
the unfolded BZ and the peak is roughly of same distance to both
$\protect\alpha$- and $\protect\beta$-band FSs. As a matter of the fact,
it couples roughly equal to both $\protect\alpha$ and $\protect\beta$ bands.
In contrast, the $B_{1g}$ vertex has a $d_{x^2-y^2}$ symmetry and is centered
around the $\Gamma$ point and the corners, $(\pm\pi,\pm\pi)$.
Thus it couples predominantly to the $\alpha$ bands in the unfolded BZ.
Consequently $B_{1g}$ Raman scattering mainly excite the QP in the $\alpha$ bands and
can give more information about the pairing symmetry in the $\alpha$ bands.
We now test the possible extended $s$-wave and $d$-wave pairings.
Our approach is the following. We will try to fit the currently available $B_{2g}$
Raman intensity \cite{PhysRevB.80.180510} for each possible pairing symmetry.
The best fitting parameters of the gap magnitudes will be quoted. Using the
best fitting parameters for $B_{2g}$ Raman spectra, the predicted $B_{1g}$ Raman
intensity will be given. While $B_{1g}$ Raman intensity is also reported in
Ref.~[\onlinecite{PhysRevB.80.180510}], a strong phonon mode has appeared at $\omega=214$ cm$^{-1}$
responsible for Fe vibration, which makes the fitting unfeasible at the present time.
In our following calculations,
the only fitting parameters are $\Delta_{\alpha}$ and $\Delta_{\alpha}$ which are in units
of $t_1$. Both $\Delta_{\alpha}$ and $\Delta_{\beta}$ are adjusted to obtain
the best fitting for the experimental $B_{2g}$ spectra.
In particular, the $B_{2g}$ peak obtained through fitting
is identified with the ratio of $\omega/2\Delta_\beta$
which in turn is compared to the actual experimental
data of $\omega=69$cm$^{-1}$. One thus obtains the fitting value of $\Delta_\beta$.
With the knowledge of $\Delta_\beta$, one can further deduce the fitting values
of $\Delta_\alpha$ and $t_1$. The results for all possible pairing candidates
associated with extended $s$-wave and $d$-waves are
listed in Table~\ref{table1}.
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=9cm ]{fig4.eps}\vspace{-0.5cm}
\end{center}
\caption{Plot of the integrand, $(\gamma_{\bf k}^{ll})^2
({|\Delta_{\mathbf{k}}^l|/2\tilde{E}_{\mathbf{k}l}})^2$,
in Eq.~(\ref{Raman intensity}) in the first BZ.
Frame (a)\& (b) correspond respectively to $l=\alpha$ and $\beta$ with extended $s$-wave gap;
frame (c)\& (d) correspond respectively to $l=\alpha$ and $\beta$ with $d_{xy}$-wave gap.}
\label{fig4}
\end{figure}
We first consider the case of the extended $s$-wave pairing:
$\Delta_{\bf k}^l=\Delta_l\cos(k_x/2)\cos(k_y/2)$ ($l=\alpha,\beta$)
shown in Fig.~\ref{fig2}(a).
In a close observation of the $B_{2g}$ Raman spectra reported in
Ref.~\cite{PhysRevB.80.180510}, it is identified that only a {\em single}
peak develops at $\omega=69$ cm$^{-1}$. This is an important point in terms of
theoretical fitting.
In our fitting, the key is thus to ensure that both bands give a peak at the same frequency
($\omega=69$ cm$^{-1}$). {\em This one-peak scenario of fitting may seem unrealistic, but
indeed it is the only way to successfully describe the presently available data.}
It is found that for the best fitting [see Fig.~\ref{fig3}(a)],
$\Delta_\alpha/\Delta_\beta=0.35$.
The Raman peak occurs at $\omega/{2\Delta_\beta}=0.38$ which corresponds to
gap magnitudes $\Delta_\beta=91$ cm$^{-1}$ and $\Delta_\alpha=32$ cm$^{-1}$. Moreover
we have obtained $t_1=455$cm$^{-1}$.
In view of Fig.~\ref{fig3}(a) for the best fitted curve, it is seen
that $\beta$-band contributes most to the overall $B_{2g}$ Raman intensity.
As mentioned before, $B_{2g}$ Raman vertex couples roughly equal
to the $\alpha$ and $\beta$ bands, thus the individual contribution to the Raman intensity
will depend crucially on the multiple effect of Raman vertex, symmetry and magnitude
of the gap function, as well as the Fermi surface topology. To see this multiple effect,
we have plotted in Fig.~\ref{fig4} the integrand, $(\gamma_{\bf k}^{ll})^2
({|\Delta_{\mathbf{k}}^l|/2\tilde{E}_{\mathbf{k}l}})^2$,
in Eq.~(\ref{Raman intensity}) in the first BZ.
It is shown that $\beta$-band has much stronger intensity near the $\beta$-band
FS which results in a
much stronger contribution to the overall Raman scattering [see Fig.~\ref{fig3}(a)].
Moreover, due to the nature of an extended $s$-wave gap which has nodes at the BZ edges,
Raman intensity is linear dependent at small frequencies.
The case of the $d_{xy}$-wave pairing in the unfolded BZ:
$\Delta_{\bf k}^l=\Delta_l\sin(k_x/2)\sin(k_y/2)$ ($l=\alpha,\beta$)
is studied next. As shown in Fig.~\ref{fig2}(b), the gap amplitude in
$\alpha_2$ FS is larger than that in $\alpha_1$ FS. Thus
the Raman intensity due to the $\alpha$-band will mainly contributed by the
$\alpha_2$-band FS. Due to the same gap amplitude near $\beta_1$ and $\beta_2$ FSs,
they will contribute equally to the Raman intensity.
In a similar approach, to obtain the same Raman peak for the two-band model,
we again use the two-gap approach. It is found that
$\Delta_\alpha/\Delta_\beta=0.3$ will give the best fitting for the experimental $B_{2g}$ data.
Moreover the $B_{2g}$ Raman peak corresponds to $\omega/{2\Delta_\beta}=0.3$
which in turn gives $\Delta_\beta=115$ cm$^{-1}$, $\Delta_\alpha=34$ cm$^{-1}$,
and $t_1=575$cm$^{-1}$.
While $\alpha$ band FSs are fully gapped, $\beta$-band FSs are gapped with a node.
Therefore the Raman intensity is powers-law dependent at low frequencies.
Moreover, as $\beta$-band contributes most to the Raman intensity because of larger gap size,
the low-energy $B_{2g}$ Raman intensity is actually linear.
The fitting to the Raman response in $B_{1g}$ channel is not feasible at the moment.
There occurs a strong phonon mode at 214 cm$^{-1}$ due to the
Fe vibration \cite{PhysRevB.80.180510}. The phonon mode is expected to be removed
by changing the crystal structure slightly.
We have calculated and predicted the $B_{1g}$ Raman responses for the extended $s$-wave
pairing symmetry shown in Fig.~\ref{fig3}(c) with the same parameters as those used in
Fig.~\ref{fig3}(a).
As shown in Fig.~\ref{fig2}(d), although $B_{1g}$ mode couples predominantly to the
$\alpha$ band in XM directions, the gap amplitude of
$\alpha$-band is smaller than that of the $\beta$-band. Consequently
it results QP excitation from the two bands with the same weight approximately.
The low-energy Raman response is predicted to be power-law dependent
due to the full gap in both FSs.
We have also calculated and predicted the $B_{1g}$ Raman responses
for the $d_{xy}$-wave pairing symmetry shown in Fig.~\ref{fig3}(d)
with the same parameters as those used in Fig.~\ref{fig3}(b).
As shown in Fig.~\ref{fig2}(b), $\alpha_1$ FS is near gap node while $\alpha_2$
FS is fully gapped. Thus $\alpha_1$-band will contribute more to the low-energy
Raman scattering than the $\alpha_2$-band. Since the gap amplitudes in $\beta_1$ FS
is equal to that of $\beta_2$ FS, they have the same contribution to Raman scattering.
Although the gap amplitude of $\alpha$-band is smaller than that of the $\beta$-band,
$B_{1g}$ mode couples predominantly to the $\alpha$ band in XM directions however.
It results that $\alpha$ bands contribute more to the Raman scattering than
the $\beta$ band. The low-energy Raman response is predicted
to be power-law dependent with frequency due to the full gap in $\alpha_1$ band FS.
\begin{table}[ptb]
\caption{Summary of the fitting results for the three possible pairing symmetries on
the $B_{2g}$ Raman intensity.\cite{PhysRevB.80.180510}
Row 1 corresponds to the energy ratio of the $B_{2g}$ peak (occurs at
69cm$^{-1}$) to $2\Delta_\beta$. Row 2 and 3 are the gap amplitudes
of $\Delta_\beta$ and $\Delta_\alpha$ in units of cm$^{-1}$ and
row 4 are their ratios. Row 5 are the deduced values of $t_1$
in units of cm$^{-1}$.}
\begin{ruledtabular}
\begin{tabular}{cccc}
Pairing Symmetry & extended $s$-wave & $d_{xy}$-wave & $d_{x^2-y^2}$-wave\\
\hline
${\omega}/{2\Delta_\beta}$& 0.38 & 0.3 & 0.24 \\
$\Delta_{\beta}$(cm$^{-1}$) & 91 & 115 & 143\\
$\Delta_{\alpha}$(cm$^{-1}$) & 32 & 34 & 14\\
$\Delta_{\alpha}/\Delta_\beta$& 0.3 & 0.3 & 0.1 \\
$t_1$(cm$^{-1}$)& 455 & 575 & 7150 \\
\end{tabular}
\end{ruledtabular}
\label{table1}
\end{table}
To complete the studies, we have also studied the case of $d_{x^2-y^2}$-wave pairing symmetry.
It is found that the best fitting of the $B_{2g}$ Raman intensity is given by
the ratio $\Delta_{\alpha}/\Delta_{\beta}=0.1$ which in turn gives $t_1=7150$cm$^{-1}$.
Based on the given unrealistically large $t_1$,
we conclude that $d_{x^2-y^2}$-wave pairing is ruled out in terms of the current available
Raman scattering data. Table~\ref{table1} is a summary of the
the fitting results for the $B_{2g}$ Raman intensity.
\section{Summary} \label{Summary}
In summary, we have studied the Raman response of iron-pnictide
superconductor in both normal and SC states based on a two-band model.
Predictions are given for the normal-state Raman intensities.
A more quantitative fitting to the currently available Raman spectra is made
to which useful fitting parameters are quoted in terms of the
gap amplitudes on both bands.
\begin{acknowledgments}
This work was supported by National Science Council of Taiwan (Grant No.
99-2112-M-003-006), Hebei Provincial Natural Science Foundation of China (Grant
No. A2010001116), and the National Natural Science
Foundation of China (Grant No. 10974169). We also acknowledge the
support from the National Center for Theoretical Sciences, Taiwan.
\end{acknowledgments}
|
\section{Introduction}
\label{SEC_INTRO}
The understanding of quark fragmentation functions has been rapidly evolving
in recent years~\cite{Accardi:2009qv}. Based on precise data on inclusive hadron production
in hard scattering processes~\cite{Buskulic:1994ft,Abreu:1998vq,Abbiendi:1999ry,Aihara:1986mv,Abe:1998zs,Adler:2003pb,Arsene:2007jd,Abelev:2006cs}, empirical fragmentation functions have
been extracted, and their behavior under scale evolutions has been
investigated in detail~\cite{deFlorian:2007aj, deFlorian:2007hc,Hirai:2007cx}. It is reasonable to expect that in the near future
an understanding of the fragmentation functions will be attained which
is as precise as the present knowledge on quark distribution functions.
Very interesting prospects for the study of fragmentation processes
were opened by the HERMES~\cite{Airapetian:2007vu} and JLab~\cite{Avakian:2010ae, Avakian:2005ps,Mkrtchyan:2007sr,Brooks:2009xg} semi-inclusive
measurements on nuclear
targets, which will be followed up by the planned Electron-Ion Collider~\cite{Brooks:2010rz,Thomas:2009ei}.
These exciting developments present a challenge for effective theories of
QCD. Because both the distribution and fragmentation functions are basically
nonperturbative quantities, effective quark theories are powerful tools for
theoretical studies.
In particular, the Nambu-Jona-Lasinio (NJL) model~\cite{Nambu:1961tp,Nambu:1961fr}, which was very successful in the description of quark distribution functions~\cite{Cloet:2005pp}, has been combined recently
with the ideas of the jet model of Field and Feynman~\cite{Field:1977fa} to give a realistic
description of quark fragmentation functions~\cite{Ito:2009zc}. In recent work we have
applied this NJL-jet model to the fragmentation functions for pions and
kaons, and compared the results to the empirical functions~\cite{Matevosyan:2010hh}. The main
advantage of this model, which is based on the multiplicative ansatz
(product ansatz) of Field and Feynman, is that the fragmentation functions automatically
satisfy the important momentum and isospin sum rules without introducing
any new parameters into the theory. Ultimately, the NJL-jet model will provide a consistent framework to describe
semi-inclusive deep inelastic lepton-hadron scattering, as well as allowing
predictions for nuclear targets.
The purpose of the present paper is to extend the NJL-jet calculations
of quark fragmentation functions in the following three directions:
First, we include the effects of secondary pions and kaons, which come from the decay of
intermediate $\rho$, $K^*$, and $\phi$ mesons. We also include the strong decays of the
vector mesons to $\pi$ and $K$ two-particle final states.
Our aim is to investigate, in particular, the role of the $\rho$ meson for the
softening of the pion momentum distribution. The effects of three-particle decays,
like the $\omega$ meson, will be left for future work, as those processes require a treatment
of nontrivial phase space factors that cannot be evaluated analytically,
which goes beyond the usual convolution formalism.
Second, we will include the fragmentation processes
to nucleons and antinucleons. For this purpose, we will use the
splitting functions obtained in the quark-diquark description of
baryons~\cite{Mineo:1999eq}, taking into account only the effects of the scalar diquark as a first step in extending the model. The need to include the axial-vector diquark to fully describe the nucleon structure and fragmentation functions has been demonstrated in the earlier work~\cite{Jakob:1997wg,Bacchetta:2003rz,Bacchetta:2008af}, thus this remains a priority for the future developments of the model. The NJL description of nucleons as
bound states of a quark and a scalar diquark has already been applied to fragmentation functions in previous work~\cite{Kitagawa:2001ig,Yang:2002gh}. In the present paper we will, however, go beyond those earlier attempts
by including the elementary fragmentations to nucleons and antinucleons
in the cascadelike processes also including pions, kaons and vector mesons.
Third, we use the Monte Carlo (MC) method to solve for the fragmentation functions within the quark-cascade model as opposed to solving integral equations in previous work. We demonstrate the viability of this approach to replace the integral equations by providing very similar order of precision in determining the fragmentation functions, at the same time allowing to relax the approximations necessary for formulating the integral equations and easily incorporating the resonance decays i nto the model.
This paper is organized as follows: In Sec. \ref{SEC_SPLITT}, we present the calculations of the vector meson elementary fragmentation functions and elementary fragmentation functions of a quark to nucleon antinucleon pair. In Sec. \ref{SEC_MC}, we describe the Monte Carlo method for calculating fragmentation functions and include the vector meson decays. In Sec. \ref{SEC_RES}, we present the resulting solutions for the fragmentation functions. Section \ref{SEC_CONC} contains some concluding remarks and a future outlook for extending the model.
\section{Elementary Fragmentation Functions}
\label{SEC_SPLITT}
In this article, we extend our previous work of Ref.~\cite{Matevosyan:2010hh} using the same notation and model parameters. In the current section, we evaluate the "elementary" fragmentation functions of quarks to hadrons as a "one-step" process in the NJL model using light-cone (LC) coordinates\footnote{ We use the following LC convention for Lorentz 4-vectors $(a^+,a^-,\mathbf{a_\perp})$, $a^\pm=\frac{1}{\sqrt{2}}(a^0\pm a^3)$ and $\mathbf{a_\perp} = (a^1,a^2)$. }. The NJL model we use includes only four-point quark interaction in the Lagrangian, with up, down, and strange quarks and no additional free parameters (see, e.g. Refs.~\cite{Kato:1993zw,Klimt:1989pm,Klevansky:1992qe} for detailed reviews of the NJL model). We employ Lepage-Brodsky (LB) ``invariant mass'' cut-off regularization for loop integrals (see Refs.~\cite{Matevosyan:2010hh} for a detailed description as applied to the NJL-jet model), except when calculating meson-quark couplings as discussed further in this section, and use our previous values for the constituent quark masses for light and strange quarks $M_u=300~\rm{MeV}$, $M_s=537~\rm{MeV}$.
\subsection{Pseudoscalar Meson Splitting Functions}
\label{SEC_PS}
\begin{figure}[h]
\includegraphics[width=0.4\textwidth]{Plots/Fragmentation-Cut.pdf}
\caption{Quark splitting function for mesons.}
\label{PLOT_FRAG_QUARK}
\end{figure}
The elementary fragmentation function of quark $q$ emitting a meson $m$ carrying light-cone momentum fraction $z$ is depicted in Fig.~\ref{PLOT_FRAG_QUARK}. For completeness of the description we present the results for the pseudoscalar mesons derived in~\cite{Matevosyan:2010hh}, leaving out the details. In the frame where the fragmenting quark has $\mathbf{k_\perp}=0$, but nonzero transverse momentum component $\mathbf{k}_T=-\mathbf{p_\perp}/z$ with respect to the direction of the produced hadron, the relation for the elementary fragmentation function is
\begin{align}
\label{EQ_QUARK_FRAG}
\nonumber
d_{q}^{m}(z)= - \frac{C_q^m}{2} g_{mqQ}^{2} \frac{z}{2} \int \frac{d^{4}k}{(2\pi)^{4}} Tr[S_{1}(k)\gamma^{+}S_{1}(k)\gamma_{5} (\slashed{k}-\slashed{p}+M_{2}) \gamma_{5}]&\\
\times \delta(k_{-} - p_{-}/z) 2 \pi \delta( (p-k)^{2} -M_{2}^{2} )
&\\
= \frac{C_q^m}{2} g_{mqQ}^{2} z \int \frac{ d^{2} \mathbf{p}_{\perp}}{(2\pi)^{3}} \frac{p_{\perp}^{2}+((z-1)M_{1}+M_{2})^{2}} {(p_{\perp}^{2}+z(z-1)M_{1}^{2}+zM_{2}^{2}+(1-z)m_{m}^{2})^{2}}.&
\end{align}
Here $\mathrm{Tr}$ denotes the Dirac trace and the subscripts on the quark propagators denote quarks of different flavor - also indicated by $q$ and $Q$, where the meson of type $m$ under consideration has the flavor structure $m=q\overline{Q}$. The $C_q^m$ is the corresponding flavor factor given in the Table~\ref{TB_FLAVOR_FACTORS}. The pseudoscalar meson-quark coupling constant, $g_{mqQ}$, is determined from the residue at the pole in the quark-antiquark t-matrix at the mass of the meson under consideration.
In LB regularization the cut-off in the transverse momentum, $P_\perp^2$, is given by
\begin{eqnarray}
\label{EQ_QUARK_FRAG_LB_PPERP}
P_\perp^2= z(1-z)\left(\sqrt{\Lambda_{3}^2 +m_m^2}+ \sqrt{\Lambda_{3}^2 +M_2^2}\right)^2 - (1-z)m_m^2 -z M_2^2,
\end{eqnarray}
where $\Lambda_3= 0.67 ~{\rm GeV}$ denotes the 3-momentum cutoff obtained in Ref.~\cite{Bentz:1999gx}, yielding the following values for the meson-quark couplings:
\begin{align}
\label{EQ_PS_Q_COUPLING}
g_{\pi qQ}=3.15 ,\ \
g_{K qQ}=3.39 .
\end{align}
A consequence of LB regularization is a limited range for $z$ in corresponding regularized functions, $0<z_{min}\leq z\leq z_{max}<1$, where $z_{min}$ and $z_{max}$ are determined by imposing the condition $P_{\perp}^2\geq 0$ in Eq.~(\ref{EQ_QUARK_FRAG_LB_PPERP}). These range limitations depend on the masses of hadrons and quarks involved. For example, the $z$ limits are very close to endpoints ($z=0$ and $z=1$) for splitting functions to pions, but are quite far from them for heavier hadrons like nucleons or $\phi$ meson.
\subsection{Vector Meson Splitting Functions}
\label{SEC_VM}
In the experimental measurements of the quark hadronization process, one usually directly detects only the relatively long lived particles: pions, nucleons, and sometimes kaons. These particles can come either as direct emissions of the fragmenting partons or as decay products of hadronic resonances. We note that in the current article we will only consider the strong decays of the hadrons. The inclusion of the vector mesons in the NJL-jet model is important, since it has a two-fold effect on the description of the previously calculated pion and kaon fragmentation functions.
First, the availability of the new emission channels decreases the normalization of the direct quark splitting functions to pions and kaons. In fact, in the early models for the quark fragmentation, it was assumed that the ratio of the vector to pseudoscalar meson fragmentations should follow from the simple spin state counting rule, $3:1$. Later, it was shown experimentally that this vector meson ratio is significantly smaller, especially in the light quark sector (see, for example, Ref.~\cite{Chliapnikov:1999qi}), and thus has been attributed to dynamic effects like the masses of the considered mesons having a strong influence on the ratio. (The above mentioned rule holds only for those mesons containing the heaviest quarks, where the differences in masses between pseudoscalar and vector mesons are negligible.)
Second, the strong decays of the vector mesons produce pseudoscalar mesons with a $z$ distribution that is distinguishable from the ones emitted directly in the quark fragmentation process, thus modifying the final $z$ dependence of the fragmentation functions.
The basic NJL interaction Lagrangian relevant for the vector meson channels has the form $-G_v \left( \bar{\psi}\gamma^\mu \lambda_\alpha \psi \right)^2$, where $\alpha = 0, 1, . . .8$ with $\lambda_0 = \sqrt{2/3}\ \mathbf{1}$. If the 4-Fermi coupling constant $G_v$ is chosen to reproduce the mass of the $\rho$ meson as the pole of the $q\bar{q}$ t-matrix, the calculated masses of $K^*$ and $\phi$ mesons agree well with the experimental values (see Appendix~\ref{SUB_SEC_VM_COUPL}). Using the vector meson - quark vertex $g_{mqQ}\gamma^\mu \lambda_\alpha$ (versus $ \imath g_{mqQ}\gamma^5 \lambda_\alpha$ for pseudoscalar mesons), the elementary fragmentation function can be written as
\begin{align}
\label{EQ_QUARK_FRAG_V}
\nonumber
d_{q}^{m}(z)= \sum \limits_{\lambda=\mp1, 0} \frac{C_{q}^m}{2} g_{mqQ}^{2} \frac{z}{2}
\int \frac{d^{4}k}{(2\pi)^{4}}
\mathrm{Tr}[S_{1}(k)\gamma^{+}S_{1}(k)\gamma_{\mu} (\slashed{k}-\slashed{p}+M_{2}) \gamma_{\nu}] \epsilon^\mu(\lambda, p) \epsilon^{*\nu}&(\lambda, p)\\
\times \delta(k_{-} - p_{-}/z) 2\pi \delta( (k-p)^{2}& -M_{2}^{2} ),
\end{align}
where the sum is over the polarization\footnote{Since in this paper we consider only spin-independent fragmentation functions, we include the spin degeneracy factor $2s_b + 1$ into the definition of the fragmentation function $d_a^b(z)$, so as to facilitate the comparison to the empirical parametrizations~\cite{Hirai:2007cx,deFlorian:2007aj}. (We note that the operator definition used in Ref.~\cite{Ito:2009zc} does not include this degeneracy factor.)} of the vector meson $m$ and $Tr$ is over Dirac indices. The details of the calculation of the $Tr$ and integration over plus and minus components of the momentum are given in Appendix~\ref{APP_VM_SPLITT}. Here we simply present the resulting expression:
\begin{eqnarray}
\label{EQ_QUARK_FRAG_F1}
d_{q}^{m}(z)
=\frac{C_{q}^m}{2} g_{mqQ}^{2} z
\int \frac{ d^{2} \mathbf{p}_\perp}{(2\pi)^{3}} &
\frac{
2 \left(p_\perp^2 +((1-z)M_1-M_2)^2\right)
+\frac{1}{z^2 m_m^2}\left( \left(p_\perp^2 - z^2 M_1 M_2+(1-z) m_m^2\right)^2+p_\perp^2 z^2 \left(M_1+M_2\right)^2 \right)
}
{ \left (p_\perp^2 + z(z-1)M_1^2+ z M_2^2 + (1-z)m_m^2 \right)^2}.& \nonumber \\
\end{eqnarray}
The vector meson-quark couplings, $g_{mqQ}$, are determined from the residue at the pole in the quark-antiquark t-matrix at the mass of the meson under consideration. Here we encounter a deficiency in LB regularization scheme, where the mesons with mass larger than the sum of constituent quark's masses are not bound. This problem is usually circumvented \cite{Ebert:1996vx} by choosing a large enough constituent quark mass accounting for the masses of all the considered hadrons. Here to keep consistency with our previous calculations, we choose to use a different approach: we will use the same constituent quark masses as in the previous NJL-jet model calculations, but in order to assess the vector meson-quark couplings we will use the proper-time (PT) regularization scheme of Refs.~\cite{Ebert:1996vx, Hellstern:1997nv, Bentz:2001vc} that mimics confinement and eliminates the unphysical decay thresholds. The details of these calculations are presented in Appendix~\ref{APP_QM_COUPLING}, where we compare the pseudoscalar meson-quark couplings calculated both in LB and PT regularization schemes and ensure that they are practically the same. Also, the strange constituent quark mass determined from the experimentally measured kaon mass is practically the same in both regularization schemes. It is therefore reasonable to use the PT scheme for the calculation of the vector meson- quark couplings in the present model, while keeping the LB scheme for the regularization of the elementary fragmentation functions.
\subsection{$N$ and $\overline{N}$ fragmentation channels}
\label{SEC_NNB}
\begin{figure}[phtb]
\centering
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Split-q-N-1A.pdf}}
\hspace{0.1cm}
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Split-q-N-1B.pdf}
}
\vspace{0.1cm}
\subfigure[] {
\includegraphics[width=0.98\textwidth]{Plots/Split-q-N-1C.pdf}
}
\caption{Two contributions to the quark fragmentation function to a nucleon and antidiquark are depicted on Figs. (a) and (b), and the quark distribution function in nucleon $f_q^N(x)$ is depicted in (c) as a cut diagram (left) and as the equivalent Feynman diagram (right).}
\label{PLOT_FRAG_Q_N}
\end{figure}
In this section, we consider the elementary splitting functions, which involve
nucleons and \\ antinucleons, i.e., $q \rightarrow N\overline{D}$ and the subsequent
process $\overline{D} \rightarrow \overline{N} q$. Here $q=u,\,d$ denotes the
nonstrange quark, $N=n,\,p$ denotes
the nucleon, and $D$ denotes a scalar
diquark ($J=T=0, \, \overline{3}_c$). The first process leads to the
elementary fragmentation functions $d_q^N(z)$, represented as a cut-
diagram in Fig.~\ref{PLOT_FRAG_Q_N}a, and $d_q^{\overline{D}}(z)$, which is shown in Fig.~\ref{PLOT_FRAG_Q_N}b.
We recall that for a general process $a \rightarrow b c$, the
fragmentation function $d_a^b(z)$ corresponds to the probability distribution
of the LC momentum carried by $b$ relative to $a$, and includes a sum over the
spin-color quantum numbers of the spectator $c$. Therefore, the two
fragmentation functions of Fig.~\ref{PLOT_FRAG_Q_N} are simply related by
\begin{eqnarray}
d_q^{\overline D}(z) = \frac{1}{3} d_q^N(1-z) \,. \label{rel1}
\end{eqnarray}
The second process $\overline{D} \rightarrow \overline{N} q$ leads to the
elementary fragmentation functions $d_{\overline D}^{\overline N}(z)$, shown in
Fig.~\ref{PLOT_FRAG_DQB_N}a, and $d_{\overline D}^q(z)$ of Fig.~\ref{PLOT_FRAG_DQB_N}b. The relation for this case is
\begin{eqnarray}
d_{\overline D}^q(z) = \frac{1}{3} d_{\overline D}^{\overline N}(1-z) \,. \label{rel2}
\end{eqnarray}
To evaluate the functions $d_q^N(z)$ and $d_{\overline D}^{\overline N}(z)$,
one can either directly consider the
cut diagrams of Figs.~\ref{PLOT_FRAG_Q_N}a and~\ref{PLOT_FRAG_DQB_N}a, or one can make use of the fact that all
elementary NJL fragmentation functions are formally related to the more familiar
distribution functions $f(x)$ by crossing and charge conjugation, which is expressed
by the relation\footnote{We emphasize that this so called ``Drell-Levy-Yan
relation'' can be used only to obtain the expressions for the {\em elementary}
{\em unregularized} fragmentation functions from the distribution functions.} \cite{Drell:1969jm,Blumlein:2000wh}
\begin{eqnarray}
d_a^b(z) = \left(-1 \right)^{2(s_a+s_b) + 1} (2s_b+1)\frac{z}{\gamma_a}
f_a^b\left(x=\frac{1}{z}\right) \,, \label{dly}
\end{eqnarray}
where $\gamma_a$ is the spin-color degeneracy of $a$.
Therefore, in order to obtain $d_q^N(z)$, we can use the well-known
expression for the quark distribution function inside a nucleon, where the
nucleon is
described as a bound state of a quark (mass $M$) and a scalar diquark\footnote{The basic NJL interaction Lagrangian in the scalar diquark (D) channel has the form $G_s \left(\bar{\psi} \gamma_5 C \tau_2 \beta^A \bar{\psi}^T\right) \left(\psi^T C^{-1} \gamma_5 \tau_2 \beta^A \psi \right)$, where $\beta^A= \sqrt{3/2}\ \lambda_A\ (A = 2, 5, 7)$ and $C = i\gamma_2 \gamma_0$. The diquark mass $M_D$ is calculated as the pole of the $qq$ t-matrix, and the nucleon mass as the pole of the $qD$ t-matrix~\cite{Bentz:2001vc}. The 4-Fermi coupling constant $G_s$ is chosen so as to reproduce the experimental nucleon mass.}
(mass $M_D$). This
distribution function $f_q^N(x)$ is shown in Fig.~\ref{PLOT_FRAG_Q_N}c. It is easily evaluated
by using the form of the nucleon vertex function \cite{Mineo:2003vc}
\begin{eqnarray}
\Gamma_N(p) = \sqrt{- Z_N \frac{M_N}{p_-}} \, u_N(p) \,,
\label{vert}
\end{eqnarray}
where the Dirac spinors are normalized as $\overline{u}_N u_N = 1$, and the normalization
factor $Z_N$ is given by
\begin{eqnarray}
{Z}_N = \left(- \frac{\partial {\Pi}_N}{\partial \slashed{p}}
\right)^{-1}_{\slashed{p}=M_N}\,, \label{zn}
\end{eqnarray}
with the quark-diquark bubble graph expressed by the Feynman
propagators of the quark and the diquark:
\begin{eqnarray}
{\Pi}_N(p) = i \int \frac{{\rm d}^4 k}{(2\pi)^4} S_F(k) \Delta_s(p-k) \,,
\label{qdbub}
\end{eqnarray}
which yields for the normalization factor
\begin{eqnarray}
Z_N^{-1} &=& \frac{1}{2} \int_0^1 {\rm d}x
\int \frac{{\rm d}^2 \mathbf{p}_{\perp}}{(2\pi)^3}
\frac{(1-x)\left(p_{\perp}^2+(M_Nx+M)^2 \right)}{\left[
M_N^2 x(1-x) - p_{\perp}^2 - M^2 - x(M_D^2-M^2)\right]^2}.
\end{eqnarray}
We obtain from the Feynman diagram of Fig.~\ref{PLOT_FRAG_Q_N}c
\begin{eqnarray}
f_q^N(x) &=& i M_N {Z}_N {\overline u}_N \int \frac{{\rm d}^4 k}
{(2\pi)^4} \left(S_F(k) \gamma^+ S_F(k) \right) \Delta_s(p-k)
\delta(k_- - p_- x) u_N \nonumber \\
&=& \frac{1}{2} {Z}_N (1-x) \int \frac{{\rm d}^2 \mathbf{k}_T}
{(2\pi)^3} \frac{{k}_T^2 + (M_N x + M)^2}{
\left[{ k}_T^2 + M^2(1-x) + M_D^2 x - M_N^2 x (1-x)\right]^2}\,.
\label{fqn}
\end{eqnarray}
Here, $(q,N) = (u,p)$ or $(d,n)$, while for the other flavor combinations the
distribution function
vanishes. The relation (\ref{dly}) then gives
\begin{eqnarray}
d_q^N(z) = \frac{1}{6} {Z}_N (1-z)
\int \frac{{\rm d}^2 \mathbf{p}_{\perp}} {(2\pi)^3}
\frac{{ p}_{\perp}^2 + (M_N + M z)^2}
{\left[{ p}_{\perp}^2 - M^2 z(1-z) + M_D^2 z + M_N^2 (1-z)\right]^2}\,.
\label{dqn}
\end{eqnarray}
The function $d_q^{\overline D}(z)$ can then be obtained from (\ref{rel1}),
which completes the evaluation of the fragmentation functions of Figs.~\ref{PLOT_FRAG_Q_N}a and~\ref{PLOT_FRAG_Q_N}b.
\begin{figure}[phtb]
\centering
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Split-DB-N-2A.pdf}}
\hspace{0.1cm}
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Split-DB-N-2B.pdf}
}
\caption{Two contributions to the antidiquark fragmentation function to an antinucleon and a quark are depicted on Figs. (a) and (b).}
\label{PLOT_FRAG_DQB_N}
\end{figure}
For the fragmentation function $d_{\overline D}^{\overline N}(z)$ of
Fig.~\ref{PLOT_FRAG_DQB_N}a, we note that charge conjugation invariance gives
$d_{\overline D}^{\overline N}(z) = d_{D}^{N}(z)$, where the latter function is
related to the diquark distribution function inside the nucleon ($f_D^N(x)$)
by the crossing relation (\ref{dly}). The function $f_D^N(x)$ in turn is simply
obtained from the quark distribution function (\ref{fqn}) by replacing
$x \rightarrow 1-x$. Summarizing, we obtain
\begin{eqnarray}
d_{\overline D}^{\overline N}(z) &=& z \frac{2}{3} f_q^N(1-x) |_{x=1/z} \nonumber \\
&=& \frac{1}{3} {Z}_N \int \frac{{\rm d}^2 \mathbf{p}_{\perp}}
{(2\pi)^3} \frac{{ p}_{\perp}^2 + \left(M z - M_N(1-z)\right)^2}
{\left[{ p}_{\perp}^2 + M^2 z - M_D^2 z (1-z) + M_N^2 (1-z)\right]^2}\,,
\label{ddn}
\end{eqnarray}
for both cases $N=p$ or $n$.
The remaining function $d_{\overline D}^q(z)$ is then obtained
from the relation (\ref{rel2}). This concludes the evaluation of the fragmentation
functions of Figs.~\ref{PLOT_FRAG_DQB_N}a,~\ref{PLOT_FRAG_DQB_N}b.
Here, we note that another possible channel for a nucleon emission within NJL formalism, $q \rightarrow D\overline{q}$ and subsequent $ D \rightarrow N\overline{q}$ is not included in the current version of the model, as numerically the norm for the splitting function of this channel is 2 orders of magnitude smaller than the norm of the splitting function for the channel described above, thus proving insignificant.
\section{Monte Carlo Approach to Calculating the Fragmentation Functions}
\label{SEC_MC}
\subsection{Monte Carlo Simulations as an Alternative to the Integral Equation Method}
\label{SEC_MC_SIMS}
The Monte Carlo method for describing quark fragmentation process using most notably the Lund string model \cite{Andersson:1983ia} has been long employed to successfully describe the hadronization process \cite{Sjostrand:1982fn} and has been implemented in very sophisticated event generator frameworks like PYTHIA~\cite{Sjostrand:2007gs}. These frameworks have been refined, extended, and tuned over several decades to cater for the needs of the experimental data analysis. Our purpose here is to develop a standalone MC simulation software that has the specific purpose of implementing the quark-cascade description of the hadronization process with quark splitting functions supplied from an effective quark model, particularly NJL-jet model in the current article, and vector meson decays to pseudoscalar mesons. This allows for flexibility and simplicity of the software platform, making it readily accessible for further development.
Previously, in the NJL-jet model the fragmentation functions were obtained as solutions of a set of coupled integral equations (\cite{Ito:2009zc,Matevosyan:2010hh}),
\begin{equation}
\label{EQ_FRAG_PROB}
D^{h}_{q}(z)dz=\hat{d}^{h}_{q}(z)dz+\sum_{Q}\int^{1}_{z}\hat{d}^{Q}_{q}(y) dy \; D^{h}_{Q}\left(\frac{z}{y}\right) \frac{dz}{y},
\end{equation}
where $D^h_q(z)$ denotes the fragmentation function of quark $q$ to hadron $h$ carrying light-cone momentum fraction $z$, and $\hat{d}_{q}^{h}(z)$, $\hat{d}_{q}^{Q}(z)$ are the elementary fragmentation functions for the process $q\to h Q$, normalized as $\sum_{h}\int \hat{d}_{q}^{h}(z) dz=\sum_{Q}\int \hat{d}_{q}^{Q}(z) dz=1$, thus allowing an interpretation as the probability of an elementary process. The sum on the right-hand side is over all possible intermediate states in the quark cascade ( in our model that includes all the active flavors of quarks and scalar antidiquarks). In formulating the integral equations, one assumes that the quark has infinite momentum and produces an infinite number of hadrons, which physically corresponds to the Bjorken limit.
We propose to calculate the fragmentation functions using Monte Carlo simulations akin to the method described in Ref.~\cite{Ritter:1979mk} using the probabilistic interpretation: $D_q^h(z)\ dz $ is the probability to emit a hadron $h$ carrying the light-cone momentum fraction $z$ to $z + dz$ of initial quark $q$ in a quark-jet picture. The quark goes through a cascade of hadron emissions, where at every emission vertex we choose the type of emitted hadron $h$ and its light-cone momentum fraction $z$ (of the fragmenting quark) by randomly sampling the corresponding probability densities of the elementary fragmentations, $\hat{d}_q^h(z)$, that are calculated within the NJL model (in general these can be calculated in any effective quark model). We keep track of the flavor and the light-cone momentum fraction of the initial quark left to the remaining quark, also recording the type and light-cone momentum fraction of the initial quark transferred to the emitted hadron in each elementary fragmentation process. We stop the fragmentation chain after the quark has emitted a predefined number of hadrons, $N_{Links}$. (Alternatively, one could stop the chain after the remnant quark in the cascade has less than a given fraction of the initial quark's light-cone momentum, $z_{Min}$.) We repeat the calculation $N_{Sims}$ times with the same initial quark flavor $q$, until we have sufficient statistics for the emitted hadrons. We extract the fragmentation functions by calculating the average number of hadrons of type $h$ with light-cone momentum fraction $z$ to $z + \Delta z$, $\left<N_q^h(z, z+ \Delta z) \right>$ and expressing it as
\begin{equation}
\label{EQ_FRAG_MC}
D_q^h(z) \Delta z = \left< N_q^h(z, z+ \Delta z) \right> \equiv \frac{ \sum_{N_{Sims}} N_q^h(z, z+ \Delta z) } { N_{Sims} }.
\end{equation}
From the construction, it is obvious that the fragmentation functions calculated using the integral equations, Eq.~(\ref{EQ_FRAG_PROB}), should be equivalent to those calculated using the MC method in the limit $N_{Links} \to \infty$ and $N_{Sims} \to \infty$. The plots in Fig.~\ref{PLOT_MC_INT_COMP} show that the solutions for fragmentation functions from both methods are indeed equivalent with a large enough number of emitted hadrons within statistical errors. It follows from Eq.~(\ref{EQ_FRAG_PROB}) that the solution to the integral equations behaves as $zD_q^h(z) \to \mathrm{const}$ as $z \to 0$. This behavior originates in the assumption made by Field and Feynman in the original jet model~\cite{Field:1977fa} that the total fragmentation function can be expressed as an infinite product of elementary fragmentation functions, which allows one to formulate the integral equations. The deviations of the MC solution from the solution of the integral equations for very small $z$ is because in the MC calculation we always use a finite number of hadrons emissions, although in Fig.~\ref{PLOT_MC_INT_COMP} this number was taken to be large enough so as to demonstrate the equivalence to the integral equations.
\begin{figure}[phtb]
\centering
\includegraphics[width=0.8\textwidth]{Plots/Frag-MC-Integ-Comp.pdf}
\caption{Comparison of the solutions for quark fragmentation function $D_u^{\pi^+}(z) $ in NJL-jet model with only nonstrange pseudoscalar mesons calculated from integral equations, Eq.~(\ref{EQ_FRAG_PROB}) and MC simulation.}
\label{PLOT_MC_INT_COMP}
\end{figure}
MC also allows us to study the dependence of the resulting fragmentation functions on the number of hadrons emitted by the quark in the cascade, which could well be relevant to many medium-energy experiments. Figure~\ref{PLOT_FRAG_VS_NLINKS} shows that the solution for $z D_u^{\pi^+}(z) $ with $N_{Links}=1$ [equivalent to the elementary fragmentation function $z \hat{d}_u^{\pi^+}(z)$] is peaked at $z \sim 0.8$. As we increase the number of emitted hadrons, the solution increases in the low $z$ region because of the hadrons emitted further in the quark jet, where the fragmenting quark typically has a small fraction of the initial quark's light-cone momentum. We can readily see that the solutions saturate after including only a few emitted hadrons, where there is virtually no difference between solutions with $N_{Links} = 8$ and $N_{Links} = 20$, and the discrepancy to the solution of the integral equations only occurs at extremely small values of $z$, approaching the limit $zD(z) \to \mathrm{const}$ in the case of large number of links. Thus we can reliably use the solutions of MC simulations with $N_{Links}\geq 8$.
\begin{figure}[phtb]
\centering
\includegraphics[width=0.8\textwidth]{Plots/Frag-MC-Integ-Comp-NLinks.pdf}
\caption{The dependence of the solutions for $z D_u^{\pi^+}$ on $N_{Links}$.}
\label{PLOT_FRAG_VS_NLINKS}
\end{figure}
\subsection{Including the Resonance Decays}
In the current version of the NJL-jet model, we included the production of vector mesons, among the other hadrons directly emitted by the quark cascade (so called "primary" hadrons). The vector mesons considered have an extremely short lifetime and decay quickly, thus in an experimental setup one usually detects only their decay products ( "secondary" hadrons), most often pions and kaons. Schematically, the process is depicted in Fig.~\ref{PLOT_JET_DECAY}. Consequently, to best describe the experimentally measured fragmentation functions of a hadron $h$ from the Eq.~(\ref{EQ_FRAG_MC}), we should not only consider the number of primary hadrons with a given range of light-cone momentum fraction of the original quark, $N_q^h(z, z+ \Delta z)$, but also add the number of hadrons $h$ satisfying the above criteria that originate from the decays of primary vector mesons (or in general other hadronic resonances). This is accomplished using the dependence of the decay probability of a hadronic resonance $h$ on the fractions of its light-cone momentum, $z_1$ to $z_1+d z_1,..., z_n$ to $z_n+dz_n; \sum_i z_i =1$, carried by the decay products $h_1, ... h_n$, denoted as $dP^{h\rightarrow h_1... h_n}(z_1,...,z_n)$. First we perform the MC simulations described in the previous section and record all the produced (primary) hadrons along with the fractions of the initial quark's light-cone momentum. We calculate the fragmentation functions without the decays using the formula in Eq.~(\ref{EQ_FRAG_MC}). Second, we consider each encountered resonance particle $h$ with a momentum fraction of the initial quark $z$ and its possible strong decay channels, randomly selecting one according to the corresponding relative branching ratio. Then we randomly generate the light-cone momentum fractions $z_1,..., z_n$ carried by the decay products $h_1, ... h_n$ according to the probability $dP^{h\rightarrow h_1... h_n}(z_1,...,z_n)$. The decaying hadron $h$ is removed from the list of the produced hadrons and the decay products $h_1,..., h_n$ are added with their respective fractions of the initial quark's light-cone momenta $z z_1, ..., z z_n$. The fragmentation functions are once again calculated using the new list of produced hadrons using the formula in Eq.~(\ref{EQ_FRAG_MC}). (In general, the decay of a primary resonance can produce another resonance of a lower mass that decays itself, so we start the simulation of the decay process from the heaviest resonances present and move on to the lighter ones.)
\begin{figure}[htb]
\centering
\includegraphics[width=0.6\textwidth]{Plots/Quark-Jet-Decay.pdf}
\caption{Quark-Jet model with the decay of the resonances.}
\label{PLOT_JET_DECAY}
\end{figure}
In the current article, we consider only the strong decays of the vector mesons to two-body pseudoscalar meson final states for simplicity. A generalization to include the three or more-body final states is straightforward, although it requires sampling the nontrivial phase space factors using Monte Carlo techniques. We consider all of the two-body strong decays of $\phi$, $K^*$, and $\rho$ mesons with the corresponding relative branching ratios as given in the "Review of Particle Physics"~\cite{Nakamura:2010zzi}. For calculation of the two-body decay probability function, let us consider the initial hadron $h$ with mass $m_h$, momentum $q$ decaying to hadron $h_1$ with a mass $m_{h1}$, and a momentum $p_1$ and hadron $h_2$ with mass $m_{h_2}$, and momentum $p_2$. We also denote the light-cone momentum fraction of the $h$ carried by $h_1$ as $z_1 \equiv p_{1}^-/q^- $ and the fraction carried by $h_2$ as $z_2 \equiv p_2^-/q^-$, where trivially $z_1+z_2=1$. Thus, the decay probability is a function of only one momentum fraction chosen to be the $z_1$. The $dP^{h\rightarrow h_1,h_2}(z_1)$ is determined as a product of the decay amplitude squared times a two-body phase space factor.
A detailed description of the decay amplitudes and the branching ratios into different channels can be calculated using specific models ( for example model Lagrangians for the interaction from Ref. \cite{Matsuyama:2006rp}). Here, we are only concerned with the dependence of the decay probability on $z_1$, where we average over the polarization of the vector meson. Thus, the Lorentz invariance requires that amplitude squares depend only on scalar products of the 4-momenta of the particles involved in the decay, which in turn are trivially expressed through on-mass-shell conditions in terms of their masses. Thus, the only dependence on $z_1$ comes from the two-body phase space factor. Our goal is to express the elementary probability for the decay as a function of $z_1$, assuming a constant $C_h^{h_1 h_2}$ for the amplitude squared of the decay. For that we integrate over all components of momenta $p_1$ and $p_2$ with exception of $p_1^-$, assuming $\bf{q_\perp=0}$ and using the light-cone form for the two-body phase space factor:
\begin{align}
\label{EQ_DIFF_CS_LF}
dP^{h\rightarrow h_1,h_2}(z_1) =& C_h^{h_1 h_2}\ d p_1^- \int \frac{ d^2 \mathbf{p}_{1\perp}}{(2\pi)^3 2 p_1^-} \frac{d p_2^- d^2 \mathbf{p}_{2\perp}}{(2\pi)^3 2 p_2^-} (2\pi)^4 \delta^4(q-p_1-p_2)\\
=& \frac{C_h^{h_1 h_2}}{8 \pi}\ dz_1 \int_0^\infty d l\ \delta \left(l - \left[z_1 z_2\ {m_h^2} - z_2 m_{h1}^2 - z_1 m_{h2}^2\right] \right)|_{z_2=1 - z_1}\\
= &\left\{
\begin{array}{l l}
\frac{C_h^{h_1 h_2}}{8\pi}\ d z_1 & \quad \text{ if $z_1 z_2\ {m_h^2} - z_2 m_{h1}^2 - z_1 m_{h2}^2 \geq 0; \ z_1+z_2=1$,}\\
0 & \quad \text{otherwise.}\\
\end{array} \right.
\end{align}
Here we note that the numerical values for the $z_1$ ranges for various decays exactly match those shown in the plots in Fig. 2 of Ref.~\cite{Andersson:1977xs}. We determine the $C_h^{h_1 h_2}$ such that the total probabilities $\int dP^{h\to h_1 h_2}$ of the decays to different pairs $\{h_1 h_2\}$ relate as the corresponding relative branching ratios and the normalization condition $\sum_{\{h_1 h_2\}}\int dP^{h\to h_1 h_2}=1$ is satisfied.
\section{Results}
\label{SEC_RES}
The plots in Fig.~\ref{PLOT_FRAG_DRV_NNB} depict the elementary fragmentation functions, $z \hat{d}_u^h(z)$ from a $u$ quark for pseudoscalar, vector meson and nucleon emission with the mass of scalar antidiquark $M_{D}=0.687~\mathrm{GeV}$ from Ref.~\cite{Cloet:2007em} and the following quark-hadron couplings (calculated in Appendix \ref{APP_QM_COUPLING}): $g_{\pi qQ}=3.15,\ g_{K qQ}=3.387,\ g_{\rho qQ}=1.5$, $g_{K^{*} qQ}=1.91$, and $g_{\phi qQ}=2.29$.
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\textwidth]{Plots/Frag-Drv.pdf}
\caption{Elementary fragmentation functions for $u$ quark, $z \hat{d}_u^{h}$, for all included emission channels.}
\label{PLOT_FRAG_DRV_NNB}
\end{figure}
In order to compare our results with experimental measurements or empirical parametrizations we evolve them at next-to-leading order (NLO) from our model scale $Q_{0}^{2}=0.2~{\rm GeV}^{2}$ using the software from Ref.~\cite{Botje:2010ay}. The details of determination of the model scale are given in \cite{Matevosyan:2010hh}.
The solutions for the favored and unfavored fragmentation functions from an $u$ quark to primary hadrons (without any resonance decays), evolved at NLO to a typical scale $Q^2=4~\rm{GeV}^2$, are shown in the plots in Fig.~\ref{PLOT_FRAG_HADRONS}. Here, we note that the fragmentation to nucleons is comparable to the case of vector mesons.
\begin{figure}[phtb]
\centering
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Evol-U-To-Fav.pdf}}
\hspace{0.1cm}
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Evol-U-To-UnFav.pdf}
}
\caption{The solutions for (a) favored and (b) unfavored fragmentation functions from $u$ quark to various hadrons evolved at NLO to $Q^2=4~\rm{GeV^2}$. }
\label{PLOT_FRAG_HADRONS}
\end{figure}
The plots in Fig.~\ref{PLOT_MOM_SUM} depict the total light-cone momentum fractions of the initial quark carried by the emitted hadrons of different species, calculated at model scale of $Q_0^2=0.2~\mathrm{GeV}^2$. (As discussed in Ref.~\cite{Matevosyan:2010hh}, the momentum and isospin sum rules can only be reliably satisfied at model scale, as the NLO QCD evolution kernels have known singularities in the low $z$ region, in practice leaving the values of the fragmentations for $z \lesssim 0.01$ undetermined.) Here we compare the fractions for elementary fragmentation functions, the solutions without the resonance decays and the solutions after the resonance decays. The sum of the fractions from elementary splitting functions calculated in the NJL model amounts to about half of the initial quark's light-cone momentum, where the rest is carried by the remnant quark. In the quark-jet picture, we sum up the light-cone momentum fractions of all hadrons emitted in the chain. This implies that the initial quark transfers all of its light-cone momentum to the produced hadrons, which we confirm in our numerical solutions (the sum of all the fractions is 1 within numerical errors). This shows that after resonance decays, the pions carry approximately $67\%$ and kaons carry about $24\%$ of the initial $u$ (or $d$) quark's light-cone momentum. In the case of an initial $s$-quark, pions carry approximately $25\%$ and kaons carry about $72\%$ of its light-cone momentum. The rest of the momentum is carried by nucleons and antinucleons, where the momentum sum rule is satisfied within $0.1\%$.
\begin{figure}[phtb]
\centering
\subfigure[] {
\includegraphics[width=0.9\textwidth]{Plots/MomSum-U.pdf}}
\vspace{0.1cm}
\subfigure[] {
\includegraphics[width=0.9\textwidth]{Plots/MomSum-S.pdf}
}
\caption{The total fractions of momenta carried by mesons of type $m$ from the jet of (a) $u$ and (b) $s$ quarks calculated using the numerical solutions for fragmentation functions at the model scale $Q_0^2=0.2~\rm{GeV^2}$. Here "splittings" denote the momentum fractions calculated using the elementary splitting functions $\langle z \hat{d}_q^m(z) \rangle$, "NJL-jet" and "with decays" denote the momentum fractions calculated from solutions for the fragmentation functions without and with inclusion of pions and kaons originating from decays of vector meson resonances (the corresponding columns arranged from left to right for each hadron). }
\label{PLOT_MOM_SUM}
\end{figure}
The results for the solutions for fragmentation function $z D_u^{\pi^+}$ and $z D_u^{\pi^-}$ prior to and after the vector meson decays are shown in Fig.~\ref{PLOT_FRAG_UPI_DECAY}. Here, we see that the inclusion the resonance decay slightly changes the shape of the function in mid- to low-$z$ region, bringing it closer to the empirical parametrizations of the experimental data from Refs.~\cite{Hirai:2007cx} (HKNS) and~\cite{deFlorian:2007aj} (DSS). Similarly, the results for $z D_u^{K}$ prior to and after the vector meson decays are shown in Fig.~\ref{PLOT_FRAG_UK_DECAY} and the results for $z D_u^{P}$ and $z D_u^{N}$ are shown in Fig.~\ref{PLOT_FRAG_UPN_DECAY}. For $z D_u^{P}$ the agreement with the empirical parametrizations from Refs.~\cite{Hirai:2007cx} (HKNS) and~\cite{deFlorian:2007hc} (DSS) is reasonable, while our results for $z D_u^{N}$ are well below the empirical curves. This is because we only include the scalar diquarks in our model, making the fragmentation of a $u$ quark to $N$ an unfavored process, while in the empirical parametrizations it is usually assumed $D_u^P(z)\simeq 2 D_u^N(z)$ based on naive $SU(2)$ flavor symmetry arguments. In the future developments of the NJL-jet model, the inclusion of the axial-vector diquark intermediate states in the quark to nucleon splitting process will include a favored channel [depicted in Fig.~\ref{PLOT_FRAG_Q_N}a ] for emission of a neutron from a $u$ quark.
\begin{figure}[htb]
\centering
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Evol-U-To-PiPl.pdf}}
\hspace{0.1cm}
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Evol-U-To-PiMi.pdf}
}
\caption{The solutions for fragmentation function $z D_u^{\pi}(z)$ evolved at NLO to $Q^2=4~\rm{GeV^2}$. The results are compared to the empirical parametrizations of the experimental data from Refs.~\cite{Hirai:2007cx} (HKNS) and~\cite{deFlorian:2007aj} (DSS). Here "NJL-jet" and "with decays" denote the fragmentation functions calculated without and with inclusion of pions originating from decays of vector meson resonances. The shadows show the uncertainties of the empirical functions of Ref.~\cite{Hirai:2007cx}.}
\label{PLOT_FRAG_UPI_DECAY}
\end{figure}
\begin{figure}[htb]
\centering
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Evol-U-To-KPl.pdf}}
\hspace{0.1cm}
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Evol-U-To-KMi.pdf}
}
\caption{Same as Fig.~\ref{PLOT_FRAG_UPI_DECAY} for the case $z D_u^{K}(z)$.}
\label{PLOT_FRAG_UK_DECAY}
\end{figure}
\begin{figure}[htb]
\centering
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Evol-U-To-P.pdf}}
\hspace{0.1cm}
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Evol-U-To-N.pdf}
}
\caption{Same as Fig.~\ref{PLOT_FRAG_UPI_DECAY} for the case $z D_u^{P}(z)$ and $z D_u^{N}(z)$.}
\label{PLOT_FRAG_UPN_DECAY}
\end{figure}
The solutions for the $z D_s^{K^-}$ and $z D_s^{K^+}$ are shown in Fig.~\ref{PLOT_FRAG_SK_DECAY}. Here we see the strong discrepancies in the global fits \cite{Hirai:2007cx} and \cite{deFlorian:2007aj} of the experimental data that illustrates the need for the model calculations providing additional insight into the quark fragmentation process.
\begin{figure}[htb]
\centering
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Evol-S-To-KMi.pdf}}
\hspace{0.1cm}
\subfigure[] {
\includegraphics[width=0.48\textwidth]{Plots/Evol-S-To-KPl.pdf}
}
\caption{Same as Fig.~\ref{PLOT_FRAG_UPI_DECAY} for the case $z D_s^{K}(z)$.}
\label{PLOT_FRAG_SK_DECAY}
\end{figure}
\section{Conclusions and Outlook}
\label{SEC_CONC}
In the current article, we added the vector meson, nucleon, and antinucleon emission channels to NJL-jet framework for calculating quark fragmentation functions. We also included the two-body decays of the vector mesons to pseudoscalars, which allows for easier comparison of the calculated fragmentation functions with the experimental measurements or their parametrizations. We employed the Monte Carlo method to obtain the fragmentation functions in a quark-cascade description. Here we demonstrated that the Monte Carlo approach to calculating the fragmentation functions in NJL-jet framework is a powerful and reliable method. We reproduced the fragmentation functions calculated as solutions of the previously employed integral equations, where only the light quarks and pions were included. Moreover, we showed that the MC approach allows for the flexibility to surpass the model limitations necessary in formulating the integral equations. That is, in the future MC studies we can assume an initial quark carrying only a finite momentum and thus emitting a finite number of hadrons. We demonstrated that the medium and low $z$ regions of the fragmentation functions are greatly affected by the number of the emitted hadrons, thus the finiteness of the quark momentum might have a noticeable effect. The future development of the NJL-jet model would also allow an access to the transverse momentum distribution of the produced hadrons, thus becoming relevant for the analysis of a large variety of semi-inclusive data. The MC approach provides a robust and efficient platform for implementing these and other possible extensions of the NJL-jet model that would allow for a much more detailed description of the physical picture.
A further advantage of the MC approach is in reducing the numerical task in solving for the fragmentation functions when including many more channels for emitted hadrons. Here, solving the integral equations requires inverting larger and larger matrices, while the MC procedure can be drastically sped up by trivially parallelizing the task and solving simultaneously on computer clusters.
The results for the fragmentation functions exhibit only slight changes with addition of the new hadronic channels. In particular, vector meson-quark couplings are relatively small in our model, lowering their relative contribution to the fragmentation process. The resulting quark fragmentation functions exhibit a good agreement with the empirical parametrizations in high $z$ region, staying reasonably close to them also in mid and low $z$ regions in part due to the slight modifications of the shape of the functions in this region brought by the effects of the vector meson decays. On the other hand, the plots in Fig.~\ref{PLOT_MOM_SUM} show that pions, kaons, nucleons, and antinucleons include the most dominant contributions in fragmentation process, thus the inclusion of higher resonance states is unlikely to be important. A notable underestimation of the fragmentation of $u$ quark to neutrons, shown on plots in Fig.~\ref{PLOT_FRAG_UPN_DECAY}b, demonstrates the need to include the axial-vector diquark for a more realistic description of the fragmentations to nucleons, which will be accomplished in the future developments of the model.
In this work, we followed our previous NJL-jet calculations and used the LB scheme to calculate the regularized fragmentation functions. However, as we have pointed out, the inclusion of vector mesons in the NJL model favors the PT regularization scheme, which is free of unphysical decay thresholds. Here, we have used this scheme only to assess the vector meson-quark couplings, but in future work on extensions of the model we will use the PT scheme consistently throughout.
The future development of the NJL-jet model using the Monte Carlo framework will allow us to study the transverse momentum dependence of the quark fragmentation functions as well as polarized fragmentation functions. This can be accomplished within the NJL framework, without inclusion of any additional parameters.
Last, the medium modifications of the fragmentation functions may be essential to our understanding of the semi-inclusive processes on nuclear targets. Medium effects have long been studied in the NJL model, yielding a successful description of modifications of nucleon properties in nuclei~\cite{Bentz:2001vc,Mineo:2003vc, Cloet:2005pp,Cloet:2006bq}. Thus the model should provide a reliable framework for the calculation of medium modifications of quark hadronization process.
\section*{Acknowledgements}
This work was supported by the Australian Research Council through Grant No. FL0992247 (A.W.T.), by the University of Adelaide and by a Subsidy for Activating Educational Institutions from the Department of Physics, Tokai University.
|
\section{Introduction}
In this paper we study properties of the horizontal foliations of certain kinds of flat bundles. The paper divides into two main parts, the first of which is concerned with closed leaves of flat surface bundles and the second of which deals with fillings of flat circle bundles.
The existence of flat bundles with compact fibre that have closed leaves with non-trivial normal bundles is well known, with explicit examples given by certain flat structures of sphere bundles (cf.\ \cite{Mit}). This leads to the question of whether similar foliations exist on more general kinds of bundles. In view of this, we show the existence of foliated surface bundles with closed leaves that have prescribed self-intersection numbers, where the fibre can be taken to be any surface of sufficiently large genus. This is proved by a variation of the stabilisation trick of \cite{KM} that was originally used to show the existence of flat surface bundles with non-zero signatures. Indeed, given a surface bundle with a section $S$, whose self-intersection number is divisible by the Euler characteristic of the fibre, we show that $S$ can be made the leaf of a flat structure after stabilisation (Theorem \ref{leaf_stabilisation}).
To show the existence of flat surface bundles possessing closed leaves with non-trivial normal bundles, we are naturally led to certain calculations involving the homology of surface diffeomorphism groups with marked points. In particular, we show the following:
\begin{thm}
Let $\Sigma_h$ be a surface of genus $h \geq 3$ and let $k \geq 2$. Then
\[H_1(Diff_{\delta}^+(\Sigma_{h,k})) = \mathbb{R}^+ \times \mathbb{Z}_2.\]
\end{thm}
\noindent We also show that the Abelianisation of the group of compactly supported diffeomorphism of $\mathbb{R}^2$ that fix the origin is $ \mathbb{R}^+$. This is a special case of a result of \cite{Fuk}, although the proof given there seems to overlook a small, but important, technical point (see discussion preceding Theorem \ref{Fukui}). Moreover, our proof, which relies on Sternberg Linearisation, is briefer than and independent of Fukui's original argument.
The stabilisation trick can also be applied to obtain foliations that have transverse symplectic structures. As in the case of flat bundles we show the existence of symplectically flat surface bundles with closed leaves having prescribed self-intersection numbers (Proposition \ref{symp_leaves_construction}). These results yield the existence of manifolds with certain kinds of symplectic pairs, as defined in \cite{KM}. Indeed, by applying the normal connected sum operation we deduce the existence of symplectic pairs both of whose foliations have closed leaves with non-trivial normal bundles (Corollary \ref{Pairs_no_triv_leaves}).
Given a flat circle bundle, one can consider the problem of extending the flat structure to the interior of a surface bundle. If the fibre is assumed to be a disc, then there is a dichotomy depending on whether one requires that the foliation is symplectic or not. For in the smooth case, it is relatively easy to show that any flat circle bundle over a surface admits a flat disc bundle filling after stabilisation (Proposition \ref{boundary_Euler_class}). However, in the symplectic case the Euler class provides an obstruction by a result of Tsuboi. In fact, Tsuboi gave the following formula for computing the Euler class of a flat circle bundle in terms of the Calabi homomorphism of certain extensions of the boundary holonomy to the interior of a disc:
\begin{thm}[\cite{Tsu}]\label{Tsuboi_int}
Let $\pi_1(\Sigma_g) \stackrel{\psi} \rightarrow Diff_0(S^1)$ be a homomorphism and let $a_i, b_i$ be standard generators of $\pi_1(\Sigma_g)$. Furthermore, let $f_i, h_i \in Symp(D^2)$ be extensions of $\psi(a_i), \psi(b_i)$ respectively and let $e(E)$ denote the Euler class of the total space of the $S^1$-bundle $E$. Then
\[-\pi^2 e(E) = Cal([f_1, h_1]...[f_g, h_g]).\]
\end{thm}
\noindent Tsuboi's result can be reformulated in terms of the five-term exact sequence in group cohomology. The advantage of this reformulation is that it can easily be generalised to the case where the fibre of the filling is an arbitrary surface and the extensions of the boundary holonomies lie in the extended Hamiltonian group (Theorem \ref{Hamiltonian_extended_Tsuboi}).
As a consequence we see that the Euler class gives an obstruction to filling a circle bundle by a flat surface bundle with holonomy in the extended Hamiltonian group. We contrast this result with the fact that any flat circle bundle can be filled by a flat symplectic bundle after stabilisation (Theorem \ref{symp_flat_extension}).
Another important topological quantity associated with a surface bundle is its signature. By considering appropriate extended Hamiltonian groups, we will derive a Tsuboi-type formula for the signature of bundles with fibre a once-punctured surface $\Sigma_h^1$.
\begin{thm}
Let $\Sigma^1_h \to E \to \Sigma_g$ be a bundle with holonomy representation $\pi_1(\Sigma_g) \stackrel{\rho} \rightarrow \Gamma_h^1$ and assume $h \geq 2$. Furthermore, let $\alpha_i = \rho(a_i)$ and $\beta_i = \rho(b_i)$ be the images of standard generators $a_i,b_i$ of $\pi_1(\Sigma_g)$. Then for any lifts $\phi_i, \psi_i \in \widetilde{Ham^c}(\Sigma_h^1)$ of $\alpha_i, \beta_i$ the signature satisfies
\[\sigma(E) = Cal([\phi_1, \psi_1]...[\phi_g, \psi_g ]).\]
\end{thm}
\noindent As is the case for fillings of circle bundles, this then implies restrictions on the topology of bundles whose holonomy groups lie in the extended Hamiltonian group.
\subsection*{Outline of paper:}
In Section \ref{closed_leaves_horizontal} we show the existence of flat bundles possessing closed leaves with non-trivial normal bundles and in Section \ref{points} we compute the Abelianisation of diffeomorphism groups with marked points. In Section \ref{symp_flat_leaves} we derive similar results for symplectic surface bundles. Then after showing that the problem of filling circle bundles is solvable after stabilisation in Section \ref{fill_poss}, we recast Tsuboi's result in the language of the five-term exact sequence and extend it to higher genus fillings in Section \ref{Flat_Ham}. In Section \ref{MMM_first} we derive the Tsuboi-type formula for the signature of surface bundles with once-punctured fibre.
For the sake of completeness we give an explicit description of the five-term exact sequence in group cohomology in Appendix \ref{App_five_term}.
\subsection*{Acknowledgments:}
The results of this article are taken from the author's doctoral thesis, which would not have been possible without the encouragement and support of Prof.\ D.\ Kotschick. The financial support of the Deutsche Forschungsgemeinschaft is also gratefully acknowledged.
\subsection*{Notation and Conventions:}
Throughout this paper $\Sigma^r_{g,k}$ will denote a compact, oriented genus $g$ surface with $k$ marked points and $r$ boundary components. By $\Gamma^r_{g,k}$ we shall denote the mapping class group of diffeomorphisms that preserve $k$ marked points and have support in the interior of $\Sigma^r_g$.
All bundles will be assumed to be oriented and all maps are smooth. Unless otherwise stated all homology groups will be taken with integral coefficients. Finally a topological group will be decorated with a $\delta$ when it is to be considered as a discrete group.
\section{Closed leaves and horizontal foliations}\label{closed_leaves_horizontal}
Interesting examples of codimension two foliations come from considering the horizontal foliations of flat surface bundles. In this section we will focus on the closed leaves of such foliations. Any surface bundle $\Sigma_h \to E \to B$ determines a holonomy representation \[\pi_1(B) \stackrel{\rho} \rightarrow \Gamma_h.\]
Such a bundle is then \emph{flat} if its holonomy representation $\rho$ admits a lift to the group of orientation preserving diffeomorphisms $Diff^+(\Sigma_h)$
\[\xymatrix{ & Diff^+(\Sigma_h) \ar[d] \\
\pi_1(B) \ar[r]^{\rho} \ar@{-->}[ur]^{\bar{\rho}} & \Gamma_h.} \]
If $B$ is a manifold, then this is equivalent to the existence of a foliation that is complementary to the fibres. Such a foliation will be called a \emph{horizontal} foliation.
For a flat bundle $E$ over a manifold a closed leaf of the horizontal foliation intersects each fibre in $k$ points, where $k$ is the homological intersection number of the leaf with a fibre. The existence of such a foliation is thus equivalent to a lift of the holonomy map $\rho$ to the group of diffeomorphisms fixing $k$ marked points $Diff^+(\Sigma_{h,k})$
\[\xymatrix{ & Diff^+(\Sigma_{h,k}) \ar[d] \\
\pi_1(B) \ar[r]^{\rho} \ar@{-->}[ur]^{\bar{\rho}} & \Gamma_h.} \]
By taking pullbacks under a suitable finite cover of the base, one obtains a horizontal foliation with a leaf $S$ that intersects each fibre exactly once, in which case the holonomy of $E$ lies in $Diff^+(\Sigma_{h,1})$. Moreover, the horizontal foliation induces a flat structure on the normal bundle $\nu_S$ of $S$, which is given by composing $\bar{\rho}$ with the derivative map at $p$:
\[Diff^+(\Sigma_{h,1}) \stackrel{D_p} \rightarrow GL^+(T_p \Sigma_g) = GL^+(2, \mathbb{R}).\]
In \cite{Mil}, Milnor constructed flat bundles with non-trivial Euler class over oriented surfaces and, hence, the image of the Euler class in $H^2(GL_{\delta}^+(2, \mathbb{R}))$ is non-trivial. In view of this, to show that there are flat bundles with horizontal leaves of non-zero self-intersection it will suffice to show that the map $D_p$ induces an injection $H^2(GL_{\delta}^+(2,\mathbb{R})) \rightarrow H^2(Diff_{\delta}^+(\Sigma_{h,1}))$.
To this end, we let $\mathcal{G}_p$ be the group of smooth diffeomorphism germs that fix the marked point $p$. We then define
\[\xymatrix{Diff^+(\Sigma_{h,1}) \ar[r]^<<<<\pi & \mathcal{G}_p \ar[r]^<<<<<<{\bar{D}_p} & \ar @/^1pc/[l]^s GL^+(T_p \Sigma_g)},\]
where $\pi$ is the map taking a diffeomorphism fixing $p$ to its germ at $p$ and $\bar{D}_p$ maps a germ to its linear part. The kernel of the map $\pi$ consists of diffeomorphisms with support disjoint from the marked point $p$ and will be denoted by $Diff^c(\Sigma_{h,1})$. The final map has an obvious section given by considering a linear map as an element of $\mathcal{G}_p$. Thus, to show that the map $H^2(GL_{\delta}^+(2,\mathbb{R})) \rightarrow H^2(Diff_{\delta}^+(\Sigma_{h,1}))$ is injective, it will be sufficient to show that $H^2(\mathcal{G}_p) \rightarrow H^2(Diff_{\delta}^+(\Sigma_{h,1}))$ is injective. To this end we first note the following lemma, which is proved by a simple cut-off argument (cf.\ \cite{Bow2}, Lemma 4.1.3).
\begin{lem}\label{exact_sequence}
The following sequence of groups is exact
\[ 1 \to Diff^c(\Sigma_{h,1}) \to Diff^+(\Sigma_{h,1}) \rightarrow \mathcal{G}_p \to 1.\]
\end{lem}
\noindent We may now prove the existence of horizontal foliations that have compact leaves with non-trivial self-intersection numbers.
\begin{prop}\label{horizontal_leaves}
If $h \geq 3$, then there exist flat surface bundles $\Sigma_h \to E \to \Sigma_g$ with horizontal foliations that have leaves of non-zero self-intersection.
\end{prop}
\begin{proof}
We consider the last three terms of the five-term exact sequence in cohomology associated to the extension of groups in Lemma \ref{exact_sequence}:
\[ H^1(Diff_{\delta}^c(\Sigma_{h,1}))^{\mathcal{G}_p} \to H^2(\mathcal{G}_p) \stackrel{\pi^*} \rightarrow H^2(Diff^+_{\delta}(\Sigma_{h,1})).\]
From our discussion above it is sufficient to show that the map $\pi^*$ is injective or by exactness that $H^1(Diff_{\delta}^c(\Sigma_{h,1}))^{\mathcal{G}_p} = 0$. We claim that $H_1(Diff_{\delta}^c(\Sigma_{h,1}))$ is in fact trivial and by the Universal Coefficient Theorem the same holds in cohomology.
Let $\Sigma^{\epsilon}_h = \Sigma_h \setminus D_{\epsilon}$ denote $\Sigma_h$ with a disc of radius $\epsilon$ removed. We note that $Diff^c(\Sigma_{h,1})$ is isomorphic to the direct limit of the groups $Diff^c(\Sigma^{\epsilon}_h)$. By the stability result of Harer $H_1(\Gamma_h^1) = H_1(\Gamma_h)$ for $h \geq 3$ (cf.\ \cite{Iva}). Moreover, $\Gamma_h$ is perfect for $h \geq 3$ (see \cite{Pow}). Finally, by the classical result of Thurston the identity component of $Diff^c(\Sigma^{\epsilon}_{h})$ is also perfect (see \cite{Th2}). The five-term sequence in homology then implies that $H_1(Diff_{\delta}^c(\Sigma^{\epsilon}_h)) = 0$. Hence, each of the groups $H_1(Diff_{\delta}^c(\Sigma^{\epsilon}_h))$ is trivial and we conclude that $H_1(Diff_{\delta}^c(\Sigma_{h,1}))$ also vanishes.
\end{proof}
One may interpret the proof of Proposition \ref{horizontal_leaves} in a more geometric fashion, which gives a sharper result, and we note this in the following proposition.
\begin{prop}\label{horizontal_leaves2}
If $h \geq 3$ and $k \in \mathbb{Z}$, then there exist flat surface bundles $\Sigma_h \to E \to \Sigma_g$ with horizontal foliations that have closed leaves of self-intersection $k$.
\end{prop}
\begin{proof}
Let $a_i, b_i \in \pi_1(\Sigma_g)$ denote the standard generators of the fundamental group and let $\xi_k$ be a flat $GL^+(2, \mathbb{R})$-bundle over $\Sigma_g$ that has Euler class $k \leq g-1$ as provided by \cite{Mil}. This corresponds to a holonomy representation
\[a_i \mapsto A_i\]
\[b_i \mapsto B_i\]
for $A_i, B_i \in GL^+(2, \mathbb{R})$. Then by Lemma \ref{exact_sequence} there are diffeomorphisms $\phi_i, \psi_i$ which agree with $A_i, B_i$ in a small neighbourhood of $p$ so that the product $\eta = \prod_{i = 1}^{g} [\phi_i, \psi_i]$ has support disjoint from $p$, that is $\eta \in Diff^c(\Sigma_{h,1})$. This group is perfect and, thus, we may write $\eta^{-1} = \prod_{i = 1}^{g'} [\alpha_i, \beta_i]$ where $\alpha_i, \beta_i \in Diff^c(\Sigma_{h,1})$. We define a flat bundle over $\Sigma_{g + g'}$ by the holonomy representation
\[a_i \mapsto \phi_i, \ \ b_i \mapsto \psi_i \text{ for } 1 \leq i \leq g\]
\[a_{g +j} \mapsto \alpha_j, \ \ b_{g +j} \mapsto \beta_j \text{ for } 1 \leq j \leq g',\]
which we denote by $\rho$. This bundle then has a compact leaf $S$ corresponding to the marked point $p$. The Euler class of the normal bundle $\nu_S$ to $S$ is computed from the induced holonomy representation $D_p \rho$:
\[a_i \mapsto D_p(\phi_i) = A_i, \ \ b_i \mapsto D_p(\psi_i) = B_i \text{ for } 1 \leq i \leq g\]
\[a_{g +j} \mapsto D_p(\alpha_j) = Id, \ \ b_{g +j} \mapsto D_p(\beta_j) = Id \text{ for } 1 \leq j \leq g'.\]
It then follows that $e(\nu_S) = k$ and thus $[S]^2 = k$.
\end{proof}
With the geometric construction of Proposition \ref{horizontal_leaves2} we are now able to say when a section $S$ of a bundle $E$ can be made a leaf of a horizontal foliation after \emph{stabilisation}.
\begin{defn}
A surface bundle $E'$ over a surface is called a stabilisation of a bundle $E$, if it is the fibre sum of $E$ with a trivial bundle $\Sigma_{h} \times \Sigma_{g'}$. This is then a bundle over the connected sum $\Sigma_g \# \Sigma_{g'} = \Sigma_{g + g'}$ that is trivial over the second factor.
\end{defn}
We will show that under certain conditions any bundle $E$ with a section $S$ of self-intersection $k$ can be stabilised to a bundle $E'$ that admits a horizontal foliation with a closed leaf $S'$ that agrees with $S$ on $E'|_{\Sigma_g}$.
If the bundle $E$ is trivial then after stabilisation it remains trivial. For a trivial bundle the Euler class of the vertical tangent bundle $e(E)$ is divisible by $2h-2$ and, hence, the same is true for the self-intersection of $S$ and its stabilisation $S'$. Thus, the condition that $[S]^2$ is divisible by $2h-2$ is, in general, necessary for the existence of a stabilisation of the desired form. It is, however, also sufficient and this is the content of the following theorem.
\begin{thm}\label{leaf_stabilisation}
Let $\Sigma_h \to E \to \Sigma_g$ be a surface bundle that has a section of self-intersection $k$, where $k$ is divisible by $2h-2$. Then after stabilisation $E$ admits a flat structure whose horizontal foliation has a closed leaf of self-intersection $k$.
\end{thm}
\begin{proof}
We first stabilise $E$ until the Milnor-Wood equality is satisfied for $S$. We let $\bar{g} = g + g'$ denote the genus of the base of the stabilisation and let $\bar{\rho}: \pi_1(\Sigma_{\bar{g}}) \to \Gamma_{h,1}$ be its holonomy representation. Since the Milnor-Wood inequality is satisfied for $S$, it has a tubular neighbourhood $\nu_S$ that is diffeomorphic to a flat $GL^+(2,\mathbb{R})$-bundle. We let $\xi$ denote the corresponding horizontal foliation on $\nu_S$. We extend $\xi$ to a horizontal distribution $\xi'$ that agrees with $\xi$ on a (possibly smaller) neighbourhood of $S$. We choose curves $a_i,b_i$ representing the standard generators of $\pi_1(\Sigma_{\bar{g}})$ and let $\phi_i, \psi_i \in Diff^+(\Sigma_h)$ be the holonomy maps induced by $\xi'$, so that $[\phi_i] = \bar{\rho}(a_i)$ and $\ [\psi_i] = \bar{\rho}(b_i)$ in $\Gamma_h$. Note that these diffeomorphisms depend on the choice of curves and not just their homotopy classes.
By construction the distribution $\xi'$ is a foliation in a neighbourhood of $S$. Hence the product of commutators $\eta= \prod_{i = 1}^{\bar{g}} [\phi_i, \psi_i]$ has compact support disjoint from the marked point corresponding to the section $S$, and is thus an element in $Diff^c(\Sigma_h^1)$. We next consider the following diagram that relates the mapping class groups $\Gamma_h^1, \Gamma_{h,1}$ and $\Gamma_h$:
\[\xymatrix{ & 1 \ar[d] & 1 \ar[d] & & \\
& \mathbb{Z} \ar[d] & \mathbb{Z} \ar[d] & &\\
1 \ar[r] & \pi_1(T_1\Sigma_h) \ar[r] \ar[d] & \Gamma_h^1 \ar[d] \ar[r] & \Gamma_h \ar[r]& 1\\
1 \ar[r] & \pi_1(\Sigma_h) \ar[r] \ar[d] & \Gamma_{h,1} \ar[d] \ar[r] & \Gamma_h \ar[r]&1.\\
& 1 & 1 & & }\]
Here $\mathbb{Z}$ is generated by a positive Dehn twist $\Delta_{\partial}$ along an embedded curve parallel to the boundary of $\Sigma_h^1$ and $T_1\Sigma_h$ denotes the unit tangent bundle of $\Sigma_h$.
Now the image of $\eta$ in $\Gamma_{h,1}$ is trivial. Thus $\eta = \Delta_{\partial}^k$ where $k$ is the self-intersection of $S$. This is because the second column is the central extension corresponding to the vertical Euler class as a characteristic class in the group cohomology of $\Gamma_{h,1}$ (cf.\ \cite{Mor}). By assumption $k$ is divisible by $2h-2$ and hence $\eta \in H_1(\pi_1(T_1\Sigma_h)) = H_1(T_1\Sigma_h)$ is trivial. Again this is because the left most column is the central extension corresponding to the Euler class of the unit tangent bundle over $\Sigma_h$ and $[\eta] $ is a multiple of the fibre class of this $S^1$-bundle that is divisible by $2h -2$. Hence $\eta^{-1} = \prod_{j = 1}^{N} [\alpha_j, \beta_j]$ is a product of commutators in $\Gamma_h^1$, each of which lie in the kernel of the natural map to $\Gamma_h$.
We let $\phi_j, \psi_j \in Diff^+(\Sigma_h^1)$ be representatives of the mapping classes $\alpha_j, \beta_j$ respectively, and consider the product $\gamma = \eta \thinspace \prod_{j = 1}^{N} [\phi_j, \psi_j]$ in $Diff_0^c(\Sigma_h^1)$. Since this group is perfect we may write $\gamma^{-1} = \prod_{l = 1}^{M} [\gamma_l, \delta_l]$. Letting $a_i, b_i$ be standard the generators for $\pi_1(\Sigma_{\bar{g} +N +M})$, we define a flat bundle via the holonomy map
\[a_i \mapsto \phi_i, \ \ b_i \mapsto \psi_i \text{ for } 1 \leq i \leq \bar{g}\]
\[a_{\bar{g} +j} \mapsto \alpha_j, \ \ b_{\bar{g} +j} \mapsto \beta_j \text{ for } 1 \leq j \leq N\]
\[a_{\bar{g} +N + l} \mapsto \gamma_l, \ \ b_{\bar{g} + N + l} \mapsto \delta_l \text{ for } 1 \leq l \leq M.\]
This gives a horizontal foliation with a closed leaf of self-intersection $k$ on a stabilisation of $E$. That the bundle is a stabilisation of the original bundle follows since the mapping classes represented by $\alpha_j,\beta_j, \gamma_l,, \delta_l$ are trivial in $\Gamma_h$.
\end{proof}
In \cite{BCS} Bestvina, Church and Souto show the non-existence of certain lifts of bundles with sections to the diffeomorphism group with marked points, using the bounds on the Euler class given by the Milnor-Wood inequality. In particular, the diagonal section in the product of two genus $h$ surfaces provides such an example. However, by Theorem \ref{leaf_stabilisation}, these examples do possess lifts after stabilisation.
\section{Abelianisation of diffeomorphism groups with marked points}\label{points}
Our discussion above will enable us to calculate the first group homology of $Diff^+(\Sigma_{h,1})$ and, in particular, we will show that this group is not perfect. In fact it is clear that the group $Diff^+(\Sigma_{h,1})$ is not perfect, as there is a surjection to $GL^+(2, \mathbb{R})$ given by the derivative map and
\[H_1(GL_{\delta}^+(2, \mathbb{R})) = H_1((SL(2, \mathbb{R}) \times \mathbb{R}^+)_{\delta}) = \mathbb{R}^+, \]
since $SL(2, \mathbb{R})$ is a perfect group. But this is the only contribution to $H_1(Diff_{\delta}^+(\Sigma_{h,1}))$ if $h \geq 3$. The proof of this fact is based on the following result due to Sternberg.
\begin{thm}[Sternberg's Linearisation Theorem, \cite{Ster}]\label{Sternberg_linearisation}
Let $\phi$ be a smooth diffeomorphism defined in a neighbourhood $U$ of the origin in $\mathbb{R}^n$ and let $\phi(0) = 0$. Further, let $s_1, ...s_n \in \mathbb{C}$ denote the eigenvalues (counted with multiplicities) of the Jacobian $D_0 (\phi)$ at the origin and assume that
\[s_i \neq s_1^{m_1}...s_n^{m_n},\]
for all non-negative integers $m_1, ..., m_n$ with $\sum m_i > 1$. Then there is a change of coordinates $\psi$ that fixes the origin so that on a possibly smaller neighbourhood $W \subset U$ the following holds
\[ \psi \phi \psi ^{-1} = D_0 (\phi).\]
\end{thm}
\begin{rem}\label{Germ_version} We note that the hypotheses of Theorem \ref{Sternberg_linearisation} hold, in particular, if $D_0 (\phi) = \lambda \thinspace Id$ for $\lambda \neq 0, 1$. Sternberg's Theorem may also be interpreted in terms of germs of diffeomorphisms, i.e. if the hypotheses of the theorem are satisfied for a germ $\phi \in \mathcal{G}_p$, then $\phi$ is conjugate to the germ represented by $D_p (\phi)$.
\end{rem}
\begin{prop}\label{not_perfect}
Let $\Sigma_h$ be a surface of genus $h \geq 3$. Then $H_1(Diff_{\delta}^+(\Sigma_{h,1})) = \mathbb{R}^+$.
\end{prop}
\begin{proof}
We consider the extension given in Lemma \ref{exact_sequence}
\[ 1 \to Diff^c(\Sigma_{h,1}) \to Diff^+(\Sigma_{h,1}) \stackrel{\pi} \rightarrow \mathcal{G}_p \to 1.\]
Since the group $Diff^c(\Sigma_{h,1})$ is perfect (cf.\ Proposition \ref{horizontal_leaves}), the associated five-term exact sequence yields $H_1(Diff_{\delta}^+(\Sigma_{h,1})) = H_1(\mathcal{G}_p)$. We next consider the exact sequence
\[ 1 \to \mathcal{G}_{p,Id} \to \mathcal{G}_p \stackrel{D_p} \rightarrow GL^+(2, \mathbb{R}) \to 1,\]
where $ \mathcal{G}_{p,Id} $ is the set of germs whose linear part is the identity. By Remark \ref{Germ_version} above, if $\phi \in \mathcal{G}_{p}$ and $D_p (\phi) = \lambda \thinspace Id$ for some $\lambda > 1$, then there is a $ \psi \in \mathcal{G}_{p}$ so that $\psi \phi \psi ^{-1} = \lambda \thinspace Id$. We set $A_{\lambda} = \lambda \thinspace Id$, then for any $\phi \in \mathcal{G}_{p,Id}$, there is a germ $\psi$ such that
\[A_{\lambda} = \psi ( A_{\lambda} \phi ) \psi ^{-1}.\]
Since $A_{\lambda}$ is central in $GL^+(2,\mathbb{R})$, we may assume that $\psi \in \mathcal{G}_{p,Id}$ after conjugating the above equation with the element $D_p \psi$. Thus $\phi = \psi^{-1} A_{\lambda}^{-1} \psi A_{\lambda} = [\psi^{-1}, A_{\lambda}^{-1}]$ and we have shown that $H_1(\mathcal{G}_{p,Id})_{GL^+(2, \mathbb{R})} = 0$.
In view of this, the five-term exact sequence gives
\[ 0 = H_1(G_{p,Id})_{GL^+(2, \mathbb{R})} \to H_1(\mathcal{G}_p ) \to H_1(GL_{\delta}^+(2, \mathbb{R})) \to 0\]
and, hence, $H_1(Diff_{\delta}^+(\Sigma_{h,1}))=H_1(GL_{\delta}^+(2,\mathbb{R})) = \mathbb{R}^+$.
\end{proof}
We let $PDiff^+(\Sigma_{h,k})$ denote the group of pure orientation preserving diffeomorphisms, i.e. an element $\phi \in PDiff^+(\Sigma_{h,k})$ is a diffeomorphism of $\Sigma_h$, which fixes a set of $k$ marked points \emph{pointwise}. With exactly the same argument as in Proposition \ref{not_perfect} we obtain the following.
\begin{prop}\label{not_perfect2}
Let $\Sigma_h$ be a surface of genus $h \geq 3$. Then $H_1(PDiff_{\delta}^+(\Sigma_{h,k})) = (\mathbb{R}^+)^k$.
\end{prop}
Using Proposition \ref{not_perfect2} it is now possible to compute the first homology of the full diffeomorphism group $Diff^+(\Sigma_{h,k})$.
\begin{thm}\label{not_perfect3}
Let $\Sigma_h$ be a surface of genus $h \geq 3$ and let $k \geq 2$. Then
\[H_1(Diff_{\delta}^+(\Sigma_{h,k})) = \mathbb{R}^+ \times \mathbb{Z}_2.\]
\end{thm}
\begin{proof}
By considering the action on the marked points induced by $Diff^+(\Sigma_{h,k})$, we obtain the following extension of groups
\[1 \to PDiff^+(\Sigma_{h,k}) \to Diff(\Sigma_{h,k}) \to S_k \to 1.\]
The five-term exact sequence for group homology then gives
\[ H_2(S_k) \stackrel{\partial} \rightarrow H_1(PDiff_{\delta}^+(\Sigma_{h,k}))_{S_k} \to H_1(Diff_{\delta}^+(\Sigma_{h,k})) \to H_1(S_k) \to 0.\]
By Proposition \ref{not_perfect2} we have that $H_1(PDiff_{\delta}^+(\Sigma_{h,k})) = (\mathbb{R}^+)^k$, which, in particular, implies that $H_1(PDiff_{\delta}^+(\Sigma_{h,k}))_{S_k} = \mathbb{R}^+$. Since $S_k$ is a finite, the group $H_2(S_k)$ consists entirely of torsion elements. Hence, as $\mathbb{R}^+$ is torsion free the connecting homomorphism $\partial$ is trivial and we obtain the following short exact sequence:
\[0 \to \mathbb{R}^+ \to H_1(Diff_{\delta}^+(\Sigma_{h,k})) \to H_1(S_k) = \mathbb{Z}_2 \to 0.\]
Finally, since in $\mathbb{R}^+$ every element has a square root, this extension has a section and we conclude that
\[H_1(Diff_{\delta}^+(\Sigma_{h,k})) = \mathbb{R}^+ \times \mathbb{Z}_2. \qedhere\]
\end{proof}
The proof of Proposition \ref{not_perfect} above will allow us to calculate the first group homology of $Diff^c(\mathbb{R}^2,0)$, which here denotes the group of diffeomorphisms of the plane that have compact support and fix the origin. This fact was stated in a more general form by Fukui in \cite{Fuk}, however his argument appears to be incomplete.
Fukui argues as follows (see \cite{Fuk}, p.\ 485). Let $\phi \in Diff^c(\mathbb{R}^n,0)$ have $D_0 \phi = Id$, then there is a product of commutators so that $\eta = \phi \prod_{i = 1}^{g'} [\alpha_i, \beta_i]$ is the identity on some neighbourhood of $0$. He then claims that by Thurston's result on the perfectness of the identity component the group of diffeomorphisms, we may write $\eta$ as a product of commutators of elements in $Diff^c(\mathbb{R}^n \setminus \{0 \})$, which denotes the group of compactly supported diffeomorphisms of $\mathbb{R}^n \setminus 0$. In order to apply the result of Thurston, one must have that $\eta$ is isotopic to the identity through diffeomorphisms with compact support away from the origin. However, it is not clear that $\eta$ is isotopic to the identity through diffeomorphisms with support disjoint from the origin. In fact for $n = 2$ the mapping class group of compactly supported diffeomorphisms on $\mathbb{R}^2 \setminus \{ 0 \}$ is isomorphic to $\mathbb{Z}$ (cf.\ \cite{Iva}, Cor.\ 2.7 E). As a corollary of the results we have obtained thus far we are now able to give a complete proof of the theorem stated by Fukui in the case $n = 2$.
\begin{thm}\label{Fukui}
$H_1(Diff^c_{\delta}(\mathbb{R}^2, 0)) = \mathbb{R}^{+}$.
\end{thm}
\begin{proof}
We have the following exact sequences
\begin{equation*}
1 \to Diff^c(\mathbb{R}^2 \setminus \{ 0 \}) \to Diff^c(\mathbb{R}^2, 0) \stackrel{\pi} \rightarrow \mathcal{G}_p \to 1
\end{equation*}
and
\begin{equation*}
1 \to Diff_0^c(\mathbb{R}^2 \setminus \{ 0 \}) \to Diff^c(\mathbb{R}^2 \setminus \{ 0 \}) \rightarrow \mathbb{Z} \to 1.
\end{equation*}
We consider the five-term sequence in cohomology associated to the first exact sequence above
\begin{align*}
0 \to H^1(\mathcal{G}_p) \to H^1(Diff^c_{\delta}(\mathbb{R}^2, 0)) &\to H^1(Diff^c_{\delta}(\mathbb{R}^2 \setminus \{ 0 \}))^{\mathcal{G}_p}\\
& \stackrel{\delta} \rightarrow H^2(\mathcal{G}_p) \to H^2(Diff^c_{\delta}(\mathbb{R}^2, 0)).
\end{align*}
By Thurston's result $Diff_0^c(\mathbb{R}^2 \setminus \{ 0 \})$ is perfect and applying the five-term exact sequence to the second exact sequence above implies that $H^1(Diff^c_{\delta}(\mathbb{R}^2 \setminus \{ 0 \})) = \mathbb{Z}$.
Next we consider the sequence of classifying spaces
\[\xymatrix{B Diff^c_{\delta}(\mathbb{R}^2, 0) \ar[r] & B \mathcal{G}_p \ar[r] & B GL_{\delta}^+(2, \mathbb{R}) \ar[r] & B GL^+(2, \mathbb{R}).}\]
The Euler class is a generator of $H^2(B GL^+(2, \mathbb{R}))$ and the pullback to $H^2(\mathcal{G}_p)$ is non-zero and primitive, as one sees by evaluating this class on a flat $GL^+(2, \mathbb{R})$-bundle with Euler class 1, thought of as an element in $H_2(\mathcal{G}_p)$. Moreover, a flat bundle with holonomy in $ Diff^c(\mathbb{R}^2, 0)$ is topologically trivial, since it admits a section with vanishing self-intersection number. Hence the pullback of $e$ to $H^2( Diff^c_{\delta}(\mathbb{R}^2, 0))$ is zero. By exactness of the five-term sequence above $e = \delta(f) \text{ for some } f \in H^1(Diff^c_{\delta}(\mathbb{R}^2 \setminus \{ 0 \}))^{\mathcal{G}_p} \subset \mathbb{Z}$. Hence as $e$ is a primitive, non-torsion class, the connecting homomorphism for the five-term exact sequence in \emph{homology} must be surjective. Thus by exactness
\[ H_1(Diff^c_{\delta}(\mathbb{R}^2, 0)) = H_1(\mathcal{G}_p) = \mathbb{R}^+,\]
where the second equality was shown in the proof of Proposition \ref{not_perfect}.
\end{proof}
\section{Closed leaves of flat bundles with symplectic holonomy}\label{symp_flat_leaves}
One may consider flat bundles with additional structure. In the context of surface bundles with horizontal foliations it is natural to consider bundles whose horizontal foliations are transversally symplectic. This is equivalent to the existence of a fibrewise symplectic form that is holonomy invariant, and such a bundle will be called \emph{symplectically flat}. In this way, a flat bundle $\Sigma_h \to E \to B$ with a transversal symplectic structure is equivalent to a holonomy representation $\pi_1(B) \stackrel{\rho} \rightarrow Symp(\Sigma_h, \omega)$, where $\omega$ is the symplectic form restricted to a fibre. We shall for the most part suppress any explicit reference to the symplectic form, since by Moser stability any two symplectic forms on $\Sigma_h$ are equivalent after rescaling.
In order to prove the existence of symplectically flat bundles, we wish to mimic the proof of Proposition \ref{horizontal_leaves2}. In order to do this we note that the analogue of Lemma \ref{exact_sequence} holds in the symplectic case (cf.\ \cite{Bow2}, Proposition 4.2.1). So if $\mathcal{G}^{Symp}_p$ the group of symplectomorphism germs that fix the marked point $p$ and $Symp^c(\Sigma_{h,1})$ consists of symplectomorphisms whose supports are disjoint from $p$, we have the following.
\begin{lem}\label{exact_sequence_Symp}
The following sequence of groups is exact
\[ 1 \to Symp^c(\Sigma_{h,1}) \to Symp(\Sigma_{h,1}) \stackrel{\pi} \rightarrow \mathcal{G}^{Symp}_p \to 1.\]
\end{lem}
The other important ingredient is the fact that the compactly supported symplectomorphism group is perfect. The proof of this fact is implicit in the articles \cite{KM},\cite{KM2} and for a detailed proof we refer to (\cite{Bow2}, Lemma 4.2.7).
\begin{lem}\label{Symp_perfect}
The group $Symp^c(\Sigma_h^k)$ is perfect for $h \geq 3$.
\end{lem}
\noindent With these facts the proof of the following is entirely analogous to the smooth case and the details are left to the reader.
\begin{prop}\label{symp_leaves_construction}
For $h \geq 3$, there exist flat bundles $\Sigma_h \to E \to \Sigma_g$ that have symplectic holonomy and whose horizontal foliations have arbitrarily many closed leaves $S_k$ of prescribed self-intersection $[S_k]^2 = m_k \leq h -1$.
In particular, if $m_k = 0$ we may assume that the horizontal foliation in some neighbourhood $\nu_k$ of $S_k$ is given by the kernel of a projection $\nu_k = \Sigma_h \times D^2 \to D^2$.
\end{prop}
In \cite{KM} the authors introduced the notion of a \emph{symplectic pair} (cf.\ also \cite{BK}). In the case of four-manifolds, a symplectic pair consists of a pair of complementary, two-dimensional transversally symplectic foliations. As a consequence of Proposition \ref{symp_leaves_construction} we may now give examples of manifolds with symplectic pairs, both of whose foliations have closed leaves of non-zero self-intersection.
\begin{cor}\label{Pairs_no_triv_leaves}
There exist 4-manifolds that admit symplectic pairs $(\omega_1, \omega_2)$ both of whose foliations $\mathcal{F}_1, \mathcal{F}_2$ have closed leaves $L_1, L_2$ with $[L_i]^2 \neq 0$.
\end{cor}
\begin{proof}
We let $E_1$ be a flat symplectic bundle with a section $s_1$ such that $[s_1]^2 \neq 0$. We further let $E_2$ be a flat symplectic bundle with two sections $s_2,t_2$, the first of which has non-trivial self-intersection and the second of which has a neighbourhood on which the horizontal foliation is given by projection to a disc. The existence of these bundles is guaranteed by Proposition \ref{symp_leaves_construction}. By a suitable choice of $E_1$ we may assume that the genus of its fibre $g(F_1)$ to be arbitrarily large. After stabilisation of $E_2$, we may also assume that the genus of the base of $E_2$ is $g(F_1)$. As noted in \cite{BK}, the assumptions on the foliation in a neighbourhood of $t_2$ imply that the normal connected sum
\[E_1 \#_{F_1 = t_2} E_2\]
admits a symplectic pair, whose foliations we denote by $\widetilde{\mathcal{F}}_i$. Then by construction the connect sums
\[\sigma_1 = s_1 \# F_2\]
\[\sigma_2 = s_2 \# F_1\]
are leaves of $\widetilde{\mathcal{F}}_i$ and $[\sigma_i]^2 = [s_i]^2 \neq 0$.
\end{proof}
For smooth diffeomorphisms we computed the Abelianisation of $Diff^+(\Sigma_{h,k})$. It is then natural to try to determine the Abelianisation of $Symp(\Sigma_{h,k})$, however one cannot mimic the proof of Proposition \ref{not_perfect} used above. The first step is still valid and thus one has that $H_1(Symp_{\delta}(\Sigma_{h,1})) = H_1(\mathcal{G}^{Symp}_p)$. There is also a version of the Sternberg Linearisation Theorem for symplectic germs, but the normal form that it yields is not linear (cf.\ \cite{Ster2}), and thus the computation of the Abelianisation of $Symp(\Sigma_{h,1})$ remains an open question.
\subsection*{The case of genus $0$:}\label{case_low_genus}
So far the results that we have obtained have been for bundles whose fibre $\Sigma_h$ has been of genus at least 3. We shall now consider the case of genus $0$, where one can give a fairly precise description of the possible compact leaves of a (symplectically) flat bundle.
Examples of sphere bundles with horizontal foliations that have closed leaves of arbitrary self-intersection have been given by Mitsumatsu (cf.\ \cite{Mit}). We shall summarise his construction here. Let $\mathbb{R}^2 \to \xi_k \to \Sigma_g$ be a flat bundle of Euler class $k \leq g -1$ as given by \cite{Mil}. Then the sphere bundle $ E_k = S(\xi_k \oplus \mathbb{R})$ is flat and has two sections $L_{\pm}$ corresponding to the north and south poles of the fibre and $[L_{\pm}]^2 = \pm k$.
We would of course like to have similar examples for flat bundles with symplectic holonomy. The flat structures that one obtains via the construction of Mitsumatsu cannot have symplectic holonomy. For if so, then one would have a vertical symplectic form $\omega_v$ that is positive on each fibre, i.e. $\omega_v([F]) \neq 0$, and vanishes identically on the leaves of the horizontal foliation. But the set $\{L_{-}, L_{+}\}$ generates $H_2(E_k, \mathbb{R})$, which is a contradiction. It is not difficult to adapt the argument of Proposition \ref{symp_leaves_construction} to produce horizontal foliations of sphere bundles with symplectic holonomy (cf.\ \cite{Bow2}, p.\ 76). Interestingly, this gives foliations with a unique compact leaf. For if $L'$ were any other leaf then $\{L , L' \}$ would generate $H_2(E_k, \mathbb{R})$ and this would contradict the existence of a vertical symplectic form.
\section{Filling flat $S^1$-bundles}\label{fill_poss}
Given a flat $S^1$-bundle, we would like to know when it bounds a flat surface bundle. To answer this question in full generality is a subtle matter. However for bundles over compact surfaces this can always be achieved after stabilisation.
\begin{prop}\label{boundary_Euler_class}
Let $h \geq 3$ or $h = 0$ and let $M$ be a flat $S^1$-bundle. Then there is a flat bundle $\Sigma_h^1 \to E \to \Sigma_g$, whose boundary is a stabilisation of $M$. In particular, there exist flat $\Sigma_h^1$-bundles whose boundaries have non-trivial Euler class.
\end{prop}
\begin{proof}
Let $a_i, b_i \in \pi_1(\Sigma_g)$ denote the standard generators of the fundamental group and let $ \phi_i = \rho(a_i)$ and $\psi_i = \rho(b_i) $ be the images of these generators in $Diff_0(S^1)$ under the monodromy homomorphism. Since $\phi_i, \psi_i$ are isotopic to the identity, we may extend them to diffeomorphisms $\bar{\phi}_i, \bar{\psi}_i$ on a collar of the boundary $[0, 1] \times S^1$ in such a way that
\[\bar{\phi}_i(t,x) = (t,\phi_i(x)) \text{ , } \bar{\psi}_i(t,x) = (t,\psi_i(x)) \text{ for } 0 \leq t < \epsilon\]
and
\[\bar{\phi}_i(t,x) = \bar{\psi}_i(t,x) = Id\text{ for } 1 - \epsilon < t \leq 1.\]
We then extend by the identity to obtain $\bar{\phi}_i, \bar{\psi}_i \in Diff^+(\Sigma_h^1)$ such that $\eta = \prod_{i = 1}^{g} [\bar{\phi}_i, \bar{\psi}_i]$ lies in $Diff^c(\Sigma_h^1)$.
For $h \geq 3$ the group $Diff^c(\Sigma_h^1)$ is perfect (cf.\ Proposition \ref{horizontal_leaves}) and for $h = 0$ this is the classical result of Thurston (cf.\ \cite{Th2}). Thus we may write $\eta^{-1} = \prod_{i = 1}^{g'} [\alpha_i, \beta_i]$, where $\alpha_i, \beta_i \in Diff^c(\Sigma_h^1)$. We define a flat bundle $E$ over $\Sigma_{g + g'}$ by the holonomy representation
\[a_i \mapsto \bar{\phi}_i, \ \ b_i \mapsto \bar{\psi}_i \text{ for } 1 \leq i \leq g\]
\[a_{g +j} \mapsto \alpha_j, \ \ b_{g +j} \mapsto \beta_j \text{ for } 1 \leq j \leq g'.\]
The boundary of $E$ is a flat $S^1$-bundle and by construction it is a stabilisation of $M$ as required.
For the second statement we note that there exist flat $GL^+(2,\mathbb{R})$-bundles with non-trivial Euler classes, with explicit examples described in \cite{Mil}. The associated projective space bundles are then flat $S^1$-bundles with non-trivial Euler classes and these can then be filled by flat disc bundles after stabilisation, which does not affect the Euler class.
\end{proof}
Proposition \ref{boundary_Euler_class} implies that any flat circle bundle can be filled in by a flat disc bundle after a suitable stabilisation. On the other hand, if we require that the bundle have symplectic holonomy, then this is no longer true (cf.\ Theorem \ref{Tsuboi_int}). However, if the fibre has genus $h \geq 3$, then one can indeed find a filling by a symplectically flat bundle after a suitable stabilisation. To this end we shall need an analogue of the extension trick of Proposition \ref{boundary_Euler_class} in the symplectic case.
\begin{prop}\label{flat_extension}
Let $\pi_1(\Sigma_g) \stackrel{\rho} \rightarrow Diff_0(S^1)$ be a flat structure on an $S^1$-bundle $M$ and let $\phi_i, \psi_i$ denote $\rho(a_i), \rho(b_i)$ respectively. Then there are symplectic extensions $\tilde{\phi}_i, \tilde{\psi}_i$ on the annulus $A = S^1 \times [0, 1]$ that are the identity in a neighbourhood of $S^1 \times \{ 1 \}$ such that $\prod_i^g [\tilde{\phi}_i, \tilde{\psi}_i]$ has support in the interior of $A$.
\end{prop}
\begin{proof}
Let $\mathcal{F}$ be the horizontal foliation given by the flat structure on $M$ and let $\alpha \in \Omega^1(M)$ be a defining 1-form for $\mathcal{F}$. We choose a function $\phi$ on $[0, 1]$, which is equal to $t$ on a neighbourhood of $0$ and is identically zero for all $t$ in a neighbourhood of $1$. Set $\omega = dt \wedge \alpha + \phi(t) d \alpha$ on $E = M \times [0, 1]$ and let $\frac{\partial}{\partial \theta}$ denote a vector field that is tangent to the fibres of $M$. Then
\[\omega(\frac{\partial}{\partial t},\frac{\partial}{\partial \theta}) = \alpha(\frac{\partial}{\partial \theta}) \neq 0,\]
since $\mathcal{F}$ is transverse to the fibres of $M$ and, thus, $\omega$ is a nowhere vanishing 2-form on $E$. Furthermore, since $\mathcal{F}$ is a foliation we compute:
\begin{align*}
\omega^2 &= (dt \wedge \alpha + \phi(t)d \alpha)^2\\
& = 2\phi(t) dt \wedge \alpha \wedge d \alpha = 0.
\end{align*}
Thus $\mathcal{F}_{\omega} = Ker(\omega)$ is a well-defined distribution that is transverse to the (annular) fibres of $E \to \Sigma_g$. Moreover, since $\omega = d(t \alpha)$ in a neighbourhood of $M \times \{ 0 \}$ this distribution is integrable and transversally symplectic on this neighbourhood, and restricts to $\mathcal{F}$ on $M \times \{ 0 \}$. On a neighbourhood of $M \times \{ 1 \}$ the form $\omega$ reduces to $dt \wedge \alpha$ and again the kernel distribution is integrable and agrees with $\mathcal{F}$ on this neighbourhood.
We choose a base point $x_0 \in \Sigma_g$ and embedded representatives $a_i,b_i$ for the standard generators of $\pi_1(\Sigma_g, x_0)$ and let $\bar{\phi}_i, \bar{\psi}_i$ be the holonomies of the curves $a_i,b_i$ given by the distribution $\mathcal{F}_{\omega}$. Then on $S^1 \times \{ 0 \}$ and near $S^1 \times \{ 1 \}$ these diffeomorphisms are given by $\phi_i \times Id$ and $\psi_i \times Id$ respectively, where $\phi_i, \psi_i$ are the images of the standard basis under the holonomy representation of $M$. Since $\phi_i, \psi_i$ lie in $Diff_0(S^1)$, we may alter the maps $\bar{\phi}_i, \bar{\psi}_i$ near $S^1 \times \{ 1 \}$ so that they restrict to the identity in a neighbourhood of $S^1 \times \{ 1 \}$. We shall continue to denote these altered maps by $\bar{\phi}_i, \bar{\psi}_i$.
We let $\Omega$ be the restriction of $\omega$ to the annular fibre over $x_0$. Then since the holonomies $\bar{\phi}_i, \bar{\psi}_i$ have support in $S^1 \times [0,1)$ and the distribution defining them was transversally symplectic in a neighbourhood of $M \times \{ 0 \}$, the forms $\bar{\phi}^*_i \Omega - \Omega$ and $\bar{\psi}^*_i \Omega - \Omega$ have compact support in the interior of $S^1 \times (0,1)$. Moreover
\[\int_A (\bar{\phi}^*_i \Omega - \Omega) = \int_A (\bar{\psi}^*_i \Omega - \Omega) = 0\]
so that the forms $\bar{\phi}^*_i \Omega$ and $\bar{\psi}^*_i \Omega$ are cohomologous to $\Omega$ in compactly supported cohomology. By applying a Moser isotopy, which will have support in the interior of $S^1 \times [0,1]$, we obtain symplectomorphisms $\tilde{\phi_i}, \tilde{\psi}_i$ that are symplectic extensions of $\phi_i, \psi_i$ respectively, and by construction $\prod_i^g [\tilde{\phi}_i, \tilde{\psi}_i]$ has support in the interior of $A$.
\end{proof}
Proposition \ref{flat_extension} is the main step in extending flat structures symplectically and the following result follows from this and the perfectness of $Symp^c(\Sigma_h^1)$.
\begin{thm}\label{symp_flat_extension}
Let $M$ be a flat $S^1$-bundle and assume that $h \geq 3$. Then some stabilisation of $M$ bounds a flat $\Sigma_h^1$-bundle with symplectic holonomy.
\end{thm}
\begin{proof}
Let $\pi_1(\Sigma_g) \stackrel{\rho} \rightarrow Diff_0(S^1)$ be the holonomy representation associated to $M$ and let $\tilde{\phi}_i, \tilde{\psi}_i \in Symp(A)$ be the extensions given by Proposition \ref{flat_extension}. After a suitable choice of symplectic form on $\Sigma_h^1$, we may symplectically embed $A = S^1 \times [0,1]$ in $\Sigma_h^1$ so that $S^1 \times \{ 0 \}$ maps to $\partial \Sigma_h^1$ . We then consider $\eta = \prod_i^g [\tilde{\phi}_i, \tilde{\psi}_i]$ as an element in $Symp^c(\Sigma_h^1)$. By Lemma \ref{Symp_perfect} this group is perfect and, thus, we may write $\eta^{-1}$ as a product of $g'$ commutators. We then define the associated flat bundle $E'$ over $\Sigma_{g + g'}$ as in the proof of Proposition \ref{boundary_Euler_class}, and by construction the boundary of $E'$ is a stabilisation of $M$.
\end{proof}
Theorem \ref{symp_flat_extension} can be interpreted in terms of the five-term exact sequence of a certain extension of groups. For this we let $Symp(\Sigma_h^1)$ as usual denote the group of symplectomorphisms of $\Sigma_h^1$. We further let $Symp(\Sigma_h^1, \partial \Sigma_h^1)$ denote those symplectomorphisms that restrict trivially to the boundary. Then as a consequence of Proposition \ref{flat_extension} the following sequence, which is given by restriction to $\partial \Sigma_h^1$, is exact:
\[1 \to Symp(\Sigma_h^1, \partial \Sigma_h^1) \to Symp(\Sigma_h^1) \to Diff^+(\partial \Sigma_h^1) = Diff_0(S^1) \to 1.\]
With this notation we have the following proposition.
\begin{prop}\label{conn_surj}
For $h \geq 3$ the connecting homomorphism in the five-term exact sequence in real cohomology associated to the following exact sequence is trivial:
\[1 \to Symp(\Sigma_h^1, \partial \Sigma_h^1) \to Symp(\Sigma_h^1) \to Diff^+(\partial \Sigma_h^1) = Diff_0(S^1) \to 1.\]
\end{prop}
\begin{proof}
By the Universal Coefficient Theorem it suffices to show that the map
\[H_2(Symp_{\delta}(\Sigma_h^1)) \to H_2( Diff_{0,{\delta}}(S^1))\]
is surjective on integral cohomology. This follows immediately from Theorem \ref{symp_flat_extension}, since any flat $S^1$-bundle extends after stabilisation and this does not change the homology class represented by this bundle in $H_2( Diff_{0,{\delta}}(S^1))$.
\end{proof}
The Godbillon-Vey class of the horizontal foliation of a flat $S^1$-bundle $M$ defines an element $GV$ in $H^2( Diff_{0,{\delta}}(S^1), \mathbb{R})$, which is non-trivial by the work of Thurston (cf.\ \cite{Bott}). It is possible that the Godbillon-Vey class provides an obstruction to the existence of a flat symplectic bundle $E$ that bounds $M$. However, by Proposition \ref{conn_surj} the image of the class $GV$ in $H^2(Symp_{\delta}(\Sigma_h^1), \mathbb{R})$ is non-trivial. Geometrically, this means that after stabilisation the horizontal foliation of any $S^1$-bundle extends to a transversally symplectic foliation on some surface bundle $E$ with fibre $\Sigma_h^1$. In particular, the Godbillon-Vey class is not an obstruction to finding a null-cobordism that extends the horizontal foliation of $M$ to the interior of $E$ symplectically.
\section{Flat bundles and the extended Hamiltonian group}\label{Flat_Ham}
In Section \ref{fill_poss} we showed that any flat circle bundle over a surface can be filled by a flat disc bundle with smooth holonomy after stabilisation. However, as was shown in \cite{Tsu}, it is not in general possible to fill in a flat circle bundle by a flat disc bundle that has symplectic holonomy, even after stabilisation, since the existence of such a filling implies that the Euler class of the circle bundle vanishes. More specifically Tsuboi proved the following theorem.
\begin{thm}[\cite{Tsu}]\label{Tsuboi}
Let $\pi_1(\Sigma_g) \stackrel{\psi} \rightarrow Diff_0(S^1)$ be a homomorphism and let $a_i, b_i$ be standard generators of $\pi_1(\Sigma_g)$. Furthermore, let $f_i, h_i \in Symp(D^2)$ be extensions of $\psi(a_i), \psi(b_i)$ respectively and let $e(E)$ denote the Euler class of the total space of the $S^1$-bundle $E$ associated to $\psi$. Then
\[-\pi^2 e(E) = Cal([f_1, h_1]...[f_g, h_g]).\]
\end{thm}
In the case of a disc, a diffeomorphism is symplectic if and only if it is Hamiltonian. For bundles with fibres of higher genus we shall generalise Tsuboi's result under the assumption that the holonomies are Hamiltonian. As usual a symplectomorphism $\psi \in Symp_0(\Sigma_h^1)$ will be called Hamiltonian if it is isotopic to the identity via an isotopy $\psi_t$ such that $\iota_{\dot{\psi}_t} \omega = dH_t$ for $0 \leq t \leq 1$. As a first step, following \cite{Tsu} we note that it is always possible to extend diffeomorphisms on the boundary of $\Sigma_h^1$ to its interior by a Hamiltonian diffeomorphism. It suffices to consider the case $M = [0, 1) \times S^1$ and to show that any diffeomorphism of the boundary extends to a Hamiltonian diffeomorphism on $M$ that has compact support. The following is essentially Lemma 2.2 of \cite{Tsu} and the proof will be omitted.
\begin{lem}\label{extend_volume}
Let $M = [0, 1) \times S^1$ and let $\omega$ be a symplectic form of finite total volume. Then the map $Ham^c(M) \to Diff_0(S^1)$ given by restriction is surjective.
\end{lem}
As a consequence of Lemma \ref{extend_volume} we obtain an exact sequence:
\begin{equation*
1 \to Ham(\Sigma_h^1, \partial \Sigma_h^1) \to Ham(\Sigma_h^1) \to Diff^+(\partial \Sigma_h^1) = Diff_0(S^1) \to 1,
\end{equation*}
where $Ham(\Sigma_h^1, \partial \Sigma_h^1)$ denotes the intersection $Symp(\Sigma_h^1, \partial \Sigma_h^1) \cap Ham(\Sigma_h^1)$.
Exactly as in the case of compactly supported symplectomorphisms, we define a Flux homomorphism $Symp_0(\Sigma_h^1) \to H^1(\Sigma_h^1, \mathbb{R})$ via the formula
\[ Flux(\psi) = \int_0^1(\iota_{\dot{\psi}_t} \omega) \thinspace dt, \]
where $\psi_t$ is an isotopy from $Id$ to $\psi$. As in the compactly supported case, one can show that $Flux(\psi) = [\lambda - \psi^* \lambda]$ for any primitive $ \lambda$ such that $\omega = -d \lambda$ (cf.\ \cite{McS}, Lemma 10.14). Hence, $Flux$ is well-defined independently of the choice of isotopy $\psi_t$ and primitive $\lambda$. Moreover, it is easy to show that $Ker(Flux) = Ham(\Sigma_h^1)$ (cf.\ \cite{Bow2}, Lemma 5.2.4).
One may also define a Calabi homomorphism $Cal$ on $Ham(\Sigma_h^1, \partial \Sigma_h^1)$. For this let $\lambda$ be a primitive such that $\omega = - d \lambda$ and define
\[Cal(\phi) = -\frac{1}{3} \int_{\Sigma_h^1} \phi^* \lambda \wedge \lambda .\]
Again this definition is independent of the choice of $\lambda$ (see \cite{McS}, Lemma 10.26).
We will now extend Tsuboi's result to bundles with fibre $\Sigma_h^1$. In order to do this we shall need to reinterpret Theorem \ref{Tsuboi} in terms of a five-term exact sequence. Now the map $Cal$ is an element of $H^1(Ham_{\delta}(D^2, \partial D^2),\mathbb{R})$ and we claim that it is invariant under the conjugation action of $Ham(D^2)$. For let $\psi \in Ham(D^2)$ and $\phi \in Ham(D^2, \partial D^2)$, and let $\lambda$ be a primitive such that $ \omega = - d \lambda$. Then $\psi^* \lambda$ is also a primitive for $\omega$ and we have
\begin{equation*
\int_{D^2} \phi^* \lambda \wedge \lambda = \int_{D^2} \phi^* (\psi^* \lambda) \wedge (\psi^*\lambda) = \int_{D^2} (\psi \phi \psi^{-1})^* \lambda \wedge \lambda. \end{equation*}
Thus, $Cal \in H^1(Ham_{\delta}(D^2, \partial D^2),\mathbb{R})^{Diff_0(S^1)}$ and we claim that Tsuboi's result can be interpreted as saying that the image of $Cal$ under the connecting homomorphism in the five-term exact sequence is a non-zero multiple of the Euler class $e$ considered as an element in the real group cohomology of $Diff_0(S^1)$. For this we need to make use of an explicit description of the connecting homomorphism, which is described in Appendix \ref{App_five_term} below.
\begin{thm}\label{Tsuboi_five}
Consider the extension of groups
\[1 \to Ham(D^2, \partial D^2) \to Ham(D^2) \to Diff_0(S^1) \to 1,\]
and let $\delta$ denote the connecting homomorphism in the five-term exact sequence in real cohomology. Then $\delta [Cal]$ is $-\pi^2 \thinspace e$, where $e$ denotes the Euler class in $H^2(Diff_{0,\delta}(S^1), \mathbb{R})$.
\end{thm}
\begin{proof}
In order to verify the equality $\delta Cal = -\pi^2 \thinspace e$ in real cohomology, it suffices to evaluate both sides on 2-cycles $Z$ in $H_2(Diff_{0,\delta}(S^1))$. Such a cycle may be thought of as the image of the fundamental class under the map induced by a representation of a surface group $\pi_1(\Sigma_g) \stackrel{\psi} \rightarrow Diff_0(S^1)$. If we let $a_i, b_i$ be standard generators of $\pi_1(\Sigma_g)$, then a generator of $H_2(\pi_1(\Sigma_g))$ may be described by the group 2-cycle
\begin{align*}
z & = (a_1,b_1) + (a_1b_1,a_1^{-1}) + ... + (a_1b_1...b_{g-1}a_{g}^{-1}, b_g^{-1})\\
& - (2g+1)(e,e) - \sum_{i = 1}^g (a_i,a_i^{-1}) + (b_i,b_i^{-1}) .
\end{align*}
Since $[a_1, b_1]...[a_g, b_g] = e$ in $\pi_1(\Sigma_g)$, we compute that
\begin{align*} \partial z = \sum_{i = 1}^g (a_i) + (a_i^{-1}) + (b_i) +(b_i^{-1}) &- \sum_{i = 1}^g [(a_i) - (e) +(a_i^{-1}) + (b_i) - (e) +(b_i^{-1})] \\
& - 2g(e) + ([a_1, b_1]...[a_g, b_g]) -(e) = 0.
\end{align*}
We let $f_i, h_i$ denote representatives of $\psi(a_i), \psi(b_i)$ in $Diff_0(S^1)$ considered as a quotient group, and let $\tilde{z}$ be the associated lift of the fundamental cycle above. Then we compute
\[\partial \tilde{z} = ([f_1, h_1]...[f_g, h_g]) - (e).\]
Thus, by Lemma \ref{connecting_hom}, there is a set-theoretic extension $Cal_S$ of $Cal$ to $Ham(D^2)$ such that
\begin{align*}
\delta Cal (Z) & = \delta Cal_S(\psi_* [\Sigma_g]) = Cal_S(\partial \tilde{z}) \\
& = Cal_S(([f_1, h_1]...[f_g, h_g]) - (e))= Cal([f_1, h_1]...[f_g, h_g]),
\end{align*}
and by Proposition \ref{extended_Tsuboi} this is equal to $-\pi^2 \thinspace e(E)$.
\end{proof}
\noindent We are now ready to generalise Theorem \ref{Tsuboi} to the case of surfaces of higher genus.
\begin{prop}\label{extended_Tsuboi}
Let $\pi_1(\Sigma_g) \stackrel{\psi} \rightarrow Diff_0(S^1)$ be a homomorphism and let $a_i,b_i$ be standard generators of $\pi_1(\Sigma_g)$. Let $f_i, h_i \in Ham(\Sigma_h^1)$ be any extensions of $\psi(a_i), \psi(b_i)$ respectively and let $e(E)$ denote the Euler class of the total space of the $S^1$-bundle $E$ associated to $\psi$. Then
\[-\pi^2 e(E) = Cal([f_1, h_1]...[f_g, h_g]).\]
\end{prop}
\begin{proof}
By Lemma \ref{extend_volume} we may assume that the extensions $f_i, h_i$ are Hamiltonian and have support in a collar $K = [0,1) \times S^1$ of the boundary. We may then consider $K \subset D^2$ with an appropriately chosen area form $\Omega$ on $D^2$ and $f_i, h_i$ as elements in $Diff_{\Omega}(D^2)$. We then compute
\begin{equation*
\psi^* \delta Cal([\Sigma_g]) = Cal^{\Sigma^1_h}([f_1, h_1]...[f_{g}, h_{g} ]) = Cal^{D^2}([f_1, h_1]...[f_{g}, h_{g} ]),
\end{equation*}
where the first equality follows as in Theorem \ref{Tsuboi_five} and the second follows from our choice of extensions. The latter value is $-\pi^2 \thinspace e(E)$ by Theorem \ref{Tsuboi_five}. Thus, since the left hand side of the equation above is independent of any choices, we conclude that for \emph{any} extensions $f_i,h_i$
\[Cal^{\Sigma^1_h}([f_1, h_1]...[f_{g}, h_{g} ]) = -\pi^2 \thinspace e(E). \qedhere\]
\end{proof}
In particular, it follows by the exactness of the five-term sequence that the boundary of any flat bundle with holonomy in $Ham(\Sigma_h^1)$ is trivial as an $S^1$-bundle. Furthermore, with our interpretation of Tsuboi's result we may extend our discussion to extended Hamiltonian groups as introduced in \cite{KM}. To this end we define $\widetilde{Flux}_{\lambda}: Symp(\Sigma_h^1) \to H^1(\Sigma_h^1, \mathbb{R})$ by
\[\widetilde{Flux}_{\lambda}(\phi) = [(\phi^{-1})^*\lambda - \lambda] \]
for some fixed primitive $ - d \lambda = \omega$. This map is a crossed homomorphism and is referred to as an \emph{extended Flux homomorphism}, since is restricts to the ordinary Flux map on $Symp_0(\Sigma_h^1)$. We note that the definition of $\widetilde{Flux}_{\lambda}$ depends in an essential way on the choice of primitive $\lambda$. For if $\lambda '$ is another primitive, then $\lambda - \lambda' = \alpha$ is closed and
\begin{equation}\label{Extended_Flux_1}
\widetilde{Flux}_{\lambda}(\phi) = \widetilde{Flux}_{\lambda'}(\phi) + [(\phi^{-1})^* \alpha - \alpha].
\end{equation}
In terms of group cohomology this means that $\widetilde{Flux}_{\lambda}$ and $\widetilde{Flux}_{\lambda'}$ are cohomologous, when considered as elements in $H^1(Symp(\Sigma_h^1),H^1(\Sigma_h^1, \mathbb{R}))$. The \emph{extended Hamiltonian group} $\widetilde{Ham}(\Sigma_h^1)$ is defined as the kernel of $\widetilde{Flux}_{\lambda}$, which is a subgroup since $\widetilde{Flux}_{\lambda}$ is a crossed homomorphism. The group $Ham(\Sigma_h^1)$ is contained in $ \widetilde{Ham}(\Sigma_h^1)$ and we may extend the Calabi homomorphism to a map $\widetilde{Cal}_{\lambda}$ on the group
$\widetilde{Ham}(\Sigma_h^1, \partial \Sigma_h^1) = Symp(\Sigma_h^1, \partial \Sigma_h^1) \cap \widetilde{Ham}(\Sigma_h^1)$ by defining
\[\widetilde{Cal}_{\lambda}(\phi) = -\frac{1}{3} \int_{\Sigma_h^1} \phi^* \lambda \wedge \lambda = \frac{1}{3} \int_{\Sigma_h^1} (\phi^{-1})^* \lambda \wedge \lambda,\]
where $\lambda$ is the primitive chosen in the definition of $\widetilde{Flux}_{\lambda}$. This is a homomorphism on $\widetilde{Ham}(\Sigma_h^1, \partial \Sigma_h^1)$, since the following holds on $Symp(\Sigma_h^1, \partial \Sigma_h^1)$ (cf.\ \cite{KM2}, Prop.\ 19):
\begin{equation}\label{Calabi_prop_1_extended}
\widetilde{Cal}_{\lambda}(\phi \psi) = \widetilde{Cal}_{\lambda}(\phi) + \widetilde{Cal}_{\lambda}(\psi) +\frac{1}{3} \widetilde{Flux}_{\lambda}(\phi) \wedge (\phi^{-1})^*\widetilde{Flux}_{\lambda}(\psi).
\end{equation}
Again the definition of $\widetilde{Cal}_{\lambda}$ depends on the choice of primitive $\lambda$. However, we do have the following technical lemma, which will be important in showing the equivariance of $\widetilde{Cal}_{\lambda}$.
\begin{lem}\label{Cal_invariance}
Let $\phi \in \widetilde{Ham}_{\lambda}(\Sigma_h^1, \partial \Sigma_h^1) \cap \widetilde{Ham}_{\lambda'}(\Sigma_h^1, \partial \Sigma_h^1)$ for two different primitives $\lambda, \lambda'$ and set $\alpha = \lambda - \lambda'$, further assume that $\phi^* \alpha - \alpha = d H_{\phi}$ is exact. Then $\widetilde{Cal}_{\lambda}(\phi) = \widetilde{Cal}_{\lambda'}(\phi)$.
\end{lem}
\begin{proof}
By assumption $ \lambda - (\phi^{-1})^*\lambda$ is exact and hence $\phi^* \lambda - \lambda = dF_{\phi}$ is also exact. Since the boundary of $\Sigma_h^1$ is connected we may further assume that $F_{\phi} = 0$ on $\partial \Sigma_h^1$. We compute
\begin{align*}
\widetilde{Cal}_{\lambda}(\phi) & = -\frac{1}{3} \int_{\Sigma_h^1} \phi^* \lambda \wedge \lambda = -\frac{1}{3} \int_{\Sigma_h^1} (\phi^* \lambda - \lambda) \wedge \lambda\\
& = -\frac{1}{3} \int_{\Sigma_h^1} dF_{\phi} \wedge \lambda = -\frac{1}{3} \int_{\Sigma_h^1} d(F_{\phi}\lambda) - F_{\phi}\wedge d \lambda = -\frac{1}{3} \int_{\Sigma_h^1} F_{\phi} \omega \text.\\
\end{align*}
Similarly one computes
\[\widetilde{Cal}_{\lambda'}(\phi) = \widetilde{Cal}_{\lambda}(\phi) -\frac{1}{3} \int_{\Sigma_h^1} \phi^* \alpha \wedge \alpha = -\frac{1}{3} \int_{\Sigma_h^1} F_{\phi}\omega - \frac{1}{3} \int_{\Sigma_h^1} H_{\phi} \omega,\]
where $\phi^* \alpha - \alpha = dH_{\phi}$ and again we assume that $H_{\phi} = 0$ on $\partial \Sigma_h^1$. We finally have that
\begin{align*}
- \int_{\Sigma_h^1} H_{\phi} \omega & = \int_{\Sigma_h^1} H_{\phi} d \lambda = \int_{\Sigma_h^1} d(H_{\phi}\lambda) - d H_{\phi}\wedge \lambda\\
& = \int_{\Sigma_h^1} ( \alpha - \phi^* \alpha) \wedge \lambda = \int_{\Sigma_h^1} \alpha \wedge \lambda - \phi^* \alpha \wedge \lambda\\
& = \int_{\Sigma_h^1} \phi^* \alpha \wedge \phi^* \lambda - \phi^* \alpha \wedge \lambda = \int_{\Sigma_h^1} \phi^* \alpha \wedge (\phi^* \lambda - \lambda)\\
& = 0
\end{align*}
since $\lambda - \phi^* \lambda = dF_{\phi}$, $d \alpha = 0$ and $F_{\phi}|_{\partial \Sigma_h^1} = 0$.
\end{proof}
\noindent Lemma \ref{Cal_invariance} then implies that $\widetilde{Cal}_{\lambda}$ is equivariant under the conjugation action of $\widetilde{Ham}(\Sigma_h^1)$.
\begin{cor}\label{Cal_equi}
Let $\psi \in \widetilde{Ham}_{\lambda}(\Sigma_h^1)$, then $\widetilde{Cal}_{\lambda} = \widetilde{Cal}_{\psi^* \lambda}$. In particular, $\widetilde{Cal}_{\lambda}$ is equivariant under the action of $\widetilde{Ham}_{\lambda}(\Sigma_h^1)$.
\end{cor}
\begin{proof}
Let $\lambda' = \psi^* \lambda$, then $\lambda'$ is also a primitive with $- d\lambda' = \omega$ and $\alpha = \lambda' - \lambda = dH_{\psi}$ is exact since $\psi \in \widetilde{Ham}_{\lambda}(\Sigma_h^1)$. By formula (\ref{Extended_Flux_1}), it follows that $\widetilde{Flux}_{\lambda} = \widetilde{Flux}_{\lambda'}$ and hence
\[\widetilde{Ham}_{\lambda}(\Sigma_h^1, \partial \Sigma_h^1) = \widetilde{Ham}_{\lambda'}(\Sigma_h^1, \partial \Sigma_h^1).\]
By applying Lemma \ref{Cal_invariance} we conclude that $\widetilde{Cal}_{\lambda} = \widetilde{Cal}_{\psi^* \lambda}$. This then implies
\begin{align*}
\widetilde{Cal}_{\lambda}(\phi) & = -\frac{1}{3} \int_{\Sigma_h^1} \phi^* \lambda \wedge \lambda = -\frac{1}{3} \int_{\Sigma_h^1} \phi^* (\psi^*\lambda) \wedge \psi^* (\lambda)\\
& = -\frac{1}{3} \int_{\Sigma_h^1} (\psi \phi \psi^{-1})^*\lambda \wedge \lambda = \widetilde{Cal}_{\lambda}(\psi \phi \psi^{-1}),
\end{align*}
which is the desired equivariance.
\end{proof}
We may now prove an analogue of Tsuboi's result for the extended Hamiltonian group. For the sake of notational expediency, we shall drop any explicit references to $\lambda$.
\begin{thm}\label{Hamiltonian_extended_Tsuboi}
Let $\pi_1(\Sigma_g) \stackrel{\psi} \rightarrow Diff_0(S^1)$ be a homomorphism and let $a_i,b_i$ be standard generators of $\pi_1(\Sigma_g)$. Let $f_i, h_i \in \widetilde{Ham}(\Sigma_h^1)$ be any extensions of $\psi(a_i), \psi(b_i)$ respectively and let $e(E)$ denote the Euler class of the total space of the $S^1$-bundle $E$ associated to $\psi$. Then
\[-\pi^2 e(E) = Cal([f_1, h_1]...[f_g, h_g]).\]
In particular, if $\Sigma_h^1 \to E \to \Sigma_g$ is a flat bundle with holonomy in the extended Hamiltonian group, then the boundary is a trivial bundle.
\end{thm}
\begin{proof}
We consider the commuting diagram
\[\xymatrix{1 \ar[r] & Ham(\Sigma_h^1, \partial \Sigma_h^1) \ar[r] \ar[d]& Ham(\Sigma_h^1) \ar[r] \ar[d] & Diff_0(S^1) \ar[r] \ar[d] & 1\\
1 \ar[r]& \widetilde{Ham}(\Sigma_h^1, \partial \Sigma_h^1) \ar[r] & \widetilde{Ham}(\Sigma_h^1) \ar[r] & Diff_0(S^1) \ar[r] & 1.}\]
The five-term sequence then gives the following commuting triangle
\[\xymatrix{ H^1(Ham_{\delta}(\Sigma_h^1, \partial \Sigma_h^1), \mathbb{R})^{Diff_0(S^1)} \ar[r]^<<<<{\delta} & H^2(Diff_{0,\delta}(S^1), \mathbb{R})\\
H^1(\widetilde{Ham}_{\delta}(\Sigma_h^1, \partial \Sigma_h^1), \mathbb{R})^{Diff_0(S^1)} \ar[u] \ar[ur]^{\delta} &}.\]
By Corollary \ref{Cal_equi} we see that $\widetilde{Cal} \in H^1(\widetilde{Ham}_{\delta}(\Sigma_h^1, \partial \Sigma_h^1), \mathbb{R})^{Diff_0(S^1)}$. If $\iota$ denotes the inclusion of $ Ham(\Sigma_h^1, \partial \Sigma_h^1)$ into $\widetilde{Ham}(\Sigma_h^1, \partial \Sigma_h^1)$, then $\iota^* \widetilde{Cal} = Cal$ and, hence, $\delta(\widetilde{Cal}) = \delta(Cal)$ is also a multiple of the Euler class by Proposition \ref{extended_Tsuboi}.
The second statement follows from the exactness of the five-term sequence.
\end{proof}
A comparison of Theorem \ref{symp_flat_extension} and Theorem \ref{Hamiltonian_extended_Tsuboi} exhibits a stark difference between the two groups $Symp(\Sigma_h^1)$ and $\widetilde{Ham}(\Sigma_h^1)$.
\section{The Calabi map and the first MMM-class}\label{MMM_first}
We have seen that the Euler class of the boundary of a surface bundle with one boundary component can be interpreted as the image of the Calabi map under the connecting homomorphism of a certain five-term exact sequence. We shall give a similar construction for the first Mumford-Miller-Morita (MMM) class $e_1$, which represents a generator of $H^2(\Gamma^1_h, \mathbb{R}) \cong \mathbb{R}$ for $h \geq 3$. We recall that the first MMM-class of a surface bundle over a surface is up to a constant just the signature of the total space.
In order to describe the first MMM-class in terms of a five-term exact sequence we shall need to consider compactly supported extended Hamiltonian groups. We first define an extended Flux homomorphism $\widetilde{Flux}_c$ on $Symp^c(\Sigma_h^1) \subset Symp(\Sigma_h^1)$ by restricting the map $\widetilde{Flux}_{\lambda}$ defined in Section \ref{Flat_Ham}. There are other possible extensions of $Flux$ to crossed homomorphisms, but on the the level of group cohomology these can be easily described and the following is a slight variant of (\cite{KM2}, Theorem 11).
\begin{lem}\label{Flux_class}
Let $\widetilde{Flux}$ be an extended Flux homomorphism on $Symp^c(\Sigma_h^1)$, then the following holds in $H^1(Symp^c(\Sigma_h^1), H_c^1(\Sigma_h^1, \mathbb{R}))$
\[[\widetilde{Flux}] = [\widetilde{Flux}_c] + a[p^*k_{\mathbb{R}}],\]
where $a \in \mathbb{R}$ and $k_{\mathbb{R}} \in H^1(\Gamma_h^1, H_c^1(\Sigma_h^1, \mathbb{R})) \cong \mathbb{R}$ is the generator defined by the extended Johnson homomorphism of Morita.
\end{lem}
\begin{proof}
We let $\Delta = \widetilde{Flux} - \widetilde{Flux}_c$. For $\phi \in Symp_0^c(\Sigma_h^1)$ and $\psi \in Symp^c(\Sigma_h^1)$ we see that on the level of cochains
\begin{align*}
\Delta(\phi.\psi) & = [\widetilde{Flux} - \widetilde{Flux}_c](\phi.\psi)\\
& = [\widetilde{Flux}(\phi) - \widetilde{Flux}_c(\phi)] + (\phi^{-1})^*[\widetilde{Flux}(\psi) - \widetilde{Flux}_c(\psi)]\\
& = [Flux(\phi) - Flux(\phi)] + [\widetilde{Flux}(\psi) - \widetilde{Flux}_c(\psi)]\\
& = [\widetilde{Flux}(\psi) - \widetilde{Flux}_c(\psi)] = \Delta(\psi).
\end{align*}
Moreover, $\Delta$ vanishes on $Symp_0^c(\Sigma_h^1)$ by definition and is coclosed, hence $[\Delta] = p^* [\beta]$ for some $[\beta] \in H^1(\Gamma^1_h, H_1(\Sigma_h^1, \mathbb{R}))$. Now this group is isomorphic to $\mathbb{R}$ and is generated by the extended Johnson homomorphism of Morita (cf.\ \cite{KM2} and \cite{Mor2}).
\end{proof}
\noindent In view of Lemma \ref{Flux_class} we set
\[\widetilde{Flux_a} = \widetilde{Flux}_c + a \thinspace p^* k_{\mathbb{R}}, \text{ for any } a \in \mathbb{R}.\]
The kernel of $\widetilde{Flux_a}$ will then be denoted by $\widetilde{Ham_a^c}(\Sigma_h^1)$ and by considering the projection to the mapping class group we obtain the following extension of groups
\begin{equation*}
1 \to Ham^c(\Sigma_h^1) \to \widetilde{Ham^c}(\Sigma_h^1) \stackrel{p} \rightarrow \Gamma_h^1 \to 1.
\end{equation*}
For any group $G$ there is a pairing $H^1(G, H_c^1(\Sigma_h^1, \mathbb{R})) \times H^1(G, H_c^1(\Sigma_h^1, \mathbb{R})) \to H^2(G, \mathbb{R})$, which we denote $[\alpha.\beta]$ for classes $[\alpha], [\beta] \in H^1(G, H_c^1(\Sigma_h^1, \mathbb{R}))$. This is defined via the following formula
\[\alpha.\beta(\phi,\psi) = \alpha(\phi) \wedge (\phi^{-1})^*\beta(\psi).\]
The induced map on cohomology is well-defined independently of the chosen representatives, and is natural with respect to pullbacks (cf.\ \cite{KM2}, Lemma 18). With these preliminaries we may now prove the following theorem.
\begin{thm}\label{MMM_Cal}
Let $h \geq 2$ and let
\[1 \to Ham^c(\Sigma_h^1) \to \widetilde{Ham_a^c}(\Sigma_h^1) \stackrel{p} \rightarrow \Gamma_h^1 \to 1\]
be the extension associated to the extended Hamiltonian group defined by the extended flux map $\widetilde{Flux_a}$.
Then the image of $[Cal]$ under the connecting homomorphism of the associated five-term exact sequence is $ \frac{1}{3}a^2 e_1$. In particular, if $a$ is non-zero, then any flat bundle with holonomy in $\widetilde{Ham_a^c}(\Sigma_h^1)$ has signature zero.
\end{thm}
\begin{proof}
We consider the following part of the five-term exact sequence associated to the extended Hamiltonian group
\[H^1(\widetilde{Ham_{a,\delta}^c}(\Sigma_h^1),\mathbb{R}) \to H^1(Ham_{\delta}^c(\Sigma_h^1),\mathbb{R})^{\Gamma_h^1} \stackrel{\delta} \rightarrow H^2(\Gamma_h^1, \mathbb{R}).\]
We first claim that the Calabi map lies in the invariant part of $H^1(Ham^c(\Sigma_h^1),\mathbb{R})$. For if $\lambda$ is a primitive for the symplectic form $\omega$ on $\Sigma_h^1$, then so is $\psi^* \lambda$ for any $\psi \in \widetilde{Ham_a^c}(\Sigma_h^1)$. Thus, since the Calabi map is independent of the choice of primitive for any $\phi \in Ham^c(\Sigma_h^1)$, we compute that
\begin{align*}
Cal(\phi) & = -\frac{1}{3} \int_{\Sigma_h^1} \phi^* \lambda \wedge \lambda = -\frac{1}{3} \int_{\Sigma_h^1} \phi^*( \psi^* \lambda) \wedge (\psi^*\lambda)\\
& = -\frac{1}{3} \int_{\Sigma_h^1}(\psi \phi \psi^{-1})^* \lambda \wedge \lambda = Cal(\psi\phi\psi^{-1}),
\end{align*}
and $[Cal]$ lies in $H^1(Ham^c(\Sigma_h^1),\mathbb{R})^{\Gamma_h^1}$ as claimed.
If $i$ denotes the inclusion $\widetilde{Ham_a^c}(\Sigma_h^1) \hookrightarrow Symp^c(\Sigma_h^1)$, then by definition $\widetilde{Flux_a}$ vanishes on $\widetilde{Ham^c_a}(\Sigma_h^1)$ and we see that $i^*[\widetilde{Flux}_c] = - a . i^*(p^* [k_{\mathbb{R}}])$. Let $\widetilde{f} = \widetilde{Cal}_{\lambda}$ and note that by formula (\ref{Calabi_prop_1_extended}) this map satisfies the hypotheses of Lemma \ref{connecting_hom2}. Thus we have an explicit description of the connecting homomorphism in terms of $\widetilde{Cal}_{\lambda}$. More precisely, let $\overline{\phi},\overline{\psi} \in \Gamma_h^1$ be considered as elements of the quotient, then we compute
\begin{align*}
\delta(Cal)(\overline{\phi},\overline{\psi}) & = \widetilde{Cal}_{\lambda}(\phi) + \widetilde{Cal}_{\lambda}(\psi) - \widetilde{Cal}_{\lambda}(\phi.\psi)\\
& = - \frac{1}{3} \widetilde{Flux}_c(\phi) \wedge (\phi^{-1})^*\widetilde{Flux}_c(\psi)\\
& = - \frac{1}{3} a . i^*(p^* k_{\mathbb{R}})(\phi) \wedge (\phi^{-1})^*a . i^*(p^* k_{\mathbb{R}})(\psi)\\
& = - \frac{1}{3} a^2 k_{\mathbb{R}}(\overline{\phi}) \wedge (\phi^{-1})^*k_{\mathbb{R}}(\overline{\psi}).
\end{align*}
Hence we have shown that $[\delta(Cal)] = -\frac{1}{3} a^2 [k_{\mathbb{R}}.k_{\mathbb{R}}]$. Now we know by (\cite{Mor1}, Proposition 4.2) that $[k_{\mathbb{R}}.k_{\mathbb{R}}] = - e_1$, where $e_1$ is the first MMM-class, and we conclude that $\delta[Cal] = \frac{1}{3}a^2 e_1$. The second claim follows by the exactness of the five-term exact sequence.
\end{proof}
We may now give an interpretation of the signature of certain surface bundles in terms of the Calabi map of commutators lying in the kernel of a particular extended Flux homomorphism. Specifically, we let $\widetilde{Flux}$ be the pullback of the extended flux map on $Symp(\Sigma_h)$ under the inclusion $Symp^c(\Sigma^1_h) \hookrightarrow Symp(\Sigma_h)$. By Theorem 12 of \cite{KM2} we know that $[\widetilde{Flux}] = [\widetilde{Flux}_c] + p^* [k_{\mathbb{R}}]$. Then as a consequence of Theorem \ref{MMM_Cal} and the calculations in the proof of Theorem \ref{Tsuboi_five} we obtain the following corollary.
\begin{cor}\label{Tsuboi_first_MMM}
Let $\Sigma^1_h \to E \to \Sigma_g$ be a bundle with holonomy representation $\pi_1(\Sigma_g) \stackrel{\rho} \rightarrow \Gamma_h^1$ and assume $h \geq 2$. Furthermore, let $\alpha_i = \rho(a_i)$ and $\beta_i = \rho(b_i)$ be the images of standard generators $a_i,b_i$ of $\pi_1(\Sigma_g)$. Then for any lifts $\phi_i, \psi_i \in \widetilde{Ham_1^c}(\Sigma_h^1)$ of $\alpha_i, \beta_i$ the signature satisfies
\[\sigma(E) = Cal([\phi_1, \psi_1]...[\phi_g, \psi_g ]).\]
\end{cor}
We contrast Corollary \ref{Tsuboi_first_MMM} with results of \cite{KM}, where it is shown that there exists $Symp^c(\Sigma_g^1)$-bundles with non-zero signature. In particular, we see that those bundles can not have holonomy in the subgroup $\widetilde{Ham_1^c}(\Sigma_g^1)$.
|
\section{Introduction}
Sentiment analysis has become popular in recent years.
Web services, such as socialmention.com, may even score microblog postings on
Identi.ca and Twitter for sentiment in real-time.
One approach to sentiment analysis starts with labeled texts and uses
supervised machine learning trained on the labeled text data to classify
the polarity of new texts \cite{PangB2008Opinion}.
Another approach creates a sentiment lexicon and scores the text based
on some function that describes how the words and phrases of the text
matches the lexicon.
This approach is, e.g., at the core of the \emph{SentiStrength}
algorithm \cite{ThelwallM2010Sentiment}.
It is unclear how the best way is to build a sentiment lexicon.
There exist several word lists labeled with emotional valence, e.g.,
ANEW \cite{BradleyM1999Affective}, General Inquirer, OpinionFinder
\cite{WilsonT2005Recognizing}, SentiWordNet and WordNet-Affect as well as the
word list included in the SentiStrength software
\cite{ThelwallM2010Sentiment}.
These word lists differ by the words they include, e.g., some do not
include strong obscene words and Internet slang acronyms, such as
``WTF'' and ``LOL''.
The inclusion of such terms could be important for reaching good
performance when working with short informal text found in Internet
fora and microblogs.
Word lists may also differ in whether the words are scored with
sentiment strength or just positive/negative polarity.
I have begun to construct a new word list with sentiment strength and
the inclusion of Internet slang and obscene words.
Although we have used it for sentiment analysis on Twitter data
\cite{HansenL2011Good_accepted} we have not yet validated it.
Data sets with manually labeled texts can evaluate the performance of the
different sentiment analysis methods.
Researchers increasingly use Amazon Mechanical Turk (AMT) for creating
labeled language data, see, e.g., \cite{AkkayaC2010Amazon}.
Here I take advantage of this approach.
\section{Construction of word list}
My new word list was initially set up in 2009 for tweets downloaded for
online sentiment analysis in relation to the United
Nation Climate Conference (COP15).
Since then it has been extended.
The version termed AFINN-96 distributed on the
Internet\footnote{http://www2.imm.dtu.dk/pubdb/views/publication\_details.php?id=59819}
has 1468 different words, including a few phrases.
The newest version has 2477 unique words, including 15 phrases that
were not used for this study.
As SentiStrength\footnote{http://sentistrength.wlv.ac.uk/} it uses a
scoring range from $-5$ (very negative) to $+5$ (very positive).
For ease of labeling I only scored for valence, leaving out, e.g.,
subjectivity/objectivity, arousal and dominance.
The words were scored manually by the author.
The word list initiated from a set of obscene words
\cite{BaudhuinE1973Obscene,SapolskyB2008Rating} as well as a few
positive words. It was gradually extended by examining Twitter
postings collected for COP15 particularly the postings which scored high on
sentiment using the list as it grew.
I included words from the public domain \emph{Original Balanced Affective Word
List}\footnote{http://www.sci.sdsu.edu/CAL/wordlist/origwordlist.html}
by Greg Siegle.
Later I added Internet slang by browsing the Urban
Dictionary\footnote{http://www.urbandictionary.com} including acronyms such
as WTF, LOL and ROFL.
The most recent additions come from the large word list by Steven
J. DeRose, \emph{The Compass DeRose Guide to Emotion
Words}.\footnote{http://www.derose.net/steve/resources/emotionwords/ewords.html}
The words of DeRose are categorized but not scored for valence with
numerical values.
Together with the DeRose words I browsed Wiktionary and the synonyms it
provided to further enhance the list.
In some cases I used Twitter to determine in
which contexts the word appeared.
I also used the Microsoft Web n-gram similarity Web service
(``Clustering words based on context
similarity''\footnote{http://web-ngram.research.microsoft.com/similarity/})
to discover relevant words.
I do not distinguish between word categories so to avoid ambiguities
I excluded words such as patient, firm, mean, power and frank.
Words such as ``surprise''---with high arousal but with variable
sentiment---were not included in the word list.
Most of the positive words were labeled with +2 and most of the negative words with
--2, see the histogram in Figure~\ref{fig:hist}.
I typically rated strong obscene words, e.g., as listed in
\cite{BaudhuinE1973Obscene}, with either --4 or --5.
The word list have a bias towards negative words (1598, corresponding
to 65\%) compared to positive words (878). A single phrase was labeled
with valence 0.
The bias corresponds closely to the bias found in the
OpinionFinder sentiment lexicon (4911 (64\%) negative and 2718
positive words).
\begin{wrapfigure}[13]{r}{0.5\textwidth}
\vspace{-0em}
\centering
\includegraphics[width=.5\textwidth]{Nielsen2011New_hist}
\caption{Histogram of my valences.}
\label{fig:hist}
\end{wrapfigure}
I compared the score of each word with mean valence of ANEW.
Figure~\ref{fig:anew} shows a scatter plot for this comparison
yielding a Spearman's rank correlation on 0.81 when words are directly
matched and including words only in the intersection of the two word lists.
I also tried to match entries in ANEW and my word list by applying
Porter word stemming (on both word lists) and WordNet lemmatization (on
my word list) as implemented in NLTK
\cite{BirdS2009Natural}.
The results did not change significantly.
\begin{wrapfigure}[16]{r}{0.5\textwidth}
\vspace{-2em}
\centering
\includegraphics[width=0.5\textwidth]{Nielsen2011New_anew}
\caption{Correlation between ANEW and my new word list.}
\label{fig:anew}
\end{wrapfigure}
When splitting the ANEW at valence 5 and my list at valence 0
I find a few discrepancies: aggressive, mischief, ennui, hard, silly,
alert, mischiefs, noisy.
Word stemming generates a few further discrepancies, e.g.,
alien/alienation, affection/affected, profit/profiteer.
Apart from ANEW I also examined General Inquirer and the OpinionFinder
word lists.
As these word lists report polarity I associated words with
positive sentiment with the valence +1 and negative with --1.
I furthermore obtained the sentiment strength from SentiStrength via
its Web service\footnote{http://sentistrength.wlv.ac.uk/} and
converted its positive and negative sentiments to one single value by
selecting the one with the numerical largest value and zeroing the
sentiment if the positive and negative sentiment magnitudes were equal.
\section{Twitter data}
For evaluating and comparing the word list with ANEW, General
Inquirer, OpinionFinder and SentiStrength a data set of 1,000 tweets
labeled with AMT was applied.
These labeled tweets were collected by Alan Mislove for the
\emph{Twittermood}/``Pulse of a
Nation''\footnote{http://www.ccs.neu.edu/home/amislove/twittermood/}
study \cite{BieverC2010Twitter}.
Each tweet was rated ten times to get a more reliable estimate of
the human-perceived mood, and each rating was a sentiment strength with an
integer between 1 (negative) and 9 (positive).
The average over the ten values represented the canonical ``ground
truth'' for this study.
The tweets were not used during the construction of the word list.
To compute a sentiment score of a tweet I identified words and
found the valence for each word by lookup in the sentiment lexicons.
The sum of the valences of the words divided by the number of words
represented the combined sentiment strength for a tweet.
I also tried a few other weighting schemes: The sum of valence
without normalization of words, normalizing the sum with the number of
words with non-zero valence, choosing the most extreme valence
among the words and quantisizing the tweet valences to +1, 0 and
--1. For ANEW I also applied a version with match using the NLTK
WordNet lemmatizer.
\section{Results}
\begin{wrapfigure}[15]{r}{0.5\textwidth}
\vspace{-6em}
\centering
\includegraphics[width=.5\textwidth]{Nielsen2011New_tweetscatter}
\caption{Scatter plot of sentiment strengths for 1,000 tweets with
AMT sentiment plotted against sentiment found by application or my
word list.}
\label{fig:tweetscatter}
\end{wrapfigure}
My word tokenization identified 15,768 words in total among the 1,000
tweets with 4,095 unique words.
422 of these 4,095 words hit my 2,477 word sized list, while the
corresponding number for ANEW was 398 of its 1034 words.
Of the 3392 words in General Inquirer I labeled with non-zero
sentiment 358 were found in our Twitter corpus and for OpinionFinder
this number was 562 from a total of 6442.
\begin{wraptable}{r}{.5\textwidth}
\vspace{.5em}
\begin{center}
\begin{tabular}{c|ccccc}
& \makebox[9mm]{My} & \makebox[9mm]{ANEW}
& \makebox[9mm]{GI} & \makebox[9mm]{OF} & \makebox[9mm]{SS} \\
\hline
AMT & .564 & .525 & .374 & .458 & .610 \\
My & & .696 & .525 & .675 & .604 \\
ANEW & & & .592 & .624 & .546 \\
GI & & & & .705 & .474 \\
OF & & & & & .512
\end{tabular}
\caption{Pearson correlations between sentiment strength
detections methods on 1,000 tweets. AMT: Amazon Mechanical Turk,
GI: General Inquirer, OF:
OpinionFinder, SS: SentiStrength.}
\label{tab:corrmatrix}
\end{center}
\end{wraptable}
I found my list to have a higher correlation (Pearson correlation:
0.564, Spearman's rank correlation: 0.596, see the scatter plot in
Figure~\ref{fig:tweetscatter}) with the labeling
from the AMT than ANEW had (Pearson: 0.525,
Spearman: 0.544).
In my application of the General Inquirer word list it did not
perform well having a considerable lower AMT correlation than my list
and ANEW (Pearson: 0.374, Spearman: 0.422).
OpinionFinder with its 90\% larger lexicon performed better than
General Inquirer but not as good as my list and ANEW (Pearson: 0.458, Spearman: 0.491).
The SentiStrength analyzer showed superior performance with a Pearson
correlation on 0.610 and Spearman on 0.616, see
Table~\ref{tab:corrmatrix}.
I saw little effect of the different tweet sentiment scoring
approaches: For ANEW 4 different Pearson correlations were in the range
0.522--0.526.
For my list I observed correlations in the range 0.543--0.581 with
the extreme scoring as the lowest and sum scoring without
normalization the highest.
With quantization of the tweet scores to +1, 0 and --1 the
correlation only dropped to 0.548.
For the Spearman correlation the sum scoring with normalization for
the number of words appeared as the one with the highest value
(0.596).
\begin{wrapfigure}[20]{r}{.5\textwidth}
\vspace{-1em}
\centering
\includegraphics[width=.5\textwidth]{Nielsen2011New_evolution}
\caption{Evolution of performance as the word list is extended with
from 5 words to the full set of words (2477). The upper panel is
for the Pearson correlation while the lower for the Spearman rank
correlation. The boxplots are generated from 50 resamples among the 2477 words.}
\label{fig:evolution}
\end{wrapfigure}
Figure~\ref{fig:evolution} plots the evolution of the performance of
the word list on the Twitter as the word list is extended from 5
words to the full set of 2477 words.
To examine whether the difference in performance between the
application of ANEW and my list is due to a different lexicon or a
different scoring I looked on the intersection between the two word
lists.
With a direct match this intersection consisted of 299 words.
Building two new sentiment lexicons with these 299 words, one with the
valences from my list, the other with valences from ANEW, and
applying them on the Twitter data I found that the Pearson
correlations were 0.49 and 0.52 to ANEW's advantage.
\section{Discussion}
On the simple word list approach for sentiment analysis I found my
list performing slightly ahead of ANEW.
However the more elaborate sentiment analysis in SentiStrength showed
the overall best performance with a correlation to
AMT labels on 0.610.
This figure is close to the correlations reported in the evaluation of
the Senti\-Strength algorithm on 1,041 {MySpace}
comments (0.60 and 0.56) \cite{ThelwallM2010Sentiment}.
Even though General Inquirer and OpinionFinder have the largest word
lists I found I could not make them perform as good as
SentiStrength, my list and ANEW for sentiment strength detection in
microblog posting.
The two former lists both score words on polarity rather than
strength and it could explain the difference in performance.
Is the difference between my list and ANEW due to better scoring or
more words?
The analysis of the intersection between the two word list indicated
that the ANEW scoring is better.
The slightly better performance of my list with the entire lexicon
may be due to its inclusion of Internet slang and obscene words.
Newer methods, e.g., as implemented in SentiStrength, use a range of
techniques: detection of negation, handling of emoticons and spelling
variations \cite{ThelwallM2010Sentiment}.
The present application of my list used none of these approaches and
might have benefited.
However, the SentiStrength evaluation showed that valence switching at
negation and emoticon detection might not necessarily increase the
performance of sentiment analyzers (Tables 4 and 5 in
\cite{ThelwallM2010Sentiment}).
The evolution of the performance (Figure~\ref{fig:evolution})
suggests that the addition of words to my list might still
improve its performance slightly.
Although my list comes slightly ahead of ANEW in Twitter sentiment
analysis, ANEW is still preferable for scientific psycholinguistic
studies as the scoring has been validated across several persons.
Also note that ANEW's standard deviation was not used in the
scoring. It might have improved its performance.
\section*{Acknowledgment}
I am grateful to Alan Mislove and Sune Lehmann for providing the 1,000
tweets with the Amazon Mechanical Turk labels and to Steven J. DeRose
and Greg Siegle for providing their word lists.
Mislove, Lehmann and Daniela Balslev also provided input to the article.
I thank the Danish Strategic Research Councils
for generous support to the `Responsible Business in the
Blogosphere' project.
\bibliographystyle{splncs}
|
\section{Introduction}
\subsection{Description of the results}
The famous theorem of Carleson and Hunt asserts that, when $f$ belongs to $L^p(\mathbb T)$, $1<p<+\infty$,
where $\mathbb T=\mathbb R/\mathbb Z$, the sequence of the partial sums of its Fourier series $(S_nf(x))_{n\geq 0}$
converges for almost every $x\in\mathbb T$. On the other hand, it can diverge at some point. This divergence cannot be too
fast since, for any $f\in L^p(\mathbb T)$ and any $x\in\mathbb T$, $|S_n
f(x)|\leq C_p n^{1/p}\Vert f\Vert_p$ (see \cite{Zyg} for instance). In view of these results, a natural question arises. How big can be the sets $F$ such that $|S_nf(x)|$ grows as fast
as possible for every $x\in F$? More generally, can we say something on the size of the sets such that $|S_nf(x)|$
behaves like (or as bad as) $n^{\beta}$ for some $\beta\in(0,1/p]$?
\smallskip
To measure the size of subsets of $\mathbb T$,
we shall use the Hausdorff dimension. Let us recall the relevant definitions (we refer to \cite{Falc} and to \cite{Mat95} for more on this subject).
If $\phi:\mathbb R_+\to\mathbb R_+$ is a nondecreasing continuous function
satisfying $\phi(0)=0$ ($\phi$ is called a \emph{dimension function} or a \emph{jauge function}),
the $\phi$-Hausdorff outer measure of a set $E\subset \mathbb R^d$ is
$$\mathcal H^{\phi}(E)=\lim_{\varepsilon\to 0}\inf_{r\in R_\varepsilon(E)}\sum_{B\in r}\phi(|B|),$$
$R_\varepsilon(E)$ being the set of countable coverings of $E$ with balls $B$ of diameter $|B|\leq\varepsilon$.
When $\phi_s(x)=x^s$, we write for short $\mathcal H^s$ instead of $\mathcal H^{\phi_s}$. The Hausdorff dimension of a set $E$
is
$$\dim_{\mathcal H}(E):=\sup\{s>0; \mathcal H^s (E)>0\}=\inf\{s>0;\ \mathcal H^s(E)=0\}.$$
The first result studying the Hausdorff dimension of the divergence sets of
Fourier series is due to J-M. Aubry \cite{Aub06}.
\begin{theorem}\label{THMAUBRY}
Let $f\in L^p(\mathbb T)$, $1<p<+\infty$. For $\beta\geq 0$, define
$$\mathcal E(\beta,f)=\left\{x\in\mathbb T;\ \limsup_{n\to+\infty}n^{-\beta}|S_nf(x)|>0\right\}.$$
Then $\dim_\mathcal{H}\big(\mathcal E(\beta,f)\big)\leq 1-\beta p$. Conversely, given a set $E$ such that $\dim_\mathcal{H}(E)<1-\beta p$, there exists
a function $f\in L^p(\mathbb T)$ such that, for any $x\in E$,
$\displaystyle\limsup_{n\to +\infty}n^{-\beta}|S_nf(x)|=+\infty$.
\end{theorem}
This result motivated us to introduce the notion of divergence index. For a
given function $f\in L^p(\mathbb T)$ and a given point $x_0\in \mathbb T$, we can define
the real number $\beta(x_0)$ as the infimum of the non negative real numbers
$\beta$ such that $|S_nf(x_0)|=O(n^\beta)$. The real number $\beta(x_0)$ will
be called the \emph{divergence index} of the Fourier series of $f$
at point $x_0$. Of
course, for any function $f\in L^p(\mathbb T)$ ($1<p<+\infty$) and any point $x_0\in\mathbb T$, $0\le
\beta(x_0)\le 1/p$. Moreover, Carleson's theorem implies that $\beta(x_0)=0$
almost surely and we would like to have precise estimates on the size of the level sets of the function
$\beta$. These are defined as
\begin{eqnarray*}
E(\beta,f)&=&\left\{x\in\mathbb T;\ \beta(x)=\beta\right\}\\
&=&\left\{x\in\mathbb T;\ \limsup_{n\to+\infty}\frac{\log |S_nf(x)|}{\log
n}=\beta\right\}.
\end{eqnarray*}
We can ask for which values of $\beta$ the sets
$E(\beta,f)$ are non-empty. This set of values will be called the domain of definition of the spectrum of singularities of $f$.
If $\beta$ belongs to the domain of definition of the spectrum of singularities, it is also interesting to estimate the
Hausdorff dimension of the sets $E(\beta,f)$. The function $\beta\mapsto
\dim_\mathcal{H}(E(\beta,f))$ will be called the spectrum of singularities of the
function $f$ (in terms of its Fourier series).
By Aubry's result, $\dim_\mathcal{H}(E(\beta,f))\leq1-\beta p$ and, for any fixed $\beta_0\in[0,1/p)$,
for any $\varepsilon>0$, one can find
$f\in L^p(\mathbb T)$ such that $\dim_\mathcal{H}\left(\bigcup_{\beta_0\leq\beta\leq
1/p}E(\beta,f)\right)\geq 1-\beta_0p-\varepsilon$. Our first main result
is that a \emph{typical} function $f\in L^p(\mathbb T)$ satisfies $\dim_\mathcal{H}(E(\beta,f))=1-\beta p$ for \emph{any} $\beta\in[0,1/p]$.
In particular, $f$ has a multifractal behavior with respect to the summation of its Fourier series, meaning that the domain of
definition of its spectrum of singularities contains an interval with non-empty interior.
\begin{theorem}\label{THMLP}
Let $1<p<+\infty$. For quasi-all functions $f\in L^p(\mathbb T)$, for any $\beta\in[0,1/p]$, $\dim_\mathcal{H}\big(E(\beta,f)\big)=1-\beta p$.
\end{theorem}
The terminology "quasi-all" used here is relative to the Baire category theorem. It means that this property is true for a residual
set of functions in $L^p(\mathbb T)$.
\medskip
In a second part of the paper, we turn to the case of $\mathcal C(\mathbb
T)$, the set of continuous functions on $\mathbb T$.
In that space, the divergence of Fourier series is controlled by a
logarithmic factor.
More precisely, if $(D_n)$ is the sequence of the Dirichlet kernels, we know that $\|S_n
f\|_\infty\leq \|D_n\|_1\|f\|_\infty$, so that
there exists some absolute constant $C>0$ such that
$\|S_n f\|_\infty\leq C\|f\|_\infty\log n$ for any $f\in\mathcal C(\mathbb T)$ and any $n> 1$.
As before, one can discuss the size of the sets such that $|S_nf(x)|$ behaves badly, namely like $(\log n)^{\beta}$,
$\beta\in[0,1]$. More precisely, mimicking the case of the $L^p$ spaces,
we introduce, for any $\beta\in[0,1]$ and any $f\in\mathcal C(\mathbb T)$, the following sets:
\begin{eqnarray*}
\mathcal F(\beta,f)&=&\left\{x\in\mathbb T;\ \limsup_{n\to+\infty}\,(\log n)^{-\beta}|S_nf(x)|>0\right\}\\
F(\beta,f)&=&\left\{x\in\mathbb T;\ \limsup_{n\to+\infty}\frac{\log |S_nf(x)|}{\log\log n}=\beta\right\}.
\end{eqnarray*}
Theorem \ref{THMAUBRY} indicates that, on $L^p(\mathbb T)$, $|S_nf(x)|$ can grow as fast as possible (namely like $n^{1/p}$)
only on small sets: for every function $f\in L^p(\mathbb T)$,
$\dim_{\mathcal H}(E(1/p,f))=0$. This property dramatically breaks down on $\mathcal C(\mathbb T)$, as the following result indicates.
\begin{theorem}\label{THMCT1}
For quasi-all functions $f\in\mathcal C(\mathbb T)$, $\dim_{\mathcal H}\big(F(1,f)\big)=1$.
\end{theorem}
Thus, for quasi-all functions $f\in\mathcal C(\mathbb T)$, the partial sums
$(S_nf(x))_{n\ge 0}$ grow as fast as possible on big sets.
We can also study the domain of the spectrum of singularities of $f$, namely the values of $\beta$ such that $F(\beta,f)$ is non-empty.
Like in the case of the space $L^p(\mathbb T)$, this domain is for quasi-all functions of $\mathcal C(\mathbb T)$ an interval with non-empty
interior, so that a typical function $f$ in $\mathcal C(\mathbb T)$ has a multifractral behavior with respect to the summation
of its Fourier series.
However, the spectrum of singularities is constant!
\begin{theorem}\label{THMCT2}
For quasi-all functions $f\in\mathcal C(\mathbb T)$, for any $\beta\in[0,1]$, $F(\beta,f)$ is non-empty
and has Hausdorff dimension 1.
\end{theorem}
Theorem \ref{THMCT2} indicates that the Hausdorff dimension is not precise enough to measure the size of the level sets $F(\beta,f)$.
This leads us to introduce a notion of \emph{precised Hausdorff dimension}, in
order to distinguish more finely sets which have the same
Hausdorff dimension. For $s> 0$ and $t\in(0,1]$, we consider
$$\phi_{s,t}(x)=x^s \exp\left[(\log 1/x)^{1-t}\right].$$
\begin{definition}
Let $E\subset\mathbb R^d$. We say that $E$ has \emph{precised Hausdorff
dimension} $(\alpha,\beta)$ if $\alpha$ is the Hausdorff dimension of $E$ and
\begin{itemize}
\item $\beta=0$ if $\mathcal H^{\phi_{\alpha,t}}(E)=0$ for every $t\in(0,1)$;
\item $\beta=\sup\big\{t\in(0,1);\ \mathcal H^{\phi_{\alpha,t}}(E)>0\big\}$ otherwise.
\end{itemize}
\end{definition}
It is not difficult to check that $\phi_{s,t}(x)\leq\phi_{s',t'}(x)$ for small values of $x$ iff
$$s>s'\textrm{ or }(s=s'\textrm{ and }t\geq t').$$
Thus the precised Hausdorff dimension is a refinement of the Hausdorff
dimension. In particular it is a tool to classify sets that have the same
Hausdorff dimension. The natural order for the precised dimension $(s,t)$ is
the lexicographical order which will be denoted by $\prec$.
With respect to this order, we can say that the
greater is the set, the greater is the precised dimension. Moreover, if
$(s,t)\prec (s',t')$ and $(s,t)\not= (s',t')$, then $\phi_{s',t'}\ll
\phi_{s,t}$. It follows that
$H^{\phi_{s',t'}}(E)=0$ as soon as $H^{\phi_{s,t}}(E)<\infty$.
\smallskip
Our main theorem on $\mathcal C(\mathbb T)$, which contains both Theorems \ref{THMCT1} and \ref{THMCT2}, is the following:
\begin{theorem}\label{THMCT3}
For quasi-all functions $f\in\mathcal C(\mathbb T)$, for any $\beta\in[0,1]$, the precised Hausdorff dimension of $F(\beta,f)$ is $(1,1-\beta)$.
\end{theorem}
\medskip
The paper is organized as follows. In the remaining part of this section, we introduce tools which will be needed during the rest of the paper.
In Section \ref{SECLP}, we prove Theorem \ref{THMLP} whereas in Section \ref{SECCT}, we prove Theorem \ref{THMCT3}.
\subsection{A precised version of Fej\'er's theorem}
Working on Fourier series, we will need results on approximation by trigonometric polynomials.
Let $k\in\mathbb Z$ and $e_k:t\mapsto e^{2\pi i kt}$, so that, for any $g\in L^1(\mathbb T)$
and any $n\in\mathbb N$,
$$S_ng:t\mapsto \sum_{k=-n}^n \langle g,e_k\rangle e_k(t).$$
Let $\sigma_ng$ be the $n-$th Fej\'er sum of $g$,
$$\sigma_n g:t\mapsto \frac 1n\sum_{k=0}^{n-1}S_kg(t).$$
$\sigma_n g$ is obtained by taking the convolution of $g$ with the Fej\'er kernel
$$F_n:t\mapsto \frac 1n\left(\frac{\sin(n\pi t)}{\sin(\pi t)}\right)^2.$$
If $g$ belongs to $\mathcal C(\mathbb T)$, $(\sigma_n g)_{n\ge 1}$ converges uniformly to $g$. For our purpose, we need to estimate how quick is the convergence. This is the content
of the next lemma (part (1) rectifies a mistake in the proof of Lemma 12
in \cite{Aub06} and requires to replace $\|\theta\|_\infty/4$ in Aubry's
version by $\|\theta\|_\infty/2$).
\begin{lemma}\label{LEMFEJER}Let $\theta$ be a Lipschitz
function on $\mathbb T$, let $n\in\mathbb N$ and let $x\in\mathbb T$. Suppose that $\|\theta'\|_\infty\le n$
and that $\theta(x)=0$. Then the two following inequalities hold:
\begin{eqnarray}\label{EQFEJER1}
|\sigma_{n}\theta(x)|\le\frac14+\frac12\|\theta\|_\infty\ \textrm{ for any }n\geq 8\\
|\sigma_{n}\theta(x)|\le 4+\frac14\|\theta\|_\infty\ \textrm{ for any }n\geq 4.\label{EQFEJER2}
\end{eqnarray}
\end{lemma}
\begin{proof}
We may assume that $x=0$. Hence,
$\sigma_{n}\theta(0)=\int_{-1/2}^{1/2}\theta(y)F_{n}(y)\,dy$. Let us consider $\delta\in(0,2]$ and $n\ge 4$.
On the one hand, for any $y\in[0,1/2)$,
$$0\leq F_{n}(y)=\frac{\sin^2(n\pi y)}{n\sin^2(\pi
y)}\leq\frac1{n(2y)^2}$$
so that
$$\left|\int_{\delta/n<|y|\le1/2}\theta(y)F_{n}(y)\,dy\right|\leq
\frac{1}{2n}\|\theta\|_\infty\int_{\delta/n}^{+\infty}\frac{dy}{y^2}=\frac{\|\theta\|_\infty}{2\delta}.$$
On the other hand,
$$\left|\int_{-\delta/n}^{\delta/n}\theta(y)F_{n}(y)\,dy\right|\leq
2\int_0^{\delta/n}\left(\frac{\sin (n\pi y)}{\sin(\pi y)}\right)^2y\,dy := u_n.$$
Using the convexity inequality $\sin\left(\frac{n}{n+1}\pi y\right)\geq
\frac{n}{n+1}\sin(\pi y)$ and a change of variables, we see that $(u_n)$ is
non-increasing. To prove (\ref{EQFEJER1}), we choose $\delta=1$ and we observe that $u_8=0.2496...\leq \frac14$.
To prove (\ref{EQFEJER2}), we choose $\delta=2$ and we observe that, since the maximum of $F_n$ is $F_n(0)=n$,
$$|u_n|\leq 2n^2\int_0^{2/n} ydy=4.$$
\end{proof}
\subsection{The mass transference principle}
The second main tool that we need in this paper is a method to produce sets with large Hausdorff dimension (Theorem \ref{THMLP})
or with large precised Hausdorff dimension (Theorem \ref{THMCT3}). An efficient way to do this
is to consider \emph{ubiquitous systems} like this was done in \cite{DMPV95,Jaf00b}. This was later refined in \cite{BV06} to obtain a general mass transference principle, which we recall in the form that we need.
\begin{theorem}[The mass transference principle]\label{THMUBI2}
Let $(x_n)_{n\ge 0}$ be a sequence of points in $[0,1]^d$ and let $(r_n)_{n\ge 0}$
be a sequence of positive real numbers decreasing to 0. Let also $\phi:\mathbb R_+\to\mathbb R_+$
be a dimension function satisfying $\phi(s)\gg s^d$ when $s$
goes to 0 and $s^{-d}\phi$ is monotonic. Define
\begin{eqnarray*}
E&=&\limsup_n B(x_n,r_n)\\
E^\phi&=&\limsup_n B\left(x_n,\phi^{-1}(r_n^d)\right)
\end{eqnarray*}
and suppose that almost every point of $[0,1]^d$ (in the sense of the
Lebesgue's measure) lies in $E$. Then,
$\mathcal{H}^\phi(E^\phi)=+\infty$.
\end{theorem}
We shall use it in the following situation.
\begin{corollary}\label{CORUBI2}
Let $(q_n)$ be a sequence of integers and, for each $n\in\mathbb N$, each $k\leq q_n$, let
$B_{k,n}=B(x_{k,n},r_n)$ be a ball with center $x_{k,n}\in[0,1]^d$ and with radius $r_{k,n}$
such that $\lim_{n\to+\infty}\max_k(r_{k,n})=0$. Let also
$\phi:\mathbb R_+\to\mathbb R_+$
be a dimension function satisfying $\phi(s)\gg s^d$ when $s$
goes to 0 and $s^{-d}\phi$ is monotonic. Define
$$\begin{array}{rclcrcl}
B_n&=&\bigcup_{k=1}^{q_n}B_{k,n}&\quad&E&=&\limsup_n B_n\\
B_n^\phi&=&\bigcup_{k=1}^{q_n}B(x_{k,n},\phi^{-1}(r_{k,n}^d))&\quad&E^\phi&=&\limsup_n B_n^\phi.
\end{array}
$$
Suppose that almost every point of $[0,1]^d$ (in the sense of the
Lebesgue's measure) lies in $E$. Then,
$\mathcal{H}^\phi(E^\phi)=+\infty$.
\end{corollary}
\begin{proof}
Reordering the sequences $(B_{k,n})$ and $(B_{k,n}^\phi)$ as $(C_j)$ and $(C^\phi_j)$, we can observe that
\begin{eqnarray*}
\limsup_n B_n&=&\limsup_j C_j=E\\
\limsup_n B_n^\phi&=&\limsup_j C_j^\phi=E^\phi.
\end{eqnarray*}
Thus the corollary follows from a direct application of Theorem \ref{THMUBI2}.
\end{proof}
\section{Multifractal analysis of the divergence of the Fourier series of functions of $L^p(\mathbb T)$}\label{SECLP}
In this section, we shall prove Theorem \ref{THMLP}. Our method, which is inspired by \cite{Jaf00a}, is divided into two parts.
During the first one, we will construct a single function, which we call the saturating function, satisfying the conclusions of
Theorem \ref{THMLP}. During the second one, we will show how to derive a residual set from this single function.
\subsection{The saturating function}\label{SECSATURATINGLP}
Our intention is to construct a function $g$ such that $|S_ng(x)|$ is big when $x$ is close to a dyadic number. The following
definition gives a precise meaning.
\begin{definition}
A real number $x$ is $\alpha$-approximable by dyadics, $\alpha\geq 1$, if there exist two sequence of integers $(k_n)$, $(j_n)$
such that
$$\left|x-\frac{k_n}{2^{j_n}}\right|\leq\frac{1}{2^{\alpha j_n}}$$
and $(j_n)$ goes to infinity. The dyadic exponent of $x$ is the supremum of the set of real numbers $\alpha$ such that $x$ is
$\alpha$-approximable by dyadics.
\end{definition}
We denote by
$$D_\alpha=\left\{x\in[0,1];\ \textrm{$x$ is $\alpha$-approximable by dyadics}\right\}.$$
It is easy to check that ${\mathcal H}^{\beta}(D_\alpha)=0$ for $\beta>1/\alpha$ so that $\dim_{\mathcal H}(D_\alpha)\le 1/\alpha$. On the other hand, it is well-known that $\dim_{\mathcal H}(D_\alpha) \geq\frac1\alpha$.
Let us nevertheless show how this
follows from Corollary \ref{CORUBI2}. Indeed, $D_\alpha$ can be described as a limsup set:
$$D_\alpha=\limsup_{j\to+\infty}\bigcup_{k=0}^{2^j-1}I_{k,j}^\alpha$$
where the $I_{k,j}$ are the dyadic intervals
$$I_{k,j}=\left[\frac{k}{2^j}-\frac1{2^{j}},\frac{k}{2^j}+\frac1{2^{j}}\right]$$
and
$$I_{k,j}^\alpha=\left[\frac{k}{2^{j}}-\frac1{2^{\alpha j}},\frac{k}{2^j}+\frac1{2^{\alpha j}}\right].$$
Since $\bigcup_{k=0}^{2^j-1}I_{k,j}\supset [0,1]$, Corollary \ref{CORUBI2} implies that
$\mathcal H^{1/\alpha}(D_\alpha)=+\infty$.
\medskip
We are going to define $g\in L^p(\mathbb T)$ such that the divergence index of $g$ at $x$ depends on the dyadic exponent of $x$. The greater the dyadic exponent will be, the greater the divergence index of $g$ at $x$ will be. To do this, we will classify the dyadic intervals following their center. Namely, each $k/2^j$ can be uniquely written $K/2^J$
with $K\notin 2\mathbb Z$ and $1\leq J\leq j$ (such a center comes into play from the $J$-th generation). Let
$\mathcal I_J=\{K/2^J;\ K\notin 2\mathbb Z,\ 0\leq K\leq 2^{J}-1\}$ and
$$\mathbf I_{J,j}=\bigcup_{\frac{k}{2^j}\in\mathcal I_J}I_{k,j}\quad\quad \mathbf I'_{J,j}=\bigcup_{\frac{k}{2^j}\in\mathcal I_J}2I_{k,j}.$$
Here and elsewhere, when $I$ is an interval and $c$ is a positive real number, $c I$ means the interval with the same
center as $I$
and with length $c |I|$. Observe that, when $1\le J<j$, the intervals
$2I_{k,j}$, $\frac{k}{2^j}\in\mathcal I_J$ don't overlap and the set $\mathbf I'_{J,j}$ has measure
$2^{J-1}2^{2-j}$. Observe also that when $J$ is small with respect to
$j$, the real numbers $x$ in $\mathbf I_{J,j}$
are well-approximated by dyadics $K/2^J$, since $|x-K/2^J|\leq 1/2^j$.
We first define a trigonometric polynomial with $L^p$-norm 1 which is almost constant on each
$\mathcal I_{J,j}$ and which is big on $\mathcal I_{J,j}$ when $J$ is small.
\begin{lemma}\label{LEMBASICLP}
Let $j\geq 1$. There exists a trigonometric polynomial $g_j\in L^p(\mathbb T)$ with spectrum contained in
$[0,j2^{j+1})$ such that
\begin{itemize}
\item $\|g_j\|_p\leq 1$;
\item For any $1\leq J\leq j$ and any $x\in \mathbf I_{J,j}$, we can find two
integers $n_1$ and $n_2$ satisfying
$0\le n_1<n_2< j2^{j+1}$ and such that
$$|S_{n_2}g_j(x)-S_{n_1}g_j(x)|\geq\frac{1}{4j}2^{-(J-j+1)/p}.$$
\end{itemize}
\end{lemma}
\begin{proof}
We set for any $1\leq J\leq j$:
\begin{itemize}
\item $\chi_{J,j}$ a continuous piecewise linear function equal to 1 on $\mathbf I_{J,j}$, equal to 0 outside $\mathbf I'_{J,j}$, and satisfying
$0\leq \chi_{J,j}\leq 1$ and $\|\chi'_{J,j}\|_\infty\le 2^{j}$;
\item $c_{J,j}=\frac{1}{j}2^{-(J-j+1)/p}$ ($c_{J,j}$ is big when $J$ is small);
\item $g_{J,j}=e_{(2J-1)2^{j}}\sigma_{2^j}\chi_{J,j}$.
\end{itemize}
It is straighforward to observe that the spectrum of $g_{J,j}$ is contained in $[n_{J,j},m_{J,j}]$ with
$$\left\{\begin{array}{rcl}
n_{J,j}&=&(2J-1)2^{j}-(2^j-1)\\
m_{J,j}&=&(2J-1)2^{j}+(2^j-1).
\end{array}\right.$$
Thus, the spectra of the $g_{J,j}$, $1\leq J\leq j$ are
disjoint. Moreover, $\Vert g_{j,j}\Vert_p=1$ and for $1\le J<j$, $\Vert
g_{J,j}\Vert_p\le\Vert\chi_{J,j}\Vert_p\le 2^{(J-j+1)/p}$.
We finally set
$$g_j=\sum_{J=1}^j c_{J,j}g_{J,j}$$
and we claim that $g_j$ is the trigonometric polynomial we are looking
for. First of all, the spectrum of $g_j$ is included in
$[n_{1,j},m_{j,j}]$ which is contained in $[0,j2^{j+1})$. Moreover, the $L^p$
norm of $g_j$ is
$$\|g_j\|_p\leq\sum_{J=1}^j \frac{1}{j}2^{-(J-j+1)/p}\|g_{J,j}\|_p\leq 1.$$
Pick now any $x\in \mathbf I_{J,j}$, $1\leq J\leq j$ so that
\begin{eqnarray*}
|S_{m_{J,j}}g_j(x)-S_{n_{J,j}-1}g_j(x)|&=&|c_{J,j}g_{J,j}(x)|\\
&=&\frac{1}j 2^{-(J-j+1)/p} |\sigma_{2^j}\chi_{J,j}(x)|.
\end{eqnarray*}
Observing that $\chi_{J,j}(x)=1$ and applying the first point of Lemma
\ref{LEMFEJER} to $1-\chi_{J,j}$, we find
$$\left|\sigma_{2^j}\chi_{J,j}(x)\right|\geq1-\left|\sigma_{2^j}(1-\chi_{J,j}(x))\right|\geq \frac14.$$
Thus,
$$|S_{m_{J,j}}g_j(x)-S_{n_{J,j}-1}g_j(x)|\geq \frac{1}{4j}2^{-(J-j+1)/p}$$
and the conclusion follows with $n_2=m_{J,j}$ and $n_1=n_{J,j}-1$.
\end{proof}
We are now ready to construct the saturating function. It is defined by
$$g=\sum_{j\geq 1}\frac1{j^2}e_{j2^{j+1}}g_j.$$
Observe in particular that the functions
$e_{j2^{j+1}}g_j$ have disjoint spectra (the spectrum of $e_{j2^{j+1}}g_j$ is
contained in $[j2^{j+1};j2^{j+2})$ ) and that $g$ belongs to $L^p(\mathbb T)$.
We then show that for any $x\in D_\alpha$, $\alpha>1$,
$$\displaystyle\limsup_{n\to+\infty}\frac{\log |S_ng(x)|}{\log
n}\geq\frac1p\left(1-\frac1\alpha\right).$$
Indeed, let $x\in D_\alpha$
and let $\varepsilon>0$ with $\alpha-\varepsilon>1$. We can find integers $K$ and $J$ with
$J$ as large as we want and $K\notin 2\mathbb Z$ such that
$$\left|x-\frac{K}{2^J}\right|\leq\frac1{2^{(\alpha-\varepsilon/2)J}}.$$
We set $j=[(\alpha-\varepsilon/2)J]$ the integer part of $(\alpha-\varepsilon/2)J$ and $k$ such that $k/2^j=K/2^J$. Hence,
$$\left|x-\frac{k}{2^j}\right|\leq\frac{1}{2^{(\alpha-\varepsilon/2)J}}\leq \frac{1}{2^{j}}.$$
Using Lemma \ref{LEMBASICLP}, we can find two integers $n_1$ and $n_2$
satisfying $j2^{j+1}\le n_1< n_2< j2^{j+2}$ and such that
\begin{eqnarray*}
|S_{n_2}g(x)-S_{n_1}g(x)|&=&\frac{1}{j^2}|S_{n_2}(e_{j2^{j+1}}g_j)(x)-S_{n_1}(e_{j2^{j+1}}g_j)(x)|\\
&\geq&\frac{1}{4j^3}2^{-(J-j+1)/p}\\
&\geq&\frac{1}{4j^3}2^{\frac1p\left(j-\frac{j+1}{\alpha-\varepsilon/2}-1\right)}\\
&\geq&C2^{\frac1p\left(1-\frac{1}{\alpha-\varepsilon}\right)j}.
\end{eqnarray*}
It follows that we can find $n\in\{ n_1,n_2\}$ such that $|S_{n}g(x)|\ge
\frac{C}2 2^{\frac1p\left(1-\frac{1}{\alpha-\varepsilon}\right)j}$. Combining the estimates on $n$ and on $|S_ng(x)|$, and since $J$ (hence $j$, hence $n$) can be taken as large as we want, we get that
$$\limsup_{n\to+\infty}\frac{\log |S_ng(x)|}{\log n}\geq\frac1p\left(1-\frac{1}{\alpha-\varepsilon}\right).$$
Since $\varepsilon>0$ is arbitrary, we obtain in fact that
$$\textrm{for any }x\in D_\alpha,\quad \limsup_{n\to+\infty}\frac{\log |S_ng(x)|}{\log n}\geq\frac1p\left(1-\frac1\alpha\right).$$
At this point, it would be nice to get a lower bound for
$\displaystyle\limsup_{n\to +\infty} \frac{\log
|S_ng(x)|}{\log n}$ for any $x$ with dyadic exponent equal to
$\alpha$. Unfortunately, this does not seem
easy and we will rather conclude by using an argument lying on Hausdorff measures. Indeed, define
\begin{eqnarray*}
D_\alpha^1&=&\left\{x\in D^\alpha;\ \limsup_{n\to +\infty} \frac{\log |S_ng(x)|}{\log n}=\frac1p\left(1-\frac{1}{\alpha}\right)\right\}\\
D_\alpha^2&=&\left\{x\in D^\alpha;\ \limsup_{n\to +\infty}\frac{\log |S_ng(x)|}{\log n}>\frac1p\left(1-\frac{1}{\alpha}\right)\right\}.
\end{eqnarray*}
We have already observed that $\mathcal{H}^{1/\alpha}(D_\alpha^1\cup
D_\alpha^2)=\mathcal{H}^{1/\alpha}(D_\alpha)=+\infty$. It suffices
to prove that $\mathcal{H}^{1/\alpha}(D_\alpha^2)=0$.
Let $(\beta_n)$ be a sequence of real numbers such that $\displaystyle\beta_n
>\frac1p\left(1-\frac1\alpha\right)$ and $\displaystyle\lim_{n\to
+\infty}\beta_n=\frac1p\left(1-\frac1\alpha\right)$.
Let us observe that
$$D_\alpha^2\subset \bigcup_{n\ge 0}\mathcal E(\beta_n,g).$$
Moreover, Theorem \ref{THMAUBRY} implies that $\mathcal{H}^{1/\alpha}(\mathcal E(\beta_n,g))=0$ for
all $n$. Hence, $\mathcal{H}^{1/\alpha}(D_\alpha^2)=0$ and $\mathcal{H}^{1/\alpha}(D_\alpha^1)=+\infty$, which proves
that
$$\dim_\mathcal{H}\left(E\left(\frac1p\left(1-\frac{1}{\alpha}\right),g\right)\right)
\ge\frac1\alpha.$$
By Theorem \ref{THMAUBRY} again, this inequality is necessarily an equality.
Finally, $g$ satisfies the conclusions of Theorem \ref{THMLP}, setting $1-\beta p=1/\alpha$.
\begin{remark}
If $\alpha=1$, then $\beta=0$ and the conclusion is a
consequence of Carleson's Theorem.
\end{remark}
\subsection{The residual set}
To build the dense $G_\delta$-set, the idea is that any function whose Fourier
coefficients are sufficiently
close to those of the saturating function $g$ on infinitely many intervals $[j2^{j+1};j2^{j+2})$ will
satisfy the conclusions of Theorem \ref{THMLP}. Precisely,
let $(f_j)_{j\ge 1}$ be a dense sequence of polynomials in $L^p(\mathbb T)$ with
spectrum contained in $[-j,j]$.
We define a sequence $(h_j)_{j\ge 1}$ as follows:
$$h_j=f_j+\frac{1}{j}e_{j2^{j+1}}g_j$$
so that $\|h_j-f_j\|_p$ goes to 0 and $(h_j)_{j\ge 1}$ remains dense in
$L^p(\mathbb T)$. Observe also that the spectra of $f_j$ and $h_j-f_j$ don't
overlap. Finally, let $(r_j)_{j\ge 1}$ be a
sequence of positive integers so small that, for any $f\in L^p(\mathbb T)$
with $\|f\|_{L^p}\leq r_j$, $\|S_n f\|_\infty\leq 1$ for any $n\le j2^{j+2}$.
The dense $G_\delta$ set we will consider is
$$A=\bigcap_{l\in\mathbb N}\bigcup_{j\geq l}B(h_j,r_j).$$
Let $f$ belong to $A$ and let $(j_l)_{l\ge 1}$ be an increasing sequence of integers such that $f$ belongs to $B(h_{j_l},r_{j_l})$ for any $l$. Then,
for any $\alpha> 1$, we define $J_l=[j_l/\alpha]+1$ (which is smaller than
$j_l$ if $l$ is large enough) and
$$E=\limsup_{l \to+\infty} \mathbf I_{J_l,j_l}.$$
For any $x\in E$ one can find $j=j_l$ as large as we want, the
corresponding $J=J_l$ and $1\leq k\leq 2^j-1$ such that $x$ belongs to $I_{k,j}$
with $k/2^j\in \mathcal I_{J}$.
Observe that $f=f_j+\frac1je_{j2^{j+1}}g_j+(f-h_j)$.
By Lemma \ref{LEMBASICLP}, we can find two
integer $n_1$ and $n_2$ satisfying $j2^{j+1}\le n_1< n_2< j2^{j+2}$ and such
that
$$|S_{n_2}(e_{j2^{j+1}}g_j)(x)-S_{n_1}(e_{j2^{j+1}}g_j)(x)|\ge
\frac{1}{4j}2^{-(J-j+1)/p}.$$
Using the definition of the $r_j$, we obtain
\begin{eqnarray*}
|S_{n_2}f(x)-S_{n_1}f(x) |&\geq&\frac1{4j^2}2^{-(J-j+1)/p}-|S_{n_2}(f-h_j)(x)|-|S_{n_1}(f-h_j)(x)|\\
&\geq&\frac{1}{4j^2}2^{-(J-j+1)/p}-2
\end{eqnarray*}
so that
$$|S_{n_2}f(x)|\ge\frac{C}{j^2}2^{-(J-j+1)/p}\quad\text{or}\quad |S_{n_1}f(x)|\ge\frac{C}{j^2}2^{-(J-j+1)/p}.$$
Observing that
$$
\left\{\begin{array}{rcl}
\max(\log n_2,\log n_1)&=&j\log 2+O(\log j)\\
\null\\
\log\left(j^{-2}2^{-(J-j+1)/p}\right)&=&\frac1p
\left(1-\frac1\alpha\right)j\log2+O(\log j)
\end{array}\right.$$
we find in particular that, for any $x\in E$,
$$\limsup_{n\to +\infty} \frac{\log |S_nf(x)|}{\log n}\geq
\frac1p\left(1-\frac 1\alpha\right).$$
On the other hand, let us write
$$\mathbf I_{J_l,j_l}=\bigcup_{1\le K<2^{J_l},\ K\notin 2{\mathbb Z}}\left[\frac{K}{2^{J_l}}-\frac1{2^{j_l}},\frac{K}{2^{J_l}}+\frac{1}{2^{j_l}}\right]$$
and remark that for any $l$, since $J_l\geq j_l/\alpha$,
$$\bigcup_{1\le K<2^{J_l},\ K\notin 2{\mathbb Z}}\left[\frac{K}{2^{J_l}}-\frac1{2^{j_l/\alpha}},\frac{K}{2^{J_l}}+\frac{1}{2^{j_l/\alpha}}\right]\supset [0,1].$$
Hence, we can apply Corollary \ref{CORUBI2} to get $\mathcal{H}^{1/\alpha}(E)=+\infty$. We now conclude exactly as in Section \ref{SECSATURATINGLP}
to get $\mathcal{H}^{1/\alpha}(E^1)=+\infty$, with
$$E^1=\left\{x\in E;\ \limsup_{n\to +\infty} \frac{\log |S_nf(x)|}{\log n}=\frac1p\left(1-\frac1\alpha\right)\right\}.$$
Finally,
$$\dim_\mathcal{H}\left(E\left(\frac1p\left(1-\frac{1}{\alpha}\right),f\right)\right)
\ge\frac1\alpha$$
and $f$ satisfies the conclusions of Theorem \ref{THMLP}, setting $1-\beta p=1/\alpha$.
\begin{remark}During the construction , we didn't use that the spectra
of the functions $e_{j2^{j+1}}g_j$ are disjoint,
because we considered each one separately. We could also define $h_j$ by $h_j=f_j+\frac 1je_{j+1}g_j$.
\end{remark}
\begin{remark}
The above construction can be carried on $L^1(\mathbb T)$. Namely, for quasi-all
$f\in L^1(\mathbb T)$, we obtain for any $\beta\in [0,1]$,
$$\dim_{\mathcal H}\left( E\left(\beta,f\right)\right)\geq 1-\beta .$$
However, we cannot go further because Carleson's Theorem dramatically breaks
down in $L^1(\mathbb T)$ and we do not have Theorem \ref{THMAUBRY} at our disposal in this context.
The study of what happens exactly on $L^1(\mathbb T)$ is a very exciting open question.
\end{remark}
\section{Multifractal analysis of the divergence of the Fourier series of functions of $\mathcal C(\mathbb T)$}\label{SECCT}
We turn to the proof of Theorem \ref{THMCT3}. We follow a strategy close to that of Section \ref{SECLP}. First of all,
we will give un upper bound for
the precised Hausdorff dimension of the sets $\mathcal F(\beta,f)$ (hence, of the sets $F(\beta,f)$) for
any $f\in\mathcal C(\mathbb T)$ and any $\beta\in(0,1)$. Second, we will build polynomials with small $L^\infty$-norms
and such that their Fourier series have big partial sums on big intervals.
These polynomials will be the blocks of our final construction. Working on $\mathcal C(\mathbb T)$ adds several difficulties which will be explained
when we will encounter them.
\subsection{The sets $\mathcal F(\beta,f)$ cannot be too big}
We shall prove the following lemma (recall that $\phi_{s,t}(x)=x^s\exp\left((\log 1/x)^{1-t}\right)$).
\begin{lemma}\label{LEMCTMAJO}
Let $\beta\in(0,1)$ and $f\in\mathcal C(\mathbb T)$. Then, for any $\gamma>1-\beta$,
$$\mathcal H^{\phi_{1,\gamma}}\big(\mathcal F(\beta,f)\big)=0.$$
In particular, the precised Hausdorff dimension of $\mathcal F(\beta,f)$ cannot exceed $(1,1-\beta)$.
\end{lemma}
\begin{proof}
A key point in Aubry's proof of Theorem \ref{THMAUBRY} is the Carleson-Hunt theorem which asserts that,
for any $g\in L^p(\mathbb T)$, $1<p<+\infty$,
$$\|S^* g\|_{p}\leq C_p\|g\|_{p}\quad\textrm{where}\quad S^*g(x)=\sup_{n\ge 0} |S_ng(x)|.$$
On $\mathcal C(\mathbb T)$, a weak inequality (also due to Hunt) remains valid (see \cite[Theorem 12.5]{Ar02}): there
are two absolute constants $A,B>0$ such that, for every $f\in\mathcal C(\mathbb T)$ and every $y>0$,
$$\mathcal \lambda\big(\left\{x\in\mathbb T\ ;\ S^* f(x)>y\right\}\big)\leq Ae^{-By/\|f\|_\infty}.$$
Here, $\lambda$ denotes the Lebesgue measure on $\mathbb T$.
So, let $\beta\in(0,1)$ and $f\in\mathcal C(\mathbb T)$. We may assume $\|f\|_\infty\leq 1$. For any $M>0$,
we introduce
$$\mathcal F(\beta,f,M)=\left\{x\in\mathbb T;\ \limsup_{n\to +\infty} \,(\log n)^{-\beta}|S_n f(x)|> M\right\}.$$
Since $\mathcal F(\beta,f)=\bigcup_{M>0} \mathcal F(\beta,f,M)$, we just need to prove that
$\mathcal H^{\phi_{1,\gamma}}\big(\mathcal F(\beta,f,M)\big)=0$ for every $M>0$.
From now on, we fix some $M>0$.
We pick any $x\in\mathcal F(\beta,f,M)$ and $n_x$ large enough such that
$$|S_{n_x}f(x)|\geq M(\log n_x)^\beta.$$
Such an inequality remains true in an interval around $x$ whose size is not so small. Precisely,
because $n_x$ can be assumed to be large and since the $L^1$-norm of the
Dirichlet kernel $D_n$
behaves like $\frac{4}{\pi^2}\log n$, we may assume that
$\|S_{n_x}f\|_\infty\leq (\log n_x)\|f\|_{\infty}\le \log n_x$.
By Bernstein's inequality, $\|(S_{n_x}f)'\|_\infty\leq n_x \log n_x$. Let
$$I_x=\left[x-\frac{M}{2n_x(\log n_x)^{1-\beta}},x+\frac{M}{2n_x(\log
n_x)^{1-\beta}}\right].$$
For any $y\in I_x$, we get
\begin{eqnarray}\label{EQLEMCTMAJO1}
|S_{n_x}f(y)|\geq \frac M2(\log n_x)^\beta.
\end{eqnarray}
$(I_x)_{x\in\mathcal F(\beta,f,M)}$ is a covering of $\mathcal F(\beta,f,M)$.
We can extract a Vitali's covering, namely a countable family of disjoint intervals $I_i$, $i\in\mathbb N$,
of length $\frac M{n_i(\log n_i)^{1-\beta}}$ such that $\mathcal F(\beta,f,M)\subset \bigcup_{i}5I_i$.
Let us finally set, for any $q\geq 1$, $\mathcal U_q=\left\{i;\ 2^{q+1}\geq \frac{M(\log n_i)^\beta}{2}>2^q\right\}$.
Without loss of generality, we may assume the $n_i$ so large that $\bigcup_q \mathcal U_q=\mathbb N$. By applying Hunt's theorem,
$$\mathcal \lambda\left(\left\{x;\ S^* f(x)>2^q\right\}\right)\leq Ae^{-B2^q}.$$
Now, by (\ref{EQLEMCTMAJO1}), the set $\{x;\ S^* f(x)>2^q\}$ contains the disjoint intervals $I_i$, for $i\in\mathcal U_q$.
Thus,
$$\sum_{i\in\mathcal U_q} |I_i|\leq Ae^{-B2^q}.$$
Moreover, for any $i\in\mathcal U_q$, it is not hard to check that
$$|I_i|\geq Ce^{-D2^{q/\beta}}$$
for some positive constants $C,D$ which do not depend on $q$. Picking any
$\alpha$ such that $1-\beta<\alpha<\gamma$,
we get
\begin{eqnarray*}
\sum_{i\in\mathcal U_q}\mathcal \phi_{1,\alpha}(5|I_i|)
&=&\sum_{i\in\mathcal U_q} 5|I_i| \exp\left((\log (1/5|I_i|))^{1-\alpha}\right)\\
&\leq&5\left(\sum_{i\in\mathcal U_q}|I_i|\right)\exp\left(\left(D2^{q/\beta}-\log 5C\right)^{1-\alpha}\right)\\
&\leq&5Ae^{-B2^q+D'2^{q(1-\alpha)/\beta}}.
\end{eqnarray*}
Since $1-\alpha<\beta$, this shows that there exists $C_0<+\infty$ such that
$$\sum_{i\in\mathbb N}\phi_{1,\alpha}(5|I_i|)=\sum_{q\in\mathbb
N}\sum_{i\in\mathcal U_q}\phi_{1,\alpha}(5|I_i|)\le C_0.$$
Remember that $\bigcup_i5I_i$ is a covering of $\mathcal F(\beta,f,M)$ and
that the $I_i$ can be chosen as small as we want. We can then conclude that
$\mathcal H^{\phi_{1,\alpha}}(\mathcal F(\beta,f,M))\le C_0$. In particular, $\mathcal H^{\phi_{1,\gamma}}\big(\mathcal F(\beta,f,M)\big)=0$,
since $\phi_{1,\alpha}\gg\phi_{1,\gamma}$.
\end{proof}
\begin{remark}
The functions $\phi_{1,\gamma}$, for $\gamma>1-\beta$, are not optimal in the statement of the previous lemma.
We can replace them by any function $\phi(x)=x\left(\exp\big((\log 1/x)^\beta\varepsilon(x)\big)\right)$ with
$\varepsilon(x)$ goes to 0 as $x$ goes to 0.
\end{remark}
\subsection{The basic construction}
When we try to build explicitely a continuous function whose Fourier series diverges at some point, say 0, a natural way is to consider
polynomials $P$ with small $L^\infty$ norm, and satisfying nevertheless that
$|S_nP(0)|$ is big for some large value of $n$. The easiest examples are
$$P_N(x)=e_N(x) \sum_{j=1}^N \frac{\sin(2\pi jx)}{j},$$
since the sequence $(\|P_N\|_\infty)_{N\ge 1}$ is bounded whereas $|S_N(P)(0)|\sim \frac12\log N$. Moreover, this example is in some sense optimal
since $\|S_N f\|_\infty\leq C(\log N)\|f\|_{\infty}$ for any $f\in \mathcal C(\mathbb T)$.
In our context, we need to find a polynomial $P$ which satisfies a similar property not only at one point, but on a set which is rather big since at the end we want to construct
sets of divergence with Hausdorff dimension 1. This does not seem to be the case for $P_N$, the reason being that $|(S_NP)'(0)|$ behaves like $N$, which is
much bigger than $S_NP(0)$.
\smallskip
To tackle this problem, we start from a construction of Kahane and Katznelson in \cite{KK66} (see also \cite{Katz}) which
they use to prove
that every subset of $\mathbb T$ of Lebesgue measure 0 is a set of divergence for $\mathcal C(\mathbb T)$. Since we want to control
both the size of the sets $E$ and the index $n$ such that $S_n P(x)$ becomes larger than some given real number for any $x\in E$,
the forthcoming lemma needs very careful estimations.
\begin{lemma}\label{LEMBASICCT}
Let $\beta\in(0,1)$, $\delta\in(0,1)$ and $K\geq 2$. Then there exist an
integer $k\ge K$, an integer $n$ as large as we want and
a trigonometric polynomial $P$ with spectrum contained in $[0,2n-1]$ such that
\begin{itemize}
\item $|P(x)|\leq 1$ for any $x\in\mathbb T$;
\item $\log |S_nP(x)|\geq (1-\delta)\beta\log\log n$ for any $x\in I_k^\beta$,
\end{itemize}
where $\displaystyle I_k^\beta=\bigcup_{j=0}^{k-1}\left[\frac jk-\frac{1}{2k\exp\big((\log k)^\beta\big)};\frac jk+\frac{1}{2k\exp\big((\log k)^\beta\big)}\right]$.
\end{lemma}
\begin{proof}
Let us first describe the idea of the proof. We shall construct a trigonometric polynomial $Q$ with spectrum in $[1,n-1]$ and with the following properties:
$|\Im m\ Q|$ is small and $|Q|$ is large on a set $E$. We then set $P=e_{n}\times \Im m\ Q$, so that $\|P\|_\infty$ is small. On the other hand,
writing $Q=\sum_{k=1}^{n-1} a_k e_k$, $2i\Im m\ Q=-\sum_{k=1}^{n-1}\overline{a_k}e_{-k}+\sum_{k=1}^{n-1} a_k e_k$, so that
$$|S_n (P)|=\frac 12\left| \sum_{k=1}^{n-1} \overline{a_k} e_{n-k}\right|=\frac12\left|\sum_{k=1}^{n-1} a_k e_k\right|=\frac12|Q|$$
is large on $E$. The construction of $Q$ will be done by taking $\log f$, the
logarithm of an holomorphic function defined on a neighbourhood of the closed
unit disk $\overline\mathbb D$ (which allows to control
the imaginary part of $\log f$ while the modulus of it can be large), and by taking a Fej\'er sum of $\log f$.
We now proceed with the details. The proof uses holomorphic functions and it is better to see $\mathbb T$ as the boundary of the unit disk $\mathbb D$.
To avoid cumbersome notations, the letter $C$ will denote throughout the proof a positive and absolute constant, whose value may change from line to line.
Let $k\geq K$ whose value will be fixed later. We set:
\begin{eqnarray*}
\varepsilon&=&\frac1{k\exp\big((\log k)^\beta\big)}\\
z_j&=&e^{\frac{2\pi ij}k},\ j=0,\dots,k-1\\
f(z)&=&\frac1k\sum_{j=0}^{k-1}\frac{1+\varepsilon}{1+\varepsilon-\overline{z_j}z}.
\end{eqnarray*}
$f$ is holomorphic in a neighbourhood of $\overline{\mathbb D}$. We claim that $f$ satisfies the following four properties.
\begin{description}
\item[(P1)]$\forall z\in\overline{\mathbb D}$,\quad $\Re ef(z)\geq C\varepsilon$;
\item[(P2)]$\forall z\in I_k^\beta,\quad |f(z)|\geq\Re ef(z)\geq C\exp\big((\log k)^\beta\big)$;
\item[(P3)]$\forall z\in\mathbb T,\quad |f(z)|\leq C\exp\big((\log k)^\beta\big)$;
\item[(P4)]$\forall z\in\mathbb T,\quad \left|\frac{f'(z)}{f(z)}\right|\leq\frac{C}{\varepsilon^3}$.
\end{description}
Indeed, for any $z\in\overline{\mathbb D}$ and any $j\in\{0,\dots,k-1\}$,
\begin{eqnarray}\label{EQBASIC1}
\Re e\left(\frac{1+\varepsilon}{1+\varepsilon-\overline{z_j}z}\right)=\frac{1+\varepsilon}{|1+\varepsilon-\overline{z_j} z|^2}\Re e\big(1+\varepsilon-z_j\overline{z}\big)\geq\frac{1+\varepsilon}{(2+\varepsilon)^2}\times\varepsilon\geq C\varepsilon,
\end{eqnarray}
which proves \textbf{(P1)}. To prove \textbf{(P2)}, we may assume that $z=e^{2\pi i\theta}$ with
$\theta\in\left[\frac{-\varepsilon}{2};\frac{\varepsilon}{2}\right]$. Then
$$\Re e\left(\frac{1+\varepsilon}{1+\varepsilon-\overline{z_0}z}\right)=\frac{1+\varepsilon}{|1+\varepsilon-z|^2}\Re e\big(1+\varepsilon-\overline{z}\big)\geq\frac C\varepsilon.$$
If we combine this with (\ref{EQBASIC1}), we get
$$\Re e f(z)\ge \frac{C}{k \varepsilon}+\frac{k-1}k C\varepsilon\ge \frac{C}{k \varepsilon}=C\exp\big((\log k)^\beta\big).$$
which gives \textbf{(P2)}.
Conversely, we want to control $\sup_{z\in\mathbb T}|f(z)|$. Pick any $z=e^{2\pi i\theta}\in\mathbb T$.
By symmetry, we may and shall assume that $|\theta|\leq\frac1{2k}$. Then we get
$$\left|\frac{1+\varepsilon}{1+\varepsilon-\overline{z_0}z}\right|\leq\frac{C}{\varepsilon}.$$
Now, for any $j\in\{1,\dots,k/4\}$, we can write
\begin{eqnarray*}
|1+\varepsilon-\overline{z_j}z|&\geq& |\Im m(\overline{z_j}z)|\\
&\geq&\sin\left(\frac{2\pi j}k-2\pi\theta\right)\\
&\geq&\frac2\pi\times2\pi \left(\frac jk-\theta\right)\\
&\geq&\frac{4}k\left(j-\frac12\right).
\end{eqnarray*}
Taking the sum,
$$\left|\sum_{j=1}^{k/4}\frac{1+\varepsilon}{1+\varepsilon-\overline{z_j}z}\right|
\leq \frac{k(1+\varepsilon)}4\sum_{j=1}^{k/4}\frac{1}{j-1/2}\leq Ck\log k.$$
In the same way, we obtain
$$\left|\sum_{j=3k/4}^{k-1}\frac{1+\varepsilon}{1+\varepsilon-\overline{z_j}z}\right|
\leq Ck\log k$$
whereas $|1+\varepsilon-\overline{z_j}z|\geq C$ for any $j\in[k/4,3k/4]$, so that
$$\left|\sum_{j=k/4}^{3k/4}\frac{1+\varepsilon}{1+\varepsilon-\overline{z_j}z}\right|
\leq Ck.$$
Putting this together, we get
$$|f(z)|=\left|\frac1k\sum_{j=0}^{k-1}\frac{1+\varepsilon}{1+\varepsilon-\overline{z_j}z}\right|\leq C\left(\frac{1}{k\varepsilon}+\log k+1\right)\leq C\exp\left((\log k)^\beta\right).$$
Finally, it remains to prove \textbf{(P4)}. We observe that
$$f'(z)=\frac1k\sum_{j=0}^{k-1}\frac{(1+\varepsilon)\overline{z_j}}{(1+\varepsilon-\overline{z_j}z)^2}.$$
We do not try to get a very precise estimate for $|f'(z)|$ (this is not useful for us). We just observe
that $|1+\varepsilon-\overline{z_j}z|^2\geq\varepsilon^2$ for any $j\in\{0,\dots,k-1\}$ and any $z\in\mathbb T$, so that
$$|f'(z)|\leq\frac{C}{\varepsilon^2}.$$
If we combine this with \textbf{(P1)}, we get \textbf{(P4)}.
\medskip
We are almost ready to construct $P$. The next step is to take $h(z)=\log(f(z))$, which defines a holomorphic function
in a neighbourhood of $\overline{\mathbb D}$ by \textbf{(P1)}. Moreover, $|\Im
m(h(z))|\leq\pi/2$ for any $z\in\overline{\mathbb D}$ and $h(0)=0$. Now, we look at
the function $h$ on the boundary of the unit disk $\mathbb D$, that is we introduce
the function $g(x)=h(e^{2i\pi x})$ defined on the circle
$\mathbb T=\mathbb R/\mathbb Z$. Properties
\textbf{(P2)}, \textbf{(P3)} and \textbf{(P4)} can be rewritten as
\begin{eqnarray*}
\forall x\in I_k^\beta,\quad |g(x)|&\geq&(\log k)^\beta-C\\
\forall x\in\mathbb T,\quad |g(x)|&\leq&(\log k)^\beta+C\\
\forall x\in\mathbb T,\quad |g'(x)|&\leq&Ck^3\exp\big(3(\log k)^\beta\big).
\end{eqnarray*}
Let now $n$ be the smallest
integer such that
${C}{k^3\exp\big(3(\log k)^\beta\big)}\leq n$. We also have
$\|g'\|_\infty\leq n$ and we can apply the second part of Lemma \ref{LEMFEJER}
to the function $\theta(t)=g(t)-g(x)$ when $x\in I_k^\beta$. Recall that
$\|\theta\|_\infty\leq 2(\log k)^\beta+C$. We get
$$|\sigma_{n}\theta(x)|\leq \frac{(\log k)^\beta}2+C$$
and we can conclude that
$$|\sigma_{n} g(x)|\geq |g(x)|-|\sigma_{n}\theta(x)|\geq \frac{(\log k)^\beta}2-C.$$
We finally set
$$P=\frac{2}\pi e_{n}\sigma_n(\Im m g)=\frac{2}\pi e_{n}\Im m (\sigma_n g).$$
It is straightforward to check that $\|P\|_\infty\leq 1$ (recall that
$\sigma_n$ is a contraction on $\mathcal C(\mathbb T)$), and that the spectrum of
$\sigma_n g$ is contained
in $[1,n-1]$ ($\hat g(0)=0$ since $h(0)=0$). Now, the simple algebraic trick exposed at the beginning of the proof shows that
$$|S_nP(x)|=\left|\frac{1}\pi \sigma_n g(x)\right|,$$
so that, for any $x\in I_k^\beta$,
$$|S_nP(x)|\geq \frac1{2\pi}(\log k)^\beta-C.$$
This leads to
$$\log |S_nP(x)|\geq\beta\log\log k-C.$$
On the other hand,
\begin{eqnarray*}
\log \log n&\leq&\log\left(3\log k+3(\log k)^\beta+\log C\right)\\
&\leq&\log\log k+C.
\end{eqnarray*}
Finally,
$$\frac{\log\log |S_nP(x)|}{\log\log n}\ge\frac{\beta\log\log k -C}{\log\log
k+C}\ge (1-\delta)\beta,$$
provided $k$ has been chosen large enough. Moreover, $n$ can be chosen as
large as we want since $n\to +\infty$ when $k\to +\infty$.
\end{proof}
\begin{remark}
The fact that we have to compare $\log\log n$ and $\log |S_n|$ helps us for the previous proof. Even if $n$ and $k$ do not have the same order of growth,
this is not apparent when we apply the iterated logarithm.
\end{remark}
\begin{remark}\label{REMARKNK}
During the construction, the integers $k$ and $n$ can't be chosen
independently~: they satisfy $n-1\le Ck^3\exp\big(3(\log k)^\beta\big)\leq n$
where $C$ is an absolute constant. If we want to construct a polynomial $P$
satisfying the conclusion of Lemma \ref{LEMBASICCT} with a large value of $n$,
we need also to choose a large value of $k$.
\end{remark}
\subsection{The conclusion}
We are now going to prove the full statement of Theorem \ref{THMCT3}. At this point, the situation is less favourable than in the
$L^p$-case. There, the basic construction done at each step $j$ did not depend on the index of divergence that we would like to get.
We had the same function $g_j$ which worked for all indices of divergence, and it was the dyadic exponent of $x$ which
decided how large was $|g_j(x)|$.
The construction done in Lemma \ref{LEMBASICCT} is less efficient, because the polynomial $P$ does depend on the expected divergence index $\beta$
(the index $\beta$ is a parameter of the definition of $f$ above). We have to overcome this new difficulty
and the solution will be to introduce redundancy in the construction of the $G_\delta$-set.
As usual, we start from a sequence $(f_j)_{j\ge 1}$ of polynomials which is dense in $\mathcal C(\mathbb T)$. For convenience,
we assume that $\|f_j\|_\infty\leq j$ for any $j$ and that the spectrum of $f_j$ is contained in $[-j,j]$.
Furthermore, we fix four sequences $(\alpha_l),\ (\beta_l),\ (\delta_l)$ and $(\varepsilon_l)$ with values in $(0,1)$ and such that:
\begin{itemize}
\item $(\beta_l)$ is dense in $(0,1)$ and $l\mapsto\beta_l$ is one to one;
\item $\sum_l \varepsilon_l\leq 1$;
\item $(\delta_l)$ and $(\alpha_l)$ go to zero.
\item $\delta_l<1/3$.
\end{itemize}
Let now $j\geq 1$. By induction on $l=1,\dots,j$, we build sequences $(P_{j,l})$, $(n_{j,l})$ and $(k_{j,l})$
satisfying the conclusions of Lemma \ref{LEMBASICCT} with $\beta=\beta_l$,
$\delta=\delta_l$ and $K=j$ (to ensure that $\lim_{j\to
+\infty}k_{j,l}=+\infty$) and we will decide how large should be $n_{j,l}$ during
the construction. According to Remark \ref{REMARKNK}, these constraints on
$n_{j,l}$ will
determine the values of the $k_{j,l}$. We then set
$$g_j:=f_j+\alpha_j\sum_{l=1}^j \varepsilon_l e_{n_{j,l}}P_{j,l}$$
so that $\|g_j-f_j\|_\infty\leq \alpha_j\sum_{l=1}^j \varepsilon_l \|P_{j,l}\|_\infty\leq \alpha_j$. In particular, the sequence
$(g_j)$ remains dense in $\mathcal C(\mathbb T)$.
Recall that the spectrum of $f_j$ is included in $[-j,j]$ and observe that
the spectrum of $e_{n_{j,l}}P_{j,l}$ lies in $[n_{j,l},3n_{j,l}-1]$. If we
suppose that $n_{j,1}=j+1$ and $n_{j,l+1}\ge 3n_{j,l}$, we can conclude that
the spectra of $f_j,e_{n_{j,1}}P_{j,1},\cdots,e_{n_{j,j}}P_{j,j}$ are disjoint.
Let now $x$ belongs to $I_{k_{j,l}}^{\beta_{l}}$ for some $l\leq j$. Then
\begin{eqnarray*}
\left|S_{2n_{j,l}}g_j(x)\right|&\geq&\alpha_j \varepsilon_l \left|S_{n_{j,l}}P_{j,l}(x)\right|-\alpha_j\sum_{m=1}^{l-1}\varepsilon_m \|P_{j,m}\|_\infty-j\\
&\geq&\alpha_j\varepsilon_l \left|S_{n_{j,l}}P_{j,l}(x)\right|-\alpha_j-j.\\
\end{eqnarray*}
Because we can choose $n_{j,l}$ as large as we want in the process,
we may always assume that the choice that we have done ensures that
$$\left|S_{2n_{j,l}}g_j(x)\right|\geq\frac{\alpha_j\varepsilon_l}2 \left|S_{n_{j,l}}P_{j,l}(x)\right|.$$
Taking the logarithm, we find
\begin{eqnarray*}
\log \left|S_{2n_{j,l}}g_j(x)\right|&\geq&\log \left|S_{n_{j,l}}P_{j,l}(x)\right|+\log \varepsilon_l+\log \alpha_j-\log 2\\
&\geq&(1-\delta_l)\beta_l \log\log(n_{j,l})+\log \varepsilon_l+\log\alpha_j-\log 2\\
&\geq&(1-2\delta_l)\beta_l \log\log(2n_{j,l})
\end{eqnarray*}
provided again that we have chosen $n_{j,l}$ very large.
We then fix $r_j>0$ so small that, for any $f\in B(g_j,r_j)$
(the balls are related to the norm $\|\cdot\|_\infty$), for any $l\leq j$,
$$\|S_{2n_{j,l}}f-S_{2n_{j,l}}g_j\|_\infty\le 1/2.$$
Observe that for every
real number $t \ge 1$, we have
$\log(t-1/2)\ge \log(t)-\log 2$.
For any $x\in I_{k_{j,l}}^{\beta_{l}}$ with $l\le j$, we get
\begin{eqnarray*}
\log \left|S_{2n_{j,l}}f(x)\right|&\ge&\log
\left|S_{2n_{j,l}}g_j(x)\right|-\log 2\cr
&\ge&(1-2\delta_l)\beta_l\log \log (2n_{j,l})-\log2\cr
&\ge&(1-3\delta_l)\beta_l\log \log (2n_{j,l})
\end{eqnarray*}
if $n_{j,l}$ are chosen sufficiently large such that $\delta_l\beta_l\log \log
(2n_{j,l})\ge\log 2$.
We finally set
$$A=\bigcap_{p\in\mathbb N}\bigcup_{j\geq p}B(g_j,r_j),$$
and we claim that $A$
is the dense $G_\delta$ set we are looking for.
Indeed, let $f$ belong to $A$ and let $(j_p)$ be an increasing sequence
of integers such that for every $p\ge 0$, $f\in B(g_{j_p},r_{j_p})$. We consider
$\beta\in(0,1)$ and choose $p_0$ such that
$$\left\{\beta_1,\cdots,\beta_{j_{p_0}}\right\}\cap\, (0,\beta)\not=\emptyset .$$
Such a $p_0$ exists since the sequence $(\beta_l)_{l\ge 1}$ is dense in
$(0,1)$. For every $p\ge p_0$, let $l_p$ be chosen in $\{1,\cdots, j_p\}$
such that
$$\beta-\beta_{l_p}=\inf\{\beta-\beta_l;\ l\leq j_p\textrm{ and }\beta >\beta_l\}.$$
Since the sequence $(\beta_l)$ is dense in $(0,1)$, $\beta_{l_p}<\beta$ for $p\geq p_0$
and $\beta_{l_p}\to \beta$. Moreover, since $l\mapsto \beta_l$ is one to one,
it is clear that $l_p$ is non decreasing and goes to $+\infty$.
Observe that, for $p\geq p_0$, $I_{k_{j_p,l_p}}^{\beta}\subset I_{k_{j_p,l_p}}^{\beta_{l_p}}$, so that,
for any $x\in I_{k_{j_p,l_p}}^{\beta}$, setting $N_p=2n_{j_p,l_p}$,
$$\log |S_{N_p}f(x)|\geq (1-3\delta_{l_p})\beta_{l_p}\log\log(N_p).$$
In particular, setting $F=\limsup_p I_{k_{j_p,l_p}}^\beta$, we get that
$$\limsup_{n\to+\infty}\frac{\log |S_n f(x)|}{\log\log n}\geq \beta$$
for any $x\in F$. Now, we can apply Corollary \ref{CORUBI2} with a jauge
function $\phi$ satisfying
$\phi^{-1}(y)=y\exp\left[-(\log(1/2y))^{\beta}\right]$ to obtain ${\mathcal
H}^\phi(F)=\infty$.
Observe that if $y=\phi(x)$, then
$$y=x\exp\left[(\log(1/2y))^{\beta}\right]
\quad\text{and}\quad \log x\le \log y.$$
It follows that $\phi(x)\le
x\exp\left[(\log(1/2x))^{\beta}\right]\le\phi_{1,1-\beta}(x)$ and
${\mathcal H}^{\phi_{1,1-\beta}}(F)=+\infty$.
We now conclude exactly as in the $L^p$-case, using Lemma \ref{LEMCTMAJO} to replace Aubry's result. Namely, we set
\begin{eqnarray*}
F^1&=&\left\{x\in F;\ \limsup_{n\to+\infty}\frac{\log |S_n f(x)|}{\log\log n}= \beta\right\}\\
F^2&=&\left\{x\in F;\ \limsup_{n\to+\infty}\frac{\log |S_n f(x)|}{\log\log n}> \beta\right\}
\end{eqnarray*}
and we observe that Lemma \ref{LEMCTMAJO} guarantees that ${\mathcal H}^{\phi_{1,1-\beta}}(F^2)=0$. Thus,
${\mathcal H}^{\phi_{1,1-\beta}}(F^1)=+\infty$ and the precised Hausdorff dimension of $F(\beta,f)$, which contains $F^1$, is at least $(1,1-\beta)$.
By Lemma \ref{LEMCTMAJO}, it is exactly $(1,1-\beta)$.
\begin{remark}
It is amazing that, with our method, it is easier to prove Theorem \ref{THMCT3} and to deduce Theorem
\ref{THMCT2} from it than to prove Theorem \ref{THMCT2} directly. Indeed, to ensure that the sets
$\mathcal F(\beta,f)$ are big, we need to know that the sets $F(\beta',f)$ are small for $\beta'>\beta$. This cannot be done if we stay within the notion of Hausdorff dimension.
\end{remark}
\begin{remark}
The method developed above allows us to construct a
``concrete function'' that satisfies the conclusion of Theorem \ref{THMCT3}.
More precisely, it suffices to consider
$$g=\sum_{j=1}^{+\infty}\frac1{j^2}\sum_{l=1}^j\varepsilon_le_{n_{j,l}}P_{j,l}$$
with the constraint $3n_{j,j}<n_{j+1,1}$ to ensure that the blocks
$\sum_{l=1}^j\varepsilon_le_{n_{j,l}}P_{j,l}$ have disjoint spectra. Such a function
is some kind of saturating function in the continuous case.
\end{remark}
|
\section{Introduction}
This work pertains to the probabilistic study of Euclidean combinatorial optimization
problems. The starting point in this field is the celebrated theorem of
Beardwood, Halton and
Hammersley \cite{BHH} about the traveling salesperson problem. Its ensures that given
a sequence $(X_i)_{i\ge 1}$ of independent random variables on $\dR^d$, $d\ge 2$ with
common law $\mu$ of bounded support, then almost surely
$$ \lim_{n\to\infty} n^{\frac1d-1}T(X_1,\ldots, X_n)=\beta_d \int f^{1-\frac1d}.$$
Here $\beta_d$ is a constant depending only on the dimension, $f$ is the density of the
absolutely continuous part of $\mu$ and
$$T(X_1,\ldots, X_n)=\inf_{\sigma\in \mathcal S_n} \sum_{i=1}^{n-1} |X_{\sigma(i+1)}-X_{\sigma(i)}|+
|X_{\sigma(1)}-X_{\sigma(n)}|$$
is the length (for the canonical Euclidean distance) of the shorstest tour through the points
$X_1,\ldots, X_n$. In the above formula $\mathcal S_n$ stands for the set of permutations of
$\{1,2,\ldots,n\}$.
Very informally, this result supports the following interpretation:
when the number of points $n$ is large, for $\mu$ almost every $x$, if the salesperson is at $X_i=x$
then the distance
to the next point in the optimal tour is
comparable to $ \beta(d) (nf(x))^{-1/d}$ if $f(x)>0$ and of lower order otherwise.
This should be compared to the fact that the distance from $X_i=x$ to $\{X_j, j\le n \mbox{ and } j \neq i\}$
also stabilizes at the same rate.
Later, Papadimitriou
\cite{P78} and Steele \cite{S81} have initiated a general theory of
Euclidean functionals $F(\{X_1,\ldots,X_n\})$ that satisfy almost sure
limits of this type. We refer the reader to the monographs of Steele
\cite{S} and Yukich \cite{Y} for a full treatment of this now mature
theory, and present a short outline. It is convenient to consider
multisets rather than sets, so throughout the paper
$\{x_1,\ldots,x_n\}$ will stand for a multiset (the elements are
unordered but may be repeated). The umbrella theorem in \cite{Y} puts forward
the following three features of a functional $F$ on finite multisets of $\dR^d$:
\begin{itemize}
\item $F$ is $1$-homogeneous if it is translation invariant and dilation covariant:
$$ F(a+\lambda \mathcal{X})=\lambda F(\mathcal{X})$$
for all finite multisets $\mathcal{X}$, all $a\in \dR^d$ and $\lambda\in \dR^+$.
\item The key assumption is subadditivity: $F$ is subadditive if there exists a constant
$C>0$ such that for all multisets $\mathcal{X},\mathcal{Y}$ in the unit cube $[0,1]^d$,
$$ F(\mathcal{X}\cup\mathcal{Y})\le F(\mathcal{X})+F(\mathcal{Y})+C.$$
As noted by Rhee in \cite{rhee1}, this assumption implies that there is another constant $C'$ such that
for all multiset in $[0,1]^d$,
\begin{equation}\label{eq:good-bound}
|F( \mathcal{X}) | \leq C' \left( \mathrm{card}( \mathcal{X}) \right)^{1- \frac{1 }{d} }.
\end{equation}
Hence the worst case for $n$ points is at most in $n^{1-\frac1d}$ and the above mentioned theorems
show that the average case is of the same order.
\item The third important property is smoothness (or regularity). A functional $F$ on finite multisets
$\dR^d$ is smooth if there is a constant $C''$ such that for all multisets $\mathcal{X},\mathcal{Y},\mathcal Z$ in
$[0,1]^d$, it holds
$$| F(\mathcal{X}\cup \mathcal Y)-F(\mathcal{X}\cup \mathcal Z)| \le C'' \left( \mathrm{card}( \mathcal{Y})^{1- \frac{1 }{d}} +
\mathrm{card}( \mathcal Z)^{1- \frac{1 }{d}} \right).$$
\end{itemize}
These three properties are enough to show upper limits for $F$, on the model of the Beardwood, Halton, Hammersley
theorem. To have the full limits, the umbrella theorem of \cite{Y} also requires to check a few more
properties of a so-called boundary functional associated with $F$.
Next, let us present a classical optimization problem which does not enter the above picture.
Given two multi-subsets of $\dR^d$
with the same cardinality, $\mathcal X=\{X_1,\ldots,X_n\}$ and
$\mathcal Y=\{Y_1,\ldots,Y_n\}$, the cost of the minimal bipartite
matching of $\mathcal X$ and $\mathcal Y$ is defined as
$$ M_1(\mathcal X,\mathcal Y)=\min_{\sigma\in \mathcal S_n} \sum_{i=1}^{n} |X_i-Y_{\sigma(i)}|,$$
where the minimum runs over all permutations of $\{1,\ldots,n\}$.
It is well-known that $n^{-1}
M_1\big(\{X_i\}_{i=1}^n, \{Y_i\}_{i=1}^n\big)$ coincides with the
power of the $L_1$-Wasserstein distance between the empirical distributions
$$ W_1\Big(\frac1n \sum_i \delta_{X_i},\frac1n \sum_i \delta_{Y_i}\Big),$$
hence it is easily seen to tend to $0$, for example when $\mu$ has
bounded support. Recall that given two finite measures $\mu_1$, $\mu_1$ on $\mathbb R^d$
with the same total mass,
$$ W_1(\mu_1,\mu_2)= \inf_{\pi\in\Pi(\mu_1,\mu_2)} \int_{\dR^d\times \dR^d} |x-y| \, d\pi(x,y) ,$$
where $\Pi(\mu_1,\mu_2)$ is the set of measures on $(\mathbb R^d)^2$ having $\mu_1$ as first marginal
and $\mu_2$ as second marginal (see e.g. \cite{rachev,villani} for more background).
Note that for all
finite multisets $\mathcal{X}$, $\mathcal{Y}$ in $[0,1]^d$ with $\mathrm{card}(\mathcal{X}) =
\mathrm{card}(\mathcal{Y})$,
$$
M_1( \mathcal{X}, \mathcal{Y}) \leq \sqrt d \, \mathrm{card}( \mathcal{X}),
$$
and equality holds for some well-chosen configurations of any cardinal
(all elements in $\mathcal{X}$ at $(0,\cdots,0)$ and all elements in $\mathcal{Y}$ at
$(1,\cdots,1)$). Hence, an interesting feature of $L$ (as well as
others bipartite Euclidean optimization functionals) is that the
growth bound assumption \eqref{eq:good-bound} fails, hence it is not subadditive in the above sense.
However Dobri\'c and Yukich have stated the following theorem:
\begin{theorem}[\cite{DY}] \label{th:main} Let $d\ge 3$ be an integer. Assume that $\mu$ is a probability measure on $\mathbb R^d$ having a bounded support.
Consider mutually independent random variables $(X_i)_{i\ge 1}$ and $(Y_j)_{j\ge 1}$ having distribution $\mu$.
Then, almost surely,
$$ \lim_{n} n^{\frac{1}{d}-1} M_1\big(\{X_1,\ldots,X_n\}, \{Y_1,\ldots,Y_n\}\big)
= \beta_1(d) \int _{\dR^d} f^{\frac{d-1}{d}},$$
where $f(x)\, dx$ is the absolutely continuous part of $\mu$ and $\beta_1(d)$ is a constant depending only on the dimension $d$.
\end{theorem}
When $f$ is not the uniform measure on the unit cube, there is an issue in the proof of \cite{DY} that apparently cannot be easily fixed (the problem lies in their Lemma 4.2 which is used for proving that the $\liminf$ is at least $ \beta_1(d) \int _{\dR^d} f^{\frac{d-1}{d}}$). In any case, the proof of Dobri\'c and Yukich is very specific to the bipartite matching as it uses from
the start the Kantorovich-Rubinstein
dual representation of the optimal transportation cost. It is not adapted to
a general treatment of bipartite functionals. The starting point of our work was recent paper
of Boutet de Monvel and Martin \cite{BM} which (independently of \cite{DY}) establishes the
convergence of the bipartite matching for uniform variables on the unit cube, without using
the dual formulation of the transportation cost. Building on their approach we are able to
propose a soft approach of bipartite functionals, based on appropriate notions of subadditivity
and regularity. These properties allow to establish upper estimates on upper limits. In order to
deal with lower limits we adapt to the bipartite setting the ideas of boundary functionals
exposed in \cite{Y}. We are able to explicitly construct such functionals for a class of optimization
problems involving families of graphs with good properties, and to establish full convergence for
absolutely continuous laws. Finally we introduce a new notion of inverse subadditivity which allows
to deal with singular parts.
This viewpoint sheds a new light on the result of Dobri\'c and Yukich, that we
extend in other respects, by considering power distance costs, and unbounded random variables
satisfying certain tail assumptions. Note that in the classical theory of Euclidean functionals, the analogous question for unbounded random variables was answered
in Rhee \cite{rhee} and generalized in \cite{Y}.
Let us illustrate our results in the case of the bipartite matching with power distance cost : given $p >0$ and two multi-subsets of $\dR^d$, $\mathcal X=\{X_1,\ldots,X_n\}$ and
$\mathcal Y=\{Y_1,\ldots,Y_n\}$, define
$$ M_p(\mathcal X,\mathcal Y)=\min_{\sigma\in \mathcal S_n} \sum_{i=1}^{n} |X_i-Y_{\sigma(i)}|^p,$$
where the minimum runs over all permutations of $\{1,\ldots,n\}$. Note that we have
the same result for the bipartite travelling salesperson problem, and that our generic approach puts forward
key properties that allow to establish similar facts for other functionals.
As mentioned in the title, our results apply to relatively high dimension. More precisely, if the length of
edges are counted to a power $p$, our study applies to dimensions $d>2p$ only.
\begin{theorem} \label{th:mainMp} Let $0 < 2p < d$. Let $\mu$ be a probability measure on $\mathbb R^d$ with absolutely continuous part $f(x)\, dx$. We assume that for some $\alpha > \frac{4dp}{ d -2p}$,
$$
\int |x|^{\alpha} d \mu (x) < +\infty.
$$
Consider mutually independent random variables $(X_i)_{i\ge 1}$ and $(Y_j)_{j\ge 1}$ having distribution $\mu$. Then
there are positive constants $\beta_p(d),\beta'_p(d) $ depending only on $(p,d)$ such that
the following convergence holds almost surely
\begin{eqnarray*}
\limsup_{n} n^{\frac{p}{d}-1} M_p \big(\{X_1,\ldots,X_n\}, \{Y_1,\ldots,Y_n\}\big)
& \leq &\beta_p(d) \int _{\dR^d} f^{1-\frac{p}{d}}, \\
\liminf_{n} n^{\frac{p}{d}-1} M_p \big(\{X_1,\ldots,X_n\}, \{Y_1,\ldots,Y_n\}\big)
& \geq & \beta'_p(d) \int _{\dR^d} f^{1-\frac{p}{d}}. \nonumber
\end{eqnarray*}
Moreover
$$
\lim_{n} n^{\frac{p}{d}-1} M_p \big(\{X_1,\ldots,X_n\}, \{Y_1,\ldots,Y_n\}\big)
= \beta_p(d) \int _{\dR^d} f^{1-\frac{p}{d}}
$$
provided one of the following hypothesis is verified:
\begin{itemize}
\item $\mu$ is the uniform distribution over a bounded set $\Omega\subset \dR^d$ with positive Lebesgue measure.
\item $d\in\{1,2\}$, $p\in(0,d/2)$ or $d\ge 3$, $p\in (0,1]$, and $f$ is up to a multiplicative constant the indicator function over a bounded set $\Omega\subset \dR^d$ with positive Lebesgue measure.
\end{itemize}
\end{theorem}
Our constant $\beta'(d)$ has an explicit expression in terms of the cost of an optimal boundary matching for the uniform measure on $[0,1]^d$ (see Lemma \ref{lem:liminfBM}). We strongly suspect that $\beta_p(d) = \beta'_p(d) $ but we have not been able to solve this important issue. Also, assuming only $\alpha>\frac{2dp}{ d -2p}$, we can establish convergence in probability.
A basic concentration inequality implies that if $\mu$ has bounded support the convergence holds also in $L^q$ for all $q \geq 1$.
The paper is organized as follows: Section~\ref{sec:general} presents the key properties for
bipartite functionals (homogeneity, subadditivity and regularity) and gathers
useful preliminary statements. Section~\ref{sec:cube} establishes the convergence
for uniform samples on the cube. Section \ref{sec:upper} proves upper bounds on the upper
limits. These two sections essentially rely on classical subadditive methods, nevertheless
a careful analysis is needed to control the differences of cardinalities of the two
samples in small domains.
In Section \ref{sec:ex}, we introduce some examples of bipartite functionals. The lower limits are harder to prove and require a new notion of penalized boundary functionals.
It is however difficult to build an abstract theory there, so in Section~\ref{sec:lower}, we will first present the proof for bipartite
matchings with power distance cost, and put forward a few lemmas which will be useful for other
functionals. We then check that for a natural family of Euclidean combinatorial optimization functionals defined in \S \ref{subsec:combopt}, the lower limit also holds. This family includes the bipartite traveling salesman tour. Finally, Section~\ref{sec:final} mentions possible variants and extensions.
\section{A general setting} \label{sec:general}
Let $\mathcal M_d$ be the set of all finite multisets contained in $\dR^d$. We consider a bipartite
functional:
$$L:\mathcal M_d\times \mathcal M_d\to \dR^+.$$
Let $p>0$. We shall say that $L$ is $p$-homogeneous if for all multisets $\mathcal X,\mathcal Y$, all $a\in \dR^d$
and all $\lambda \ge 0$,
\begin{equation*}\label{eq:Hp}
L(a+\lambda \mathcal X, a+\lambda \mathcal Y)=\lambda^p L(\mathcal X,\mathcal Y). \tag{$\mathcal H_p$}
\end{equation*}
Here $a+\lambda\{x_1,\ldots,x_k\}$ is by definition $\{a+\lambda x_1,\ldots, a+\lambda x_k\}$.
For shortness, we call the above property $(\mathcal H_p)$. Note that a direct consequence is that $L(\emptyset,\emptyset)=0$.
The functional $L$ satisfies the regularity property $(\mathcal R_p)$ if there exists a number $C$ such that for all multisets $\mathcal X,\mathcal Y,
\mathcal X_1,\mathcal Y_1,\mathcal X_2,\mathcal Y_2$, denoting by $\Delta$ the diameter of their union, the following
inequality holds
\begin{equation*}\label{eq:Rp}
L(\mathcal X\cup \mathcal{X}_1,\mathcal{Y} \cup\mathcal{Y}_1)\le L(\mathcal X\cup \mathcal{X}_2,\mathcal{Y} \cup\mathcal{Y}_2)+C \Delta^p \big( \mathrm{card}(\mathcal{X}_1)+
\mathrm{card}(\mathcal{X}_2)+\mathrm{card}(\mathcal{Y}_1)+\mathrm{card}(\mathcal{Y}_2)\big).\tag{$\mathcal R_p$}
\end{equation*}
The above inequality implies in particular an easy size bound: $L(\mathcal{X},\mathcal{Y})\le C \Delta^p (\mathrm{card}(\mathcal{X})+
\mathrm{card}(\mathcal{Y}))$ when $L(\emptyset,\emptyset)=0$.
Eventually, $L$ verifies the subbaditivity property $(\mathcal S_p)$ if there exists a number $C$ such that for every
$k\ge 2$ and all multisets $(\mathcal{X}_i,\mathcal{Y}_i)_{i=1}^k$, denoting by $\Delta$ the diameter of their union, the following
inequality holds
\begin{equation*}\label{eq:Sp}
L\Big( \bigcup_{i=1}^k \mathcal{X}_i, \bigcup_{i=1}^k \mathcal{Y}_i\Big) \le \sum_{i=1}^k L(\mathcal{X}_i,\mathcal{Y}_i)+ C\Delta^p \sum_{i=1}^k
\Big( 1+\big|\mathrm{card}(\mathcal{X}_i)-\mathrm{card}(\mathcal{Y}_i)\big|\Big).\tag{$\mathcal S_p$}
\end{equation*}
\begin{remark}
A less demanding notion of "geometric subadditivity" could be introduced by requiring the above
inequality only when the multisets $\mathcal{X}_i\cup\mathcal{Y}_i$ lie in disjoint parallelepipeds (see \cite{Y}
where such a notion is used in order to encompass more complicated single sample functionals). It is
clear from the proofs that some of our results hold assuming only geometric subadditivity (upper limit
for bounded absolutely continuous laws for example). We will not push this idea further in this paper.
\end{remark}
We will see later on that suitable extensions of the bipartite matching, of the bipartite traveling salesperson
problem, and of the minimal bipartite spanning tree with bounded maximal degree satisfy all these properties. Our main generic result on bipartite functionals is the following.
\begin{theorem}
\label{th:mainup}
Let $d>2p>0$ and let $L$ be a bipartite functional on $\dR^d$ with the properties $(\mathcal H_p)$, $(\mathcal R_p)$ and $(\mathcal S_p)$.
Consider a probability measure $\mu$ on $\mathbb R^d$ such that there exists $\alpha>\frac{4dp}{d-2p}$ with
$$\int |x|^\alpha d\mu(x)<+\infty.$$ Consider mutually independent random variables $(X_i)_{i\ge 1}$ and $(Y_j)_{j\ge 1}$ having distribution $\mu$.
Let $f$ be a density function for the absolutely continuous part of $\mu$,
then, almost surely,
\begin{equation*}
\limsup_{n\to \infty} \frac{L(\{X_1 , \cdots, X_n\}, \{Y_1, \cdots, Y_n\} )}{n^{1-\frac{p}{d}}}\le \beta_L \int f^{1-\frac{p}{d}} ,
\end{equation*}
for some constant $\beta_L$ depending only on $L$. Moreover, if $\mu$ is the uniform distribution over a bounded set $\Omega$ with positive Lebesgue measure, then there is equality: almost surely,
\begin{equation*}
\lim_{n\to \infty} \frac{L(\{X_1 , \cdots, X_n\}, \{Y_1, \cdots, Y_n\} )}{n^{1-\frac{p}{d}}} = \beta_L \mathrm{Vol}(\Omega)^{\frac p d}.
\end{equation*}
\end{theorem}
Beyond uniform distributions, lower limits are harder to obtain. In Section~\ref{sec:lower}, we will state a matching lower bound for a subclass of bipartite functionals which satisfy the properties $(\mathcal H_p)$, $(\mathcal R_p)$ and $(\mathcal S_p)$ (see the forthcoming Theorem \ref{th:Llower} and, for the bipartite traveling salesperson tour, Theorem \ref{th:bTSP}).
\begin{remark}
Let $B(1/2) = \{x \in \dR^d: |x|\leq 1/2 \}$ be the Euclidean ball of radius $1/2$ centered at the origin. It is immediate that the functional $L$ satisfies the regularity property $(\mathcal R_p)$ if it satisfies property $(\mathcal H_p)$ and if for all multisets $\mathcal X,\mathcal Y,
\mathcal X_1,\mathcal Y_1,\mathcal X_2,\mathcal Y_2$ in $B(1/2)$,
\begin{equation*}\label{eq:R}
L(\mathcal X\cup \mathcal{X}_1,\mathcal{Y} \cup\mathcal{Y}_1)\le L(\mathcal X\cup \mathcal{X}_2,\mathcal{Y} \cup\mathcal{Y}_2)+C \big( \mathrm{card}(\mathcal{X}_1)+
\mathrm{card}(\mathcal{X}_2)+\mathrm{card}(\mathcal{Y}_1)+\mathrm{card}(\mathcal{Y}_2)\big). \tag{$\mathcal R$}
\end{equation*}
Similarly, $L$ will enjoy the subbaditivity property $(\mathcal S_p)$ if it satisfies property $(\mathcal H_p)$ and if for every
$k\ge 2$ and all multisets $(\mathcal{X}_i,\mathcal{Y}_i)_{i=1}^k$ in $B(1/2)$,
\begin{equation*}\label{eq:S}
L\Big( \bigcup_{i=1}^k \mathcal{X}_i, \bigcup_{i=1}^k \mathcal{Y}_i\Big) \le \sum_{i=1}^k L(\mathcal{X}_i,\mathcal{Y}_i)+ C \sum_{i=1}^k
\Big( 1+\big|\mathrm{card}(\mathcal{X}_i)-\mathrm{card}(\mathcal{Y}_i)\big|\Big). \tag{$\mathcal S$} \end{equation*}
The set of assumptions $(\mathcal H_p)$, $(\mathcal R_p)$, $(\mathcal S_p)$ is thus equivalent to the set of assumptions $(\mathcal H_p)$, $(\mathcal R)$, $(\mathcal S)$.
\end{remark}
\subsection{Consequences of regularity}
\subsubsection{Poissonization}
For technical reasons, it is convenient to consider the poissonized
version of the above problem.
Let $(X_i)_{i\ge 1}, (Y_i)_{i\ge 1}$ be mutually independent variables with distribution $\mu$.
Considering independent variables $N_1$,
$N_2$ with Poisson distribution $\mathcal P(n)$, the randoms sets
$\{X_1,\ldots, X_{N_1}\}$ and $\{Y_1,\ldots, Y_{N_2}\}$ are
independent Poisson point processes with intensity measures $n\mu$.
For shortness, we set
$$ L(n\mu):= L\big(\{X_1,\ldots, X_{N_1}\},\{Y_1,\ldots, Y_{N_2}\}\big).$$
When $d\mu(x)=f(x)\, dx$ we write $L(nf)$ instead of $L(n\mu)$.
Note that whenever we are dealing with Poisson processes, $n\in
(0,+\infty)$ is not necessarily an integer. More generally $L(\nu)$
makes sense for any finite measure, as the value of the functional $L$ for
two independent Poisson point processes with intensity $\nu$.
Assume for a moment that the measure $\mu$ has a bounded support, of diameter $\Delta$.
The regularity property ensures that
$$ \left| L(\{X_1,\ldots,X_n\} ,\{Y_1,\ldots,Y_n\} )-L(\{X_1,\ldots,X_{N_1}\} ,\{Y_1,\ldots,Y_{N_2}\} )\right|
\le C \Delta^p \big( |N_1-n|+|N_2-n|\big).$$
Note that $\E |N_i-n|\le \big(\E(N_i-n)^2\big)^{1/2}= \mathrm{Var}(N_i)=\sqrt n$. Hence the difference between
$ \E L(\{X_i\}_{i=1}^n,\{Y_i\}_{i=1}^n)$ and $\E L(n\mu)$ is at most a constant times $\sqrt n =o(n^{1-p/d})$ when
$d>2p$. Hence in this case, the original quantity and the poissonized version are the same in average at the relevent scale $n^{1-p/d}$.
The boundedness assumption can actually be relaxed. To show this, we need a lemma.
\begin{lemma}
\label{lem:ET}
Let $\alpha > 0$, $n > 0$ and let $\mu $ be a probability measure on $\dR^d$ such that for all $t>0$,
$\mu\big(\{x;\; |x| \geq t\}\big) \le c\, t ^{-\alpha}.$ Let $\mathcal{X}$, $\mathcal{Y}$ be two independent Poisson point processes of intensity $n \mu$ and $T_n = \max \{ | Z| : Z \in \mathcal{X} \cup \mathcal{Y} \}$. Then, for all $0 < \gamma < \alpha$ there
exists a constant $K=K(c,\alpha,\gamma)$ such that for all $n\ge 1$,
$$ \E [T_n^\gamma ]^{\frac1\gamma}\le K n^{ \frac{1}{\alpha}}.
$$
Moreover the same conclusion holds if $\mathcal{X} = \{X_1,\ldots,X_n\}$, $\mathcal{Y} = \{Y_1,\ldots,Y_n\}$ are two mutually independent sequences of $n$ variables with distribution $\mu$.
\end{lemma}
\begin{proof}
For $t \geq 0$, let $A_t = \{ x \in \dR^d : |x| \geq t \}$ and $g(t) = \int_{ A_t } d\mu$. By assumption, $\mu(A_t) \leq c t^{-\alpha}$. We start with the Poisson case.
Since $\mathcal{X}$, $\mathcal{Y}$ are independent, we have $\dP ( T_n < t) = \dP ( \mathcal{X} \cap A_t = \emptyset ) ^2=e^{-2n\mu(A_t)}$.
Therefore, using $1-e^{-u}\le \min(1,u)$,
\begin{eqnarray*}
\E [T_n^\gamma] & = & \gamma \int_0 ^\infty t^{\gamma -1} \dP ( T_n \geq t) dt \\
&=& \gamma \int_0 ^\infty t^{\gamma -1} (1 - e^{- 2 n \mu(A_t) }) dt \\
& \leq & \gamma \int_0 ^{n ^{1/\alpha}} t^{\gamma -1} dt + \int_{n ^{1/\alpha}}^\infty 2 n c t^{\gamma - \alpha-1} dt \\
& =& n ^{\gamma /\alpha} + \frac{2c }{\alpha -\gamma} n^{\gamma/\alpha},
\end{eqnarray*}
For the second case, since $\dP ( T_n\ge t)=1-(1-\mu(A_t))^{2n}\le \min(1, 2n\mu(A_t))$ the same conclusion holds.
\end{proof}
\begin{proposition}\label{prop:poisson}
Let $d>2p>0$.
Let $\mu$ be a probability measure on $\dR^d$ such that $\int |x|^\alpha \, d\mu(x)<+\infty$ for some $\alpha> \frac{2dp}{d-2p}$.
Let $(X_i)_{i\ge 1}, (Y_i)_{i\ge 1}$ be mutually independent variables with distribution $\mu$.
If $L$ satisfies the regularity property $(\mathcal R_p)$ then
$$ \lim_{n\to \infty}\frac{\E L(\{X_i\}_{i=1}^n,\{Y_i\}_{i=1}^n)-\E L(n\mu)}{n^{1-\frac{p}{d}}}=0.$$
\end{proposition}
\begin{remark}
We have not yet proved the finiteness of the above integrals. This will be done later.
The proof below show that the expectations are finite at the same time. So the above statement is
established with the convention $\infty-\infty=0$.
\end{remark}
\begin{proof}
Let
$N_1$ and $N_2$ be Poisson random variables with mean value $n$. Let $T = \max \{ | Z| : Z \in \{X_1, \cdots, X_{N_1}\} \cup \{Y_1, \cdots, Y_{N_2}\} \}$ and $S= \max \{ | Z| : Z \in \{X_{1} , \cdots, X_{n}\} \cup \{Y_{1}, \cdots, Y_{n}\} \}$, with the convention that the maximum over an empty set is $0$. The regularity property ensures
that
$$
\left| L\big(\{X_1,\ldots,X_n\}, \{Y_1,\ldots,Y_n\}\big) -
L\big(\{X_1,\ldots,X_{N_1}\}, \{Y_1,\ldots,Y_{N_2} \}\big) \right|
\leq C( T +S)^p \left(|N_1 - n|+|N_2 - n |\right).
$$
Taking expectation gives, using Cauchy-Schwarz inequality and the bound $(a+b)^q \le \max(1,2^{q-1}) (a^q+b^q)$ valid
for $a,b,q>0$
\begin{align*}
& \left|\E L\big(\{X_1,\ldots,X_n\}, \{Y_1,\ldots,Y_n\}\big) -
L\big(\{X_1,\ldots,X_{N_1}\}, \{Y_1,\ldots,Y_{N_2} \}\big) \right| \\
& \leq c_p \Big(\E [ T^{2p}]+\E [ S^{2p}]\Big)^{\frac12} \Big(\E[|N_1 - n|^2]+\E[|N_2 - n|^2]\Big)^{\frac12} \\
& =c_p \sqrt{2n} \Big(\E [ T^{2p}]+\E [ S^{2p}]\Big)^{\frac12}
\end{align*}
Since $\alpha > 2p$, by Lemma \ref{lem:ET}, for some $c >0$ and all $n \geq 1$, $ \E [ T^{2p}] \leq c n ^{2p/\alpha}$ and $ \E [ S^{2p}] \leq c n ^{2p/\alpha}$. Hence the above difference of expectations is at most a constant times
$n^{\frac{p}{\alpha}+\frac12}$, which is negligeable with respect to $n^{1-\frac{p}{d}}$ since $\alpha$ is assumed
to be large enough.
\end{proof}
\subsubsection{Approximations}
\begin{proposition}\label{prop:lipschitz}
Assume that a bipartite functional $L$ satisfies the regularity property $(\mathcal R_p)$.
Let $m,n>0$ and $\mu$ be a probability measure with support included in a set $Q$. Then
$$ \E L(n\mu)\le \E L(m\mu)+C \mathrm{diam}(Q)^p |m-n|.$$
\end{proposition}
\begin{proof}
Assume $n<m$ (the other case is treated in the same way). Let $(X_i)_{i\ge 1}, (Y_i)_{i\ge 1},N_1,N_2,K_1,K_2$
be mutually independent random variables, such that for all $i\ge 1$, $X_i$ and $Y_i$ have law $\mu$, and for $j\in \{1,2\}$, the law of $N_j$ is $\mathcal P(n)$ and the law of $K_j$ is $\mathcal P(m-n)$.
Then $M_i=N_i+K_i$ is $\mathcal P(m)$-distributed. Then $\{X_1,\ldots,X_{N_1}\}$ and $\{Y_1,\ldots,Y_{N_2}\}$
are independent Poisson point processes of intensity $n\mu$, while Then $\{X_1,\ldots,X_{M_1}\}$ and $\{Y_1,\ldots,Y_{M_2}\}$
are independent Poisson point processes of intensity $m\mu$.
By the regularity property,
$$ L\big(\{X_1,\ldots,X_{N_1}\},\{Y_1,\ldots,Y_{N_2}\}\big)\le L\big(\{X_1,\ldots,X_{N_1+K_1}\},\{Y_1,\ldots,Y_{N_2+K_2}\} \big)+ C \mathrm{diam}(Q)^p (K_1+K_2).$$
Taking expectations gives the claim.
\end{proof}
Applying the above inequality for $m=0$ gives a weak size bound on $\E L(\nu)$.
\begin{corollary}\label{cor:trivialbound}
Assume that $L$ satisfies $(\mathcal R_p)$ and $L(\emptyset,\emptyset)=0$ (a consequence of e.g. $(\mathcal H_p)$),
then if $\nu$ is a finite measure with support included in a set $Q$,
$$ \E L(\nu)\le C \mathrm{diam(Q)}^p \,\nu(Q).$$
\end{corollary}
Recall the total variation distance of two probability measures on $\dR^d$ is defined as
$$
d_\mathrm{TV}(\mu,\mu') = \sup \{|\mu ( A) - \mu'(A)| : A \hbox{ Borel set of $\dR^d$}\}.
$$
\begin{proposition}\label{prop:approx}
Assume that $L$ satisfies $(\mathcal R_p)$. Let $\mu,\mu'$ be two probability measures on $\dR^d$ with bounded supports.
Set $\Delta$ be the diameter of the union of their supports. Then
$$ \E L(n\mu)\le \E L(n\mu') + 4C \Delta^p\, n \, d_\mathrm{TV}(\mu,\mu').$$
\end{proposition}
\begin{proof}
The difference of expectations is estimated thanks to a proper
coupling argument. Let $\pi$ be a probability measure on
$\dR^d\times \dR^d$ having $\mu$ as its first marginal and $\mu'$ as
its second marginal. We consider mutually independent random
variables $N_1,N_2, (X_i,X_i')_{i\ge 1}, (Y_i,Y_i')_{i\ge 1}$ such
that $N_1, N_2$ are $\mathcal P(n)$ distributed and for all $i\ge
1$, $ (X_i,X_i')$ and $(Y_i,Y_i')$ are distributed according to
$\pi$. Then the random multisets
$$ \mathcal X=\{X_1,\ldots,X_{N_1}\} \quad \mathrm{and} \quad \mathcal Y=\{Y_1,\ldots,Y_{N_2}\}$$
are independent Poisson point processes with intensity measure
$n\mu$. Similarly $ \mathcal X'=\{X'_1,\ldots,X'_{N_1}\}$ and $
\mathcal Y'=\{Y'_1,\ldots,Y'_{N_2}\}$ are independent Poisson point
processes with intensity measure $n\mu'$.
The regularity property ensures that
\begin{eqnarray*}
\lefteqn{ L\big(\{X_1,\ldots,X_{N_1}\},\{Y_1,\ldots,Y_{N_2}\}\big)}\\
&\le& L\big(\{X'_1,\ldots,X'_{N_1}\},\{Y'_1,\ldots,Y'_{N_2}\}\big)+2C\Delta^p \left(\sum_{i=1}^{N_1} \mathbbm 1_{X_i\neq X'_i}+
\sum_{j=1}^{N_2} \mathbbm 1_{Y_j\neq Y'_j} \right).
\end{eqnarray*}
Taking expectations yields
\begin{eqnarray*}
\E L(n\mu)&\le& \E L(n\mu') +2C\Delta^p \E \left(\sum_{i=1}^{N_1} \mathbb P (X_i\neq X'_i)+
\sum_{j=1}^{N_2} \mathbb P(Y_j\neq Y'_j) \right) \\
&=& \E L(n\mu') +4C\Delta^p\, n \, \pi\big(\{(x,y)\in (\dR^d)^2;\; x\neq y\}\big).
\end{eqnarray*}
Optimizing the later term on the coupling $\pi$ yields the claimed inequality involving the total variation distance.
\end{proof}
\begin{corollary}\label{cor:couplingaverage}
Assume that the functional $L$ satisfies the regularity property $(\mathcal R_p)$.
Let $m>0$, $Q\subset \dR^d$ be measurable with positive Lebesgue
measure and let $f$ be a nonnegative locally integrable function on
$\dR^d$. Let $\alpha=\int_Q f/\mathrm{vol}(Q)$ be the average value
of $f$ on $Q$. It holds
$$ \E L(m\,f \mathbbm 1_Q) \le \E L(m\alpha \mathbbm 1_Q) + 2C m\, \mathrm{diam}(Q)^p \, \int_Q|f(x)-\alpha|\, dx.$$
\end{corollary}
\begin{proof}
We simply apply the total variation bound of the previous lemma
with $n=m\int_Q f=m\alpha\, \mathrm{vol}(Q)$, $d\mu(x)= f(x)\mathbbm 1_Q(x)
dx/\int_Q f $ and $d\mu'(x)= \mathbbm 1_Q(x)dx/\mathrm{vol}(Q)$. Note that
$$ 2d_{TV}(\mu,\mu')=\int \Big| \frac{f(x) \mathbbm 1_Q(x)}{\int_Q f}-\frac{\mathbbm 1_Q(x)}{\mathrm{vol(Q)}}\Big|\, dx=
\frac{\int_Q|f(x)-\alpha|\, dx}{\int_Q f}\cdot$$
\end{proof}
\subsubsection{Average is enough}
It is known since the works of Rhee and Talagrand that concentration inequalities often allow to deduce almost sure
convergence from convergence in average. This is the case in our general setting.
\begin{proposition}\label{prop:conc}
Let $L$ be a bipartite functional on multisets of $\dR^d$, satisfying the regularity property $(\mathcal R_p)$.
Assume $d>2p>0$. Let $\mu$ be a probability measure $\mu$ on $\dR^d$ with
$\int |x|^\alpha d\mu(x)<+\infty.$
Consider independent variables
$(X_i)_{i\ge 1}$ and $(Y_i)_{i\ge 1}$ with distribution $\mu$.
If $\alpha>2dp/(d-2p)$ then the following convergence holds in probability:
$$\lim_{n\to \infty} \frac{L\big(\{X_i\}_{i=1}^n,\{Y_i\}_{i=1}^n\big)-\E L\big(\{X_i\}_{i=1}^n,\{Y_i\}_{i=1}^n\big)}{n^{1-\frac{p}{d}}}=0.$$
Moreover if $\alpha>4dp/(d-2p)$, the convergence happens almost surely, and if $\mu$ has bounded support, then it also holds in $L^q$ for any $q \geq 1$.
\end{proposition}
\begin{proof}
This is a simple consequence of Azuma's concentration inequality. It is convenient to $Z(n)=(X_1,\ldots,X_n,Y_1,\ldots,Y_n)$.
Assume first that the support of $\mu$ is bounded and let $\Delta$ denote its diameter. By the regularity property, modifying one point changes the
value of the functional by at most a constant:
$$ |L(Z_1,\ldots,Z_{2n})-L(Z_1,\ldots, Z_{i-1}, Z'_i,Z_{i+1},\ldots,Z_{2n})|\le 2C \Delta^p.$$
By conditional integration, we deduce that the following martingale difference:
$$ d_i:=\E\big( L(Z(n))\,|\, Z_1,\ldots,Z_i\big) -\E\big( L(Z(n))\,|\, Z_1,\ldots,Z_{i-1}\big) $$
is also bounded $|d_i|\le 2C \Delta^p$ almost surely.
Recall that Azuma's inequality states that
$$ \mathbb P \left( \big|\sum_{i=1}^{k} d_i\big|>t\right) \le 2 e^{-\frac{t^2}{2 \sum_i \|d_i\|_\infty^2}}.$$
Therefore, we obtain that
\begin{equation}\label{eq:conc}
\mathbb P \Big( \big|L(\{X_i\}_{i=1}^n,
\{Y_i\}_{i=1}^n)-\E L(\{X_i\}_{i=1}^n,
\{Y_i\}_{i=1}^n)\big|>t\Big) \le 2 e^{-\frac{t^2}{16n C^2 \Delta^{2p}}},
\end{equation}
and there is a number $C'$ (depending on $\Delta$ only) such that
$$ \mathbb P \left( \frac{\big|L(\{X_i\}_{i=1}^n,
\{Y_i\}_{i=1}^n)-\E L(\{X_i\}_{i=1}^n,
\{Y_i\}_{i=1}^n)\big|}{n^{1-\frac{p}{d}}}>t\right) \le 2 e^{-C't^2 n^{1-\frac{2p}{d}}}.$$
When $d>2p$, we may conclude by the Borel-Cantelli lemma.
If $\mu$ is not assumed to be of bounded support, let $S:=\max\{|Z_i|;\; i\le 2n\}$. A conditioning argument
allows to use the above method.
Let $s > 0$ and $B(s)=\{x;\; |x|\le s\}$. Given $\{ S \leq s \}$, the variables $\{X_1, \cdots, X_n\}$ and $\{Y_1, \cdots, Y_n\}$ are mutually independent sequences with distribution $\mu_{| B(s)} / \mu( B(s))$.
Hence, applying \eqref{eq:conc} for $\mu_{| B(s)} / \mu( B(s))$ instead of $\mu$ and $2s$ instead of $\Delta$,
for any $t >0$,
$$
\dP \left( \left| \frac{L\big(\{X_i\}_{i=1}^n,
\{Y_i\}_{i=1}^n\big)}{ n^{1-\frac{p}{d}}} - \frac { \E
L\big(\{X_i\}_{i=1}^n, \{Y_i\}_{i=1}^n\big) }{
n^{1-\frac{p}{d}}} \right| > t \Bigm| S \leq s \right) \leq 2 \exp \left(
-\frac{ n ^{ 1 - \frac{ 2p}{ d}} t ^2 }{c_p s^{2p} } \right).
$$
Hence for $\delta>0$ to be chosen later,
\begin{align*}
u_n:&= \dP \left( \left| \frac{L\big(\{X_i\}_{i=1}^n,
\{Y_i\}_{i=1}^n\big)}{ n^{1-\frac{p}{d}}} - \frac { \E
L\big(\{X_i\}_{i=1}^n, \{Y_i \}_{i=1}^n\big) }{
n^{1-\frac{p}{d}}} \right| > t \right) \\
& \leq \dP( S > n^{\frac 1 \delta} ) + 2 \exp \left( -\frac{ n ^{ 1 - \frac {2p}{ d} - \frac{ 2p}{ \delta} } t ^2 }{ c_p} \right).
\end{align*}
Since $\dP(S>u)=1-(1-\mu(B(s))^{2n}\le 2n \mu(B(s))\le 2n (\int |x|^\alpha d\mu(x))/u^\alpha$, we get that for
some constant $c$ and any $\delta >0$,
$$ u_n \le c n^{1-\frac \alpha \delta } + 2 \exp \left( -\frac{ n ^{ 1 - \frac{ 2p}{ d} - \frac{ 2p}{ \delta} } t ^2 }{ c_p} \right).$$
Since $\alpha > 2dp/(d-2p)$ we may choose $\delta\in [2 dp / ( d -2p),\alpha]$, which ensures that the latter quantities tend
to zero as $n$ increases. This shows the convergence in probability
to $0$ of $$ \frac{L\big(\{X_i\}_{i=1}^n,
\{Y_i\}_{i=1}^n\big)}{ n^{1-\frac{p}{d}}} - \frac { \E
L\big(\{X_i\}_{i=1}^n, \{Y_i \}_{i=1}^n\big) }{ n^{1-\frac{p}{d}}}.$$
If $\alpha > 4 dp / ( d -2p)$ we may choose we may choose $\delta\in [2 dp / ( d -2p),\alpha/2]$, which ensures
that $\sum_n u_n<+\infty$. The Borel-Cantelli lemma yields the almost
sure convergence to $0$.
\end{proof}
\subsection{Consequences of subadditivity}
We start with a very general statement, which is
however not very precise when the measures do not have disjoint
supports.
\begin{proposition}\label{cor:mu1mu2} Let $L$ satisfy $(\mathcal S_p)$.
Let $\mu_1,\mu_2$ be finite measures on $\dR^d$ with
supports included in a set $Q$. Then
$$ \E L( \mu_1+\mu_2 ) \le \E L( \mu_1)+ \E L( \mu_2)
+2C \mathrm{diam}(Q)^p \left(1+ \sqrt{\mu_1(Q)}+
\sqrt{\mu_2(Q)}\right).$$
\end{proposition}
\begin{proof} Consider four independent Poisson point processes $\mathcal{X}_1,\mathcal{Y}_1,\mathcal{X}_2,\mathcal{Y}_2 $ such that
for $i\in \{1,2\}$, the intensity of $\mathcal{X}_i$ and of $\mathcal{Y}_i$ is $\mu_i$.
It is classical \cite{K} that the random
multiset $\mathcal{X}_1\cup \mathcal{X}_2$ is a Poisson point process with intensity $\mu_1+\mu_2$. Also,
$\mathcal{Y}_1\cup\mathcal{Y}_2$ is an independent copy of the latter process.
Applying the subadditivity property,
\begin{eqnarray*}
\lefteqn{ L(\mathcal{X}_1\cup\mathcal{X}_2,\mathcal{Y}_1\cup \mathcal{Y}_2)} \\
& \le & L(\mathcal X_1,\mathcal Y_1)+ L(\mathcal X_2,\mathcal Y_2)+
C \mathrm{diam}(Q)^p \left(1+|\mathrm{card}(\mathcal X_1)
-\mathrm{card}(\mathcal Y_1)|+ 1+|\mathrm{card}(\mathcal X_2)
-\mathrm{card}(\mathcal Y_2)| \right).
\end{eqnarray*}
Since $\mathrm{card}(\mathcal X_i)$ and $\mathrm{card}(\mathcal Y_i)$ are independent with
Poisson law of parameter $\mu_i(Q)$ (the total mass of $\mu_i)$,
$$ \E |\mathrm{card}(\mathcal X_i)-\mathrm{card}(\mathcal Y_i)|\le \left(\E \big(\mathrm{card}(\mathcal X_i)-\mathrm{card}(\mathcal Y_i)\big)^2\right)^{\frac12}=
\sqrt{2\mathrm{var}(\mathrm{card}(\mathcal X_i))}=\sqrt{2\mu_i(Q)}.$$
Hence, taking expectations in the former estimate leads to the claimed inequality.\end{proof}
Partition techniques are essential in the probabilistic theory of Euclidean functionals.
The next statement allows to apply them to bipartite functionals. In what follows, given a multiset $\mathcal{X}$
and a set $P$, we set $\mathcal{X}(P):=\mathrm{card}(\mathcal{X}\cap P)$.
If $\mu$ is a measure and $f$ a nonnegative function, we write $f
\cdot \mu$ for the measure having density $f$ with respect to $\mu$.
\begin{proposition} \label{prop:partition}
Assume that the functional $L$ satisfies $(\mathcal S_p)$.
Consider a finite partition $Q=\cup_{P\in \mathcal P} P$ of a subset of $\dR^d$ and
let $\nu$ be a measure on $\dR^d$ with
$\nu(Q)<+\infty$. Then
$$ \E L(\mathbbm 1_Q\cdot \nu) \le \sum_{P\in \mathcal P} \E L(\mathbbm 1_P\cdot \nu)+
3C\mathrm{diam}(Q)^p \sum_{p\in \mathcal P}
\sqrt{\nu(P)}.$$
\end{proposition}
\begin{proof}
Consider $\mathcal X,\mathcal Y$ two independent Poisson point
processes with intensity $\nu$. Note that $\mathcal X\cap P$ is a
Poisson point process with intensity $\mathbbm 1_P\cdot \nu$, hence
$\mathcal X(P)$ is a Poisson variable with parameter $\nu(P)$.
We could apply the subadditivity property to $(\mathcal X\cap P)_{P\in\mathcal
P}$, $(\mathcal Y\cap P)_{P\in \mathcal P}$, which yields
$$ L(\mathcal X \cap Q, \mathcal Y\cap Q) \le \sum_{P\in \mathcal P} L(\mathcal X \cap P, \mathcal Y\cap P)
+ C \mathrm{diam}(Q)^p \sum_{p\in \mathcal P} \big( 1+| \mathcal X(P)-
\mathcal Y(P)|\big).$$
Nevertheless, doing this gives a contribution at least $ C \mathrm{diam}(Q)^p$ to
cells which do not intersect the multisets $\mathcal{X},\mathcal{Y}$.
To avoid this rough estimate, we consider the cells which meet at least one of the multisets:
$$\widetilde{\mathcal P}:=\{P\in \mathcal P;\; \mathcal{X}(P)+\mathcal{Y}(P)\neq 0\}.$$
We get that
\begin{eqnarray*}
L(\mathcal X \cap Q, \mathcal Y\cap Q) &\le& \sum_{P\in \widetilde{\mathcal P}}
L(\mathcal X \cap P, \mathcal Y\cap P)
+ C \mathrm{diam}(Q)^p \sum_{p\in \widetilde{\mathcal P}} \big( 1+| \mathcal X(P)-
\mathcal Y(P)|\big) \\
&\le& \sum_{P\in \mathcal P} L(\mathcal X \cap P, \mathcal Y\cap P)
+ C \mathrm{diam}(Q)^p \sum_{p\in \mathcal P} \mathbbm 1_{\mathcal{X}(P)+\mathcal{Y}(P)\neq 0}\big( 1+| \mathcal X(P)-
\mathcal Y(P)|\big)\\
&\le& \sum_{P\in \mathcal P} L(\mathcal X \cap P, \mathcal Y\cap P)
+ C \mathrm{diam}(Q)^p \sum_{p\in \mathcal P} \big(\mathbbm 1_{\mathcal{X}(P)+\mathcal{Y}(P)\neq 0} +| \mathcal X(P)-
\mathcal Y(P)|\big).
\end{eqnarray*}
Since $\mathcal{X}(P)$ and $\mathcal{Y}(P)$ are independent Poisson variables with parameter $\nu(P)$,
$$ \mathbb P\big(\mathcal{X}(P)+\mathcal{Y}(P) \neq 0\big)=1-e^{-2\nu(P)} \mbox{ and }
\E \big|\mathcal X(P)-\mathcal Y(P)\big|\le \sqrt{2\nu(P)}.$$
Hence, taking expectation and using the bound $1-e^{-t}\le \min(1,t)\le \sqrt{t}$,
$$ \E L(\mathbbm 1_Q\cdot \nu) \le \sum_{P\in \mathcal P} \E L(\mathbbm 1_P\cdot \nu)+
2 \sqrt 2 C\mathrm{diam}(Q)^p \sum_{p\in \mathcal P} \sqrt{\nu(P)}.
$$
\end{proof}
The next statement deals with iterated partitions, which are very useful in the study
of combinatorial optimisation problems, see e.g. \cite{S,Y}.
If $\mathcal P$ is a partition,
we set $\mathrm{diam}(\mathcal P)=\max_{P\in \mathcal P} \mathrm {diam}( P)$ (the maximal
diameter of its cells).
\begin{corollary}\label{cor:part}
Assume that the functional $L$ satisfies $(\mathcal S_p)$.
Let $Q\subset \dR^d$ and $\mathcal Q_1, \ldots, \mathcal Q_k$ be a sequence of finer and finer
finite partitions of $Q$. Let $\nu $ be a measure on $\dR^d$ with $\nu(Q)<+\infty$. Then
$$ \E L(\mathbbm 1_Q\cdot \nu) \le \sum_{q\in \mathcal Q_k} \E L(\mathbbm 1_q\cdot \nu)+
3C \sum_{i=1}^k \mathrm{diam}(\mathcal Q_{i-1})^p \sum_{q \in \mathcal Q_i}
\sqrt{\nu(q)},$$
where by convention $\mathcal Q_0=\{Q\}$ is the trivial partition.
\end{corollary}
\begin{proof}
We start with applying Proposition~\ref{prop:partition} to the partition $\mathcal Q_1$ of $Q$:
$$ \E L(\mathbbm 1_Q\cdot \nu) \le \sum_{q\in \mathcal Q_1} \E L(\mathbbm 1_q\cdot \nu)+
3C \mathrm{diam}(\mathcal Q_{0})^p \sum_{q \in \mathcal Q_1}
\sqrt{\nu(q)}.$$
Next for each $q\in \mathcal Q_1$ we apply the proposition again for the partition of $q$ induced
by $\mathcal Q_2$ and iterate the process $k-2$ times.
\end{proof}
\section{Uniform cube samples}\label{sec:cube}
We introduce a specific
notation for $n\in(0,+\infty)$,
$$ \bar L(n):=\E L\big( n\mathbbm 1_{[0,1]^d}\big).$$
We point out the following easy consequence of the homogeneity properties of Poisson point processes.
\begin{lemma}\label{lem:homo} If $L$ satisfies the homogeneity property $(\mathcal H_p)$ then
for all $a\in \dR^d$, $\rho>0$ and $n>0$
$$ \E L\big(n\mathbbm 1_{a+[0,\rho]^d}\big)=\rho^p \bar L \big(n \rho^d \big).$$
\end{lemma}
The following theorem is obtained by adapting to our abstract setting the line of reasoning of Boutet de Monvel and Martin in the paper \cite{BM} which was devoted to the bipartite matching:
\begin{theorem}\label{th:cube-poisson}
Let $d>2p$ be an integer. Let $L$ be a bipartite functional on $\dR^d$ satisfying the properties $(\mathcal H_p)$, $(\mathcal R_p)$ and $(\mathcal S_p)$. Then there exists $\beta_L\ge 0$ such
that
$$\lim_{n\to\infty} \frac{\bar L(n)}{n^{1-\frac{p}{d}}}=\beta_L.$$
\end{theorem}
\begin{proof}
Let $m\ge 1$ be an integer. Let $K\in \mathbb N$ such that $2^K \le m< 2^{K+1}$. Set $Q_0=[0,a]^d$ where
$a:=2^{K+1}/m>1$. Let $\mathcal Q_0=\{Q_0\}$. We consider a sequence of finer and finer partitions
$\mathcal Q_j$, $j\ge 1$ where $\mathcal Q_j$ is a partition
of $Q_0$ into $2^{jd}$ cubes of size $a 2^{-j}$ (throughout the paper, this means that the interior of
the cells are open cubes of such size, while their closure is a closed cube of the same size. We do not
describe precisely how the points in the boundaries of the cubes are partitioned, since it is not relevent
for the argument). One often says that $\mathcal Q_j$, $j\ge 1$ is a sequence of dyadic partitions of $Q_0$.
A direct application of Corollary~\ref{cor:part} for the partitions $\mathcal Q_1,\ldots, \mathcal Q_{K+1}$
and the measure $n\mathbbm 1_{[0,1]^d}(x) \, dx$ gives
$$\bar L(n) = \E L(n\mathbbm 1_{[0,1]^d}) \leq \sum_{q\in \mathcal Q_{K+1}} \E L(n \mathbbm 1_{q\cap[0,1]^d}) + 3C \sum_{j=1}^{K+1}
\mathrm{diam}(\mathcal Q_{j-1})^p \sum_{q\in \mathcal Q_j} \sqrt{n \,\mathrm{Vol}(q\cap [0,1]^d)}.$$
Note that $\mathcal Q_{K+1}$ is a partition into cubes of size $1/m$, so that its intersection
with $[0,1]^d$ induces an (essential) partition of the unit cube into $m^d$ cubes of side-length $1/m$.
Hence, in the first sum, there are $m^d$ terms which are equal, thanks to translation invariance and Lemma~\ref{lem:homo}
to $\E L(n\mathbbm 1_{[0,m^{-1}]^d})=m^{-p} \bar L( nm^{-d})$. The remaining terms of the first sum vanish.
In order to deal with the second
sum of the above estimate, we simply use the fact that $\mathcal Q_j$ contains $2^{jd}$ cubical cells
of size $a 2^{-j}=2^{K+1-j}/m \le 2^{1-j}$. Hence their indidual volumes are at most $2^{d(1-j)}$.
These observations allow to rewrite the above estimate as
\begin{eqnarray*}
\bar L(n)&\le & m^{d-p} \bar L(nm^{-d}) +3C \sum_{j=1}^{K+1}
\mathrm{diam}( [0,2^{2-j}]^d)^p 2^{jd} \sqrt{n \, 2^{d(1-j)}}\\
&=& m^{d-p} \bar L(nm^{-d}) + 3C \sqrt n \,\mathrm{diam}( [0,1]^d)^p \sum_{j=1}^{K+1}
2^{p(2-j)+\frac{d}{2}(j+1)}.
\end{eqnarray*}
Hence, there is a number $D$ depending only on $p,d$ and $C$ such that
$$ \bar L(n) \le m^{d-p} \bar L(nm^{-d}) + D \sqrt n \,2^{K (\frac{d}{2}-p)} \le m^{d-p} \bar L(nm^{-d}) + D\sqrt n \, m^{\frac{d}{2}-p}.$$
Let $t>0$. Setting, $n=m^d t^d$ and $f(u)=\bar L(u^d)/u^{d-p}$, the latter inequality reads as
$$f(mt)\le f(t)+ D t^{p-\frac{d}{2}},$$
and is valid for all $t>0$ and $m\in \mathbb N^*$. Since $f$ is continuous (Proposition~\ref{prop:lipschitz}
shows that $u\mapsto \bar L(u)$ is Lipschitz) and $\lim_{t\to +\infty} t^{p-\frac{d}{2}}=0$, it follows that $\lim_{t\to +\infty} f(t)$ exists
(we refer to \cite{BM} for details).
\end{proof}
\begin{remark}
The above constant $\beta_L$ is positive as soon as
$L$ satisfies the following natural condition: for all $x_1,\ldots,x_n,y_1,\ldots y_n$ in $\dR^d$, $L(\{x_1,\ldots,x_n\},\{y_1,\ldots,y_n\}) \ge c \sum_i
\mathrm{dist}(x_i,\{y_1,\ldots,y_n\})^p$.
To see this, one combines Proposition~\ref{prop:poisson} and the lower estimate given in \cite{T}.
\end{remark}
\section{Upper bounds, upper limits} \label{sec:upper}
\subsection{A general upper bound}
\begin{lemma}
\label{lem:upperL} Let $d>2p$ and let $L$ be a bipartite functional satisfying $(\mathcal S_p)$,
$(\mathcal R_p)$ and $L(\emptyset,\emptyset)=0$.
Then there exists a constant $D$ such that, for all finite measures $\nu$,
$$
\E L ( \nu) \leq D \, \mathrm{diam}(Q)^p \min\big( \nu(Q), \nu(Q) ^{1-\frac{p}{d}}\big),$$
where $Q$ contains the support of $\nu$.
\end{lemma}
\begin{proof}
Thanks to corollary~\ref{cor:trivialbound}, it is enough to deal with the case $\nu(Q)\ge 2^d$ (or any other
positive number).
First note that we may assume that $Q$ is a cube (given a set of diameter $\Delta$, one can find
a cube containing it, with diameter no more than $c$ times $\Delta$ where $c$ only depends on the
norm).
We consider a sequence of dyadic partitions of $Q$, $(\mathcal P_\ell)_{\ell\ge 0}$, where
for $\ell \in\mathbb N$,
$\mathcal P_\ell$ divides $Q$ into $2^{\ell d}$ cubes of
side-length $2^{-\ell}$ times the one of $Q$. Let $k\in \mathbb N^*$ to be chosen later.
By Corollary~\ref{cor:part}, we have the following estimate
\begin{equation}\label{eq:upper20} \E L( \nu )\le \sum_{P\in \mathcal
P_k} \E L(\mathbbm 1_P\cdot \nu)+
3C \sum_{\ell=1}^k \big(2^{-\ell+1}\mathrm{diam}(Q)\big)^p \sum_{P\in\mathcal P_\ell} \sqrt{\nu (P) }.
\end{equation}
Thanks to Corollary~\ref{cor:trivialbound},
the first term of the right-hand side of \eqref{eq:upper20} is at most
$$
\sum_{P\in \mathcal P_k} C\, \big(2^{-k}\mathrm{diam}(Q)\big)^p \nu (P)
=C\, 2^{-kp} \big(\mathrm{diam}(Q)\big)^p
\nu (Q) .
$$
By the Cauchy-Schwarz inequality
$$\sum_{P\in\mathcal P_\ell} \sqrt{ \nu (P)} \le \left(2^{\ell d} \right)^{\frac12}
\left( \sum_{P\in\mathcal P_\ell} \nu (P) \right)^{\frac12}= 2^{\frac{\ell d}{2}} \sqrt{\nu(Q)}.$$
Hence the second term of the right-hand side of \eqref{eq:upper20} is at most
$$ 3C \big(2\mathrm{diam}(Q)\big)^p \sum_{\ell=1}^k 2^{\ell \big(\frac{d}{2}-p\big)} \sqrt{\nu(Q)}
\le C' 2^{k\big(\frac{d}{2}-p\big)} \big(\mathrm{diam}(Q)\big)^p \sqrt{\nu(Q)}.$$
This leads to
$$ \E L(\nu)\le \big(\mathrm{diam}(Q)\big)^p \Big( C 2^{-kp} \nu(Q) + C' 2^{k\big(\frac{d}{2}-p\big)}
\sqrt{\nu(Q)}\Big).$$
Choosing $ k = \big\lfloor \frac 1 d \log_2 \nu(Q) \big\rfloor\ge 1$ completes the proof.
\end{proof}
\subsection{The upper limit for densities}
\begin{theorem}\label{th:upper}
Let $d>2p$. Let $L$ be a bipartite functional on $\dR^d$ satisfying the properties $(\mathcal H_p)$,
$(\mathcal R_p)$, $(\mathcal S_p)$. Let $f:\dR^d\to \dR^+$ be an integrable function with
bounded support. Then
$$ \limsup_{n\to \infty } \frac{\E L(n\, f)}{n^{1-\frac{p}{d}}} \le \beta_L \int_{\dR^d} f^{1-\frac{p}{d}},$$
where $\beta_L$ is the constant appearing in
Theorem~\ref{th:cube-poisson}.
\end{theorem}
\begin{proof}
By a scaling argument, we may assume that the support of $f$ is
included in $[0,1]^d$ and $\int f =1$ (the case $\int f =0$ is
trivial). We consider a sequence of dyadic partitions $(\mathcal P_\ell)_{\ell\in\mathbb N}$ of
$[0,1]^d$: for $\ell \in\mathbb N$, $\mathcal P_\ell$ divides $[0,1]^d$ into $2^{\ell d}$ cubes of
side-length $2^{-\ell}$. Let $k\in \mathbb N^*$ to be chosen later.
Corollary~\ref{cor:part} gives
\begin{equation}\label{eq:upper} \E L(n\, f)\le \sum_{P\in \mathcal
P_k} \E L(n\, f\mathbbm 1_P)+
3C\sum_{\ell=1}^k\big( 2^{-\ell+1} \mathrm{diam}([0,1]^d)\big)^p
\sum_{P\in\mathcal P_\ell} \sqrt{n\int_P f}.
\end{equation}
By the Cauchy-Schwarz inequality
$$\sum_{P\in\mathcal P_\ell} \sqrt{\int_P f} \le \left(2^{\ell d} \right)^{\frac12}
\left( \sum_{P\in\mathcal P_\ell} \int_P f \right)^{\frac12}=
2^{\frac{\ell d}{2}} \left(\int f\right)^ \frac12= 2^{\frac{\ell d}{2}}.$$
Hence the
second term of the right-hand side of \eqref{eq:upper} is at most
$$ 3C\big( 2 \mathrm{diam}([0,1]^d)\big)^p \sqrt{n} \sum_{\ell=1}^k 2^{\ell \big(\frac{d}{2}-p\big)}
\le c_d n^{\frac12} 2^{k\big(\frac{d}{2}-p\big)}.$$
Let $\alpha_P$ be the average of $f$ on $P$, then applying
Corollary~\ref{cor:couplingaverage} to the first terms of
\eqref{eq:upper} leads to
$$ \E L(n\, f)\le \sum_{P\in \mathcal P_k} \left( \E L(n\, \alpha_P \mathbbm 1_P)+2C\, n\, \mathrm{diam}(P)^p \int_P|f-\alpha_P| \right) +c_d n^{\frac12} 2^{k\big(\frac{d}{2}-p\big)}.$$
Each $P$ in the sum is a square of side length $2^{-k}$,
hence using homogeneity (see Lemma~\ref{lem:homo})
\begin{equation}\label{eq:upper2}
\E L(n\, f)\le \sum_{P\in \mathcal P_k} \left(2^{-kp} M\big(n\, \alpha_P 2^{-kd}\big)+n\,c'_d\, 2^{-kp} \int_P|f-\alpha_P| \right) +c_d n^{\frac12} 2^{k\big(\frac{d}{2}-p\big)}.
\end{equation}
Let us recast this inequality with more convenient notation. We set
$g(t)=\bar L (t)/t^{1-p/d}$ and we define the piecewise constant function
$$ f_k=\sum_{P\in \mathcal P_k} \alpha_P \mathbbm 1_P
= \sum_{P\in \mathcal P_k} \frac{\int_P f(x)\, dx }{\mathrm{Vol}(P)} \mathbbm 1_P.$$
It is plain that $\int f_k=\int f<+\infty$. Moreover, by Lebesgue's
theorem, $\lim_{k\to \infty} f_k=f$ holds for almost every point
$x$. Inequality \eqref{eq:upper2} amounts to
\begin{eqnarray*}
\frac{\E L(n\, f)}{n^{1-\frac{p}{d}}} &\le &
\sum_{P\in \mathcal P_k} \left(g\big(n\,\alpha_P 2^{-kd}\big) \alpha_P^{1-\frac{p}{d}} 2^{-kd}+n^{\frac{p}{d}}\,
c'_d\, 2^{-kp} \int_P|f-f_k| \right) +c_d n^{\frac{p}{d}-\frac12} 2^{k\big(\frac{d}{2}-p\big)}\\
&=&\sum_{P\in \mathcal P_k} \left( \int_P g\big(n\, f_k 2^{-kd}\big) f_k^{1-\frac{p}{d}}
+n^{\frac{p}{d}}\, c'_d\, 2^{-kp} \int_P|f-f_k| \right) +c_d n^{\frac{p}{d}-\frac12} 2^{k\big(\frac{d}{2}-p\big)}\\
&=& \int g\big(n\,2^{-kd} f_k \big) f_k^{1-\frac{p}{d}} + c'_d\,n^{\frac{p}{d}}\, 2^{-kp} \int |f-f_k|
+c_d \big(n^{\frac1d}2^{-k}\big)^{p-\frac{d}{2}}.
\end{eqnarray*}
If there exists $k_0$ such that $f=f_{k_0}$ then we easily get the
claim by setting $k=k_0$ and letting $n$ go to infinity (since $g$
is bounded and converges to $\beta_L$ at infinity, see
Lemma~\ref{lem:upperL} and Theorem~\ref{th:cube-poisson}). On the
other hand, if $f_k$ never coincides almost surely with $f$, we use
a sequence of numbers $k(n)\in \mathbb N$ such that
\begin{equation}\label{eq:kn}
\lim_n k(n)=+\infty,\quad \lim_n n^{\frac1d} 2^{-k(n)}=+\infty \quad\mbox{and}\quad
\lim_n n^{\frac1d} 2^{-k(n)}\left( \int |f-f_{k(n)}|\right)^{\frac{1}{p}}=0.
\end{equation}
Assuming its existence, the claim follows easily: applying the
inequality for $k=k(n)$ and taking upper limits gives
$$ \limsup_{n } \frac{\E L(n\, f)}{n^{1-\frac{p}{d}}} \le \limsup_n \int g\big(n\,2^{-k(n)d} f_{k(n)} \big) f_{k(n)}^{1-\frac{p}{d}} .$$
Since $\lim f_{k(n)}=f$ a.e., it is easy to see that the limit of the
latter integral is $\beta_L\int f^{1-\frac{p}{d}}$: first the
integrand converges almost everywhere to $\beta_L f^{1-\frac{p}{d}}$
(if $f(x)=0$ this follows from the boundedness of $g$; if $f(x)\neq 0$
then the argument of $g$ is going to infinity). Secondly, the sequence
of integrands is supported on the unit cube and is uniformly
integrable since
$$\int \left(g\big(n\,2^{-k(n)d} f_{k(n)} \big) f_k^{1-\frac{p}{d}} \right)^{\frac{d}{d-p}}
\le (\sup g)^{\frac{d}{d-p}} \int f_{k(n)}= (\sup g)^{\frac{d}{d-p}}
\int f <+\infty.$$
It remains to establish the existence of a sequence of integers
$(k(n))_n$ satisfying \eqref{eq:kn}. Note that since $f_k\ge 0$,
$\int f_k=\int f=1$ and a.e. $\lim f_k=f$, it follows from
Scheff\'e's lemma that $\lim_k \int |f-f_k|=0$. Hence
$\varphi(k)=(\sup_{j\ge k} \int |f-f_j|)^{-d/p}$ is non-decreasing with
an infinite limit. We derive the existence of a sequence with the
following stronger properties
\begin{equation}\label{eq:kn2}
\lim_n k(n)=+\infty,\quad \lim_n \frac{n }{ (2^d)^{ k(n)}}=+\infty \quad\mbox{and}\quad
\lim_n \frac{ n}{(2^d )^{k(n)}\varphi (k(n))}=0
\end{equation}
as follows. Set $\gamma=2^d$. Since $\gamma^k \sqrt{\varphi(k-1)}$ is
increasing with infinite limit
$$ [\gamma\sqrt{\varphi(0)},+\infty)
=\cup_{k\ge 1} \big[\gamma^k\sqrt{\varphi(k-1)},\gamma^{k+1}\sqrt{\varphi(k)}\big).$$
For $n\ge \gamma\sqrt{\varphi(0)}$, we define $k(n)$ as the integer such
that
$$ \gamma^{k(n)}\sqrt{\varphi(k(n)-1)}\le n < \gamma^{k(n)+1}\sqrt{\varphi(k(n))}.$$
This defines a non-decreasing sequence. It is clear from the above
strict inequality that $\lim_n k(n)=+\infty$. Hence $n \gamma^{-k(n)}\ge
\sqrt{\varphi(k(n)-1)}$ tends to infinity at infinity. Eventually
$n/(\gamma^{k(n)}\varphi(k(n))\le \gamma/\sqrt{\varphi(k(n))}$ tends to zero
as required. The proof is therefore complete.
\end{proof}
\subsection{Purely singular measures}
\begin{lemma}\label{le:sing} Let $d> 2p$. Let $L$ be a bipartite functional on $\dR^d$
with properties $(\mathcal R_p)$ and $(\mathcal S_p)$.
Let $\mu$ be a finite singular measure on $\mathbb R^d$ having a bounded support.
Then
$$\lim_{n\to \infty} \frac{\E L(n\mu)}{n^{1-\frac{p}{d}}}=0 . $$
\end{lemma}
\begin{proof}
Let $Q$ be a cube which contains the support of $\mu$.
We consider a sequence of dyadic partitions of $Q$, $(\mathcal P_\ell)_{\ell\in\mathbb N}$.
For $\ell\in\mathbb N$, $\mathcal P_{\ell}$ divides $Q$ into
$2^{\ell d}$ cubes of side length $2^{-\ell}$ times the one of $Q$.
As in the proof of Lemma~\ref{lem:upperL}, a direct application of Corollary~\ref{cor:part}
gives for $k\in \mathbb N^*$:
\begin{equation}\label{eq:upper3}
\E L(n\mu)\le \sum_{P\in \mathcal
P_k} \E L(n\mathbbm 1_P\cdot \mu)+
3C \sum_{\ell=1}^k \big(2^{-\ell+1} \mathrm{diam}(Q)\big)^p \sum_{P\in\mathcal P_\ell} \sqrt{n\mu(P)}.
\end{equation}
The terms of the first sum are estimated again thanks to the easy bound of Corollary~\ref{cor:trivialbound}:
since each $P$ in $\mathcal P_k$ is a cube of side length $2^{-k}$ times the one of $Q$, it holds
$$ \sum_{P\in \mathcal
P_k} \E L(n\mathbbm 1_P\cdot \mu) \le \sum_{P\in \mathcal
P_k} C \big(2^{-k} \mathrm{diam}(Q) \big)^p n\mu(P) =c_{p,Q}\, 2^{-kp} n|\mu|.$$
Here $|\mu|$ is the total mass of $\mu$. We rewrite the second term in \eqref{eq:upper3} in terms
of the function
$$ g_\ell=\sum_{P\in \mathcal P_\ell} \frac{\mu(P)}{\lambda(P)} \mathbbm 1_P
,$$
where $\lambda$ stands for Lebesgue's measure. Since $\lambda(P)=2^{-\ell d} \lambda(Q)$, we get that
\begin{eqnarray*}
\E L(n\mu)&\le&c_{p,Q}\, 2^{-kp} n|\mu| +
3C\, \big(2 \mathrm{diam}(Q)\big)^p \sqrt{n}\,\sum_{\ell=1}^k 2^{-\ell p}
\sum_{P\in\mathcal P_\ell} 2^{\frac{\ell d}{2}} \lambda(Q)^{-\frac12} \lambda(P)\sqrt{\frac{\mu(P)}{\lambda(P)}} \\
&=& c_{p,Q}\, 2^{-kp} n|\mu| + 3C\, \big(2 \mathrm{diam}(Q)\big)^p \lambda(Q)^{-\frac12}
\sqrt{n}\,\sum_{\ell=1}^k 2^{\ell\big(\frac{d}{2}- p\big)}
\int\sqrt{g_\ell} .
\end{eqnarray*}
By the differentiability theorem, for Lebesgue-almost every $x$, $g_\ell(x)$ tends to zero when $\ell$ tends
to infinity (since $\mu$ is singular with respect to Lebesgue's measure). Moreover, $g_\ell$ is supported on
the unit cube and $\int (\sqrt{g_\ell})^2=\int g_\ell=|\mu|<+\infty$. Hence the sequence of functions
$\sqrt{g_\ell}$ is uniformly integrable and we can conclude that $\lim_{\ell \to \infty}\int \sqrt{g_\ell}=0$.
By Cesaro's theorem, the sequence
$$\varepsilon_k=\frac{ \sum_{\ell=1}^{k} 2^{\ell\left( \frac{d}{2}-p\right)} \int \sqrt{g_\ell}}
{ \sum_{\ell=1}^{k} 2^{\ell\left( \frac{d}{2}-p\right)}}$$
also converges to zero, using here that $d> 2p$. By an obvious upper bound of the latter denominator,
we obtain that there exists a number $c$ which does not depend on $(k,n)$ (but depends on $C,p,d,Q,|\mu|$)
such that for all $k\ge 1$
$$ \E L(n\mu) \le c\left( n 2^{-kp} + \sqrt{n }\, 2^{k\left(\frac{d}{2}-p\right)} \varepsilon_k\right),$$
where $\varepsilon_k\ge 0$ and $\lim_k \varepsilon_k=0$. We may also assume that $(\varepsilon_k)$ is non-increasing (the inequality remains valid if one replaces $\varepsilon_k$ by $\sup_{j\ge k} \varepsilon_j$).
It remains to choose $k$ in terms of $n$ in a proper way. Define
$$ \varphi(n)=\sqrt{\varepsilon_{\lfloor \frac1d \log_2 n \rfloor}} ^{\frac{-1}{\frac{d}{2}-p}}.$$
Obviously $\lim_n \varphi(n)=+\infty$. For $n$ large enough, define $k(n)\ge 1$ as the unique integer
such that
$$ 2^{k(n)} \le n^{\frac1d}\varphi(n) < 2^{k(n)+1}.$$
Setting $k=k(n)$, our estimate on the cost of the optimal matching yields
$$ \frac{\E L(n\mu)}{n^{1-\frac{p}{d}}} \le c(d) \left(\frac{2 }{\varphi(n)^p}+\varepsilon_{k(n)}
\varphi(n)^{\frac{d}{2}-p} \right) .$$
It is easy to check that the right hand side tends to zero as $n$ tends to infinity. Indeed,
$\lim_n \varphi(n)=+\infty$, hence for $n$ large enough
$$ k(n) \ge \left\lfloor \log_2\left(n^{\frac1d} \varphi(n)/2 \right)\right\rfloor \ge
\left\lfloor \frac1d \log_2 n\right\rfloor.$$
Since the sequence $(\varepsilon_k)$ is non-increasing, it follows that
$$ \varepsilon_{k(n)} \varphi(n)^{\frac{d}{2}-p} \le \varepsilon_{ \left\lfloor \frac1d \log_2 n\right\rfloor} \varphi(n)^{\frac{d}{2}-p} = \sqrt{ \varepsilon_{ \left\lfloor \frac1d \log_2 n\right\rfloor} }$$
tends to zero when $n\to \infty$. The proof is therefore complete.
\end{proof}
\subsection{General upper limits}
The first statement of Theorem \ref{th:mainup} is a consequence of Propositions \ref{prop:poisson}, \ref{prop:conc}, and the following result.
\begin{theorem}\label{th:up-poisson-general} Let $d>2p>0$. Let $L$ be a bipartite functional
on $\dR^d$ with the properties $(\mathcal H_p)$, $(\mathcal R_p)$ and $(\mathcal S_p)$.
Consider a finite measure $\mu$ on $\mathbb R^d$ such that there exists $\alpha>\frac{2dp}{d-2p}$ with
$$\int |x|^\alpha d\mu(x)<+\infty.$$ Let $f$
be a density function for the absolutely continuous part of $\mu$,
then
\begin{equation}\label{eq:glupper}
\limsup_{n\to \infty} \frac{\E L(n\mu)}{n^{1-\frac{p}{d}}}\le \beta_L \int f^{1-\frac{p}{d}} \cdot
\end{equation}
\end{theorem}
\begin{remark}
Observe that the hypotheses ensure the finiteness of $ \int f^{1-\frac{p}{d}}$. Indeed H\"older's inequality
gives
$$ \int_{\dR^d} f^{1-\frac{p}{d}} \le \left(\int_{\dR^d} (1+|x|^\alpha)f(x) dx \right)^{1-\frac{p}{d}}
\left(\int_{\dR^d} (1+|x|^\alpha)^{1-\frac{d}{p}} \right)^{\frac{p}{d}}$$
where the latter integral converges since $\alpha > \frac{2dp}{d-2p}>\frac{dp}{d-p}.$
\end{remark}
\begin{proof}
Assume first that $\mu$ has a bounded support. Write $\mu=\mu_{ac}+\mu_s$ where $\mu_s$ is the singular part and $d\mu_{ac}(x)=f(x)\, dx$.
Applying Proposition~\ref{cor:mu1mu2} to $\mu_{ac}$ and $\mu_s$, dividing by $n^{1-p/d}$, passing to the
limit and using Theorem~\ref{th:upper} and Lemma~\ref{le:sing} gives
$$ \limsup_n \frac{\E L(n\mu)}{n^{1-\frac{p}{d}}} \le \limsup_n \frac{\E L(n\mu_{ac})}{n^{1-\frac{p}{d}}}
+ \limsup_n \frac{\E L(n\mu_s)}{n^{1-\frac{p}{d}}} \le \beta_L \int f^{1-\frac{p}{d}}.$$
Hence the theorem is established for measures with bounded supports.
Now, let us consider the general case.
Let $B (t) = \{x \in \dR^d : |x | \leq t\}$. Let $A_0 = B (2)$ and for integer $\ell \geq 1$, $A_\ell = B (2^{\ell+1})\backslash B (2^{\ell})$. Now, let $\mathcal{X} = \{ X_1, \cdots, X_{N_1}\}$, $\mathcal{Y} = \{ Y_1, \cdots, Y_{N_2}\}$ be two independent Poisson process of intensity $n\mu$, and $T = \max \{ | Z| : Z \in \mathcal{X} \cup \mathcal{Y} \}$.
Applying the subadditivity property like in the proof of Proposition~\ref{prop:partition}, we obtain
\begin{eqnarray}
L ( \mathcal{X} , \mathcal{Y} ) & \leq & \sum_{\ell \geq 0} L ( \mathcal{X} \cap A_\ell , \mathcal{Y} \cap A_\ell ) + C T^p \sum_{\ell \geq 0} \mathbbm 1_{\mathcal{X} (A_\ell) +\mathcal{Y}( A_\ell)\neq 0}\big(1+|\mathcal{X} (A_\ell) -\mathcal{Y}( A_\ell)|\big) . \label{eq:UP1}
\end{eqnarray}
Note that the above sums have only finitely many non-zero terms, since $\mu$ is finite.
We first deal with the first sum in the above inequality. By Fubini's Theorem,
$$
\E \sum_{\ell \geq 0 } \frac{L ( \mathcal{X} \cap A_\ell , \mathcal{Y} \cap A_\ell ) }{n^{1-\frac{p}{d}}} = \sum_{\ell \geq 0} \E \frac{L ( \mathcal{X} \cap A_\ell , \mathcal{Y} \cap A_\ell ) }{n^{1-\frac{p}{d}}}.
$$
Applying \eqref{eq:glupper} to the compactly supported measure $\mu_{|A_\ell}$ for every integer $\ell$ gives
\begin{equation}\label{eq:limsupAl}
\limsup_n \E \frac{L ( \mathcal{X} \cap A_\ell , \mathcal{Y} \cap A_\ell ) }{n^{1-\frac{p}{d}}} \le \beta_L \int_{A_l} f^{1-\frac{p}{d}}.
\end{equation}
By Lemma \ref{lem:upperL}, for some constant $c_d$,
$$
\E \frac{L ( \mathcal{X} \cap A_\ell , \mathcal{Y} \cap A_\ell ) }{n^{1-\frac{p}{d}}} \leq c_d 2^{\ell p} \mu (A_\ell)^{1-\frac{p}{d}}.
$$
From Markov inequality, with $m_\alpha = \int |x|^\alpha d \mu (x)$,
$$
\mu (A_\ell) \leq \mu ( \dR^d \backslash B (2^{\ell})) \leq 2^{- \ell \alpha} m_\alpha.
$$
Thus, since $\alpha > 2pd/(d-2p)>dp/(d-p)$, the series $\sum_{\ell} 2^{\ell p} \mu (A_\ell)^{1-\frac{p}{d}}$ is convergent. We may then apply the dominated convergence theorem, we get from \eqref{eq:limsupAl} that
$$
\limsup_n \E \sum_{\ell \geq 0 } \frac{L ( \mathcal{X} \cap A_\ell , \mathcal{Y} \cap A_\ell ) }{n^{1-\frac{p}{d}}} \le
\beta_L \int f^{1-\frac{p}{d}}.
$$
For the expectation of the second term on the right hand side of (\ref{eq:UP1}), we use Cauchy-Schwartz inequality,
\begin{eqnarray*}
\lefteqn{\E \left[T^p \sum_{\ell \geq 0} \mathbbm 1_{\mathcal{X} (A_\ell) +\mathcal{Y}( A_\ell)\neq 0}\big(1+|\mathcal{X} (A_\ell) -\mathcal{Y}( A_\ell)|\big) \right] }\\
& \leq & \sum_{\ell \geq 0} \sqrt {\E[T^{2p} ]} \sqrt{ \E \big( \mathbbm 1_{\mathcal{X} (A_\ell) +\mathcal{Y}( A_\ell)\neq 0}\big(1+|\mathcal{X} (A_\ell) -\mathcal{Y}( A_\ell)|\big)^2\big) } \\
&\le & \sqrt2 \sqrt {\E[T^{2p} ]} \sum_{\ell \geq 0} \sqrt{ \mathbb P(\mathcal{X} (A_\ell) +\mathcal{Y}( A_\ell)\neq 0)
+ \E \big[|\mathcal{X} (A_\ell) -\mathcal{Y}( A_\ell)|^2\big] } \\
&=& \sqrt2 \sqrt {\E[T^{2p} ]} \sum_{\ell \geq 0} \sqrt{ 1-e^{-2n\mu(A_\ell)} + 2 n\mu(A_\ell)} \\
&\le & 2 \sqrt {\E[T^{2p} ]}\, \sqrt{n} \sum_{\ell \geq 0} \sqrt{\mu(A_\ell)},
\end{eqnarray*}
where we have used $1-e^{-u}\le u$.
As above, Markov inequality leads to
$$
\sum_{\ell \geq 0} \sqrt{ \mu (A_\ell) } \leq \sqrt{m_\alpha} \sum_{\ell \geq 0} 2^{-\ell \frac{\alpha}{2} }
<+\infty.
$$
Eventually we apply Lemma \ref{lem:ET} with $\gamma:=2p < 2pd/(d-2)<\alpha$ to upper bound $\E[T^{2p}]$. We get
that for some constant $c >0$ and all $n >0$,
$$
n^{-1+\frac{p}{d} } \E \left[ T^p \sum_{\ell \geq 0} \mathbbm 1_{\mathcal{X} (A_\ell) +\mathcal{Y}( A_\ell)\neq 0}\big(1+|\mathcal{X} (A_\ell) -\mathcal{Y}( A_\ell)|\big) \right] \leq c n^{ - \frac 1 2 + \frac{p}{d} + \frac{p}{ \alpha}}.
$$
Since $\alpha > 2 dp / ( d -2p)$, the later and former terms tend to zero as $n$ tends to infinity.
The upper bound (\ref{eq:glupper}) is proved. \end{proof}
\section{Examples of bipartite functionals}
\label{sec:ex}
The minimal bipartite matching is an instance
of a bipartite Euclidean functional $M_1(\mathcal{X},\mathcal{Y})$ over the multisets
$\mathcal X = \{X_1,\ldots,X_n\}$ and $\mathcal Y =
\{Y_1,\ldots,Y_n\}$. We may mention at least two other interesting
examples: the bipartite traveling salesperson problem over $\mathcal
X$ and $\mathcal Y$ is the shortest cycle on the multiset $\mathcal{X} \cup
\mathcal{Y}$ such that the image of $\mathcal{X}$ is $\mathcal{Y}$. Similarly, the bipartite
minimal spanning tree is the minimal edge-length spanning tree on $\mathcal{X}
\cup \mathcal{Y}$ with no edge between two elements of $\mathcal{X}$ or two elements
of $\mathcal{Y}$.
\subsection{Minimal bipartite matching}\label{sec:preparation}
Fix $p>0$. Given two multi-subsets of $\dR^d$
with the same cardinality, $\mathcal X=\{X_1,\ldots,X_n\}$ and
$\mathcal Y=\{Y_1,\ldots,Y_n\}$, the $p$-cost of the minimal bipartite
matching of $\mathcal X$ and $\mathcal Y$ is defined as
$$ M_p(\mathcal X,\mathcal Y)=\min_{\sigma\in \mathcal S_n} \sum_{i=1}^{n} |X_i-Y_{\sigma(i)}|^p,$$
where the minimum runs over all permutations of $\{1,\ldots,n\}$.
It is useful to extend the definition to sets of different
cardinalities, by matching as many points as possible: if $\mathcal
X=\{X_1,\ldots,X_m\}$ and $\mathcal Y=\{Y_1,\ldots,Y_n\}$ and $m\le
n$ then
$$ M_p(\mathcal X,\mathcal Y)=\min_{\sigma} \sum_{i=1}^{m} |X_i-Y_{\sigma(i)}|^p,$$
where the minimum runs over all injective maps from
$\{1,\ldots,m\}$ to $\{1,\ldots,n\}$. When $n\le m$ the symmetric
definition is chosen $ M_p(\mathcal X,\mathcal Y):=M_p(\mathcal Y,\mathcal X) $.
The bipartite functional $M_p$ is obviously homogeneous of degree $p$, i.e. it satisfies $(\mathcal H_p)$.
The next lemma asserts that it is also verifies the subadditivity
property $(\mathcal S_p)$. In the case $p=1$, this is the starting point of the paper \cite{BM}.
\begin{lemma}\label{lem:partition}
For any $p >0$, the functional $M_p$ satisfies property \eqref{eq:Sp} with constant $C = 1/2$. More precisely, if $\mathcal X_1, \ldots,\mathcal X_k$ and $\mathcal Y_1,\ldots,\mathcal Y_k$ are multisets in a bounded subset $Q \subset \dR^d$, then
$$ M_p\Big(\bigcup_{i=1}^k\mathcal X_i , \bigcup_{i=1}^k\mathcal Y_i\Big) \le \sum_{i=1}^k M_p(\mathcal X_i , \mathcal Y_i)
+ \frac{\mathrm{diam}(Q)^p}{2} \sum_{i=1}^k | \mathrm{card}(\mathcal
X_i)- \mathrm{card}(\mathcal Y_i) |.$$
\end{lemma}
\begin{proof}
It is enough to upper estimate of the cost of a particular matching
of $\bigcup_{i=1}^k\mathcal X_i $ and $ \bigcup_{i=1}^k\mathcal
Y_i$. We build a matching of these multisets as follows. For each
$i$ we choose the optimal matching of $\mathcal X_i$ and $\mathcal
Y_i$. The overall cost is $ \sum_{i} M_p(\mathcal X_i , \mathcal
Y_i)$, but we have left $ \sum_{i} | \mathrm{card}(\mathcal X_i)-
\mathrm{card}(\mathcal Y_i)|$ points unmatched (the number of
excess points). Among these points, the less numerous species
(there are two species: points from $\mathcal X_i$'s, and points
from $\mathcal Y_i$'s) has cardinality at most $\frac{1}{2}
\sum_{i} | \mathrm{card}(\mathcal X_i)- \mathrm{card}(\mathcal
Y_i)|$. To complete the definition of the matching, we have to
match all the points of this species in the minority. We do this
in an arbitrary manner and simply upper bound the distance between
matched points by the diameter of $Q$.
\end{proof}
The regularity property is established next.
\begin{lemma}\label{lem:Rp}
For any $p >0$, the functional $M_p$ satisfies property \eqref{eq:Rp} with constant $C = 1$.
\end{lemma}
\begin{proof}
Let $\mathcal{X},\mathcal{X}_1,\mathcal{X}_2,\mathcal{Y},\mathcal{Y}_1,\mathcal{Y}_2$ be finite multisets contained in $Q = B(1/2)$. Denote by $x,x_1,x_2,y,y_1,y_2$ the cardinalities of the multisets and $a\wedge b$ for $\min(a,b)$. We start with an optimal matching
for $ M_p(\mathcal{X}\cap \mathcal{X}_2,\mathcal{Y}\cap \mathcal{Y}_2)$. It comprises $(x+x_2)\wedge (y+y_2)$ edges. We remove the ones
which have a vertex in $\mathcal{X}_2$ or in $\mathcal{Y}_2$. There are at most $x_2+y_2$ of them, so we are left with
at least $\big((x+x_2)\wedge (y+y_2)-x_2-y_2\big)_+$ edges connecting points of $\mathcal{X}$ to points of $\mathcal{Y}$.
We want to use this partial matching in order to build a (suboptimal) matching of $\mathcal{X}\cap \mathcal{X}_1$ and
$\mathcal{Y}\cap \mathcal{Y}_1$. This requires to have globally $(x+x_1)\wedge (y+y_1)$ edges. Hence we need to add at most
$$ (x+x_1)\wedge (y+y_1)-\big((x+x_2)\wedge (y+y_2)-x_2-y_2\big)_+$$
new edges. We do this in an arbitrary way, and simply upper bound their length by the diameter of $Q$.
To prove the claim it is therefore sufficient to prove the following inequalities for non-negative numbers:
\begin{equation}\label{eq:xy}
(x+x_1)\wedge (y+y_1)-\big((x+x_2)\wedge (y+y_2)-x_2-y_2\big)_+\le x_1+x_2+y_1+y_2.
\end{equation}
This is obviously equivalent to
\begin{eqnarray*}
&&x+x_1 \le x_1+x_2+y_1+y_2+\big((x+x_2)\wedge (y+y_2)-x_2-y_2\big)_+\\
& \mathrm{or} & y+y_1\le x_1+x_2+y_1+y_2+ \big((x+x_2)\wedge (y+y_2)-x_2-y_2\big)_+.
\end{eqnarray*}
After simplification, and noting that $y_1\ge 0$ appears only on the right-hand side of the first inequation
(and the same for $x_1$ in the second one), it is enough to show that
$$ x\wedge y \le x_2+y_2+ \big((x+x_2)\wedge (y+y_2)-x_2-y_2\big)_+.$$
This is obvious, as by definition of the positive part, $ x\wedge y \le x_2+y_2+ \big((x\wedge y)-x_2-y_2\big)_+.$
\end{proof}
\subsection{Bipartite traveling salesperson tour}
Fix $p>0$. Given two multi-subsets of $\dR^d$ with the same cardinality, $\mathcal X=\{X_1,\ldots,X_n\}$ and
$\mathcal Y=\{Y_1,\ldots,Y_n\}$, the $p$-cost of the minimal bipartite
traveling salesperson tour of $(\mathcal X,\mathcal Y)$ is defined as
$$T_p(\mathcal X,\mathcal Y)=\min_{(\sigma,\sigma') \in S_n ^2 } \sum_{i=1}^{n} |X_{\sigma(i)} -Y_{\sigma'(i)}|^p + \sum_{i=1}^{n-1} |Y_{\sigma'(i)} -X_{\sigma(i+1)}|^p + |Y_{\sigma'(n)} -X_{\sigma(1)}|^p ,$$
where the minimum runs over all pairs of permutations of $\{1,\ldots,n\}$.
We extend the definition to sets of different cardinalities, by completing the longest possible bipartite tour : if $\mathcal
X=\{X_1,\ldots,X_m\}$ and $\mathcal Y=\{Y_1,\ldots,Y_n\}$ and $m\le
n$ then
$$ T_p(\mathcal X,\mathcal Y)=\min_{(\sigma,\sigma')} \sum_{i=1}^{m} |X_{\sigma(i)} -Y_{\sigma'(i)}|^p + \sum_{i=1}^{m-1} |Y_{\sigma'(i)} -X_{\sigma(i+1)}|^p + |Y_{\sigma'(m)} -X_{\sigma(1)}|^p $$
where the minimum runs over all pairs $(\sigma, \sigma')$, with $\sigma \in S_m$ and $\sigma'$ is an injective maps from
$\{1,\ldots,m\}$ to $\{1,\ldots,n\}$. When $n\le m$ the symmetric
definition is chosen $ T_p(\mathcal X,\mathcal Y):=T_p(\mathcal Y,\mathcal X) $. This traveling salesperson functional is an instance of a larger class of functionals that we now describe.
\subsection{Euclidean combinatorial optimization over bipartite graphs}
\label{subsec:combopt}
For integers $m,n$, we define $[n] = \{1, \cdots n\}$ and $[n]_m = \{m +1 , \cdots , m + n\}$. Let $\mathcal B_n$ be the set of bipartite graphs with common vertex set $([n], [n]_n)$ : if $G \in \mathcal B_n$, the edge set of $G$ is contained is the set of pairs $\{i , n+j\}$, with $i, j \in [n]$.
We should introduce some graph definitions. If $G_1 \in \mathcal B_n$ and $G_2 \in \mathcal B_m$ we define $G_1 + G_2$ as the graph in $\mathcal B_{n+m}$ obtained by the following rule : if $\{i , n + j\}$ is an edge of $G_1$ then $\{i , n +m+ j\}$ is an edge of $G_1 + G_2$, and if $\{i , m + j\}$ is an edge of $G_2$ then $\{n+i , 2n +m+ j\}$ is an edge of $G_1 + G_2$. Finally, if $G \in \mathcal B_{n+m}$, the restriction $G'$ of $G$ to $\mathcal B_n$ is the element of $\mathcal B_n$ defined by the following construction rule:
if $\{i , n + m + j\}$ is an edge of $G$ and $(i,j) \in [n]^2$ then add $\{i, n+ j \}$ as an edge of $G'$.
We consider a collection of subsets $\mathcal G_n \subset \mathcal B_n$ with the following properties, there exist constants $\kappa_0, \kappa \geq 1$ such that for all integers $n,m$,
\begin{itemize}
\item[(A1)] {\em (not empty)} If $n \geq \kappa_0$, $\mathcal G_n$ is not empty.
\item[(A2)]
{\em (isomorphism)} If $G \in \mathcal G_n$ and $G' \in \mathcal B_n$ is isomorphic to $G$ then $G' \in \mathcal G_n$.
\item[(A3)] {\em (bounded degree)} If $G \in \mathcal G_n$, the degree of any vertex is at most $\kappa$.
\item[(A4)]
{\em (merging)} If $G \in \mathcal G_n$ and $G' \in \mathcal G_m$, there exists $G'' \in \mathcal G_{n+m}$ such that $G + G'$ and $G''$ have all but at most $\kappa$ edges in common. For $1 \leq m < \kappa_0$, it also holds if $G'$ is the empty graph of $\mathcal B_m$.
\item[(A5)]
{\em (restriction)} Let $G \in \mathcal G_n$ and $\kappa_0+1\leq n$ and $G'$ be the restriction of $G$ to $\mathcal B_{n-1} $. Then there exists $G'' \in \mathcal G_{n-1}$ such that $G'$ and $G''$ have all but at most $\kappa$ edges in common.
\end{itemize}
If $|\mathcal X |= |\mathcal Y| = n$, we define
$$
L ( \mathcal X , \mathcal Y ) = \min_{G \in \mathcal G_n} \sum_{(i,j)\in [n]^2: \{i , n+ j \} \in G } | X_i - Y_{j} |^p .
$$
With the convention that the minimum over an empty set is $0$. Note that the isomorphism property implies that $ L ( \mathcal X , \mathcal Y ) = L ( \mathcal Y , \mathcal X )$. If $m = |\mathcal X | \leq |\mathcal Y| = n$, we define
\begin{equation}
\label{eq:combopt}
L ( \mathcal X , \mathcal Y ) = \min_{(G,\sigma)} \sum_{(i,j)\in [m]^2: \{i , m+j \} \in G } | X_i - Y_{\sigma(j)} |^p,
\end{equation}
where the miminum runs over all pairs $(G, \sigma)$, $G \in \mathcal G_m$ and $ \sigma$ is an injective maps from
$\{1,\ldots,m\}$ to $\{1,\ldots,n\}$. When $n\le m$ the symmetric
definition is chosen $ L(\mathcal X,\mathcal Y):=L(\mathcal Y,\mathcal X) $.
\medskip
The case of bipartite matchings is recovered by choosing $\mathcal G_n$ as the set of graphs in $\mathcal B_n$ where all vertices have degree $1$. We then have $\kappa_0 =1$ and $\mathcal G_n$ satisfies the merging property with $\kappa = 0$. It also satisfies the restriction property with $\kappa = 1$. The case of the traveling salesperson tour is obtained by choosing $\mathcal G_n$ as the set of connected graphs in $\mathcal B_n$ where all vertices have degree $2$, this set is non-empty for $n \geq \kappa_0 = 2$. Also this set $\mathcal G_n$ satisfies the merging property with $\kappa = 4$ (as can be checked by edge switching). The restriction property follows by merging strings into a cycle.
For the minimal bipartite spanning tree, we choose $\mathcal G_n$ as the set of connected trees of $[2n]$ in $\mathcal B_n$. It satisfies the restriction property and the merging property with $\kappa = 1$. For this choice, however, the maximal degree is not bounded uniformly in $n$. We could impose artificially this condition by defining $\mathcal G_n$ as the set of connected graphs in $\mathcal B_n$ with maximal degree bounded by $\kappa \geq 2$. We would then get the minimal bipartite spanning tree with maximal degree bounded by $\kappa $. It is not hard to verify that the corresponding functional satisfies all the above properties.
Another interesting example is the following. Fix an integer $r \geq 2$. Recall that a graph is $r$-regular if the degree of all its vertices is equal to $r$. We may define $\mathcal G_n$ as the set of $r$-regular connected graphs in $\mathcal B_n$. This set is not empty for $n \geq \kappa_ 0 = r$. It satisfies the first part of the merging property (A4) with $\kappa = 4$. Indeed, consider two $r$-regular graphs $G$, $G'$, and take any edge $e = \{x,y\} \in G$ and $e' = \{x',y'\} \in G'$. The merging property holds with $G''$, the graph obtained from $G+G'$ by switching $(e,e')$ in $(\{x,y'\},\{x',y\})$. Up to increasing the value of $\kappa$, the second part of the merging property is also satisfied. Indeed, if $n$ is large enough, it is possible to find $r m < r \kappa_0 = r^2$ edges $e_{1}, \cdots, e_{rm}$ in $G$ with no-adjacent vertices. Now, in $G''$, we add $m$ points from each species, and replace the edge $e_{r i + q} = \{x , n + y\}$, $1 \leq i \leq m$, $0 \leq q < r$, by two edges : one between $x$ and the $i$-th point of the second species, and one between $y$ and the $i$-th point of the first species. $G''$ is then a connected $r$-regular graph in $\mathcal B_{n+m}$ with all but at most $2r^2$ edges in common with $G$. Hence, by taking $\kappa$ large enough, the second part of the merging property holds.
Checking the restriction property (A5) for $r$-regular graphs requires a little more care. Let $r = \kappa_0 +1\leq n$ and consider the restriction $G_1$ of $G \in \mathcal B_n$ to $\mathcal B_{n-1}$. Our goal is to show that by modifying a small number of edges of $G_1$, one can obtained a connected $r$-regular bipartite
graph on $\mathcal B_{n-1}$. We first explain how to turn $G_1$ into a possibly non-connected $r$-regular graph. Let us observe that $G_1$ was obtained from
$G$ by deleting one vertex of each spieces and the edges to which these points belong. Hence $G_1$ has vertices of degree $r$, and vertices of degree $r-1$
($r$ blue and $r$ red vertices if the removed points did not share an edge, only $r-1$ points of each spieces if the removed points shared an edge).
In any case $G_1$ has at most $2r$ connected components and $r$ vertives of each color with one edge missing.
The simplest way to turn $G_1$ into a $r$ regular graph is to connect each blue vertex missing an edge with a red vertex missing an edge. However this is
not always possible as these vertices may already be neighbours in $G_1$ and we do not allow multiple edges. However given a red vertex $v_R$ and a blue vertex
$v_B$ of degree
$r-1$ and provided $n-1 > 2r^2$ there exists a vertex $v$ in $G_1$ which is at graph distance at least 3 from $v_B$ and $v_R$. Then open up an edge to which
$v$ belongs and connect its end-points to $v_R$ and $v_B$ while respecting the bipartite structure. In the new graph $v_B$ and $v_R$ have degree $r$.
Repeating this operation no more than $r$ times turns $G_1$ into a $r$ regular graphs with at most as many connected components (and the initial and the final
graph differ by at most $3r$ edges). Next we apply the merge operation at most $2r-1$ times in order to glue together the connected componented (this leads
to modifying at most $4(2r-1)$ edges. As a conclusion, provided we choose $\kappa_0>2r^ 2$, the restriction property holds for $\kappa=11r$.
\medskip
We now come back to the general case. From the definition, it is clear that $L$ satisfies the property $\eqref{eq:Hp}$. We are going to check that it also satisfies properties \eqref{eq:Sp} and \eqref{eq:Rp}.
\begin{lemma}\label{le:Sp}
Assume (A1-A4). For any $p >0$, the functional $L$ satisfies property \eqref{eq:Sp} with constant $C = ( 3 + \kappa_0 ) \kappa /2 $.
\end{lemma}
\begin{proof} The proof of is an extension of the proof of Lemma \ref{lem:partition}. We can assume without loss of generality $k \geq 2$. Let $\mathcal X_1, \ldots,\mathcal X_k$ and $\mathcal Y_1,\ldots,\mathcal Y_k$ be multisets in $Q = B(1/2)$. For ease of notation, let $x_i = |\mathcal X_i|$, $y_i = |\mathcal Y_i|$ and $n = \sum_{i=1}^k x_i \wedge \sum_{i=1}^k y_i $. If $n < \kappa_0$, then from the bounded degree property (A3),
$$L \Big(\bigcup_{i=1}^k\mathcal X_i , \bigcup_{i=1}^k\mathcal Y_i\Big) \leq n \kappa \leq \kappa \kappa_0. $$
If $n \geq \kappa_0$, it is enough to upper bound the cost for an element $G$ in $\mathcal G_n$. For each $1 \leq i \leq k$, if $n_i = x_i \wedge y_i \geq \kappa_0$, we consider the element $G_i$ in $\mathcal G_{n_i}$ which reaches the minimum cost of $L(\mathcal X_i, \mathcal Y_i)$. From the merging property (A4), there exists $G'$ in $\mathcal G_{\sum_{i} \ind_{n_i \geq \kappa_0} n_i}$ whose total cost is at most
$$
L' := \sum_{i} L(\mathcal X_i , \mathcal Y_i) + \kappa k .
$$
It remains at most $ \sum_{i} \kappa_0 + |x_i - y_i|$ vertices that have been left aside. The less numerous species has cardinal $ m_0 \leq m = ( \sum_{i} \kappa_0 + |x_i - y_i| ) / 2 $. If $m_0 \geq \kappa_0$, from the non-empty property (A1), there exists a graph $G'' \in \mathcal G_{m_0}$ that minimizes the cost of the vertices that have been left aside. From the merging and bounded degree properties, we get
$$
L \Big(\bigcup_{i=1}^k\mathcal X_i , \bigcup_{i=1}^k\mathcal Y_i\Big) \leq L' + \kappa + \kappa m \leq \sum_{i} L(\mathcal X_i , \mathcal Y_i) + \frac{\kappa}{2} \sum_i \left( 3 + \kappa_0 + |x_i - y_i | \right).
$$
If $ m_0 < \kappa_0$, we apply to $G'$ the merging property with the empty graph : there exists an element $G$ in $\mathcal G_n$ whose total cost is at most
$$
L \Big(\bigcup_{i=1}^k\mathcal X_i , \bigcup_{i=1}^k\mathcal Y_i\Big) \leq L' + \kappa \leq \sum_{i} L(\mathcal X_i , \mathcal Y_i) + (k +1 ) \kappa .
$$
We have proved that property \eqref{eq:Sp} is satisfied for $ C =( 3 + \kappa_0 ) \kappa /2$.
\end{proof}
\begin{lemma}
\label{le:Rp}
Assume (A1-A5). For any $p >0$, the functional $L$ satisfies property \eqref{eq:Rp} with constant $C =C(\kappa, \kappa_0) $.
\end{lemma}
\begin{proof}
Let $\mathcal{X},\mathcal{X}_1,\mathcal{X}_2,\mathcal{Y},\mathcal{Y}_1,\mathcal{Y}_2$ be finite multisets contained in $B(1/2) = Q$. Denote by $x,x_1,x_2,y,y_1,y_2$ the cardinalities of the multisets.
As a first step, let us prove that
\begin{equation}\label{eq:RpsansXY2}
L(\mathcal{X}\cup\mathcal{X}_1 ,\mathcal{Y}\cup\mathcal{Y}_1) \leq L(\mathcal{X},\mathcal{Y}) + C ( x_1 + y_1).
\end{equation}
By induction, it is enough to deal with the cases $(x_1,y_1)=(1,0)$ and $(x_1,y_1)=(0,1)$.
Because of our symmetry assumption, our task is to prove that
\begin{equation}\label{eq:RpsansY1XY2}
L(\mathcal{X}\cup\{a\} ,\mathcal{Y}) \leq L(\mathcal{X},\mathcal{Y}) + C .
\end{equation}
If $ \mathrm{card}(\mathcal{Y})\le \mathrm{card}(\mathcal{X})$, then the latter is obvious: choose an optimal graph for $L(\mathcal{X},\mathcal{Y})$ and use it to upper estimate
$ L(\mathcal{X}\cup\{a\} ,\mathcal{Y})$.
Assume on the contrary that $\mathrm{card}(\mathcal{Y})\ge \mathrm{card}(\mathcal{X})+1$. Then there exists $\mathcal{Y}'\subset \mathcal{Y}$ with $ \mathrm{card}(\mathcal{Y}')= \mathrm{card}(\mathcal{X})$ and
$L(\mathcal{X},\mathcal{Y}')$. Let $b\in \mathcal{Y}\setminus \mathcal{Y}'$. In order to establish \eqref{eq:RpsansY1XY2}, it is enough to show that
$$ L(\mathcal{X}\cup\{a\} ,\mathcal{Y}'\cup\{b\}) \leq L(\mathcal{X},\mathcal{Y}') + C ,$$
but this is just an instance of the subadditivity property. Hence \eqref{eq:RpsansXY2} is established.
\medskip
In order to prove the regularity property, it remains to show that
\begin{equation}\label{eq:RpsansXY1}
L(\mathcal{X} ,\mathcal{Y}) \leq L(\mathcal{X}\cup \mathcal{X}_2,\mathcal{Y}\cup \mathcal{Y}_2) + C ( x_2 + y_2).
\end{equation}
Again, using induction and symmetry, it is sufficient to establish
\begin{equation}\label{eq:RpsansY2XY1}
L(\mathcal{X},\mathcal{Y}) \leq L(\mathcal{X}\cup\{a\},\mathcal{Y}) + C .
\end{equation}
If $\mathrm{card}(\mathcal{X})\wedge\mathrm{card}(cY)<\kappa_0$, then by the bounded degree property $ L(\mathcal{X},\mathcal{Y}) \leq \kappa\kappa_0 \mathrm{diam}(Q)^p$ and we are done.
Assume next that $\mathrm{card}(\mathcal{X}),\mathrm{card}(\mathcal{Y})\ge \kappa_0$.
Let us consider an optimal graph for $ L(\mathcal{X}\cup\{a\},\mathcal{Y}) $. If $a$ is not a vertex of this graph (which forces
$\mathrm{card}(\mathcal{X})\ge \mathrm{card}(\mathcal{Y})$) then one can use the same graph to upper estimate $ L(\mathcal{X},\mathcal{Y})$ and obtain \eqref{eq:RpsansY2XY1}.
Assume on the contrary that $a$ is a vertex of this optimal graph. Let us distinguish two cases:
if $\mathrm{card}(\mathcal{X})\ge \mathrm{card}(\mathcal{Y})$, then in the optimal graph for $ L(\mathcal{X}\cup\{a\},\mathcal{Y}) $, at least a point $b\in \mathcal{X}$ is not used.
Consider the isomorphic graph obtained by replacing $a$ by $b$ while the other points remain fixed
(this leads to the deformation of the edges out of $a$. There are at most $\kappa$ of them by the bounded degree assumption).
This graph can be used to upper estimate $ L(\mathcal{X},\mathcal{Y})$, and gives
$$ L(\mathcal{X},\mathcal{Y})\le L(\mathcal{X}\cup\{a\},\mathcal{Y}) + \kappa \,\mathrm{diam}(Q)^p .$$
The second case is when $a$ is used but $\mathrm{card}(\mathcal{X})+1\le \mathrm{card}(\mathcal{Y})$. Actually, the optimal graph for $ L(\mathcal{X}\cup\{a\},\mathcal{Y})$
uses all the points of $\mathcal{X}\cup\{a\}$ and of a subset of same cardinality $\mathcal{Y}'\subset \mathcal{Y}$.
Choose an element $b$ in $\mathcal{Y}'$. Then $\mathcal{Y}''=\mathcal{Y}'\setminus\{b\}$ has the same cardinality as $\mathcal{X}$.
Obviously $L(\mathcal{X}\cup\{a\},\mathcal{Y})=L(\mathcal{X}\cup\{a\},\mathcal{Y}''\cup\{b\})$. Consider the corresponding optimal bipartite graph.
By the restriction property, if we erase $a$ and $b$ and their edges, we obtain a bipartite graph on $(\mathcal{X},\mathcal{Y}'')$
which differs from an admissible graph of our optimization problem by at most $\kappa$ edges. Using this new graphs
yields
$$ L(\mathcal{X},\mathcal{Y})\le \kappa\, \mathrm{diam}(Q)^p + L(\mathcal{X}\cup\{a\},\mathcal{Y}''\cup\{b\})= \kappa\, \mathrm{diam}(Q)^p + L(\mathcal{X}\cup\{a\},\mathcal{Y}).$$
This concludes the proof.\end{proof}
\section{Lower bounds, lower limits}\label{sec:lower}
\subsection{Uniform distribution on a set}
In order to motivate the sequel, we start with the simple case where
$f$ is an indicator function. The lower bound is then a direct
consequence of Theorem~\ref{th:cube-poisson} and
Theorem~\ref{th:upper}.
\begin{theorem} \label{th:lowerset}
Let $d>2p>0$. Let $L$ be a bipartite functional on $\dR^d$ satisfying the properties $(\mathcal H_p)$,
$(\mathcal R_p)$, $(\mathcal S_p)$.
Let $\Omega\subset \dR^d$ be a bounded set with positive Lebesgue measure. Then
$$ \lim_{n\to \infty} \frac{\E L(n\mathbbm 1_\Omega)}{n^{1-\frac{p}{d}}}= \beta_L \mathrm{Vol}(\Omega).$$
\end{theorem}
\begin{proof}
Theorem~\ref{th:upper} gives directly $\limsup \E
L(n\mathbbm 1_\Omega)/n^{1-\frac{p}{d}}\le \beta_L \mathrm{Vol}(\Omega)$.
By translation and dilation invariance, we may assume without loss
of generality that $\Omega\subset [0,1]^d$. Let $\Omega_c:=
[0,1]^d \setminus \Omega$. Applying Proposition~\ref{prop:partition} for
the partition $[0,1]^d=\Omega \cup \Omega_c$, gives after division
by $n^{1-p/d}$
$$ \frac{\E L\big(n\mathbbm 1_{[0,1]^d}\big)}{n^{1-\frac{p}{d}}}- \frac{\E L\big(n\mathbbm 1_{\Omega_c}\big)}{n^{1-\frac{p}{d}}}
\le \frac{\E L\big(n\mathbbm 1_{\Omega}\big)}{n^{1-\frac{p}{d}}} + 3C \mathrm{diam}([0,1]^d)
n^{\frac{p}{d}-\frac12}
\Big(\mathrm{Vol}(\Omega)^{\frac12}+\mathrm{Vol}(\Omega_c)^{\frac12}\Big).
$$
Since $d>2p$, letting $n$ go to infinity gives
\begin{eqnarray*}
\liminf_n \frac{\E L\big(n\mathbbm 1_{\Omega}\big)}{n^{1-\frac{p}{d}}} &\ge& \lim_n \frac{\E L\big(n\mathbbm 1_{[0,1]^d}\big)}{n^{1-\frac{p}{d}}}- \limsup_n \frac{\E L\big(n\mathbbm 1_{\Omega_c}\big)}{n^{1-\frac{p}{d}}}\\
&\ge&\beta_L-\beta_L \mathrm{Vol}(\Omega_c)=\beta_L \mathrm{Vol}(\Omega),
\end{eqnarray*}
where we have used Theorem~\ref{th:cube-poisson} for the limit and
Theorem~\ref{th:upper} for the upper limit.
\end{proof}
The argument of the previous proof relies on the fact that the
quantity $\lim n^{1-p/d}\E L(n\mathbbm 1_\Omega)
=\beta_L\mathrm{Vol}(\Omega)$ is in a sense additive in $\Omega$.
This line of reasoning does not pass to functions since $f\mapsto
\int f^{1-p/d}$ is additive only for functions with disjoint
supports. The lower limit result requires more work for general
densities.
\subsection{Lower limits for matchings}
In order to establish a tight estimate on the lower limit, it is
natural to try and reverse the partition inequality given in
Proposition~\ref{prop:partition}. This is usually more difficult and there does not exist a general method to perform this lower bound. We shall first restrict our attention to the case of the matching functional $M_p$ with $p >0$, we define in this subsection
$$
L = M_p.
$$
\subsubsection{Boundary functional}
Given a matching on the unit
cube, one needs to infer from it matchings on the subcubes of a dyadic
partition and to control the corresponding costs. The main difficulty
comes from the points of a subcube that are matched to points of
another subcube. In other words some links of the optimal matching
cross the boundaries of the cells. As in the book by Yukich \cite{Y},
a modified notion of the cost of a matching is used in order to
control the effects of the boundary of the cells of a partition. Our
argument is however more involved, since the good bound
\eqref{eq:good-bound} used by Yukich is not available for the
bipartite matching. \medskip
We define
$$
q = 2^{p-1} \wedge 1.
$$
Let $S\subset \mathbb R^d$ and $\varepsilon \ge 0$. Given multisets
$\mathcal{X}=\{X_1,\ldots,m\}$ and $\mathcal{Y}=\{Y_1,\ldots,Y_n\}$ included in $S$ we
define the penalized boundary-matching cost as follows
\begin{eqnarray}\label{eq:Lboundary}
\lefteqn{L_{\partial S,\varepsilon}(X_1,\ldots,X_m;Y_1,\ldots,Y_n)} \\
&=& \min_{A,B,\sigma}\left\{
\sum_{i\in A} |X_i-Y_{\sigma(i)}|^p + \sum_{i\in A^c}q \Big( d(X_i,\partial S)^p +\varepsilon^p \Big)
+\sum_{j\in B^c} q \Big( d(Y_j,\partial S)^p +\varepsilon^p \Big) \right\}, \nonumber
\end{eqnarray}
where the minimum runs over all choices of subsets
$A\subset\{1,\ldots, m\}$, $B\subset\{1,\ldots, n\}$ with the same
cardinality and all bijective maps $\sigma:A\to B$. When $\varepsilon=0$
we simply write $L_{\partial S}$. Notice that in our
definition, and contrary to the definition of optimal matching, all
points are matched even if $m\neq n$. If $\mathcal{X}$ and $\mathcal{Y}$ are
independent Poisson point processes with intensity $\nu$ supported in
$S$ and with finite total mass, we write $L_{\partial S,
\varepsilon}(\nu)$ for the random variable $L_{\partial S,
\varepsilon}(\mathcal{X},\mathcal{Y})$.
The main interest of the notion of boundary matching is that it allows
to bound from below the matching cost on a large set in terms of
contributions on cells of a partition. The following Lemma establishes a superadditive property of $L_{\partial S}$ and it can be
viewed as a counterpart to the upper bound provided by
Proposition~\ref{prop:partition}.
\begin{lemma}\label{prop:part2}
Assume $L = M_p$. Let $\nu$ be a finite measure on $\mathbb R^d$ and consider a
partition $Q=\cup_{P\in\mathcal P} P$ of a subset of $\mathbb
R^d$. Then
$$ \mathrm{diam}(Q)^p \sqrt{2\nu(\mathbb R^d)} + \E L(\nu)\ge \E L_{\partial Q} (\ind_{Q} \cdot \nu ) \geq \sum_{P\in\mathcal P} \E L_{\partial P} (\mathbbm 1_P\cdot
\nu).$$
\end{lemma}
\begin{proof}
Let $\mathcal X=\{X_1,\ldots,X_m\},\mathcal
Y=\{Y_1,\ldots,Y_n\}$ be multisets included in $Q$ and $\mathcal{X}'=\{X_{m+1},\ldots,X_{m+m'}\}$, $\mathcal{Y}'=\{Y_{n+1},\ldots,Y_{n+n'}\}$ be multisets included in $Q^c$. By considering
an optimal matching of $\mathcal{X}\cup \mathcal{X}'$ and $\mathcal{Y} \cup \mathcal{Y}'$, we have the lower bound
$$ \mathrm{diam}(Q)^p |m+m'-n - n'|+ L(\mathcal{X}\cup \mathcal{X}',\mathcal{Y}\cup \mathcal{Y}') \geq L_{\partial Q} (\mathcal{X},\mathcal{Y}). $$
Indeed, if $1 \leq i \leq m$ and a pair $(X_i, Y_{n+j})$, is matched then $|X_i - Y_{n+j}| \geq d ( X_i , \partial Q)$ and similarly for a pair $(X_{m+i}, Y_{j})$, with $1 \leq j \leq n$, $|X_{m+i} - Y_{j}| \geq d ( Y_j , \partial Q)$. The term $\mathrm{diam}(Q)^p |m+m'-n - n'| $ takes care of the points of $\mathcal{X} \cup \mathcal{Y}$ that are not matched in the optimal matching of $\mathcal{X}\cup \mathcal{X}'$ and $\mathcal{Y} \cup \mathcal{Y}'$. We apply the above inequality to $\mathcal{X}$, $\mathcal{Y}$ independent Poisson processes of intensity $\ind_Q \cdot \nu$, and $\mathcal{X}'$, $\mathcal{Y}'$, two independent Poisson processes of intensity $\ind_{Q^c} \cdot \nu$, independent of $(\mathcal{X},\mathcal{Y})$. Then $\mathcal{X} \cup \mathcal{X}'$, $\mathcal{Y} \cup \mathcal{Y}'$ are independent Poisson processes of intensity $\nu$. Taking expectation and
bounding the average of the difference of cardinalities in
the usual way, we obtain the first inequality.
Now, the second inequality will follow from the superadditive property of the boundary functional:
\begin{equation} \label{eq:SLb} L_{\partial Q} (\mathcal{X},\mathcal{Y}) \geq \sum_{P \in\mathcal P} L_{\partial P} ( \mathcal{X} \cap P,\mathcal{Y} \cap P).
\end{equation}
This is proved as follows. Let $(A,B,\sigma)$ be an optimal triplet for $L_{\partial Q} (\mathcal{X},\mathcal{Y})$:
$$
L_{\partial Q} (\mathcal{X},\mathcal{Y}) = \sum_{i\in A} |X_i-Y_{\sigma(i)}|^p + \sum_{i\in A^c}q d(X_i,\partial Q)^p
+\sum_{j\in B^c} q d(Y_j,\partial Q)^p.
$$
If $x \in Q$, we denote by $P(x)$ the unique $P \in \mathcal{P}$ that contains $x$. If $P(X_i)=P(Y_{\sigma(i)})$ we leave the term
$|X_i-Y_{\sigma(i)}|$ unchanged. On the other hand if $P(X_i)\neq P(Y_{\sigma(i)})$, from H\"older's inequality,
$$ |X_i-Y_{\sigma(i)}|^p \ge q \, d(X_i,\partial P(X_i))^p + q \, d(Y_{\sigma(i)}, \partial P(Y_{\sigma(i)}))^p .$$
Eventually, we apply the inequality
$$ d(x,\partial Q) \geq d(x,\partial P(x))$$
in order to take care of the points in $A^c \cup B^c$. Combining
these inequalities and grouping the terms according to the
cell $P\in\mathcal P$ containing the points, we obtain that
\begin{eqnarray*}
L_{\partial Q} (\mathcal{X},\mathcal{Y}) &\ge & \sum_{P\in\mathcal P}\left(\sum_{i\in A;\; X_i\in P, Y_{\sigma(i)}\in P} |X_i-Y_{\sigma(i)}| ^p + \sum_{i\in A;\; X_i\in P, Y_{\sigma(i)}\notin P}q \, d(X_i,\partial P)^p \right.\\
&& \left. + \sum_{i \in A^c;\; X_i\in P} q \, d(X_i,\partial P)^p +
\sum_{j\in B ;\; Y_j \in P,\; j\not\in \sigma(\{i; \;X_i\in P\})} q \, d(Y_j,\partial P) ^p + \sum_{j \in B^c;\; Y_j\in P} q \, d(Y_j,\partial P)^p\right)\\
&\ge & \sum_{P\in\mathcal P} L_{\partial P}(\mathcal{X}\cap P,\mathcal{Y}\cap P),
\end{eqnarray*}
and we have obtained the inequality \eqref{eq:SLb}.
\end{proof}
The next lemma will be used to reduce to uniform distributions
on squares.
\begin{lemma}\label{lem:coupling2}
Assume $L = M_p$. Let $\mu,\mu'$ be two probability measures on $\dR^d$ with supports in $Q$ and $ n >0$. Then
$$ \E L_{\partial Q}(n\mu)\le \E L_{\partial Q}(n\mu') + 4 n \,\mathrm{diam}(Q)^p\, d_\mathrm{TV}(\mu,\mu').$$
Consequently, if $f$ is a nonnegative locally integrable function on
$\dR^d$, setting $\alpha=\int_Q f/\mathrm{vol}(Q)$, it holds
$$ \E L_{\partial Q}(nf \mathbbm 1_Q) \le \E L_{\partial Q}(n\alpha \mathbbm 1_Q) + 2 n\, \mathrm{diam}(Q)^p \, \int_Q|f(x)-\alpha|\, dx.$$
\end{lemma}
\begin{proof}
The functional $ L_{\partial Q}$ satisfies a slight modification of property $(\mathcal R_p)$ : for all multisets $\mathcal X,\mathcal Y,
\mathcal X_1,\mathcal Y_1,\mathcal X_2,\mathcal Y_2$ in $Q$, it holds
\begin{equation*}
L_{\partial Q}(\mathcal X\cup \mathcal{X}_1,\mathcal{Y} \cup\mathcal{Y}_1)\le L_{\partial Q}(\mathcal X\cup \mathcal{X}_2,\mathcal{Y} \cup\mathcal{Y}_2) + \mathrm{diam}(Q) ^p \big( \mathrm{card}(\mathcal{X}_1)+
\mathrm{card}(\mathcal{X}_2)+\mathrm{card}(\mathcal{Y}_1)+\mathrm{card}(\mathcal{Y}_2)\big).
\end{equation*}
Indeed, we start from an optimal boundary matching of $L_{\partial Q}(\mathcal X\cup \mathcal{X}_2,\mathcal{Y} \cup\mathcal{Y}_2)$, we match to the boundary the points of $(\mathcal{X}, \mathcal{Y})$ that are matched to a point in $(\mathcal{X}_2, \mathcal{Y}_2)$. There are at most $\mathrm{card}(\mathcal{X}_2) + \mathrm{card}(\mathcal{Y}_2)$ such points. Finally we match all points of $(\mathcal{X}_1,\mathcal{Y}_1)$ to the boundary and we obtain a suboptimal boundary matching of $L_{\partial Q}(\mathcal X\cup \mathcal{X}_1,\mathcal{Y} \cup\mathcal{Y}_1)$. This establishes the above inequality. The statements follow then from the proofs of Proposition \ref{prop:approx} and Corollary \ref{cor:couplingaverage}.
\end{proof}
We will need an asymptotic for the boundary matching for the uniform
distribution on the unit cube. Let $Q=[0,1]^d$ and denote $$\bar L_{\partial Q}(n)=\E L_{\partial Q}(n\mathbbm 1_Q).$$
\begin{lemma}
\label{lem:liminfBM}
Assume $L = M_p$ and $0 < p<d/2$, then
$$ \lim_{n\to \infty} \frac{\bar L_{\partial Q}(n)}{n^{1-\frac{p}{d}}} = \beta'_L,$$
where $\beta'_L >0$ is a constant depending on $p$ and $d$.
\end{lemma}
\begin{proof}
Let $m\ge 1$ be an integer. We consider a dyadic partition $\mathcal P$ of $Q$ into $m^d$ cubes of size $1/m$. Then, Lemma \ref{prop:part2} applied to the measure $n\mathbbm 1_{[0,1]^d}(x) \, dx$ gives
$$\bar L_{\partial Q} (n) \geq \sum_{q\in \mathcal P} \E L_{\partial q} (n \mathbbm 1_{q\cap[0,1]^d}).$$
However by scale and translation invariance, for any $q \in \mathcal P$ we have $\E L_{\partial q} (n \mathbbm 1_{q\cap[0,1]^d}) = m^{-p} \E L_{\partial Q} (n m^{-d} \mathbbm 1_Q)$. It follows that
$$
\bar L_{\partial Q} (n) \geq m^{d-p} \bar L_{\partial Q} (n m^{-d}).
$$
The proof is then done as in Theorem \ref{th:cube-poisson} where superadditivity here replaces subadditivity there.
\end{proof}
\subsubsection{General absolutely continuous measures}
We are ready to state and prove
\begin{theorem} \label{th:lower} Assume $L = M_p$ and $0 < p<d/2$. Let $f: \dR^d \to \dR^+$
be an integrable function. Then $$ \liminf_{n}
\frac{\E L(n f )}{n^{1 - \frac{p}{d}}}\geq \beta'_L \int_{\dR^d}
f^{1 - \frac{p}{d}}.$$
\end{theorem}
\begin{proof}
Assume first that the support of $f$ is bounded.
By a scaling argument, we may assume that the support of $f$ is
included in $Q=[0,1]^d$. The proof is now similar to the one of
Theorem~\ref{th:upper}. For $\ell \in\mathbb N$, we consider the
partition $\mathcal P_\ell$ of $[0,1]^d$ into $2^{\ell d}$ cubes of
side-length $2^{-\ell}$. Let $k\in \mathbb N^*$ to be chosen later.
For $P\in \mathcal P_k$, $\alpha_P$ denotes the average of $f$ over
$P$. Applying Lemma~\ref{prop:part2}, Lemma~\ref{lem:coupling2} and
homogeneity, we obtain
\begin{eqnarray*}
2 d^{\frac p 2} \sqrt{n\int f}+\E L(nf)\; \geq \; \E L_{\partial Q} (nf) &\ge& \sum_{P\in \mathcal P_k} \E L_{\partial P} (nf \mathbbm 1_P) \\
&\ge& \sum_{P\in \mathcal P_k}\left( \E L_{\partial P} (n \alpha_P \mathbbm 1_P)- 2 n d^{\frac p 2} \, 2^{-kp} \int_{P} |f-\alpha_P| \right)\\
&=& \sum_{P\in \mathcal P_k}\left(2^{-kp} \E L_{\partial Q} (n \alpha_P 2^{-kd}\mathbbm 1_Q)- 2 nd^{\frac p 2} \, 2^{-kp} \int_{P}|f-\alpha_P| \right).
\end{eqnarray*}
Setting as before $f_k= \sum_{P\in \mathcal P_k} \alpha_P\mathbbm 1_P$ and
$h(t)=\bar L_{\partial Q} (t)/t^{\frac{d-1}{d}}$ where $\bar L_{\partial Q}(t)=\E L_{\partial Q}
(t\mathbbm 1_Q)$, the previous inequality reads as
$$ 2 n^{\frac{p}{d}-\frac12} d^{\frac p 2} \sqrt{ \int f} + \frac{\E L(nf)}{n^{1 - \frac{p}{d}}} \geq \E L_{\partial Q} (nf) \ge \int h(n2^{-kd} f_k)f_k^{1 - \frac{p}{d}}- 2 d ^{\frac p 2 }\, n^{\frac{p}{d}}2^{-k p }
\int|f-f_k|.$$ As in the proof of Theorem~\ref{th:upper} we may choose
$k=k(n)$ depending on $n$ in such a way that $\lim_n k(n)=+\infty$,
$\lim_n n^{1/d}2^{-k(n)}=+\infty$ and $\lim_n n^{\frac{1}{d}}2^{-k(n)}
\big( \int|f-f_{k(n)}| \big)^{\frac 1 p}=0$. For such a choice, since $\liminf_{t\to +\infty }
h(t)\ge\beta'_L$ by Lemma~\ref{lem:liminfBM} and a.e. $\lim_k f_k=f$,
Fatou's lemma ensures that
$$\liminf_n \int h(n2^{-k(n)d} f_{k(n)})f_{k(n)}^{1 - \frac{p}{d}} \ge \liminf_n
\int_{\{f>0\}} h(n2^{-k(n)d} f_{k(n)})f_{k(n)}^{1 - \frac{p}{d}} \ge
\beta'_L \int f^{1 - \frac{p}{d}}.$$ Our statement easily follows.
\medskip
Now, let us address the general case where the support is not bounded. Let $ \ell \geq 1$ and $Q = [-\ell, \ell]^d$. By Lemma~\ref{prop:part2},
$$
2 \mathrm{diam}(Q)^p \sqrt{ n \int f} + \E L( n f ) \geq \E L_{\partial Q}( n f \ind_Q ).
$$
Also, the above argument has shown that
$$
\liminf_{n} \frac{ \E L_{\partial Q}( n f \ind_Q ) }{n ^{1 - \frac{p}{d}}} \geq \beta'_L \int_Q f^{1 - \frac{p}{d}} .
$$
We deduce that for any $Q = [-\ell, \ell]^d$,
$$
\liminf_{n} \frac{ \E L( n f ) }{n ^{1 - \frac{p}{d}}} \geq \beta'_L \int_Q f^{1 - \frac{p}{d}} .
$$
Taking $\ell$ arbitrary large we obtain the claimed lower bound.
\end{proof}
\subsubsection{Dealing with the singular component}\label{sec:singular}
In this section we explain how to extend Theorem~\ref{th:lower} from measures with densities
to general measures. Given a measure $\mu$, we consider its decomposition $\mu=\mu_{ac}+\mu_s$
into an absolutely continuous part and a singular part.
Our starting point is the following lemma, which can be viewed as an inverse subbadditivity property.
\begin{lemma}\label{le:mino-point}
Let $p\in (0,1]$ and $L=M_p$. Let $\mathcal{X}_1,\mathcal{X}_2,\mathcal{Y}_1,\mathcal{Y}_2$ be four finite multisets included in a bounded
set $Q$. Then
$$L(\mathcal{X}_1,\mathcal{Y}_1)\le L(\mathcal{X}_1\cup\mathcal{X}_2,\mathcal{Y}_1\cup\mathcal{Y}_2)+L(\mathcal{X}_2,\mathcal{Y}_2)+\mathrm{diam}(Q)^p
\Big(|\mathcal{X}_1(Q)-\mathcal{Y}_1(Q)|+ 2 |\mathcal{X}_2(Q)-\mathcal{Y}_2(Q)|\Big).$$
\end{lemma}
\begin{proof}
Let us start with an optimal matching achieving $L(\mathcal{X}_1\cup\mathcal{X}_2,\mathcal{Y}_1\cup\mathcal{Y}_2)$ and an optimal matching achieving $L(\mathcal{X}_2,\mathcal{Y}_2)$. Let us view them as bipartite graphs $G_{1,2}$ and $G_2$ on the vertex sets $(\mathcal{X}_1\cup\mathcal{X}_2,\mathcal{Y}_1\cup\mathcal{Y}_2)$ and $(\mathcal{X}_2,\mathcal{Y}_2)$ respectively (note that if a point appears more than
once, we consider its instances as different graph vertices).
Our goal is to build a possibly suboptimal matching of $\mathcal{X}_1$ and $\mathcal{Y}_1$. Assume without loss of generality that $\mathcal{X}_1(Q)\le \mathcal{Y}_1(Q)$. Hence we need to build an injection from $\sigma:\mathcal{X}_1\to \mathcal{Y}_1$ and to upper bound
its cost $\sum_{x\in\mathcal{X}_1} |x-\sigma(x)|^p$.
To do this, let us consider the graph $G$ obtained as the union of $G_{1,2}$ and $G_2$ (allowing multiple edges when
two points are neighbours in both graphs).
It is clear that in $G$ the points from $\mathcal{X}_1$ and $\mathcal{Y}_1$ have degree at most one, while the points
from $\mathcal{X}_2$ and $\mathcal{Y}_2$ have degree at most 2.
For each $x\in \mathcal{X}_1$, let us consider its connected component $C(x)$ in $G$. Because of the above degree
considerations (and since no point is connected to itself in a bipartite graph) it is obvious that
$C(x)$ is a path.
It could be that $C(x)=\{x\}$, in the case when $x$ is a leftover point in the matching corresponding
to $G_{1,2}$. This means that $x$ is a point in excess and there are at most
$|\mathcal{X}_1(Q)+\mathcal{X}_2(Q)-(\mathcal{Y}_1(Q)+\mathcal{Y}_2(Q))|$ of them.
\begin{figure}[htb]
\begin{center}
\begin{psfrags}
\psfrag{x}{\Large $x$}
\psfrag{y}{\Large $y $}
\includegraphics[angle=0,width = 5cm]{Matchcolor.eps}
\end{psfrags}
\caption{\label{fig:matchcolor}The three possibilities for the path $C(x)$. In blue, $G_{1,2}$, in red $G_2$, the points in $\mathcal{X}_1 \cup \mathcal{X}_2$ are represented by a cross, points in $\mathcal{Y}_1 \cup \mathcal{Y}_2$ by a circle.}
\end{center}\end{figure}
Consider now the remaining case, when $C(x)$ is a non trivial path. Its first edge belongs to $G_{1,2}$. If there
is a second edge, it has to be from $G_2$ (since $G_{1,2}$ as degree at most one). Repeating the argument,
we see that the edges of the path are alternately from $G_{1,2}$ and from $G_2$. Note also
that the successive vertices are alternately from $\mathcal{X}_1\cup\mathcal{X}_2$ and from $\mathcal{Y}_1\cup\mathcal{Y}_2$ (see Figure \ref{fig:matchcolor}). There are three possibilities:
\begin{itemize}
\item The other end of the path is a point $y\in \mathcal{Y}_1$. In this case we are done, we have associated a point
$y\in \mathcal{Y}_1$ to our point $x\in \mathcal{X}_1$. By the triangle inequality and since $(a+b)^p\le a^p +b^p$ due to the
assumption $p\le 1$, $|x-y|^p$ is upper bounded by the sum of the $p$-th powers of the length of the edges
in $C(x)$.
\item The other end of the path is a point $y\in \mathcal{Y}_2$. The last edge is from $G_{1,2}$. So necessarily,
$y$ has no neighbour in $G_2$. This means that it is not matched. There are at most $|\mathcal{X}_2(Q)-\mathcal{Y}_2(Q)|$
such points in the matching $G_2$.
\item The other end of the path is a point $x'\in \mathcal{X}_2$. The last edge is from $G_{2}$. So necessarily,
$x'$ has no neighbour in $G_{1,2}$. This means that it is not matched in $G_{1,2}$. As already
mentionend there are at most $|\mathcal{X}_1(Q)+\mathcal{X}_2(Q)-(\mathcal{Y}_1(Q)+\mathcal{Y}_2(Q))|$
such points.
\end{itemize}
Eventually we have found a way to match the points from $\mathcal{X}_1$, apart maybe $|\mathcal{X}_2(Q)-\mathcal{Y}_2(Q)|+|\mathcal{X}_1(Q)+\mathcal{X}_2(Q)-(\mathcal{Y}_1(Q)+\mathcal{Y}_2(Q))|$ of them. We match the latter points arbitrarily
to (unused) points in $\mathcal{Y}_1$ and upper bound the distances between matched points by $\mathrm{diam}(Q)$.
\end{proof}
As a direct consequence, we obtain:
\begin{lemma}\label{lem:is}
Let $\mu_1$ and $\mu_2$ be two finite measures supported in a bounded set $Q$.
Let $p\in (0,1]$ and $L=M_p$ be the bipartite matching functional. Then
$$ \E L(\mu_1)\le \E L(\mu_1+\mu_2) +\E L( \mu_2)+ 3 \, \mathrm{diam}(Q)^p \big(\sqrt{\mu_1(Q)}+\sqrt{\mu_2(Q)}\big).$$
\end{lemma}
\begin{proof}
Let $\mathcal{X}_1,\mathcal{X}_2,\mathcal{Y}_1,\mathcal{Y}_2$ be four independent Poisson point processes. Assume that for $i\in\{1,2\}$,
$\mathcal{X}_i$ and $\mathcal{Y}_i$ have intensity measure $\mu_i$. Consequently $\mathcal{X}_1\cup\mathcal{X}_2$ and $\mathcal{Y}_1\cup \mathcal{Y}_2$
are independent Poisson point processes with intensity $\mu_1+\mu_2$. Applying the preceeding lemma~\ref{le:mino-point}
and taking expectations yields
$$ \E L(\mu_1)\le \E L(\mu_1+\mu_2)+\E L(\mu_2)+ 2\mathrm{diam}(Q)^p \big(\E |\mathcal{X}_1(Q)-\mathcal{Y}_1(Q)|+ \E |\mathcal{X}_2(Q)-\mathcal{Y}_2(Q)|\big).$$
As usual, we conclude using that
$$ \E |\mathcal{X}_i(Q)-\mathcal{Y}_i(Q)|\le \sqrt{\E\big( (\mathcal{X}_i(Q)-\mathcal{Y}_i(Q))^2\big)}=\sqrt{2\mathrm{var}(\mathcal{X}_i(Q))}=\sqrt{2\mu_i(Q)}.$$
\end{proof}
\begin{theorem}\label{th:poisson-general}
Assume that $d\in\{1,2\}$ and $p\in(0,d/2)$, or that $d\ge 3$ and $p\in(0,1]$.
Let $L=M_p$ be the bipartite matching functional. Let $\mu$ be a finite measure on $\dR^d$ with bounded support.
Let $f$ be the density of the absolutely continuous part of $\mu$. Assume that there exists $\alpha>\frac{2dp}{d-2p}$ such that $\int |x|^\alpha d\mu(x) <+\infty$. Then
$$\liminf_n \frac{\E L(n\mu)}{n^{1-\frac{p}{d}}} \geq \beta'_L \int_{\dR^d} f^{1-\frac{p}{d}}.$$
Moreover if $f$ is proportional to the indicator function of a bounded set with positive Lebesgue measure
$$
\lim_n \frac{\E L(n\mu)}{n^{1-\frac{p}{d}}} = \beta_L \int_{\dR^d} f^{1-\frac{p}{d}}.
$$
\end{theorem}
\begin{proof} Note that in any case, $p\le 1$ is assumed. Let us write $\mu=\mu_{ac}+\mu_s$ where $d\mu_{ac}(x)=f(x)dx$ is the absolutely continuous
part and $\mu_s$ is the singular part of $\mu$.
The argument is very simple if $\mu$ has a bounded support:
apply the previous lemma with $\mu_1=n\mu_{ac}$ and $\mu_2=n\mu_s$. When $n$ tends to infinity, observing
that $\sqrt n$ is negligible with respect to $n^{1-\frac{p}{d}}$, we obtain that
$$ \liminf_n \frac{\E L(n\mu_{ac})}{n^{1-\frac{p}{d}}}\le \liminf_n \frac{\E L(n\mu )}{n^{1-\frac{p}{d}}}+
\limsup_n \frac{\E L(n\mu_{s})}{n^{1-\frac{p}{d}}}.$$
Observe that the latter upper limit is equal to zero thanks to Theorem~\ref{th:up-poisson-general} applied to a purely singular measures. Eventually
$\liminf_n \frac{\E L(n\mu_{ac})}{n^{1-\frac{p}{d}}} \geq \beta'_L \int_{\dR^d} f^{1-\frac{p}{d}}$
by Theorem~\ref{th:lower} about absolutely continuous measures.
If $f$ is proportional to an indicator function, we simply use scale invariance and Theorem \ref{th:lowerset} in place of Theorem~\ref{th:lower}.
Let us consider the general case of unbounded support. Let $Q=[-\ell,\ell]^d$ where $\ell>0$ is arbitrary. Let $\mathcal{X}_1,\mathcal{Y}_1,\mathcal{X}_2,\mathcal{Y}_2$ be four independent Poisson point processes, such that $\mathcal{X}_1$ and $\mathcal{Y}_1$
have intensity measure $n \mathbbm 1_Q \cdot \mu_{ac}$, and $\mathcal{X}_2$ and $\mathcal{Y}_2$
have intensity measure $n (\mu_s+ 1_{Q^c}\cdot \mu_{ac})$. It follows that
$\mathcal{X}_1\cup \mathcal{X}_2$ and $\mathcal{Y}_1\cup\mathcal{Y}_2$ are independent Poisson point processes with intensity $n\mu$.
Set $T:= \max\{ |z|;\; z\in \mathcal{X}_1\cup \mathcal{X}_2\cup \mathcal{Y}_1\cup\mathcal{Y}_2\}$. Applying Lemma~\ref{le:mino-point} gives
$$ L(\mathcal{X}_1,\mathcal{Y}_1) \le L(\mathcal{X}_1\cup\mathcal{X}_2,\mathcal{Y}_1\cup\mathcal{Y}_2)+ L(\mathcal{X}_2,\mathcal{Y}_2)+c_p T^p
\big(|\mathrm{card}(\mathcal{X}_1)-\mathrm{card}(\mathcal{Y}_1)|-|\mathrm{card}(\mathcal{X}_2)-\mathrm{card}(\mathcal{Y}_2)| \big).$$
Taking expectations, applying the Cauchy-Schwarz inequality twice and Lemma~\ref{lem:ET} (note that
$\alpha>2p$) gives
\begin{eqnarray*}
\E L(n f \mathbbm 1_Q)& \le& \E L(n\mu)+ \E L\big(n(\mu_s+\mathbbm 1_{Q^c}\mu_{ac})\big)+ c_p \sqrt{E\big[T^{2p}\big]} \left(
\sqrt{2n \mu_{ac}(Q)}+\sqrt{2n (\mu_s(\dR^d)+\mu_{ac}(Q^c))} \right)\\
&\le & \E L(n\mu)+ \E L\big(n(\mu_s+\mathbbm 1_{Q^c}\mu_{ac})\big)+ c'_p n^{\frac{p}{\alpha}+\frac{1}{2}}.
\end{eqnarray*}
Since $\alpha>\frac{2dp}{d-2p}$ we obtain
$$ \liminf_n \frac{\E L(n\mu)}{n^{1-\frac{p}{d}}} \ge \liminf_n \frac{\E L(n f \mathbbm 1_Q)}{n^{1-\frac{p}{d}}}
- \limsup_n \frac{\E L(n(\mu_s+ \mathbbm 1_Q^c\cdot \mu_{ac}))}{n^{1-\frac{p}{d}}}
\ge \beta'_L \int_Q f^{1-\frac{p}{d}}- \beta_L\int_{Q^c} f^{1-\frac{p}{d}},$$
where we have used Theorem~\ref{th:lower} for the lower limit for bounded absolutely continuous measures
and Theorem~\ref{th:up-poisson-general} for the upper limit. Recall that $Q=[-\ell,\ell]^d$. It remains
to let $\ell$ tend to infinity.
\end{proof}
Actually, using classical duality techniques (which are specific to the bipartite matching) we can derive
the following improvement of Lemma~\ref{lem:is}, which can be seen as an average monotonicity property:
\begin{lemma}\label{le:mino}
Let $p\in(0,1]$ and $L=M_p$.
Let $\mu_1$ and $\mu_2$ be two finite measures supported on a bounded subset $Q\subset \dR^d$. Then
$$ \E L( \mu_1) \le \E L (\mu_1+\mu_2) +3 \mathrm{diam}(Q)^p \big( \sqrt{\mu_1(Q)}+ \sqrt{\mu_2(Q)} \big).$$
\end{lemma}
\begin{proof}
Since $p\in(0,1]$, the unit cost $c(x,y):=|x-y|^p$ is a distance on $\dR^d$. The Kantorovich-Rubinstein
dual representation of the minimal matching cost (or optimal transportation cost) is particularly simple in this case (see e.g. \cite{rachev, villani,T}): for $\{x_1,\ldots,x_n\}$, $\{y_1,\ldots,y_n\}$ two multisets in $Q$,
$$ L\big(\{x_1,\ldots,x_n\} , \{y_1,\ldots,y_n\}\big)=\sup_{f\in \mathrm{Lip}_{1,0}} \sum_i f(x_i)-f(y_i),$$
where $\mathrm{Lip}_{1,0}$ denotes the set of function $f:Q\to \dR$ which are 1-Lipschitzian for the distance
$c(x,y)$ (hence they are $p$-H\"olderian for the Euclidean distance) and vanish at a prescribed point $x_0\in Q$.
Observe that any function in $\mathrm{Lip}_{1,0}$ is bounded by $\mathrm{diam}(Q)^p$ pointwise.
Let $\mathcal{X}=\{X_1,\ldots,X_{N_1}\}$ and $\mathcal{Y}=\{Y_1,\ldots,Y_{N_2}\}$
be independent Poisson point processes with intensity $\mu$ of finite mass and supported on a set $Q$
of diameter $D<+\infty$.
By definition, on the event $\{N_1\le N_2\}$,
\begin{eqnarray*}
L(\mathcal{X},\mathcal{Y})&=& \inf_{A\subset \{1,\ldots, N_2\}; \mathrm{ card}(A)=N_1} L\big( \{X_i,1\le i\le N_1\},
\{Y_j,j\in A\}\big) \\
&=& \inf_{A\subset \{1,\ldots, N_2\}; \mathrm{ card}(A)=N_1} \sup_{f\in \mathrm{Lip}_{1,0}}
\left( \sum_{i\le N_1} f(X_i)-\sum_{j\in A} f(Y_j)\right) \\
&\ge& \sup_{f\in \mathrm{Lip}_{1,0}}
\left( \sum_{i\le N_1} f(X_i)-\sum_{j\le N_2} f(Y_j)\right) -D^p |N_1-N_2|
\end{eqnarray*}
where we have used Kantorovich-Rubinstein duality to express the optimal matching of two samples of the same size
and used that every $f\in \mathrm{Lip_{1,0}}$ satisfies $|f|\le D^p$ pointwise on $Q$.
A similar lower bound is valid when $N_1\ge N_2$. Hence, taking expectation and bounding
$E|N_1-N_2|$ in terms of the variance of the number of points in one process, one gets
\begin{equation}\label{eq:dual1}
\E L(\mu) \ge \E \sup_{f\in \mathrm{Lip}_{1,0}} \left( \sum_{i\le N_1} f(X_i)-\sum_{j\le N_2} f(Y_j)\right)
-D^p \sqrt{2 |\mu|}.
\end{equation}
A similar argument also gives the following upper bound
\begin{equation}\label{eq:dual2}
\E L(\mu) \le \E \sup_{f\in \mathrm{Lip}_{1,0}} \left( \sum_{i\le N_1} f(X_i)-\sum_{j\le N_2} f(Y_j)\right)
+D^p \sqrt{2 |\mu|}.
\end{equation}
Let $\mathcal{X}_1,\mathcal{X}_2,\mathcal{Y}_1,\mathcal{Y}_2$ be four independent Poisson point processes. Assume that for $i \in \{1,2\}$,
$\mathcal{X}_i$ and $\mathcal{Y}_i$ have intensity $ \mu_i$. As already mentioned, $\mathcal{X}_1\cup \mathcal{X}_2$ and $\mathcal{Y}_1\cup\mathcal{Y}_2$
are independent with common intensity $ \mu_1+\mu_2 $. Given a compact set $Q$ containing the
supports of both measures, and $x_0\in Q$ we define the set $\mathrm{Lip}_{1,0}$.
Using \eqref{eq:dual1},
\begin{eqnarray*}
\lefteqn{ \E L(\mu_1+\mu_2) = \E L(\mathcal{X}_1\cup \mathcal{X}_2, \mathcal{Y}_1\cup\mathcal{Y}_2)} \\
&\ge & \E \sup_{f\in \mathrm{Lip}_{1,0}} \left( \sum_{x_1\in \mathcal{X}_1} f(x_1) - \sum_{y_1\in \mathcal{Y}_1} f(y_1)
+ \sum_{x_2\in \mathcal{X}_2} f(x_2) - \sum_{y_2\in \mathcal{Y}_2} f(y_2) \right) -D^p \sqrt{2|\mu_1+\mu_2|}
\end{eqnarray*}
Now we use the easy inequality $\E \sup \ge \sup \E$ when $\E$ is
the conditional expectation given $\mathcal{X}_1,\mathcal{Y}_1$. Since $(\mathcal{X}_2,\mathcal{Y}_2)$ are independent from
$(\mathcal{X}_1,\mathcal{Y}_1)$, we obtain
\begin{eqnarray*}
\lefteqn{ \E L(\mu_1+\mu_2)+D^p \sqrt{2|\mu_1+\mu_2|}}\\
&\ge& \E \sup_{f\in \mathrm{Lip}_{1,0}} \left( \sum_{x_1\in \mathcal{X}_1} f(x_1) - \sum_{y_1\in \mathcal{Y}_1} f(y_1)
+\E\Big( \sum_{x_2\in \mathcal{X}_2} f(x_2) - \sum_{y_2\in \mathcal{Y}_2} f(y_2) \Big) \right)\\
&=& \E \sup_{f\in \mathrm{Lip}_{1,0}} \left( \sum_{x_1\in \mathcal{X}_1} f(x_1) - \sum_{y_1\in \mathcal{Y}_1} f(y_1)
\right)\\
&\ge & \E L(\mu_1) - D^p \sqrt{2 |\mu_1|},
\end{eqnarray*}
where we have noted that the inner expectation vanishes and used \eqref{eq:dual2}.
The claim easily follows. \end{proof}
\subsection{Euclidean combinatorial optimization}
Our proof for the lower bound for matchings extends to some combinatorial optimization functionals $L$ defined by \eqref{eq:combopt}. In this paragraph, we explain how to adapt the above argument at the cost of ad-hoc assumptions on the collection of graphs $(\mathcal G_n)_{n \in \mathbb N}$. As motivating example, we will treat completely the case of the bipartite traveling salesperson tour.
\subsubsection{Boundary functional}
\label{subsubsec:BFECO}
Let $S \subset \mathbb R^d$ and $\varepsilon , p\geq 0$. Set $q=2^{p-1}\wedge 1$. In what follows, $p$ is fixed and will be omitted in most places
where it would appear as an index. Given multisets
$\mathcal{X}=\{X_1,\ldots,X_n\}$ and $\mathcal{Y}=\{Y_1,\ldots,Y_n\}$ included in $\mathbb R^d$, we first set
$$
L^0_{\partial S,\varepsilon}(\mathcal{X} , \mathcal{Y})
= \min_{G \in \mathcal G_n}\left\{
\sum_{(i,j)\in [n]^2 : \{i , n+j \} \in G } d_{S,\varepsilon,p}(X_i,Y_{j})
\right\},
$$
where
\begin{equation}
d_{S,\varepsilon,p}(x,y) =\left\{
\begin{array}{ccc}
|x-y|^p & \mathrm{if}& x,y\in S,\\
0 & \mathrm{if}& x,y\not\in S,\\
q \big(\mathrm{dist}(x,S^c)^p+\varepsilon^ p\big) & \mathrm{if}& x\in S,\, y\not\in S\\
q \big(\mathrm{dist}(y,S^c)^p+\varepsilon^ p\big) & \mathrm{if}& y\in S,\, x\not\in S
\end{array}
\right.
\end{equation}
Now, if $\mathcal{X}$ and $\mathcal{Y}$ are in $S$, we define the penalized boundary functional as
\begin{equation}
L_{\partial S,\varepsilon}(\mathcal{X} , \mathcal{Y}) = \min_{ A , B \subset S^c }{L^0_{\partial S,\varepsilon}(\mathcal{X} \cup A , \mathcal{Y} \cup B)} , \label{eq:LLboundary}
\end{equation}
where the minimum is over all multisets $A$ and $B$ in $S^c$ such that $\mathrm{card}(\mathcal{X}\cup A) = \mathrm{card}(\mathcal{Y}\cup B) \geq \kappa_0$.
When $\varepsilon=0$ we simply write $L_{\partial S}$. The main idea of this definition is to consider all possible configurations outside
the set $S$ but not to count the distances outside of $S$ (from a metric view point, all of $S^c$ is identified to a point which is at distance
$\varepsilon$ from $S$).
The existence of the minimum in \eqref{eq:LLboundary} is due to the fact that $L^0_{\partial S}(\mathcal{X} \cup A , \mathcal{Y} \cup B)$ can only take finitely many values less than any positive value (the quantities involved are just sums of distances between points of $\mathcal{X},\mathcal{Y}$ and of their distances
to $S^c$). Notice that definition \eqref{eq:LLboundary} is consistent with the definition of the boundary functional for the matching functional $M_p$, given by \eqref{eq:Lboundary}. If $\mathcal{X}$ and $\mathcal{Y}$ are
independent Poisson point processes with intensity $\nu$ supported in
$S$ and with finite total mass, we write $L_{\partial S,
\varepsilon}(\nu)$ for the random variable $L_{\partial S,
\varepsilon}(\mathcal{X},\mathcal{Y})$.
Also note that $d_{S,0,p}(x,y)\le |x-y|^p$. Consequently if $\mathrm{card}(\mathcal{X})=\mathrm{card}(\mathcal{Y})$ then
\begin{equation}\label{eq:LLd}
L^0_{\partial S}(\mathcal{X},\mathcal{Y})\le L(\mathcal{X},\mathcal{Y}).
\end{equation}
The next lemma will be used to reduce to uniform distributions on squares.
\begin{lemma}\label{lem:Lcoupling2}
Assume (A1-A5). Let $\mu,\mu'$ be two probability measures on $\dR^d$ with supports in $Q$ and $ n >0$. Then, for some constant $c$ depending only on $\kappa, \kappa_0$,
$$ \E L_{\partial Q}(n\mu)\le \E L_{\partial Q}(n\mu') + 2 c n \,\mathrm{diam}(Q)^p\, d_\mathrm{TV}(\mu,\mu').$$
Consequently, if $f$ is a non-negative locally integrable function on
$\dR^d$, setting $\alpha=\int_Q f/\mathrm{vol}(Q)$, it holds
$$ \E L_{\partial Q}(nf \mathbbm 1_Q) \le \E L_{\partial Q}(n\alpha \mathbbm 1_Q) + c n\, \mathrm{diam}(Q)^p \, \int_Q|f(x)-\alpha|\, dx.$$
\end{lemma}
\begin{proof}
The functional $ L_{\partial Q}$ satisfies a slight modification of property $(\mathcal R_p)$ : for all multisets $\mathcal X,\mathcal Y,
\mathcal X_1,\mathcal Y_1,\mathcal X_2,\mathcal Y_2$ in $Q$, it holds
\begin{equation}
L_{\partial Q}(\mathcal X\cup \mathcal{X}_1,\mathcal{Y} \cup\mathcal{Y}_1)\le L_{\partial Q}(\mathcal X\cup \mathcal{X}_2,\mathcal{Y} \cup\mathcal{Y}_2) + C \mathrm{diam}(Q) ^p \big( \mathrm{card}(\mathcal{X}_1)+
\mathrm{card}(\mathcal{X}_2)+\mathrm{card}(\mathcal{Y}_1)+\mathrm{card}(\mathcal{Y}_2)\big), \label{eq:regLbound}
\end{equation}
with $C = C(\kappa , \kappa_0 )$. The above inequality is established as in the proof of Lemma \ref{le:Rp}. Indeed, by linearity and symmetry we should check \eqref{eq:RpsansY1XY2} and \eqref{eq:RpsansY2XY1} for $L_{\partial Q}$. To prove \eqref{eq:RpsansY1XY2}, we consider an optimal triplet $(G,A,B)$ for $(\mathcal{X}, \mathcal{Y})$ and apply the merging property (A4) to $G$ with the empty graph and $m=1$ : we obtain a graph $G''$ and get a triplet $(G'',A,B \cup \{b\})$ for $(\mathcal{X} \cup \{a\}, \mathcal{Y})$, where $b$ is any point in $\partial Q$. To prove \eqref{eq:RpsansY2XY1}, we now consider an optimal triplet $(G,A,B)$ for $(\mathcal{X} \cup \{a\}, \mathcal{Y})$ and move the point $a$ to the $a'$ in $\partial Q$ in order to obtain a triplet $(G,A \cup\{a'\},B)$ for $(\mathcal{X}, \mathcal{Y})$.
With \eqref{eq:regLbound} at hand, the statements follow from the proofs of Proposition \ref{prop:approx} and Corollary \ref{cor:couplingaverage}.
\end{proof}
The next lemma gives a lower bound on $L$ in terms of its boundary functional and states an important superadditive property of $L_{\partial S}$.
\begin{lemma}\label{prop:Lpart2}
Assume (A1-A5). Let $\nu$ be a finite measure on $\mathbb R^d$ and consider a
partition $Q=\cup_{P\in\mathcal P} P$ of a bounded subset of $\mathbb
R^d$. Then, if $c = 4 \kappa ( 1 + \kappa_0)$, we have
\begin{eqnarray*}
c \sqrt{\nu(\mathbb R^d)} \, \mathrm{diam}(Q)^p + \E L( \nu) \ge \E L_{\partial Q} (\mathbbm 1_Q \cdot \nu) \geq \sum_{P\in\mathcal P} \E L_{\partial P} (\mathbbm 1_P\cdot \nu).
\end{eqnarray*}
\end{lemma}
\begin{proof}
We start with the first inequality. Let $\mathcal X=\{X_1,\ldots,X_m\},\mathcal
Y=\{Y_1,\ldots,Y_n\}$ be multisets included in $Q$ and $\mathcal{X}'=\{X_{m+1},\ldots,X_{m+m'}\}$, $\mathcal{Y}'=\{Y_{n+1},\ldots,Y_{n+n'}\}$ be multisets included in $Q^c$. First, let us show that
\begin{equation}\label{eq:LRbQ}
c_1 |m+m'-n - n'| \mathrm{diam}(Q)^p + L(\mathcal{X}\cup \mathcal{X}',\mathcal{Y}\cup \mathcal{Y}') \geq L_{\partial Q} (\mathcal{X},\mathcal{Y}),
\end{equation}
with $c_1 = \kappa ( 1 + \kappa_0)$. To do so, let us consider an optimal graph $G$ for $L(\mathcal{X}\cup \mathcal{X}',\mathcal{Y}\cup \mathcal{Y}')$.
It uses all the points but $|m+m'-n-n'|$ points in excess.
We consider the subsets $\mathcal{X}_0 \subset \mathcal{X}$ and $\mathcal{Y}_0 \subset \mathcal{Y}$ of points that are used in $G$ and belong to $Q$.
By definition there exist subsets $A,B\subset Q^ c$ such that $\mathrm{card}(\mathcal{X}_0\cup A)=\mathrm{card}(\mathcal{Y}_0\cup B)$ and
$L(\mathcal{X}\cup \mathcal{X}',\mathcal{Y}\cup \mathcal{Y}')=L(\mathcal{X}_0\cup A,\mathcal{Y}_0\cup B)$.
By definition of the boundary functional and using \eqref{eq:LLd},
$$ L_{\partial Q}(\mathcal{X}_0,\mathcal{Y}_0)\le L^0_{\partial Q}(\mathcal{X}_0\cup A,\mathcal{Y}_0\cup B) \le L (\mathcal{X}_0\cup A,\mathcal{Y}_0\cup B)= L(\mathcal{X}\cup \mathcal{X}',\mathcal{Y}\cup \mathcal{Y}').$$
Finally, since there are at most $|n +n' - m - m'|$ points in $\mathcal{X}\cup \mathcal{Y}$ which are not in $\mathcal{X}_0\cup \mathcal{Y}_0$ (i.e. points of $Q$ not used for the optimal
$G$), the modified \eqref{eq:Rp} property
given by Equation \eqref{eq:regLbound} yields \eqref{eq:LRbQ}.
We apply the latter inequality to $\mathcal{X}$, $\mathcal{Y}$ independent Poisson processes of intensity $\ind_Q \cdot \nu$, and $\mathcal{X}'$, $\mathcal{Y}'$, two independent Poisson processes of intensity $\ind_{Q^c} \cdot \nu$, independent of $(\mathcal{X},\mathcal{Y})$. Then $\mathcal{X} \cup \mathcal{X}'$, $\mathcal{Y} \cup \mathcal{Y}'$ are independent Poisson processes of intensity $\nu$. Taking expectation, we obtain the first inequality, with $c = 4 c_1$.
We now prove the second inequality. As above, let $\mathcal X=\{X_1,\ldots,X_m\},\mathcal
Y=\{Y_1,\ldots,Y_n\}$ be multisets included in $Q$. Let $G \in \mathcal G_k$ be an optimal graph for $L_{\partial Q}( \mathcal{X}, \mathcal{Y})$ and $A = \{X_{m+1}, \cdots , X_{k}\}$, $B = \{Y_{n+1}, \cdots , Y_{k}\}$ be optimal sets in $ Q^ c$.
Given this graph $G$ and a set $S$, we denote by $E^0_S$ the set of edges $\{i,k+j\}$ of $G$ such that $X_i \in S$ and $Y_j \in S$, by $E^1_S$ the set of edges $\{i,k+j\}$ of $G$ such that $X_i \in S$ and $Y_j \in S^c$, and by $E^2_S$ the set of edges $\{i,k+j\}$ of $G$ such that $X_i \in S^c$ and $Y_j \in S$. Then by definition of the boundary functional
\begin{eqnarray*}
L_{\partial Q} ( \mathcal{X} , \mathcal{Y} ) & =& L^0_{\partial Q} ( \mathcal{X} \cup A , \mathcal{Y} \cup B) \\
&=& \sum_{\{ i , k+j\} \in E^0_Q} |X_i-Y_{j}| ^p + \sum_{ \{ i , k+j\} \in E^1_Q} q \, d(X_i,Q^c)^p +
\sum_{ \{ i , k+j\} \in E^2_Q} q\, d(Y_j,Q^c) ^p .
\end{eqnarray*}
Next, we bound these sums from below by considering the cells of the partition $\mathcal P$.
If $x \in Q$, we denote by $P(x)$ the unique $P \in \mathcal P$ that contains $x$.
If an edge $e=\{i , k+j \} \in G $ is such that
$X_i,Y_j$ belong to the same cell $P$, we observe that $e\in E^0 _P$ and we leave the quantity $|X_i-Y_j|^p$ unchanged.
If on the contrary, $X_i$ and $Y_j$ belong to different cells,
from H\"older inequality,
$$ |X_i-Y_{j}|^p \ge q \, d(X_i, P(X_i)^c)^p + q \, d(Y_j, P(Y_{j})^c)^p .$$
Eventually, for any boundary edge in $E^1_Q$, we lower bound the contribution $d(X_i,Q^c)^p$ by $d(X_i,P(X_i)^c)^p$ and we do the same for
$E^2_Q$.
Combining these inequalities and grouping the terms according to the
cell $P\in\mathcal P$ to which the points belong,
\begin{eqnarray*}
L_{\partial Q} ( \mathcal{X} , \mathcal{Y} ) &\ge & \sum_{P\in\mathcal P}\left( \sum_{\{ i , k+j\} \in E^0_P} |X_i-Y_{j}| ^p + \sum_{ \{ i , k+j\} \in E^1_P} q \, d(X_i,\partial P)^p +
\sum_{ \{ i , k+j\} \in E^2_P} q\, d(Y_j,\partial P) ^p \right).
\end{eqnarray*}
For a given cell $P$, set $A' = ( \mathcal{X} \cup A) \cap P^c$ and $B' = ( \mathcal{Y} \cup B) \cap P^c$. We get
\begin{eqnarray*}
&& \sum_{\{ i , k+j\} \in E^0_P} |X_i-Y_{j}| ^p + \sum_{ \{ i , k+j\} \in E^1_P} q \, d(X_i,\partial P)^p +
\sum_{ \{ i , k+j\} \in E^2_P} q\, d(Y_j,\partial P) ^p \\
& = & L^0_{P^c} ( ( \mathcal{X} \cap P ) \cup A' , ( \mathcal{Y} \cap P) \cup B' )
\geq L_{\partial P} ( \mathcal{X} \cap P , \mathcal{Y} \cap P ). \end{eqnarray*}
So applying these inequalities to $\mathcal{X}$ and $\mathcal{Y}$ two independent Poisson point processes with intensity $\nu \ind_Q$ and taking expectation, we obtain the claim.
\end{proof}
Let $Q=[0,1]^d$ and denote $$\bar L_{\partial Q}(n)=\E L_{\partial Q}(n\mathbbm 1_Q).$$
\begin{lemma}
\label{lem:LliminfBM}
Assume (A1-A5). Let $Q\subset \dR^d$ be a cube of side-length 1.
If $0< 2p < d$, then
$$ \lim_{n\to \infty} \frac{\bar L_{\partial Q}(n)}{n^{1-\frac{p}{d}}} = \beta'_L,$$
where $\beta'_L >0$ is a constant depending on $L$, $p$ and $d$.
\end{lemma}
\begin{proof}
The proof is the same than the proof of Lemma \ref{lem:liminfBM}, with Lemma \ref{prop:Lpart2} replacing Lemma \ref{prop:part2}.
\end{proof}
\subsubsection{General absolutely continuous measures with unbounded support}
\begin{theorem} \label{th:Llower} Assume (A1-A5) and that $0 < 2 p < d$. Let $f: \dR^d \to \dR^+$
be an integrable function. Then $$ \liminf_{n}
\frac{\E L(n f )}{n^{1 - \frac{p}{d}}} \geq \beta'_L \int_{\dR^d}
f^{1 - \frac{p}{d}}.$$
\end{theorem}
\begin{proof}
The proof is now formally the same than the proof of Theorem \ref{th:lower}, invoking Lemmas \ref{lem:Lcoupling2}, \ref{prop:Lpart2} and \ref{lem:LliminfBM} in place of Lemmas \ref{lem:coupling2}, \ref{prop:part2} and \ref{lem:liminfBM} respectively.
\end{proof}
\begin{remark} Finding good lower bounds for a general bipartite functional $L$ on $\mathbb R^d$ satisfying the properties $(\mathcal H_p)$, $(\mathcal R_p)$, $(\mathcal S_p)$ could be significantly more difficult. It is far from obvious to define a proper boundary functional $L_{\partial Q}$ at this level of generality. However if there exists a bipartite functional $L_{\partial Q}$ on $\mathbb R^d$ indexed on sets $Q \subset \dR^d$ such that for any $t >0$, $\E L_{\partial (t Q) } (n \ind_{t Q}) = t^{p} \E L_{\partial Q} (n t^d \ind_{Q})$ and such that the statements of Lemmas \ref{prop:Lpart2}, \ref{lem:Lcoupling2}, \ref{lem:LliminfBM} hold, then the statement of Theorem \ref{th:lower} also holds for the functional $L$. Thus, the caveat of this kind of techniques lies in the good definition of a boundary functional $L_{\partial Q}$.
\end{remark}
\subsubsection{Dealing with the singular component. Example of the traveling salesperson problem.}
Let $p\in(0,1]$. We shall say that a bipartite functional $L$ on $\dR^d$ satisfies the inverse subadditivity property
$(\mathcal I_p)$ if there is a constant $C$ such that for all finite multisets $\mathcal{X}_1,\mathcal{Y}_1,\mathcal{X}_2,\mathcal{Y}_2$ included in a bounded set $Q\subset \dR^d$,
$$ L(\mathcal{X}_1,\mathcal{Y}_1) \le L(\mathcal{X}_1\cup\mathcal{X}_2,\mathcal{Y}_1 \cup\mathcal{Y}_2)+ L(\mathcal{X}_2,\mathcal{Y}_2) + C \mathrm{diam}(Q)^p \big(1+ |\mathcal{X}_1(Q)-\mathcal{Y}_1(Q)|
+ |\mathcal{X}_2(Q)-\mathcal{Y}_2(Q)|\big).$$
Although it makes sense for all $p$, we have been able to check this property on examples only for
$p\in(0,1]$. Also we could have added a constant in front of $L(\mathcal{X}_2,\mathcal{Y}_2)$.
It is plain that the argument of Section~\ref{sec:singular} readily adapts to a functional satisfying
$(\mathcal I_p)$, for which one already knows a general upper limit result and a limit result for
absolutely continuous laws. It therefore provides a limit result for general laws. In the remainder of this section, we show that the traveling salesperson bipartite tour functional $L=T_p$,
$p \in(0, 1]$ enjoys the inverse subadditivity property.
This allows to prove the following result:
\begin{theorem} \label{th:bTSP} Assume that either $d\in \{1,2\}$ and $0 < 2 p < d$, or $d\ge 3$ and $p\in (0,1]$.
Let $L = T_p$ be the traveling salesperson bipartite tour functional. Let $\mu$ be a finite measure such that for some $\alpha > \frac{2dp}{ d -2p}$,
$
\int |x|^{\alpha} d \mu < +\infty.
$
Then, if $f$ is a density function for the absolutely continuous part of $\mu$,
$$ \liminf_{n}
\frac{\E L(n \mu )}{n^{1 - \frac{p}{d}}} \geq \beta'_L \int_{\dR^d}
f^{1 - \frac{p}{d}}.$$
Moreover if $f$ is proportional to the indicator function of a bounded set with positive Lebesgue measure
$$
\lim_n \frac{\E L(n\mu)}{n^{1-\frac{p}{d}}} = \beta_L \int_{\dR^d} f^{1-\frac{p}{d}}.
$$
\end{theorem}
All we have to do is to check property $(\mathcal I_p)$. More precisely:
\begin{lemma}\label{le:euler}
Assume $p\in(0,1]$ and $L = T_p$. For any set $\mathcal{X}_1, \mathcal{X}_2, \mathcal{Y}_1, \mathcal{Y}_2$ in a bounded set $Q$
\begin{eqnarray*}
L ( \mathcal{X}_1 , \mathcal{Y}_1 ) & \leq & L ( \mathcal{X}_1 \cup \mathcal{X}_2 , \mathcal{Y}_1 \cup \mathcal{Y}_2 ) + L ( \mathcal{X}_2, \mathcal{Y}_2 ) \\
& & \quad + \; 2 \, \mathrm{diam}(Q)^p\left( 1 + | \mathrm{\mathrm{card}}(\mathcal{X}_1)- \mathrm{\mathrm{card}}(\mathcal{Y}_1)|+ | \mathrm{\mathrm{card}}(\mathcal{X}_2) - \mathrm{\mathrm{card}}(\mathcal{Y}_2)|\right).
\end{eqnarray*}
\end{lemma}
\begin{proof}
We may assume without loss of generality that $\mathrm{\mathrm{card}} ( \mathcal{X}_1) \wedge \mathrm{\mathrm{card}}(\mathcal{Y}_1) \geq 2$, otherwise, $L ( \mathcal{X}_1, \mathcal{Y}_1 ) = 0$ and there is nothing to prove. Consider an optimal cycle $G_{1,2}$ for $L(\mathcal{X}_1 \cup \mathcal{X}_2 , \mathcal{Y}_1 \cup \mathcal{Y}_2)$. In $G_{1,2}$, $m = | \mathrm{\mathrm{card}}(\mathcal{X}_1)+ \mathrm{\mathrm{card}}(\mathcal{Y}_1) - \mathrm{\mathrm{card}}(\mathcal{X}_2) - \mathrm{\mathrm{card}}(\mathcal{Y}_2)| \leq | \mathrm{\mathrm{card}}(\mathcal{X}_1)- \mathrm{\mathrm{card}}(\mathcal{Y}_1)|+ | \mathrm{\mathrm{card}}(\mathcal{X}_2) - \mathrm{\mathrm{card}}(\mathcal{Y}_2)|$ points have been left aside. We shall build a bipartite tour $G_1$ on $(\mathcal{X}'_1, \mathcal{Y}'_1)$, the points of $(\mathcal{X}_1, \mathcal{Y}_1)$ that have not been left aside by $G_{1,2}$.
\begin{figure}[htb]
\begin{center}
\includegraphics[angle=0,width = 6cm]{TSPcolor.eps}
\caption{\label{fig:matchcolor} In blue, the oriented cycle $G_{1,2}$, in red $G_2$, in black $G'_{1,2}$. The points in $\mathcal{X}_1 \cup \mathcal{X}_2$ are represented by a cross, points in $\mathcal{Y}_1 \cup \mathcal{Y}_2$ by a circle.}
\end{center}\end{figure}
We consider an optimal cycle $G_2$ for $L( \mathcal{X}'_2 , \mathcal{Y}'_2)$, where $(\mathcal{X}'_2, \mathcal{Y}'_2)$ are the points of $(\mathcal{X}_2, \mathcal{Y}_2)$ that have not been left aside by $G_{1,2}$. We define $(\mathcal{X}''_2, \mathcal{Y}''_2) \subset (\mathcal{X}'_2, \mathcal{Y}'_2)$ as the sets of points that are in $G_2$. Since $\mathrm{\mathrm{card}} ( \mathcal{X}'_1) + \mathrm{\mathrm{card}} ( \mathcal{X}'_2) = \mathrm{\mathrm{card}} ( \mathcal{Y}'_1) + \mathrm{\mathrm{card}} ( \mathcal{Y}'_2)$, we get $\mathrm{\mathrm{card}} ( \mathcal{X}'_1) - \mathrm{\mathrm{card}} ( \mathcal{Y}'_1) = - \mathrm{\mathrm{card}} ( \mathcal{X}'_2) + \mathrm{\mathrm{card}} ( \mathcal{Y}'_2)$. It implies that the same number of points from the opposite type need to be removed in $( \mathcal{X}'_1, \mathcal{Y}'_1)$ and $( \mathcal{X}'_2 , \mathcal{Y}'_2)$ in order to build a bipartite tour. We fix an orientation on $G_{1,2}$. Assume for example that $\mathrm{\mathrm{card}} ( \mathcal{X}'_2) \geq \mathrm{\mathrm{card}} ( \mathcal{Y}'_2)$, if a point $x \in \mathcal{X}'_2 \backslash \mathcal{X}''_2$, we then remove the next point $y$ on the oriented cycle $G_{1,2}$ of $\mathcal{Y}'_1$. Doing so, this defines a couple of sets $(\mathcal{X}''_1, \mathcal{Y}''_1) \subset (\mathcal{X}'_1, \mathcal{Y}'_1)$ of cardinality $\mathrm{\mathrm{card}} ( \mathcal{X}'_1) \wedge \mathrm{\mathrm{card}}(\mathcal{Y}'_1)$ and
$$
L (\mathcal{X}'_1, \mathcal{Y}'_1) \leq L ( \mathcal{X}''_1, \mathcal{Y}''_1).
$$
We define $G'_{1,2}$ as the cycle on $(\mathcal{X}''_1 \cup \mathcal{X}''_2, \mathcal{Y}''_1\cup \mathcal{Y}''_2)$ obtained from $G_{1,2}$ by saying that the point after $x \in(\mathcal{X}''_1 \cup \mathcal{X}''_2, \mathcal{Y}''_1\cup \mathcal{Y}''_2)$ in the oriented cycle $G'_{1,2}$ is the next point $y \in(\mathcal{X}''_1 \cup \mathcal{X}''_2, \mathcal{Y}''_1\cup \mathcal{Y}''_2)$ in $G_{1,2}$. By construction, $G'_{1,2}$ is a bipartite cycle. Also, since $p \in (0, 1]$, we may use the triangle inequality : the distance between two successive points in the circuit $G'_{1,2}$ is bounded by the sum of the length of the intermediary edges in $G_{1,2}$. We get
$$
L (\mathcal{X}''_1 \cup \mathcal{X}''_2, \mathcal{Y}''_1\cup \mathcal{Y}''_2) \leq L(\mathcal{X}_1 \cup \mathcal{X}_2 , \mathcal{Y}_1 \cup \mathcal{Y}_2).
$$
Now consider the (multi) graph $G = G'_{1,2} \cup G_2$ obtained by adding all edges of $G'_{1,2}$ and $G_2$. This graph is bipartite, connected, and points in $(\mathcal{X}''_1, \mathcal{Y}''_1)$ have degree $2$ while those in $(\mathcal{X}''_2, \mathcal{Y}''_2)$ have degree $4$. Let $k$ be the number of edges in $G$, we recall that an eulerian circuit in $G$ is a sequence $E = (e_1, \cdots, e_{k})$ of adjacent edges in $G$ such that $e_{k}$ is also adjacent to $e_1$ and all edges of $G$ appears exactly once in the sequence $E$. By the Euler's circuit theorem, there exists an eulerian circuit in $G$. Moreover, this eulerian circuit can be chosen so that if $e_i = \{u_{i-1}, u_i\} \in G_2$ then $e_{i+1} = \{u_{i+1} , u_i\} \in G'_{1,2}$ with the convention that $e_{k+1} =e_1$.
This sequence $E$ defines an oriented circuit of points. Now we define an oriented circuit on $(\mathcal{X}''_1, \mathcal{Y}''_1)$, by connecting a point $x$ of $(\mathcal{X}''_1, \mathcal{Y}''_1)$ to the next point $y$ in $(\mathcal{X}''_1, \mathcal{Y}''_1)$ visited by the oriented circuit $E$. Due to the property that $e_i \in G_2$ implies $e_{i+1} \in G$, if $x \in \mathcal{X}''_1$ then $y \in \mathcal{Y}''_1$ and conversely, if $x \in \mathcal{Y}''_1$ then $y \in \mathcal{X}''_1$. Hence, this oriented circuit defines a bipartite cycle $G_1$ in $(\mathcal{X}''_1, \mathcal{Y}''_1)$.
By the triangle inequality, the distance between two successive points in the circuit $G_1$ is bounded by the sum of the length of the intermediary edges in $E$. Since each edge of $G$ appears exactly once in $E$, it follows that
$$
L( \mathcal{X}'_1, \mathcal{Y}'_1) \leq L(\mathcal{X}_1 \cup \mathcal{X}_2 , \mathcal{Y}_1 \cup \mathcal{Y}_2) + L(\mathcal{X}'_2 , \mathcal{Y}'_2).
$$
To conclude, we merge arbitrarily to the cycle $G_1$ the remaining points of $(\mathcal{X}_1,\mathcal{Y}_1)$, there are at most $m$ of them (regularity $(\mathcal R_p)$ property). \end{proof}
\section{Variants and final comments}\label{sec:final}
As a conclusion, we briefly discuss variants and possible extensions of Theorem~\ref{th:mainMp}. For $d > 2p$ and when $\mu$ is
the uniform distribution on the cube $[0,1]^d$, there exists a
constant $\beta_p(d) >0$ such that almost surely
$$ \lim_{n\to \infty } n^{\frac{p}{d}-1} M_p \big(\{X_1,\ldots,X_n\}, \{Y_1,\ldots,Y_n\}\big)=\beta_p(d).$$
A natural question is to understand what happens below the critical line $d = 2p$, i.e. when $d \leq 2p$. For example for $d =2$ and $p =1$, a similar convergence is also expected in dimension 2 with scaling $\sqrt{n \ln n}$, but this is a
difficult open problem. The main result in this direction goes back
to Ajtai, Koml\'os and Tusn\'ady \cite{AKT}. See also the improved upper bound of Talagrand and Yukich in \cite{TY}. In dimension 1, there is
no such stabilization to a constant.
Recall that $$\left(\frac 1 n M_p \big( \{X_i \}_{i = 1}^n , \{Y_i \}_{i = 1}^n \big) \right)^{\frac 1 p} = W_p \left( \frac 1 n \sum_{i =1}^n \delta_{X_i} , \frac 1 n \sum_{i =1}^n \delta_{Y_i} \right).$$
where $W_p$ is the $L_p$-Wasserstein distance. A variant of Theorem~\ref{th:mainup} can be obtained along the same lines,
concerning the convergence of
$$ n^{\frac{1}{d}} W_p \left( \frac{ 1 } {n }\sum_{i=1}^{n} \delta_{X_i},\, \mu\right),$$
where $\mu$ is the common distribution of the $X_i$'s. Such results are of fundamental
importance in statistics. Also note that combining the triangle inequality and Jensen
inequality, it is not hard to see that
$$ \E W_1 \Big( \frac1n \sum_{i=1}^{n} \delta_{X_i},\,\mu \Big)
\leq \E W_1\Big( \frac1n \sum_{i=1}^{n} \delta_{X_i},\,\frac 1 n \sum_{i =1}^n \delta_{Y_i} \Big)
\le 2 \E W_1\Big( \frac1n \sum_{i=1}^{n} \delta_{X_i},\, \mu \Big),$$
(similar inequalities hold for $p \geq 1$). Hence it is clear
that the behaviour of this functional is quite close to the one
of the two-sample optimal matching. However, the extension of Theorem~\ref{th:mainMp} would require some care in the definition of the boundary functional.
Finally, it is worthy to note that the case of uniform distribution for $L = M_p$ has a connection with stationary matchings of two independent Poisson point processes of intensity $1$, see Holroyd, Pemantle, Peres and Schramm \cite{HPPS}. Indeed, consider mutually independent random variables $(X_i)_{i\ge 1}$ and $(Y_j)_{j\ge 1}$ having uniform distribution on $Q = [-1/2,1/2]^d$. It is well known that for any $x$ in the interior of $Q$, the pair of point processes
$$\left( \frac 1 n \sum_{i = 1} ^n \delta_{n^{\frac 1 d} ( X_i - x ) } , \frac 1 n \sum_{i = 1} ^n \delta_{n^{\frac 1 d} ( Y_i - x) } \right)$$ converges weakly for the topology of vague convergence to
$
(\Xi_1 , \Xi_2 )
$, where $\Xi_1$ and $\Xi_2$ are two independent Poisson point processes of intensity $1$. Also, we may write
$$
n^{\frac p d - 1} \E M_p ( \{X_i \}_{i = 1} ^n , \{Y_i \}_{i = 1} ^n ) = \frac 1 n \E \sum_{i = 1} ^n \left| n^{\frac 1 d} ( X_i - x) - n^{\frac 1 d} ( Y_{\sigma_n ( i )}- x) \right| ^p.
$$
where $ \sigma_n$ is an optimal matching. Now, the fact that for $0 < p < 2d$, $\lim_n n^{\frac p d - 1} \E M_p ( \{X_i \}_{i = 1} ^n , \{Y_i \}_{i = 1} ^n ) = \beta_p ( d)$ implies the tightness of the sequence of matchings $\sigma_n$ and it can be used to define a stationary matching $\sigma$ on $(\Xi_1 , \Xi_2 )$, see the proof of Theorem 1 (iii) in \cite{HPPS} for the details of such an argument. In particular, this matching $\sigma$ will enjoy a local notion of minimality for the $L_p$-norm, as defined by Holroyd in \cite{H} (for the $L_1$-norm).
See also related work of Huesmann and Sturm \cite{Hu}.
\section*{Acknowledgments}
We are indebted to Martin Huesmann for pointing an error in the proof of a previous version of Theorem \ref{th:mainMp}. This is also a pleasure to thank for its hospitality the Newton Institute where part of this work has been done (2011 Discrete Analysis programme).
\medskip
\noindent
|
\section{Introduction}
Fundamental processes, such as transcription and replication require a transient, local opening of the DNA double helix \cite{calladine,alberts}. Such {\it bubbles} occur spontaneously under physiological conditions \cite{kowalski88,kowalski89}, while complete melting and the separation of the two complementary strands requires temperatures around $80^o$C \cite{FK}. Bubbles have been implicated as an explanation for high cyclization rates in short DNA fragments \cite{cloutier,yan}, but their main interest lies in biology and in the physical mechanism underlying the functioning and control of transcription and replication start sites: the stability of DNA is sequence-dependent \cite{SL} and opening is strongly influenced by superhelical stress \cite{benham92,fye99,benham06} and the binding of regulatory proteins \cite{sobellPNAS85,ambjornsson05}. In particular, Benham and coworkers showed significant correlations between the positions of strongly stress-induced destabilized regions and regulatory sites \cite{wang04,ak05,wang08}. \\
Several models exist to describe the internal opening of DNA.
The Peyrard-Bishop-Dauxois \cite{peyrard89,dauxois93} and the Poland-Scheraga (PS) models \cite{poland,jostBJ09} have already been used to quantify bubble statistics \cite{kalosakas04,vanerp05,ares07,kafri02,blossey03,garel05,coluzzi07,monthus07} and for the {\it ab initio} annotation of genomes on the basis of correlations between biological function and thermal melting \cite{yeramian02,carlon07,jostJPCM09}.
Here, we describe the bubble statistics of random and biological sequences at physiological conditions using (i) the Zimm-Bragg (ZB) model, an efficient approximation of the PS model \cite{jostJPCM09}, and (ii) the Benham model \cite{fye99}, a generalization of the ZB model accounting for superhelical density but neglecting writhe.
After exploring the bubble statistics of unconstrained DNA using the ZB model, we introduce an asymptotically exact self-consistent linearization \cite{vologodskii} of the Benham model as a precise and convenient tool to study the huge impact of superhelicity on local bubble opening. The numerical efficiency of the method allows us to investigate the bubble statistics for entire genomes and random sequences of sufficient length ($10^6-10^9$ bp) to obtain statistically significant results for sequence effects on bubble statistics and positioning.
In a final step, we correlate the positions of highly probable bubbles within the genome of {\it E. coli} with the position of transcription start sites (TSS) and start codons.
\section{Models and Methods}
\subsection{Unconstrained DNA}
\subsubsection{The Zimm-Bragg model}
The most widely-used model to treat the denaturation of DNA chains is the PS model \cite{poland} which offers predictive power for thermal melting and strand dissociation for DNA of arbitrary length, strand concentration and a wide range of ionic conditions.
For long heterogeneous sequences, whose local denaturation is dominated by the quenched sequence disorder \cite{garel05,coluzzi07,monthus07}, the related and computationally faster ZB model gives surprisingly good results \cite{jostJPCM09}.
In the PS and ZB formalisms, the free energy of a given configuration is decomposed into formation free energies for closed base-pair steps and free energy penalties for the nucleation of unpaired regions (or bubbles). Using periodic boundary conditions, the ZB Hamiltonian for a circular chain of length $N$ can be expressed as \cite{jostJPCM09}
\begin{equation}
\mathcal{H} _{ZB} = \sum_{i=1}^N \left\{ (\Delta g_{i}-\Delta g_{loop})\theta_i\theta_{i+1}+\Delta g_{loop}\theta_i\right\} \label{eq:zb}
\end{equation}
where $\theta_i=0$ ($1$) if base-pair $i$ is open (closed). $\Delta g_{i}$ ($\sim k_b T $) is the nearest-neighbor (NN) free energy to form the base-pair step $(i,i+1)$.
$\Delta g_{loop}\equiv-k_B T\log[\sigma_{PS}\bar{L}^{-c}]$ is the loop nucleation penalty and depends on the PS cooperativity $\sigma_{PS}\sim 10^{-4}$, on the interacting self-avoiding loop exponent $c\sim2.1$ \cite{kafri02,blossey03} and on a typical bubble length $\bar{L}$ \cite{jostJPCM09} ($\Delta g_{loop}\sim 20 k_b T$ for $\bar{L}=170$). Parameters are taken at physiological conditions for temperature $T=37^o$ C and salt concentration $[Na^+]=0.1$ M \cite{jostJPCM09}.
\subsubsection{Transfer-matrix method}
Being formulated as a 1D Ising model, the model is easily solved analytically (numerically) for homogeneous (heterogeneous) sequences using transfer matrix methods. The partition function of the system is then given by \cite{jostJPCM09}
\begin{equation}\label{eq:zbz}
\mathcal{Z}=\textrm{Trace}\left[\prod_{i=1}^{N} T_i\right],
\end{equation}
the individual closing probability by
\begin{equation}\label{eq:zbthet}
\langle \theta_i\rangle=\frac{1}{\mathcal{Z}}\textrm{Trace}\left[\left(\prod_{j=1}^{i-1} T_j \right) T'_i, \left(\prod_{j=i+1}^{N} T_j \right)\right].
\end{equation}
In particular, it is possible to calculate position resolved opening probabilities, $P_{i,l}$, for bubbles containing $l$ open bps and beginning at the closed bp $i$, by
\begin{equation}\label{eq:zbpb}
P_{i,l}=\frac{1}{\mathcal{Z}}\textrm{Trace}\left[\left(\prod_{j=1}^{i-1} T_j \right) T'_i S T'_{i+l+1} \left(\prod_{j=i+l+2}^{N} T_j \right)\right]
\end{equation}
with
\begin{displaymath}
T_i = \left( \begin{array}{cc}
e^{-\beta \Delta g_{i}} & e^{-\beta\Delta g_{loop}} \\
1 & 1
\end{array} \right),
\end{displaymath}
\begin{displaymath}
T'_i = \left( \begin{array}{cc}
e^{-\beta \Delta g_{i}} & e^{-\beta\Delta g_{loop}} \\
0 & 0
\end{array} \right)\quad\textrm{and}\quad S = \left( \begin{array}{cc}
0 & 0 \\
1 & 1
\end{array} \right).
\end{displaymath}
The global closing probability $\Theta$ and the probability per bp $P_l$ to observe a bubble of length $l$ are given by $\Theta=(\sum_{i=1}^N \langle \theta_i\rangle)/N$ and $P_l=(\sum_{i=1}^N P_{i,l})/N$. Basic summing rules on the bubble probabilities imply that $1-\Theta = \sum_l l P_l$, and that $\mathcal{L}=(\sum_l l P_l)/(\sum P_l)=(1-\Theta)/P_b$,
\begin{equation}
\mathcal{L}=\frac{\sum_l l P_l}{\sum P_l}=\frac{1-\Theta}{P_b}
\end{equation}
where $\mathcal{L}$ is the average bubble length and $P_b=\sum_l P_l$ is the probability per bp to observe a bubble of arbitrary length.
For homogeneous sequences, these equations could be solved easily. In the asymptotic limit $N\rightarrow \infty$, it leads to
\begin{eqnarray}
\mathcal{Z} & = & \left[\frac{1+g+\sqrt{(g-1)+4s}}{2}\right]^N, \label{eq:homoz}\\
\langle\theta\rangle & = & \frac{1}{2}\left[1+\frac{g-1}{\sqrt{(g-1)^2+4s}}\right], \label{eq:homozb}\\
P_l & = & \frac{1}{2}\left[\frac{2}{1+g+\sqrt{(g-1)^2+4s}}\right]^{l+2} \nonumber \\
& &\times\left[\frac{2s+g\left(g-1+\sqrt{(g-1)^2+4s}\right)}{\sqrt{(g-1)^2+4s}}\right] \label{eq:homop}
\end{eqnarray}
with $g=\exp[-\beta\Delta g]$ and $s=\exp[-\beta\Delta g_{loop}]$.
The bubble probability is therefore a decreasing exponential function of $l$ for homogeneous sequences.
For heterogeneous sequences, we numerically compute the different desired observables in a $\mathcal{O}(N)$-algorithm. In a first step, we iteratively compute the forward and backward products, $F_i=\left(\prod_{j=1}^{i} T_j \right)$ and $B_i=\left(\prod_{j=i}^{N} T_j \right)$. The second step consists in applying equations \ref{eq:zbz}, \ref{eq:zbthet} and \ref{eq:zbpb}.
\subsection{Superhelical DNA}
In living organisms, DNA is highly topologically constrained into circular domains (closed-loops or circular molecules) \cite{calladine}.
Each closed domain is defined by a topological invariant $L$, the so-called linking number. $L$ represents the algebraic number of turns either strand of the DNA makes around the other. It can be decomposed in two contributions: the twist which is the number of turns of the double-helix around its central axis, and the writhe which is the number of coils of the double-helix.
In the majority of living systems, the average linking number $L$ is below the characteristic linking number value $L_o$ of the corresponding unconstrained linear DNA, due to a negative superhelical density $\sigma=\alpha A/N\approx -0.06$ imposed by protein machineries, where $\alpha=L-L_o$ is the linking number difference.
\subsubsection{The Benham model}
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{figure1.eps}\caption{Effective superhelical hamiltonian $\mathcal{H}_{TW}$ per bp as a function of the relative opening $n_o/N$ in the asymptotic limit, for different stress values: $\sigma=0$ (squares), -0.03 (stars), -0.06 (circles) and -0.1 (crosses).}\label{fig:qeff}
\end{center}
\end{figure}
In this article, we consider bubble openings in superhelically constrained circular DNA using the Benham model for an imposed superhelical density $\sigma$, where the standard thermodynamic description of base-pairing ($\mathcal{H}_{ZB}$) is coupled with torsional stress energetics \cite{benham92,fye99}.
For each state, if one neglects the writhe contribution, the imposed linking difference $\alpha$ can be decomposed in three contributions: 1) the denaturation of $n_o$ base-pairs relaxes the helicity by $-n_o/A$ where $A=10.4$ bp/turn is the number of base-pair in a helical turn; 2) the resulting single-strand regions can twist around each other inducing a global over-twist of $\mathcal{T}$; 3) then, the bending and twisting of double-stranded parts is put in the residual linking number $\alpha_r$. Therefore, due to the topological invariance of $\alpha$, we get the closure relation
\begin{equation}\label{eq:she}
\alpha=-\frac{n_o}{A}+\mathcal{T}+\alpha_r
\end{equation}
For denatured regions, the high flexibility of single-stranded DNA allows unpaired strands to interwind. The energy associated with the helical twist $\tau_i$ (in rad/bp) of open base-pair $i$ is
\begin{equation}
\mathcal{H}_{tw}(\theta_i,\tau_i)=\frac{C}{2} (1-\theta_i)\tau_i^2
\end{equation}
where the torsional stiffness $C$ is known from experiments, $C\approx 3.09 k_B T$.
The individual twist $\tau_i$ are related to the global over-twist $\mathcal{T}$ via the relation $\mathcal{T}=\sum_i (1-\theta_i)\tau_i/(2\pi)$.
For paired helical regions, it has been experimentally found that superhelical deformations induce an elastic energy, quadratic in the residual linking difference
\begin{equation}
\mathcal{H}_r = \frac{K}{2}\alpha_r^2=\frac{K}{2}\left(\alpha+\frac{n_o}{A}-\mathcal{T}\right)^2
\end{equation}
where $n_o=\sum_i (1-\theta_i)$ is the number of open base-pairs and $K\equiv K'/N \approx 2220 k_B T /N$.
By integrating over the $n_o$ continuous degrees of freedom $\tau_i$, the superhelical stress energetics is represented by a non-linear effective Hamiltonian \cite{fye99}:
\begin{eqnarray}
\mathcal{H}_{TW}(n_o) & = &\frac{2\pi^2 C K}{4\pi^2C+Kn_o}\left[\alpha+\frac{n_o}{A}\right]^2 \nonumber \\
& &-\frac{1}{2\beta} \log\left(\left[\frac{2\pi}{ \beta C}\right]^{n_o}\frac{4\pi^2 C}{4\pi^2 C+K n_o}\right) \label{eq:qo}
\end{eqnarray}
$\mathcal{H}_{TW}$ is minimal for a non-zero number of opening base-pairs which increases as the stress strength increases (see figure \ref{fig:qeff}).\\
The total effective Hamiltonian is given by $\mathcal{H}_{eff}(\{\theta_i\})=\mathcal{H}_{ZB}(\{\theta_i\}) +\mathcal{H}_{TW}(n_o)$. We fix $\Delta g_{loop}=20 k_b T$.
\subsubsection{Torque-imposed ensemble}
The Benham Hamiltonian is defined in a superhelical density ($\sigma$)-imposed ensemble defined by equation \ref{eq:she}. In this section, we briefly discuss similar model but in the torque-imposed ensemble. \\
In this ensemble, a linear DNA segment (length $N$) is constraint by a weak torque $\Gamma$ applied on base $N$, the first bp being fixed. For each bp $i$, we define $\omega_i$ its orientation in the plane perpendicular to the average axis of the double-helix and oriented in the 5' to 3' direction (for a denatured bp-step, $\delta \omega_i\equiv\omega_{i+1}-\omega_i=\tau_i/(2\pi)$ of the Benham model). The total Hamiltonian of the system is then given by \cite{orland}
\begin{eqnarray}
\mathcal{H}&=&H_{ZB}+\frac{C}{2}\sum_i (1-\theta_i)(\delta\omega_i)^2 \nonumber \\
& &+\frac{K_r}{2}\sum_i \theta_i (\delta\omega_i-\omega_0)^2-\Gamma(\omega_N-\omega_1)
\end{eqnarray}
with $K_r=K'/(4\pi^2)$ and $\omega_0=2\pi/A$ the natural helical twist.
Writing $(\omega_N-\omega_1)=\sum_i \delta\omega_i$, and integrating over the $\delta\omega_i$, leads to the effective Hamiltonian (relatively to the situation without torque):
\begin{equation}
\mathcal{H}_{eff}=\mathcal{H}_{ZB}+\left[\Gamma^2\left(\frac{1}{2C}-\frac{1}{2K_r}\right)-\Gamma \omega_0\right]\sum_i \theta_i
\end{equation}
Applying the constant torque $\Gamma$ is therefore equivalent to adding an external field $h_e=-[\Gamma^2(1/(2C)-1/(2K_r))-\Gamma \omega_0]$ in the ZB formalism.
\subsection{Self-consistent linearization}
\subsubsection{Solving the Benham model}
Recently, Jeon et al \cite{jeonPRL10} derived formulas to solve the Benham model for homogeneous and random heterogeneous sequences, including the computation of sequence-average bubble properties. The computation is based on a reorganization of the partition function sum into partial sums for fixed numbers of bubbles.
For heterogeneous sequences, Fye and Benham \cite{fye99} proposed an exact $\mathcal{O}(N^2)$-algorithm by decomposing the prefactor $\exp[-\beta\mathcal{H}_{TW}(n_o)]$ in discrete Fourier modes, and by using the transfer-matrix method.
Benham and coworkers \cite{benham04,bi03,bi04} have also developed an approximate method which first involves a windowing procedure to find the minimum free energy and then consider only the states whose energies do not exceed the minimum one by more than a given threshold. At high threshold values or high negative superhelicity, the computation time for this algorithm scales exponentially with the threshold and the superhelical stress.
Both schemes are still time demanding for very long sequences.
\subsubsection{Self-consistent field}
In order to speed up the resolution of the Benham model and to access directly to position-dependent opening properties of bps {\it and} bubbles, we develop an efficient variational method \cite{vologodskii} allowing us to use the transfer-matrix solution of the ZB model. For long sequences, assuming that fluctuations of $n_o$ are small around the value $\bar{n}_o$, we can expand $\mathcal{H}_{TW}(n_o)$ around $\bar{n}_o$. The approximated effective Hamiltonian then takes the typical ZB form:
\begin{equation}
\mathcal{H}_{eff}\approx \mathcal{H}_{ZB}(\{\theta_i\})+\mathcal{H}_{TW}(\bar{n}_o)+(N-\bar{n}_o)h -h\sum_{i=1}^N \theta_i \label{eq:heff}
\end{equation}
where $h=(\partial \mathcal{H}_{TW}/\partial n_o)(\bar{n}_o)$ represents the mean-field of our approximation. If one imposes the superhelical stress $\Gamma$ (ie, the torque) instead of the superhelical density, $h$ is related to the effective field $(4\pi^2N/K-1/C)(\Gamma^2/2)+2\pi\Gamma/A$ generated by the torque (see above). \\%This representation has the advantage to decouple the opening of different parts of the molecule and is probably the simplest way to express the destabilizing effect of undertwisting on DNA stability.\\
In the following, we employ the Benham ensemble of an imposed superhelical density, $\sigma$, and determine $h$ (or $\Gamma$) self-consistently. Self-consistency requires $\langle n_o \rangle(\bar{n}_o)=\bar{n}_o$, and is equivalent to solving
\begin{equation}\label{eq:fh}
h=(\partial \mathcal{H}_{TW}/\partial n_o)(\langle n_o \rangle(h)),
\end{equation}
because the function $(\partial \mathcal{H}_{TW}/\partial n_o)(n_o)$ is monotonic.
\subsubsection{Homogeneous sequences}
For homogeneous sequences, the general solution of the self-consistency equation \ref{eq:fh} cannot be computed analytically. However, at low temperatures, weak superhelical densities and in the limit of infinitely long chains, a small perturbation development is valid and leads to analytical expressions for $h$.
For infinitely long chains, $\mathcal{H}_{TW}$ becomes
\begin{equation}
\mathcal{H}_{TW}(x)/N=\frac{2 C K' \pi^2 (\sigma+x)^2}{A^2(4C \pi^2+K' x)}-f\,x
\end{equation}
with $x=n_o/N$, $K'=N K$ and $f=-\log[2\pi/(\beta C)]/(2\beta)$.
Assuming that the fraction of open base-pairs $x$ is small ($x<<1$), $\partial \mathcal{H}_{TW}/\partial n_o=\partial (\mathcal{H}_{TW}/N)/\partial x$ is given by
\begin{equation}\label{eq:dh}
\left(-f+\frac{K' \sigma}{A^2}-\frac{K'^2\sigma^2}{8 A^2 C \pi^2}\right)+K'\left(\frac{[K'\sigma-4\pi^2 C]^2}{A^2 (4\pi^2 C)^2}\right)x+o(x^2)
\end{equation}
Using Eq.\ref{eq:homozb} and noting $y=\exp[\beta h]$, we also have
\begin{equation}
x\approx \frac{s y }{(g y -1)^2}
\end{equation}
In the limit $g y >>1$, inserting $x\approx s/(g^2 y)$ in Eq.\ref{eq:dh} and solving Eq.\ref{eq:fh}, leads to
\begin{equation}\label{eq:h1}
h^* = -f+\frac{K'\sigma (8C \pi^2-K'\sigma)}{8A^2 C \pi^2}
\end{equation}
This expression is valid until $g y \sim 1 $ ($h\approx \Delta g$), i.e.
\begin{equation}
\sigma\approx \left(\frac{2\pi}{K'}\right)\left(2\pi C-\sqrt{2C(2\pi^2 C-A^2[f-\Delta g])}\right)\equiv \sigma_l
\end{equation}
For $\sigma < \sigma_l$, we could write $g y =1+\epsilon $ (with $\epsilon >0 $). Then, $x\approx (s/g)/\epsilon^2$ and $h\approx \Delta g +\epsilon$. Solution of Eq.\ref{eq:fh} leads to
\begin{widetext}
\begin{equation}\label{eq:h2}
h^* = \Delta g+\left(\sqrt{\frac{K' s}{2\pi^2 C g}}\right)(4 C \pi^2-K'\sigma) \left(\frac{4 C \pi^2-K'\sigma}{\sqrt{8 A^2 \pi^2 f C+K'\sigma(K'\sigma-8\pi^2 C)+8A^2 \pi^2 C \Delta g}}\right)
\end{equation}
\end{widetext}
Figure \ref{fig:approx} shows that the numerical solution of Eq.\ref{eq:fh} agrees very well with the two expressions found above (Eq.\ref{eq:h1} and \ref{eq:h2}). \\
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{figure2}\caption{Numerical (dots) or analytical (lines) solutions of the self-consistency equation Eq.\ref{eq:fh} for a homogeneous sequence ($\Delta g=-3.14 k_B T$).}\label{fig:approx}
\end{center}
\end{figure}
To compute bp or bubbles properties, formulas \ref{eq:homoz}, \ref{eq:homozb} and \ref{eq:homop} are still available if one replaces $g$ by $g\,y$ and $s$ by $s\, y$.
\subsubsection{Heterogeneous sequences}
For heterogeneous sequences, we use the bisection method coupled to the Newton-Raphson method \cite{press} to numerically solve the self-consistency equation, for fixed values of the temperature and of the superhelical density. Knowing that an evaluation of the function $f$ requires one transfer matrix method computation ($\mathcal{O}(N)$ each), it takes typically $10-20$ evaluations to determine the root with a relative precision of $10^{-4}$. This allows numerically efficient computation of denaturation profiles. For example, computing the local closing probabilities for the {\it E.coli} genome ($N\sim4.6$ Mbps) takes about $70$ seconds on a 2.4 GHz computer with the self-consistent method, whatever the density is. On the same computer, it would take about $10^{10}$ s with the exact method ($\mathcal{O}(N^2)$ algorithm) for any $\sigma$ values (interpolation of data given in Ref.\cite{fye99}), and about $6.10^4$ s with the approximate method for $\sigma=-0.055$ ( and around 40 times more for $\sigma=-0.075$) \cite{benham04,bi03,bi04}.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{figure3}\caption{(Color online) Opening free energy $-\log[1-\langle \theta_i\rangle]$ for the extended neighborhood of the TSS of Fig. \ref{fig:land}, computed using the Benham's web server \cite{benhamweb} (A) or using our formalism (B).}\label{fig:compa}
\end{center}
\end{figure}
\subsubsection{Comparison with the Benham model}
The self-consistent linearization consists in working in the {\em torque}-imposed ensemble and determining the torque self-consistently to better approximate the {\em superhelical density}-imposed ensemble.
This representation has the advantage of decoupling the opening of different parts of the molecule and is probably the simplest way to express the destabilizing effect of undertwisting on DNA stability.
In the thermodynamic limit, i.e., for very long (genomic) sequences and at low temperature (well below the melting temperature), we expect our linearization to be a quasi-exact solution of the Benham model due to the asymptotic decrease of fluctuations within the system.
For shorter sequences, however, the non-linearity of the effective superhelical Hamiltonian $\mathcal{H}_{TW}$ (see Eq.\ref{eq:qo}) may significantly couple remote domains along the sequence, leading for example, to the closing of an open domain as one increases the superhelical stress (as observed in Fig.\ref{fig:compa}A). Although, the self-consistent linearization neglects such effects, it gives a reliable general picture of the superhelically-stressed destabilization of DNA sequences. In figure \ref{fig:compa}, we show a comparison of results obtained from the Benham web server \cite{bi04,benhamweb} and by our method. The agreement is excellent and the small deviations are mostly due to the slightly different parametrizations of the ZB parts and to the absence of finite size effect in our approach as discussed above.
\section{Results and Discussion}
In the following, we discuss results obtained for random homogeneous and heterogeneous sequences, as well as for some bacterial genomes ({\it E.coli}, {\it T. whipplei}, {\it A. Baumanii}, {\it B. subtilis} and {\it S. coelicolor}, see Table \ref{tab:zs}).
The results shown for random heterogeneous sequences were obtained by compiling profiles from 100 random sequences each containing $10^6$ bp.
\subsection{Melting of superhelical DNA}
The melting properties of constrained DNAs reflect a balance between the two parts of $\mathcal{H}_{eff}$. At $37^o$ C, $\mathcal{H}_{ZB}$ opposes local opening the more strongly the higher the GC-content of the sequence. In contrast, $\mathcal{H}_{TW}$ is minimal for a finite number of open base-pairs, which increases as $\sigma$ becomes more negative. At biological levels, the free energy of completely closed DNA becomes prohibitively large (see Fig.\ref{fig:qeff}). As a consequence, superhelicity leads to a notable level of base-pair opening at physiological temperatures, where unconstrained DNA exhibits negligible breathing.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{figure4.eps}\caption{Evolution of the global opening probability $1-\Theta$ as a function of temperature for an unconstrained (stars) or constrained ($\sigma=0$: squares, $\sigma=-0.06$: circles, $\sigma=-0.2$: crosses) random sequence (GC=0.5).}\label{fig:thetatm}
\end{center}
\end{figure}
Figure \ref{fig:thetatm} shows the overall impact on the opening probability of imposing a superhelical constraint. For comparison, we have also included a melting curve for unconstrained DNA. At physiological temperatures, the unconstrained DNA is very stable whereas the superhelicity significantly contributes to opening of base pairs and bubbles. However, for intermediate and weak negative stresses, the destabilizing effect of an imposed superhelical density is reversed close to the melting temperature due to the overall stabilizing impact of untwisting on the rest of the DNA ($\mathcal{H}_{TW}(n_o=N/2)>0$), resulting in a slowdown of the melting process via a change of $\Gamma$ (or $h$) with temperature. This effect has already been pointed out for positively-stressed homogeneous molecules \cite{benham96}. For stronger stresses, the effective twisting Hamiltonian $\mathcal{H}_{TW}$ at the melting transition ($n_o=N/2$) is still negative and then results in a decrease of the melting temperature. A quantification on this salt-, GC-dependent effect is described below.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{figure5}\caption{(A) Melting temperature for random sequences (GC$\in[0,1]$) at different salt concentrations ($[Na^+]\in[0.05,1]$ M) as a function of the superhelical density $\sigma$ (dots), respectively to the situation with $\sigma=0$. (B) Melting temperature at $\sigma=0$ for random sequences as a function of their GC-content, at $[Na^+]=0.05$ (circles), 0.1 (squares) and 0.3 M (crosses), respectively to the corresponding unconstrained (ZB) situation. Dashed lines represent the fitted relation given in Eq.\ref{eq:tmsiggc}.}\label{fig:salt}
\end{center}
\end{figure}
On figure \ref{fig:salt} A, we observe that, under the different considered GC-contents and salt concentrations, the melting temperature of random superhelical DNA is a quadratic function of $\sigma$, and only the intercept of this function ($T_m(\sigma=0)$) depends on the GC and on $[Na^+]$ (see figure \ref{fig:salt} B). From a systematic study of the melting temperature $T_m$ as a function of the GC-content and the salt concentration, we fitted the empirical relation (in $^o$ C):
\begin{eqnarray}\label{eq:tmsiggc}
\Delta T_m([Na^+],\textrm{GC},\sigma) & = & 2.35 -0.54\log[Na^+]-3.42\textrm{GC} \nonumber \\
& & -0.93\textrm{GC}^2+4.7\sigma-46.8\sigma^2
\end{eqnarray}
where $\Delta T_m$ is the difference in melting temperature between a constrained and an unconstrained random DNA polymers.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{figure6}\caption{(Color online) Evolution at $T=37^o$ C of the global opening probability $1-\Theta$ for superhelically constrained random heterogeneous (full lines) and homogeneous (dashed lines) DNAs under different $\sigma$ (from -0.02 to -0.1 every 0.01: color scale from light to dark red, $\sigma=-0.06$: blue line) and for several bacterial genomes at $\sigma=-0.06$: {\it E. coli} (triangles), {\it T. whipplei} (squares), {\it A. Baumanii} (diamonds), {\it B. subtilis} (circles) and {\it S. coelicolor} (stars). Inset: proportion of the different contributions to the linking number difference $\alpha$: relaxation due to denaturation $-n_o/N$ (I), over-twist of the denatured regions $\mathcal{T}$ (II) and residual linking number $\alpha_r$ (III) (see Eq.\ref{eq:she}).}\label{fig:stresrandt}
\end{center}
\end{figure}
Figure \ref{fig:stresrandt} shows the dependance of the opening probability $1-\Theta$ as a function of the GC-content. Results are reported for different levels of superhelical density and at $T=37^o$ C, in the biological relevant range, i.e the overall degree of opening is small yet strongly increased relative to the unconstrained case (see also Fig.\ref{fig:thetatm}). We have included results for both homogeneous and heterogeneous systems. In the former case, the employed NN-free energies $\Delta g$ were determined as composition dependent averages over the tabulated step parameters \cite{SL,jostBJ09}.\\
Passing a threshold \cite{jeonPRL10} (see Inset in Fig.\ref{fig:stresrandt}) depending on the GC-content, strong $\sigma$-values allow the opening of many bp along the sequences.
For a fixed superhelical stress, as expected, $1-\Theta$ is a decreasing function of the GC-content.
Differences between homogeneous and heterogeneous are weak, meaning that sequence heterogeneity self-averages and has only a small effect on the total degree of opening.\\
The inset in figure \ref{fig:stresrandt} shows the different contributions to the linking number difference as a function of the superhelical density (see Eq.\ref{eq:she}) for random sequences with GC=0.5. The over-twist $\mathcal{T}$ contribution is estimated by integrating over the $n_o$ continuous degrees of freedom $\tau_i$ \cite{fye99} and applying the self-consistent linearization:
\begin{eqnarray}
\langle \mathcal{T} \rangle & = &\sum_{\{\theta_i \}} \frac{K n_o}{4\pi^2 C+K n_o}(\alpha+n_o/A) \frac{e^{-H_{eff}(\{\theta_i\})}}{\mathcal{Z}}\\
& \approx & \frac{K \langle n_o \rangle }{4\pi^2 C+K \langle n_o \rangle}(\alpha+\langle n_o \rangle/A)
\end{eqnarray}
We observe that the over-twist increases linearly with the opening probability. While for weak stresses, almost all the superhelical energy is stored in the (residual) deformations of the double-helix, for strong stresses, more than 50 \% of the imposed superhelical constraint is used to drive the local denaturation of bps. Accounting for the writhe in the model should however decrease the contributions due to bp-denaturation and over-twisting since a part of the constraint would be absorbed by coils of the double-helix.
\subsection{Sequence heterogeneity and bubble statistics in superhelical DNA}
\subsubsection{Bubble statistics in unconstrained DNA}
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{figure7}\caption{(Color online) Evolution of the bubble probability per bp $P_l$ to observe a bubble of length $l$. (A) For unconstrained molecules: random sequences of different GC-content (GC from 0 to 1 every 0.1: color scale from light to dark red, GC=0.5: squares) and the genome of {\it E. coli} (dashed line with crosses, GC=0.51). Inset: average bubble length $\mathcal{L}$ as a function of the GC-content. (B) For superhelically-stressed molecules: a random sequence ($GC=0.5$, squares), the genome of {\it E. coli} (crosses) and a homogeneous sequence (with $\Delta g=-1.9 k_B T$ equals to the average NN-free energies in {\it E. coli}, circles), and $\sigma=0$ (dotted and dashed lines), $-0.06$ (full lines) and -0.1 (dashed lines). }\label{fig:probzb}
\end{center}
\end{figure}
Figure \ref{fig:probzb}A shows the bubble probabilities computed with the ZB model for random DNAs of different GC-content, with $\Delta g_{loop}=10k_bT$ consistent with very small loops ($\bar{L}\sim \mathcal{L}\sim 1$, see inset in Fig.\ref{fig:probzb} A). We remark an exponential decrease of $P_l$ with very short decay lengths, corresponding to fairly closed molecules. Increasing GC-content stabilizes the DNA and reduces average bubble lengths. Even accounting for the weak biological sequence effect apparent in our results, the absolute level of bubble opening in unconstrained DNA seems too small for natural breathing to play a direct biological role \cite{benham06}.
Whereas we obtain similar $\mathcal{L}$-values as reported for the Peyrard-Bishop-Dauxois (PBD) model \cite{ares07}, absolute probabilities $P_l$ are far lower (for example $P_l\sim 10^{-11}$ (ZB) versus $\sim 10^{-5}$ (PBD) for $l=10$ and GC=0.5), mainly due to the absence of a cooperativity penalty factor preventing bubble formation in the PBD model.
\subsubsection{Bubble statistics in superhelical DNA}
Figure \ref{fig:probzb}B highlights the huge impact of the superhelical stress on the bubble statistics. For physiological levels, the opening probability per bp remains significant even for large bubbles and has increased by many orders of magnitude compared to the unconstrained situation. For example, for GC=0.5 and $\sigma=-0.06$, $P_{l=10}\approx 10^{-9}$ and $P_{l=60}\approx 10^{-7}$ for superhelical DNA, while for the corresponding unconstrained DNA, we found $P_{l=10}\approx 10^{-11}$ and $P_{l=60}\approx 10^{-41}$.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{figure8}\caption{(Color online) Evolution of bubble occurrence probability $P_b$ (A,B) and of the average bubble length $\mathcal{L}$ (C,D) for superhelically constrained random heterogeneous (B,D) and homogeneous (A,C) DNAs under different stresses $\sigma$ (from -0.02 to -0.1 every 0.01: color scale from light to dark red, $\sigma=-0.06$: blue line) and for several bacterial genomes at $\sigma=-0.06$: {\it E. coli} (triangles), {\it T. whipplei} (squares), {\it A. Baumanii} (diamonds), {\it B. subtilis} (circles) and {\it S. coelicolor} (stars).}\label{fig:stresrand}
\end{center}
\end{figure}
Figure \ref{fig:stresrand} shows the evolution of the bubble occurence $P_b$ and of the average bubble length $\mathcal{L}$ for different $\sigma$-values and GC-content.
With $\mathcal{H}_{TW}(n_0)$ only being a function of the total number of open base-pairs, bubble sizes in homogeneous systems are determined by a competition between the bubble initiation penalty, $\Delta g_{loop}$, favoring the opening of a small number of large bubbles, and entropy, favoring the opening of a large number of small bubbles. In heterogeneous systems, it is possible to lower the fraction of stable GC-steps in the open domains by denaturing a larger number of smaller bubbles in particularly AT-rich regions (see Fig.\ref{fig:gcloc}) \cite{hwa03}. The comparison in Figs. \ref{fig:stresrand} and \ref{fig:gcloc} shows that the disorder effect dominates in the present case \cite{garel05,coluzzi07,monthus07} with the number of bubbles being maximal around GC=0.5. For biological superhelicities, $\mathcal{L}\approx700$ for random AT- and GC-DNAs, while $\mathcal{L}\approx150$ for $0.2<f_{GC}<0.9$.
Sequence-heterogeneity plays an essential role by lowering the fraction of stable GC-steps in the open domains and leads to a {\em localization} of open base-pairs in (AT-enriched) stress-induced duplex destabilized regions \cite{wang04,ak05,wang08}, whose length in turn limits the bubble sizes.
Indeed, figure \ref{fig:gcloc} shows that bubbles appear mainly in AT-enriched regions compared to the background GC-content. Interestingly, for intermediate (biological) $\sigma$ values, the evolution of the GC-composition of bubbles remains flat over a large range of global GC-content.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{figure9}\caption{(Color online) Evolution of the average GC-content of bubbles for superhelically constrained random heterogeneous DNAs under different stresses $\sigma$ (from 0 to -0.1 every 0.01: color scale from light to dark red, $\sigma=-0.06$: blue line). The dashed line represents results for homogeneous sequences.}\label{fig:gcloc}
\end{center}
\end{figure}
\subsection{Bubble statistics in biological DNA}
Compared to random sequences with identical GC-content, bacterial genomes ({\it E.coli}, {\it T. whipplei}, {\it A. Baumanii}, {\it B. subtilis} and {\it S. coelicolor}) exhibit higher averages degrees of opening and increased average bubble sizes (symbols in Figs.\ref{fig:stresrandt} and \ref{fig:stresrand}).
Figure \ref{fig:probzb}B shows striking discrepancies in the bubble distribution $P_l$ between random and biological sequences, with a significant enrichment of large bubbles. In the latter case, the nearly flat distributions resemble those expected for homogeneous sequences and can be viewed as a signature of large stress-induced destabilized domains, which concentrate the DNA breathing into large regions with homogeneous opening profiles. We also note that the level of opening in the biological domains is comparatively insensitive to the precise level of the superhelical density.
\begin{table}
\caption{Z-scores computed relatively to the average bubble length for 5 prokaryotic organisms.}\label{tab:zs}
\begin{tabular}{l|ccc}
\hline
Organism & $N$ (Mbp) & GC-content & Z-score \\
\hline
{\it A. baumanii} & 3.98 & 0.40 & 23.0 \\
{\it B. subtilis} & 4.21 & 0.44 & 21.0 \\
{\it E. coli} & 4.64 & 0.50 & 25.7 \\
{\it S. coelicolor} & 8.67 & 0.70 & 17.6\\
{\it T. whipplei} & 0.93 & 0.46 & 10.6 \\
\hline
\end{tabular}
\end{table}
In general, biological sequences are more destabilized than random sequences with a same GC-content, leading to less but longer bubbles. We estimate these differences by computing, for all genomes, the Z-score relatively to $\mathcal{L}$ (see Tab.\ref{tab:zs}). The Z-score is an estimation of the non-randomness of a specific sample. For the average bubble length $\mathcal{L}$, it can be computed by
\begin{equation}
Z=\frac{\mathcal{L}(\mathrm{genome})-\langle\mathcal{L}(\mathrm{random})\rangle}{\sigma_{\mathcal{L}(\mathrm{random})}}
\end{equation}
where $\langle\mathcal{L}(\mathrm{random})\rangle$ and $\sigma_{\mathcal{L}(\mathrm{random})}$ are the mean and the standard deviation of $\mathcal{L}$ computed on 100 random sequences with the same GC and length as the corresponding genome. The Z-score represents therefore the distance between a specific point and the mean value obtained for random sequences, given in standard deviation units.
Z-scores of 20 for $\mathcal{L}$ suggests the presence of over-represented long AT-rich domains in bacterial genomes which have the ability to easily open under superhelical stress.
Interestingly, the organism with the lowest Z-score ({\it T. whipplei}) was shown to exhibit a random-like behavior relatively to the local melting temperature distribution \cite{jostJPCM09}.
\subsection{Bubble positioning in the promoter regions of biological DNA}
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{figure10}\caption{(Color online) Free energies to open a bubble of a given length centered around a given nucleotide ($-k_B T\log[P_{i,l}]$) for the neighborhood of a transcription start site (bp n$^o$ 259382) in the {\it E. coli} genome, in the presence (B, C, D) or in the absence (A) of an imposed superhelical density $\sigma=-0.03$ (B), -0.06 (C) and -0.10 (D).}\label{fig:land}
\end{center}
\end{figure}
Positions of strongly stress-induced destabilized regions have been shown to correlate with many regulatory regions \cite{wang08} including transcription start sites \cite {wang04} or origins of replication \cite{ak05}. In this section, we focus on bubble positioning inside the promoter regions of the bacterium {\it E. coli}.\\
The transcription of DNA is initiated by the local opening of the double-helix at transcription start sites. Figure \ref{fig:compa} illustrates their association with strongly stress-induced destabilized regions. In addition to position-dependent opening probabilities, our approach allows us to calculate the complete bubble free-energy landscape, $G_{i,l}=-k_BT \log P_{i,l}$. Figure \ref{fig:land} shows the effect of superhelicity on $G_{i,l}$ for the neighborhood of the same TSS in {\it E. coli}. The analysis reveals that opening is the result of the meta-stable unwinding of a large bubble and not of enhanced small scale breathing. We note that knowledge of $G_{i,l}$ is essential for modeling the dynamics of bubble nucleation and growth \cite{ambjornsson07}.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{figure11}\caption{(A) Opening probability as a function of the position relatively to the TSS, averaged on 760 promoter regions of {\it E. coli} \cite{regulondb} (black line), or on 760 randomly picked regions inside the genome (blue line), for a constrained (top) or unconstrained (down) DNA. (B) Histogram of the highly probable ($P_{i,l} \ge 10^{-5}$) bubble centers included in the regions [TSS-300,TSS+300] for the 760 studied TSS (black bars) or for randomly picked regions (blue line). (C) Probability distribution function of the distance between highly probable bubble centers and the nearest start codons for the 1158 actually found (black line) or randomly situated (blue line) bubbles.}\label{fig:tss}
\end{center}
\end{figure}
Figure \ref{fig:tss} analyses the statistical relation beween TSSs and bubbles induced by superhelical stress for the entire genome of {\it E. coli}.
Figure \ref{fig:tss}A shows a significant and non-random destabilization around TSSs, with a maximal opening around $-80$. The same computation using the ZB model shows insignificant and orders-of-magnitude smaller opening probabilities. However, we find a non-random signal around TSS-10 corresponding to the position of the AT-rich Pribnow box, an essential motif to start transcription in bacteria \cite{pribnowPNAS75}.
Figure \ref{fig:tss}B gives the relative positions of highly probable bubbles included in TSS neighborhoods.
The centers of these bubbles are mainly localized in the [TSS-200,TSS] region where many transcriptional and promoter factors are recruited and bind to DNA \cite{regulondb}.
Conversely, figure \ref{fig:tss}C confirms that the majority of highly probable bubbles are situated upstream and close to start codons of genes. Actually, these bubbles are composed by around $36\%$ of coding bps, significantly lower than the percentage of coding bps in {\it E. coli} ($88\%$).
\section{Summary and Conclusion}
We have developed a numerically efficient, self-consistent solution of the Benham model of bubble opening in superhelically constrained DNA. In particular, we are able to go beyond the calculation of opening probabilities for base pairs and to address the full, position-dependent bubble statistics for entire genomes. Our results indicate, that negative supercoiling leads to (meta-) stable unwinding of bubbles comprising ${\cal O}(100-1000)$ base-pairs. In heterogeneous sequences, bubbles are strongly localized in AT-rich domains with sequence disorder dominating the bubble statistics. As we have shown, large bubbles open with a significantly larger probability in biological sequences, than in random sequences with identical GC-content. In the case of {\it E. coli}, the most likely bubbles are located directly upstream from transcription start sites, highlighting the biological importance of this now well understood, physical property of DNA.
|
\section{Introduction}
Polymers in solution exposed to shear flow exhibit a remarkably
rich dynamical behavior, as has been shown by direct observation
using fluorescence
microscopy~\cite{smith:99,schr:05,teix:05,schr:05_1,gera:06}. In
particular, polymers exhibit tumbling motion, i.e., they undergo a
cyclic stretch and collapse dynamics, with a characteristic
frequency which depends on the shear rate and their internal
relaxation time. This nonequilibrium behavior has intensively been
studied for polymers in dilute solution
~\cite{smith:99,aust:99,schr:05,teix:05,gera:06,doyl:00,cela:05,cher:05,puli:05,schr:05_1,delg:06,wink:06_1,zhan:09}.
The dynamical behavior of a polymer in \textit{semidilute}
solution under shear flow has received far less attention
\cite{babc:00,hur:01,huan:10_1}. Insight into the behavior of such
systems is of fundamental importance in a wide spectrum of systems
ranging from biological cells, where transport appears in dense
environments, to turbulent drag reduction in fluid flow. While the
dynamical behavior of polymers in dilute solution is strongly
affected by hydrodynamic
interactions~\cite{doi:86,kapr:08,gomp:09}, their relevance in
semidilute solutions is less clear.
The complex interactions in semidilute solutions hamper an
analytical treatment. Here, computer simulations are an important
tool to shed light on the rich and intricate dynamical behavior of
such systems. The large length- and time-scale gap between the
solvent and macromolecular degrees of freedom requires a mesoscale
simulation approach in order to assess their structural,
dynamical, and rheological properties. We apply a hybrid
simulation approach, combining molecular dynamics simulations (MD)
for the polymers with the multiparticle collision dynamics (MPC)
method describing the solvent~\cite{male:99,kapr:08,gomp:09}.
By this approach, we demonstrated that polymers in dilute and
semidilute solutions exhibit large deformations and a strong
alignment along the flow direction in simple shear flow
\cite{huan:10_1}. More importantly, in the stationary state, the
{\em conformational} and {\em rheological properties} for various
concentrations are universal functions of the Weissenberg number
$\mathrm{Wi}_c = \dot \gamma \tau(c)$, where $\dot \gamma$ is the shear
rate and $\tau(c)$ the concentration-dependent polymer end-to-end
vector relaxation time at equilibrium. Hence, with increasing
concentration, hydrodynamic interactions affect the conformational
and rheological properties only via the increasing relaxation time
$\tau(c)$. Experiments on DNA in shear flow \cite{hur:01} and
simulations of polymer brushes \cite{galu:10} lead to a similar
conclusion. Then, the question arises to what extent hydrodynamic
interactions are relevant in non-equilibrium systems.
In this letter, results are presented for the concentration
dependence of the non-equilibrium {\em dynamical properties} of
polymers in shear flow by calculating tumbling times and
orientational distribution functions. These quantities exhibit a
dependence on hydrodynamic interactions in dilute solution, and
are independent of such interactions in semidilute solution where
hydrodynamic interactions are screened. This is supported by a
comparison of non-draining and free-draining simulations. As a
result, hydrodynamic interactions ar found to clearly contribute
to the non-equilibrium polymer dynamics in dilute solution beyond
the change of relaxation times.
\section{Model and Parameters}
A solution is considered of $N_p$ linear flexible polymer chains
embedded in an explicit solvent. Each polymer is comprised of
$N_m$ beads of mass $M$, which are connected by linear springs of
equilibrium bond length $l$ \cite{gomp:09,huan:10_1}. Inter- and
intramolecular excluded-volume interactions are taken into account
by the repulsive, shifted and truncated Lennard-Jones potential,
with the parameter $\sigma$ characterizing the bead size and
$\epsilon$ the energy \cite{huan:10_1}. The monomer dynamics is
determined by Newton's equations of motion, which are integrated
by the velocity Verlet algorithm with time step $h_p$
\cite{alle:87}.
The solvent is simulated by the multiparticle collision dynamics
(MPC) method \cite{male:99,kapr:08,gomp:09}. It is composed of
$N_s$ point-like particles of mass $m$. The algorithm consists of
alternating streaming and collision steps. In the streaming step,
the solvent particles move ballistically for a time $h$. In the
collision steps, particles are sorted into cubic cells of side
length $a$ and their relative velocities, with respect to the
center-of-mass velocity of their cell, are rotated around a
randomly oriented axis by a fixed angle $\alpha$.
The solvent-polymer coupling is achieved by taking the monomers
into account in the collision step. To insure Galilean invariance,
a random shift is performed at every collision
step~\cite{ihle:01}. The collision step is a stochastic process,
where mass, momentum and energy are conserved, which leads to the
build-up of correlations between the particles and gives rise to
hydrodynamic interactions \cite{gomp:09}.
Three-dimensional periodic boundary conditions are considered for
the simulation of shorter chains. Here, Lees-Edwards boundary
conditions are applied to impose a shear flow \cite{alle:87}. A
local Maxwellian thermostat is used to maintain the temperature of
the fluid at the desired value~\cite{huan:10}. A parallel MPC
algorithm is exploited for systems of longer chains, which is
based on a three-dimensional domain-decomposition approach
\cite{huan:10_1}. In such a system, shear flow is imposed by the
opposite movement of two confining walls, and periodic boundary
conditions are applied parallel to them. We impose no-slip
boundary conditions at walls for both, fluid particles and
monomers~\cite{lamu:01,wink:09,gomp:09}.
Non-hydrodynamic simulations are performed by Brownian MPC, where
each monomer independently performs stochastic collisions with a
phantom particle which mimics a fluid element of size $a^3$. In
shear flow, the phantom-particle momentum is taken from a
Maxwell-Boltzmann distribution with mean $ \langle p_x \rangle = m
\left\langle N_c \right\rangle \dot \gamma r_y$ and variance
$\langle p^2_{\beta}\rangle = m \left\langle N_c \right\rangle
k_BT$ ($\beta \in \{x,y,z\}$), where $\left\langle N_c
\right\rangle$ is the average number of solvent particles per
collision cell, $r_y$ the particle position along the gradient
direction, and $\langle p_x \rangle$ the momentum along the flow
direction \cite{ripo:07,gomp:09}.
We employ the parameters $\alpha = 130^{\circ}$, $h= 0.1 \tilde
\tau$, with $\tilde \tau = \sqrt{ma^2/(k_BT)}$ ($k_B$ is
Boltzmann's constant and T is temperature), $\left\langle N_c
\right\rangle = 10$, $M=m\left\langle N_c \right\rangle$,
$l=\sigma =a$, $k_BT/\epsilon =1$, $h/h_p = 50$, and the bond
spring constant $\kappa =5\times10^3 k_BT/a^2$. The polymer
lengths $N_m=50$ and $250$ are considered in the concentration
ranges $c/c^* = 0.16 - 2.08$ and $0.17 - 10.38$, respectively. The
corresponding overlap concentrations are $c^* =0.098 l^3$ and
$c^*=0.029 l^3$, determined by their radii of gyration. In dilute
solution, the equilibrium end-to-end vector relaxation times are
$\tau_0 = 6169 \tilde \tau$ and $78330 \tilde \tau$
\cite{huan:10_1}. The Brownian MPC simulations for $N_m=50$ yield
an approximately five times larger relaxation time than
hydrodynamic MPC.
\section{Tumbling Dynamics}
\begin{figure}[t]
\begin{center}
\includegraphics*[width=.23\textwidth,angle=270]{figure_1.eps}
\caption{Snapshot of a system of $N_p=800$ polymers of length $N_m=250$ for the
Weissenberg number $\mathrm{Wi}_c=184$ at the concentration $c/c^* = 2.77$.
For illustration, some chains are highlighted in red. Animations are provided as
supplementary material. low.mov shows the polymer dynamics
under shear for the Weissenberg number $\mathrm{Wi}_c = 18$ and heigh.mov for $\mathrm{Wi}_c = 184$.}
\label{snap}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics*[width=8cm]{figure_2.eps}\\[1ex]
\caption{(a) Monomer density distribution in the
flow-gradient plane for $N_p=3000$, $N_m=250$, i.e., $c/c^*=10.38$, and
$\mathrm{Wi}_c=569$.
The contour lines for the densities $0.1$ (outer) and $0.5$ (inner) are highlighted
to emphasize the non-ellipticity of the shape.
(b) Illustration of polymer stretching (right)
and recoiling (left).
$\theta$ is the angle between the end-to-end vector and
its projection onto the flow-gradient plane and $\varphi$ is the angle
between this projection and the flow direction. }
\label{den_illu}
\end{center}
\end{figure}
The snapshot of a semidilute solution, displayed in
Fig.~\ref{snap}, indicates large conformational differences
between the various polymers in flow. The average shape of an
individual chain is illustrated in Fig.~\ref{den_illu}(a) by the
monomer density distribution in the flow-gradient plane. Their
conformational and rheological properties are discussed in detail
in Ref. \cite{huan:10_1}. The large conformational variations are
due to a continuous end-over-end tumbling motion
\cite{smith:99,schr:05}, with stretched and coiled states as
depicted in Fig.~\ref{den_illu}(b).
The instantaneous shape of a polymer is characterized by the
radius of gyration tensor $G_{\beta \beta'}$ ($\beta, \beta' \in
\{x,y,z\}$), which is defined as
$G_{\beta\beta'}=\sum^{N_m}_{i=1}\langle\Delta r_{i,\beta} \Delta
r_{i,\beta'}\rangle /N_m$, where $\Delta {\bm r}_{i}$ is the
position of monomer $i$ in the center-of-mass reference frame of
the polymer. To find a characteristic time for the tumbling
dynamics, we determine the cross-correlation function
\begin{equation} \label{correlation}
C_{xy}(t)=\frac{\left\langle G_{xx}'(t_0)G_{yy}'(t_0+t)\right\rangle }
{\sqrt{ \left\langle G_{xx}'^{2}(t_0) \right\rangle \left\langle G_{yy}'^{2}(t_0) \right\rangle }} ,
\end{equation}
for deviations from average stationary values
$G_{\beta\beta}'(t)=G_{\beta\beta}(t)- \left\langle G_{\beta\beta} \right\rangle$.
Figure~\ref{cross}
shows cross-correlation functions for several shear rates and
concentrations. Each of the curves exhibits a deep minimum at time
$t_{+}>0$ and a pronounced peak at time $t_{-}<0$, and decays to
zero at large time-lags. Hence, the tumbling dynamics is not
periodic, but cyclic. The latter has been questioned for tethered
polymers \cite{zhan:09}. The minimum at $t_{+}$ indicates that
positive values of $G_{\beta \beta}'$ are linked with negative
ones of the orthogonal directions, i.e., polymer shrinkage in the
$y$-direction is linked with its extension in $x$-direction, and
similarly, an extension in $y$-direction is linked to shrinkage in
$x$-direction. The maximum of $C_{xy}(t)$ reveals that positive
deviations $G_{xx}'$ are correlated with positive values $G_{yy}'$
at earlier times, or a collapsed state along the $x$-direction
($G_{xx}' <0$) is correlated with a previous collapsed state in
$y$-direction \cite{schr:05}. Hence, the time difference $t_{+}-
t_{-}$ is related to conformational changes that a polymer
undergoes during tumbling. We therefore characterize tumbling by
the time $\tau_t = 2(t_{+}-t_{-})$. The factor two is introduced,
because two non-equivalent conformations lead to a maximum and a
minimum, respectively, and will be (more or less) assumed during a
cycle.
As shown in Fig.~\ref{cross}, the positions of the peaks and
minima are rather close for equal Weissenberg numbers, when the
lag-time is scaled by the relaxation time $\tau(c)$. Hence, the
tumbling times exhibit a strong concentration dependence due to
the strong concentration dependence of the relaxation times, which
is shown in the inset of Fig. 3 \cite{huan:10_1}.
\begin{figure}[t]
\begin{center}
\includegraphics*[width=.38\textwidth]{figure_3.eps}
\caption{Cross-correlation functions [Eq. (\ref{correlation})]
for a polymer of length $N_m=250$ and the concentrations $c/c^*=2.77$
(black solid line), $c/c^*=10.38$, (red dashed line),
and $c/c^*=0.35$ (blue solid line),
corresponding to the Weissenberg numbers $\mathrm{Wi}_c=5520$, $\mathrm{Wi}_c=5690$,
and $\mathrm{Wi}_c=2670$, respectively, as well as $N_m=50$ with $c/c^*=2.08$
(green dashed-dotted line) and $c/c^*=0.8$ (green dashed line),
the Weissenberg numbers
are $\mathrm{Wi}_c = 220$ and $\mathrm{Wi}_c=233$, respectively.
Inset: Longest polymer relaxation times for
$N_m=50$ ($\bullet$) and $N_m=250$ (${\color{red}\blacksquare}$).} \label{cross}
\end{center}
\end{figure}
Normalized tumbling frequencies $f= \tau(c)/\tau_t$, with tumbling
times extracted from the correlation functions and scaled by the
corresponding relaxation times $\tau(c)$, are presented in
Fig.~\ref{tumb_freq} for a wide range of shear rates and
concentrations. For comparison, the theoretical prediction for a
polymer in dilute solution is presented as well
\cite{wink:06_1,wink:10}. The results are in excellent agreement,
when the Weissenberg number $\mathrm{Wi}^*$ of the theoretical model is
identified with $\mathrm{Wi}^*=\mathrm{Wi}_c/2$, where the factor two accounts for
the approximately twice larger relaxation time of the theoretical
model. The short chain results clearly show the crossover from
unity, assumed in the limit $\dot \gamma \to 0$, to the asymptotic
dependence $\sim \mathrm{Wi}_c^{2/3}$ at high shear rates. We obtain a
chain-length dependence in close agreement with the theoretical
prediction. More importantly, we find a slight and gradual shift
of $f$ to larger values with increasing concentration at a given
$\mathrm{Wi}_c$, until a saturation is reached in the semidilute regime
$c/c^*>1$. This is seen for the two largest concentrations for
$N_m=50$ and the three largest ones for $N_m=250$. As a
consequence, the polymers exhibit a universal behavior, both in
dilute ($c \ll c^*$) as well as in semidilute solution as function
of $\mathrm{Wi}_c$, with the same power-law dependence on $\mathrm{Wi}_c$ (for $\mathrm{Wi}_c
>1$).
Theory~\cite{wink:06_1}, simulations~\cite{puli:05}, and
experiments~\cite{gera:06} suggest that tumbling is an aperiodic
process with an exponential distribution of intervals between
tumbling events, i.e., $P_t(t)\sim \exp(-t/\tau^e_{t})$, where
$\tau^e_{t}$ is defined as the characteristic tumbling time. By
calculating the distribution functions of times between
successive gradient-vorticity plane crossings of the end-to-end
vector, we determined the tumbling times $\tau^e_{t}$ presented in
the inset of Fig~\ref{tumb_freq}. Their dependence on Weissenberg
number and concentration is in perfect accord with that obtained
for $\tau_t$; the absolute values are only approximately $20\%$
smaller.
Alternatively, relaxation times under shear flow can be obtained
by the end-to-end vector auto-correlation function $\left\langle
R_{\beta} (t) R_{\beta}(0)\right\rangle$, where $R_{\beta} =
r_{N,\beta}-r_{1,\beta}$ \cite{wink:06_1,delg:06,jose:08}. Similar
to the results presented in Ref. \cite{jose:08}, we find a damped
oscillatory behavior for $\mathrm{Wi}_c \gtrsim 1$. A fit of $g(t) =
e^{-t/\tau_{r}}[\cos(\omega t) + a \sin(\omega t)]$ to the
correlation functions along the flow and gradient directions,
yields, within the accuracy of the simulations, relaxation times
$\tau_{r} (\dot \gamma)$ and normalized tumbling frequencies $f=
\tau(c)/\tau_{r}$ equal to the values presented in
Fig.~\ref{tumb_freq}. The parameter $\omega$ of $g(t)$ is
independent of shear rate for the considered systems. This is in
contrast to results of Ref.~\cite{jose:08}, where very short
polymers ($N_m=10$) have been considered only.
The simulation results for the tumbling time lead to the following
conclusions. (i) The correlation function (\ref{correlation}) can
be used to obtain a characteristic time $\tau_t$ for the tumbling
motion (see also Ref. \cite{teix:05}). (ii) The non-equilibrium
end-to-end distance relaxation time $\tau_{r}$ along the flow
direction is equal to the tumbling time $\tau_t$ extracted from
the correlation function (\ref{correlation}). Hence, the tumbling
time $\tau_t$ is equal to the non-equilibrium relaxation time
$\tau_{r}$, as also predicted in Refs.~\cite{wink:06_1,zhan:09}.
(iii) There is a weak chain-length dependence of the tumbling
time. (iv) Screening of hydrodynamic interactions in semidilute
solutions leads to a (small) increase of the normalized tumbling
frequencies. The screening aspects are discussed in more detail in
then next section.
\begin{figure}[t]
\begin{center}
\includegraphics*[width=.38\textwidth]{figure_4.eps}
\caption{Normalized tumbling frequencies $f=\tau(c)/\tau_t$ obtained from
cross-correlation functions.
Open symbols correspond to the polymer length
$N_m=50$ for $c/c^*=0.16$ ($\circ$), $c/c^*=0.4$
($\triangledown$), $c/c^*=0.8$ ($\color{red}\triangle$), $c/c^*=1.6$
(${\Diamond}$), and $c/c^*=2.08$
(${\color{blue}\square}$). Filled symbols indicate results for
$N_m=250$ and $c/c^*=0.35$ ($\bullet$),
$c/c^*=0.69$ ($\blacktriangleright$), $c/c^*=1.38$
($\blacktriangledown$), $c/c^*=2.77$
(${\color{green}\blacklozenge}$), $c/c^*=5.19$
($+$), $c/c^*=10.38$
(${\color{blue}\blacksquare}$). The lines present
the theoretical predictions \cite{wink:06_1,wink:10}.
Inset: Relaxation times $\tau^e_t$ for $N_m=50$ and various concentrations.}
\label{tumb_freq}
\end{center}
\end{figure}
\section{Angular Probability Distribution Functions}
Further insight into the tumbling and orientational behavior of
polymers is gained by the probability distribution functions
(PDFs) $P(\varphi)$ and $P(\theta)$ for the orientation angles
$\varphi$ and $\theta$ (cf. Fig.~\ref{den_illu}(b))
\cite{wink:06_1,puli:05,gera:06,cela:05}. For a \textit{dilute}
solution, $P(\varphi)$ is shown in Fig.~\ref{P_phi_dilute},
together with theoretical lines obtained from
Ref.~\cite{wink:06_1} (note, $\mathrm{Wi}^* = \mathrm{Wi}_c/2$). Evidently, the
simulation results agreement well with the analytical approach,
as is expected for a dilute solution in which the intermolecular
interactions are irrelevant. $P(\varphi)$ exhibits a significant
shear-rate dependence. At zero shear, no angle is preferred. An
increasing shear rate leads to the appearance of a peak, which
shifts to smaller values with increasing $\dot \gamma$ and, at the
same time, the width $\Delta \varphi$ of $P(\varphi)$ decreases.
\begin{figure}[t]
\begin{center}
\includegraphics*[width=.38\textwidth]{figure_5.eps}
\caption{Probability distribution functions of the angle $\varphi$ for a
dilute solution with $c/c^*=0.16$ of polymers of length $N_m=50$ for
$\mathrm{Wi}_c=616.9$
(${\color{blue}\blacksquare}$), $\mathrm{Wi}_c=61.7$ ($\circ$), and $\mathrm{Wi}_c=12.3$
(${\color{green}\blacktriangle}$). Inset: Distribution functions
for
$N_m=250$, $c/c^*=0.17$, and $\mathrm{Wi}_c=2350$
(${\color{blue}\circ}$), $\mathrm{Wi}_c=235$ (${\color{green}\times}$),
$\mathrm{Wi}_c=23.5$ (${\color{red}\square}$).
The solid lines are theoretical results for
Weissenberg numbers $\mathrm{Wi}^*=\mathrm{Wi}_c/2$.} \label{P_phi_dilute}
\end{center}
\end{figure}
In a \textit{semidilute} solution, the universality observed for
the tumbling time is also reflected in $P(\varphi)$.
Figure~\ref{P_phi_semidilute} displays distribution functions for
various concentrations and Weissenberg numbers $\mathrm{Wi}_c$. For every
Weissenberg number distributions are compared for three
concentrations. Evidently, the distributions are independent of
concentration for the considered Weissenberg numbers.
However, we observe a clear concentration dependence, when we
compare distributions from dilute and semidilute solutions. The
inset of Fig.~\ref{P_phi_semidilute}, displays distributions for
the concentrations $c/c^*=0.35, 5.19$ and the Weissenberg numbers
$\mathrm{Wi}_c \approx 267$, $2670$ and $270$, $2700$, respectively.
Clearly, the increase in concentration from a dilute solution
beyond the overlap concentration leads to a broadening of the
distribution function.
Interestingly, the value $\varphi_m$ at the peak of the
distribution function is independent of concentration at a given
$\mathrm{Wi}_c$.
Figure~\ref{angle_e_e} displays $\tan(2 \varphi_m)$ as function of
$\mathrm{Wi}_c$ for various concentrations of the two studied chain
lengths. The simulation data are consistent with the analytical
predictions for both chain lengths \cite{wink:06_1,wink:10}. The
results show that a universal behavior is obtained for the various
concentration at a given $N_m$. In the asymptotic limit of high
shear rates, the dependence
\begin{equation}
\tan(2\varphi_m)\sim (\mathrm{Wi}_c)^{-1/3}
\end{equation}
is obtained, which has also been found in Ref.~\cite{huan:10_1}
for the alignment angle determined from the gyration tensor. In
the limit of $\mathrm{Wi}_c\rightarrow 0$, theory predicts
$\tan(2\varphi_m)\sim \mathrm{Wi}_c^{-1}$, whereas the simulations yield
$\tan(2\varphi_m)\sim \mathrm{Wi}_c^{-0.8}$ for the considered range of
$\mathrm{Wi}_c$, which might be due to excluded volume interactions not
taken into account in the theoretical calculations. The inset of
Fig.~\ref{angle_e_e} displays $\Delta \varphi$, the full width at
half maximum. For dilute solutions, $\Delta \varphi$ agrees with
the prediction of the theoretical model, which yields the
asymptotic dependence $\Delta \varphi \sim \mathrm{Wi}_c^{-1/3}$ in the
limit $\mathrm{Wi}_c \to \infty$. The widths of the distributions are
larger for semidilute solutions, but the asymptotic Weissenberg
number dependence seems to be the same. At $c/c^*>1$ a universal
curve is adopted, as already pointed out above in connection with
Fig~\ref{P_phi_semidilute}.
\begin{figure}[t]
\begin{center}
\includegraphics*[width=.38\textwidth]{figure_6.eps}
\caption{Probability distribution functions $P(\varphi)$ of polymers of
length $N_m = 250$ for
$c/c^*=2.77$ with $\mathrm{Wi}_c=5520$ (${\color{red}\blacksquare}$),
$\mathrm{Wi}_c=552$ (${\color{blue}\blacklozenge}$),
and $\mathrm{Wi}_c=55.2$ (${\color{green}\blacktriangle}$), for
$c/c^*=5.19$ with $\mathrm{Wi}_c=5423$ ($\circ$), $\mathrm{Wi}_c=542.3$ ($\square$),
and $\mathrm{Wi}_c=54.23$ ($+$), as well as $c/c^*=10.38$ with $\mathrm{Wi}_c=5691$ (solid line),
$\mathrm{Wi}_c=569.1$ (dashed line), $\mathrm{Wi}_c=56.91$ (dashed-dotted line).
Inset: Distribution functions for $c/c^*=0.35$ and
$\mathrm{Wi}_c=2670$ (black solid line), $\mathrm{Wi}_c=267$ (red solid line) as well as
for $c/c^*=5.19$
with $\mathrm{Wi}_c=2700$ (black dashed line) and $\mathrm{Wi}_c=270$ (red
dashed line).} \label{P_phi_semidilute}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics*[width=.38\textwidth]{figure_7.eps}
\caption{Maximum angle $\varphi_m$ as a
function of $\mathrm{Wi}_c$ for $N_m=250$ and
$c/c^*=0.17$ ($\circ$), $0.35$ ($\blacktriangledown$), $2.77$
(${\color{green}\blacklozenge}$), $5.19$
(${\color{red}\blacktriangle}$), $10.38$
(${\color{blue}\blacksquare}$), as well as $N_m=50$ for $c/c^*=0.16$
($\blacktriangleright$), $1.6$ ($\times$), and $2.08$ ($\square$).
Inset: Widths $\Delta \varphi$
of the distribution functions.
For both, the lines are theoretical predictions.}
\label{angle_e_e}
\end{center}
\end{figure}
Probability distribution functions of the angle $\theta$ are
displayed in Fig.~\ref{theta_fig} for various Weissenberg numbers.
Theoretical calculations for dilute solutions predict a crossover
from a Gaussian shape of the distribution function to a power-law
decay according to $P(\theta)\sim\theta^{-2}$ with increasing
shear rate, within a certain range of angles, which is confirmed
by the simulations. The dependence of $P(\theta)$ on concentration
is in accord with that of $P(\varphi)$. In the semidilute regime,
$P(\theta)$ is independent of concentration for a given $\mathrm{Wi}_c$,
while a comparison of distributions of \textit{semidilute} and
\textit{dilute} solutions yields a broadening for the semidilute
case.
\begin{figure}[t]
\begin{center}
\includegraphics*[width=.38\textwidth]{figure_8.eps}
\caption{Probability distribution functions $P(\theta)$ of polymers of
length $N_m = 250$ for
$c/c^*=2.77$ with $\mathrm{Wi}_c=5520$ (${\color{red}\blacksquare}$),
$552$ (${\color{blue}\blacklozenge}$), $55.2$
(${\color{green}\blacktriangle}$), and $0.552$ ($\circ$),
as well as $c/c^*=10.38$ with $\mathrm{Wi}_c=5691$ (solid line),
$569$ (dashed line), $57$ (dashed-dotted line), and $0.57$
(thin line).
The slope of the short solid lines is $-2$.
Inset: $P(\theta)$s of a dilute solution with $c/c^*=0.35$ for
$\mathrm{Wi}_c=2670$ (black solid line) and $\mathrm{Wi}_c=267$ (red solid line) and a
semidilute solution with $c/c^*=5.19$
for $\mathrm{Wi}_c=2700$ (black dashed line) and $\mathrm{Wi}_c=270$ (red dashed line).}
\label{theta_fig}
\end{center}
\end{figure}
We attribute the concentration independence of the tumbling time
and the probability distribution functions for $c/c^*> 1$ to
screening of hydrodynamic interactions. To confirm our hypothesis,
we performed Brownian MPC simulations for dilute and semidilute
solutions. Using similar Weissenberg numbers, we find, within the
accuracy of the simulations, identical distribution functions
$P(\varphi)$ for both cases. Moreover, the distributions agree
with those of semidilute systems of the same concentration and
Weissenberg number in the presence of hydrodynamic interactions.
Hence, the differences between distribution functions at low and
high concentrations, as displayed in the insets of Figs.~6 and 8,
are due to hydrodynamic interactions. In dilute solutions,
hydrodynamic interactions are present, whereas in systems with $c
> c^*$, hydrodynamic interactions are screened. Naturally, at
larger concentrations friction is higher. This aspect is captured
in the relaxation time $\tau(c)$, which increases considerably
with concentration \cite{huan:10_1}.
The fact that the tumbling frequencies and the widths of the
distribution functions are larger in semidilute solutions, i.e.,
when hydrodynamic interactions are screened, might be explained as
follows. (i) The observed broadening of the distribution function
$P(\varphi)$ in semidilute solution implies that hydrodynamic
interactions favor polymer alignment and lead to a faster dynamics
during the collapse and stretching part of the tumbling motion.
(ii) At the same Weissenberg number, the shear rate of a
non-draining polymer is larger than that of a free-draining one,
due to differences in equilibrium relaxation times, i.e., for a
coiled conformation. As a consequence, the effective Weissenberg
number $\mathrm{Wi}_R= \dot \gamma \tau_R$, where $\tau_R$ is the
rotational relaxation time in the stretched rodlike conformation,
of the non-draining polymer is larger than that of the
free-draining one. This could explain the larger probability of
angles in the vicinity of $\varphi_m$ for non-draining polymers as
well as their faster collapse dynamics. Overall, the tumbling time
is larger in a non-draining system.
The broadening of the distribution functions with increasing
concentration or screening of hydrodynamic interactions is not
captured by standard theories employing the preaveraging
approximation \cite{doi:86,wink:06_1,wink:10}. Here, hydrodynamic
interactions are included in the relaxation times and hence the
Weissenberg number only. Additional, ``higher order effects'' are
neglected. Therefore, one might expect that the theoretical
description would reproduce results of simulations without
hydrodynamic interactions, in contrast, the model calculations
rather reproduce the simulation data for systems with hydrodynamic
interactions.
\section{V. Conclusions}
We have found in Ref. \cite{huan:10_1} that in shear flow the
stationary-state conformational polymer properties are independent
of concentration when expressed in terms of the Weissenberg number
$\mathrm{Wi}_c = \dot \gamma \tau(c)$. This is remarkable, since the
longest polymer relaxation time $\tau(c)$ increases significantly
with concentration and indicates that an effective local friction
determines the stationary-state properties.
Here, we have analyzed dynamical properties---orientational
distribution functions and tumbling times---of semidilute polymer
solutions in shear flow and have found that they depend on
concentration (in excess of $\tau(c)$), a dependence which we
attribute to screening of hydrodynamic interactions in semidilute
solution. Compared to the dilute case, such a screening causes a
broadening of orientational angle distribution functions and an
increasing ratio $f=\tau_t/\tau(c)$ at the same Weissenberg number
$\mathrm{Wi}_c$ in semidilute solution. The effect itself is small
($f(c=0)/f(c>c^*) \approx 1.3$ at $\mathrm{Wi}_c = 10^3$). More
importantly, the same asymptotic dependencies are obtained as
function of the Weissenberg number $\mathrm{Wi}_c$ in dilute and
semidilute solutions. This explains the previously obtained
agreement of power spectral densities obtained from free-draining
and non-draining computer simulations \cite{schr:05}.
\acknowledgments
Financial support by the Deutsche Forschungsgemeinschaft
within SFB TR6 is gratefully acknowledged. We are grateful to the
J\"ulich Supercomputer Centre (JSC) for allocation of a special
CPU-time grant.
|
\section{Introduction}
\label{sec:introduction}
Dusty disks made up of rocky and icy debris have been observed around
other stars, both in reflected optical light \citep{Smith:1984} and in
long wavelength thermal radiation \citep{Aumann}. Multiple surveys
have reported that a significant fraction of main-sequence stars
harbor detectable infrared excesses: $\sim 15\%$ solar-type stars
\citep{Trillingetal:2008,Lawler}, and $\sim 30\%$ for A-stars
\citep{Suetal06}. The infrared luminosity, when compared to the
luminosity of the central star, ranges from $\sim 10^{-5}$ to $\sim
10^{-3}$. In contrast, the fractional dust luminosity from the Kuiper
belt is estimated to be $\sim 10^{-7}$ \citep{Teplitzetal:1999} and
remains undetected.
The observed excess luminosities arise primarily from small ($\sim \mu
m - mm$) dust grains. Due to their short survival time \citep{Pawel}, these grains are
believed to be continuously produced by collisions between large parent
bodies (`planetesimals'). These planetesimals, analogous to the Kuiper
belt objects in our own system, are in turn left-overs from the epoch
of planet formation.
In this article, we describe how we can use debris disks to test
theories of planetesimal formation. We first focus our attention on
the primordial size spectrum of planetesimals, often characterized by a
single power-law, $dn/ds \propto s^{-q}$, where $s$ is the size. In
the following, we briefly summarize theoretical understandings and
observational evidences for the value of $q$.
The conventional picture of planetesimal formation is composed of a
number of steps. The formation of the first generation planetesimals
is not yet well-understood and is an area of active research
\citep[see, e. g.][]{Youdin:2002,Dominik:2007,
Johansen:2007,Garaud:2007}. If these are sufficiently massive,
gravity dominates their subsequent growth \citep{Weidenschilling}. At
first, objects grow in an orderly fashion, where collisions and
conglomerations occur at rates that are proportional to their
geometric cross sections. But when these bodies become so massive
that the effect of gravitational focusing becomes significant,
run-away growth commences where the largest bodies accrete small
planetesimals at the highest rate and quickly distance themselves from
their former peers \citep{WetherillStewart,Kokubo96}. The run-away
phase is succeeded by the oligarchic phase where individual large
bodies are responsible for stirring the small bodies that they accrete
\citep{KokuboIda,Kokubo:1998}. At the end of these steps, an entire
size spectrum of planetesimals are produced. This is the `primordial
spectrum'.
During the run-away phase, N-body simulations have typically produced
a slope of $q \sim 6$ \citep{Kokubo96,Morishima:2008}.
This slope is naturally explained if there is energy equi-partition
among planetesimals of different sizes \citep{Makino:1998}. Moreover,
one expects that the distribution becomes shallower (smaller $q$) if
larger planetesimals have higher kinetic energies. This indeed occurs
during the oligarchic phase when all small and intermediate-sized
planetesimals are stirred to the same velocity dispersion. The value of
$q$ is then reduced to $\approx 4$ \citep{Morishima:2008}.
Using particles-in-a-box simulations and later hybrid simulations,
\citet{KenyonandLuu:1999, Kenyon:2004, 2008ApJS..179..451K} followed
the growth of planetesimals. They also found that $q$ decreases with
time after the run-away phase, finishing up with $3.75 \leq q \leq 4.5$
for planetesimals of sizes between $10$ and $1000$~kms. Recently,
\citet{2011ApJ...728...68S} argued analytically that a $q=4$
spectrum is the natural outcome of conglomeration.
Observational constraints on the value of $q$ currently come
exclusively from counting large Kuiper belt objects. Kuiper belt
objects larger than about $30-50$~kms are commonly believed to be
primordial. Collision timescales for these bodies well exceed that of
the Solar system age
\citep{1997Icar..125...50D,2010AJ....139.1499B}. The size distribution
for these bodies can be probed by present-day surveys. Published
values for $q$ are scattered: $q =4.0_{-0.6}^{+0.5}$
\citep{Trujillo:2001}, $q = 4.25 \pm 0.25$ \citep{Fraser:2008b}, $q =
4.5\pm 0.4$ \citep{2009AJ....137...72F} and $q = 4.5_{-0.5}^{+1.0}$
\citep{2008AJ....136...83F}. This scatter may be intrinsic and reflect
both the different size ranges and the different dynamical populations
emphasized by various surveys
\citep{Bernstein,2006P&SS...54..243D,FraserBrown}. For bodies
smaller than $\sim 30$~kms, the size distribution adopts a shallower
power-law \citep{Bernstein,2008AJ....136...83F,2009Natur.462..895S}.
This break in the power-law index has been argued to be due to
collisional erosion \citep{PanSari}, but a different opinion has
surfaced \citep{morbidelli}.
So at least for the value of $q$, current coagulation models appear to
be vindicated by the observations. These models enjoy a further
success. In the Kuiper belt region, the solid mass of the so-called
Minimum Mass Solar Nebula is $\sim 10~M_\oplus$
\citep{Hayashi,Weidenschilling:1977}, while the mass in large Kuiper
belt objects is estimated to be $\lesssim 0.1 M_\oplus$ \citep[see,
e.g.][]{Gladman,Bernstein}. This large difference, however, is
explained by current models where the formation of large planetesimals
has a very low efficiency
\citep{2006AJ....131.2737B,2011ApJ...728...68S}.
With these two remarkable concordances, one wonders if debris disks
will ever tell us anything new and unexpected. Furthermore, every
debris disk likely has a different initial condition and evolves in a
different dynamical environment. For instance, dynamical interactions
with Neptune or other planets may have qualitatively affected the
evolution of the Kuiper belt \citep{Levison:2008}. It seems
difficult, therefore, to extract any universal truth about the
formation process from these disparate objects.
However, based only on a modest sample of debris disks, we argue in
this paper that there is already a serious issue in current
coagulation models.
To achieve this, we first construct a simple collisional model (\S
\ref{sec:luminosityevolution}) to compare against the set of debris
disks reported in \citet{Hillenbrand:2008}. Our collisional model
does not differ in essence from previous works
\citep{Krivov05,Wyattetal:2007,Lohne:2008}, but we interpret the
observations in a new way. This allows us to measure the value of $q$
as well as the initial masses of planetesimal belts (\S
\ref{sec:results}). The latter result challenges the current models of
planetesimal formation (\ref{sec:discussions}). We summarize in \S
\ref{sec:summary}.
\section{Model: Luminosity Evolution of a Debris Disk}
\label{sec:2}
\label{sec:luminosityevolution}
The debris phase commences when eccentricities of the primordial
planetesimals are further increased so that they no longer coalesce at
encounter, but are instead broken into
fragments.\footnote{\citet{2008ApJS..179..451K} find that
fragmentation begins once Pluto-sized bodies form.} In this phase,
the smallest primordial planetesimals enter into a collisional cascade
first, followed by progressively larger bodies. During the collisional
cascade, a primordial body is broken down into smaller and smaller
fragments until all its mass ends up in small grains. The small grains
may spiral in towards the star due to Poynting-Robertson drag, as
happens in the Solar system, or, be ground down by frequent collisions
to sizes so small that they are promptly removed by radiation
pressure, as happens in bright debris disks \citep{Wyatt:2005}.
\subsection{Debris Rings}
We model the debris disk as a single, azimuthally smooth ring composed
of planetesimals of different sizes. The ring is centered at a
semi-major axis $a$ with a full radial width of $\Delta a$ and a
constant surface density. We take $\Delta a/a = 0.1 \ll 1$ as our
standard input.
This is motivated by the following observations. Spatially resolved
debris disks often appear as narrow rings. Examples are, $\Delta a/a$
$\sim 0.1$ for AU Microscopii \citep{2007ApJ...670..536F}, $\sim 0.5$
for HD 10647 \citep{2010A&A...518L.132L}, $\sim 0.3$ for HD 92945
\citep{2007lyot.confE..46G}, $\sim 0.3$ for HD 139664
\citep{2006ApJ...637L..57K}, $\sim 0.2$ for HD 207129,
\citep{2010AJ....140.1051K}, $\sim 0.5$ for $\epsilon$ Eridani
\citep{2000MNRAS.314..702D}, $\sim 0.1$ for Fomalhaut
\citep{Kalas:2005}, $\sim 0.2$ for Vega \citep{2005ApJ...628..487S}.
Similarly, unresolved disks often exhibit spectral energy distribution
that is well fit by a single temperature blackbody
\citep{Hillenbrand:2008, 2010A&A...518A..40N,Moor}. This ring-like
topology also show up in our own Solar system, hence the name the
asteroid ``belt'' and the Kuiper ``belt''.
\subsection{Initial Size Distribution of the Planetesimals}
\label{subsubsec:dnds}
We adopt the following power-law forms for the initial size
distributions,
\begin{equation}
\left. \frac{dn}{ds}\right|_{t=0} \propto
\begin{cases}
s^{-q_3} \quad \quad s_{\rm small} < s < s_{\rm big} , \\
s^{-q_1} \quad \quad s_{\rm min} < s < s_{\rm small} .\\
\end{cases}
\label{eq:time0}
\end{equation}
The index $q_3$ is the primordial size index for large bodies, like
one that arises out of conglomeration models. Previous studies of
collisional debris disks have taken this value to be a given, in fact
it is commonly set to be the power law one expects from collisional
equilibrium \citep{Krivov05,Krivov06,Wyattetal:2007,Lohne:2008}.
In contrast, in this contribution we use the observed sample to
measure this value.
In equation \refnew{eq:time0}, $s_{\rm big}$ is the size of the biggest
planetesimals, $s_{\rm min}$ the smallest. The intermediate size
$s_{\rm small}$ is introduced for the purpose of mass accounting: the
original mass counts only those between $s_{\rm big}$ and $s_{\rm
small}$,
\begin{equation}
M_0 = \int_{s_{\rm small}}^{s_{\rm{big}}} \frac {4\pi}{3}\rho s^3\, n_3s^{-q_3}\,ds.
\label{eq:m0defined}
\end{equation}
While $s_{\rm min}$ is naturally taken to be the size at which
radiation pressure unbinds dust grains from the star ( $\sim \mu m$
for a Sun-like star), we discuss our choice for $s_{\rm big}$ and
$s_{\rm small}$ below.
Motivated by the observational and numerical results discussed in \S
\ref{sec:introduction}, we investigate values of $q_3$ between $3.5$
and $5$. The value $q_3 = 4$ has the special property that mass is
distributed equally among all logarithmic size ranges, while masses in
systems with $q_3 > 4$ diverge toward the small end. The intermediate
size $s_{\rm small}$ is introduced, partly to avoid dealing with this
divergence. For sizes below $s_{\rm small}$, we assume that
collisions have set up an equilibrium power law with index $q_1$ (see
Appendix). So, the intermediate size $s_{\rm small}$ can also be
interpreted as the collisional break size at time zero. For our study,
we set $s_{\rm small} = 100$ m.
For our typical disks, we find that, within a few million years,
collisional equilibrium is established for bodies up to sizes $\sim 1$
km. So the choice of $s_{\rm small}$ is not important for late time
evolution.
The choice of size for the largest bodies, $s_{\rm big}$, deserves
some discussion, as it affects the qualitative character of the
evolution. As a collisional cascade progresses, bodies of larger and
larger sizes come into collisional equilibrium, opening up fresh mass
reserve to produce the small particles. Once the largest bodies enter
into collisional equilibrium, the dust production rate decays with
time as $L_{\rm IR} \propto t^{-1}$ \citep{2007ApJ...658..569W}.
Two previous studies \citep{Wyattetal:2007,Lohne:2008} have adopted
sizes for the largest bodies of $s_{\rm big} = 30$ and $74$ km,
respectively. For some of their disks, the largest bodies can enter
collision equilibrium during the lifetime of the system.
Both Kuiper belt observations and numerical studies of coagulation
favor a largest size of $\sim 1000$~km. The largest object yet found
in the Kuiper Belt, (136199) Eris, has a radius of $1200 \pm 50$~km
\citep{Brown:2006}. In the simulations of \citet{Kenyon:2004b},
coagulation of planetesimals at 30 - 150 AU produces bodies as large
as $1000$ - $3000$ km. When the largest bodies reach this size,
self-stirring increases the velocity dispersion and collisions become
destructive rather than conglomerating.
Therefore, we adopt a maximum body size of $1000$ km in our study. Our
quoted masses reflect this choice of $s_{\rm big}$. Our largest
bodies never enter into collisional equilibrium. If this assumption
turns out to be erroneous, namely, $s_{\rm big}$ is much smaller and
enters into collisional cascade within system lifetime, our model
would underestimate the initial masses for old disks. As a result, we
would overestimate the value for $q_3$.
\subsection{Collisions}
\label{subsubsec:collision}
We only consider collisions that are catastrophically destructive. A
catastrophic collision is defined as one that removes at least $50\%$
of the mass of the primary body. In so doing, we have implicitly
assumed that both cratering collisions and conglomerating collisions
are unimportant. When a destructive collision occurs, the total mass
(bullet plus target) is redistributed to all smaller sizes according
to $dn/ds \propto s^{-4}$. This choice is somewhat arbitrary and we
have confirmed that modifying it (within reasonable bounds) does not
change our results.
We do not model evolution of the orbital dynamics as bodies
collide. This is justified by the discussions in \S
\ref{subsec:eccentricity}.
Let the chance of collisions between two bodies of sizes $s$ and
$s^{\prime}$ be,
\begin{equation}
f_{\rm{collision}} = \frac{\pi \left(s + s^{\prime}\right)^2}
{2 \pi a \Delta a \, t_{\rm{orb}}},
\label{eq:fcol}
\end{equation}
Here, $2\pi a \Delta a$ is the surface area spanned by the debris ring
in the orbital plane, and $t_{\rm orb}$ is the orbital period.
Gravitational focusing is negligible for the high random velocities
we consider here. The typical encounter velocity, for particles with
eccentricity $e$ and inclination $i$, is \citep{Wetherill:1993}
\begin{equation}
v_{\rm{col}} = \sqrt{1.25e^2+i^2}\, v_{\rm{kep}},
\label{eq:vimpact}
\end{equation}
where $v_{\rm kep}$ is the local Keplerian velocity. We adopt $i
\approx e/2$ so $v_{\rm col} \approx 1.32\, e \, v_{\rm kep}$. As
argued in \S \ref{subsec:eccentricity}, it is reasonable to assume a
constant eccentricity (and inclination) for all bodies. We take a
value of $e=0.1$ as the standard input, and discuss this
assumption in \S \ref{sec:discussions}.
We denote the specific impact energy required to catastrophically
disrupt a body (target) as ${Q}^*$. The scaling of $Q^*$ with the size
of the target depends on whether its strength is dominated by material
cohesion or self-gravity. We adopt the following form
\citep{BenzAndAsphaug},
\begin{equation}
{Q}^*~=~A\left(\frac{s}{1~\mbox{cm}}\right)^\alpha+B\rho\left(\frac{s}{1~\mbox{cm}}\right)^\beta
\label{BenzAndAsphaug}
\end{equation}
where $\rho$ is the bulk density which we take to be $2.5
\rm{g}/\, \rm cm^3$. The first term on the right-hand-side describes the
internal strength limit, important for small bodies, while the second
term the self-gravity limit, important for larger bodies.
The strength law sets the size of the smallest bullets required to
destroy a target. Since these are also the most numerous, they
determine the downward conversion rate of mass during a collisional
cascade. As such, the power indexes in the strength law directly
determine the size spectrum at collisional equilibrium. For a strength
law of the form ${Q}^* \propto s^c$, the equilibrium size spectrum is
$dn/ds \propto s^{-q}$, with \citep{1997Icar..130..140D}:
\begin{equation}
q=(21+c)/(6+c).
\label{eq:whatisq}
\end{equation}
The famous Dohnanyi-law \citep{Dohnanyi:1969}, $dn/ds \propto
s^{-3.5}$, obtains from $c = 0$.
The value and form for $Q^*$ are notoriously difficult to assess. It
depends on, among other factors, material composition, porosity and
impact velocity. A number of computations and compilations have
appeared in the literature. We select three representative
formulations for our study (Fig. \ref{fig:Qstarcomp}).
Based on a variety of experimental data and SPH simulations,
\citet{Krivov05,Lohne:2008} advocated the following choices, $A = 2
\times 10^7 \, \rm erg/\rm g$, $\alpha = -0.3$, $B = 0.158$, $\beta = 1.5$. We
call this the 'hard' strength law. In this case, the collision
spectrum satisfies $q \approx 3.6$ and $3.0$, in the strength and
gravity regimes respectively.
Based on energy conservation, \citet{PanSari} calculated a destruction
threshold for bodies that have zero internal strength and obtained $B
= 3.3 \times 10^{-8}$, $\beta = 2$. So bodies at $100 \, \rm km$ is weaker
by a factor $\sim 1000$ than their counterparts in the
\citet{Krivov05} formulation. We refer to this as the 'soft' strength
law. A softer strength implies smaller bullets and therefore more
frequent destruction of the targets. \citet{PanSari} did not consider
smaller bodies that are strength bound. We adopt $A = 2 \times 10^{7}
\, \rm erg/g$ and $\alpha = -0.3$ in this range to complete the soft
prescription.
\cite{2009ApJ...691L.133S} proposed a strength law that depends on
impact velocity,
\begin{equation}
Q^* = \left(500\, s^{-0.33}+10^{-4}\, s^{1.2}\right)v_{\rm col}^{0.8},
\label{eq:sllaw}
\end{equation}
For a typical velocity $v_{\rm col}= 500$m/s and for bodies
greater than 1km, this gives rise to a strength law that falls
in-between that of the hard and the soft case. We call this the medium
strength law. Note that this strength law is much weaker than
the other two for small bodies.
For the strength laws we consider, transitions from material strength
domination to self-gravity domination occur at size $s \approx s_1$,
with $s_1$ ranging between $100$ m (the hard and the medium laws) and
$10$ km (the soft law).
\begin{figure}[t]
\begin{center}
\includegraphics[scale=.65, trim = 0 0 0 0, clip]{figure1.eps}
\caption{ Prescriptions for specific strength from \citet{Lohne:2008},
\citet{PanSari} and \citet{2009ApJ...691L.133S}, plotted here as
functions of target sizes. We insert an impact velocity of $500
\, \rm m/\, \rm s$ to evaluate the last prescription. Strength of small bodies
are dominated by material cohesion, while that of larger bodies by
self-gravity. Transitions between the two limits occur around
$100$ m (the hard and the medium laws) or around $10$ km (the soft
law). Strength for bodies smaller than $1 \, \rm cm$ are extrapolations
as both laboratory and numerical experiments only concern bodies of
larger sizes.}
\label{fig:Qstarcomp}
\end{center}
\end{figure}
\subsection{Luminosity Evolution}
\begin{figure}[]
\begin{center}
\includegraphics[scale=.65]{figure4.eps}
\caption{Time evolution of the break-size in a model system, with
$M_0 = 12.4 M_{\oplus}$, $q_3 = 4.0$, $e = 0.1$, $a = 31\rm{AU}$,
and $\Delta a/a = 0.1$. Here, break-size is defined as the size
at which all bodies initially at that size have encountered of
order one destructive collision. Break size increases with time
monotonically as larger bodies enter into collisional cascade.
The numerical results are shown as solid curves, while the
analytical scaling relations (see Appendix) are plotted as dashed
lines. The bends in the curves occur at $s \approx s_1$, i.e.,
sizes for which material cohesion and gravity binding are
comparable. The set of thick curves are for the case of hard
material strength, while the thin lines for soft strength. }
\label{fig:s2}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[scale=.65, trim = 0 0 0 0, clip]{figure2.eps}
\caption{Evolution of fractional luminosity, $L_{\rm IR}/L_*$, for the
system in Fig. \ref{fig:s2}. The thick line is obtained using the
hard strength law, and the thin line the soft one. The evolution
proceeds in two stages: the flatter early stage when collisional
cascade only involves small bodies that are bound by material
cohesion; and a steeper later stage where bodies bound by
self-gravity enter the cascade. At late times, fractional
luminosity decays as $t^{-0.5}$ (eq. \refnew{eq:ftime}). While
break-sizes differ for the two adopted strength laws
(Fig. \ref{fig:s2}), this appears to have little influence on the
overall luminosity. }
\label{fig:flHD1}
\end{center}
\end{figure}
The planetesimal disk, starting from an initial disk mass of $M_0$,
and an initial size spectrum (eq. \ref{eq:time0}), is numerically
collided and ground down. We divide the particles between $s_{\rm
small}$ and $s_{\rm big}$ into $500$ equal logarithmic size bins.
The time-step for the simulations is adaptively set so that over one
time-step, the maximum mass gain (from larger bodies) or loss (to
smaller bodies) per bin falls below $5\%$. The net mass change is
substantially smaller than this due to the cancellation between gain
and loss.
We calculate the fractional brightness of the dust disk, $L_{IR}/L_*$,
by integrating the geometrical cross section over all grains. This
assumes that grains are perfect absorbers at the optical and can emit
efficiently in the infrared.
An example of such a calculation is reported in Figs. \ref{fig:s2} \&
\ref{fig:flHD1}. To understand these results, a simple analytical
model (see Appendix) is introduced. Scaling relations obtained using
this analytical model compares well with our numerical results.
Fig. \ref{fig:s2} shows that, with time, larger and larger
planetesimals enter into collisional cascade. Within a million years or
so, the cascade has advanced to size of order one kilometer. Beyond
this time, bodies bound by self-gravity can be gradually eroded. By 1
Gyrs, bodies with sizes $10-100$ kms may be affected. The exact value
depends on the strength law. The dust luminosity is related to the
dust mass, which is in turn related to the dust production rate. The
dust production rate, on the other hand, is simply the primordial mass
stored at the break-size divided by the system age. If the primordial
spectrum is such that a large amount of mass is piled at the large
end, debris disks would not exhibit significant fading even up to a
few billion years.
Fig. \ref{fig:flHD1} shows that dust luminosity $L_{\rm IR}/L_*
\propto t^{-0.5}$ for $q_3 = 4$, consistent with equation
\refnew{eq:frac2}. That same equation also demonstrates that the value
of $B$, strength constant for bodies bound by self-gravity, affects
the luminosity only minorly. This is born out by
results shown in Fig. \ref{fig:flHD1}.
An important result on which we base our later analysis is shown in
Fig. \ref{fig:varyq}. Luminosity evolution for disks with the same
initial mass but different $q_3$ are depicted. As
eq. \refnew{eq:ftime} predicts, $L \propto t^{(q_3-3)/(2-q_3)}$. If
$q_3$ is shallow (e.g. $q_3 \leq 4$), most of the initial mass is
deposited at the largest planetesimals. This mass reservoir is harder
to reach by collision and allows the disk to remain brighter at later
times. In comparison, disks with a steeper $q_3$ decay faster.
If one observes a collection of debris disk all at the same age,
intrinsic scatter in, e.g., initial masses, makes it impossible to
differentiate between models of different $q_3$. However, a collection
of disks with a large age spread can be used to constrain $q_3$. This
we proceed to demonstrate.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=.65,trim=0 0 0 0,clip]{figure3.eps}
\caption{Luminosity evolution for disks with different $q_3$ but the
same initial mass ($8 M_\oplus$). Systems with a steeper primordial
size spectrum (larger $q_3$) exhibit a more pronounced decline of
luminosity with time, since at a given time, a bigger fraction of
their mass reservoir has been depleted. Systems with shallower
$q_3$ (e.g., $q_3 = 3.5$), on the other hand, are initially dimmer
due to the relative shortage of smaller rocks, but eventually
outshine the higher $q_3$ disks as they can hold on to their mass
reservoir for longer. The luminosity decay of observed disks that
span a large range of ages can thus be used to infer the value of
$q_3$. All other parameters here are similar to those used in
Fig. \ref{fig:flHD1} and we adopt the hard strength law.}
\label{fig:varyq}
\end{center}
\end{figure}
\section{Observed Ensemble}
\label{sec:ensemble}
Several debris disks surveys have been carried out \citep[see,
e.g.][]{Suetal06,Trillingetal:2008,Lawler, Moor}. The sample of most
interest to us is that reported in \citet{Hillenbrand:2008}. Together
with updates in \citet{2009ApJS..181..197C}, \citet{Hillenbrand:2008}
presented a collection of debris disks around F/G/K type stars,
obtained as part of the Spitzer program on Formation and Evolution of
Planetary Systems (FEPS). This sample is unique in that both the
stellar age and the radial distance of the dust ring are determined:
isochrone fitting provides the age for the host stars (spanning from
$\sim 10^{7}$~years to a few $10^{9}$~years), while multi-band
photometry and spectral energy fitting yield the semi-major axis of
the dust ring. Together with fractional luminosity of the dust belt,
these provide the most important constraints to infer the primordial
properties of parent planetesimals.
To obtain the blow-out size ($s_{\rm min}$) for each system, we take
luminosity values for the central stars as given in
\citep{Hillenbrand:2008}, and we assign stellar masses by assuming
that $M_* \propto L_*^{1/3}$, as appropriate for solar type main-sequence
stars.
Out of the $31$ disks listed in \citet{Hillenbrand:2008}, we focus
only on a sub-sample of 13 disks that appear radially unextended and
are around main-sequence stars. In \citet{Hillenbrand:2008}, emission
from each disk is initially fitted with a single temperature blackbody
(a ring). If agreement between the ${24 \mu \rm{m}}/{33\mu \rm{m}}$
fit and the ${33 \mu \rm{m}}/{70 \mu \rm{m}}$ fit is poor,
they argue that the disk is likely radially extended and fit the data
instead with two radial components.
Since our numerical model is a one-zone model, we find that including
the extended sources into our analysis causes significant scatter in
the results. This leads us to discard them for the current
analysis. We have excluded HD 191089 from our sample. Its fluxes in
13 $\mu$m and 33 $\mu$m are not measured, and cannot be reliably
identified as an unextended source. In all, we are left with 13
sources.
It is interesting to note that most of the extended sources are
relatively young, all younger than a few hundred million years. In
contrast, the unextended sources have a larger age spread, lasting
till a few billion years (Fig. \ref{fig:compage}). All systems may be
born with more than one debris rings, but after a sufficiently long
time, only the outermost ring, which has the longest erosion
timescale, remains shining. The extended system are also brighter than
the average, likely related to their relative youth.
\begin{figure}[]
\centering
\includegraphics[scale=.65]{figure5.eps}
\caption{Cumulative distribution of stellar age (left panel) and
dust luminosity (right panel) for the \citet{Hillenbrand:2008}
sample. The extended systems (solid curve) tend to be younger and
brighter than the unextended systems (dashed curves). }
\label{fig:compage}
\end{figure}
\begin{figure*}
\centering
\includegraphics[scale=1.1]{figure6.eps}
\caption{Inferred disk initial masses, plotted against system ages,
for the unextended systems in \citet{Hillenbrand:2008}. The four
panels present four different choices of $q_3$. The other parameters
chosen are $e = 0.1$, $s_{\rm small} = 10^{4} \rm{cm}$,~and $\Delta
a/a = 0.1$. A value of $q_3 \in [3.5 ,4.0]$ is preferred: the upper
envelopes for the disk mass remain constant at all ages in the two
top plots. Models with higher $q_3$ are excluded as they require a
rising upper envelope. In addition, the $q_3 = 5$ model requires
unphysically large disk masses for very old disks. }
\label{fig:main}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale=1.1]{figure7.eps}
\caption{ Similar to Fig. \ref{fig:main} except $q_3$ is fixed at $4$
and a number of parameters are varied to test how the inferred
initial masses depend on them. Specific values for the inferred
mass may change, when the small end of the primordial size spectrum
($s_{\rm small}$, top-left), the eccentricity of particles
(bottom-left), the fractional width of the debris ring (top-right),
and the adopted strength law (bottom-right) are varied for all
systems. However, the important indicator for our study, the upper
envelope of the masses as a function of system age, remains flat.
So the conclusion that $q_3 \sim 4$ remains valid. }
\label{fig:9to13}
\end{figure*}
\section{The Primordial Size Spectrum Revealed}
\label{sec:results}
We have a simple strategy. Knowing the luminosity, the age and the
semi-major axis of each debris ring, we use our collisional model to
backtrack the evolution to infer its initial mass in the planetesimal
belt. These initial masses, when plotted against system ages, should
show a spread. One expects this spread to be constant across all ages,
as disks formed at different cosmic times likely have the same
distribution of disk properties. This property could be used to test
model assumptions. However, using the spread is difficult due to
selection effects. For instance, low mass disks may become too dim at
late times to be observable. So we propose instead to study the upper
envelope of this spread. The upper envelope should be flat with age
for the correct model. From our analytical scaling relations (see
Appendix), we find that the most important parameter in our model that
affects this mass slope is $q_3$, the power-law index in the
primordial size spectrum.
The results of such a procedure are shown in
Fig. \ref{fig:main}. Models with $q_3 = 4.5$ or greater appear to be
excluded by data, as they would require a rise of initial disk mass
with stellar ages. The reason behind this is transparent by studying
Fig. \ref{fig:varyq}. Models with $q_3 = 3.5$ and $4$ are compatible
with observations. Models with smaller $q_3$ lead to a
decreasing initial mass with system age and are excluded as
well.
Our model employs a number of other parameters, such as the
radial position and extent of the debris ring, the dynamical
excitation and break-up strength of the particles. We have studied
the robustness of our results when these parameters are varied
(Fig. \ref{fig:9to13}). As long as the values for these
parameters remain constant over age, varying them do not affect our
conclusion on $q_3$. The assumption that the dynamical excitation is
constant over age is suspicious, in light of results from
coagulation models showing that stirring by large plantesemals
increases gradually eccentricities of the disk particles. This is
discussed in \S \ref{sec:discussions}.
There is significant uncertainty in our conclusion due to the small
sample size. However, we argue that a larger sample may still not
favor models with, e.g., $q_3 = 5$. If $q_3 = 5$ (lower-right panel in
Fig. \ref{fig:varyq}), the system that remains easily detectable at 2
Gyrs of age requires an initial solid mass of $\sim 200 M_\oplus \sim
1 M_J $ in the planetesimal belt. The initial gas mass in such a belt
will be higher than the total disk mass of a typical T-Tauri star
($0.01 M_\odot$).
By focusing on dust luminosities, we are sensitive only to bodies that
lie below the break-size. As seen in Fig. \ref{fig:s2}, break-size
marches up to few tens to a hundred kilometers by the end of a few
billion years, if the disk has a mass of $M_0 = 12.4 M_\oplus$.
\section{Discussions}
\label{sec:discussions}
\subsection{Coagulation Models vs. Debris Disks}
\label{subsec:compare}
\begin{figure}[t]
\centering
\includegraphics[scale=0.65]{figure8.eps}
\caption{Inferred initial masses for a broken power-law
size-distribution. We investigate two particular forms, motivated by
coagulation simulations by \citet{2008ApJS..179..451K} and \citet{2011ApJ...728...68S}, respectively.
Other parameters adopted are $e = 0.1$, $\Delta a/a = 0.1$, and the
hard strength law. The inferred disk mass rises sharply with system
age. Moreover, to make old and bright systems, we require disk
masses that approach the mass of Jupiter.}
\label{fig:kenyon}
\end{figure}
In our exercise, we have assumed a simple initial size distribution
(eq. \ref{eq:time0}), with all bodies larger than a few hundred meters
described by a single power law index $q_3$. We relax this assumption
here.
Simulations of planetesimal coagulation produce typically more
complicated size distributions. For example, \citet{2008ApJS..179..451K}
started their simulations with all bodies at $\leq 1$ km. After tens
of millions of years of growth, most of the mass still remains at or
below $1$ km, with only $\sim 8\%$ of the mass being accreted into
bodies $10$ km or larger, $\sim 6\%$ into bodies $100$~kms or larger,
and $\sim 3\%$ into bodies of order $1000$~kms. We use a broken
power-law to replicate this kind of primordial spectrum. We set $q_3 =
5.5$~from $1$~km~to $10$~km, and $q_3 = 4$~from $10$~km~to $1000$~km. Motivated by \citet{2011ApJ...728...68S},
we also consider a slightly different initial distribution with $q_3 =
7$~from $1$~km~to $10$~km, and $q_3 = 4$~from $10$~km~to $1000$~km. Both sets of size spectrum deposit mass
mostly at the low end ($\leq 1$ km) and little at the large sizes. As
expected, when initial masses are determined for different systems
(Fig. \ref{fig:kenyon}), we find that young systems require
exceedingly low initial masses, while old systems require unphysically
large initial masses.
If we follow the luminosity evolution of such a disk, we will see that
the disk flares brightly in the first tens of millions of years, due
to the large mass reservoir at the $1$ km-range. Then the luminosity
decays as $t^{-1/2}$ (as expected of a $q_3 = 4$ spectrum) but with a
low normalization -- most of the disk mass has been ground down in the
early stage and we are now left with but a scrap remnant of the
original. Conglomeration simulations typically find that only a small
fraction of the mass can be accreted to make large bodies, before viscous stirring effectively stalls the growth.
\citet{2011ApJ...728...68S} showed that the fraction in large
bodies can only be of order $10^{-3}$ in Kuiper-belt-like
environments.
Does results in Fig. \ref{fig:kenyon} allow us to exclude current
conglomeration models? One possible caveat in our analysis is the
eccentricity. We discuss this below.
\subsection{Eccentricity}
\label{subsec:eccentricity}
We assume a static, high eccentricity ($e=0.1$) for all systems at all
times. In realistic systems, eccentricities can be a function of time.
One possible cause of eccentricity evolution is collisional cooling.
Collisions dissipate energy, so collisional products have in average
lower velocity dispersion than their parent bodies. In a single
collision, two bodies with masses $m_1$ and $m_1^\prime$ (assume $m_1
\gg m_1^\prime$) impact with typical velocities\footnote{ Velocities
here refer to the random component.}
\begin{equation}
v_1 \sim v_1^\prime \sim e_1 \, v_{\rm{kep}},
\end{equation}
where the subscript $1$ indicates that this is a first generation
collision in our counting. Assuming that all collision debris fly
away from the collision site with the velocity of the center-of-mass,
i.e., all relative velocities in the center-of-mass frame is
dissipated during the collision, collisional cooling can be expressed
as
\begin{equation}
v_2 \sim \frac{\sqrt{m_1^2v_1^2+m_1^{\prime 2}v_1^{\prime 2}}}{m_1+m_1^\prime}
\sim v_1\left(1 - \frac{m_1^\prime}{2m_1}\right).
\end{equation}
So the closer in mass the two colliding bodies are, the more cooling
their debris experiences. If cooling dominates the eccentricity
evolution, we find that a minimum eccentricity of $e_1 = 0.13$ is
required (for the hard strength law, and $e_1 = 0.02$ for the medium
law) to allow collisional cascade to proceed all the way to micron
range.
However, even if collisional cooling is severe, we argue that viscous
stirring by large planetesimals dominates the eccentricity evolution.
This is able to raise the eccentricity of collisional debris to values
comparable to that of their parents in a time shorter than a
collisional time. So the condition for a successful collisional
cascade is reduced to $e \geq 0.05$ for the hard strength law and $e
\geq 0.01$ for the medium strength law, i.e., the minimum random
motion necessary to break up the hardest grains (the smallest ones).
\footnote{This constraint can be reduced by a factor of unity when
radiation pressure on small grains are considered
\citep{Thebault09}.}
In fact, stirring is likely to gradually raise the eccentricity of all
bodies. Stirring by large bodies in the disk goes as
$e \propto t^{1/4}$
\citep[c.f.][]{GLS}.
In the simulations of \citet{2008ApJS..179..451K}, planetesimals are
continuously stirred by Pluto-like bodies, but they only reach $e\sim
0.1$ at about a Gyrs.\footnote{ An eccentricity of $e\sim 0.3$ at $40$
AU corresponds to the surface escape velocity of
Pluto. Planetesimals have to have a near-surface encounter before
they can reach such a high eccentricity. This takes time.}
Under such a scenario, the inferred initial disk mass is similar
to the original result (Fig. \ref{fig:newfig}),
but it is clear that we prefer the same range of values for $q_3$.
\begin{figure}[t]
\centering
\includegraphics[scale=0.65]{figure10.eps}
\caption{Same as the top-right panel in Fig. \ref{fig:main} but
instead of a constant eccentricity ($e=0.1$), here we assume that
the eccentricity rises as $e(t) \approx 0.1 (t/10^9
\rm{yrs})^{1/4}$. There is little difference to the inferred mass,
and if anything, the data seems to argue that a $q_3= 4$ model
slightly overestimate the value of $q_3$. So our conclusion that
$q_3 \in [3.5,4]$ remains unchanged even considering eccentricity
growth. }
\label{fig:newfig}
\end{figure}
\section{Summary}
\label{sec:summary}
Using an ensemble of bright debris disks around Sun-like stars, we
have measured the size spectrum of their embedded planetesimals. We
parametrize the size spectrum as $dn/ds \propto s^{-q_3}$ and find
$q_3 \approx 3.5-4$, where $q_3 = 4$ corresponds to equal mass per
logarithmic decade. The planetesimal sizes our technique probes lie
between a couple kms to $\sim 100 \, \rm km$.
While this size spectrum appears consistent with results of
coagulation simulations ($q_3 \sim 4$), there are two lines of
evidences that suggest problems in current coagulation models.
The first line of evidence is related to the inferre disk mass. The
inferred initial masses for these bright disks are surprisingly
high. We find total masses reaching as high as $10
M_\oplus$.\footnote{This is for $q_3 = 4$, and even higher values are
required if $q_3 = 3.5$.} This is comparable to the total solid mass
in the Kuiper belt region of Minimum Mass Solar Nebula model, and
about a factor of $100$ higher than the mass in large Kuiper belt
objects. Current coagulation models require an MMSN-like total mass
to produce the observed density of large Kuiper belt objects. If the
same inefficiency persists for our disks, one would require a total
disk mass of $\sim 100$ MMSN to produce those embedded planetesimals.
This is difficult to imagine.
The second line of evidence regards the size spectrum. We experiment
with size distributions that arise from coagulation simulations. We
find that these distributions could not reproduce the luminosity
distribution of the observed disks. Current coagulation models are
highly inefficient in making large planetesimals. So most of the mass
remains at where they started, presumably $\sim 1$ km. This leads to
debris disks that are too bright at early times and that are too dim
at late times, by a couple orders of magnitude.
We do not believe these discrepancies can be resolved by relaxing some
of our model assumptions. In particular, we argue that our estimate
for $q_3$ is unchanged even taking into account the fact that disk
eccentricity may rise with time. Our results are also insensitive to the width of the debris
ring, to the strength of bodies, and to the assumed upper and lower
sizes.
Because we restrict our attention to the upper envelope of inferred
masses, our result is dominated by a handful of systems. Our analysis
may be vulnerable to errors. However, the evidence is solid that
debris disks remain fairly bright even at a few billion years. This
alone dictates that there ought to be lots of mass stored in large
(10-100 kms) planetesimals. We address how this is accomplished by
revisiting coagulation model in an upcoming publication.
\begin{acknowledgements}
Y. Wu thanks Y. Lithwick, H. Schlichting and P. Sari for
discussions. We acknowledge financial support by NSERC.
\end{acknowledgements}
|
\section{\label{sec1}Introduction}
The presence of hard multiple parton interactions in high-energy hadron-hadron collisions has been convincingly de\-mo\-nstrated by the AFS~\cite{AFS}, UA2~\cite{UA2}, CDF~\cite{cdf4jets,cdf}, and D0~\cite{D0} Collaborations, using events with the four-jets and $\gamma+3$-jets final states, thus, providing new and complementary information on the proton structure.
In general, four high-$E_T$ jets (or $\gamma+3$-jets) may be produced either in the collision of one pair of partons in $2\to 4$ hard subprocess or via the simultaneous interaction of two parton pairs, that is in two $2\to 2$ subprocesses. In the last case in each dijet (or $\gamma+$ jet) system the transverse momenta of two high-$E_T$ jets more or less balance each other.
The possibility of observing two separate hard collisions was proposed long ago. Their theoretical investigation has a long history and goes back to the early days of the parton model~\cite{landshoff,takagi,goebel} with subsequent extension to perturbative QCD~\cite{paver,humpert,odorico,Mekhfi:1983az,Ametller:1985tp,Halzen:1986ue,sjostrand,Mangano:1988sq,Godbole:1989ti,Drees:1996rw,trelani,trelani2,sjostrand2,sjostrand3,strikman,Gaunt:2009re,Calucci:2010wg,Blok:2010ge,Strikman:2010bg,dremin,Ceccopieri:2010kg,del,DelFabbro:2002pw,Hussein:2006xr,Hussein:2007gj,Kulesza:1999zh,Cattaruzza:2005nu,maina,Berger:2009cm,Bahr:2008dy,corke,Gaunt:2010pi,Maina:2010vh,Diehl:2011tt,stir}.
Nevertheless, the phenomenology of multiple parton interactions relies on the models that are physically intuitive but involve significant simplifying assumptions. The\-re\-fore, it is extremely important to combine theoretical efforts in order to achieve a better description of multiple interactions, in particular, double scattering, which will be the dominant multiple scattering mode at the LHC. In this letter we consider some steps towards this purpose. The cross section formulae currently used to calculate double scattering processes are revised in terms of the modified collinear two-parton distributions extracted from deep inelastic scattering (DIS).
Let us recall, that with only the assumption of factorization of the two hard parton processes $A$ and $B$, the inclusive cross section of a double parton scattering process in a hadron collision is written in the following for
\begin{eqnarray}
\label{hardAB}
\sigma^D_{(A,B)} = \frac{m}{2} \sum \limits_{i,j,k,l} \int \Gamma_{ij}(x_1, x_2; {\bf b_1},{\bf b_2}; Q^2_1, Q^2_2)\nonumber\\
\times\hat{\sigma}^A_{ik}(x_1, x_1^{'},Q^2_1)
\hat{\sigma}^B_{jl}(x_2, x_2^{'},Q^2_2)\nonumber\\
\times\Gamma_{kl}(x_1^{'}, x_2^{'}; {\bf b_1} - {\bf b},{\bf b_2} - {\bf b}; Q^2_1, Q^2_2)\nonumber\\
\times dx_1 dx_2 dx_1^{'} dx_2^{'} d^2b_1 d^2b_2 d^2b,
\end{eqnarray}
where ${\bf b}$ is the usual impact parameter, that is the distance between centres of incoming (the beam and the tar\-get) hadrons in trans\-verse plane. $\Gamma_{ij}(x_1, x_2;{\bf b_1},{\bf b_2}; Q^2_1, Q^2_2)$ are the double parton distribution functions, depending on the longitudinal momentum fractions $x_1$ and $x_2$ and on the transverse position ${\bf b_1}$ and ${\bf b_2}$ of the two parton undergoing the hard processes $A$ and $B$ at the scales $Q_1$ and $Q_2$. $\hat{\sigma}^A_{ik}$ and $\hat{\sigma}^B_{jl}$ are the parton-level subprocess cross sections. The factor $m/2$ is a consequence of the symmetry of the expression for interchanging parton species $i$ and $j$. $m=1$ if $A=B$ and $m=2$ otherwise.
The double parton distribution functions $\Gamma_{ij}(x_1, x_2; {\bf b_1},{\bf b_2}; Q^2_1, Q^2_2)$ are the main reason of interest in multiple parton interactions. These distributions contain in fact all the information of probing the hadron in two different points simultaneously, through the hard processes $A$ and $B$.
It is typically taken that the double parton distribution functions may be decomposed in terms of longitudinal and transverse components as follows:
\begin{eqnarray}
\label{DxF}
\Gamma_{ij}(x_1, x_2;{\bf b_1},{\bf b_2}; Q^2_1, Q^2_2)\nonumber\\
= D^{ij}_h(x_1, x_2; Q^2_1, Q^2_2) f({\bf b_1}) f({\bf b_2}),
\end{eqnarray}
where $f({\bf b_1})$ is supposed to be an universal function for all kind of partons with the fixed normalization,
\begin{eqnarray}
\label{f}
\int f({\bf b_1}) f({\bf b_1 -b})d^2b_1 d^2b = \int T({\bf b})d^2b = 1,
\end{eqnarray}
and $T({\bf b}) = \int f({\bf b_1}) f({\bf b_1 -b})d^2b_1 $ is the overlap function.
If one makes the further assumptions that the longitudinal components $D^{ij}_h(x_1, x_2; Q^2_1, Q^2_2)$ reduce to the product of two independent one parton distributions,
\begin{eqnarray}
\label{DxD}
D^{ij}_h(x_1, x_2; Q^2_1, Q^2_2) = D^i_h (x_1; Q^2_1) D^j_h (x_2; Q^2_2),
\end{eqnarray}
the cross section of double parton scattering can be expressed in the simple form
\begin{eqnarray}
\label{doubleAB}
& \sigma^D_{(A,B)} = \frac{m}{2} \frac{\sigma^S_{(A)} \sigma^S_{(B)}}{\sigma_{\rm eff}}, \\
& \sigma_{\rm eff}=[ \int d^2b (T({\bf b}))^2]^{-1}.
\end{eqnarray}
In this representation and at the factorization of longitudinal and transverse components, the inclusive cross section of single hard scattering is written as
\begin{eqnarray}
\label{hardS}
\sigma^S_{(A)} = \sum \limits_{i,k} \int \Gamma_{i}(x_1; {\bf b_1}; Q^2_1)
\hat{\sigma}^A_{ik}(x_1, x_1^{'}) \nonumber\\
\times\Gamma_{k}(x_1^{'}; {\bf b_1} - {\bf b}; Q^2_1) dx_1 dx_1^{'} d^2b_1 d^2b \nonumber\\
=\sum \limits_{i,k} \int D^{i}_h(x_1; Q^2_1) f({\bf b_1})
\hat{\sigma}^A_{ik}(x_1, x_1^{'})\nonumber\\
\times D^{k}_{h'}(x_1^{'}; Q^2_1)f( {\bf b_1} - {\bf b}) dx_1 dx_1^{'} d^2b_1 d^2b \nonumber\\
=\sum \limits_{i,k} \int D^{i}_h(x_1; Q^2_1)
\hat{\sigma}^A_{ik}(x_1, x_1^{'}) D^{k}_{h'}(x_1^{'}; Q^2_1) dx_1 dx_1^{'}.
\end{eqnarray}
These simplifying assumptions, even though rather common in the literature and quite convenient from a computational point of view, are not enough justified and should be revised~\cite{Blok:2010ge,Diehl:2011tt,stir}, while the starting cross section formula~(\ref{hardAB}) was found (derived) in the momentum representation using light-cone variables in a number of works (see, e.g., Refs.~\cite{paver,Mekhfi:1983az,Blok:2010ge,Diehl:2011tt,stir}) at the same approximations as in processes with a single hard scattering.
\maketitle
\section{\label{sec2} Momentum representation}
All the previous formulae were written in the mixed (momentum and coordinate) representation. Recall that in general for the case of the multiple parton interactions we have to use the
{\it Generalized} Parton Distribution Functions (GPDF). In other words in the Feynman diagram (ladder) which describes the GPDF the parton momenta $k_L$ (in the left part of diagram corresponding to the amplitude $A^*$) and $k_R$ (in the right part of the diagram corresonding to amplitude $A$) may be different. Let us denote $k_L=k+q/2$ and $k_R=k-q/2$ where $q$ is the momentum transfered through the whole ladder. Since the ladders in Figs.~1 and 2 form a loop we will call $q$ --- the loop momentum. In the previous formulae instead of transverse momentum $q_t$ we had used the conjugated coordinate ${\bf b}$~\footnote{In the case of conventional DIS the cross section is given by the integral over ${\bf b}$ corresponding to $q_t=0$.}.
For our further goal the momentum representation is more convenient:
\begin{eqnarray}
\label{hardAB_p}
\sigma^D_{(A,B)} = \frac{m}{2} \sum \limits_{i,j,k,l} \int \Gamma_{ij}(x_1, x_2; {\bf q}; Q^2_1, Q^2_2)\nonumber\\
\times \hat{\sigma}^A_{ik}(x_1, x_1^{'})
\hat{\sigma}^B_{jl}(x_2, x_2^{'})\nonumber\\
\times\Gamma_{kl}(x_1^{'}, x_2^{'}; {\bf -q}; Q^2_1, Q^2_2) dx_1 dx_2 dx_1^{'} dx_2^{'} \frac{d^2q}{(2\pi)^2}.
\end{eqnarray}
The hard subprocesses {\bf A} and {\bf B} are originated by two different branches of the parton cascade. Note that only the sum of the parton momenta (in both branches) is conserved, while in each individual branch there may be some difference, $q$,
of transverse (parton) momenta in the initial state wave function and the conjugated wave function.
\begin{figure}
\includegraphics[width=0.45\textwidth]{fig1.eps}
\caption{ A graph for double parton scattering due to the first term. $A$ and $B$ are the hard parton subprocesses. $q$ is the momemtum transfered through the ladders $L1, L2, L1'$ and $L2'$.
\label{fig:fig1}}
\end{figure}
The main problem is in the correct calculation of these two-parton functions $\Gamma_{ij}(x_1, x_2; {\bf q}; Q^2_1, Q^2_2)$ without the simplifying assumptions~(\ref{DxF}) and (\ref{DxD}). The case is that they are known in the current literature~\cite{Kirschner:1979im,Shelest:1982dg,snig03,snig04,snig08} only at ${\bf q}=0$ in the collinear approximation. In this approximation the two-parton distribution functions, $\Gamma_{ij}(x_1, x_2; {\bf q}=0; Q^2, Q^2)=D^{ij}_h(x_1, x_2; Q^2, Q^2)$ with the two hard scales set equal, satisfy the generalized Dokshitzer-Gribov-Lipatov-Altarelli-Parisi (DGLAP) evolution equations, derived for the first time in Refs.~\cite{Kirschner:1979im,Shelest:1982dg}, as well as single distributions satisfy more known and cited DGLAP equations~\cite{gribov,lipatov,dokshitzer,altarelli}. The functions in question have a specific interpretation in the leading logarithm approximation of perturbative QCD: they are inclusive probabilities that in a hadron $h$ one finds two bare partons of types $i$ and $j$ with the given longitudinal momentum fractions $x_1$ and $x_2$.
\begin{figure}
\includegraphics[width=0.45\textwidth]{fig2.eps}
\caption{ A graph for double parton scattering due to the second evolution term. $A$ and $B$ are the hard parton subprocesses. $q$ is the momemtum transfered through the ladders $L1, L2, L1'$ and $L2'$
\label{fig:fig2}}
\end{figure}
The evolution equation for $\Gamma_{ij}$ contains two terms. The first term describes independent (simultaneous) evolution of two branches of parton cascade: one branch contains the parton $x_1$ and another branch --- the parton $x_2$. The second term accounts for the possibility to split the one parton evolution (one branch $k$) into the two different branches, $i$ and $j$. It contains the usual splitting function $P_{k\to ij}(z)$. The solutions of the generalized DGLAP evolution equations with the given initial conditions at the reference scales $\mu^2$ may be written~\cite{Gaunt:2009re,Ceccopieri:2010kg,snig11}
in the form:
\begin{eqnarray}
\label{solutiontwoscale}
D_h^{j_1j_2}(x_1,x_2;\mu^2, Q_1^2, Q_2^2)\nonumber\\
= D_{h1}^{j_1j_2}(x_1,x_2;\mu^2, Q_1^2, Q_2^2) +
D_{h2}^{j_1j_2}(x_1,x_2;\mu^2, Q_1^2, Q_2^2),\nonumber\\
D_{h1}^{j_1j_2}(x_1,x_2;\mu^2, Q_1^2, Q_2^2)\nonumber\\
= \sum\limits_{j_1{'}j_2{'}}
\int\limits_{x_1}^{1-x_2}\frac{dz_1}{z_1}
\int\limits_{x_2}^{1-z_1}\frac{dz_2}{z_2}~
D_h^{j_1{'}j_2{'}}(z_1,z_2;\mu^2)\nonumber\\
\times D_{j_1{'}}^{j_1}(\frac{x_1}{z_1};\mu^2,Q_1^2)
D_{j_2{'}}^{j_2}(\frac{x_2}{z_2},\mu^2,Q_2^2) ,\nonumber\\
D_{h2}^{j_1j_2}(x_1,x_2;\mu^2, Q_1^2, Q_2^2)\nonumber\\
=\sum\limits_{j{'}j_1{'}j_2{'}} \int\limits_{\mu^2}^{\min(Q_1^2,Q_2^2)}dk^2\frac{\alpha_s(k^2)}{2\pi k^2}
\int\limits_{x_1}^{1-x_2}\frac{dz_1}{z_1}
\int\limits_{x_2}^{1-z_1}\frac{dz_2}{z_2}\nonumber\\
\times D_h^{j{'}}(z_1+z_2;\mu^2,k^2)
\frac{1}{z_1+z_2}P_{j{'} \to
j_1{'}j_2{'}}\Bigg(\frac{z_1}{z_1+z_2}\Bigg)\nonumber\\
\times D_{j_1{'}}^{j_1}(\frac{x_1}{z_1};k^2,Q_1^2)
D_{j_2{'}}^{j_2}(\frac{x_2}{z_2};k^2,Q_2^2),\nonumber\\
\end{eqnarray}
where
$\alpha_s(k^2)$ is the QCD coupling,
$D_{j_1{'}}^{j_1}(z;k^2,Q^2)$ are the known single distribution functions (the Green's functions) at the parton level with the specific $\delta$-like initial conditions at $Q^2=k^2$. $D_h^{j'_1,j'_2}(z_1,z_2,\mu^2)$ is the initial (input) two-parton distribution at a relatively low scale $\mu$. The one parton distribution (before the slitting into the two branches at some scale $k^2$) is given by $D_h^{j'}(z_1+z_2,\mu^2,k^2)$.
Note that in Eq.~(\ref{solutiontwoscale}) we assume that the loop momentum $q<\mu$ is small and due to a strong ordering of parton transverse momenta during the collinear DGLAP evolution it may be neglected.\\
The first term is the solution of homogeneous evolution equation (independent evolution of two branches) where in general the input two-parton distribution at low scale $\mu$ is not known. For this non-perturbative two-parton function at low $z_1,z_2$ one may assume the factorization $D_h^{j_1{'}j_2{'}}(z_1,z_2,\mu^2) \simeq D_h^{j_1{'}}(z_1,\mu^2)D_h^{j_2{'}}(z_2,\mu^2)$
neglecting the influence of momentum conservation ($z_1+z_2<1$).
This leads to
\begin{eqnarray}
\label{DxD_Q}
D^{ij}_{h1}(x_1, x_2; \mu^2, Q^2_1, Q^2_2)\nonumber\\
\simeq D^i_h (x_1; \mu^2, Q^2_1) D^j_h (x_2;\mu^2, Q^2_2).
\end{eqnarray}
\noindent
Since the multiple interactions take place at relatively low transverse momenta, the contribution to the cross section from double scattering can be separated experimentally in the region of relatively low longitudinal momentum fraction $x_1$ and $x_2$, where the factorization hypothesis~(\ref{DxD_Q}) for the first term is a good approximation. In this case the cross section for double parton scattering can be estimated, using the two-gluon form factor of the nucleon $F_{2g}(q)$~\cite{Blok:2010ge,Frankfurt:2002ka} for the dominant gluon-gluon scattering mode (or something similar for other parton scattering modes),
\begin{eqnarray}
\label{hardAB1}
\sigma^{D,1\times1}_{(A,B)} = \frac{m}{2} \sum \limits_{i,j,k,l} \int D^i_h (x_1; \mu^2, Q^2_1) D^j_h (x_2;\mu^2, Q^2_2) \nonumber\\
\times \hat{\sigma}^A_{ik}(x_1, x_1^{'})
\hat{\sigma}^B_{jl}(x_2, x_2^{'}) D^k_{h'} (x'_1; \mu^2, Q^2_1)
D^l_{h'} (x'_2;\mu^2, Q^2_2)\nonumber\\
\times dx_1 dx_2 dx_1^{'} dx_2^{'} \int F_{2g}^4(q)\frac{d^2q}{(2\pi)^2}.\nonumber\\
\end{eqnarray}
From the dipole fit $F_{2g}(q)=1/(q^2/m^2_g+1)^2$ to the two-gluon form factor follows that the characteristic value of $q$ is of the order of gluon mass $m_g$, and therefore the initial conditions for the single distributions can be fixed at some reference scale $\mu \sim m_g$ because of the weak logarithmic dependence of these distributions on the scale value. In this approach $\int F_{2g}^4(q)\frac{d^2q}{(2\pi)^2}$ gives the estimation of $[\sigma_{\rm eff}]^{-1}$.
The second term in Eq.~(\ref{solutiontwoscale}) is the solution of complete evolution equation with the evolution from one ``nonperturbative'' parton at the reference scale. Here the independent evolution of two branches starts at scale $k^2$ from a point-like parton $j'$. Therefore we have no the form factor $F_{2g}(q)$ which suppresses the large $q_t$ domain. The corresponding contribution to the cross section reads
\begin{eqnarray}
\label{D2xD2}
\sigma^{D,2\times2}_{(A,B)} \nonumber\\
= \frac{m}{2} \sum \limits_{i,j,k,l} \int dx_1 dx_2 dx_1^{'} dx_2^{'} \int^{\min(Q_1^2,Q_2^2)} \frac{d^2q}{(2\pi)^2}\nonumber\\
\times \sum\limits_{j{'}j_1{'}j_2{'}} \int\limits_{q^2}^{\min(Q_1^2,Q_2^2)}dk^2 \frac{\alpha_s(k^2)}{2\pi k^2}
\int\limits_{x_1}^{1-x_2}\frac{dz_1}{z_1}
\int\limits_{x_2}^{1-z_1}\frac{dz_2}{z_2}\nonumber\\
\times D_h^{j{'}}(z_1+z_2;\mu^2,k^2)
\frac{1}{z_1+z_2}P_{j{'} \to
j_1{'}j_2{'}}\Bigg(\frac{z_1}{z_1+z_2}\Bigg)\nonumber\\
\times D_{j_1{'}}^{i}(\frac{x_1}{z_1};k^2,Q_1^2)
D_{j_2{'}}^{j}(\frac{x_2}{z_2};k^2,Q_2^2)
\nonumber\\
\times\hat{\sigma}^A_{ik}(x_1, x_1^{'}) \hat{\sigma}^B_{jl}(x_2, x_2^{'})\nonumber\\
\times\sum\limits_{j{'}j_1{'}j_2{'}} \int\limits_{q^2}^{\min(Q_1^2,Q_2^2)}dk^{'2} \frac{\alpha_s(k^{'2})}{2\pi k^{'2}}
\int\limits_{x'_1}^{1-x'_2}\frac{dz_1}{z_1}
\int\limits_{x'_2}^{1-z_1}\frac{dz_2}{z_2}\nonumber\\
\times D_{h'}^{j{'}}(z_1+z_2;\mu^2,k^{'2})
\frac{1}{z_1+z_2}P_{j{'} \to
j_1{'}j_2{'}}\Bigg(\frac{z_1}{z_1+z_2}\Bigg)\nonumber\\
\times D_{j_1{'}}^{k}(\frac{x'_1}{z_1};k^{'2},Q_1^2)
D_{j_2{'}}^{l}(\frac{x'_2}{z_2};k^{'2},Q_2^2),
\end{eqnarray}
or in less transparent, but substantially shorter form:
\begin{eqnarray}
\label{D2xD2_s}
\sigma^{D,2\times2}_{(A,B)} \nonumber\\
= \frac{m}{2} \sum \limits_{i,j,k,l} \int dx_1 dx_2 dx_1^{'} dx_2^{'} \int^{\min(Q_1^2,Q_2^2)} \frac{d^2q}{(2\pi)^2}\nonumber\\
\times D_{h2}^{ij}(x_1,x_2;q^2, Q_1^2, Q_2^2)
\hat{\sigma}^A_{ik}(x_1, x_1^{'}) \nonumber\\
\times \hat{\sigma}^B_{jl}(x_2, x_2^{'})
D_{h'2}^{kl}(x'_1,x'_2;q^2, Q_1^2, Q_2^2).
\end{eqnarray}
There is yet the combined (``interference'') contribution, which is written by analogy,
\begin{eqnarray}
\label{D1xD2_s}
\sigma^{D,1\times2}_{(A,B)}\nonumber\\
= \frac{m}{2} \sum \limits_{i,j,k,l} \int dx_1 dx_2 dx_1^{'} dx_2^{'} \int^{\min(Q_1^2,Q_2^2)} F_{2g}^2(q)\frac{d^2q}{(2\pi)^2}\nonumber\\
\times [D^i_h (x_1; \mu^2, Q^2_1) D^j_h (x_2;\mu^2, Q^2_2)
\hat{\sigma}^A_{ik}(x_1, x_1^{'}) \nonumber\\
\times \hat{\sigma}^B_{jl}(x_2, x_2^{'})
D_{h'2}^{kl}(x'_1,x'_2;q^2, Q_1^2, Q_2^2)\nonumber\\
+D_{h2}^{ij}(x_1,x_2;q^2, Q_1^2, Q_2^2)
\hat{\sigma}^A_{ik}(x_1, x_1^{'}) \nonumber\\
\times \hat{\sigma}^B_{jl}(x_2, x_2^{'})
D^k_{h'} (x'_1; \mu^2, Q^2_1) D^l_{h'} (x'_2;\mu^2, Q^2_2)].\nonumber\\
\end{eqnarray}
The equations~(\ref{hardAB1}), (\ref{D2xD2_s}) and (\ref{D1xD2_s}) are our solution of the problem --- the estimation of the inclusive cross section for double parton scattering, taking into account the QCD evolution, in terms of the well-known collinear distributions extracted from deep inelastic scattering.
However, here one should note that
the input two-parton distribution $D_h^{j'_1,j'_2}(z_1,z_2,\mu^2)$ may be more complicated than that given by factorization ansatz (\ref{DxD_Q}). Let us discuss in more detail the second term, that is the $2\times 2$ contribution.
\maketitle
\section{\label{sec3} Discussion and conclusions}
The contribution to the cross section from the second term induced by the QCD evolution does not reduce to the simple contribution with a some new, other, constant effective cross section $\sigma_{\rm eff}$, as it was done in first estimations~\cite{Cattaruzza:2005nu,Gaunt:2010pi,Maina:2010vh}. The QCD evolution effects for the cross section are anticipated to be larger than for the two-parton distribution functions, for which they were estimated in Refs.~\cite{Gaunt:2009re,snig04} on the level of 10$\%$ -30$\%$ in comparison with the ``factorization'' components at $x \sim 0.1$ and $Q \sim 100$~GeV. Indeed, in Eq.~(\ref{D2xD2}) the integration over $q$ has no a strong suppression factor $F_{2g}(q)$ and the phase space integral may be estimated as,
\begin{eqnarray}
\label{pspace}
\int^{Q^2} dq^2 \int_{q^2}^{Q^2} \frac{dk^2}{k^2}\int_{q^2}^{Q^2} \frac{dk^{'2}}{k^{'2}} \simeq 2Q^2 ,~~~Q^2 \gg \mu^2,
\end{eqnarray}
where within the leading order (LO) accuracy we take $q^2$ as the lower limit in $k^2$ and $k^{'2}$ integrations; at $q^2>k^2$ the loop momentum $q_t$ destroys the logarithmic structure of the integrals in collinear evolution from $k^2$ to $Q^2$.
We see that for a large final scale $Q^2$ the second ($2\times 2$) contribution dominates being proportional to $Q^2$ in comparison with the $1\times 1$ or $1\times 2$ components which contributions $\sim m^2_g\sim 1/\sigma_{eff}$ are limited by the nucleon (hadron) form factor $F_{2g}$~\footnote{In terms of impact parameters ${\bf b}$ this means that in the second ($2\times 2$) term two pairs of partons are very close to each other; $|{\bf b}_1-{\bf b}_2|\sim 1/Q$.}.
The real gain is, of course, less, since the running coupling constant and the distribution functions are logarithmically dependent on the integration variables and we have the additional suppression factor, which is inversely proportional~\cite{Snigirev:2010ds} to the initial gluon and quark multiplicities squared; due to the different ``normalization'': the second term evolves from one ``nonperturbative'' parton unlike the first factorization term having the two initial independent ``nonperturbative'' partons at the reference scale.
As a result, the experimental effective cross section, $\sigma_{\rm eff}^{\rm exp}$, which is
not measured directly but is extracted, using the normalization to the product of two single cross sections:
\begin{eqnarray}
\label{dps}
\frac{\sigma_{DPS}^{\gamma+3j}}{\sigma^{\gamma j}\sigma^{jj}}= [\sigma_{\rm eff}^{\rm exp}]^{-1}
\end{eqnarray}
in both the CDF and D0 experiments, should be dependent on the probing hard scale. It should decrease with the growth of the resolution scale due to the fact that all additional contributions to the cross section of double parton scattering are positive and increase. Here $\sigma^{\gamma j}$ and $\sigma^{j j}$ are the inclusive $\gamma +$ jet and dijets cross sections, $\sigma_{DPS}^{\gamma+3j}$ is the inclusive cross section of the $\gamma + 3$ jets events produced in the double parton process. It is worth noticing that the CDF and D0 Collaborations extract $\sigma_{\rm eff}^{\rm exp}$ without any theoretical predictions on the $\gamma +$ jet and dijets cross sections, by comparing the number of observed double parton $\gamma + 3$ jets events in one $p{\bar p}$ collision to the number of $\gamma + 3$ jets events with hard interactions occurring in the case of two separate $p{\bar p}$ collisions.
The recent D0 measurement~\cite{D0} of this effective cross section, $\sigma_{\rm eff}^{\rm exp}$, done as a function of the second (ordered in the transverse momentum $p_T$) jet $p_T$, $p^{\rm jet2}_T$, that can serve as a resolution scale, shows a tendency to be dependent on this scale. In Ref.~\cite{Snigirev:2010tk}
this fact was interpreted as the first indication to the QCD evolution of double parton distributions.\\
We have to emphasize that the dominant contribution to the phase space integral~(\ref{pspace}) comes from a large $q^2\sim Q^2$ and strictly speaking we have no place for the collinear (DGLAP) evolution of two independent branches of the parton cascade (i.e., in the ladders $L1, L2, L1'$ and $L2'$) in the $2\times 2$ term. Formally in the framework of collinear approach this contribution should be considered as that caused by the interaction of {\it one} pair of partons with the $2\to 4$ hard subprocess~\footnote{This is in agreement with the statement~\cite{stir} that ``the structure of Fig.~2 should not be included in the leading logarithmic {\it double} parton scattering cross section''.}. Recall however that in the estimate (\ref{pspace}) we neglect the anomalous dimension, $\gamma$, of the parton distributions $D^k_j(x/z,k^2,Q^2)\propto (Q^2/k^2)^\gamma$. In collinear appoach the anomalous dimensions
$\gamma\propto \alpha_s<<1$ are assumed to be small. On the other hand in a low $x$ region the value of anomalous dimension is enhanced by the $\ln(1/x)$ logarithm and may be rather large numerically. So the integral over $q^2$ is {\it slowely} convergent and the major contribution to the cross section is expected to come actually from some characteristic intermediate region, $m^2_g<<q^2<< Q^2_1 (Q_1<Q_2)$, with the not such strong sensitivity to the upper limit of $q$-integration as in the case of the phase space integral. Therefore it makes sense to consider the numerical contribution of the $2\times 2$ term even within the collinear approach at the LHC kinematics with the available large enough $Q_1$ and $Q_2$ with respect to $m_g$ (providing the wide enough integration region and the characteristic intermediate momenta $q$).
Next in a configuration with two quite different scales (say, $Q^2_1<<Q^2_2$) the upper limit of $q^2$ integral is given by a smaller scale (at $q>Q_1$ the hard matrix element corresponding to $\sigma^A$ starts to fall down with $q_t$). In this case the collinear evolution from the scale $q=Q_1$ up to the scale $Q_2$ in the ladders (parton branches) $L2$ and $L2'$ is well justified.\\
Moreover the integrals over $z_1$, $z_2$ in Eq.~(\ref{D2xD2}) are concentrated at a rather large $z$ giving enough space for the Balitsky-Fadin-Kuraev-Lipatov (BFKL) evolution~\cite{bfkl,bfkl2,bal,ryskin}. Thus it may be intersting (and even more justified theoretically) to consider the multiple parton interactions in the framework of the BFKL approach (see the recent paper~\cite{flesburg}).\\
In summary, we suggest a practical constructive me\-thod how to estimate the inclusive cross section for double parton scattering, taking into account the QCD evolution, in terms of the well-known collinear distributions extracted from deep inelastic scattering. We support also the conclusion of Refs.~\cite{Snigirev:2010tk,flesburg} that the experimentally measured effective cross section, $\sigma_{\rm eff}^{\rm exp}$, at a normalization~(\ref{dps}) and the presence of the evolution (correlation) term in the two-parton distributions, should decrease with the growth of the resolution scale $Q^2$.
\begin{acknowledgments}
Discussions with D.V.~Bandurin, E.E.~Boos, S.V.~Mo\-lo\-dtsov, M.~Strikman, D.~Treleani, G.M.~Zinovjev and N.P.~Zotov are gratefully acknowledged. This work is partly supported by Russian Foundation for Basic Research Grants No. 10-02-93118, the President of Russian Federation for support of Leading Scientific Schools Grant No 4142.2010.2. and by the Federal Program of the Russian State
RSGSS=65751.2010.2.
\end{acknowledgments}
|
\section{Introduction}\label{sec:intro}
\IEEEPARstart{P}{ublic-key} cryptosystems based on coding theory, known for nearly as long as the very concept of asymmetric cryptography itself, have recently been attracting renewed interest because of their apparent resistance even against attacks mounted with the help of quantum computers, constituting a family of so-called post-quantum cryptosystems~\cite{bernstein-buchmann-dahmen}. However, not all error-correcting codes are suitable for cryptographic applications. The most commonly used family of codes for such purposed is that of Goppa codes, which remain essentially unharmed by cryptanalysis efforts despite considerable efforts and progress in the area.
Introduced in 1970, Goppa codes~\cite{goppa} are a subfamily of alternant codes, i.e. subfield subcodes of Generalized Reed-Solomon codes.
Let $q = p^m$ for some prime $p$ and some $m > 0$. A Goppa code $\Gamma(L, g)$ over $\mathbb{F}_p$ is determined by a sequence $L \in \mathbb{F}_q^n$ of distinct values, and a polynomial $g \in \mathbb{F}_q[x]$ of degree $t := \deg(g)$ whose roots are disjoint from $L$. Goppa codes have by design a minimal distance at least $t + 1$ by virtue of being alternant. Certain codes are known to have better minimum distances than this lower bound. Thus, binary Goppa codes where $g$ is square-free are known to have a larger minimum distance of at least $2t + 1$ instead. A family of codes where $g$ is not square-free have minimum distance at least $t + \gamma - 1$ for some $2 < \gamma < t - 1$, which is known as the Hartmann-Tzeng bound for Goppa codes~\cite{tzeng-hartmann,hartmann-tzeng}.
The class of Sugiyama-Kasahara-Hirasawa-Namekawa codes~\cite{sugiyama-kasahara-hirasawa-namekawa} where $g = h^{r-1}$ for some square-free monic polynomial $h \in \mathbb{F}_q[x]$ and some power $r$ of $p$ dividing $q$, which constitute a proper superclass of the so-called ``wild'' codes where $h$ is restricted to being irreducible~\cite{bernstein-lange-peters:wild}, have minimum distance at least $r\deg(h) + 1$ rather than $(r - 1)\deg(h) + 1$.
Although it is known that the minimum distance of a Goppa code of degree $t$ is at least $t + 1$ and there are known cases where it is higher (up to $2t + 1$, as it happens for binary square-free Goppa codes), systematically determining the true minimum distance of any given subfamily of Goppa codes remains largely an open problem, yet it is an important metric as it determines not only how many errors can always be uniquely corrected, but indirectly the security level and the key sizes of the cryptosystems based on each given code.
Apart from brute force, known decoding methods for alternant codes can in general correct only about half as many errors as a binary square-free Goppa code is in principle able to correct~\cite{berlekamp,sugiyama-kasahara-hirasawa-namekawa:decoding} (see also~\cite{macwilliams-sloane}). Even the Guruswami-Sudan algorithm~\cite{guruswami-sudan}, which exceeds the $t/2$ limit, can only correct about $n - \sqrt{n(n - t)} \approx t/2 + (t/2)^2/(2n - t)$ errors. In contrast, Patterson's algorithm can correct all $t$ design errors of binary Goppa codes, as can an alternant decoder using the equivalence $\Gamma(L, g) = \Gamma(L, g^2)$ albeit at a larger computational cost. Bernstein's list decoding method~\cite{bernstein} goes somewhat further, attaining a correction capability of $n - \sqrt{n(n - 2t - 2)} \approx t + 1 + (t + 1)^2/2(n - t - 1)$ errors for binary irreducible Goppa codes, although decoding is ambiguous if the actual distance is not proportionally higher. Similar techniques can in principle correct about $n - \sqrt{n(n - rt)} \approx rt/2 + (rt/2)^2/(2n - rt)$ errors for wild codes~\cite{bernstein-lange-peters:wild}. Bernstein's method does not reach the $q$-ary Johnson radius, but a more recent algorithm by Augot \emph{et al.} does so in the binary case \cite{augot-barbier-couvreur}.
\subsection{Our Results}\label{sec:our-contrib}
Our contribution in this paper is a non-deterministic decoding algorithm for square-free Goppa codes over $\mathbb{F}_p$ for any prime $p$. The method generalizes Patterson's approach and can potentially correct up to $(2/p)t$ errors, on the condition that a suitable short vector can be found in a certain polynomial lattice. In particular, our method corrects $(2/3)t$ errors in characteristic~3, exceeding the $t/2$ barrier when the average distance to the closest codeword is at least $(4/3)t + 1$.
In experiments conducted to assess the practical behaviour of our proposal, the result of the decoding is observed to be unique with overwhelming probability for irreducible ternary Goppa codes chosen uniformly at random,
hinting that, for the vast majority of such codes, the average distance to the closest codeword is sufficiently higher than the ensured minimum distance.
Besides, our proposal can probabilistically correct a still larger number of errors that approaches and reaches $t$ depending on the distribution of error magnitudes. For instance, the method corrects up to $t$ errors with high probability if all error magnitudes are known to be equal.
This feature outperforms even Sugiyama-Kasahara-Hirasawa-Namekawa and wild codes and the associated decoding methods, and is particularly interesting for cryptographic applications like McEliece encryption~\cite{mceliece} under the Fujisaki-Okamoto or similar semantic security transform~\cite{fujisaki-okamoto}, where error magnitudes can be chosen by convention to be all equal. In that case, even if an attacker could somehow derive a generic alternant decoder from the public code that is typical in such systems (a strategy exploited e.g. in~\cite{faugere-otmani-perret-tillich}), he will not be able to correct more than about $t/2$ errors out of roughly $t$ that can be corrected with the private trapdoor enabled by our proposal, facing an infeasible workload of about $(p-1)\binom{n}{t/2}/\binom{t}{t/2}$ guesses to mount a complete attack. This makes Goppa codes in odd characteristic, which have already been shown to sport some potential security advantages over binary ones~\cite{peters}, even more attractive in practice.
For the benefit of implementors, we describe a dedicated version of the Mulders-Storjohann algorithm to convert the particular lattice basis encountered during the decoding process to weak Popov form.
The computational complexity of this step is then shown to be $O(p^3 t^2)$.
\subsection{Organization of the Paper}
The remainder of this document is organized as follows. We provide basic notions in Section~\ref{sec:prelim}. We recapitulate Patterson's decoding algorithm for binary irreducible Goppa codes in Section~\ref{sec:patterson}, and extend it to square-free codes in characteristic~$p$ in Section~\ref{sec:extension}, showing that it can correct $(2/p)t$ errors in general and up to $t$ errors depending on the distribution of error magnitudes. We conclude in Section~\ref{sec:conclusion}.
\section{Preliminaries}\label{sec:prelim}
Matrix indices will start from 0 throughout this paper, unless otherwise stated. Let $p$ be a prime and let $q = p^m$ for some $m > 0$. The finite field of $q$ elements is written $\mathbb{F}_q$. For sequences of elements $(g_1, \dots, g_t) \in \mathbb{F}_q^t$, $(L_0, \dots, L_{n-1}) \in \mathbb{F}_q^n$ and $(d_0, \dots, d_{n-1}) \in \mathbb{F}_q^n$ for some $t, n \in \mathbb{N}$, we denote by $\toep(g_1, \dots, g_t)$ the $t \times t$ Toeplitz matrix with elements $T_{ij} := g_{t - i + j}$ for $j \leqslant i$ and $T_{ij} := 0$ otherwise; by $\vdm_t(L_0, \dots, L_{n-1})$ the $t \times n$ Vandermonde matrix with elements $V_{ij} := L_j^i$, $0 \leqslant i < t$, $0 \leqslant j < n$; and by $\diag(d_0, \dots, d_{n-1})$ the diagonal matrix with diagonal elements $D_{jj} := d_j$, $0 \leqslant j < n$.
\subsection{Error Correcting Codes}
Let $L = (L_0, \dots, L_{n-1}) \in \mathbb{F}_q^n$ be a sequence (called the \emph{support}) of $n \leqslant q$ distinct elements, and let $g \in \mathbb{F}_q[x]$ be an irreducible monic polynomial of degree $t$ such that $g(L_i) \neq 0$ for all $i$. For any word $e \in \mathbb{F}_p^n$ we define the corresponding \emph{Goppa syndrome} polynomial $s_e \in \mathbb{F}_q[x]$ to be:
\[
s_e(x) = \sum_{i=0}^{n-1}{\dfrac{e_i}{x - L_i} \mod{g(x)}}.
\]
Thus the syndrome is a linear function of $e$.
The $[n, \geqslant n - mt, \geqslant t+1]$ \emph{Goppa code} over $\mathbb{F}_p$ with support $L$ and generator polynomial $g$ is the kernel of the syndrome function applied to elements from $\mathbb{F}_p$, i.e. the set $\Gamma(L, g) := \{e \in \mathbb{F}_p^n \mid s_e \equiv 0 \mod{g}\}$.
Writing $s_e(x) := \sum_i{s_i x^i}$ for some $s \in \mathbb{F}_q^{t}$, one can show that $s^{\textrm{\tiny T}} = H e^{\textrm{\tiny T}}$ where the \emph{parity-check matrix} $H$ has the form
\begin{equation}\label{eq:parity-check}
\begin{array}{rcl}
H &=& \toep(g_1, \dots, g_t)\\
&\cdot& \vdm_t(L_0, \dots L_{n-1})\\
&\cdot& \diag(g(L_0)^{-1}, \dots, g(L_{n-1})^{-1})
\end {array}
\end{equation}
Thus $H = TVD \in \mathbb{F}_q^{t \times n}$, where $T \in \mathbb{F}_q^{t \times t}$ is a Toeplitz matrix, $V \in \mathbb{F}_q^{t \times n}$ is a Vandermonde matrix, and $D \in \mathbb{F}_q^{n \times n}$ is a diagonal matrix.
Since a Goppa code is a $\mathbb{F}_p$-subfield subcode, it is possible to express the syndrome function in terms of a parity-check matrix $\bar{H} \in \mathbb{F}_p^{mt \times n}$ using the so-called \emph{trace construction} (see e.g.\cite[Ch.~7, \S~7]{macwilliams-sloane}). This is useful to obtain a syndrome $\bar{s} \in \mathbb{F}_p^{mt}$ equivalent to $s \in \mathbb{F}_q^t$ above while keeping the arithmetic operations in $\mathbb{F}_p$ rather than $\mathbb{F}_q$, even though it is not immediately useful for decoding, at which point a syndrome over $\mathbb{F}_q$ has to be assembled by inverting the trace construction.
The \emph{syndrome decoding problem} consists of computing the error pattern $e$ given its syndrome $s_e$. Knowledge of the code structure in the form of the support $L$ and the polynomial $g$ makes this problem solvable in polynomial time, with some constraints relating the weight of $e$ to the degree of $g$.
\subsection{Polynomial Lattices}
Let $A \in \mathbb{F}_q[x]^{n \times m}$ be a polynomial matrix, and let $r$ denote its rank (i.e. assume that $A$ has $r$ linearly independent rows). The (polynomial) lattice $\lat(A)$ over $\mathbb{F}_q[x]$ spanned by the rows of $A$ is
\[
\lat(A) = \{ (u_0, \dots, u_{n-1}) A \in \mathbb{F}_q[x]^m \mid (u_0, \dots, u_{n-1}) \in \mathbb{F}_q[x]^n \}.
\]
The notion of length which we will use for $f \in \mathbb{F}_q[x]$ is $|f| = \deg(f)$. For polynomial vectors $v \in \mathbb{F}_q[x]^n$ we adopt the notion of maximal degree length: $|v| = \max_i |v_i|$. This notion is coarse enough that, contrary to integer lattices where finding even an approximation to the shortest vector by a constant factor is a hard problem~\cite{micciancio:focs}, reducing a basis for a polynomial lattice can be achieved in polynomial time. The following result by Mulders and Storjohann holds~\cite{mulders-storjohann}:
\begin{theorem}\label{thm:mulders-storjohann}
There exists an algorithm which finds the shortest nonzero vector in the $\mathbb{F}_q[x]$-module generated by the rows of $A$ with $O(m n r d^2)$ operations in $\mathbb{F}_q$, where $d = \max \{ \deg(A_{ij}) \mid 1 \leqslant i \leqslant n, \; \leqslant j \leqslant m \}$.
\end{theorem}
The algorithm whose existence is established by Theorem~\ref{thm:mulders-storjohann} is based on converting a given lattice basis to the so-called weak Popov form, formally defined in Appendix~\ref{app:weak-Popov}, which also contains a description of Algorithm \ref{alg:weak-Popov} and its cost behavior in the context of decoding.
The weak Popov form is not the only way to find short vectors in a polynomial lattice, and in fact this is not critical to our proposal in this paper; for instance, the method by Lee and O'Sullivan \cite{lee-osullivan}, which is related to Gr\"{o}bner bases, would appear to be an alternative. Our choice of the Mulders-Storjohann method, which computes the weak Popov form, derives from its conceptual simplicity and ease of implementation, since the result turns out to be a natural generalization of Patterson's decoding algorithm described in the next section.
\section{Patterson's Decoding Method}\label{sec:patterson}
We briefly recapitulate Patterson's decoding algorithm~\cite{patterson}, which will provide the basis for the general algorithm we propose. The goal, of course, is to compute the error pattern $e$ given its syndrome $s_e$ and the structure of $\Gamma(L, g)$.
Let $q = 2^m$, and assume we are given a binary Goppa code $\Gamma(L, g)$ where the monic polynomial $g$ is irreducible. We define the Patterson locator polynomial $\sigma \in \mathbb{F}_q[x]$ as:
\begin{equation}\label{eq:patterson-sigma}
\sigma(x) := \prod_{e_i = 1}{(x - L_i)}.
\end{equation}
The name \emph{locator polynomial} comes from the fact that the roots of $\sigma$ clearly indicate where errors occurred, since $\sigma(L_j) = 0 \Leftrightarrow e_j = 1$. Taking the derivative of the formal power series underlying $\sigma$, we obtain
\begin{eqnarray*}
\sigma'(x) &=& \sum_{e_i = 1}{\prod_{\substack{e_j = 1\\j \neq i}}{(x - L_j)}}\\
&=& \sum_{e_i = 1}{\dfrac{1}{x - L_i}\prod_{e_j = 1}{(x - L_j)}}\\
&=& \sigma(x)\sum_i{\dfrac{e_i}{x - L_i}},
\end{eqnarray*}
and hence, in $\mathbb{F}_q[x]/g(x)$,
\begin{equation}\label{eq:key-binary}
\sigma'(x) = \sigma(x) s_e(x) \mod{g(x)}.
\end{equation}
This is called the \emph{key equation}, and now we discuss how to solve it.
Being a polynomial in characteristic~2, $\sigma(x)$ modulo $g(x)$ can be written as
\[
\sigma(x) = a_0(x)^2 + x a_1(x)^2
\]
for some $a_0(x)$, $a_1(x)$ with $\deg(a_0) \leqslant \lfloor t/2 \rfloor$ and $\deg(a_1) \leqslant \lfloor (t - 1)/2 \rfloor$, and hence
\[
\sigma'(x) = 2\,a_0(x)\,a_0'(x) + a_1(x)^2 + 2\,a_1(x)\,a_1'(x)\,x = a_1(x)^2,
\]
since the characteristic is~2.
Therefore
\begin{eqnarray*}
a_1(x)^2 &=& \sigma'(x) = \sigma(x) s_e(x)\\
&=& \left(a_0(x)^2 + x a_1(x)^2\right) s_e(x) \mod g(x),
\end{eqnarray*}
whence
\begin{equation}\label{eq:bezout}
a_0(x) = a_1(x) v(x) \mod{g(x)}
\end{equation}
where $v(x)$ is a polynomial satisfying $v(x)^2 = x + 1/s_e(x) \mod g(x)$. Such a polynomial surely exists in characteristic~2 if $g(x)$ is square-free: if $g(x) = \prod_i{g_i(x)}$ where each $g_i(x)$ is irreducible, then $v(x) \bmod g_i(x)$ can be computed as a square root of $x + 1/s_e(x) \bmod g_i(x)$ viewed as an element of the finite field $\mathbb{F}_q[x]/g_i(x)$, and $v(x) \bmod g(x)$ can then be obtained by combining the results via the Chinese Remainder Theorem. We can thus assume that $\deg(v) < t = \deg(g)$.
The last equation is actually a B\'{e}zout relation $a_0(x) = a_1(x)v(x) + \lambda(x)g(x)$, which can be solved for $a_0(x)$ and $a_1(x)$ with the restriction $\deg(a_j) \leqslant \lfloor (t - j)/2 \rfloor$ (and also $\lambda(x)$ but it is not used) using the extended Euclidean algorithm. Solutions $(a_0,a_1)$ can also be seen as short vectors in the lattice spanned by the rows of the following matrix:
\[
A =
\left[
\begin{array}{cc}
g & 0\\
v & 1
\end{array}
\right]
\]
in the sense that the degrees of these polynomials are much smaller than uniformly random vectors, since $(\lambda, a_1)A = (\lambda g + a_1 v,a_1) = (a_0, a_1)$ for some $\lambda \in \mathbb{F}_q[x]$, by virtue of Equation~\ref{eq:bezout}. Therefore, Algorithm \ref{alg:weak-Popov} can be used to find candidate solutions $(a_0, a_1)$.
At first glance there is no guarantee that a short vector in the lattice generated by $A$ yields the desired solution; in other words, being short is a necessary condition, but in principle not a sufficient one.
However, the fact that in the binary case the minimum code distance is known~\cite{goppa} to be at least $2t + 1$ actually restricts $\sigma$ to a single candidate, so that Algorithm \ref{alg:weak-Popov} is bound to find it. Thus, decoding is always successful up to $t$ introduced errors.
\section{Decoding Codes over $\mathbb{F}_p$}\label{sec:extension}
We now show how to generalize Patterson's decoding algorithm so as to correct errors for codes defined over general prime fields. Thus, let $q = p^m$ for some prime $p$ and some $m > 0$, and assume we are given an irreducible Goppa code $\Gamma(L, g)$ over $\mathbb{F}_p$.
Let $\phi \in \mathbb{F}_p \setminus \{0\}$ be a constant scalar. We define the generalized error locator polynomial to be
\begin{equation}\label{eq:locator}
\sigma_\phi(x) := \prod_i{(x - L_i)^{e_i/\phi}}
\end{equation}
where the value $e_i/\phi$ is lifted from $\mathbb{F}_p$ to $\mathbb{Z}$ (i.e. the value $e_i/\phi \in \mathbb{F}_p$ that occurs as an exponent is taken to mean its corresponding least non-negative integer representative, which lies in range $0 \dots p-1$).
One can easily see that this definition actually coincides with Patterson error locator polynomials as defined by Equation~\ref{eq:patterson-sigma} for $p=2$. Lifting Equation~\ref{eq:locator} to the field of rational functions in characteristic 0 and taking the derivative, we have
\begin{eqnarray*}
\sigma_\phi'(x) &=& \sum_{j}{(e_j/\phi)(x - L_j)^{e_j/\phi - 1}\prod_{i \neq j}}{(x - L_i)^{e_i/\phi}}\\
&=& (1/\phi)\sum_{j}{\dfrac{e_j}{x - L_j}\prod_i{(x - L_i)^{e_i/\phi}}}\\
&=& (1/\phi)\sigma_\phi(x)\sum_{j}{\dfrac{e_j}{x - L_j}},
\end{eqnarray*}
which over $\mathbb{F}_q[x]$ reduces to
\begin{equation}\label{eq:modif-key}
\phi \sigma_\phi'(x) = \sigma_\phi(x) s_e(x) \mod{g(x)}.
\end{equation}
This is the $\phi$-th key equation of the proposed method, which generalizes Equation~\ref{eq:key-binary} to Goppa codes over $\mathbb{F}_p$.
The actual $\phi$ must be chosen so as to minimize the degree of $\sigma_\phi$ (and hence maximize the number of correctable errors). One cannot expect to know \emph{a priori} the value of $\phi$, but since there are only $p-1$ possibilities, the error correction strategy will be to try each of them in turn.
Notice that the maximum number of correctable errors can be, and usually is, less than $t$, since the degree of $\sigma_\phi$ exceeds the number of roots in the presence of multiple roots. The following theorem provides an upper bound for how many errors can be corrected by solving Equation~\ref{eq:modif-key} when the distribution of error magnitudes is not taken into account.
\begin{theorem}
The maximum number of errors that can be corrected by solving Equation~\ref{eq:modif-key} independently of the distribution of error magnitudes is $w = (2/p)t$.
\end{theorem}
\begin{proof}
Let $w_v$ denote the number of times the magnitude $v$ occurs in an error pattern of weight $w$, so that $\sum_v{w_v} = w$.
Since we are working with a Goppa code, the constraint for correctability is $\deg(\sigma_\phi) = \sum_v{(v/\phi)w_v} \leqslant t$. In the extreme situation when the weight of the error pattern reaches $w$, the most often error magnitude occurs $w_{\max} \geqslant w/(p-1)$ times, attaining the lower bound when all error magnitudes occur with equal frequency.
In that case, $\deg(\sigma_\phi) \leqslant \sum_v{(v/\phi)w_{\max}} = (1 + 2 + \cdots + (p-1)) w/(p-1) = wp/2 \leqslant t$, and hence no more than errors than $w = (2/p)t$ can be corrected independently of the distribution of magnitudes, as claimed.
\end{proof}
Since the proposed method coincides with Patterson's for $p = 2$, it is not surprising that $t$ errors can be corrected in that case. However, in characteristic~3 the number of potentially correctable errors is $(2/3)t$, non-deterministically exceeding the limit of $t/2$ errors attainable by previously known decoding methods for codes of degree $t$, except in the case of so called ``wild codes'' \cite{bernstein-lange-peters:wild} whereby the Goppa polynomial is a $(p-1)$-th power of an irreducible polynomial (our method, by contrast, applies when that polynomial is square-free, as we will see in Section~\ref{sec:solving}).
Despite the low general limit of $(2/p)t$ correctable errors for $p > 3$, it is still possible to exceed that limit in any odd characteristic if the distribution of error magnitudes is unbalanced. Indeed, all that is required to get a chance of uniquely decoding a word containing $w \leqslant t$ errors is that $\deg(\sigma_\phi) \leqslant t$ for some choice of $\phi$ and that the actual distance from the right codeword to any other codeword be at least $2w + 1$.
The actual number of correctable errors depends heavily on the distribution of error magnitudes and has to be computed in a case-by-case basis, always laying in the range $(2/p)t$ to $t$. In particular, if all error magnitudes are equal, in principle one could correct $t$ errors, even though this is a statistical rather than deterministic behavior.
This is especially useful for cryptographic applications involving an all-or-nothing transform~\cite{rivest}, as it happens e.g. for a semantically secure encryption scheme involving the McEliece one-way trapdoor function~\cite{mceliece} and the Fujisaki-Okamoto conversion~\cite{fujisaki-okamoto}. In such scenarios, the magnitudes of the introduced errors can be chosen to be all or nearly all equal by convention, making the proposed decoder attractive for its higher decodability bound, under the explicit assumption that decoding them remains hard.
\subsection{Solving the Key Equation}\label{sec:solving}
We now focus on actually solving Equation~\ref{eq:modif-key}.
Being a polynomial in characteristic~$p$, $\sigma_\phi(x)$ can be written as
\begin{equation}\label{eq:decompose}
\sigma_\phi(x) = \sum_{k=0}^{p-1}{x^k a_k(x)^p}
\end{equation}
for some $a_k(x)$ with $\deg(a_k) \leqslant \lfloor (t - k)/p \rfloor$, $0 \leqslant k \leqslant p-1$, and hence
\begin{eqnarray*}
\sigma_\phi'(x) &=& \sum_{k=0}^{p-1}{\left(k x^{k-1} a_k(x)^p + p x^k a_k(x)^{p-1} a_k'(x)\right)}\\
&=& \sum_{k=1}^{p-1}{k x^{k-1} a_k(x)^p}
\end{eqnarray*}
since the characteristic is $p$.
Therefore
\begin{eqnarray*}
\phi\sum_{k=1}^{p-1}{k x^{k-1} a_k(x)^p} &=& \phi\sigma_\phi'(x) \;=\; \sigma_\phi(x) s_e(x)\\
&=& \left(\sum_{k=0}^{p-1}{x^k a_k(x)^p}\right) s_e(x)
\bmod{g(x)},
\end{eqnarray*}
whence
\begin{equation}\label{eq:lattice}
a_0 + \sum_{k=1}^{p-1}{a_k(x) v_k(x)} = 0 \mod g(x)
\end{equation}
where the $v_k(x)$ are polynomials satisfying $v_k(x)^p = x^k - \phi k x^{k-1} / s_e(x) \mod g(x)$. Such polynomials surely exist in characteristic~$p$ if $g(x)$ is square-free: if $g(x) = \prod_i{g_i(x)}$ where each $g_i(x)$ is irreducible, then $v_k(x) \bmod g_i(x)$ can be computed as a $p$-th root of $x^k - \phi k x^{k-1} / s_e(x) \bmod g_i(x)$ viewed as an element of the finite field $\mathbb{F}_q[x]/g_i(x)$, and $v_k(x) \bmod g(x)$ can then be obtained by combining the results via the Chinese Remainder Theorem. We can thus assume that $\deg(v_k) < t = \deg(g)$.
The Diophantine equation~\ref{eq:lattice} has to be solved for $a_k(x)$ with the stated restriction on their degrees. Solutions $(a_0, a_1, \dots, a_{p-1})$ can be seen as short vectors in the lattice spanned by the rows of the matrix
\begin{equation}\label{eq:lattice-basis}
A_\phi =
\left[
\begin{array}{ccccc}
g & 0 & 0 & \dots & 0\\
-v_1 & 1 & 0 & \dots & 0\\
-v_2 & 0 & 1 & \dots & 0\\
\vdots & \vdots & \vdots & \ddots & \vdots\\
-v_{p-1} & 0 & 0 & \dots & 1
\end{array}
\right],
\end{equation}
since, by virtue of Equation~\ref{eq:lattice}, one has $(\lambda, a_1, \dots, a_{p-1}) A_\phi = (\lambda g - \sum_{k=1}^{p-1} a_k(x) v_k(x), a_1, \dots, a_{p-1}) = (a_0, a_1, \dots, a_{p-1})$ for some $\lambda \in \mathbb{F}_q[x]$.
Therefore, Algorithm \ref{alg:weak-Popov} can be used to find candidate solutions $(a_0, \ldots, a_{p-1})$.
The method is applicable whenever one can actually invert $s \bmod g$ and then compute the $p$-roots needed to define the $v_k$ polynomials. This is always the case when $g$ is irreducible, but not exclusively so. Indeed, to compute the $v_k$ it suffices that $g$ is square-free and that $s$ is invertible modulo each of the irreducible factors of $g$, since in this case the $v_k$ can be easily computed modulo those irreducible factors and finally recovered via the Chinese Remainder Theorem.
Theorem~\ref{thm:mulders-storjohann} ensures a cost not exceeding $O(p^3 t^2)$ $\mathbb{F}_q$ operations for computing short vectors in $\lat(A_\phi)$.
\subsection{Estimating the Success Probability}\label{sec:success-prob}
Regrettably the ability to find shorts vectors in lattice $\lat(A_\phi)$ does not mean that any such vector yields a solution to Equation~\ref{eq:modif-key}. We will now see that, fortunately, the proposed method has a surprisingly favourable probability of finding the right $(a_0,\ldots,a_{p-1})$ that solves Equation~\ref{eq:modif-key}.
As a cautionary note, we stress that a full theoretical analysis of the failure probability remains, to the best of our knowledge, an open problem. Because the distance notion here is not Euclidean, but rather that of Hamming, Minkowski's theorem on the existence of a lattice point in any large enough convex set does not appear to apply in our case. Also, the analysis would also appear to require a detailed theory on the distribution of the average distance between a vector and the closest codeword whose magnitudes satisfy some constraint (like being all equal, or following a highly skewed distribution), which to the best of our knowledge is too an open problem.
As a consequence, the failure probability estimates we provide are conjectured on an empirical basis. Namely, they result from experiments conducted on a large number of random Goppa codes at which our decoding method is targeted, and for each of those codes, on a large number of decoding attempts on random error patterns following the particular magnitude constraint (equal for all error positions) for which the decoder works best.
In a successful decoding, the reduced basis for lattice $\lat(A_\phi)$ leads to candidates for $\sigma_\phi$ with degree $t$ onward, of which of course only the candidate with the smallest degree is the correct one. Spurious candidates of degree close to $t$ result from random-looking short (albeit not shortest) vectors in the reduced basis and are usually harmless. But the fact that those short vectors are ``random-looking'' means they are also a threat: if by chance they are such that the coefficient of the highest-degree term in the associated spurious $\sigma_\phi$ vanishes, $\deg(\sigma_\phi)$ becomes $t$ or less. Since this is connected with the vanishing of a coefficient from $\mathbb{F}_q$, this event happens with probability $1/q$ assuming that short spurious vectors in $\lat(A_\phi)$ are approximately uniformly distributed.
In general, when trying to correct $w < t$ errors of equal magnitude for a uniformly random irreducible Goppa code, the top $t + 1 - w$ coefficients in the spurious $\sigma_\phi$ must vanish to interfere with the decoding process, whence the probability of successful decoding is roughly $1 - 1/q^{t + 1 - w}$. This matches the empirically observed behaviour of the proposed method in odd characteristic. Not surprisingly, the method always succeeds for binary codes, since it reduces to Patterson's algorithm and the minimum code distance is known to be at least $2t + 1$.
Table~\ref{tab:results} illustrates the results of experiments in Magma~\cite{magma} supporting the conjecture that the probability of successful decoding is roughly $\Pr_{suc} := 1 - 1/q^{t + 1 - w}$.
For each quadruple $(p, m, t, w)$, a set of 10000 Goppa codes of maximum length $n = p^m$ and degree $t$ plus an error pattern of length $n$, weight $w$ and all magnitudes equal (to a single random value in $\mathbb{F}_p \setminus \{0\}$) were randomly generated, and the proposed method was then applied to decode the syndrome of that error pattern. The predicted number of successful decodings, $N_{pre}$, is then compared with the actually observed number $N_{obs}$ of successful decodings.
For each combination of $p$ and $m$, the first listed $t$ is the largest integer satisfying $p^m - mt > 0$. The number $w$ of introduced errors in then decreased starting from $t$ until the probability of success exceeds 0.9999.
Since $\Pr_{suc}$ is close to 1 for large $q$, reasonably small values are chosen for all parameters so that the probability of decoding failure is large enough to be easily discernible. This is also the reason why more detail is provided for smaller parameters. We omit the results for characteristic~2, since in all tests we conducted no decoding failure was observed, as expected. We stress that these examples are not meant by any means for practical cryptographic use.
\begin{table}[htp]\centering
\caption{Experimental assessment of the probability of decoding success}\label{tab:results}
\begin{tabular}{ccccccc}\hline
$p$ & $m$ & $t$ & $w$ & $\Pr_{suc}$ & $N_{pre}$ & $N_{obs}$\\
\hline
3 & 3 & 8 & 8 & 0.962963 & 9630 & 9670\\
3 & 3 & 8 & 7 & 0.998628 & 9986 & 9992\\
3 & 3 & 8 & 6 & 0.999949 & 9999 & 9999\\
\\
3 & 3 & 7 & 7 & 0.962963 & 9630 & 9639\\
3 & 3 & 7 & 6 & 0.998628 & 9986 & 9989\\
3 & 3 & 7 & 5 & 0.999949 & 9999 & 10000\\
\\
3 & 3 & 6 & 6 & 0.962963 & 9630 & 9645\\
3 & 3 & 6 & 5 & 0.998628 & 9986 & 9991\\
3 & 3 & 6 & 4 & 0.999949 & 9999 & 10000\\
\\
3 & 4 & 20 & 20 & 0.987654 & 9877 & 9883\\
3 & 4 & 20 & 19 & 0.999848 & 9998 & 9997\\
3 & 4 & 20 & 18 & 0.999998 & 10000 & 10000\\
\\
5 & 2 & 12 & 12 & 0.960000 & 9600 & 9612\\
5 & 2 & 12 & 11 & 0.998400 & 9984 & 9985\\
5 & 2 & 12 & 10 & 0.999936 & 9999 & 10000\\
\\
5 & 3 & 41 & 41 & 0.992000 & 9920 & 9924\\
5 & 3 & 41 & 40 & 0.999936 & 9999 & 10000\\
\\
7 & 2 & 24 & 24 & 0.997085 & 9971 & 9989\\
7 & 2 & 24 & 23 & 0.999992 & 9999 & 10000\\
\\
11 & 2 & 60 & 60 & 0.991736 & 9917 & 9922\\
11 & 2 & 60 & 59 & 0.999932 & 9999 & 9999\\
\hline
\end{tabular}
\end{table}
Decoding $w \leqslant (2/p)t$ errors of uniformly random magnitude for a uniformly random irreducible code is of course always successful for $p > 3$, since in that case $w$ does not exceed half the minimum code distance, which is at least $t + 1 > (4/p)t + 1 \geqslant 2w + 1$.
\subsection{Computing the Error Magnitudes}\label{sec:error-eval}
In contrast to generic alternant decoding methods, there is no need to compute an error evaluator polynomial to obtain the error magnitudes in the current proposal. After obtaining $\sigma_\phi(x)$ and finding its roots $L_j$, all that is needed to compute the corresponding error values $e_j$ is to determine the multiplicity $\mu_j$ of each root, since one can see from Equation~\ref{eq:locator} that $e_j = \phi\mu_j$.
Computing $\mu_j$ is accomplished by determining how many times $(x - L_j) \mid \sigma_\phi(x)$ whenever $\sigma_\phi(L_j) = 0$, or alternatively by finding the highest derivative of $\sigma_\phi$ such that $\sigma_\phi^{(h)}(L_j) = 0$ (and setting $\mu_j \gets h$).
Since the value of $\phi$ is not known \emph{a priori}, and even in scenarios where it is actually known beforehand, an additional syndrome check is necessary for each guessed $\phi$, and the process usually must check all possible values of $\phi$ anyway since more than one solution may exist.
\subsection{The Completed Decoder}\label{sec:decoder}
We are finally ready to state the full decoding method explicitly in Algorithm~\ref{alg:decoder}.
It can be seen as a list decoding algorithm with possible failures.
The polynomial decomposition of Equation~\ref{eq:decompose} immediately suggests a simple and efficient way to compute the $p$-th roots needed at Step~\ref{step:p-th-root}, namely, precompute $r(x) \gets \sqrt[p]{x} \mod g(x)$ and $r(x)^k \bmod g(x)$, and then compute the $p$-th root of $z(x) := \sum_k{x^k z_k(x)^p}$ as $\sqrt[p]{z(x)} \bmod g(x) \gets \sum_k{r(x)^k z_k(x)}$. The test in Step~\ref{step:test-inverse} is unnecessary if $g$ is irreducible.
To find the zeroes of $\sigma_\phi$ in Step~\ref{step:find-zeroes} one can use the Chien search technique~\cite{chien}, in which case the multiplicities of each root can be determined as part of the search, or the Berlekamp trace algorithm~\cite{berlekamp:zeroes}.
\begin{algorithm
\caption{Decoding $p$-ary square-free Goppa codes}\label{alg:decoder}
\begin{algorithmic}[1]
\item[\textsc{Input:}] $\Gamma(L, g)$, a Goppa code over $\mathbb{F}_p$ where $g$ is square-free.
\item[\textsc{Input:}] $H \in \mathbb{F}_q^{r \times n}$, a parity-check matrix in the form of Equation~\ref{eq:parity-check}.
\item[\textsc{Input:}] $c' = c + e \in \mathbb{F}_p^n$, the received codeword with errors.
\item[\textsc{Output:}] a list of corrected codewords $c \in \Gamma(L, g)$ ($\varnothing$ upon failure).
\State $s^{\textrm{\tiny T}} \gets H c'^{\textrm{\tiny T}} \in \mathbb{F}_q^n$, $s_e(x) \gets \sum_i{s_i x^i}$. \ZComment N.B. $H c'^{\textrm{\tiny T}} = H e^{\textrm{\tiny T}}$.
\If{$\nexists \; s_e^{-1}(x) \bmod g(x)$}\label{step:test-inverse}
\State \Return $\varnothing$
\EndIf
\State $S \gets \varnothing$
\For{$\phi \gets 1 \; \textbf{to} \; p-1$}\label{step:guess-scale} \ZComment guess the correct scale factor $\phi$
\For{$k \gets 1 \; \textbf{to} \; p-1$}
\State $u_k(x) \gets x^k + \phi k x^{k-1} / s_e(x) \mod {g(x)}$
\State\label{step:p-th-root} $v_k(x) \gets \sqrt[p]{u_k(x)} \bmod g(x)$
\EndFor
\State Build the lattice basis $A_\phi$ defined by Equation~\ref{eq:lattice-basis}.
\State Apply Algorithm~\ref{alg:weak-Popov} to reduce the basis of $\lat(A_\phi)$.
\For{$i \gets 1 \; \textbf{to} \; p$}\label{step:take-shortest}
\State Let $a$ denote the $i$-th row of the reduced basis of $\lat(A_\phi)$
\For{$j \gets 0 \; \textbf{to} \; p-1$}
\If{$\deg(a_j) > \lfloor (t - j)/p \rfloor$}
\State \textbf{try next} $i$ \ZComment not a solution
\EndIf
\EndFor
\State $\sigma_\phi(x) \gets \sum_j{x^j a_j(x)^p}$
\State\label{step:find-zeroes} Compute the set $J$ such that $\sigma_\phi(L_j) = 0$, $\forall j \in J$.
\For{$j \in J$}
\State Compute the multiplicity $\mu_j$ of $L_j$.
\State $e_j \gets \phi \mu_j$
\EndFor
\If{$H e^{\textrm{\tiny T}} = s^{\textrm{\tiny T}}$}
\State $S \gets S \cup \{c' - e\}$
\EndIf
\EndFor
\State \Return $S$
\EndFor
\end{algorithmic}
\end{algorithm}
\section{Conclusion}\label{sec:conclusion}
We described a new decoding algorithm for square-free (in particular, irreducible) Goppa codes of degree $t$ over $\mathbb{F}_p$ that can correct $(2/p)t$ errors in general, and up to $t$ errors for certain distributions of error magnitudes of cryptographic interest. By attaining an correction capability of $(2/3)t$ errors in characteristic~3 with high probability, our method outperforms the best previously known decoder for that case, and suggests that the corresponding average distance to the closest codeword is at least $(4/3)t + 1$ for most irreducible ternary Goppa codes. Regardless of the characteristic, our proposal can correct a still larger number of errors that approaches (and probabilistically reaches) $t$ as the distribution of error magnitudes becomes ever more skewed toward the predominance of some individual value. The method can be viewed as generalizing Patterson's binary decoding procedure, and is similarly efficient in practice.
A further increase in the number of correctable errors may be possible by resorting to list decoding and by extracting more information from the decoding process along the lines proposed by Bernstein~\cite{bernstein}.
This in principle might enable the correction of approximately $n - \sqrt{n(n - (4/p)t)}$ errors in general, and possibly as many as $n - \sqrt{n(n - 2t - 2)}$ errors depending on the distribution of error magnitudes.
Furthermore, the ability to correct close to $t$ errors with high probability means that smaller keys might be adopted for coding-based cryptosystems. Properly chosen parameters would keep the probability of decoding failure below the probability of breaking the resulting schemes by random guessing, while maintaining the security at the desired level.
We leave the investigation of such possibilities for future research.
\bibliographystyle{IEEEtran}
|
\section{Motivation}
Recent experiments, such as the CHOOZ experiment [1], have been able to set lower bounds on the determination of $\theta_{13}$:
\[ \sin^2\theta_{13} < 0.19 \]
The experiment has also indicated that reactor-based uncertainties have become the dominant systematic error in extracting $\theta_{13}$. In a breakdown of contributions to the overall systematic errors [2], it is clearly seen that the error on the antineutrino flux outweighs the others. Thus, if one is to decrease the $\theta_{13}$ lower bound, one must contend with and accurately model this flux; this requires detailed simulation of the reactor internals. Prediction of the flux is our final goal; to be confident in our predictions, however, we must benchmark our code against existing real data. To expedite the benchmarking process, instead of using flux data, we use measurements of the antineutrino rate from a reactor since we were in ready possession of this data. Simulation of the rate and flux both require our code to predict a set of key quantities: the \textit{fission rates} of the four radioisotopes that contribute $> 99$\% of the fissions in a nuclear reactor. If our rate prediction is adequate, then we can immediately apply our results to a flux prediction when that data is available. Thus, for the rest of this report, we will focus only on the rate prediction, particularly on the determination of the fission rates. First, we give a brief overview of the reactor and detector from which we have obtained our rate data.
\section{SONGS: Reactor and Detector}\label{subsec:songs}
\subsection{Reactor}
SONGS Unit 2 is a PWR (pressurized water reactor) in California. It is operated by Southern California Electric (SCE) and has a thermal power output of 3.46 GW. Our colleagues at Lawrence Livermore National Laboratory (LLNL) were able to obtain very detailed inputs and outputs of their simulation package that were then employed as inputs to our own reactor simulation. These SCE values were not accompanied by uncertainties, so it will be up to us in the future to assign and / or estimate the associated uncertainties to these variables.
The reactor core is divided into 217 subunits called \textit{fuel assemblies}. Each of these fuel assemblies is a bundle of fuel rods (explaining a common alternate name, the \textit{fuel bundle}). The fuel assembly is a ``degree of freedom" for the reactor core; after a fuel-burning cycle, the assembly as a whole can be removed, refreshed, and reloaded. This subdivision is also convenient since our DRAGON code also specializes in assembly simulation.
This reactor is labeled ``16 x 16" because it is comprised of a 16-by-16 grid of fuel rods (236) and instrumentation rods (20). These fuel rods contain ``fresh" fuel (only uranium) as well as ``burned" mixtures of fuel (uranium with plutonium and other heavy metals). Detailed information about each assembly, from their temperature during fission to their composition of the fuel cladding, was provided for the simulation.
\subsection{Detector}
The SONGS detector, commissioned and operated by LLNL and Sandia Laboratory, contains approximately 0.64 metric ton of gadolinium-doped scintillator, located 24.5 meters from the SONGS reactor core. The detection mechanism uses the tried-and-true inverse $\beta$ decay process, whose products produce a delayed coincidence signal between the annihilating positron and the neutron capture; the very high thermal capture cross section of gadolinium enhances this signal. The overall detection efficiency is 10.7\%, which is itself known to approximately 10\%.
With this information, we can now simulate the SONGS reactor, the topic to which we turn next.
\section{The DRAGON Code}
DRAGON is an open-source simulation package that allows one to study the behavior of neutrons in a nuclear reactor. It allows one to determine the isotopic concentrations of radionuclides during the burnup cycle, as well as to perform isotopic depletions. The open source nature of this code is attractive not only because of cost factors, but also due to the relative ease of its modification. Indeed, \textbf{our DRAGON simulation package has been modified to output the fission rates of the four most important isotopes in fissions, $^{235}U, ^{238}U, ^{239}Pu$, and $^{241}Pu$}. Obtaining these fission rates as a function of time is the primary use of DRAGON.
Since DRAGON is a deterministic code, not a Monte Carlo code (such as MCNP), there is no error associated with its calculations and we must estimate this ourselves. DRAGON is also known as a ``2D" code, since it only simulates a 2D cross section of the fuel assembly. A full 3D solution of the multidimensional neutron diffusion equation is very expensive deterministically, so DRAGON provides the option of interfacing its 2D solution with its companion code, DONJON, to do a 3D calculation. However, as a first simplification, we have merely scaled the fission rates from DRAGON by the height of the reactor ($H$ = 381 cm). It has been checked that this procedure is appropriate unit-wise. While it makes the assumption that the fission rates--and thus the flux--will be uniform along the axial direction when they are not, we note that the CHOOZ reactor simulations also made this uniform axial flux assumption [3], and ours can be relaxed in the future with DONJON.
With the fission rates, we can calculate the antineutrino rate.
\section{The SONGS Detected Antineutrino Rate}
\begin{figure}
\begin{center}
\includegraphics[scale=0.6]{Songs.png}
\caption{SONGS Rate Prediction With DRAGON}
\end{center}
\end{figure}
DRAGON's prediction of the rate detected at SONGS is given by
\[ \frac{dN_\nu}{dt} = \frac{\epsilon N_p}{4\pi D^2} \sum_i f_i \int_{1.806}^{\infty} \,dE_\nu \sigma(E_\nu) S_i(E_\nu) \]
Here, $i$ sums over the aforementioned isotopes, $f_i$ are fission rates, $\sigma(E_\nu)$ is the inverse $\beta$ decay cross section, $D$ is the distance to the reactor, $N_p$ is the number of target protons (= 4.35 $\times 10^{28}$) in the detector, and $\epsilon$ is the detection efficiency. The functions $S_i$ are the parameterizations of the \textit{Schreckenbach spectra}, which provide the number of antineutrinos produced per fission, per radionuclide. This parameterization is provided by Petr Vogel [4].
\subsection*{Discussion}
The work on this rate calculation has progressed since the time of the conference and the plot shown here differs greatly from the one shown there. The normalization concerns mentioned during the author's talk have been corrected. Also, this time two different fuel cycles are shown. The SONGS data were taken at the latter half of Cycle 12; after the reactor was shut down and refueled, data were taken for the first half of the subsequent Cycle 13. These cycles had different fuel loadings and thus different rate predictions.
The plot here shows the prediction for Cycle 12 with a 2\% error band (red band); 2\% was chosen as a broad estimate since American reactors must be certain that their power output is within 2\%. Due to the 10\% uncertainty in the detector efficiency (blue band), it is difficult to estimate DRAGON's effectiveness at predicting the rate. However, there are other comparisons that we can make; one of these is to ascertain if DRAGON predicts the same isotopic concentrations at the end of its fuel cycle as predicted by SCE itself. Work of this nature is underway. Also, even if absolute normalization is lacking, information about the slope will suffice for some rudimentary nonproliferation efforts.
\section*{Acknowledgments}
The author would to thank parties at both MIT (Professor Janet Conrad and Dr. Lindley Winslow) and at LLNL (Drs. Adam Bernstein, Gregory Keefer, and Nathaniel Bowden) for their support and frequent discussions. The author is especially grateful for the generous assistance of Professor Guy Marleau, a principal author of the DRAGON code, for support in the code modification. The author is supported by NSF grant PHY-0847843.
|
\section{Introduction}
The presence of ionized gas in early-type galaxies, part of a complex
multiphase interstellar medium (ISM), has now been firmly established. Among
other issues, the source of ionization that powers the observed emission lines
is still a matter of debate. Several alternatives have been advanced,
ascribing it to active galactic nuclei (AGN), to hot evolved stars, and to
shocks (e.g. \citealt{heckman80,binette94,dopita95}). A better understanding
of this phenomenon is important for several reasons, e.g. in the contexts of
the origin and evolution of early-type galaxies and of their ISM and
of a correct census and demography of active galaxies.
In this paper we explore this issue by studying homogeneous sample of
galaxies, restricted to a rather narrow range of galaxy properties. We isolated
nearby, quiescent (from the point of view of star formation), massive
early-type galaxies from the SDSS \citep{york00}. More specifically, from the
main sample of $\sim$800000 galaxies with spectra available from the SDSS,
Data Release (DR) 7, we used the MPA-JHU DR7 release of spectrum measurements,
available at {\sl http://www.mpa-garching.mpg.de/SDSS/DR7/}, to select all
sources with the following spectroscopic criteria: 1) redshift between 0.01
and 0.1, 2) Ca break strength $D_n(4000)>1.7$, and 3) stellar velocity
dispersion $\sigma_* > 156$ $\rm{\,km \,s}^{-1}$. If adopting the scaling relation between mass
and velocity dispersion this corresponds to galaxies with $M_* \gtrsim 5
\times 10^{10} M_{\sun}$, \citep{hyde09}. We only retain early-type
galaxies, i.e. objects with a concentration index (the ratio of the radii
including 90\% and 50\% of the $r$ band light) $C_{r}\geq 2.86$
\citep{nakamura03,shen03}. This selection yields 27244 red, giant, early-type
galaxies (hereafter RGEs).
\begin{figure*}
\centerline{
\includegraphics[scale=0.75,angle=0]{16388f1a.epsi}
\medskip
\includegraphics[scale=0.76,angle=0]{16388f1b.epsi}
}
\includegraphics[scale=1.15,angle=0]{16388f1c.epsi}
\caption{Left panel: logarithm of the [O~III] emission line flux (in units of
$10^{-17}\>{\rm erg}\,{\rm s}^{-1}\,{\rm cm}^{-2}$) versus the k-corrected $i$ band magnitude within the
SDSS fiber, both quantities corrected for galactic absorption. Green dots
mark upper limits in line flux. The dashed red line corresponds to a
constant ratio between the two quantities. Contours (blue for the objects
with an [O~III] detection, yellow for upper limits) represent the
iso-densities of DR7 galaxies in the same redshift range; levels are in
geometric sequence with a common ratio of 4. Right panel: histogram of the
residuals from the median line. The dashed histogram is the contribution of
upper limits. The dispersion of the distribution is 0.18 dex. Bottom:
residuals from the median vs redshift.}
\label{riga}
\end{figure*}
\section{Emission lines and stellar continuum}
\label{el}
We considered the [O~III]$\lambda5007$ emission line flux measured after
subtracting a starlight template (see \citealt{kauffmann03} for a detailed
description of the method used for the continuum subtraction). The [O~III]
line is detected at a significance higher than 3$\sigma$ in $\sim$ 53\% of
RGEs of the sample. Figure\ref{riga} compares the [O~III] flux with the
k-corrected $i$ band magnitude within the SDSS fiber, with both quantities
corrected for galactic absorption. A strong connection emerges, and the vast
majority of the objects are clustered in a narrow stripe. We used the
algorithms proposed by \citet{dempster77} and \citet{buckley79}, implemented
as {\sl emmethod} and {\sl buckleyjames} in IRAF, to deal with censored data
in order to derive the best-fitting logarithmic slope of the $F_{\rm[O~III]}$
versus $i$ relation. The resulting values are $0.41\pm0.01$ and $0.40\pm0.01$,
respectively, both consistent with a constant ratio between emission line and
stellar continuum.
In the right panel we show the histogram of the residuals from the median line
that is reproduced well by a Gaussian distribution with a dispersion of only
0.18 dex (similar results, although with slightly broader distributions, are
obtained using different lines and/or continuum bands). Nonetheless, there are
outliers with a line excess, R\forb{O}{III}\footnote{Defined as the ratio between the
\forb{O}{III}\ flux of a galaxy with respect to its median value at a given
magnitude.}, of up to 100. Setting a threshold at R\forb{O}{III} = 5, we find 0.5\%
outliers, a fraction that increases to 1.2\% (3.5\%) when lowering the limit
to 3 (2).
The EW of the [O~III] line, estimated by considering the continuum level 200
pixels around the line, has a median value of $\sim 0.75$ \AA\ including
upper limits in the analysis) and the EW dispersion is of 0.20 dex; 95
\% of RGEs with an \forb{O}{III}\ detection have 0.3 $<$ EW $<$ 1.7 \AA.
Contours in Figure\ref{riga} represent the iso-densities of the $\sim$340000
DR7 galaxies in the same redshift range, $0.01<z<0.1$, but without any further
selection of the spectro-photometric properties. The general galaxies
population shows a concentration in the same locus as covered by the RGEs, but
also a substantial population extending toward high line excesses ($\sim$ 30\%
of the sample has an excess of R\forb{O}{III} $>$3).
We looked for trends between the residuals and the spectrophotometric
parameters used for the sample selection (namely redshift, $D_n(4000)$,
concentration index, and velocity dispersion), but we failed to find any
statistically significant link. In particular, there are no differences
between RGEs at low and high redshifts: the best linear fit of the residuals
against redshift indicates that objects at z=0.01 have a ratio between line
and continuum only $(20\pm11)$\% higher than those at z=0.1.
\begin{figure*}
\centerline{
\includegraphics[scale=1.,angle=0]{16388f2.epsi}
}
\caption{Spectroscopic diagnostic diagrams for RGEs. The solid lines are from
\citet{kewley06} and separate star-forming galaxies, LINER, and Seyfert; in
the first panel the region between the two curves is populated by the
composite galaxies. Blue dots mark the ``strong'' outliers (objects with a
line excess with respect to the median value R\forb{O}{III}$>$5) in Figure\ref{riga},
and red dots the ``weak'' outliers (3$<$R\forb{O}{III}$<$5). Contours represent
the iso-densities of all DR7 emission line galaxies.}
\label{diag}
\end{figure*}
\begin{figure}
\centerline{
\includegraphics[scale=0.5,angle=0]{16388f3.epsi}
}
\caption{[N~II]/\Ha\ ratio for all sources with both lines detected. Emission
lines of the objects above the horizontal dashed line are unlikely to be
powered by star formation.}
\label{n2ha}
\end{figure}
\section{Spectroscopic diagnostic diagrams}
Fig.~\ref{diag} shows the location in the spectroscopic diagnostic diagrams
(e.g. \citealt{heckman80,baldwin81,veilleux87,kewley06}) for the RGEs that
have all the relevant emission lines detected at SNR$>$3 separately for each
diagram. Starting from the left side, the percentages with respect to the
whole sample are of 29, 23, and 8\% in the three diagrams, respectively. The
vast majority of the objects fall in the LINERs region, while the Seyfert and
star-forming regions are scarcely populated.
Considering the properties of outliers first, `strong' outliers (i.e. the
objects with R\forb{O}{III}\ $>$ 5) represent essentially all of the Seyferts;
nonetheless, most of them are LINERs. `Weak' outliers (with $3 <$ R\forb{O}{III} $< 5$)
do not differ significantly, on the one hand, from the `stronger' ones (but
there are fewer Seyfert among them), but on the other, they cannot be
readily separated from the bulk of the RGEs population.
We now turn our attention to the objects that have no optical classification
because at least one of the key emission lines is undetected. This subsample
is clearly very important, because it is composed of $\sim$ 3/4 of the
galaxies. Nonetheless, [N~II] and \Ha\ are both detected in $\sim$50\% of the
galaxies and this allows us to derive at least a tentative classification. In
fact, log([N~II]/\Ha) $> -0.3$ in 90 \% of them (see Figure\ref{n2ha}), a
threshold above which no star-forming galaxies are found. On the other hand,
Seyfert galaxies usually have bright emission lines, thence they are expected
to be all properly cataloged by the diagnostic diagrams in
Fig.~\ref{diag}. This leads to the conclusion that these galaxies, amounting
to about half of the RGEs, are generally LINERs.
\section{Error budget of the emission line fluxes}
\begin{figure}
\centerline{
\includegraphics[scale=0.6,angle=0]{16388f4.epsi}
}
\caption{Difference between pairs of duplicated [O~III] flux measurements
divided by the error estimated propagating the individual uncertainties.
The solid line is a Gaussian distribution with a dispersion of 1.}
\label{dupl}
\end{figure}
Since we are studying emission lines generally of very low EW, it is important
to assess the reliability of such measurements and of the error
estimates. Two sources of uncertainties should be considered, the first
related to statistical errors, the second to the effects of starlight
modeling.
We deal with statistical uncertainties by studying the galaxies with
multiple SDSS spectroscopic observations. There are 3103 such objects: 45\%
of them have the [O~III] line detected in both datasets, 35\% are upper
limits in both observations, while 20\% change from detected to undetected
(or vice-versa). Considering the objects detected in both observations, the
differences between the independent measurements closely follow a Gaussian
distribution with a dispersion equal to what is expected based on the
uncertainties in each flux estimate (see Fig ~\ref{dupl}). We conclude that
the error assessment is robust and that most (85\%) of the [O~III]
detections are confirmed by multiple observations.
In addition to the statistical uncertainties, we must also consider the
effects related to the accuracy of the subtraction of the stellar
emission. As discussed in detail by \citet{annibali10}, because of the
degeneracy between age, metallicity, and extinction, fits to the continuum
emission of similarly good quality can produce significantly different
residual spectra. This is particular relevant for the \Hb\ line (and to a
lesser extent for \Ha) which is superimposed on the stellar Balmer
absorption feature. The method proposed by \citet{annibali10} to cope with
this problem in their study of a sample of nearby ETG is to explore the
changes in the emission lines fluxes related to the various stellar
templates, limiting it to those that yield a fit with $\chi^2<2$. The
adopted emission line flux is a weighted average of those derived from all
acceptable fits, while its error is a weighted value of the dispersion
between the individual estimates.
The very large number of SDSS objects considered prevents a detailed
analysis of the error budget considered by these authors. However, this
provides us with guidance on the typical uncertainties that should be
associated with the line measurements due to a template mismatch. From their
published values for their smaller synthetic aperture (comparable in size
with the SDSS fiber), we estimated that the median errors in the EW for \Hb,
[O III], and \Ha\ are 0.16, 0.08, and 0.08 \AA, respectively. We adopt these
reference values also for the SDSS sources. The errors in EW have been been
converted into a line flux error and then added quadratically to the
uncertainties provided by the SDSS database.
Our results for the connection between \forb{O}{III}\ flux and $z$ magnitude are only
marginally changed by adopting this more conservative error treatment. The
fraction of detected sources decreases from 53 to 50\%, but the slope of the
relation remains unchanged. The number of outliers is not affected.
We then reconsidered the diagnostic diagrams, excluding all sources that do
not meet the 3$\sigma$ criterion with the revised errors. The fraction of
sources in the various diagrams decreases by 25 to 40 \%, but their visual
appearance is essentially preserved. In order to quantify more subtle
differences, we estimated the changes of median values and dispersions of
the various ratios. Not surprisingly log$\,$\forb{O}{III}/\Hb\ is the more affected
ratio, because it changes from 0.15 (with an rms of 0.20) to 0.09, (rms =
0.18 dex). This is most likely due to the preferred exclusion of galaxies
with low \Hb\ EW. Nonetheless, the result that the vast majority of the
sources are LINERs is confirmed.
Finally, we explored how crucial the precise value of the adopted errors
on EW is. Doubling the reference values, the number of sources in the
diagnostic diagrams keeps decreasing, but their location is effectively
unchanged, with a median of log$\,$\forb{O}{III}/\Hb\ = 0.05.
\section{Discussion}
Most nearby, red, giant early-type galaxies are emission line galaxies. In
particular, in 53\% of their SDSS spectra, the [O~III]$\lambda5007$ line is
detected. The majority of the objects for which this analysis was
possible show emission line ratios characteristic of LINERs.
The [O~III] flux shows a strong correlation with the flux measured within the
3$\arcsec$ SDSS fiber in the $i$ band. The $i$ band magnitude (being less
affected by uncertainties in the k-correction, absorption, and by differences
in the stellar population) is a good estimator of the stellar mass within the
fiber. Thus the ratio between lines and stellar mass is essentially constant
(showing a dispersion of only 0.18 dex) while both quantities vary by a factor
of $\sim$ 30. Furthermore, there is only a marginal trend for a change of EW
with redshift: objects at z=0.01 have a ratio between line and continuum
$(20\pm11)$\% higher than those at z=0.1 despite the change in the size of the
region covered by the SDSS fiber from $\sim$0.6 to 5.5 kpc.
These two results are not straightforward for explaining if the emission lines
are powered by an AGN. In fact, they require a fine tuning between the
strength of the nuclear ionizing field, the spatial distribution of the
emission lines, and the stellar mass. A general link between AGN activity and
the {\sl total} stellar mass can be envisaged, via the constant ratio between
the mass of stars and of the black hole \citep{marconi03}, while we find a
connection with the stellar mass covered by the SDSS fiber. Furthermore, also
the accretion rate, the second parameter that together with the black hole
mass sets the radiative output of the AGN, should be closely linked with the
amount of stars within the central 3$\arcsec$ of each given galaxy.
Conversely, these findings strongly suggest that processes related to the
stellar population are at the origin of the observed emission lines.
As argued by several authors
(e.g. \citealt{binette94,macchetto96,stasinska08,sarzi10}), hot evolved
stars, such as post-asymptotic giant branch stars (pAGB) and white dwarves
(WD), can produce a substantial diffuse field of ionizing photons. More
quantitatively, the observed median EW for \Ha\ in the sample considered
here is EW$_{\rm H\alpha} \sim 0.8$ \AA, well within the range $0.6-1.7$
\AA\ predicted by \citet{binette94} for a stellar population of age
$\sim10^{10}$ years, assuming that the cold gas intercepts all the ionizing
photons. \citet{stasinska08} have explored photoionization models in which
the Lyman continuum radiation is directly estimated from a stellar
population analysis, and find that, by varying the metallicity and
ionization parameter $U$, the resulting emission line ratios cover the whole
region typical of LINERs in the spectroscopic diagnostic diagrams. In
particular, the location of RGEs is reproduced well by models with
$0.5\lesssim Z_*/Z_{\sun}\lesssim 2$ and log $U \sim -3.7$.
Shocks can represent an additional source of ionization. It has been
shown by, e.g., \citet{allen08} that they can produce emission lines with
ratios mimicking those of active nuclei and of LINERs, in particular. The
connection between stellar mass and emission line flux suggests that such
shocks must be driven by the stellar population, as in the case of
supernovae or of stellar ejecta, rather than by bars of triaxial
perturbations.
We conclude that the properties of emission lines in RGEs can in general be
satisfactorily accounted for by photoionization related to hot evolved
stars or by shocks driven by the stellar processes (but below we discuss the
few and important exceptions of likely AGN dominated objects). The narrow
spread of observed EW, very difficult to explain in an AGN framework, set
strong constraints even when adopting a ``stellar'' interpretation. For example,
in the scenario of lines powered by an evolved stellar population, the
apparently stringent requirement that the cold gas must intercept essentially
all of the ionizing photons, becomes probably instrumental in producing a
narrow distribution of lines EW. In fact this effectively removes the
fundamental degree of freedom related to the ISM optical depth to ionizing
photons. The selection of galaxies within a rather narrow range of the various
spectro-photometric parameters space is also a key ingredient in reducing the
scatter between F\forb{O}{III}\ and $i$ magnitude (see Figure\ref{riga}).
These results were obtained by focusing on RGEs, but the consequences are
probably more far reaching. In fact, in the F\forb{O}{III}\ vs. $i$ plane a strong
concentration is also seen for the general population of galaxies, in the same
region covered by RGEs. This suggests a wider application, possibly extending
to less massive early-type galaxies and to the bulges of spiral galaxies.
Nonetheless, there are outliers from the general link between stellar and line
emission. Only among these sources we found Seyfert-like spectra. Furthermore,
a strong connection emerges between the excess in line emission and the
presence of a nuclear radio source, a characteristic signature of an AGN. Our
selection criteria isolate galaxies whose properties are closely matched to
the typical hosts of low-luminosity radio-loud AGN \citep{baldi10b}.
Considering the outliers with R\forb{O}{III} $>5$ (3), 49\% (45\%) have a radio
counterpart in the FIRST\footnote{Faint Images of the Radio Sky at Twenty
centimeters survey \citep{becker95}} catalog within 2$\arcsec$ from the
optical position. In the rest of the sample, only 6\% of the sources have a
radio counterpart. The connection between outliers and radiosources is
preserved even at very small line excesses, since 31\% of those with $2 <
$R\forb{O}{III} $<3$ and 17\% of the objects with an excess by only a factor of 1.5 to
2, are associated with a FIRST source. Thus, the galaxies that exceed the
rather narrow range of line emission EW set by the stellar processes are very
often associated with an active nucleus. The observed line excess is likely to
stem from the additional source of ionizing photons represented by the
AGN. The rather loose relation between radio power and line luminosity seen in
these low-luminosity AGN (see the results presented by \citealt{baldi10b})
might explain the incomplete coincidence between line excess and the presence
of a radio source and, on the other hand, the reason not all radio sources
produce a detectable increase in the line EW.
Apparently, the location in the spectroscopic diagnostic diagrams is not
sufficient to establish the dominant source of ionizing photons, which,
conversely, is best predicted by the EW of the emission lines. In fact, the
likely AGN show line ratios that are essentially indistinguishable (leaving
aside the few Seyfert-like objects) from those probably dominated by stellar
processes. As already pointed out by \citet{stasinska08}, this has important
ramifications where the census and the properties of AGN are concerned. To
isolate and explore the properties of genuine low-luminosity active galaxies,
spectroscopic observations obtained with a smaller aperture are needed to
increase the contrast between AGN and stellar induced line emission.
\acknowledgements
We thank the David J. Axon and anonymous referee for comments that improved
the clarity of this paper.
|
\section{Introduction}\label{sec:intro}
Flares from luminous blue variables (LBV)
are studied for about four hundred years, since the 1600
flare of P~Cyg \citep{degroot}. However, these objects are rare and thus
relatively unstudied.
As variable objects, LBV are characterised by a complex hierarchy of
variability timescales and amplitudes \citep{HD94}. Usually, three
empirical variability scales are distinguished:
\begin{enumerate}
\item microvariations, with amplitudes $\sim 0\magdot{.}1$ and
relatively small
variability timescales, from days to months
\item S~Dor cycles, or eruptions (following \citet{HD94}) with
amplitudes $\sim 1\div 2\magdot{\,}$, years to
tens of years in length
\item giant eruptions with similar and larger (up to $\sim 100\yr$)
timescales and larger amplitudes ($\gtrsim 2\magdot{\,}$)
\end{enumerate}
Physically, giant eruptions may differ from the variations of the
second type by significant changes in the bolometric luminosity (see
discussion in section \ref{sec:disc:lvar}).
Some authors like \citet{genderen_perceptions} propose division of
the second type variability into normal and very long timescale S~Dor
cycles. Longer-timescale variations indeed seem to represent a
separate mode of LBV variability, that is most evident in the
XXth-century light curve of P~Cyg (see next section)
that is free from ordinary flares but dominated by
microvariations superimposed over slow variations $\gtrsim 10 \rm yr$
in length.
Several well-studied LBVs
exhibit periodic variability components.
In particular, \citet{genderen_new} mention a $\sim 1 \rm yr $
period for AG~Car and a $\sim 7\rm yr$ period for S~Dor. Probably, these
periods do not correspond to any coherent process but
rather reflect some characteristic timescales close to the duration
timescale of a single flare.
$\eta$ Carinae is an exceptional case in this sense. The well-known
5-year period is stable on the timescales of tens of years that
supports the idea that this object is indeed a (rather broad)
binary \citep{damineli}.
Note however that the orbital separation of a massive binary with a 5-year
period should be $a\simeq 2\times 10^{14} (M/100\Msun )^{1/3}
(T/5\yr)^{2/3} \rm cm$. Hydrostatic radii for LBV and related stars
are $\sim 10^{12}\rm cm$ that makes the expected effects of binary
interaction on the central machine (-s) of $\eta$~Car very small. In
this study, we find the broad-band PDS of $\eta$~Car quite similar to
that of other flaring LBVs such as AG~Car.
All the observed variability time scales of LBV stars lie between their
dynamic and thermal times.
Dynamic time scale is determined by the mean density of the
stellar hydrostatic core. Outflowing matter can not take part in
pulsations because the wind rapidly becomes
supersonic. Straightforward estimate for the dynamic timescale as the
free-fall time yields:
\begin{equation}\label{E:tdyn}
t_{dyn} \simeq 0.6 \left( \frac{R_\star}{10^{12}\rm cm}\right)^{3/2}
\left( \frac{M_\star}{100 \Msun}\right)^{-1/2} \rm d
\end{equation}
Due to period doubling (see for example \citet{buchler93}) and instability of
pulsational modes, lower-frequency modes are excited during
pulsations. Probably, this effect is observed in ``ordinary'' hot
supergiant stars, where microvariations occur at unexpectedly long
time scales of several days \citep{genderen_OBA}. Flaring LBV
variability is still much longer.
On the other hand, Kelvin-Helmholtz
time scale for a typical hypergiant is:
\begin{equation}\label{E:tKH}
t_{KH} \sim \frac{GM}{RL} \simeq 2\times 10^4 \frac{M_\star}{100
\Msun} \left( \frac{R_\star}{10^{12}\rm cm}\right)^{-1}
\left( \frac{L_\star}{10^{6}\Lsun}\right)^{-1} \rm yr
\end{equation}
Unfortunately, these time scales are unreachable for modern
observations, but may be, in principle,
studied indirectly by observations of the
ejecta, possibly light echoes from strong flares and using the
statistical properties of LBV ensembles.
All the considered
variability scales are somewhere in between and may correspond either
to secular effects in dynamic-timescale evolution (``slow pulsations'')
or to effects in the expanding atmosphere that has, evidently, longer
characteristic timescales than the hydrostatic core. Stellar wind, ascribed some
velocity $v_w$, mass loss rate $\dot{M}$ and opacity $\kappa$, is
characterised by the radius of the (pseudo-) photosphere $R_{ph} \simeq \kappa
\dot{M}/ 4\pi v_w$ and corresponding time scale that has the physical
meaning of the replenishment time for the matter in the wind:
\begin{equation}\label{E:tref}
t_{ref} \sim R_{ph} / v_w \simeq 2 \frac{\kappa}{\kappa_T}
\frac{\dot{M}}{10^{-5} \Msunyr} \left( \frac{v_w}{100\kms}\right)^{-2} \rm d
\end{equation}
Here, $\kappa_T \simeq 0.4 \rm cm^2 g^{-1}$ is
the Thomson cross-section for the scattering by free electrons in a
hydrogen-rich gas. Real
opacity is higher and contains some contribution of true absorption
processes that may increase the characteristic time scales by about an order
of magnitude (see, for example, \citet{OPAL}, bearing in mind
that, for a typical LBV, $\lg T \sim 4\div 5 $ and $\lg
R \sim -5\div -7$ for the inner wind).
Despite the apparently small value of $t_{ref}$, it is an important
estimate for the characteristic variability time of an optically thick
hot stellar wind.
Longer time scales arise
due to higher opacities
Mass loss modulation during strong outbursts leads to even longer time
scales, up to years.
Bolometric luminosity is not
conserved during strong flares (see discussion in section
\ref{sec:disc:lvar}). Besides this, modeling
LBV photospheres provides evidence \citep{koter}
for more complicated nature of
the variability of these objects, incorporating pulsations, mass loss
variations and some additional mechanisms (such as unstable pulsational
modes) leading to tremendous energy release during giant eruptions.
The overall power density spectrum (PDS) of LBV
variability has not been considered as a whole. It is tempting to
compare LBV stars to
active galactic nuclei where different time scales
of a single flicker noise were for a long time interpreted as
physically distinct variability processes (see for example
\citet{terebizh4151}). Below we show that
indeed for several LBV stars, broad-band power spectra have steep
power-law shapes.
In this work we aim for the broad-band PDS of LBV variability. In
the next section, we make a compilation of observational data on
several reasonably well-studied LBV stars and present their PDS in the
following section \ref{sec:AR}. A possible explanation for the observed
broad-band PDS shapes is proposed in section \ref{sec:varshape}. Results
are discussed in section \ref{sec:disc}.
\section{Observational Data}\label{sec:obs}
We selected three well-studied and representative Galactic LBV stars.
For these objects (see table \ref{tab:objs}),
relatively long observational series exist, spanning periods of time
in excess of 50 years. The
data were taken from the archives of the American Association for Variable
Star Observers. Earlier data are primarily visual magnitude estimates,
therefore we use only visual magnitudes. We also checked that the
data series do not have gaps longer than several months and binned the
observational data points by five to diminish the effect of single
erroneous estimates and the round-off effect connected to the low
precision ($\sim 0\magdot{.}1$) of visual magnitude measurements. Note
that using a median filter would not diminish the round-off effect.
We also consider V-band light curves of comparable length obtained for
two Hubble-Sandage variables in M33, Romano's star (the light curve itself
is described in \citet{V532_photo}) and Hubble-Sandage Variable C
\citep{varc88}.
\begin{table}
\centering
\caption{ Objects selected for analysis. For Galactic objects, the
numbers of data points are given after binning. }\label{tab:objs}
\center{\small
\begin{tabular}{lccc}
Object ID & Variability Limits, mag &
Time span & Number of Data Points \\
\hline
\noalign{\smallskip}
\multicolumn{4}{c}{Galactic objects} \\
\noalign{\smallskip}
P~Cyg & 4.5$\div$ 5.6 & 1917-08-13..2010-05-07 & 3396 \\
$\eta$~Car & 4.5$\div$ 8.0 & 1943-07-24..2010-07-19 & 4637 \\
AG~Car & 5.6$\div$ 8.9 & 1939-12-05..2010-05-03 & 2186 \\
\noalign{\smallskip}
\multicolumn{4}{c}{M33} \\
\noalign{\smallskip}
Romano's~star & 16.1$\div$ 18.6 & 1949-12-20..2009-11-08 & 802 \\
Var~C & 15.3$\div$ 17.9 & 1961-09-13..2005-11-08 & 635 \\
\noalign{\smallskip}
\end{tabular}
}
\end{table}
\subsection{Notes on Individual Objects}
\subsubsection*{P~Cyg}
P~Cyg is known for its outbursts in 1600 and later
\citep{degroot}. However, in the considered nearly 100-year time span it
hardly shows any signatures of activity, except for rare short-lasting
low-amplitude excursions \citep{markova01}.
Variability amplitude is primarily contributed by
low-frequency variability.
Several periodicities, from 17 to 100 days, were reported
\citep{ikolka}.
\subsubsection*{$\eta$~Car}
During the considered time span, its behaviour is as well relatively
quiet, especially if compared with its tremendous outbursts in the XIX
century. In our sample, $\eta$~Car is the only proven binary. Its
orbital period does not show up strongly in the optical
variability, probably because of the large optical depth of the outflows in the
optical. The variability pattern is dominated by the gradual increase
of the optical luminosity.
\subsubsection*{AG~Car}
Variability of this prototypical LBV star was considered in
\citet{genderen_discoveries} together with that of $\eta$~Car. The
object exhibits several strong flares. Periodicity of $\sim 370^{\rm
d}$ (somewhat smaller than the
characteristic flare duration time scale) is known to be present.
\subsubsection*{Romano's Star, or V532 (M33)}
One strong flare with an amplitude of about 2$\magdot{\,}$ and a
couple of smaller maxima represent a typical picture of S~Dor
variability.
The V-band data and data reduction process are described
in \citet{V532_photo}. The object is hotter than all the other stars
of the sample
and reaches the spectral class of WN8 in its low/hot state
\citep{polcaro,us10}. Its spectrum in the quiet state (BIa$^+$e) is similar to
that of P~Cyg.
\subsubsection*{Var C (M33)}
Behaviour of this variable denoted as Variable C by \citet{HS53} was
studied by \citet{varc88}, who conclude that the object behaves as a
typical S~Dor variable with a higher than average mass loss rate. The
variable spends about a half of its time in a high
and cool phase of the S~Dor cycle. Amplitudes of the flares reach
2$\magdot{.}$5 in the visual band. The data were provided by Alla
Zharova (private communication).
\section{Analysis and Results}\label{sec:AR}
All the time series are non-uniform but free from strong aliases,
therefore we use extirpolation method \citep{PR89}, where the initially
non-uniform time series is interpolated upon a regular grid,
to compute the Fourier spectra. Fourier transform for uniform series was
performed using the {\tt fftw } library \citep{FFTW05}. Software
written in C and IDL was used for extirpolation and binning of the
power density spectra (PDS). PDS were binned by 10. We use relative
normalization defined as follows:
\begin{equation}
P_j = \frac{1}{\Delta T} \langle F \rangle^{-2} \left| \tilde{F}_j\right|^2
\end{equation}
where $\langle F\rangle$ is the mean flux, $ \tilde{F}_j$ is the
$j$-th component of digital Fourier transform of the flux $F$ (already
interpolated over a regular grid), $\Delta T= N \Delta t$ is the
total effective time span, $N$ is
the number of data points, $j=0..N-1$, $\Delta t$ is the
spacing of the regular grid. This normalisation has an evident
physical meaning of the relative variability power in a
unit frequency range.
Its expectation does not depend on the time span considered
and on data binning. Variability amplitude may be estimated by integrating the PDS
over frequency domain and taking a square root.
All the obtained PDS (see figure \ref{fig:pds}) have approximately
power-law shapes with putative flattenings at lower ($\lesssim 10^{-4}\rm
d^{-1}$) and higher ($\gtrsim 0.01\rm d^{-1}$) frequencies. The
lower-frequency turnovers are smoothed out by the binning in the
frequency domain, but are better seen at finer binning. The data
were fitted with power-law ($PDS=N\cdot f^{-\alpha}$),
Lorentzian ($PDS=N/\left(1+\left(f/\gamma\right)^2\right)$, with zero
resonance frequency) and power-law
+ white noise ($PDS=N\cdot f^{-\alpha}+N_0$) models. Our choice of the
third model is justified by existence of the observational errors that create an
uncorrelated (white) noise component.
Lorentzian function is relevant because it is expected to represent the PDS
of a Poissonian sequence of exponentially decaying flares (and,
approximately, rapidly-decaying flares of other shapes; see section
\ref{sec:flaring}). It indeed proves to be a reasonable fit for the
objects that demonstrate strong flaring activity.
Fitting results are given is table \ref{tab:objfits}.
\begin{table}
\centering
\footnotesize
\caption{ PDS Fitting Results. }\label{tab:objfits}
\begin{tabular}{p{2.5cm}ccccc}
& \multicolumn{3}{c}{Galactic} & \multicolumn{2}{c}{M33} \\
& P~Cyg & $\eta$~Car & AG~Car & V532 & Var~C\\
\hline
\noalign{\smallskip}\\
\noalign{Power Law: }\\
$N$, $10^{-4}$d$^{-1}$ & $1.9\pm 0.2 $ & $5.3\pm 0.4 $ & $1.6\pm
0.3 $ & $0.59\pm 0.17 $ & $0.42\pm 0.09 $\\
$\alpha$ & $1.17\pm 0.02$ & $1.31\pm 0.01$ & $1.56\pm 0.04$ & $1.63\pm
0.05$ & $1.86\pm 0.05$ \\
$\chi^2/DOF$ & $549/168$ & $1393/230$ & $545/107$ & $166/38$
& $213/30$ \\
\noalign{\bigskip}\\
\hline
\noalign{\smallskip}\\
\noalign{Lorentzian:}\\
$N$, d$^{-1}$ & $38\pm 4 $ & $770\pm 30 $ & $1900\pm 1200 $ &
$1.23\pm 1.07$ & $1.51\pm 0.18 $ \\
$\gamma$, $10^{-4}$d$^{-1}$ & $2.16\pm 0.07 $ & $1.69\pm 0.03 $ & $1.4\pm 1.3 $ & $1.23\pm 1.07$ & $1.51\pm 0.18 $ \\
$\chi^2/DOF$ & $1602/168$ & $2574/230$ & $2167/107$ & $169/38$ & $227/30$ \\
\noalign{\bigskip}\\
\hline
\noalign{\smallskip}\\
\noalign{Power Law with White Noise Component:}\\
$N$, $10^{-4}$d$^{-1}$ & $0.47\pm 0.07 $ & $0.67\pm 0.06$ & $0.090\pm 0.007$ & $0.008\pm 0.004 $ & $0.08\pm 0.03 $ \\
$\alpha$ & $1.43\pm 0.03$ & $1.75\pm 0.02$ & $2.12\pm
0.01$ & $2.34\pm 0.01$ & $2.14\pm 0.07$ \\
$N_W$, $10^{-4}$d$^{-1}$ & $45.2\pm 4.9 $ & $139\pm 6 $ & $302\pm 15 $ & $550\pm 20$ & $500\pm 100 $ \\
$\chi^2/DOF$ & $536/167$ & $1344/229$ & $455/106$ & $123/37$ & $199/29$ \\
\end{tabular}
\normalsize
\end{table}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{var6}
\caption{Power density spectra of the five variables fitted by the power-law +
white noise model. }
\label{fig:pds}
\end{figure*}
P~Cyg here is unique in its spectral shape that has the hardest
power-law slope spanning a frequency range of two orders of
magnitude. It is also unique in being the least active among all the
objects. Other stars that exhibit S~Dor cycles and outbursts have
steeper spectral slopes close to 2 with possible indications for a
low-frequency flattening. Average LBV power spectrum clearly has a
power-law shape between tens of days to years or tens of years and
possibly further toward hundreds of years. Gaining
statistics and increasing the time spans to hundreds of years may allow
to clearly resolve the turnover at about $f\sim 10^{-4}\,\rm
d^{-1}$ that definitely should take place because the light curve
should remain reasonably uniform in time.
All the spectra exhibit curvatures and diffuse peaks that make all the
fits rather poor ($\chi^2/DOF \sim 2\div 5$). The overall broad-band shape
is, however, in good agreement with a single power law (with a
high-frequency contribution from observational uncertainties) in all
the cases.
\section{Variability Pattern of a Pseudo-Photosphere}\label{sec:varshape}
The idea of pseudo-photospheric nature of LBV cycles may be traced down
to the note of \citet{lamers87} who proposed that the primary source of the
variability in these objects is the variability of the mass-loss
rate. Below, we will assume that this variability is fast and
uncorrelated (white noise in mass loss rate) and show that such
variations indeed lead to a Brownian-like noise in the PDS, but with a
turnover at a higher frequency. Lower-frequency parts of LBV PDS require
some additional mechanism of stochastic variability.
\subsection{Basic Assumptions}
Excluding brightest outbursts, variability of luminous blue
variables is generally assumed to be purely spectral. According to
\citet{lamers87}, the basic mechanism responsible for variable
luminosities of LBV stars in the optical is through variations in the mass
loss rate resulting in variable pseudo-photosphere radius and
effective temperature.
In the relatively hot, relatively rarefied atmospheres of massive
stars opacities are dominated by
Thomson scattering, therefore the (radial) optical depth through the
wind may be calculated as:
\begin{equation}\label{E:tau}
\tau(R, t) = \frac{\kappa_T}{4\pi} \int_R^{+\infty}
\frac{\dot{M}(t-r/v(r)) dr}{r^2 v(r)}
\end{equation}
Here, $v=v(r)$ is the velocity of
the wind. Different true absorption processes also contribute to the opacity of
the wind, but their effect depends on the temperature structure of the
wind and on the effects of clumping. We neglect all the effects
of true absorption though they are able to increase the opacity of the wind
by a factor of several. This will mimic a larger
mass accretion rate in the model.
In its simplest form, the radius and effective temperature
as a function of time may be estimated by equating the optical depth to
some fixed value (below we use $\tau=2/3$).
Equation (\ref{E:tau}) may be then solved for $r$.
\subsection{Numerical Simulations}
Six model light curves (see table \ref{tab:sim:lc}) were calculated using a pseudo-photosphere
model with a blackbody photosphere defined by the condition
of $\tau=2/3$. Bolometric luminosity was everywhere set to $L_{bol}=3\times
10^{39}\ergl$ that is close to its value for P~Cyg
as estimated by \citet{PP90}. All the curves are 50 years in length
and contain between 195 and 9774 data points. Model F was aimed on reproducing the basic
variability properties of P~Cyg. Others differ in the unperturbed mass
loss rate $M_0$, mass-loss rate dispersion and wind velocity.
We use the semi-empirical $\beta$-law
$v(r)=\left(v_{\infty}-v_0\right)\left( 1-r/R\right)^\beta+v_0$ with
$\beta=1$. Wind acceleration spatial scale $R$ is assumed identical to
the hydrostatic inner radius
$R$, and we set $v_0=2.5 \kms$ everywhere.
Wind velocity at infinity was fixed, while the mass
loss rate experiences log-normal variations that we assume uncorrelated
on the time scales of interest. Of course, real stellar winds have
variable velocities, but simultaneous variations of wind velocity and
mass loss rate do not change the picture qualitatively. Variable wind
velocity would however lead to formally divergent solutions for
density distribution. In real winds, shock waves are expected to be
formed. Hence the effects of variable wind velocity should be
considered together with deviations from spherical symmetry and
quasi-stationarity.
Log-normality of the mass-loss rate distribution
is a useful assumption because is allows to
easily account for dramatic changes in the value of
$\dot{M}$ without producing unphysical negative values.
If the mass-loss rate varies on
dynamical time scales, the wind will work as an integrator, making the
observed power density spectrum softer than the spectrum of $\dot{M}$
variations. For most of the models, we
assume the mass loss rate logarithm dispersion $\sigma\left(\ln
\dot{m}\right) = 2$ that corresponds to about a factor of $7$ change
in the mass loss rate itself and may account for the observed $\sim
1\div 2\magdot{\,}$ variations of S~Dor cycles. Models D and E
illustrate the effect of variable amplitude of mass loss variations.
Inner (hydrostatic) radius was everywhere set to $2\times 10^{12}\rm
cm$, the scale height of the $\beta$ law for velocity is ascribed the
same value.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{wind_slow}
\caption{
Simulated light curves.
}
\label{fig:sim:lc}
\end{figure*}
In figure \ref{fig:sim:lc}, we show the simulated light curves in the
Johnson V band. Corresponding power density spectra are shown in
figures \ref{fig:sim:pds1} and \ref{fig:sim:pds2} together with
modified power law and Lorentzian
approximations. Modified power law is defined as follows:
\begin{equation}\label{E:modPL}
F_{MPL} = N f^{p}\exp(-T f)+N_0
\end{equation}
This spectral law differs from the {\tt power law + white noise} model used
above by an exponential decay factor with characteristic time $T$,
close to the wind replenishment time.
If the mass loss is uncorrelated at
higher frequencies, small fast variations should have a flat spectrum
as well (see section \ref{sec:linear}).
Basic parameters of the light curves and fitting results are
given in tables \ref{tab:sim:lc} and \ref{tab:sim:pds}, respectively.
$N(t_{dyn} < t_{dif})$ in the table
is the fraction of time when dynamical timescale defined as $t_{dyn} =
R_0/v$ is smaller than the radial diffusion time scale defined as $t_{diff} = \tau
R_0/c$, where $\tau$ is the total optical depth throughout the wind. If this
inequation is satisfied, the assumption of instantaneous energy
transfer vital for the pseudo-photosphere approximation
is violated (see section \ref{sec:disc:lvar}).
Power density spectra of the simulated light curves are shown in
figures~\ref{fig:sim:pds1} and \ref{fig:sim:pds2}
\begin{table}
{\centering\footnotesize
\caption{ Light Curve Simulations }\label{tab:sim:lc}
\begin{tabular}{lcccccc}
\noalign{\smallskip}\\
Model ID & A & B & C & D & E & F \\
\noalign{\smallskip}\\
\hline
\noalign{\smallskip}\\
$\dot{M}_0$, $\Msunyr$ & $10^{-4}$ & $10^{-4}$ & $10^{-5}$ & $10^{-4}$ & $10^{-4}$ &
$ 10^{-5}$ \\
$\sigma\left(\ln \dot{m}\right)$ & 2 & 2 & 2 & 3.5 & 1 & 1 \\
$v/100\kms$ & 1 & 0.5 & 1 & 1 & 2 & 2 \\
$R_0$, $10^{12}\rm cm$ & 28 & 14 & 3.7 & 28 & 55 & 3.7 \\
No of points & 3910 & 1955 & 9774 & 195 & 1955 & 5585 \\
$M_V$ range, mag & -6.6..-10.0 & -6.7..-10.0 & -5.7..-8.2 & -6.1..-10.0 &
-6.6..-10.0 & -5.7..-7.9 \\
$\langle M_V\rangle$, mag & -8.2 & -9.0 & -6.0 & -9.0 & -9.0 & -6.0
\\
$\sigma\left( M_V\right)$, mag & 0.5 & 0.5 & 0.24 & 0.67 & 0.5 &
0.22 \\
$N(t_{dyn} < t_{diff})$, \% & 0.2 & 0 & 0 & 4 & 0.2 & 0\\
\end{tabular}}
\end{table}
\begin{table}
\centering \footnotesize
\caption{ PDS of the Simulated Light Curves }\label{tab:sim:pds}
\begin{tabular}{p{2.5cm}cccccc}
\noalign{\smallskip}\\
Model ID & A & B & C & D & E & F \\
\noalign{\smallskip}\\
\hline
\noalign{\smallskip}\\
\noalign{Lorentzian:}\\
$N$, d & $30\pm 1 $ & $0.468\pm 0.007 $ & $0.91\pm 0.09$ & $120\pm
10 $ & $42\pm 2$ & $1.2\pm 0.1 $ \\
$\gamma$, $10^{-3}$d$^{-1}$ & $1.44\pm 0.05 $ & $3.92\pm 0.19$ &
$14\pm 3 $ & $1.2\pm 0.3$ & $0.9\pm 0.1 $ & $6.48\pm 0.05$ \\
$\chi^2/DOF$ & $2527/115$ & $1304/298$ & $1299/291$ & $55/24$ & $947/27$ & $1700/333$\\
\noalign{\bigskip}\\
\hline
\noalign{\smallskip}\\
\noalign{Modified Power Law:}\\
$N$, d & $550\pm 40$ & $2\pm 4 $ & $40\pm 40$ & $284\pm 3 $ & $104\pm 1$ & $12\pm 7 $ \\
$T$, d & $361\pm 8 $ & $100\pm 28 $ & $160\pm 20 $ & $614\pm 17 $ & $8.55\pm 5$ & $103\pm 11 $ \\
$p$ & $0.41\pm 0.03$ & $0.3\pm 0.2$ & $0.46\pm 0.16$ & $0.00\pm
0.01$ & $0.00\pm 0.01$ & $0.41\pm 0.10$ \\
$N_0$, d & $17\pm 9 $ & $0.631\pm 2 $ & $1\pm 3 $ & $0.6\pm 0.2 $ &
$110\pm 10 $ & $1\pm 2 $ \\
$\chi^2/DOF$ & $1161/113$ & $1128/296$ & $1443/289$ & $57/22$ & $173/25$ & $1362/331$\\
\end{tabular}
\normalsize
\end{table}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{pdsmod1}
\caption{
PDS fitting by Lorentzian (upper panels) and modified power law (lower
panels) models for simulated light curves, models A-C.
}
\label{fig:sim:pds1}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{pdsmod2}
\caption{
PDS fitting by Lorentzian (upper panels) and modified power-law (lower
panels) models for simulated light curves, models D-F.
}
\label{fig:sim:pds2}
\end{figure*}
\subsection{Approximate Solution for Small Variations}\label{sec:linear}
For the linear pseudo-photosphere described by the equation
(\ref{E:tau}), it is easy to find the PDS for $\tau$ in the assumption
$v=const$. For small variations, $\tau$, $r(\tau)$ and the resulting
luminosity will have identical spectral shapes.
Let $R_0$ be the radius corresponding to the photospheric radius
calculated for $\dot{M}_0$:
$$
R_0=\frac{3\kappa_T \dot{M}_0}{8\pi v}
$$
If $\dot{M}(t)$ is uncorrelated noise with the mean of $\dot{M}_0$ and
dispersion $\dot{M}_0 D\dot{m} \ll \dot{M}_0^2$, the resulting PDS will
be identical to the spectrum of the equivalent $\delta$-function response:
$$
\tau_\delta= \frac{2}{3} \sqrt{D\dot{m}} \frac{1}{t^2}
$$
Here, $t>R_0/v$, because only visible matter influences the opacity.
Fourier image:
$$
\begin{array}{l}
\tilde{\tau}(f) = \frac{4\pi}{3} \sqrt{D\dot{m}} f \int_{2\pi f R_0 /
v} \frac{e^{-i\eta}}{\eta^2} d\eta = \\
\qquad{} = \frac{4\pi}{3} \sqrt{D\dot{m}}
f \left( -\frac{\cos a}{a} - \si(a)+i \left(\frac{\sin a}{a} +
\ci(a)\right)\right), \\
\end{array}
$$
where integral sine and cosine functions (si and ci) are defined as integrals from
$a=2\pi f R_0 / v$ to positive infinity. Finally, power spectrum in
the linear case behaves as:
$$
\begin{array}{l}
PDS \propto D\dot{m} f^2 \left(a^{-2} + 2 a^{-1} (\cos a \si a +
\sin a \ci a) + \ci^2 a + \si^2 a \right) \propto\\
\qquad{ } \qquad{} \propto
\left\{
\begin{array}{lc}
f & f\ll v/2\pi R_0 \\
1 & f\gg v/2\pi R_0 \\
\end{array}
\right. \\
\end{array}
$$
The resulting PDS should be peaked, similar to the (\ref{E:modPL}) law
with $p=1$. The predicted power spectrum is vastly different from the
quiet-state variability spectrum of LBVs.
\subsection{PDS in the Flaring Case}\label{sec:flaring}
The strongly-variable \textbf{D} model has a
PDS qualitatively similar to those of strongly variable objects like
AG~Car. In fact, any source exhibiting smooth light variations
during individual uncorrelated flare events is expected to have flat
spectrum at the smallest frequencies and a Brownian-shaped $\propto
f^{-2}$ noise in the opposite limit ($f\to \infty$). If one considers
a flare of duration $t_f$ (smooth in its maximum but with rapid rise
and decay), its PDS shape in the
high-frequency limit may be estimated as:
$$
|\tilde{F}(f)|^2 \propto \left|\int_{-t_f/2}^{t_f/2}l(t) e^{-i 2\pi f
t} dt\right|^2 \propto \frac{\sin^2\left( \pi f t_f\right)}{f^2}
\sim f^{-2}
$$
This approximation works for $f\gg 1/t_f$. In the opposite frequency limit,
variability is a Poissonian series of short events that implies white
noise. The spectral slope changes somewhere around $f\sim 1/t_f$, and
the exact shape of the knee in the PDS depends on the shape of a
single flare. The best known example is the Lorentzian spectrum for
exponential flares.
Qualitatively, this spectral shape is recognisable in objects like
AG~Car and Romano's star. The steep $\alpha \sim 2$ slope is probably
connected to the smooth shape of their flares. Yet the slope remains
steep in the lower-frequency domain, and
the predicted turnover in the power spectrum has a frequency a couple of
orders of magnitude higher. This may be attributed to the existing
longer-time variability trends and to correlated behavior of outbursts and S~Dor
cycles.
Though there are indications for periodicities such as the
one-year period in AG~Car, a slope holding for two orders
of magnitude requires some power-law scaling present. The
duration-amplitude relation reported by \citet{genderen_discoveries} may be the
key. Let us assume that the shape of a flare is determined by a single
parameter $t_f$ (its duration), but its amplitude is $\propto
t_f^p$. There are indications for $p\simeq 1$. A Poissonian series of
such events will be characterised by the following power spectrum:
$$
\left|\tilde{F}(f)\right|^2 \propto f^{-2} \left| \sum t_{f, i}^p
\sin(\pi f t_{f,i}) \right|^2\simeq f^{-2(1+p)} \left|\int x^p
sin(\pi x) n(x/f) dx \right|^2
$$
Here, $n(t_f) = dN/dt_f$ is distribution of flares per unit duration
time. If $p\simeq 1$, distribution of the form $n(t_f) \propto t_f^{-1}$ (uniform if
one considers $\ln t_f$ interval) may explain
the observed spectral slopes.
\section{Discussion}\label{sec:disc}
\subsection{Different Kinds of Variability?}
One principal question about LBV variability is whether
micro-variability (with temporal scales less than several months and amplitudes
$\lesssim 0\magdot{.}2$), S~Dor cycles, slow luminosity variations
and giant eruptions are produced by a
single physical mechanism. As we show in this
article, the basic properties of LBV variability at shorter timescales
(up to $\sim 1\yr$) may be reproduced by
the toy model described in the previous section. Complex shapes of PDS
at frequencies $f\sim 10^{-3}\div 10^{-2}\rm d^{-1}$ (approximately
corresponding to the duration
timescale of an average flare) are possibly connected to transition
from the wind-smoothed pseudo-photospheric noise at higher frequencies to a
very similar Brownian noise at lower frequencies, probably connected
to slow pulsations or some other internal
processes.
Pulsations studied by the hydrodynamical modeling in
\citet{fadeyev} are similar to the observed micro-variations in
periods and amplitudes. They are also a good candidate for the process
driving mass loss variations. Viewed at $\sim 1\yr$ timescales, irregular
pulsations are both fast and uncorrelated, that justifies the usage of
white noise approximation. Longer-timescale variations may appear if
some instabilities lead to secular evolution of LBV pulsations
\citep{buchler93}.
Pseudo-photosphere approach is able to reproduce some of the key
points of LBV variability such as the steep power spectra. Even if
the higher opacity of the matter is taken into account, it is short to
explain the low-frequency part of the PDSs. This is consistent with the
conclusion of \citet{koter} who find that the observed variability of
LBV stars is too strong to be explained solely by velocity and
density variations in the wind.
\subsection{Variable Bolometric Luminosity}\label{sec:disc:lvar}
Strongest deviations from this simple picture are the giant eruptions and ``supernova
impostors'' (see \citet{snimpostors} for review) characterised by significant changes in bolometric
luminosity (by two magnitudes and more, see for example
\citet{clark09}). Variable bolometric luminosities are impossible to
explain in the assumptions of the pseudo-photosphere approach.
Note however that variable bolometric luminosity may still arise from
variations in mass loss rate if the characteristic diffusion
timescale in the wind becomes large enough. Generally, the
replenishment timescale (\ref{E:tref}) is much longer than
the radiation diffusion timescale that may be roughly estimated as $t_{diff}
\sim \tau R/c$. This makes the expanding atmosphere practically
quasi-stationary from the point of view of energy transfer.
However, if the total optical depth in the wind becomes significantly larger than
the inverse wind velocity in $c$ units, $\tau > c/v \sim 10^3$, the energy
output from the pseudo-photosphere will be strongly modulated by the
envelope evolution. Adiabatic losses should be taken into account as
well in this case.
In these energy output modulation, there should
be both increases and decreases of bolometric luminosity during the
outburst. The mean level of energy release in the stellar interior is
expected to be constant at much larger (thermal and nuclear) timescales of
$\gtrsim 10^4\yr$
Bolometric luminosity should
become \emph{ smaller} during giant eruptions because of adiabatic
losses close to the photon-tiring limit \citep{photon_tiring}. During
strong flares, deviations from the mean luminosity level should hold on
timescales smaller than the survival time of an optically-thick
outburst of the mass $M$ expelled at the maximal velocity of $v_{out}$:
\begin{equation}
t_{out} \sim \frac{1}{v_{out}} \sqrt{\frac{\varkappa_T M_{out}}{4\pi}}
\sim 20\, \left(\frac{v_{out}}{100\,\kms}\right)^{-1}
\left(\frac{M_{out}}{1\,\Msun}\right)^{1/2} \yr
\end{equation}
This estimate naturally arises in a gas that remains hot. Changes in
opacity and in particular gas recombination produce a variety of
similar time scales proposed for the photosphere phase or
plateau stage durations in supernova light curves
(see for example \citet{SNlc} and references therein).
Summarizing, we conclude that LBV eruptions should be accompanied
by additional energy input inside the hydrostatic radius.
\section{Conclusions}
Variability of LBV stars is generally consistent with a single steep
power law, but the processes forming its low- and high-frequency parts
are probably different. The XX-th century light curve of P~Cyg does not
show any noticeable flares but is correlated at long timescales up to
decades. The overall power spectrum is consistent with a
$p=1.3\div1.4$ power law.
For all the flaring sources,
the slope is higher, $p\simeq 2$.
Emergence of this steep power-law spectral shape is reasonable in the
pseudo-photosphere approach. Short-timescale variations in mass loss
rate are effectively smeared and integrated over the line of sight
that leads to a Brownian noise with a break at the ``replenishment''
time of about one year. Smaller amplitude variability should produce
peaked noise with the maximum at several months.
The resulting variability lacks two important
features of the real light curves: it is not correlated at
longer timescales (slow variability
component) and conserves bolometric luminosity.
Longer-timescale correlations may arise partially from the higher
opacity of the outflowing matter (that affects the observed luminosity
changes in a complex way not implemented in our model), variable
wind velocity or some internal mechanisms.
\section*{Acknowledgements}
Our article makes use of the AAVSO International Database, hence we
would like to thank all the AAVSO observers who made the present study
possible. Author also thanks Alla Zharova for providing the V-band
light curves of the two objects in M33.
\bibliographystyle{model2-names}
|
\section{Introduction}
Elliptic flow is one of the most important observables in relativistic energy heavy ion collisions. Elliptic flow measure the azimuthal correlation of produced particles with respect to the reaction plane.
Finite elliptic flow in
a non-central collision is now regarded as a definitive signature of collective effect in relativistic energy nuclear collisions \cite{Ollitrault:1992bk,Poskanzer:1998yz}.
Qualitatively, elliptic flow is best explained in a relativistic hydrodynamical model,
rescattering of secondaries generates pressure and drives the subsequent collective motion. In a non-central collision, the reaction zone is asymmetric (almond shaped), and pressure gradient
is large in one direction and small in the other. The asymmetric pressure gradient generates the elliptic flow. As the fluid evolve and
expands, asymmetry in the reaction zone decreases and a stage arise when reaction zone become symmetric and system no longer generate elliptic flow.
Elliptic flow is an early time phenomena. It is a sensitive probe to, (i) degree of
thermalisation, (ii) transport coefficients and (iii) equation of state of the early stage of the fluid \cite{Ollitrault:1992bk,Kolb:2001qz,Hirano:2004en}.
Relativistic hydrodynamics has been successfully applied to analyse various experimental data in Au+Au collisions at RHIC energy ($\sqrt{s}$=200 GeV).
Hydrodynamical evolution of near ideal QGP fluid, with shear viscosity to entropy ratio $\eta/s\approx$0.1, initialized to central energy density $\varepsilon_i\approx$ 30 $GeV/fm^3$ at initial time $\tau_i\approx$0.6 fm/c, is consistent with experimental elliptic flow in central and mid-central collisions \cite{QGP3,Luzum:2008cw,Chaudhuri:2008sj,Chaudhuri:2008ed,Chaudhuri:2009uk}.
Small viscosity is consistent with string theoretical calculation based on
ADS/CFT correspondence \cite{Kovtun:2004de} that viscosity of a strongly interacting matter is bounded by the KSS bound
$\eta/s \geq 1/4\pi$. One wonders about the viscosity to entropy ratio ($\eta/s$) at LHC energy.
Present paradigm is that $\eta/s$ has a minimum, possibly with a cusp, around the critical temperature $T=T_c$ \cite{Csernai:2006zz}. Since, at LHC, produced fluid will be at a higher temperature than the fluid at RHIC, one expects higher value of $\eta/s$ at LHC.
Recently, ALICE collaboration measured charged particles
elliptic flow in 10-50\% Pb+Pb collisions at LHC energy, $\sqrt{s}_{NN}$=2.76 TeV \cite{Aamodt:2010pa}.
Differential elliptic flow in 10-50\% collisions at LHC is also rather similar to that at RHIC. Integrated elliptic flow on the other hand increased by $\sim$30\%.
ALICE collaboration also measured charged particles
multiplicity \cite{Collaboration:2010cz}, $p_T$ spectra \cite{Aamodt:2010jd} in Pb+Pb collisions. At LHC multiplicity increase by a factor of $\sim$2.2, though the centrality dependence is rather similar to that in $\sqrt{s}_{NN}$=200 GeV Au+Au collisions. High $p_T$ suppression is also more at LHC energy. In 0-5\% collisions, $p_T$ spectra of charged particle's are suppressed by a factor $R_{AA}\sim 0.14$ at $p_T$=6-7 GeV, which is smaller than at lower energies. In peripheral collision, suppression is modest, $R_{AA}\approx$0.6-0.7.
Increase of elliptic flow at LHC energy collisions has been predicted in
several hydrodynamical and hybrid models. In \cite{Luzum:2009sb,Hirano:2010jg,Drescher:2007uh} it was predicted that flow will increase by 20-30\%. Some models also predicted decreased elliptic flow in LHC energy \cite{Krieg:2007sx,Molnar:2007an,Chaudhuri:2008je}. For example, in \cite{Chaudhuri:2008je}, one of the present author predicted that at $\sqrt{s}_{NN}$=5 TeV Pb+Pb collisions, elliptic flow will decrease by $\sim$15\%. The prediction was based on several assumptions, which are incorrect as of now. For example, contrary to the present understanding that multiplicity increases with energy by a power law, in \cite{Chaudhuri:2008je} a logarithmic dependence was assumed. This led to much reduced multiplicity than observed experimentally. Also, in \cite{Chaudhuri:2008je}, confinement-deconfinement transition was assumed to be 1st order, contrary to the present understanding that the transition is a cross-over.
In the present paper, in a viscous hydrodynamic model, we have analysed the ALICE data for the centrality dependence of the charged particles multiplicity, $p_T$ spectra in 0-5\% collisions, centrality dependence of integrated and differential elliptic flow and extracted viscosity to entropy ratio for the fluid produced in $\sqrt{s}_{NN}$=2.76 TeV Pb+Pb collisions.
Analysis indicate that the ALICE data on the centrality dependence of charged particles multiplicity and 0-5\% $p_T$ spectra are best explained in ideal hydrodynamics. Elliptic flow data on the otherhand demand viscosity.
Small viscosity $\eta/s=0.06\pm 0.02$ is required if all the data sets are analysed simultaneously.
The paper is organised as follows: in section \ref{sec2}, we briefly explain the second order Israel-Stewarts theory of dissipative hydrodynamics.
Hydrodynamics equations are closed with an equation of state (EoS).
In section \ref{sec3} we discuss the EoS used in the present study. Section \ref{sec4} deals with the initial conditions used for hydrodynamic simulation of $\sqrt{s}_{NN}$=2.76 TeV Pb+Pb collisions. Simulation results are compared with experiments in section \ref{sec5}. Finally in section \ref{sec6} we summarise the results.
\section{Hydrodynamic model}\label{sec2}
In Israel-Stewart's second order theory \cite{IS79}, thermodynamic space is extended to include the dissipative fluxes. Relaxation equations for the dissipative fluxes
are obtained from the entropy law, $\partial_\mu s^\mu \geq 0$. Equation of motion of the fluid is then obtained by solving simultaneously the conservation equations, e.g. energy-momentum and the relaxation equations for the dissipative fluxes. In $\sqrt{s}_{NN}$=2.76 TeV Pb+Pb collisions, the central rapidity region is approximately baryon free and dissipative effect like conductivity can be neglected. We also neglect the bulk viscosity. In general bulk viscosity is much less than the shear viscosity. But in QCD, near the transition point bulk viscosity can be large \cite{Kharzeev:2007wb,Karsch:2007jc}. Effect of bulk viscosity on elliptic was studied in \cite{Song:2008si}. The effect is not large. We have also extended the code AZHYDRO-KOLKATA to include bulk viscosity. Preliminary results are reported in \cite{Chaudhuri:2011dp}. We also find that inclusion of bulk viscosity marginally modify the spectra or elliptic flow.
Equation of motion of the fluid in $\sqrt{s}_{NN}$=2.76 TeV Pb+Pb collisions is then obtained by solving,
\begin{eqnarray}
\partial_\mu T^{\mu\nu} & = & 0, \label{eq1} \\
D\pi^{\mu\nu} & = & -\frac{1}{\tau_\pi} (\pi^{\mu\nu}-2\eta \nabla^{<\mu} u^{\nu>}) \nonumber \\
&-&[u^\mu\pi^{\nu\lambda}+u^\nu\pi^{\mu\lambda}]Du_\lambda. \label{eq2}
\end{eqnarray}
Eq.\ref{eq1} is the conservation equation for the energy-momentum tensor, $T^{\mu\nu}=(\varepsilon+p)u^\mu u^\nu - pg^{\mu\nu}+\pi^{\mu\nu}$,
$\varepsilon$, $p$ and $u$ being the energy density, pressure and fluid velocity respectively. $\pi^{\mu\nu}$ is the shear stress tensor (we have neglected bulk viscosity and heat conduction). Eq.\ref{eq2} is the relaxation equation for the shear stress tensor $\pi^{\mu\nu}$.
In Eq.\ref{eq2}, $D=u^\mu \partial_\mu$ is the convective time derivative, $\nabla^{<\mu} u^{\nu>}= \frac{1}{2}(\nabla^\mu u^\nu + \nabla^\nu u^\mu)-\frac{1}{3}
(\partial . u) (g^{\mu\nu}-u^\mu u^\nu)$ is a symmetric traceless tensor. $\eta$ is the shear viscosity and $\tau_\pi$ is the relaxation time. It may be mentioned that in a conformally symmetric fluid relaxation equation can contain additional terms \cite{Song:2008si}.
Assuming boost-invariance, Eqs.\ref{eq1} and \ref{eq2} are solved in $(\tau=\sqrt{t^2-z^2},x,y,\eta_s=\frac{1}{2}\ln\frac{t+z}{t-z})$ coordinates, with the code
"`AZHYDRO-KOLKATA"', developed at the Cyclotron Centre, Kolkata.
Details of the code can be found in \cite{Chaudhuri:2008sj}.
\section{Equation of State} \label{sec3}
Hydrodynamic equations are closed with an equation of state $p=p(\varepsilon)$.
Most reliable information about the QCD equation of state is obtained from lattice simulations. It is now established that the confinement-deconfinement transition is not a thermodynamic phase transition, rather a cross-over \cite{Aoki:2006we}. Since it is cross-over, there is no unambiguous temperature where the transition take place. Inflection point of the Polyakov loop is generally quoted as the (pseudo)critical temperature for the confinement-deconfinement transition. Currently, there is consensus that
pseudo critical temperature for confinement-deconfinement phase transition is
$T_c\approx$170 MeV \cite{Aoki:2009sc,Fodor:2010zz,Borsanyi:2010cj}.
\begin{figure}[t]
\center
\resizebox{0.35\textwidth}{!}
\includegraphics{eos.eps}
}
\caption{(color online) Black and red circles are Wuppertal-Budapest lattice simulations \cite{Borsanyi:2010cj} for entropy density and pressure. Solid lines are the fit to the lattice simulation with model the EOS (Eq.\ref{eq3}). For completeness energy density in the model EOS is also shown.}
\label{F1}
\end{figure}
We have constructed EoS by complementing Wuppertal-Budapest lattice simulations \cite{Borsanyi:2010cj} with a hadron resonances gas comprising all the resonances upto mass $m_{res}\approx$ 2.5 GeV. Entropy density is parameterised as,
\begin{equation} \label{eq3}
s=0.5[(1-\tanh(x)]s_{HRG}+0.5[1+\tanh(x)]s_{lattice}
\end{equation}
\noindent where $s_{HRG}$ and $s_{lattice}$ are entropy density of the confined and deconfined phase, $x=\frac{T-T_c}{\Delta T_c}$, $\Delta T_c=0.1 T_c$.
From the parametric form of the entropy density, pressure and energy density can be obtained using the thermodynamic relations,
\begin{eqnarray} \label{eq4}
p(T)&=&\int_0^T s(T^\prime) dT^\prime \label{eq2a} \\
\varepsilon(T)&=&Ts -p.
\end{eqnarray}
As stated earlier for $s_{HRG}$ we have used hadronic resonance gas comprising all the resonances with mass $m_{res}\leq$ 2.5 GeV.
For $s_{lattice}$ we use a parametric form for the Wuppertal-Budapest simulations for entropy density,
\begin{equation}
\frac{s_{lattice}}{T^3}=[\alpha +\beta T][1+\tanh\frac{T-T_c}{\delta}],
\end{equation}
\noindent with $\alpha$=6.74, $\beta$=0.32, $T_c$=174 MeV and $\delta$=70 MeV.
In Fig.\ref{F1},
Wuppertal-Budapest simulations for the entropy density and pressure, as a function of temperature are shown. The solid lines are fit obtained to the lattice simulations for entropy density and pressure in the model EoS. The model EoS correctly reproduces lattice simulations for entropy and pressure.
\section{Initialisation of the fluid}\label{sec4}
Solution of partial differential equations (Eqs.\ref{eq1},\ref{eq2}) requires initial conditions, e.g. transverse profile of the energy density ($\varepsilon(x,y)$), fluid velocity ($v_x(x,y),v_y(x,y)$) and shear stress tensor ($\pi^{\mu\nu}(x,y)$) at the initial time $\tau_i$. One also need to specify the viscosity ($\eta$) and the relaxation time ($\tau_\pi$). A freeze-out prescription is also needed to convert the information about fluid energy density and velocity to particle spectra and compare with experiments.
In a recent publication, we have analysed the centrality dependence of charged particles multiplicity in $\sqrt{s}_{NN}$=2.76 TeV Pb+Pb collisions \cite{Chaudhuri:2011hv}. It was shown that very short thermalisation time $\tau_i$=0.2 fm or very large thermalisation time $\tau_i$=4 fm, is not consistent with the ALICE measurements. Presently, we assume that the fluid is thermalised at the time scale $\tau_i$=0.6 fm. Initial fluid velocity is assumed to be zero, $v_x(x,y)=v_y(x,y)=0$. At the initial time $\tau_i$, in an impact parameter ${\bf b}$ collisions, we distribute the initial energy density as in a Glauber model \cite{QGP3},
\begin{equation} \label{eq7}
\varepsilon({\bf b},x,y)=\varepsilon_i[(1-x) N_{part}({\bf b},x,y) + x N_{coll}({\bf b},x,y)],
\end{equation}
\noindent $\varepsilon_i$ in Eq.\ref{eq7} is the central energy density in impact parameter ${\bf b}=0$ collision. $N_{part}$ and $N_{coll}$ are the transverse profile of the average number of participants and average number binary collisions respectively, calculated in a Glauber model. $x$ is the fraction of hard scattering and is an important parameter for elliptic flow development. Initial spatial eccentricity is more if initial density is dominated by the hard scattering, so is the elliptic flow \cite{Roy:2010zd}. In Glauber model of initialisation with energy density, unless the hard scattering fraction is large,
ALICE data on the centrality dependence of the charged particles multiplicity is not reproduced. We assume 90\% hard scattering fraction.
The shear stress tensor was initialised to boost-invariant values, $\pi^{xx}=\pi^{yy}=2\eta/3\tau_i$, $\pi^{xy}$=0. For the relaxation time, we used the Boltzmann estimate $\tau_\pi=3\eta/2p$.
Finally, the freeze-out temperature was fixed at $T_F$=130 MeV. Elliptic flow is an early time phenomena and depend little on the freeze-out temperature.
In viscous hydrodynamics, particle production has contributions from the non-equilibrium part of the distribution function. We have
checked that at the freeze-out, non-equilibrium contribution to particle production is much smaller than the equilibrium contribution.
We note that the freeze-out condition does not account for the validity condition for viscous hydrodynamics, i.e. relaxation time for dissipative fluxes are much greater than the inverse of the expansion rate, $\tau_R \partial_\mu u^\mu << 1$. Recently, Dusling and Teaney \cite{Dusling:2007gi} implemented dynamical freeze-out condition.
In dynamical freeze-out, non-equilibrium effects are stronger than obtained at fixed temperature freeze-out.
\begin{figure}
\center
\resizebox{0.35\textwidth}{!}
\includegraphics{dndy_npart.eps}
}
\caption{ALICE measurements for centrality dependence of charged particles multiplicity are compared with hydrodynamical simulations for $\eta/s$=0-0.16.}
\label{F2}
\end{figure}
\begin{table}[h
\caption{\label{table1} Central energy density ($\varepsilon_i$) and
temperature ($T_i$) at the initial time $\tau_i$=0.6 fm/c, for different
values of viscosity to entropy ratio ($\eta/s$). The bracketed values are estimated central energy density and temperature in $\sqrt{s}_{NN}$=200 GeV Au+Au collisions \cite{Chaudhuri:2009uk}.
}
\begin{ruledtabular}
\begin{tabular}{|c|c|c|c|c|}\hline
$\eta/s$ & 0 & 0.08 & 0.12 & 0.16 \\ \hline
$\varepsilon_i$ & $143.0\pm 6.0$ & $126.0\pm 5.5$ & $115.0\pm 5.2$ & $103.0\pm 4.6$ \\
($\frac{GeV}{fm^3}$) & ($35.5\pm 5.0$) & ($29.1\pm3.6$) & ($25.6\pm 4.0$) & ($20.8\pm 2.7$) \\ \hline
$T_i$ & $548\pm 5$ & $531\pm 7$ & $520\pm 6$ & $504\pm 7$ \\
(MeV) & ($377\pm 14$) & ($359\pm 12$) & ($348\pm 14$) & ($331\pm 11$)
\end{tabular}\end{ruledtabular}
\end{table}
The only parameters remain to be fixed is the central energy density $\varepsilon_i$ and viscosity coefficient $\eta$. We assume that the ratio of viscosity to entropy $\eta/s$ remain a constant through out the evolution and simulate Pb+Pb collisions for four values of viscosity, (i) $\eta/s$=0 (ideal fluid) (ii) $\eta/s=1/4\pi\approx 0.08$ (KSS bound), (iii)$\eta/s$=0.12 and (iv) $\eta/s$=0.16. During the evolution, fluid cross over from QGP phase to hadronic phase. The assumption that $\eta/s$ remain unchanged during the evolution violates the present understanding that $\eta/s$ has a minimum around the critical temperature $T=T_c$ \cite{Csernai:2006zz}. Constant $\eta/s$ should be understood as space-time averaged $\eta/s$. The remaining parameter, central energy density $\varepsilon_i$ at the initial time $\tau_i$=0.6 fm is fixed to reproduce ALICE measurements for charged particles multiplicity in 0-5\% Pb+Pb collisions, $(\frac{dN_{ch}}{dy})_{ex}=1601\pm 60$ \cite{Collaboration:2010cz}. We simulate Pb+Pb collisions and compute the
negative pion multiplicity. Resonance contributions are included. Assuming that pions constitue $\approx$85\% of the charged particles, negative pion multiplicity is multiplied by the factor $2\times 1.15$ to compare with experimental charged particles multiplicity.
Irrespective of fluid viscosity, in the hydrodynamic model, experimental multiplicity can be fitted by changing the initial energy density, more viscous fluid requiring less energy density. Multiplicity is proportional to final state entropy. In viscous fluid, during evolution, entropy is generated. Entropy generation increases with viscosity and more viscous fluid require less initial energy density to reach a fixed final state entropy. In table.\ref{table1}, central energy density and temperature of the fluid required to reproduce charged particles multiplicity in 0-5\% collisions are listed. ALICE collaboration measured multiplicity quite accurately and initial energy density of QGP fluid in $\sqrt{s}_{NN}$=2.76 Pb+Pb collisions can be estimated accurately, within 5\%.
In table.\ref{table1}, the bracketed values are estimated central energy density and temperature in $\sqrt{s}_{NN}$=200 GeV Au+Au collisions \cite{Chaudhuri:2009uk}. From RHIC to LHC, collision energy increases by a factor of $\sim$14 but central energy density of the fluid increase by a much lower factor $\sim$ 4-5.
\begin{figure}
\center
\resizebox{0.35\textwidth}{!}
\includegraphics{ptdis.eps}
}
\caption{ALICE measurements for charged particles $p_T$ distribution
0-5\% Pb+Pb collisions are compared with hydrodynamic simulations for $\eta/s$=0-0.16.}
\label{F3}
\end{figure}
\section{Results}\label{sec5}
\subsection{Centrality dependence of charged particles multiplicity}
With all the model parameters fixed we can study the centrality dependence of simulated charged particles multiplicity in $\sqrt{s}_{NN}$=2.76 TeV Pb+Pb collisions. In Fig.\ref{F2} simulation results for charged particles multiplicity per participant pairs are compared with the ALICE measurements.
Even though, fluid was initialised to reproduce experimental multiplicity in 0-5\% collisions, in peripheral collisions, viscous fluid produces more particles than an ideal fluid.
The reason is understood. Viscous effects depend on the velocity gradients. Velocity gradients are more in a peripheral collisions than in a central collisions, so is the viscous effect.
To be quantitative about the fit to the data in ideal and viscous hydrodynamics, we have computed $\chi^2$ values for the fits, they are listed in table.\ref{table2}. Best fit to the data is obtained in the ideal fluid approximation. In viscous evolution $\chi^2$ values increase by a factor of 2-3.5. We can conclude that ALICE charged particles multiplicity data in $\sqrt{s}_{NN}$=2.76 TeV Pb+Pb collisions do not demand any viscosity.
\begin{figure}[t]
\center
\resizebox{0.35\textwidth}{!}
\includegraphics{integrated_v2.eps
}
\caption{ ALICE measurements for integrated elliptic flow are compared with hydrodynamic simulations for viscosity to entropy ratio $\eta/s$=0-0.16.}
\label{F4}
\end{figure}
\begin{table}[h]
\caption{\label{table2} $\chi^2/N$ values for the fit to the ALICE data for different values of viscosity to entropy ratio.}
\begin{tabular}{|c|c|c|c|c|}\hline
& \multicolumn{4}{|c|} {$\chi^2/N$} \\ \hline \
Data set & $\eta/s=0.0$ & $\eta/s$=0.08 & $\eta/s$=0.12 &$\eta/s$=0.16 \\ \hline
\hline
charged particles & 1.46 & 3.12 & 4.44 & 5.24 \\
multiplicity & & & &\\ \hline
0-5\% & 4.38 & 7.14 & 9.78 & 14.84 \\
$p_T$ spectra & & & &\\ \hline
$v_2$ integrated & 9.68 & 3.02 & 12.46 & 41.44 \\ \hline
$v_2$: 10-20\% & 7.07 &4.69 &7.68 &15.14 \\ \hline
$v_2$: 20-30\% & 23.64&8.32 &8.46 &18.97 \\ \hline
$v_2$: 30-40\% &44.60& 13.14 &9.89 &24.47 \\ \hline
$v_2$: 40-50\% & 33.62 &9.86 &4.18 &9.00 \\ \hline
\end{tabular
\end{table}
\subsection{Charged particles $p_T$ spectra in 0-5\% collision}
ALICE measurements for charged particles $p_T$ spectra in 0-5\% collisions are compared with hydrodynamical simulations in Fig.\ref{F3}.
We have initialised the fluid to reproduce charged particles multiplicity in 0-5\% collision. Such a initialisation do reproduces experimental $p_T$ spectra reasonably well. Effect of viscosity on spectra is also evident. Compared to ideal fluid, in viscous evolution, high $p_T$ production is enhanced. In table.\ref{table2}, $\chi^2$ values for the fits are listed.
Here also, we can conclude that ALICE data on charged particles $p_T$ spectra in 0-5\% collisions, do not demand any viscosity.
\begin{figure}
\center
\resizebox{0.4\textwidth}{!}
\includegraphics{v2diff.eps}
}
\caption{In four panels, ALICE measurements for charged particles
elliptic flow in 10-20\%, 20-30\%, 30-40\% and 40-50\% Pb+Pb collisions
are shown. The solid, short dashed, dashed dotted and long dashed lines are hydrodynamic model simulation for elliptic flow for fluid viscosity $\eta/s$=0, 0.08, 0.12 and 0.16 respectively.}
\label{F5}
\end{figure}
\subsection{Centrality dependence of integrated elliptic flow}
ALICE collaboration employed various methods, e.g. 2 and 4 particle cumulant, q-distribution, Lee-Yang zero method etc, to measure charged particles elliptic flow in Pb+Pb collisions \cite{Aamodt:2010pa}. Flow in 4-particle cumulant method, Lee-Yang zero method and q-distribution is consistent with each other. 2-particle cumulant method however results in $\sim$15\% higher flow. In Fig.\ref{F4} ALICE measurements for integrated elliptic flow in 4-particle cumulant method are compared with hydrodynamical simulations for elliptic flow. Ellitic flow depend sensitively on viscosity, flow reducing with increasing viscosity. Reductiuon is more in peripheral than in central collisions. As argued previously, viscous effects are more in peripheral than in central collisions. For high viscosity, $\eta/s$=0.16, in very peripheral collisions, experimental centrality dependence is not reproduced. Instead of decreasing $v_2$ increases. $\chi^2/N$ values for the fits to the data are noted in table.\ref{table2}. Unlike the centrality dependence of charged particles multiplicity or $p_T$ spectra, best fit to the ALICE data on integrated elliptic flow is obtained for $\eta/s$=0.08.
\subsection{Differential elliptic flow in 10-50\% collision}
In Fig.\ref{F5}, ALICE measurements of elliptic flow $v_2\{4\}$, in 4-particle cumulant method, in 10-20\% ,20-30\%, 30-40\% and
40-50\% collision centralities are shown. ALICE collaboration measured elliptic flow upto $p_T\approx$5 GeV. Differential $v_2$ reaches a maximum $\approx$0.2 at $p_T$=3 GeV. In Fig.\ref{F5}, we have shown the measurements only upto $p_T$=3 GeV. Hydrodynamical model are not well suited for large $p_T$. Sources other than thermal also contribute to the particle production at large $p_T$. The continuous, small dashed, dashed dot, and medium dashed lines in Fig.\ref{F5} are simulated elliptic flow for pions in hydrodynamic evolution of fluid with viscosity to entropy density ratio $\eta/s$=0.0,0.08,0.12 and 0.16 respectively.
\begin{figure}[t]
\center
\resizebox{0.3\textwidth}{!}
\includegraphics{chi2_all.eps
}
\caption{ $\chi^{2}/N$ values for the fits to the ALICE combined data set (see text). The solid line is a parabolic fit to the $\chi^2$ values. }
\label{F6}
\end{figure}
Unlike the centrality dependence of multiplicity or the $p_T$ spectra in 0-5\% collisions, differential elliptic flow is not best explained in ideal fluid evolution.
$\chi^2/N$ values for the fits, in each collision centralities, are noted in table.\ref{table2}. Experimental data only in the $p_T$ range, $0.25 \leq p_T \leq 3.25$ GeV are included in the $\chi^2$ analysis. $\chi^2/N$ values indicate that the data are better explained in viscous evolution than in ideal fluid evolution. It is also apparent that more peripheral collisions demand more viscosity. For example, in 10-20\%
collision, $\chi^2$ is minimised for the KSS bound of viscosity to entropy ratio, $\eta/s \approx$0.08. In more peripheral 30-40\% collision, data demand $\eta/s\approx$0.12. The result is interesting. In RHIC energy collisions also, it was seen that elliptic flow in peripheral collisions demand higher viscosity than in central collisions \cite{Chaudhuri:2009hj}. As discussed in detail in \cite{Chaudhuri:2009hj}, centrality dependence of viscosity can be understood in terms of continuously increasing role of hadronic phase of the fluid in peripheral collisions. The space-time averaged $\eta/s$ has contributions, both from the QGP and hadronic phase. Compared to a central collisons, in a peripehral collision, contribution of hadronic phase to the space-time averaged $\eta/s$ increases, while that of QGP decreases.
In the present analysis different $\eta/s$ are indicated by different data sets.
For example, ALICE data on multiplicity and 0-5\% $p_T$ spectra are best explained with $\eta/s\approx$0, the integrated $v_2$ data with $\eta/s\approx$0.08, the differential $v_2$ data in 10-50\% collision with $\eta/s$=0.08-0.12. To obtain $\eta/s$ which best explain the ALICE data,
$\chi^2/N$ values for the combined data set, (i) centrality dependence of multiplicity, (ii) centrality dependence of integrated elliptic flow, (iii) $p_T$ spectra in 0-5\% collisions and (iv) differential elliptic flow in 10-50\% collisions, are computed. They are shown in Fig.\ref{F6}. $\chi^2/N$ values exhibit a broad minima, indicating that the data are not very sensitive to viscosity to entropy ratio. The solid line in Fig.\ref{F6} is a parabolic fit to the $\chi^2/N$ values. From the minimum of the fit we extract, $\eta/s=0.06 \pm 0.02$. As it was at RHIC enery collisions, at LHC also, experimental data are consistent with nearly ideal fluid.
We may note that the extracted value of $\eta/s$ at LHC energy is compare well with $\eta/s$ extracted in other hydrodynamic model simulation
\cite{Schenke:2011tv,Bozek:2011wa}. However, hybrid models, i.e. hydrodynamics coupled to hadron cascade models indicate higher viscosity \cite{Song:2011qa,Hirano:2010je}.
In the present analysis we have used some specific initial conditions, e.g. initial time $\tau_i$=0.6 fm, initial zero fluid velocity, hard scattering fraction 90\%, boost-invariant shear stress tensor etc. All possible initial conditions are not explored. In \cite{Chaudhuri:2009uk}, systematic uncertainty in $\eta/s$ due to various uncertainties in initial conditions was estimated as large as 175\%. The presently extracted value $\eta/s=0.06\pm 0.02$ will be even more uncertain, if all possible initial conditions are accounted for.
\section{Summary and Conclusion}\label{sec6}
In Israel-Stewart's second order theory of hydrodynamics, we have
analysed ALICE data on the centrality dependence of charged particles multiplicity, integrated and differential elliptic flow, charged particles $p_T$ spectra in 0-5\% collisions. ALICE data for charged paricles multiplicity or 0-5\% $p_T$ spectra do not demand any viscosity, data are best explained in ideal fluid approximation. On the other hand, ALICE data on integrated or differential elliptic flow demand small viscosity. From a simulataneous analysis of all the data sets we obtain viscosity to entropy ratio $\eta/s=0.06\pm 0.02$, in $\sqrt{s}_{NN}$=2.76 TeV Pb+Pb collisions.
The extracted values of viscosity is rather similar to the value extracted at RHIC energy, even though the fluid at LHC is at higher temperature than at RHIC. It appear that both at RHIC and LHC, nearly perfect fluid is produced.
$\bf{Acknowledgement}$: Authors would like to thank Dr. Bedangadas Mohanty for providing the
ALICE data on charged particle elliptic flow for different centralities.
|
\section*{Supplementary Material for EPAPS}
\section{I. Details of tight-binding parametrization}For the five orbital, tetragonal symmetry preserving tight binding model, we adopt the parametrization of Ref.\onlinecite{Graser}. We have fitted the LDA
band structure with the band dispersion found from the tight-binding model, to determine the hopping parameters. For iron pnictides we use the tight binding parameters of Ref.\onlinecite{Graser}. For $(\mathrm{K},\mathrm{Tl})\mathrm{Fe}_{x}\mathrm{Se}_2$, we have performed calculations with two different sets of hopping parameters. These two sets respectively were derived from fitting the LDA results
for $\mathrm{K}\mathrm{Fe}_2\mathrm{Se}_2$, and $\mathrm{Tl}\mathrm{Fe}_2\mathrm{Se}_2$. In the main text we have shown the electron pockets and pairing phase diagram for the band structure corresponding to $\mathrm{K}\mathrm{Fe}_x\mathrm{Se}_2$. In this section we provide the explicit parametrization and band dispersions for both tight-binding models, and also show the electron pockets derived from $\mathrm{Tl}\mathrm{Fe}_2\mathrm{Se}_2$.
\begin{figure}[ht]
\centering
\subfigure[]{
\includegraphics[scale=0.31]{SFig1a_BDKFeSeN615.eps}
\label{fig:5a}
}
\subfigure[]{
\includegraphics[scale=0.31]{SFig1b_BDTlFeSeN615.eps}
\label{fig:5b}
}
\subfigure[]{
\includegraphics[scale=0.31]{Fig1c_FSTlFeSeN615.eps}
\label{fig:5c}
}
\label{fig:5}
\caption[]{The band dispersions of (a) $\mathrm{K}\mathrm{Fe}_2\mathrm{Se}_2$ and (b) $\mathrm{Tl}\mathrm{Fe}_2\mathrm{Se}_2$, along the high-symmetry directions of the extended Brillouin zone (one iron/ unit cell). The band dispersions have been obtained by diagonalizing the appropriate five orbital tight binding model. In panel (c) we show the Fermi surfaces of $\mathrm{Tl}\mathrm{Fe}_x\mathrm{Se}_2$ with electron doping $\delta=0.15$, which consist of only electron pockets at M points.
}
\end{figure}
\paragraph{Tight-binding model for $\mathrm{\bf K}\mathrm{\bf Fe}_{\bf 2}\mathrm{\bf Se}_{\bf 2}$:}
The following $5\times 5$ tight binding matrix $\hat{\xi}_k$ for $\mathrm{K}\mathrm{Fe}_{x}\mathrm{Se}_2$ has been used. The orbital indices 1, 2, 3, 4 and 5 respectively correspond to $d_{xz}$, $d_{yz}$, $d_{x^2-y^2}$, $d_{xy}$, and $d_{3z^2-r^2}$ orbitals.
\begin{eqnarray}
\xi^{11}_{\bf k} & = & -0.36559 -0.1095 \cos k_x -0.81736 \cos k_y + 0.83524 \cos k_x \cos k_y
+0.03114(\cos 2 k_x -\cos 2 k_y)\nonumber \\
& & -0.03464\cos 2 k_x \cos k_y -0.12572 \cos 2 k_y \cos k_x + 0.07596 \cos 2 k_x \cos 2 k_y \nonumber \\
\xi^{22}_{\bf k} & = & -0.36559 -0.81736 \cos k_x -0.1095 \cos k_y + 0.83524 \cos k_x \cos k_y
-0.03114 (\cos 2 k_x -\cos 2 k_y) \nonumber \\
& &-0.12572 \cos 2 k_x \cos k_y -0.03464 \cos 2 k_y \cos k_x + 0.07596 \cos 2 k_x \cos 2 k_y \nonumber \\
\xi^{33}_{\bf k} & = & -0.56466+ 0.65046(\cos k_x + \cos k_y) -0.39132 \cos k_x \cos k_y
-0.01074 (\cos 2 k_x + \cos 2 k_y) \nonumber \\
\xi^{44}_{\bf k} & = & -0.00096+ 0.41266 (\cos k_x + \cos k_y) + 0.26328 \cos k_x \cos k_y
-0.0705 (\cos 2 k_x + \cos 2 k_y) \nonumber \\
& & -0.08756 (\cos 2 k_x \cos k_y + \cos 2 k_y \cos k_x) + 0.01692 \cos 2 k_x \cos 2 k_y \nonumber \\
\xi^{55}_{\bf k} & = &-0.91583 -0.0854 (\cos k_x + \cos k_y) + 0.02234 (\cos 2 k_x + \cos 2 k_y) \nonumber \\
& & + 0.00708 (\cos 2 k_x \cos k_y + \cos 2 k_y \cos k_x) -0.05396 \cos 2 k_x \cos 2 k_y \nonumber \\
\xi^{12}_{\bf k} & = & -0.40644 \sin k_x \sin k_y + 0.08068 (\sin 2 k_x \sin k_y + \sin 2 k_y \sin k_x)
-0.13092 \sin 2 k_x \sin 2 k_y =\xi^{21 \ast}_{\bf k} \nonumber \\
\xi^{13}_{\bf k} & = & -0.62894i \sin k_{y} + 0.249i \sin k_{y} \cos k_{x}
-0.0412i(\sin 2 k_{y} \cos k_{x} - \cos 2 k_{x} \sin k_{y})=\xi^{31 \ast}_{\bf k} \nonumber \\
\xi^{23}_{\bf k} & = & 0.62894i \sin k_{x} -0.249i \sin k_{x} \cos k_{y}
+0.0412i(\sin 2 k_{x} \cos k_{y} - \cos 2 k_{y} \sin k_{x})=\xi^{32 \ast}_{\bf k} \nonumber \\
\xi^{14}_{\bf k} & = & 0.2757i \sin k_{x} + 0.0042i \cos k_{y} \sin k_{x} + 0.0416 i\sin 2 k_{x} \cos k_{y}=\xi^{41 \ast}_{\bf k} \nonumber \\
\xi^{24}_{\bf k} & = & 0.2757i \sin k_{y} + 0.0042i \cos k_{x} \sin k_{y} + 0.0416 i\sin 2 k_{y} \cos k_{x} =\xi^{42 \ast}_{\bf k}\nonumber \\
\xi^{15}_{\bf k}& = & -0.0965i \sin k_{y} +0.40384i \sin k_{y} \cos k_{x} +i(0.04816)
\sin 2 k_{y} \cos 2 k_{x}=\xi^{51 \ast}_{\bf k} \nonumber \\
\xi^{25}_{\bf k}& = & -0.0965 i \sin k_{x} +0.40384 i \sin k_{x} \cos k_{y} + 0.04816 i
\sin 2 k_{x} \cos 2 k_{y}=\xi^{52 \ast}_{\bf k} \nonumber \\
\xi^{34}_{\bf k}& = & -0.1918 (\sin 2 k_y \sin k_x - \sin 2 k_x \sin k_y )=\xi^{43}_{\bf k} \nonumber \\
\xi^{35}_{\bf k}& = & -0.61932(\cos k_x - \cos k_y) -0.05992 (\cos 2 k_x \cos k_y - \cos 2 k_y \cos k_x)=\xi^{53}_{\bf k} \nonumber \\
\xi^{45}_{\bf k}& = & -0.33436 \sin k_x \sin k_y -0.03064 \sin 2 k_x \sin 2 k_y =\xi^{54}_{\bf k}
\end{eqnarray}
\paragraph{$\mathrm{\bf Tl}\mathrm{\bf Fe}_{\bf 2}
\mathrm{\bf Se}_{\bf 2}$ :} The following $5\times 5$ tight binding matrix $\hat{\xi}_k$
has been deriving from fitting the LDA results of $\mathrm{Tl}\mathrm{Fe}_{2}\mathrm{Se}_2$.
\begin{eqnarray}
\xi^{11}_{\bf k} & = & -0.35956 -0.48396 \cos k_x -0.69426 \cos k_y +0.93156 \cos k_x \cos k_y
+0.20428(\cos 2 k_x -\cos 2 k_y)\nonumber \\
& & -0.23556 \cos 2 k_x \cos k_y +0.2714 \cos 2 k_y \cos k_x +0.02912 \cos 2 k_x \cos 2 k_y \nonumber \\
\xi^{22}_{\bf k} & = & -0.35956 -0.69426 \cos k_x -0.48396 \cos k_y +0.93156 \cos k_x \cos k_y
-0.20428 (\cos 2 k_x -\cos 2 k_y) \nonumber \\
& &+0.2714 \cos 2 k_x \cos k_y -0.23556 \cos 2 k_y \cos k_x +0.02912 \cos 2 k_x \cos 2 k_y \nonumber \\
\xi^{33}_{\bf k} & = & -1.11574 + 0.76802(\cos k_x + \cos k_y) -0.10188 \cos k_x \cos k_y
+0.02612 (\cos 2 k_x + \cos 2 k_y) \nonumber \\
\xi^{44}_{\bf k} & = & 0.09324 + 0.76338 (\cos k_x + \cos k_y) + 0.63348 \cos k_x \cos k_y
-0.06164 (\cos 2 k_x + \cos 2 k_y) \nonumber \\
& & -0.18652 (\cos 2 k_x \cos k_y + \cos 2 k_y \cos k_x) -0.1074 \cos 2 k_x \cos 2 k_y \nonumber \\
\xi^{55}_{\bf k} & = &-0.74545 + 0.07376 (\cos k_x + \cos k_y) -0.05124 (\cos 2 k_x + \cos 2 k_y) \nonumber \\
& & -0.12872 (\cos 2 k_x \cos k_y + \cos 2 k_y \cos k_x) +0.08408 \cos 2 k_x \cos 2 k_y \nonumber \\
\xi^{12}_{\bf k} & = & -0.97632 \sin k_x \sin k_y -0.01744 (\sin 2 k_x \sin k_y + \sin 2 k_y \sin k_x)
-0.04692 \sin 2 k_x \sin 2 k_y =\xi^{21 \ast}_{\bf k} \nonumber \\
\xi^{13}_{\bf k} & = & -0.73408i \sin k_{y} -0.08848i \sin k_{y} \cos k_{x}
+0.10508i(\sin 2 k_{y} \cos k_{x} - \cos 2 k_{x} \sin k_{y})=\xi^{31 \ast}_{\bf k} \nonumber \\
\xi^{23}_{\bf k} & = & 0.73408i \sin k_{x} +0.08848i \sin k_{x} \cos k_{y}
-0.10508i(\sin 2 k_{x} \cos k_{y} - \cos 2 k_{y} \sin k_{x})=\xi^{32 \ast}_{\bf k} \nonumber \\
\xi^{14}_{\bf k} & = & 0.46354i \sin k_{x} -0.22408i \cos k_{y} \sin k_{x} + 0.15492 i\sin 2 k_{x} \cos k_{y}=\xi^{41 \ast}_{\bf k} \nonumber \\
\xi^{24}_{\bf k} & = & 0.46354i \sin k_{y} -0.22408i \cos k_{x} \sin k_{y} + 0.15492 i\sin 2 k_{y} \cos k_{x} =\xi^{42 \ast}_{\bf k}\nonumber \\
\xi^{15}_{\bf k}& = & -0.26344i \sin k_{y} +0.24328i \sin k_{y} \cos k_{x} -0.11724 i
\sin 2 k_{y} \cos 2 k_{x}=\xi^{51 \ast}_{\bf k} \nonumber \\
\xi^{25}_{\bf k}& = & -0.26344 i \sin k_{x} +0.24328 i \sin k_{x} \cos k_{y} -0.11724 i
\sin 2 k_{x} \cos 2 k_{y}=\xi^{52 \ast}_{\bf k} \nonumber \\
\xi^{34}_{\bf k}& = & -0.20372 (\sin 2 k_y \sin k_x - \sin 2 k_x \sin k_y )=\xi^{43}_{\bf k} \nonumber \\
\xi^{35}_{\bf k}& = & -0.4617(\cos k_x - \cos k_y) +0.13144 (\cos 2 k_x \cos k_y - \cos 2 k_y \cos k_x)=\xi^{53}_{\bf k} \nonumber \\
\xi^{45}_{\bf k}& = & -0.6828 \sin k_x \sin k_y -0.1088 \sin 2 k_x \sin 2 k_y =\xi^{54}_{\bf k}
\end{eqnarray}
\section{II. Pairing amplitudes for $\mathrm{\bf{Tl}}\mathrm{\bf{Fe}}_{\bf{x}}\mathrm{\bf{Se}}_{\bf 2}$ and iron pnictides }In this section we first show the pairing phase diagram
$\mathrm{Tl}\mathrm{Fe}_x\mathrm{Se}_2$
in Fig.~\ref{fig:6a}, for electron doping $\delta=0.15$. Also shown, Fig.~\ref{fig:6b}, are the dominant pairing amplitudes projected onto the $xz$/$yz$ 3d-orbitals.
The phase diagram and the strength of the pairing amplitudes are similar to those of $\mathrm{K}\mathrm{Fe}_x\mathrm{Se}_2$, described in the main text. In Fig.~\ref{fig:7a} and Fig.~\ref{fig:7b}, we respectively show three dimensional plots of the pairing amplitudes for $\mathrm{Tl}\mathrm{Fe}_x\mathrm{Se}_2$ and iron pnictides, both with an electron doping $\delta=0.15$.
\begin{figure}[ht]
\centering
\subfigure[]{
\includegraphics[scale=0.31]{Fig2b_PhDTlFeSe.eps}
\label{fig:6a}
}
\subfigure[]{
\includegraphics[scale=0.31]{Fig4c_PAJTlFeSe.eps}
\label{fig:6b}
}
\label{fig:6}
\caption[]{In panel (a) we show the zero temperature phase diagram of $\mathrm{Tl}\mathrm{Fe}_x\mathrm{Se}_2$ for an electron doping $\delta=0.15$. The phase diagram is similar to that of $\mathrm{K}\mathrm{Fe}_x\mathrm{Se}_2$ (Fig.~\ref{fig:2a}). The regions I, II, III, and IV respectively correspond to an $\mathrm{A}_{1g}$ state with $s_{x^2y^2}$ as the dominant pairing channel, a time reversal symmetry breaking $\mathrm{A}_{1g}+i\mathrm{B}_{1g}$ state with $s_{x^2y^2}$ and $d_{x^2-y^2}$ as the dominant $\mathrm{A}_{1g}$ and $\mathrm{B}_{1g}$ pairing channels, a likewise $\mathrm{A}_{1g}+i\mathrm{B}_{1g}$ state with $s_{x^2+y^2}$ and $d_{x^2-y^2}$ as the dominant $\mathrm{A}_{1g}$ and $\mathrm{B}_{1g}$ pairing channels, and a pure $\mathrm{B}_{1g}$ state with $d_{x^2-y^2}$ pairing channel. In panel (b) we show the competing dominant pairing amplitudes $\mathrm{A}_{1g}$ $s_{x^2y^2}$, $\mathrm{A}_{1g}$ $s_{x^2+y^2}$, $\mathrm{B}_{1g}$ $s_{x^2y^2}$ for
${xz}$/$d_{yz}$ orbitals of $\mathrm{Tl}\mathrm{Fe}_x\mathrm{Se}_2$ with an electron doping $\delta=0.15$, and $\mathrm{J_2/D}=0.1$.
}
\end{figure}
\begin{figure}[ht]
\centering
\subfigure[]{
\includegraphics[scale=0.4]{T1c.eps}
\label{fig:7a}
}
\subfigure[]{
\includegraphics[scale=0.4]{A1c.eps}
\label{fig:7b}
}
\label{fig:7}
\caption[]{The pairing amplitudes for $xz$/$yz$ orbitals of (a) $\mathrm{Tl}\mathrm{Fe}_x\mathrm{Se}_2$ and (b) iron pnictides with electron doping $\delta=0.15$. Compared to $\mathrm{Tl}\mathrm{Fe}_x\mathrm{Se}_2$, iron pnictides have stronger $\mathrm{A}_{1g}$ $s_{x^2+y^2}$ pairing.
}
\end{figure}
\section{III. Pairing phase diagrams of a two orbital model with and without doping dependent band renormalization}
To demonstrate the robustness of our pairing phase diagram against the doping dependent band renormalization effects, we consider a two orbital model involving only $xz$ and $yz$ orbitals. For simplicity we consider hole doping; the electron doping case can be treated in a similar manner after performing a particle-hole transformation. The Hamiltonian of interest is
\begin{equation}
H=-\sum_{i<j,\alpha,\beta,\sigma} t_{ij}^{\alpha \beta}c^{\dagger}_{i\alpha \sigma}c_{j\beta \sigma}+h.c. -\mu \sum_{i,\alpha,\sigma}c^{\dagger}_{i\alpha \sigma}c_{i\alpha \sigma}+\sum_{i<j,\alpha,\beta} J_{ij}^{\alpha \beta} \left(\vec{S}_{i\alpha}\cdot \vec{S}_{j\beta}-\frac{1}{4}n_{i\alpha} n_{j\beta}\right)
\end{equation}
with the double occupancy prohibiting constraint $\sum_{\sigma}c^{\dagger}_{i\alpha \sigma}c_{i\alpha \sigma}\leq1$ for each orbital.
The constraint is imposed by introducing slave boson operator $b_i$ and fermionic spinon operator $f_i$ for each orbital \cite{Kotliar} and the Hamiltonian is transformed to
\begin{eqnarray}
H &=& -\sum_{i<j,\alpha,\beta,\sigma} t_{ij}^{\alpha \beta}f^{\dagger}_{i\alpha \sigma}b_{i\alpha}b_{j\beta}^{\dagger}f_{j\beta \sigma}+h.c. -\mu \sum_{i,\alpha,\sigma}f^{\dagger}_{i\alpha \sigma}f_{i\beta \sigma}-\sum_{i<j,\alpha,\beta} \frac{J_{ij}^{\alpha \beta}}{2}B_{ij,\alpha \beta}^{\dagger}B_{ij,\alpha \beta} \nonumber\\
& & +\sum_{i,\alpha} \lambda_{i\alpha}\left(\sum_{\sigma}f^{\dagger}_{i\alpha \sigma}f_{i\alpha \sigma}+b_{i\alpha}^{\dagger}b_{i\alpha}-1\right)
\end{eqnarray}
where $B_{ij,\alpha \beta}^{\dagger}=(f^{\dagger}_{i\alpha \uparrow}f^{\dagger}_{j\beta \downarrow}-f^{\dagger}_{i\alpha \downarrow}f^{\dagger}_{j\beta\uparrow})$ is the spin-singlet pairing operator for the fermionic spinons, and $\lambda_{i,\alpha}$'s are Lagnrage multipliers which enforce the occupancy constraints. At zero temperature the slave bosons are bose condensed and the bose operators have finite expectation values. In the tetragonal symmetry breaking particle-hole channel order, we have $\langle b_{i,1}\rangle= \langle b_{i,2}\rangle=\sqrt{\delta/2}$, where $\delta$ is the hole doping and $\langle \lambda_{i,1}\rangle=\langle \lambda_{i,2}\rangle=\lambda$, which is absorbed into the chemical potential. The expectation value of the bose operators renormalizes the kinetic energy term by a factor of $\delta/2$. As in the main text we choose the interactions to be diagonal in the orbital space. After mean-field decoupling, the free energy density
\begin{eqnarray}
f=\frac{J_1}{2}\sum_{\alpha}(|\Delta_{x,\alpha \alpha}|^2+|\Delta_{y,\alpha \alpha}|^2)+\frac{J_2}{2}\sum_{\alpha}(|\Delta_{x+y, \alpha \alpha}|^2+|\Delta_{x-y,\alpha \alpha}|^2)-\int\frac{d^2k}{4\pi^2}(\mathcal{E}_{\mathbf{k}+}+\mathcal{E}_{\mathbf{k}-}-\frac{\delta}{2}E_{\mathbf{k}+}-\frac{\delta}{2}E_{\mathbf{k}-}+2\mu)
\end{eqnarray}
is minimized with respect to pairing amplitudes with the constraint $n_1=n_2=1-\frac{\delta}{2}$.
The quasiparticle dispersions $\mathcal{E}_{\mathbf{k}\pm}$ in the paired state are calculated from the $4\times4$ Nambu matrix
\begin{equation}
\hat{h}_{\mathbf{k}}=\left[\begin{array}{cc}\frac{\delta\hat{\xi}_{\mathbf{k}}}{2}-\mu \mathbb{1}_{2\times2}
& \mathbf{\Delta}_{\mathbf{k}}\\\mathbf{\Delta}^{\ast}_{\mathbf{k}}
& -\frac{\delta \hat{\xi}_{\mathbf{k}}}{2}+\mu \mathbb{1}_{2\times2}\end{array}\right]
\end{equation}
where
\begin{eqnarray}
\tilde{\Delta}_{\mathbf{k},\alpha \alpha}&=&J_1\left(\Delta_{x,\alpha \alpha}\cos(k_x)+\Delta_{y,\alpha \alpha}\cos(k_y)\right)+J_2\left(\Delta_{x+y,\alpha \alpha}\cos(k_x+k_y)+\Delta_{x-y,\alpha \alpha}\cos(k_x-k_y)\right)
\end{eqnarray}
For the band structure we choose the two orbital tight-binding model of Ref. \onlinecite{Raghu}. The $2\times2$ tight-binding matrix $\hat{\xi}_{\mathbf{k}}=\hat{\xi}_{\mathbf{k}}=\xi_{{\mathbf{k}}+}\mathbb{1}_{2\times2}+\xi_{{\mathbf{k}}-}\tau_{z}
+\xi_{{\mathbf{k}}xy}\tau_{x}$, where Pauli matrices $\tau_i$)
operate on the orbital indices, and $\xi_{{\mathbf{k}}+}
=-(t_1+t_2)(\cos k_x+\cos k_y)-4t_3\cos k_x \cos k_y$,
$\xi_{{\mathbf{k}}-}=-(t_1-t_2)(\cos k_x-\cos k_y)$,
$\xi_{{\mathbf{k}}xy}=-4t_4\sin k_x \sin k_y$ are respectively $\mathrm{A}_{1g}$, $\mathrm{B}_{1g}$ and $\mathrm{B}_{2g}$ functions. The band dispersion relations $E_{{\mathbf{k}}\pm}
=\xi_{{\mathbf{k}}+}
\pm \sqrt{\xi_{{\mathbf{k}}-}^2+\xi_{{\mathbf{k}}xy}^2}$, give rise to two electron pockets
at ${\mathbf{k}}=(\pi,0)$ and $(0,\pi)$, and two hole pockets
at ${\mathbf{k}}=(0,0)$ and
$(\pi,\pi)$. The following values of the hopping parameters, $t_1=-t$, $t_2=1.3t$, $t_3=t_4=-0.85t$ were obtained in Ref.~\onlinecite{Raghu}, by a fitting of the LDA bands. The quasiparticle dispersions in the paired state are given by
\begin{eqnarray}
\mathcal{E}_{\mathbf{k}\pm}&=&\Bigg[\left(\left(\frac{\delta \xi_{\mathbf{k}+}}{2}-\mu \right)^{2}+\frac{\delta^2}{4}\left(\xi_{\mathbf{k}-}^{2}+\xi_{\mathbf{k}xy}^{2}\right)+\frac{|\tilde{\Delta}_{\mathbf{k},11}|^2}{2}+\frac{|\tilde{\Delta}_{\mathbf{k},22}|^2}{2}\right)\pm \Bigg\{\bigg(\delta \xi_{\mathbf{k}-}\left(\frac{\delta \xi_{\mathbf{k}+}}{2}-\mu \right)\nonumber \\&&+\frac{|\tilde{\Delta}_{\mathbf{k},11}|^2}{2}-\frac{|\tilde{\Delta}_{\mathbf{k},22}|^2}{2}\bigg)^2+\delta^2\xi_{\mathbf{k}xy}^2\left(\left(\frac{\delta \xi_{\mathbf{k}+}}{2}-\mu \right)^2+\frac{1}{4}|\tilde{\Delta}_{\mathbf{k},11}-\tilde{\Delta}_{\mathbf{k},22}|^2\right)
\Bigg\}^{\frac{1}{2}}\Bigg]^{\frac{1}{2}}
\end{eqnarray}
\begin{figure}[t!]
\centering
\subfigure[]{
\includegraphics[scale=0.3]{SFig3a_PhD2bFeAsRenN186.eps}
\label{fig:8a}
}
\subfigure[]{
\includegraphics[scale=0.3]{SFig3b_PhD2bFeAsN214.eps}
\label{fig:8b}
}
\label{fig:8}
\caption[]{Qualitatively similar zero temperature phase diagrams of a two orbital $\mathrm{t}-\mathrm{J}_1-\mathrm{J}_2$ model obtained (a) by considering a doping dependent renormalization of the band dispersions due to the local constraint of zero double occupancies and (b) by ignoring the doping dependent band renormalization. In (a) the effective hopping parameters $\bar{t}_i=t_i\frac{\delta}{2}$. The regions I, II, and III respectively correspond to an $\mathrm{A}_{1g}$ state with $s_{x^2y^2}$ as the dominant pairing channel, a time reversal symmetry breaking $\mathrm{A}_{1g}+i\mathrm{B}_{1g}$ state with $s_{x^2y^2}$ and $d_{x^2-y^2}$ as the dominant $\mathrm{A}_{1g}$ and $\mathrm{B}_{1g}$ pairing channels, and a likewise $\mathrm{A}_{1g}+i\mathrm{B}_{1g}$ state with $s_{x^2+y^2}$ and $d_{x^2-y^2}$ as the dominant $\mathrm{A}_{1g}$ and $\mathrm{B}_{1g}$ pairing channels.
}
\end{figure}
After explicitly accounting for the occupancy constraints and associated band renormalization we obtain the pairing phase diagram shown in Fig.~\ref{fig:8a}. A qualitatively similar phase diagram for this model, obtained by ignoring the band renormalization effects, is shown in Fig.~\ref{fig:8b}.
\section{IV. Orbital character of the Fermi surface and pairing amplitudes for $\mathbf{xy}$ orbital}
In this section we first show the orbital weights on the electron pockets for both $\mathrm{K}\mathrm{Fe}_x\mathrm{Se}_2$ and pnictides, for
electron doping $\delta=0.15$ respectively in Fig.~\ref{fig:9a} and Fig.~\ref{fig:9b}. For pnictides, the $xz$/$yz$ orbitals have the dominant orbital weights on the two hole pockets (not shown), and also contribute significantly on the electron pockets. Whereas for $\mathrm{K}\mathrm{Fe}_x\mathrm{Se}_2$ (and also for $\mathrm{Tl}\mathrm{Fe}_x\mathrm{Se}_2$), there are no hole pockets, and the contribution from $xy$ orbital is considerably enhanced on the electron pockets.
This indicates that the $xy$ orbital plays a more important role in building the electron pocket of $\mathrm{(K,Tl)}\mathrm{Fe}_x\mathrm{Se}_2$ than in pnictides.
\begin{figure}[ht]
\centering
\subfigure[]{
\includegraphics[scale=0.3]{OWKFeSe.eps}
\label{fig:9a}
}
\subfigure[]{
\includegraphics[scale=0.3]{OWFeAsE.eps}
\label{fig:9b}
}
\label{fig:9}
\caption[]{The orbital weights on the electron pocket centered at $(\pi,0)$ of (a) $\mathrm{K}\mathrm{Fe}_x\mathrm{Se}_2$ and (b) iron pnictides for electron doping $\delta=0.15$. $\theta$ is the winding angle of the pocket with respect to its center. The weights on the electron pocket centered at $(0,\pi)$ can be obtained by interchanging the $xz$ and $yz$ components and shifting $\theta$ by $\pi/2$.
}
\end{figure}
The enhanced $xy$ orbital character on Fermi surface of $\mathrm{K}\mathrm{Fe}_x\mathrm{Se}_2$ affects the electron pairing. To see this we show the pairing amplitudes at $J_2/D=0.1$ of $xy$ orbital for both $\mathrm{K}\mathrm{Fe}_x\mathrm{Se}_2$ and pnictides at electron doping $\delta=0.15$ in Fig.~\ref{fig:10a} and Fig.~\ref{fig:10b}, respectively.
Comparing with Fig.~\ref{fig:4a} and Fig.~\ref{fig:4b} we see that for pnictides, the contribution to the three competing dominant pairings from the $xz$/$yz$ orbital is comparable with the one from $xy$ orbital. But for $\mathrm{K}\mathrm{Fe}_x\mathrm{Se}_2$, the contribution from $xy$ orbital can be stronger. This is especially true in the regime $J_1/D\lesssim 0.07$, where only $\mathrm{A}_{1g}$ pairing is present.
Overall, however, the pairing amplitudes for the $xy$ orbital are similar between
$\mathrm{(K,Tl)}\mathrm{Fe}_x\mathrm{Se}_2$ and iron pnictides. This is seen
in the three-dimensional plots given in Figs.~\ref{fig:11a} and \ref{fig:11b}, which
are similar to their counterparts for the $xz$/$yz$ orbitals (Figs.~\ref{fig:3} and \ref{fig:7b}).
\begin{figure}[ht]
\centering
\subfigure[]{
\includegraphics[scale=0.3]{PAxyKFeSe.eps}
\label{fig:10a}
}
\subfigure[]{
\includegraphics[scale=0.3]{PAxyFeAs.eps}
\label{fig:10b}
}
\label{fig:10}
\caption[]{The competing dominant pairing amplitudes $\mathrm{A}_{1g}$ $\mathrm{s}_{x^2y^2}$, $\mathrm{A}_{1g}$ $\mathrm{s}_{x^2+y^2}$, $\mathrm{B}_{1g}$ $\mathrm{d}_{x^2-y^2}$ at $J_2/D=0.1$ of $d_{xy}$ orbital for (a) $\mathrm{K}\mathrm{Fe}_x\mathrm{Se}_2$ and (b) iron pnictides for electron doping $\delta=0.15$.
}
\end{figure}
\begin{figure}[ht]
\centering
\subfigure[]{
\includegraphics[scale=0.4]{K2c.eps}
\label{fig:11a}
}
\subfigure[]{
\includegraphics[scale=0.4]{A2c.eps}
\label{fig:11b}
}
\label{fig:11}
\caption[]{The pairing amplitudes for the $xy$ orbital of (a) $\mathrm{K}\mathrm{Fe}_x\mathrm{Se}_2$ and (b) iron pnictides, both for electron doping $\delta=0.15$.
}
\end{figure}
\end{document} |
\section{Introduction}
In the companion paper \citep{Dimant:Magnetosphere2011_Budget},
we have developed a theoretical description of
magnetosphere-ionosphere (MI) coupling through electrostatic
plasma turbulence in the lower ionosphere in the region where
field-aligned magnetospheric currents close and dissipate
energy. In global MHD computer codes intended for predictive
modeling of space weather, the entire ionosphere plays the role
of the inner boundary condition. The ionospheric conductances
employed in these codes are usually based on simple laminar
models of ionospheric plasma modified by the effects of
precipitating particles. These global magnetospheric MHD codes
with normal conductances often overestimate the cross-polar cap
potential during magnetic (sub)storms up to a factor of two
\citep[e.g.,][]{winglee:1997,raeder:1998,raeder.monograph:2001,
Siscoe:Hill2002,Ober:Testing2003,
Merkin:Global2005,Merkin:Anomalous2005,guild:2008a,
Merkin:Predicting2007,wang.h:2008}.
At the same time, strong DC electric field penetrates to the
\textit{E} region, roughly between 90 and 130~km of altitude,
and drives plasma instabilities. The \textit{E}-region
instabilities create turbulence that consists of density
perturbations coupled to electric field modulations. This
turbulence causes anomalous conductivity which could account
for the discrepancy.
Anomalous conductivity manifests itself in two ways. One is
associated with average anomalous heating by turbulent electric
fields, and the other is due to a net non-linear current (NC)
formed by plasma density turbulence. While the former effect
has been considered before in detail (see references below),
the latter has only been discussed but never quantitatively
investigated with application to ionospheric conductivity. This
is the main objective of this paper. In the two following
paragraphs, we outline each of these effects.
Small-scale fluctuations generated by the \textit{E}-region
instabilities can cause enormous anomalous electron heating
(AEH)
\citep{Schlegel:81a,Providakes:88,Stauning:89,St.-Maurice:90a,Williams:1992,
Foster:Simultaneous00,Bahcivan:Plasma2007}, largely because the
turbulent electrostatic field,
$\delta\vec{E}=-\nabla\delta\Phi$, has a small component
parallel to the geomagnetic field $\vec{B}_0$
\citep{StMaLaher:85,Providakes:88,DimantMilikh:JGR03,MilikhDimant:JGR03,
Bahcivan:Parallel2006}.
Anomalous electron heating directly affects the
temperature-dependent electron-neutral collision frequency and,
hence, the electron part of the Pedersen conductivity. This
part, however, is usually small compared to the dominant
electron Hall and ion Pedersen conductivities. At the same
time, AEH causes a gradual elevation of the mean plasma density
within the anomalously heated regions through the thermal
reduction of the plasma recombination rate
\citep{Gurevich:78,St.-Maurice:90b,DimantMilikh:JGR03,MilikhGoncharenko:Anom2006}.
The AEH-induced plasma density elevations increase all
conductivities in proportion by as much as a factor of two.
This mechanism requires tens of seconds or even minutes because
of the slow development of the ionization-recombination
equilibrium. If $\vec{E}_0$ changes faster than the
characteristic recombination timescale then its time-averaged
effect on density will be smoothed over field variations.
Also, low-frequency turbulence in the compressible ionospheric
plasma can directly modify local ionospheric conductivities via
a wave-induced non-linear current (NC) associated with plasma
density irregularities
\citep{RogisterJamin:1975,Oppenheim:Evidence97,Buchert:Effect2006}.
The physical nature of NC has been explained in the companion
paper \citep{Dimant:Magnetosphere2011_Budget}. While the rms
turbulent field $\langle\delta\vec{E}^{2}\rangle^{1/2}$ is
comparable to $E_{0}$ (the angular brackets denote spatial and
temporal averaging), the NC is proportional to the density
perturbations that, in saturated FB turbulence, may reach tens
percent at most. As a result, the total NC, considered as a
plasma response to the external electric field $\vec{E}_{0}$,
amounts to only a fraction of the regular electrojet Hall
current. However, in most of the electrojet the NC is directed
largely parallel to $\vec{E}_{0}$, so that it will increase
significantly the smaller Pedersen conductivity. This is
critically important because it is the Pedersen conductivity
that allows the field-aligned currents to close and dissipate
energy. The combined effect of the NC and AEH makes the
ionosphere less resistive. These anomalous conductances may, at
least partially, account for systematic overestimates of the
total cross-polar cap potential in global MHD models that
employ laminar conductivities.
In this paper, using expressions for the NC obtained in
\citet{Dimant:Magnetosphere2011_Budget}, we quantify a feedback
of developed \emph{E}-region turbulence on the global behavior
of the magnetosphere by calculating the corresponding
modifications of local conductivities. In Sect.~\ref{Anomalous
conductivity}, we do this in general terms of given spectra of
density irregularities. In Sect.~\ref{Heuristic Model of
Turbulence}, invoking the appendix, we employ a heuristic model
of non-linearly saturated turbulence \citep{DimantMilikh:JGR03}
which allows us to easily estimate the turbulent
conductivities. The anomalous heating-induced effect manifests
itself via the mean density variations so that it contributes
proportionally into both the laminar and turbulent
conductivities. In Sect.~\ref{Turbulent Frictional Heating in
Global Modeling}, we calculate the turbulent frictional heating
sources for possible inclusion into global
ionosphere-thermosphere models. In Sect.~\ref{Discussion}, we
demonstrate that all anomalous effects combined may nearly
double the undisturbed Pedersen conductance and hence they
should be taken into account when applying inner boundary
conditions in global MHD codes employed for space weather
modeling. In the appendix, using a novel and compact formalism,
we derive the general fluid-model dispersion relation for waves
in arbitrarily magnetized linearly unstable plasma.
\section{Turbulent Ionospheric Conductivity:
\allowbreak Quasilinear Approximation\label{Anomalous
conductivity}}
In this section, we calculate
the turbulent ionospheric conductivity associated with the
\textit{E}-region non-linear current, using the general
quasilinear expressions for the NC obtained in the companion
paper \citep{Dimant:Magnetosphere2011_Budget}. This quasilinear
approach is justified by the relatively small plasma density
perturbations, even though electrostatic field modulations can
be comparable to the driving electric field
\citep{DimantMilikh:JGR03,Oppenheim:Fully2011}.
The turbulent conductivity tensor can be constructed similar to
the laminar conductivity tensor,
\begin{equation}
\overleftrightarrow{\sigma^{\mathrm{L}}}\equiv\left[
\begin{array}
[c]{ccc
\sigma_{\mathrm{P}}^{\mathrm{L}} & \sigma_{\mathrm{H}}^{\mathrm{L}} & 0\\
-\sigma_{\mathrm{H}}^{\mathrm{L}} & \sigma_{\mathrm{P}}^{\mathrm{L}} & 0\\
0 & 0 & \sigma_{\parallel}^{0
\end{array}
\right] , \label{sigma_tensor
\end{equation}
where the general parallel, Pedersen, and Hall conductivities in a Cartesian
system $x_{1,2,3}$ with the axis $\hat{x}_{3}$ directed along $\vec{B}$ are
given by \citep[e.g.,][]{Kelley:Ionosphere2009}
\begin{subequations}
\label{sigma_II,P,H
\begin{align}
\sigma_{\parallel}^{\mathrm{L}} & \equiv\frac{\vec{j}_{L\parallel}\cdot\vec
{E}_{0\parallel}}{E_{0\parallel}^{2}}=\frac{(\kappa_{e}+\kappa_{i})ne
{B},\label{sigma_II}\\
\sigma_{\mathrm{P}}^{\mathrm{L}} & \equiv\frac{\vec{j}_{L\perp}\cdot\vec{E}_{0\perp}
}{E_{\perp}^{2}}=\frac{(\kappa_{e}+\kappa_{i})(1+\kappa_{i}\kappa_{e
)ne}{(1+\kappa_{e}^{2})(1+\kappa_{i}^{2})B},\label{sigma_P}\\
\sigma_{\mathrm{H}}^{\mathrm{L}} & \equiv\frac{\vec{j}_{L\perp}\cdot(\vec{E
_{0}\times\hat{b})}{E_{0\perp}^{2}}=-\ \frac{\left( \kappa_{e}^{2}-\kappa
_{i}^{2}\right) ne}{(1+\kappa_{e}^{2})(1+\kappa_{i}^{2})B}, \label{sigma_H
\end{align}
\end{subequations}
respectively. Here $n$ is the density of the quasi-neutral plasma;
$\vec{j}_{L\perp,\parallel}$
are the perpendicular and parallel to $\vec{B}_0$ components of the laminar
current density $\vec{j}_{L}$; $\kappa_s\equiv \Omega_s/\nu_{s}$ are the
magnetization parameters for particles of the $s$-kind ($s=e,i$) with
$\Omega_s=eB/m_s$ and $\nu_{sn}$ being the corresponding gyro-frequency and
mean collision frequency with neutrals; $e$ is the (positive) elementary charge;
$m_s$ are the particle-$s$ masses. These conductivities neglect Coulomb collisions,
as compared to electron and ion collisions with neutrals, so that they are
largely applicable to altitudes below the \textit{F}-region ionosphere.
Similarly to Eq.~(\ref{sigma_tensor}), the turbulent
conductivity tensor determined by the non-linear current,
$\vec{j}^{\mathrm{NC}}\equiv\overleftrightarrow{\sigma^{\mathrm{NC}}
\cdot\vec{E}_{0}$, can be expressed as
\begin{equation}
\overleftrightarrow{\sigma^{\mathrm{NC}}}\equiv\left[
\begin{array}
[c]{ccc
\sigma_{\mathrm{P}}^{\mathrm{NC}} & \sigma_{\mathrm{H}}^{\mathrm{NC}} & 0\\
-\sigma_{\mathrm{H}}^{\mathrm{NC}} & \sigma_{\mathrm{P}}^{\mathrm{NC}} & 0\\
0 & 0 & \sigma_{\parallel}^{\mathrm{NC}
\end{array}
\right]\!. \label{sigma_NC_tensor_3D
\end{equation}
We can find $\overleftrightarrow{\sigma^{\mathrm{NC}}}$ using
quasilinear Eq.~(42) from
\citet{Dimant:Magnetosphere2011_Budget} obtained for
arbitrarily magnetized particles,
\begin{align}
& \vec{j}^{\mathrm{NC}}=\frac{en_{0}}{\kappa_{i}\kappa_{e}}\nonumber\sum_{\vec{k},
\omega\neq0}(\vec{k}\cdot\vec{U}_{0})\left\vert \frac{\delta n_{\vec{k},\omega}}
{n_{0}}\right\vert^{2} \\
& \times\frac{(1+\kappa_{e}^{2})\left(
1+\kappa_{i}^{2}\right) \vec{k}_{\parallel}+\left(1+\kappa_{i}\kappa
_{e}\right) \vec{k}_{\perp}-\left( \kappa_{e}-\kappa_{i}\right) (\vec
{k}\times\hat{b})}{(1+\psi_{\vec{k}
})k_{\perp}^{2}}.\label{j_NL
\end{align}
This yields
\begin{subequations}
\label{sigma_NC_II,P,H
\begin{align}
\sigma_{\parallel}^{\mathrm{NC}} & =\vec{j}_{\parallel}^{\mathrm{NC}}\cdot
\vec{E}_{0\parallel}/E_{0\parallel}^{2}=
\frac{en_{0}(1+\kappa_{e}^{2})\left( 1+\kappa_{i}^{2}\right) }
{\kappa_{i}\kappa_{e}}\nonumber\\
& \times \sum_{\vec{k},\omega\neq0}\frac{(\vec{k}_{\parallel
}\cdot\vec{E}_{0\parallel})(\vec{k}\cdot\vec{U}_{0})}{(1+\psi_{\vec{k
})k_{\perp}^{2}E_{0\parallel}^{2}}\left\vert \frac{\delta n_{\vec{k},\omega}
}{n_{0}}\right\vert ^{2},\label{sigma_NC_II}\\
\sigma_{\mathrm{P}}^{\mathrm{NC}} & =\vec{j}_{\perp}^{\mathrm{NC}}\cdot\vec
{E}_{0\perp}/E_{0\perp}^{2}
=\frac{en_{0}}{\kappa_{i}\kappa_{e}}\sum_{\vec{k},\omega\neq0}
(\vec{k}\cdot\vec{U}_{0})\left\vert \frac{\delta n_{\vec{k},\omega}}{n_{0}
}\right\vert ^{2}\nonumber\\
&\times\frac{\left(\kappa_{e}-\kappa_{i}\right)
\vec{k}\cdot(\vec{E}_{0} \times\hat{b})+\left(
1+\kappa_{i}\kappa_{e}\right) \vec{k}_{\perp}\cdot
\vec{E}_{0}}{(1+\psi_{\vec{k}})k_{\perp}
^{2}E_{0\perp}^{2}},\label{sigma_NC_P}\\
\sigma_{\mathrm{H}}^{\mathrm{NC}} & =\vec{j}_{\perp}^{\mathrm{NC}}\cdot\left(
\vec{E}_{0}\times\hat{b}\right) /E_{0\perp}^{2}
=\frac{en_{0}}{\kappa_{i}\kappa_{e}}\sum_{\vec{k},\omega\neq0}
(\vec{k}\cdot\vec{U}_{0})\left\vert \frac{\delta n_{\vec{k},\omega}}{n_{0}
}\right\vert ^{2}\nonumber\\
&\times\frac{\left( 1+\kappa_{i}\kappa_{e}\right) \vec{k}\cdot(\vec{E}_{0}
\times\hat{b})-\left( \kappa_{e}-\kappa_{i}\right) \vec{k}_{\perp}\cdot
\vec{E}_{0}}{(1+\psi_{\vec{k}})k_{\perp}
^{2}E_{0\perp}^{2}},\label{sigma_NC_H}
\end{align}
\end{subequations}
where, according to Eq.~(24) from \citet{Dimant:Magnetosphere2011_Budget},
\begin{align}
\vec{U}_{0\parallel} & =-\ \frac{\left( \kappa_{e}+\kappa_{i}\right)
\vec{E}_{0\parallel}}{B},\nonumber\\
\vec{U}_{0\perp} & =\frac{\left( \kappa_{e}+\kappa_{i}\right) [\left(
\kappa_{e}-\kappa_{i}\right) (\vec{E}_{0}\times\hat{b})-\left( 1+\kappa
_{i}\kappa_{e}\right) \vec{E}_{0\perp}]}{\left( 1+\kappa_{e}^{2}\right)
\left( 1+\kappa_{i}^{2}\right) B}\label{k.U_0
\end{align}
are the parallel and perpendicular to $\vec{B}_0$ components of
the laminar relative velocity between the undisturbed electron
and ion streams; $\vec{U}_0\equiv\vec{V}_{e0}-\vec{V}_{i0}$. In
the quasilinear approach, the total conductivity,
$\overleftrightarrow {\sigma^{\mathrm{tot}}}$, is determined by
merely adding the laminar and turbulent conductivities,
\begin{equation}
\overleftrightarrow{\sigma^{\mathrm{tot}}}=\overleftrightarrow{\sigma^{\mathrm{L}
}+\overleftrightarrow{\sigma^{\mathrm{NC}}}
\label{sigma_tot
\end{equation}
where each of them is proportional to the same plasma density
$n_0$. Density increases caused by the AEH-induced reduction of
the recombination rate are automatically included if one allows
for the corresponding temperature modifications of
$n_0(T_{e})$.
Global MHD codes employed for magnetosphere modeling assume
equipotential magnetic field lines,
$\vec{E}_{0\parallel}=0$,
across the ionosphere due to high
electron mobility along $\vec{B}_{0}$. In the \textit{E/D}
regions, this approximation is valid above the $80$\,km
altitude where $\kappa_{e}\gg1,\kappa_{i}$, i.e., almost in
the entire electrojet. Hence, only conductivities perpendicular
to $\vec{B}_{0}$ matter, while the parallel currents can be
determined using the continuity of the total current density,
$\nabla_{\parallel}\cdot\vec{j}_{\parallel}=-\nabla_{\perp
}\cdot\vec{j}_{\perp}$. In this approach, the high parallel
electron conductivity is not invoked unless one needs to
estimate the tiny parallel field, $E_{0\|}$.
Integration of the quasi-neutral current conservation equation,
$\nabla\cdot\vec{j}^{\mathrm{tot}}\equiv-\nabla\cdot(\overleftrightarrow
{\sigma^{\mathrm{tot}}}\cdot\nabla\Phi_{0})=0$, along nearly
vertical equipotential magnetic field lines yields a 2-D
second-order differential relation for the potential,
$\nabla_{\perp}\left(\overleftrightarrow
{\Sigma^{\mathrm{tot}}}\nabla_{\perp}\Phi_{0}\right)=j_{\|}$,
where $\overleftrightarrow
{\Sigma^{\mathrm{tot}}}=\overleftrightarrow{\Sigma^{\mathrm{L}}}+\overleftrightarrow
{\Sigma^{\mathrm{NC}}}$ is the total height-integrated
ionospheric conductance tensor and $j_{\|}$ is the parallel
current density on top of the conducting ionosphere. This
height-integrated Ohm's law serves as an approximate inner
boundary condition for the global MHD codes
\citep[e.g.,][]{Merkin:Global2005}.
With neglect of $\vec{E}_{0\parallel}$,
$\vec{E}_{0\perp}\approx\vec{E}_{0}=-\nabla\Phi_{0}$, while
assuming $\kappa_{e}\gg1\gtrsim\kappa_{i}$, the normal
conductivity tensor given by Eq.~(\ref{sigma_II,P,H}) reduces
to
\begin{subequations}
\label{sigma_II,P,H_reduced
\begin{align}
\sigma_{\mathrm{P}}^{\mathrm{L}} & \approx\frac{(1+\kappa_{i}\kappa_{e})ne
{\kappa_{e}(1+\kappa_{i}^{2})B}=\frac{\kappa_{i}(1+\psi_{\perp})ne
{(1+\kappa_{i}^{2})B},\label{sigma_P_reduced}\\
\sigma_{\mathrm{H}}^{\mathrm{L}} & \approx-\ \frac{ne}{(1+\kappa_{i}^{2})B}.
\label{sigma_H_reduced
\end{align}
\end{subequations}
Under the same conditions, while additionally presuming
field-aligned irregularities, $k_{\parallel}\ll
k_{\perp}\approx k$, Eqs.~(\ref{sigma_NC_II,P,H}) and
(\ref{k.U_0}) simplify to
\begin{subequations}
\label{sigma_red_NC_P,H
\begin{align}
\sigma_{\mathrm{P}}^{\mathrm{NC}} & \approx-\ \frac{en_{0}}{\kappa_{i}
\kappa_{e}}\sum_{\vec{k},\omega\neq0}(\vec{k}\cdot\vec{U}_{0})\left\vert
\frac{\delta n_{\vec{k},\omega}}{n_{0}}\right\vert ^{2}\nonumber\\
&\times\frac{\kappa_{e}\vec{k}\cdot(\vec{E}
_{0}\times\hat{b})+\left( 1+\kappa_{i}\kappa_{e}\right) \vec{k}\cdot\vec
{E}_{0}}{(1+\psi_{\vec{k}})k^{2}E_{0}^{2}},\label{sigma_red_NC_P}
\\
\sigma_{\mathrm{H}}^{\mathrm{NC}} & \approx-\ \frac{en_{0}}{\kappa_{i
\kappa_{e}}\sum_{\vec{k},\omega\neq0}(\vec{k}\cdot\vec{U}_{0})\left\vert
\frac{\delta n_{\vec{k},\omega}}{n_{0}}\right\vert ^{2}\nonumber\\
&\times\frac{\left(1+\kappa_{i}\kappa
_{e}\right) \vec{k}\cdot(\vec{E}_{0}\times\hat{b})-\kappa_{e}(\vec{k
\cdot\vec{E}_{0})}{(1+\psi_{\vec{k}})k^{2}E_{0}^{2}
},
\label{sigma_red_NC_H
\end{align
\end{subequations}
\begin{equation}
\vec{U}_{0}=\vec{U}_{0\perp}\approx\frac{\kappa_{e}(\vec{E}_{0}\times\hat
{b})-\left( 1+\kappa_{i}\kappa_{e}\right) \vec{E}_{0}]}{\kappa_{e}\left(
1+\kappa_{i}^{2}\right) B}\simeq\frac{\vec{E}_{0}\times\hat{b}-\kappa_{i
\vec{E}_{0}}{\left( 1+\kappa_{i}^{2}\right) B}, \label{k>U_0_red
\end{equation}
where the last approximation of Eq.~(\ref{k>U_0_red})
represents the simplest interpolation between lower E/D-region
altitudes with $\kappa_{i}\ll1$, $\kappa_{e} \gg
1\gtrsim\kappa_{i}\kappa_{e}$ and higher \textit{E}-region
altitudes with $\kappa_{i} \sim1$, $\kappa_{i}\kappa_{e}\gg1$.
Quasilinear Eq.~(\ref{sigma_red_NC_P,H}) applies to arbitrary
spectra of \textit{E}-region turbulence, provided the density
perturbations are reasonably small (see the discussion below
related to Fig.~\ref{Fig:Delta n/n_0}). In the next section, we
employ a specific model of non-linearly saturated turbulence to
calculate $\sigma_{\mathrm{P,H}}^{\mathrm{NC}}$.
\section{Turbulent Ionospheric Conductivity: \allowbreak
Heuristic Model
of Turbulence\label{Heuristic Model of Turbulence}}
The current state of \textit{E}-region instability theory does
not give us accurate spectra of density irregularities $\delta
n_{\vec{k},\omega}$ as functions of the external electric field
and ionospheric parameters. In order to estimate the turbulent
conductivities we are forced to use simplified models of
non-linearly saturated turbulence. Here we employ a heuristic
model of turbulence (HMT) developed previously for qualitative
explanation of AEH by \citet{DimantMilikh:JGR03}. The HMT
provides approximate rms values of the coupled turbulent
electric field and density irregularities. This analytical
model based on simple physical reasoning has been successfully
tested by quantitative comparisons with AEH observations
\citep{MilikhDimant:JGR03,MilikhGoncharenko:Anom2006} and our
recent 3-D PIC simulations \citep{Oppenheim:Fully2011}. A
significant advantage of employing the HMT is that this model
enables one to estimate the turbulent conductivities in terms
of the rms density irregularities, without any knowledge of
their detailed $\vec{k},\omega$-spectrum.
The HMT consists of two major heuristic assumptions of
developed turbulence during the non-linearly saturated stage of
the FB instability \citep{DimantMilikh:JGR03}. First, the model
assumes the effective values of the major perpendicular
component of the turbulent electric field,
$\delta\vec{E}_{\vec{k}\perp}\perp\vec{B}$. Second, it assumes
the effective aspect angles determined by effective
$k_{\parallel }/k_{\perp}$. Both these assumptions refer to a
modified FB-instability threshold field within the developed
turbulence, where the combined linear growth/damping rate
$\gamma_{\vec{k}}$ equals zero. This modified threshold field,
hereinafter referred to as merely the threshold field, is
determined by the same expression as the actual threshold of
instability excitation where the plasma temperatures are
replaced by the elevated temperatures due to anomalous heating.
We determine this threshold field using general expressions for
the Farley-Buneman linear growth rate obtained in the appendix.
To calculate the threshold electric field necessary to maintain
turbulence, $E_{\mathrm{Thr}}^{\min}$, we can write
$\gamma_{\vec{k}}$ in the limit of $\kappa_e\gg 1 \gtrsim
\kappa_i$ as
\begin{equation}
\gamma_{\vec{k}}\approx\frac{\psi_{\vec{k}}k_{\perp}^{2}}{(1+\psi_{\vec{k
})\nu_{i}}\left[ \frac{\left( 1-\kappa_{i}^{2}\right) U_{0}^{2}\cos^{2
\chi_{\vec{k}}}{(1+\psi_{\vec{k}})^{2}}-C_{s}^{2}\right] ,
\label{gamma_total_reduced
\end{equation}
where $\chi_{\vec{k}}$ is the angle between $\vec{k}$ and
$\vec{U}_{0}$ \citep{DimOppen2004:ionthermal1},
$C_{s}=(T_{e}+T_{i})^{1/2}/m_{i}$ is the isothermal
ion-acoustic speed, $U_{0}\equiv|\vec{U}_{0}|$,
\begin{subequations}
\label{psi
\begin{align}
\psi_{\vec{k}} & \equiv\psi_{\perp}\left[ 1+(1+\kappa_{e}^{2})(1+\kappa
_{i}^{2})\frac{k_{\parallel}^{2}}{k_{\perp}^{2}}\right] ,\label{psi_k}\\
\psi_{\perp} & \equiv\frac{1}{\kappa_{i}\kappa_{e}}=\frac{\nu_{e}\nu_{i
}{\Omega_{e}\Omega_{i}}, \label{psi_perp
\end{align}
\end{subequations}
and we presume a strict inequality of $\kappa_{i}<1$. Using the relation
\begin{equation}
U_{0}\simeq\frac{E_{0}}{(1+\kappa_{i}^{2})^{1/2}B}, \label{U_0_via_E_0
\end{equation}
following from Eq.~(\ref{k>U_0_red}), we can rewrite
Eq.~(\ref{gamma_total_reduced}) a
\begin{equation}
\gamma_{\vec{k}}=\frac{\psi_{\vec{k}}(1-\kappa_{i}^{2})k_{\perp}^{2}
{(1+\psi_{\vec{k}})(1+\kappa_{i}^{2})B^{2}\nu_{i}}\left[ \frac{E_{0}^{2
\cos^{2}\chi_{\vec{k}}}{(1+\psi_{\vec{k}})^{2}}-\frac{(E_{\mathrm{Thr}}^{\min
})^{2}}{(1+\psi_{\perp})^{2}}\right] , \label{gamma_total_reduced_2
\end{equation}
where $E_{\mathrm{Thr}}^{\min}$ is the minimum FB-instability threshold
electric field at a given altitude for optimally directed waves with $\vec
{k}_{\perp}\parallel\vec{U}_{0}$ and $k_{\parallel}=0$ (i.e., for $\chi
_{\vec{k}}=0$ and $\psi_{\vec{k}}=\psi_{\perp}$)
\begin{equation}
E_{\mathrm{Thr}}^{\min}=(1+\psi_{\perp})\left( \frac{1+\kappa_{i}^{2
}{1-\kappa_{i}^{2}}\right) ^{1/2}E_{\mathrm{Thr}}^{(0)}
\label{E_Thr}
\end{equation}
Her
\begin{equation}
E_{\mathrm{Thr}}^{(0)}=C_{s}B\approx20\left( \frac{T_{e}+T_{i}
{600\mathrm{K}}\right) ^{1/2}\left( \frac{B}{5\times10^{4}\mathrm{nT
}\right) \mathrm{mV/m
\label{E_Thr_0}
\end{equation}
is the absolute threshold-field minimum which can only be
reached at optimum altitudes for the FB instability excitation
where the conditions of $\kappa_{i}^{2}\ll1$ and
$\psi_{\perp}=\Theta_{0}^{2}/\kappa_{i}^{2}\ll1$ overlap (at
high latitudes, this occurs at 100--105~km altitude
\citep[][Fig.~2]{DimOppen2004:ionthermal1};
$\Theta_{0}^{2}\equiv m_{e}\nu_{e}/(m_{i}\nu
_{i})\simeq1.8\times10^{-4}$). Be advised that notations in
this paper slightly differ from those in \citet[][Eqs. (14),
(15) and others]{DimantMilikh:JGR03}. Namely, $E_{0}$ in this
paper corresponds to $E_{C}$ in \citet{DimantMilikh:JGR03},
while our $E_{\mathrm{Thr}}^{(0)}$ corresponds to $E_{0}$ and
$E_{\mathrm{Thr}}^{\min}$ corresponds to $E_{\mathrm{Thr}}$ in
that paper.
The first heuristic assumption specifies the rms perpendicular
turbulent electric field,
\begin{equation}
\langle\delta E_{\perp}^{2}\rangle\simeq\alpha_{1}(E_{0}-E_{\mathrm{Thr}
}^{\min})^{2},
\label{E_perp_heuristic}
\end{equation}
where $\alpha_{1}$ is a dimensionless factor of order unity
\citep[][Eq.~(25)]{DimantMilikh:JGR03}. The logic behind
Eq.~(\ref{E_perp_heuristic}) is that in the non-linearly
saturated state the major electron non-linearity
$\propto(\delta\vec{E}_{\perp }\times\delta n)$ balances, on
average, the linear instability growth for optimally directed
waves. For $E_{0}\gg E_{\mathrm{Thr}}^{\min}$, the rms
turbulent field, $\langle\delta E_{\perp}^{2}\rangle^{1/2}$, is
similar in magnitude to $E_{0}$, while near the minimum FB\
instability threshold, $E_{0}\approx E_{\mathrm{Thr}}^{\min}$,
it reduces linearly with $(E_{0}-E_{\mathrm{Thr}}^{\min})$.
This heuristic assumption makes no distinction between the 2-D
and 3-D cases, implying that the rms perpendicular turbulent
fields in both cases are approximately the same. This has been
confirmed by our recent fully kinetic simulations
\citep{Oppenheim:Fully2011}.
The other heuristic assumption distinguishes the 3-D case from
the purely 2-D one by quantifying the effective magnitude of
$k_{\parallel}$. In \textit{E}-region turbulence,
$k_{\parallel}$ is always much less than $k_{\perp}\approx k$,
but even small $k_{\parallel}$ matter because for highly mobile
electrons with $\kappa_e\gg 1$ they significantly modify the
parameter $\psi_{\vec{k}}$ defined by Eq.~(\ref{psi_k}) and,
hence, the linear instability threshold. Also, it is the
parallel turbulent field, $\delta E_\|\propto k_\|$, that
largely cause AEH. The idea of the second HMT assumption is
that the effective values of $k_{\parallel}$ in developed
turbulence, on average, settle the system on the margin of
linear stability where $\gamma_{\vec{k}}=0$. Given the optimum
value of the flow angle for the pure FB instability, the
marginal linear stability yields
\begin{equation}
\psi_{\vec{k}}\simeq\psi_{\vec{k}}^{\mathrm{m}}\equiv\frac{E_{0
}{E_{\mathrm{Thr}}^{\min}}(1+\psi_{\perp})-1 \label{1+psi_m
\end{equation}
(cf. \citet[][Eq.~(26)]{DimantMilikh:JGR03} with our
$\psi_{\vec{k}}^{\mathrm{m}}$ corresponding to
$(1+\kappa_{i}^{2})\psi_{\max}$ in \citet{DimantMilikh:JGR03}
(the conditions in the two papers look formally different, but
they are essentially the same because of the negligible
difference $\sim\psi_{\perp}\kappa_{i}^{2}=\Theta_{0}^{2}
\simeq1.8\times10^{-4}$). Equations (27)--(29) from
\citet{DimantMilikh:JGR03} yield specific values for the
effective rms values of the parallel turbulent field in terms
of a second unknown constant of order unity, $\alpha_{2}$. That
was important for determining the average heating source
responsible for AEH but is not required for our current
purposes. Equations~(\ref{E_perp_heuristic}) and
(\ref{1+psi_m}) is all we need from HMT to quantify the rms
density perturbations and hence the corresponding non-linear
current.
In the quasilinear approximation, the turbulent electric field spectrum,
$\delta\vec{E}_{\vec{k},\omega}=-i\vec{k}\delta\Phi_{\vec{k},\omega}$, is
proportional to the density irregularity spectrum $\delta n_{\vec{k},\omega}$.
In our current limit of $\kappa_{e}\gg1>\kappa_{i}$,
Eq.~(28) from \citet{Dimant:Magnetosphere2011_Budget} yields
\begin{equation}
\delta\vec{E}_{\vec{k},\omega} =
\frac{(1+\kappa_{i}^{2})BU_{0}
\cos{\chi_{\vec{k}}}}{\kappa_{i}(1+\psi_{\vec{k}})}\left( \frac{\delta
n_{\vec{k},\omega}}{n_{0}}\right) ,
\label{delta_E_versus_delta_n}
\end{equation}
where the proportionality coefficient between
$\delta\vec{E}_{\vec{k},\omega}$ and $\delta
n_{\vec{k},\omega}$ depends on the wavevector direction via
$\cos{\chi_{\vec{k}}}=\vec{k}\cdot\vec{U}_{0}/(kU_0)$ and
$\psi_{\vec{k}}$, but is independent of its absolute value,
$k$. This fact allows us to directly relate the rms value of
the entire turbulent field, $\langle\delta E_{\perp}^{2}
\rangle^{1/2}\equiv(\sum_{\vec{k},\omega\neq0}|\delta\vec{E}_{\vec{k},\omega
}|^{2})^{1/2}$, to the rms of the total density fluctuations,
$\langle\delta
n^{2}\rangle^{1/2}\equiv(\sum_{\vec{k},\omega\neq0}|\delta
n_{\vec{k},\omega }|^{2})^{1/2}$. Indeed, if we assume in
accord with simulations
\citep[e.g.,][]{OppenDim2004:ionthermal2,Oppenheim:Large-Scale2008,Oppenheim:Fully2011}
that most of the $\vec{k}_{\perp}$-spectrum is concentrated
within a narrow angular sector around a preferred wavevector
direction
\begin{equation}
\left(\frac{\vec{k}}{k}\right)^{\mathrm{pr}}\simeq\frac{\vec{U}_{0}\cos\chi_{\vec
{k}}^{\mathrm{pr}}+\vec{U}_{0}\times\vec{b}\ \sin\chi_{\vec{k}}^{\mathrm{pr
}} {U_{0}} , \label{k_pref
\end{equation}
where the flow angle $\chi_{\vec{k}}^{\mathrm{pr}}$ with the
positive sign of $\sin\chi_{\vec{k}}^{\mathrm{pr}}$ corresponds
to the tilt in the $(-\vec {E}_{0})$-direction
\citep{DimOppen2004:ionthermal1}, then, setting for the entire
spectrum $\vec{k}/k\simeq (\vec{k}/k)^{\mathrm{pr}}$,
$\psi_{\vec{k}} \simeq\psi_{\vec{k}}^{\mathrm{m}}$ and using
Eq.~(\ref{U_0_via_E_0}), we obtain from
Eq.~(\ref{delta_E_versus_delta_n}) an approximate relation
\begin{align}
\frac{\langle\delta n^{2}\rangle^{1/2}}{n_{0}} & \simeq\frac{\kappa
_{i}(1+\psi_{\vec{k}}^{\mathrm{m}})\langle\delta E_{\perp}^{2}\rangle
^{1/2}\cos\chi_{\vec{k}}^{\mathrm{pr}}}{(1+\kappa_{i}^{2})BU_{0}}\nonumber\\
& \simeq\frac{\kappa_{i}\alpha_{1}^{1/2}(1+\psi_{\perp})\cos\chi_{\vec{k
}^{\mathrm{pr}}}{(1+\kappa_{i}^{2})^{1/2}}\left( \frac{E_{0}}{E_{\mathrm{Thr
}^{\min}}-1\right) . \label{|del_n|_vs_|del_E|
\end{align}
According to Eq.~(\ref{E_Thr}), we have
\begin{equation}
\frac{E_{0}}{E_{\mathrm{Thr}}^{\min}}=\frac{E_{0}}{(1+\psi_{\perp
})E_{\mathrm{Thr}}^{(0)}}\left( \frac{1-\kappa_{i}^{2}}{1+\kappa_{i}^{2
}\right) ^{1/2} \label{E_0/E_Thr
\end{equation}
where $E_{\mathrm{Thr}}^{(0)}$ is defined by Eq.~(\ref{E_Thr_0})
\begin{figure}
\noindent\includegraphics[width=21pc]{E_0-E_Thr}
\caption{Relative density perturbations, $\langle\delta n^2\rangle^{1/2}/n_0$,
vs. the ion magnetization parameter $\kappa_i\equiv \Omega_i/\nu_i$ for different
values of $E_0/E_{\mathrm{Thr}}^{\min}$ (shown near the curves).}
\label{Fig:Delta n/n_0}
\end{figure}
Figure~\ref{Fig:Delta n/n_0}\ based on
Eqs.~(\ref{|del_n|_vs_|del_E|}) and (\ref{E_0/E_Thr}) with
$\alpha_{1}=1$ and $\chi_{\vec{k}}^{\mathrm{pr}}=0$
\citep{DimantMilikh:JGR03} shows that if
$E_{0}/E_{\mathrm{Thr}}^{(0)}$ is large enough then relative
rms density perturbations may even exceed unity. Such
non-physical occasions break the validity of our quasilinear
approximation. We should bear in mind, however, that large
values of $E_{0}$ automatically lead to strong electron and ion
temperature elevations due to combined regular and anomalous
heating. This raises
$E_{\mathrm{Thr}}^{(0)}\propto(T_{e}+T_{i})^{1/2}$ and prevents
$E_{0}/E_{\mathrm{Thr}}^{(0)}$ from reaching too large values
(we discuss this issue in more detail in the next section). For
example, in the extreme case of $E_{0}=160\,$mV/m, the electron
temperature reaches above $4000$\thinspace K
\citep{Bahcivan:Plasma2007} at an altitude around 110~km
corresponding to $\kappa_{i}\simeq 0.3$
\citep[][Fig.~2]{DimOppen2004:ionthermal1}, not counting the
simultaneously increasing ion temperature. According to
Eq.~(\ref{E_Thr_0}), this raises $E_{\mathrm{Thr}}^{(0)}$ above
$50\,$mV/m, making $E_{0}/E_{\mathrm{Thr}}^{(0)}<3$. We believe
that this extreme value of $E_{0}/E_{\mathrm{Thr} }^{(0)}$
imposes the top restriction on possible density perturbations,
keeping them below $50\%$. Less extreme but more typical
density perturbations are expected to be largely within the
reliable limits of the quasilinear approach.
Applying $\vec{k}/\vec{k}\simeq
(\vec{k}/\vec{k})^{\mathrm{pr}}$, while replacing $(1+\psi
_{\vec{k}}^{\mathrm{m}})$ according to Eq.~(\ref{1+psi_m}) and
$\langle\delta n^{2}\rangle^{1/2}$ according to
Eq.~(\ref{|del_n|_vs_|del_E|}), we use
Eqs.~(\ref{sigma_red_NC_P,H}) and (\ref{k>U_0_red}) to express
the turbulent conductivities in terms of the corresponding
laminar conductivities given by
Eq.~(\ref{sigma_II,P,H_reduced}) as
\begin{subequations}
\label{sigma_final_heuristic
\begin{align}
\sigma_{\mathrm{P}}^{\mathrm{NC}} & \simeq\frac{\alpha_{1}\left[ \left(
1-\kappa_{i}^{2}\right) \cos\chi_{\vec{k}}-2\kappa_{i}\sin\chi_{\vec{k
}\right] \cos^{3}\chi_{\vec{k}}}{1+\kappa_{i}^{2}}\nonumber\\
& \times\left( \frac{E_{0}}{E_{\mathrm{Thr}}^{\min}}-1\right) \left(
1-\frac{E_{\mathrm{Thr}}^{\min}}{E_{0}}\right) \sigma_{\mathrm{P}}
^{\mathrm{L}},\label{sigma_final_heuristic_P}\\
\sigma_{\mathrm{H}}^{\mathrm{NC}} & \simeq-\ \frac{\alpha_{1}\left[
2\kappa_{i}\cos\chi_{\vec{k}}+\left( 1-\kappa_{i}^{2}\right) \sin\chi
_{\vec{k}}\right] (1+\psi_{\perp})\cos^{3}\chi_{\vec{k}}}{1+\kappa_{i}^{2
}\nonumber\\
& \times\left( \frac{E_{0}}{E_{\mathrm{Thr}}^{\min}}-1\right) \left(
1-\frac{E_{\mathrm{Thr}}^{\min}}{E_{0}}\right) \sigma_{\mathrm{H}}^{\mathrm{L}}.
\label{sigma_final_heuristic_H
\end{align}
\end{subequations}
These relations are only applicable for $E_{0}\geq
E_{\mathrm{Thr}}^{\min}$ and $\kappa_i\leq 1$, otherwise
$\sigma_{\mathrm{P,H}}^{\mathrm{NC}}=0$. Furthermore, according
to Eqs.~(\ref{E_Thr}) and (\ref{E_Thr_0}), the threshold field
$E_{\mathrm{Thr}}^{\min}\propto E_{\mathrm{Thr}}^{(0)}\propto
(T_e + T_i)^{1/2}$ depends strongly on the electron and ion
temperatures. This means that the self-consistent inclusion of
the turbulent conductivity in the total conductivity tensor
requires simultaneously including AEH
\citep{DimantMilikh:JGR03,MilikhDimant:JGR03}, otherwise the
effect of NC-induced turbulent conductivity will be exaggerated
dramatically.
The AEH-induced increase in the plasma density due to the
temperature reduction of the recombination rate
\citep{DimantMilikh:JGR03,MilikhGoncharenko:Anom2006} will
raise all values in the conductivity tensor in proportion to
the increased plasma density. This slowly-developing effect,
however, can slightly decrease if the strong external electric
field $\vec{E}_0$ varies too fast
\citep{Codrescu:Importance1995,Codrescu:Electric2000,Matsuo:Effects2008,
Cosgrove:Comparison2009,
CosgroveCodrescu:Electric2009}.
\section{Turbulent Frictional Heating in Global Modeling
\label{Turbulent Frictional Heating in Global Modeling}}
An accurate description of strong electric-field perturbations,
such as storms, substorms, and sub-auroral polar streams,
requires including macroscopic effects of \textit{E}-region
turbulence into global computer models, like the Coupled
Magnetosphere Ionosphere Thermosphere (CMIT) model
\citep{Wiltberger:Initial2004,Wang:Initial2004}. This model has
been created by combining the magnetosphere MHD solver LFM
\citep{LyonFedder:Lyon-Fedder-Mobarry2004} and
ionosphere-thermosphere NCAR solver TIEGCM
\citep{RobleRidley:Thermosphere1994,Wang:High-resolution1999}.
The ring-current model RCM \citep{Toffoletto:Inner2003Review}
has been added recently, as well as TIEGCM has been extended to
cover the mesosphere (TIMEGCM). Currently, the CMIT-RCM model
includes ionospheric processes associated only with laminar
conductivities. The turbulence-induced non-linear current (NC)
can be incorporated by adding turbulent conductivities given by
Eq~(\ref{sigma_final_heuristic}). Self-consistency requires
including also anomalous plasma heating caused by turbulent
fields. The AEH increases (through the reduced recombination
rate) the plasma density and, hence, all conductivities in
proportion. However, the anomalous temperature elevations exert
a negative feedback on the saturated level of the plasma
turbulence, reducing the NC and its non-linear Pedersen
conductivity.
Effects of \textit{E}-region turbulence can be included
directly in TIMEGCM and other codes that that model
high-latitude neutral and plasma dynamics and chemical
reactions. Given accurate heating sources, TIMEGCM includes all
ionization/recombination and collisional cooling processes
automatically. To account for temperature inputs caused by both
laminar and turbulent fields, the energy balance equation for
each species should include the corresponding source of
frictional heating. Currently, TIMEGCM includes the laminar ion
and neutral frictional heating but neglects the electron one.
This is a reasonable approximation for altitudes above 100 km
under quiet ionospheric conditions. However, during strong
electric-field events, the \textit{E}-region turbulence via AEH
dramatically raises the electron temperature, meaning that the
total electron frictional heating should also be included. For
the convection field well above the instability threshold field
given by Eqs.~(\ref{sigma_final_heuristic}), ion turbulent
heating can also be appreciable (for $E_{0}\gg E_{\mathrm{Thr}
}^{\min}$, it is comparable to the ion laminar heating).
Accuracy of the entire global model requires including the
corresponding neutral frictional heating as well. Note that
highly anisotropic \textit{E}-region turbulence also causes
average momentum changes in the plasma. Plasma-neutral
collisions in turn can transfer these changes to neutral
particles. For the weakly ionized \textit{E} region, however,
relative momentum changes are much less important than relative
energy changes, so that we will ignore the former and focus
entirely on the latter.
In the multi-component \textit{E}-region ionosphere where
electron-neutral (\emph{e-n}) and ion-neutral (\emph{i-n})
collisions dominate, we can calculate the total laminar and
turbulent frictional heating of electrons,
$H_{e}^{\mathrm{tot}}$, ions, $H_{i}^{\mathrm{tot}}$, and
neutrals, $H_{n}^{\mathrm{tot}}$, using a quasilinear
approximation to quadratic accuracy. In terms of the average
turbulent perturbations, these are given by
\begin{align}
H_{e}^{\mathrm{tot}} & =m_{e}\nu_{e}\left[ n_{0}\left( V_{e0
^{2}+\left\langle \delta V_{e}^{2}\right\rangle \right) +2(\vec{V}_{e0
\cdot\langle\delta n\delta\vec{V}_{e}\rangle)\right] ,\nonumber\\
H_{i}^{\mathrm{tot}} & =\sum_{n}\frac{m_{i}m_{n}\nu_{in}}{m_{i}+m_{n
}\nonumber\\
& \times\left[ n_{i0}\left( V_{i0}^{2}+\left\langle \delta V_{i
^{2}\right\rangle \right) +2(\vec{V}_{i0}\cdot\langle\delta n_{i}\delta
\vec{V}_{i}\rangle)\right] ,\label{Heatingi_all}\\
H_{n}^{\mathrm{tot}} & =\sum_{i}\frac{m_{i}^{2}\nu_{in}}{m_{i}+m_{n
}\nonumber\\
& \times\left[ n_{i0}\left( V_{i0}^{2}+\left\langle \delta V_{i
^{2}\right\rangle \right) +2(\vec{V}_{i0}\cdot\langle\delta n_{i}\delta
\vec{V}_{i}\rangle)\right] ,\nonumber
\end{align}
with the summations taken over all possible collisions between ions and neutrals,
where indices $i$ and $n$ refer to separate groups of ions or neutrals; $\sum
_{i}n_{i0}=n_{0}$. The physical meaning of various terms in the square
brackets is explained in \citet{Dimant:Magnetosphere2011_Budget}. The
expressions for $H_{i,n}^{\mathrm{tot}}$ explicitly
take into account the fact that only $m_{n}/(m_{i}+m_{n})$ of the partial
\emph{i-n} electric field heating, $m_{i}\nu_{in}V_{i}^{2}$, goes to ions
($H_{i}^{\mathrm{tot}}$), while the remaining fraction goes directly to the
colliding neutrals ($H_{n}^{\mathrm{tot}}$). The total electron collision
frequency $\nu_{e}$ includes all electron-neutral (\emph{e-n}) collisions,
$\nu_{e}=\sum_{n}\nu_{en}$; but because $m_{e}\ll m_{n}$, the
contributions to $H_{n}^{\mathrm{tot}}$ from the \emph{e-n} collisions are
negligible. Equation~(\ref{Heatingi_all}) is written in the neutral frame of
reference, presuming that all thermosphere components move with a common
neutral-wind velocity.
In a multi-fluid plasma, the laminar and turbulent velocities,
$\vec{V}_{s0}$ and $\delta\vec{V}_{s}$, can be expressed in
terms of the convection field $\vec {E}_{0}$ and spectral
harmonics of the wave potential $\delta\Phi_{\vec {k},\omega}$
similarly to the two-fluid plasma
\citep{Dimant:Magnetosphere2011_Budget}. However, the
first-order linear relations between the harmonics of the ion
spectral densities, $\delta n_{i\vec{k},\omega}$, and
potential, $\delta\Phi_{\vec{k},\omega}$, which are crucial for
obtaining equations like Eq.~(\ref{delta_E_versus_delta_n}),
(\ref{delta_Phi_via_delta_n_final}), can be rather complicated
or cannot even be expressed in any closed analytical form. The
reason is that the algebraic order of the underlying linear
dispersion relation for the first-order wave frequency,
$\omega_{\vec{k}}$, in the general case equals the number of
separate ion species. Even for the two dominant \emph{E}-region
ion species, NO$^{+}$ and O$_{2}^{+}$, when the dispersion
relation reduces to a quadratic equation, the first-order
relations become cumbersome. These relations, however, simplify
dramatically if all ion species have a common magnetization
parameter, $\kappa_{i}=\Omega_{i}/\nu_{i}$
($\nu_{i}\equiv\sum_{n}\nu_{in}$), so that all ion species
respond to the fields equally. As a result, the multi-fluid
relation between $\delta\Phi_{\vec {k},\omega}$ and $\delta
n_{\vec{k},\omega}$ and various terms in $n_{s0}\left(
V_{s0}^{2}+\left\langle \delta V_{s}^{2}\right\rangle \right)
+2(\vec{V}_{s0}\cdot\langle\delta
n_{s}\delta\vec{V}_{s}\rangle)$ ($s=e,i$) reduce to two-fluid
Eqs.~(49) and (51) from
\citet{Dimant:Magnetosphere2011_Budget}. For
$\kappa_{e}\gg1\gtrsim\kappa_{i}$, these become
\begin{equation}
\nu_{e}n_{0}V_{e0}^{2}\approx\frac{\nu_{e}n_{0}E_{0}^{2}}{B^{2}},\qquad\nu
_{i}n_{i0}V_{i0}^{2}=\frac{\nu_{i}n_{i0}\kappa_{i}^{2}E_{0}^{2}}{\left(
1+\kappa_{i}^{2}\right) B^{2}},\label{zero-order-heat
\end{equation
\begin{subequations}
\label{diag_psi
\begin{align}
\nu_{e}n_{0}\left\langle \delta V_{e}^{2}\right\rangle & \approx\frac
{m_{i}\nu_{i}n_{0}(1+\kappa_{i}^{2})}{m_{e}}\sum_{\vec{k},\omega\neq0
\frac{\psi_{\vec{k}}(\vec{k}\cdot\vec{U}_{0})^{2}}{(1+\psi_{\vec{k}
)^{2}k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec{k},\omega}}{n_{0
}\right\vert ^{2},\label{diag_psi_ee}\\
\nu_{i}n_{i0}\left\langle \delta V_{i}^{2}\right\rangle & \approx\nu
_{i}n_{i0}(1+\kappa_{i}^{2})\sum_{\vec{k},\omega\neq0}\frac{(\vec{k}\cdot
\vec{U}_{0})^{2}}{(1+\psi_{\vec{k}})^{2}k_{\perp}^{2}}\left\vert \frac{\delta
n_{\vec{k},\omega}}{n_{0}}\right\vert ^{2},\label{diag_psi_ii
\end{align
\end{subequations}
\begin{subequations}
\label{``diag''_2
\begin{align}
2\nu_{e}(\vec{V}_{e0}\cdot\langle\delta n\delta\vec{V}_{e}\rangle) &
\approx\frac{2en_{0}(1+\kappa_{i}^{2})}{m_{e}\kappa_{i}\kappa_{e}}\nonumber\\
&\sum
_{\vec{k},\omega\neq0}\frac{(\vec{k}\cdot\vec{E}_{0})(\vec{k}\cdot\vec{U}
_{0})}{(1+\psi_{\vec{k}})k_{\perp}^{2}}\left\vert \frac{\delta
n_{\vec
{k},\omega}}{n_{0}}\right\vert ^{2},\label{``diag''_2_e}\\
2\nu_{i}(\vec{V}_{i0}\cdot\langle\delta n_{i}\delta\vec{V}_{i}\rangle) &
\approx\frac{2en_{i0}}{m_{i}}\sum_{\vec{k},\omega\neq0}\frac{(\vec{k}\cdot
\vec{E}_{0})(\vec{k}\cdot\vec{U}_{0})}{(1+\psi_{\vec{k}})k_{\perp}^{2
}\left\vert \frac{\delta n_{\vec{k},\omega}}{n_{0}}\right\vert ^{2
.\label{``diag''_2_i
\end{align}
where $\delta n_{\vec{k},\omega}=\sum_{i}\delta
n_{i\vec{k},\omega}$ is the electron density, $\psi_{\vec{k}}$
and $\vec{U}_{0}$ are given by Eqs.~(\ref{psi}) and
(\ref{U_0_via_E_0}), and
$(1+\kappa_{i}^{2})(1+\kappa_{e}^{2}k_{\parallel}^{2}/k_{\perp}^{2})$
emerging in the derivation of Eq.~(\ref{diag_psi_ee}) was
approximated by
$\psi_{\vec{k}}/\psi_{\perp}=\kappa_{e}\kappa_{i}\psi_{\vec{k}}$
\citep[][see Eq.~(34b) and text below
it]{Dimant:Magnetosphere2011_Budget}.
To calculate heating rates in terms of \textit{E}-region
parameters, we need estimates of the turbulent density
perturbations. Using the heuristic model of saturated
turbulence, along with the approach that lead us from
Eq.~(\ref{sigma_red_NC_P,H}) via Eqs.~(\ref{E_perp_heuristic}),
(\ref{1+psi_m}), and (\ref{|del_n|_vs_|del_E|}) to
Eq.~(\ref{sigma_final_heuristic}), we obtain
\end{subequations}
\begin{align}
&\sum_{\vec{k},\omega\neq0}\frac{\psi_{\vec{k}}(\vec{k}\cdot\vec{U}_{0})^{2
}{(1+\psi_{\vec{k}})^{2}k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec
{k},\omega}}{n_{0}}\right\vert ^{2} \simeq\frac{\alpha_{1}\kappa_{i
^{2}(E_{0}-E_{\mathrm{Thr}}^{\min})^{2}\cos^{4}\chi_{\vec{k}}^{\mathrm{pr}
}{(1+\kappa_{i}^{2})^{3}B^{2}}\nonumber\\
& \times\left[ \frac{E_{0}}{E_{\mathrm{Thr}}^{\min}}\left( 1+\psi_{\perp
}\right) -1\right], \nonumber\\
&\sum_{\vec{k},\omega\neq0}\frac{(\vec{k}\cdot\vec{U}_{0})^{2}}{(1+\psi
_{\vec{k}})^{2}k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec{k},\omega}
{n_{0}}\right\vert ^{2} \simeq\frac{\alpha_{1}\kappa_{i}^{2}(E_{0
-E_{\mathrm{Thr}}^{\min})^{2}\cos^{4}\chi_{\vec{k}}^{\mathrm{pr}}
{(1+\kappa_{i}^{2})^{3}B^{2}}, \nonumber\\
&\sum_{\vec{k},\omega\neq0} \frac{(\vec{k}\cdot\vec{E}_{0})(\vec{k}\cdot\vec
{U}_{0})}{(1+\psi_{\vec{k}})k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec
{k},\omega}}{n_{0}}\right\vert ^{2} \nonumber\\
& \simeq\frac{\alpha_{1}\kappa_{i
^{2}(E_{\mathrm{Thr}}^{\min}-E_{0})^{2}\cos^{3}\chi_{\vec{k}}^{\mathrm{pr
}\sin\chi_{\vec{k}}^{\mathrm{pr}}}{(1+\kappa_{i}^{2})^{5/2}B}. \label{podstavy
\end{align}
Equations (\ref{Heatingi_all})--(\ref{podstavy}) provide
frictional heating terms for the potential inclusion into the
corresponding energy-balance equations of TIMEGCM or similar
ionosphere-thermosphere models.
Equations~(\ref{sigma_final_heuristic})--(\ref{Heatingi_all})
provide a reasonable estimate of turbulent conductivity and
heating and should greatly improve modeling of the geospace
environment during disturbed conditions.
\section{Discussion\label{Discussion}}
Figures~\ref{Fig:Temper_150}--\ref{Fig:conduct_80} illustrate
anomalous heating and give examples of the anomalously modified
Pedersen conductivity for two cases: an extreme field of
$150$~mV/m and a still large but more modest case of $80$~mV/m.
These figures are based on typical high-latitude ionospheric
parameters and the same values of the major HMT parameter,
$\alpha_1=1$, as in \citet{MilikhDimant:JGR03}.
Figure~\ref{Fig:Temper_150} shows the effect of AEH and normal
ion heating. The combined temperature, $T_e+T_i$, shows kinks
at altitudes where AEH sharply disappears, while ion heating
continues its steep rising. These altitudes are slightly below
the ion magnetization boundary ($\kappa_i=1$) located at
$h_{1}\simeq 122$~km. These kinks translate to those in the
threshold electric field, $E_{\mathrm{Thr}}^{\min} \propto\
(T_e+T_i)^{1/2}$, as seen in Fig.~\ref{Fig:thresh_field_150}.
Figure~\ref{Fig:thresh_field_150} (a) shows that above 100~km
of altitude the threshold field increases dramatically due to
plasma heating, keeping $E_{0}/E_{\mathrm{Thr}}^{\min}$ at a
modest level of 2--3. The much weaker field of $E_{0}=80$~mV/m
(b) gives a slightly lower ratio because the plasma heating in
this case is also much lower. According to Fig.~\ref{Fig:Delta
n/n_0}, such moderate ratio of $E_{0}/E_{\mathrm{Thr}}^{\min}$,
even for the extreme case of $E_0=150$~mV/m, keeps the total
density fluctuation level below 50\%, i.e., roughly within the
framework of the quasilinear approach (see above). This also
reduces the entire effect of NC on the conductivity, as seen
from Fig.~\ref{Fig:conduct_150}.
Figure~\ref{Fig:conduct_150} shows a typical altitudinal
profile of the undisturbed conductivity (curve 0) along with
two curves including the anomalous conductivity. The NC effect
alone (curve 1) nearly doubles the conductivity in a broad
altitude range between 90~km and the ion magnetization boundary
located at an altitude $h_{1}\simeq 122$~km, fairly close to
the maximum of the normal Pedersen conductivity. The
AEH-induced ionization-recombination effect doubles the total
conductivity once again (curve 2), but this occurs in a
slightly narrower range restricted from above by the AEH upper
boundary, $h_{0}\simeq 119.5$~km. Presuming that the lower
altitudes located below the undisturbed conductivity maximum
near $h_{1}$ include about a half of the entire
height-integrated conductance, we see that the NC-induced
anomalous conductivity alone (curve 1) can contribute up to a
half of the undisturbed conductance, while both anomalous
effects combined (curve 2) can nearly double the entire
conductance. We should bear in mind, however, the caveats
associated with slow evolution of ionization-recombination
processes superposed on rapidly varying electric field (see the
discussion above).
Largely the same situation takes place for the smaller field of
$E_0=80$~mV/m (Fig.~\ref{Fig:conduct_80}). Since in this case
the particle-temperature elevation is more modest than that for
$E_0=150$~mV/m, the instability threshold field shown in
Figure~\ref{Fig:thresh_field_150}~(b) remains noticeably lower
than that in Fig.~\ref{Fig:thresh_field_150}~(a). As a result,
the ratio $E_0/E_{\mathrm{Thr}}^{\min}$ changes much less
dramatically compared to $E_0=150$~mV/m. That is why in
Fig.~\ref{Fig:conduct_80} the NC-induced effect for
$E_0=80$~mV/m turns out to be nearly as strong as that in
Fig.~\ref{Fig:conduct_150} for $E_0=150$~mV/m, while the more
fragile AEH-induced recombination effect is much less
pronounced because of lower electron heating.
Above we presumed a rather conservative value for the major
HMT parameter, $\alpha_1=1$, the same as in
\citet{DimantMilikh:JGR03,MilikhDimant:JGR03,Merkin:Anomalous2005,
MilikhGoncharenko:Anom2006}. Recent 2D and 3D PIC simulations
\citep{Oppenheim:Fully2011} suggest that $\alpha_{1}$ can
actually be closer to $1.5$. This would make
$\sigma_{\mathrm{P}}^{\mathrm{NC}}/\sigma_{\mathrm{P}}^{\mathrm{L}}$
roughly 50\% larger, although due to larger $\langle\delta
n^{2}\rangle^{1/2}/n_{0}$ one may have stronger restrictions on
the quasilinear approach. We should also note that our current
heuristic model of developed turbulence does not include the
ion thermal driving mechanism (ITM)
\citep{DimOppen2004:ionthermal1,OppenDim2004:ionthermal2}.
\onecolumn
\begin{figure}
\noindent\includegraphics[width=0.48\textwidth]{temper_150}
\hfill
\noindent\includegraphics[width=0.48\textwidth]{temper_80}
\caption{\label{Fig:Temper_150} Altitude dependence of the
electron and ion temperatures, $T_e$ and $T_i$, for two
different values of $E_0$ (shown in figures). The electron
temperature includes self-consistent AEH according to
\citet{MilikhDimant:JGR03}. The kinks in $T_e$ and $T_e+T_i$
near the 119.5~km altitude (a) and 117.5~km (b) are caused by
sharp disappearance of instability and, hence, AEH.}
\end{figure}
\begin{figure}
\noindent\includegraphics[width=0.48\textwidth]{thresh_field_150} \hfill
\noindent\includegraphics[width=0.48\textwidth]{thresh_field_80}
\caption{Altitude dependence of the FB-instability threshold electric field,
$E_{\mathrm{Thr}}^{\min}$, Eqs.~(\ref{E_Thr}) and (\ref{E_Thr_0}), elevated
due to particle heating for the same values $E_0$ as in Fig.~\ref{Fig:Temper_150}.
The kinks in $E_{\mathrm{Thr}}^{\min}\propto (T_e+T_i)^{1/2}$ correspond to those
in Fig.~\ref{Fig:Temper_150}.}
\label{Fig:thresh_field_150}
\end{figure}
\twocolumn
\begin{figure}
\noindent\includegraphics[width=21pc]{conduct_150}
\caption{Altitude dependence of the total -- normal and anomalous --
Pedersen conductivity for $E_0=150$~mV/m (in relative units). Curve 0
shows the undisturbed conductivity. Curve 1 includes additionally the
NC-induced anomalous conductivity calculated according to
Eq.~(\ref{sigma_final_heuristic_P}) with $\alpha_1=0$ and
$\chi_{\vec{k}}=0$. Curve 2 shows the total Pedersen conductivity
with the AEH-affected plasma density elevated according to the
steady-state ionization recombination model by \citet{MilikhGoncharenko:Anom2006}.
The NC-induced anomalous conductivity disappears above the ion magnetization
boundary $\simeq 122$~km; the AEH-recombination effect vanishes above the top
boundary of anomalous heating $\simeq 119.5$~km (see the corresponding kinks
in Figs.~\ref{Fig:Temper_150} and \ref{Fig:thresh_field_150}).}
\label{Fig:conduct_150}
\end{figure}
\begin{figure}
\noindent\includegraphics[width=21pc]{conduct_80}
\caption{The same as Fig.~\ref{Fig:conduct_150}, but for $E_0=80$~mV/m.}
\label{Fig:conduct_80}
\end{figure}
That is why all our anomalous effects occur strictly below the
ion magnetization boundary, $\kappa_{i}=1$. Possible inclusion
of the ITM would play a two-fold role. On the one hand, this
would expand the altitudinal range of anomalous conductivity to
at least a few kilometers higher
\citep{DimOppen2004:ionthermal1} and increase the total
ionospheric conductance accordingly. A modest reduction in
$\kappa_{i}$ due to anomalously heated ions might also help. On
the other hand, near the ion magnetization boundary the
preferred values of the modified flow angle $\chi_{\vec{k}}$
deviate from $\chi_{\vec{k}}=0$, also due to the ITM as
explained in \citet{DimOppen2004:ionthermal1}. This deviation
of $\chi_{\vec{k}}$ would slightly reduce the
$\chi_{\vec{k}},\kappa_{i}$-dependent factor in the RHS of
Eq.~(\ref{sigma_final_heuristic_P}) and hence lead to a smaller
Pedersen conductivity compared to that for $\chi_{\vec{k}}=0$.
An accurate inclusion of the ITM to the heuristic model of
saturated turbulence, however, would complicate the entire
model. We plan to improve the treatment of the top electrojet
altitudes in future by using both analytical theory and PIC
simulations. At this moment, a better accuracy is of less
importance than the mere fact that the instability-induced
conductivity occupies roughly the entire lower half of the
Pedersen conductive layer and can nearly double the whole
conductance. Note also that polarized sporadic-\textit{E}
clouds \citep{Dimant:Interaction2010}, or even ubiquitous
meteor trails \citep{Dimant:Meteor:External2009}, can also make
additional contributions to anomalous conductances. All these
effects combined can, at least partially, explain why global
MHD codes developed for predictive modeling of space weather
which use normal conductances often overestimate the
cross-polar cap potentials by close to a factor of two.
\section{Summary and Conclusions}
Plasma turbulence generated by \textit{E}-region instabilities
significantly modifies the ionospheric Pedersen conductance by
exciting a net non-linear current,
$\vec{j}^{\mathrm{NC}}=\overleftrightarrow
{\sigma_{\mathrm{P}}^{\mathrm{NC}}}\cdot\vec{E}_{0}$. The
magnitude of $\vec {j}^{\mathrm{NC}}$ can only be a fraction of
the undisturbed Hall current, but can be comparable in
magnitude to the Pedersen current and pointing in that
direction. This creates an additional turbulent conductivity.
Also, anomalous electron heating, via reduced plasma
recombination, increases the mean plasma density $n_0$ and
all, laminar and turbulent, conductances in proportion with
$n_0$. Estimates based on a quasilinear theory and heuristic
model of non-linearly saturated plasma turbulence show that the
anomalous effects combined can nearly double the Pedersen
conductance, as predicted by Eqs.~(\ref{sigma_red_NC_P,H}) and
(\ref{sigma_final_heuristic}) and illustrated by
Figs.~\ref{Fig:conduct_150} and \ref{Fig:conduct_80}. Such
ionospheric response to the magnetospheric field $\vec{E}_{0}$
may efficiently reduce the high-latitude ionospheric resistance
and decrease the cross-polar cap potential. This effect might
explain, at least partially, why routine MHD simulations with
the normal ionospheric conductances systematically overestimate
this potential by a factor of two. The anomalous effects on the
conductances, as well as turbulent frictional heating, should
be included in global MHD codes developed for space weather
prediction.
\section*{Appendix: General FB/GD Dispersion Relation for
Arbitrarily Magnetized Plasmas}
This appendix develops the two-fluid linear theory of the
Farley-Buneman (FB) and gradient drift (GD) instabilities for
arbitrary magnetization parameters,
$\kappa_{e,i}\equiv\Omega_{e,i}/\nu_{e,i}$, that vary from
small values at lower altitudes to large values at higher
altitudes. This analysis has never been published but is
relevant to the lower ionosphere and enables one to accurately
estimate conventionally neglected terms that may matter in some
cases.
The approach described below allows one to obtain the
fluid-model dispersion relation in a reasonably compact and
physically clear way. Compared to Sect.~3.1 of
\citet{Dimant:Magnetosphere2011_Budget}, here we additionally
include the particle pressure, inertia, regular gradients in
the background plasma density, and recombination. These factors
are responsible for the FB and GD instability drivers, as well
as for the stabilizing diffusion and ionization balance. For
simplicity, we do not include non-isothermal processes
responsible for driving of thermal instabilities
\citep{Dimant:Physical97,Kagan:thermal00,DimOppen2004:ionthermal1}
and ignore the stabilizing effect of a weak non-quasineutrality
\citep{RosenbergChow:98,Kovalev:Modeling2008}. These factors
can be added using essentially the same approach.
The quasineutral two-fluid continuity equations including
ionization and recombination balance are
\begin{equation}
\partial_{t}n+\nabla\cdot(n\vec{V}_{e})=\partial_{t}n+\nabla\cdot(n\vec{V
_{i})=Q-\alpha n^{2}, \label{contiki
\end{equation}
where $Q$ is the total source of ionization and $\alpha$ is the recombination
constant. Equation~(\ref{contiki}) includes the standard quasineutral relation
$\nabla\cdot(n\vec{U})=0$, where $\vec{U}\equiv\vec{V}_{e}-\vec{V}_{i}$.
The continuity equations for a stationary ($\partial_t n_0 =0$) but inhomogeneous
background density, $n_{0}$, with the corresponding undisturbed fluid
velocities, $\vec{V}_{e,i0}$, are
\begin{equation}
\nabla\cdot(n_{0}\vec{V}_{e0})=\nabla\cdot(n_{0}\vec{V}_{i0})=Q-\alpha
_{0}n_{0}^{2}. \label{contiki_zero
\end{equation}
From the left equality, the divergence of the relative velocity, $\vec{U
_{0}\equiv\vec{V}_{e0}-\vec{V}_{i0}$, can be expressed in terms of the
background density gradient a
\begin{equation}
\nabla\cdot\vec{U}_{0}=-\vec{U}_{0}\cdot\nabla n_{0}/n_{0}.
\label{div_U_0_via_grad
\end{equation}
In what follows, we presume that both $\vec{U}_{0}$ and $n_{0}$
are known functions of spatial coordinates that satisfy
Eq.~(\ref{div_U_0_via_grad}). Then, to express the zero-order
drift velocities of electrons and ions in terms of $\vec{U}_0$,
we can use Eqs.~(26) and (27) from
\citet{Dimant:Magnetosphere2011_Budget} to obtain
\begin{equation}
\vec{V}_{e0\parallel}=\frac{\kappa_{e}\vec{U}_{0\parallel}}{\kappa_{e
+\kappa_{i}},\qquad\vec{V}_{i0\parallel}=-\ \frac{\kappa_{i}\vec
{U}_{0\parallel}}{\kappa_{e}+\kappa_{i}}, \label{V_0_II_via_U
\end{equation}
\begin{subequations}
\label{V_0_perp_via_U
\begin{align}
\vec{V}_{e0\perp} & =\frac{\kappa_{e}[\vec{U}_{0\perp}-\kappa_{i}(\vec
{U}_{0}\times\hat{b})]}{\kappa_{e}+\kappa_{i}},\label{V_0_perp_via_U_e}\\
\vec{V}_{i0\perp} & =-\ \frac{\kappa_{i}[\vec{U}_{0\perp}+\kappa_{e}(\vec
{U}_{0}\times\hat{b})]}{\kappa_{e}+\kappa_{i}}. \label{V_0_perp_via_U_i
\end{align}
\end{subequations}
Note that these expressions are only valid when neglecting for
the background plasma pressure gradients and gravity,
comparably small effects for almost all ionospheric conditions.
We will now consider Fourier harmonics of wave perturbations, $\delta n_{\vec{k
}$, $\delta\vec{E}_{\vec{k}}$, $\delta\vec{V}_{e,i\vec{k}}\propto\exp
[i(\vec{k}\cdot\vec{r}-\omega_{\vec{k}}t)]$, where the wavevector $\vec{k}$
and the complex linear wave frequency,
$\omega_{\vec{k}}=\omega_{\vec{k}}^{\prime}+i\gamma_{\vec{k}}$,
are locally defined. Here $\omega_{\vec{k}}^{\prime}$ is the
real wave frequency, while $\gamma_{\vec{k}}\equiv
-i\operatorname{Im}\,\omega_{\vec{k}}$ is the wave total growth
or damping rate.
Linearizing Eq.~(\ref{contiki_zero}) with respect to
$\delta n_{\vec{k}}$, $\delta\vec{V}_{e,i\vec{k}}$ and introducing two shifted
complex wave frequencies,
\begin{subequations}
\label{shifted _Omega_k
\begin{align}
\Omega_{\vec{k}}^{(e)} & \equiv\omega_{\vec{k}}-\hat{K}\cdot\vec{V
_{e0}+2i\alpha n_{0},\label{shifted _Omega_k_e}\\
\Omega_{\vec{k}}^{(i)} & \equiv\omega_{\vec{k}}-\hat{K}\cdot\vec{V
_{i0}+2i\alpha n_{0}=\Omega_{\vec{k}}^{(e)}+q, \label{shifted _Omega_k_i
\end{align}
\end{subequations}
with
\begin{equation}
\hat{K}\equiv\vec{k}-i\nabla,\qquad q\equiv\hat{K}\cdot\vec{U}_{0}, \label{Kq
\end{equation}
we obtai
\begin{equation}
\Omega_{\vec{k}}^{(e)}\ \frac{\delta n_{\vec{k}}}{n_{0}}=\vec{p}\cdot
\delta\vec{V}_{e\vec{k}},\qquad\Omega_{\vec{k}}^{(i)}\ \frac{\delta n_{\vec
{k}}}{n_{0}}=\vec{p}\cdot\delta\vec{V}_{i\vec{k}}, \label{Omegas_prom
\end{equation}
wher
\begin{equation}
\vec{p}\equiv\frac{\vec{K}n_{0}}{n_{0}}=\vec{k}-\frac{i\nabla n_{0}}{n_{0}}.
\label{p
\end{equation}
Now we express $\delta\vec{V}_{e,i\vec{k}}$ in terms of $\delta
n_{\vec{k}}$ and $\delta\Phi_{\vec{k},\omega}$,
$\delta\vec{E}_{\vec{k}}=-i\vec{k}\delta\Phi_{\vec{k},\omega}$,
using the momentum equations for the individual electron and
ion mean fluid velocities in the neutral frame,
\begin{subequations}
\label{fluid_momentum
\begin{align}
m_{e}\ \frac{\mathrm{d}_{e}\vec{V}_{e}}{\mathrm{d}t} & =-e(\vec{E}+\vec{V}_{e}
\times\vec
{B})-\frac{\nabla\left( nT_{e}\right) }{n}-m_{e}\nu_{e}\vec{V
_{e},\label{fluid_momentum_e}\\
m_{i}\ \frac{\mathrm{d}_{i}\vec{V}_{i}}{\mathrm{d}t} & =e(\vec{E}+\vec{V}_{i}
\times\vec
{B})-\frac{\nabla\left( nT_{i}\right) }{n}-m_{i}\nu_{i}\vec{V}_{i},
\label{fluid_momentum_i
\end{align}
\end{subequations}
where
$\mathrm{d}_{e,i}/\mathrm{d}t\equiv\partial_{t}+\vec{V}_{e,i}\cdot\nabla$
are the full derivatives for both electrons and ions. For
low-frequency \emph{E}/\emph{D}-region plasma processes, the
electron inertia described by the LHS\ of
Eq.~(\ref{fluid_momentum_e}), unlike the ion inertia in
Eq.~(\ref{fluid_momentum_i}), is usually neglected, but we will
keep it for completeness and symmetry. Neglecting in
Eq.~(\ref{fluid_momentum}) temperature perturbations, we
express the velocity perturbations as
\begin{subequations}
\label{delta_V_prom
\begin{align}
\delta\vec{V}_{e\vec{k}} & =\vec{G}_{e}\left( \delta\Phi_{\vec{k},\omega
-\frac{T_{e}}{e}\ \frac{\delta n_{\vec{k}}}{n_{0}}\right)
,\label{delta_V_prom_e}\\
\delta\vec{V}_{i\vec{k}} & =\vec{G}_{i}\left( \delta\Phi_{\vec{k},\omega
+\frac{T_{i}}{e}\ \frac{\delta n_{\vec{k}}}{n_{0}}\right) ,
\label{delta_V_prom_i
\end{align}
\end{subequations}
where the vector-functions $\vec{G}_{e,i}$,
\begin{subequations}
\label{G
\begin{align}
\vec{G}_{e\parallel} & \approx\frac{i\tilde{\kappa}_{e}\vec{k}_{\parallel
}{B},\qquad\vec{G}_{e\perp}\approx\frac{i\tilde{\kappa}_{e}(\vec{k}_{\perp
}-\tilde{\kappa}_{e}\vec{k}\times\hat{b})}{\left( 1+\tilde{\kappa}_{e
^{2}\right) B},\label{G_e}\\
\vec{G}_{i\parallel} & \approx-\ \frac{i\tilde{\kappa}_{i}\vec{k
_{\parallel}}{B},\qquad\vec{G}_{i\perp}\approx-\ \frac{i\tilde{\kappa
_{i}(\vec{k}_{\perp}+\tilde{\kappa}_{i}\vec{k}\times\hat{b})}{\left(
1+\tilde{\kappa}_{i}^{2}\right) B}, \label{G_i
\end{align}
\end{subequations}
describe anisotropic two-fluid responses to the harmonic perturbations of the
potential and pressures combined. Here, in accord with Eq.~(\ref{ner-vo_T}),
\begin{subequations}
\label{tilde_kappa
\begin{align}
\tilde{\kappa}_{e} & \equiv\frac{\Omega_{e}}{\nu_{e}-i\Omega_{\vec{k}
^{(e)}}\approx\kappa_{e}\left( 1+\frac{i\Omega_{\vec{k}1}^{(e)}}{\nu_{e
}\right) ,\label{tilde_kappa_e}\\
\tilde{\kappa}_{i} & \equiv\frac{\Omega_{i}}{\nu_{i}-i\Omega_{\vec{k}
^{(i)}}\approx\kappa_{i}\left( 1+\frac{i\Omega_{\vec{k}1}^{(i)}}{\nu_{i
}\right) \label{tilde_kappa_i
\end{align}
\end{subequations}
are modified magnetization ratios that include small contributions from the
particle inertia; $\Omega_{\vec{k}1}^{(e,i)}=\operatorname{Re}\Omega_{\vec{k
}^{(e,i)}=\omega_{\vec{k}}-\vec{k}\cdot\vec{V}_{e,i0}$.
The small ion inertia contribution described by
$i\Omega_{\vec{k}1}^{(,i)}/\nu_{i}$ is crucial for excitation
of the FB instability.
Now we substitute the expressions for
$\delta\vec{V}_{e,i\vec{k}}$ from Eq.~(\ref{delta_V_prom}) to
Eq.~(\ref{Omegas_prom}), expressing $\Omega
_{\vec{k}}^{(e)}=\Omega_{\vec{k}}^{(i)}-q$ via
$q=\vec{k}\cdot\vec{U}
_{0}-i\nabla\cdot\vec{U}_{0}=\vec{U}_{0}\cdot(\vec{p})^{\ast}$,
where we used Eq.~(\ref{div_U_0_via_grad}) and (\ref{Kq}). This
yields two independent linear relations between $\delta
n_{\vec{k}}$ and $\delta\Phi_{\vec{k},\omega}$:
\begin{align*}
\left[ \Omega_{\vec{k}}^{(e)}+\frac{T_{e}}{e}(\vec{p}\cdot\vec{G
_{e})\right] \frac{\delta n_{\vec{k}}}{n_{0}} & =(\vec{p}\cdot\vec{G
_{e})\delta\Phi_{\vec{k},\omega},\\
\left[ \Omega_{\vec{k}}^{(e)}+q-\frac{T_{i}}{e}(\vec{p}\cdot\vec{G
_{i})\right] \frac{\delta n_{\vec{k}}}{n_{0}} & =(\vec{p}\cdot\vec{G
_{i})\delta\Phi_{\vec{k},\omega}.
\end{align*}
From these relations and Eq.~(\ref{shifted _Omega_k_i}), we
obtain two symmetric expressions for the coupled shifted
frequencies,
\begin{subequations}
\label{Omega_general
\begin{align}
\Omega_{\vec{k}}^{(e)} & =-\ \frac{(\vec{p}\cdot\vec{G}_{e})[q-(\vec{p
\cdot\vec{G}_{i})\left( T_{e}+T_{i}\right) /e]}{\vec{p}\cdot(\vec{G
_{e}-\vec{G}_{i})},\label{Omega_general_e}\\
\Omega_{\vec{k}}^{(i)} & =-\ \frac{(\vec{p}\cdot\vec{G}_{i})[q-(\vec{p
\cdot\vec{G}_{e})\left( T_{e}+T_{i}\right) /e]}{\vec{p}\cdot(\vec{G
_{e}-\vec{G}_{i})}, \label{Omega_general_i
\end{align}
\end{subequations}
and the general relation between harmonics $\delta\Phi_{\vec{k},\omega}$ and $\delta
n_{\vec{k}}$
\begin{equation}
\delta\Phi_{\vec{k},\omega}=-\ \frac{q-\vec{p}\cdot(\vec{G}_{e}T_{e}+\vec{G}_{i
T_{i})/e}{\vec{p}\cdot(\vec{G}_{e}-\vec{G}_{i})}\ \frac{\delta n_{\vec{k}
}{n_{0}}. \label{delta_Phi_via_delta_n_final
\end{equation}
Using the definitions of $\Omega_{\vec{k}}^{(e)}$ or $\Omega_{\vec{k}}^{(i)}$
given by Eq.~(\ref{shifted _Omega_k}), we obtain the complex wave frequency in the
neutral frame, $\omega_{\vec{k}}=\Omega_{\vec{k}}^{(e)}+\hat{K}\cdot\vec
{V}_{e0}-2i\alpha n_{0}=\Omega_{\vec{k}}^{(i)}+\hat{K}\cdot\vec{V
_{i0}-2i\alpha n_{0}$
\begin{align}
& \omega_{\vec{k}}\equiv\omega_{\vec{k}}^{\prime}+i\gamma_{\vec{k}
=\frac{(\vec{p}\cdot\vec{G}_{e})(\hat{K}\cdot\vec{V}_{i0})-(\vec{p}\cdot
\vec{G}_{i})(\hat{K}\cdot\vec{V}_{e0})}{\vec{p}\cdot(\vec{G}_{e}-\vec{G}_{i
)}\nonumber\\
& +\frac{(\vec{p}\cdot\vec{G}_{i})(\vec{p}\cdot\vec{G}_{e})(T_{e}+T_{i
)}{\vec{p}\cdot(\vec{G}_{e}-\vec{G}_{i})e}-2i\alpha n_{0}.
\label{omega_k_general
\end{align}
Equations~(\ref{Omega_general}) to (\ref{omega_k_general}) give
the general expressions for the complex wave frequencies. To
find the combined linear growth of instabilities or wave
dissipation, $\gamma_{\vec{k}}=-i\,\operatorname{Im}
\omega_{\vec{k}}$, they should be further split into the real
and imaginary parts.
The approximate two-fluid description of local quasi-harmonic waves is valid
provided the characteristic wavevectors and frequencies satisfy
\begin{subequations}
\label{ner-vo
\begin{align}
& L_{\parallel,\perp}^{-1} \ll k_{\parallel,\perp}\ll l_{e,i}^{-1}
,\rho_{e,i}^{-1},\label{ner-vo_L}\\
& T^{-1},|\gamma
_{\vec{k}}| \ll|\omega_{\vec{k}}^{\prime}|\ll\nu_{e,i}. \label{ner-vo_T}
\end{align}
\end{subequations}
Here $l_{e,i}$ are the typical mean free paths of the
corresponding particles with respect to ion-neutral and
electron-neutral collisions, $\rho_{e,i}$ are the gyroradii (if
the corresponding particles are magnetized), and $T$ and
$L_{\parallel,\perp}$ are typical temporal and spatial
(parallel and perpendicular to $\vec{B}_{0}$) scales of
ionospheric density variation. Equation~(\ref{ner-vo_L}) should
be satisfied separately for the parallel and perpendicular
directions. Assuming these conditions, we will treat the wave
pressure gradients, local gradients in the background density,
and particle inertia as second-order effects with respect to
the small parameters defined by Eq.~(\ref{ner-vo}). In this
treatment, the first-order factors discussed in
\citet{Dimant:Magnetosphere2011_Budget} define the real wave
frequency, $\omega_{\vec{k}}^{\prime}$, while all second-order
factors combined define the total linear damping or growth
rate, $\gamma_{\vec{k}}$.
Under conditions specified by Eq.~(\ref{ner-vo}), the dominant
real part of the wave frequency, $\omega_{\vec{k}}^{\prime}$,
is determined by the first-order accuracy,
$\omega_{\vec{k}}^{\prime}\approx\omega_{\vec{k}1}$, when
neglecting the pressure gradients, particle inertia, regular
gradients of the background plasma density, and recombination.
To this accuracy, we have $\hat{K}
\approx\vec{p}\approx\vec{k}$,
$q\approx\vec{k}\cdot\vec{U}_{0}$,
$\tilde{\kappa}_{e,i}=\kappa_{e,i}$, and hence
\begin{subequations}
\label{G_1
\begin{align}
\vec{G}_{e\parallel} & \approx\vec{G}_{e1\parallel}=\frac{i\kappa_{e}\vec
{k}_{\parallel}}{B},\ \ \ \ \vec{G}_{e\perp}\approx\vec{G}_{e1\perp
=\frac{i\kappa_{e}(\vec{k}_{\perp}-\kappa_{e}\vec{k}\times\hat{b})}{\left(
1+\kappa_{e}^{2}\right) B},\label{G_1_e}\\
\vec{G}_{i\parallel} & \approx\vec{G}_{i1\parallel}=-\ \frac{i\kappa_{i
\vec{k}_{\parallel}}{B},\ \,\vec{G}_{i\perp}\approx\vec{G}_{i1\perp
}=-\ \frac{i\kappa_{i}(\vec{k}_{\perp}+\kappa_{i}\vec{k}\times\hat{b
)}{\left( 1+\kappa_{i}^{2}\right) B}, \label{G_1_i
\end{align
\end{subequations}
\begin{align}
\vec{p}\cdot\vec{G}_{e} & \approx\vec{k}\cdot\vec{G}_{e1}=\frac{i\kappa
_{e}k_{\perp}^{2}}{(1+\kappa_{e}^{2})B}\left[ 1+(1+\kappa_{e}^{2
)\frac{k_{\parallel}^{2}}{k_{\perp}^{2}}\right] ,\nonumber\\
\vec{p}\cdot\vec{G}_{i} & \approx\vec{k}\cdot\vec{G}_{i1}=-\ \frac
{i\kappa_{i}k_{\perp}^{2}}{(1+\kappa_{i}^{2})B}\left[ 1+(1+\kappa_{i
^{2})\frac{k_{\parallel}^{2}}{k_{\perp}^{2}}\right] . \label{proma
\end{align}
Using Eqs.~(\ref{psi}), we obtain
\begin{equation}
\vec{p}\cdot(\vec{G}_{e}-\vec{G}_{i})\approx\vec{k}\cdot(\vec{G}_{e1}-\vec
{G}_{i1})=\frac{i\kappa_{i}\kappa_{e}\left( \kappa_{e}+\kappa_{i}\right)
\left( 1+\psi_{\vec{k}}\right) k_{\perp}^{2}}{\left( 1+\kappa_{e
^{2}\right) \left( 1+\kappa_{i}^{2}\right) B}, \label{proma_2
\end{equation}
After neglecting small terms proportional to $\left( T_{e}+T_{i}\right) $
and $\alpha n_{0}$, Eq.~(\ref{omega_k_general}) yield
\[
\omega_{\vec{k}}\approx\omega_{\vec{k}1}=\frac{(\vec{k}\cdot\vec{G}_{e1
)(\vec{k}\cdot\vec{V}_{i0})-(\vec{k}\cdot\vec{G}_{i1})(\vec{k}\cdot\vec
{V}_{e0})}{\vec{k}\cdot(\vec{G}_{e1}-\vec{G}_{i1})
\]
and reduces to Eq.~(37) from
\citet{Dimant:Magnetosphere2011_Budget} with the corresponding
real shifted frequencies,
$\Omega_{\vec{k}1}^{(e,i)}\equiv\operatorname{Re}\Omega_{\vec{k}
}^{(e,i)}=\omega_{\vec{k}1}-\vec{k}\cdot\vec{V}_{e,i0}$
\begin{equation}
\Omega_{\vec{k}1}^{(e)}=-\ \frac{(\vec{k}\cdot\vec{U}_{0})(\vec{k}\cdot\vec
{G}_{e1})}{\vec{k}\cdot(\vec{G}_{e1}-\vec{G}_{i1})},\qquad\Omega_{\vec{k
1}^{(i)}=-\ \frac{(\vec{k}\cdot\vec{U}_{0})(\vec{k}\cdot\vec{G}_{i1})}{\vec
{k}\cdot(\vec{G}_{e1}-\vec{G}_{i1})}, \label{Omega_k_1
\end{equation}
given explicitly by the companion paper Eqs.~(36) and (39).
Now we will develop the dispersion relationships to the
second-order accuracy by taking into account all previously
neglected factors: the wave pressure gradients, particle
inertia, gradients of the zero-order plasma parameters, and
recombination. All second-order terms in
Eq.~(\ref{omega_k_general}), representing linear corrections to
the first-order real frequency $\omega_{\vec{k}1}$, are purely
imaginary, so that their linear combination determines the
total linear growth/damping rate, $\gamma_{\vec{k}}$.
We start by discussing the two last terms in the RHS of
Eq.~(\ref{omega_k_general}). The second-order term proportional to
$(T_{e}+T_{i})$ originates from perturbations of the particle fluid pressure.
To the leading-order accuracy, the term
\begin{align}
& \frac{(\vec{p}\cdot\vec{G}_{i})(\vec{p}\cdot\vec{G}_{e})(T_{e}+T_{i})
{\vec{p}\cdot(\vec{G}_{e}-\vec{G}_{i})e}\approx\frac{(\vec{k}\cdot\vec{G
_{i1})(\vec{k}\cdot\vec{G}_{e1})(T_{e}+T_{i})}{\vec{k}\cdot(\vec{G}_{e1
-\vec{G}_{i1})e}\nonumber\\
& =-\ \frac{ik_{\perp}^{2}\left[ 1+(1+\kappa_{e}^{2})k_{\parallel
^{2}/k_{\perp}^{2}\right] \left[ 1+(1+\kappa_{i}^{2})k_{\parallel
^{2}/k_{\perp}^{2}\right] (T_{e}+T_{i})}{(\kappa_{e}+\kappa_{i})(1+\psi
_{\vec{k}})eB} \label{T_ei_multiplier
\end{align}
contributes directly to $\gamma_{\vec{k}}$. The combination of the last two
terms describes the major linear wave dissipation due to particle diffusion
and recombination. For fully magnetized electrons, $\kappa_{e}\equiv\Omega
_{e}/\nu_{e}\gg1$, and unmagnetized ions, $\kappa_{i}\equiv\Omega_{i}/\nu
_{i}\ll1$, Eq.~(\ref{T_ei_multiplier}) reduces to the conventional loss term,
\[
\frac{(\vec{p}\cdot\vec{G}_{i})(\vec{p}\cdot\vec{G}_{e})(T_{e}+T_{i})}{\vec
{p}\cdot(\vec{G}_{e}-\vec{G}_{i})e}\approx-\ \frac{i\psi_{\vec{k}}k_{\perp
}^{2}C_{s}^{2}}{(1+\psi_{\vec{k}})\nu_{i}},
\]
where in this limit $\psi_{\vec{k}}\approx(1+\kappa_{e}^{2}k_{\parallel
^{2}/k_{\perp}^{2})\nu_{e}\nu_{i}/(\Omega_{e}\Omega_{i})$.
The first term in the RHS of Eq.~(\ref{omega_k_general}) includes the major
instability drivers. Unlike the two last terms, to the leading accuracy it is
a first-order term. To retrieve second-order corrections, we have to linearize
it with respect to small perturbations $\propto i\Omega_{\vec{k
}^{(e,i)}/\nu_{e,i}$ and $i\nabla$. Denoting the corresponding linear
corrections by $\delta(\cdots)$, we hav
\begin{align}
& \delta\left[ \frac{(\vec{p}\cdot\vec{G}_{e})(\hat{K}\cdot\vec{V
_{i0})-(\vec{p}\cdot\vec{G}_{i})(\hat{K}\cdot\vec{V}_{e0})}{\vec{p}\cdot
(\vec{G}_{e}-\vec{G}_{i})}\right] \nonumber\\
& \approx(\vec{k}\cdot\vec{V}_{i0})\delta\left[ \frac{\vec{p}\cdot\vec
{G}_{e}}{\vec{p}\cdot(\vec{G}_{e}-\vec{G}_{i})}\right] -(\vec{k}\cdot\vec
{V}_{e0})\delta\left[ \frac{\vec{p}\cdot\vec{G}_{i}}{\vec{p}\cdot(\vec{G
_{e}-\vec{G}_{i})}\right] \nonumber\\
& +\frac{(\vec{k}\cdot\vec{G}_{e1})(\delta\hat{K}\cdot\vec{V}_{i0})-(\vec
{k}\cdot\vec{G}_{e1})(\delta\hat{K}\cdot\vec{V}_{i0})}{\vec{k}\cdot(\vec
{G}_{e1}-\vec{G}_{i1})}. \label{del_pervyj
\end{align}
According to Eq.~(\ref{Kq}), in the last term of Eq.~(\ref{del_pervyj}) we
have $\delta\hat{K}=-i\nabla$. Presuming a uniform magnetic field, $\vec
{B}=B\hat{b}$, expressing $\vec{V}_{e,i0}$ in terms of $\vec{U}_{0}$ according
to Eqs.~(\ref{V_0_II_via_U}), (\ref{V_0_perp_via_U}), and using
Eqs.~(\ref{proma}), (\ref{proma_2}), we obtai
\begin{align}
& \frac{(\vec{k}\cdot\vec{G}_{e1})(\delta\hat{K}\cdot\vec{V}_{i0})-(\vec
{k}\cdot\vec{G}_{e1})(\delta\hat{K}\cdot\vec{V}_{i0})}{\vec{k}\cdot(\vec
{G}_{e1}-\vec{G}_{i1})}\nonumber\\
& =\frac{i}{\kappa_{e}+\kappa_{i}}\left[ \frac{(\kappa_{i}-\kappa_{e
)\nabla\cdot\vec{U}_{0}}{1+\psi_{k}}+\kappa_{e}\kappa_{i}\hat{b}\cdot
(\nabla\times\vec{U}_{0})\right] . \label{term_prop_delta_K
\end{align}
Recall that $\nabla\cdot\vec{U}_{0}$ can be expressed in terms of $\nabla
n_{0}$ according to Eq.~(\ref{div_U_0_via_grad}). In this sense, the term
proportional to $\nabla\cdot\vec{U}_{0}$ can be considered as an additional
small contributor to the GD instability driving. The last vortex term
$\propto\nabla\times\vec{U}_{0}$ is unrelated to any density gradients, but it
may also contribute to $\gamma_{\vec{k}}$.
Next we turn to the combination of the two first terms in the RHS of
Eq.~(\ref{del_pervyj}). The difference between the two fractions $\left.
\vec{p}\cdot\vec{G}_{e}\right/ \vec{p}\cdot(\vec{G}_{e}-\vec{G}_{i})$ and
$\left. \vec{p}\cdot\vec{G}_{i}\right/ \vec{p}\cdot(\vec{G}_{e}-\vec{G
_{i})$ equals a constant value of $1$, so that their linear
perturbations are equal and we have
\begin{subequations}
\begin{eqnarray}
&&(\vec{k}\cdot\vec{V}_{i0})\ \delta\left[ \frac{\vec{p}\cdot\vec{G}_{e}}{\vec
{p}\cdot(\vec{G}_{e}-\vec{G}_{i})}\right]\nonumber\\
&-&(\vec{k}\cdot\vec{V}_{e0})\ \delta\left[
\frac{\vec{p}\cdot\vec{G}_{i}}{\vec{p}\cdot(\vec{G}_{e}-\vec
{G}_{i})}\right] =-\beta(\vec{k}\cdot\vec{U}_{0}), \label{via_beta
\end{eqnarray}
\end{subequations}
where
\begin{subequations}
\label{betas
\begin{align}
\beta & \equiv\delta\left[ \frac{\vec{p}\cdot\vec{G}_{e}}{\vec{p}\cdot
(\vec{G}_{e}-\vec{G}_{i})}\right] =\delta\left[ \frac{\vec{p}\cdot\vec
{G}_{i}}{\vec{p}\cdot(\vec{G}_{e}-\vec{G}_{i})}\right] =\beta_{\mathrm{FB
}+\beta_{\mathrm{GD}},\label{beta_full}\\
\beta_{\mathrm{FB}} & \approx\frac{(\vec{k}\cdot\vec{G}_{e1})(\vec{k
\cdot\delta\vec{G}_{i})-(\vec{k}\cdot\vec{G}_{i1})(\vec{k}\cdot\delta\vec
{G}_{e})}{[\vec{k}\cdot(\vec{G}_{e1}-\vec{G}_{i1})]^{2}},\label{beta_FB}\\
\beta_{\mathrm{GD}} & \approx\frac{(\vec{k}\cdot\vec{G}_{e1})(\vec{G
_{i1}\cdot\delta\vec{p})-(\vec{k}\cdot\vec{G}_{i1})(\vec{G}_{e1}\cdot
\delta\vec{p})}{[\vec{k}\cdot(\vec{G}_{e1}-\vec{G}_{i1})]^{2}}.
\label{beta_GD
\end{align}
\end{subequations}
Here the coefficient $\beta_{\mathrm{FB}}$ includes particle-inertia
contributions responsible for the FB instability, while $\beta_{\mathrm{GD}}$
includes the background density gradients responsible for the GD instability.
First, we calculate the FB-instability driving term, $-\beta_{\mathrm{FB
}(\vec{k}\cdot\vec{U}_{0})$. According to Eqs.~(\ref{G}) and
Eq.~(\ref{tilde_kappa}), we obtain
\begin{align*}
\vec{k}\cdot\delta\vec{G}_{e} & =-\ \frac{\kappa_{e}k_{\perp}^{2}}{B}\left[
\frac{1-\kappa_{e}^{2}}{(1+\kappa_{e}^{2})^{2}}+\frac{k_{\parallel}^{2
}{k_{\perp}^{2}}\right] \frac{\Omega_{\vec{k}1}^{(e)}}{\nu_{e}},\\
\vec{k}\cdot\delta\vec{G}_{i} & =\frac{\kappa_{i}k_{\perp}^{2}}{B}\left[
\frac{1-\kappa_{i}^{2}}{(1+\kappa_{i}^{2})^{2}}+\frac{k_{\parallel}^{2
}{k_{\perp}^{2}}\right] \frac{\Omega_{\vec{k}1}^{(i)}}{\nu_{i}}.
\end{align*}
Using Eq.~(\ref{Omega_k_1}), we have $\vec{k}\cdot\vec{G}_{e,i1}$
\begin{align}
\vec{k}\cdot\vec{G}_{e1} & =-\ \frac{\vec{k}\cdot\left( \vec{G}_{e1}-\vec
{G}_{i1}\right) \Omega_{\vec{k}1}^{(e)}}{\vec{k}\cdot\vec{U}_{0}},\nonumber\\
\vec{k}\cdot\vec{G}_{i1} & =-\ \frac{\vec{k}\cdot\left( \vec{G}_{e1}-\vec
{G}_{i1}\right) \Omega_{\vec{k}1}^{(i)}}{\vec{k}\cdot\vec{U}_{0}},\nonumber
\end{align}
where according to Eq.~(\ref{proma_2})
\[
\vec{k}\cdot(\vec{G}_{e1}-\vec{G}_{i1})=\frac{i\kappa_{i}\kappa_{e}\left(
\kappa_{e}+\kappa_{i}\right) \left( 1+\psi_{\vec{k}}\right) k_{\perp}^{2
}{\left( 1+\kappa_{e}^{2}\right) \left( 1+\kappa_{i}^{2}\right) B}.
\]
As a result, we obtain from Eq.~(\ref{beta_FB}
\begin{align}
& -\beta_{\mathrm{FB}}(\vec{k}\cdot\vec{U}_{0})\nonumber\\
& =\frac{i\Omega_{\vec{k}1}^{(e)}\Omega_{\vec{k}1}^{(i)}}{\left( \kappa
_{e}+\kappa_{i}\right) \left( 1+\psi_{\vec{k}}\right)
\label{beta_FB_final}\\
& \times\!\!\left[ \frac{(1-\kappa_{i}^{2})(1+\kappa
_{e}^{2}) }{(1+\kappa_{i}^{2}) \kappa_{e}\nu_{i}}
+\frac{(1-\kappa_{e}^{2})(1+\kappa_{i}^{2})
}{(1+\kappa_{e}^{2}) \kappa_{i}\nu_{e}}+\left( \frac{\kappa
_{e}}{\nu_{e}}+\frac{\kappa_{i}}{\nu_{i}}\right) \frac{k_{\parallel}^{2
}{k_{\perp}^{2}}\right]\!\! .\nonumber
\end{align}
Here the first and second terms in the square bracket originate
from the ion and electron inertia, respectively. These terms
correspond to particle oscillation in the perpendicular to
$\vec{B}_{0}$ plane. Notice that for $\kappa_{e}>1$, the
electron inertia terms becomes negative, meaning that it
opposes the FB instability. The physical nature of this
electron-inertia stabilization is fully analogous to that for
the ion inertia above the magnetization boundary,
$\kappa_{i}>1$ \citep{DimOppen2004:ionthermal1}. The last term
in the square bracket of Eq.~(\ref{beta_FB_final}) includes the
effects of the electron ($\kappa_{e}/\nu_{e}$) and ion
($\kappa_{i}/\nu_{i}$) inertia in the parallel to $\vec{B}_{0}$
direction. The second and third terms in the square bracket are
usually neglected, but the conditions for such neglect have not
being properly analyzed and justified. We do this below.
Similarly, we calculate the term $-\beta_{\mathrm{GD}}(\vec{k}\cdot\vec{U
_{0})$ that describes the local gradient drift instability or
stabilization. According to Eq.~(\ref{p}),
$\delta\vec{p}=-i\nabla n_{0}/n_{0}$, and we obtain from
Eq.~(\ref{beta_GD}),
\begin{align}
& -\beta_{\mathrm{GD}}(\vec{k}\cdot\vec{U}_{0})=\frac{i(1+\kappa_{e
^{2})(1+\kappa_{i}^{2})(\vec{k}\cdot\vec{U}_{0})}{\kappa_{e}\kappa_{i
(\kappa_{e}+\kappa_{i})(1+\psi_{k})^{2}k_{\perp}^{2}}\nonumber\\
& \times\left\{ (\kappa_{e}-\kappa_{i})\left( \vec{k}_{\parallel}\cdot
\frac{\nabla_{\parallel}n_{0}}{n_{0}}-\vec{k}_{\perp}\cdot\frac{\nabla_{\perp
}n_{0}}{n_{0}}\ \frac{k_{\parallel}^{2}}{k_{\perp}^{2}}\right) \right.
\label{beta_GD_final}\\
& -\left. \left[ 1+(1+\kappa_{i}\kappa_{e})\frac{k_{\parallel}^{2
}{k_{\perp}^{2}}\right] (\vec{k}\times\hat{b})\cdot\frac{\nabla_{\perp}n_{0
}{n_{0}}\right\} .\nonumber
\end{align}
To summarize, we obtain for the combined FB and GD linear growth rate the
general expressio
\begin{equation}
\gamma_{\vec{k}}=\gamma_{\mathrm{FB}}+\gamma_{\mathrm{GD}}-\gamma_{T}-2\alpha
n_{0}, \label{gamma_total
\end{equation}
wher
\begin{align}
\gamma_{\mathrm{FB}} & =-\ \frac{\Omega_{\vec{k}1}^{(e)}\Omega_{\vec{k
1}^{(i)}}{\kappa_{e}\left( 1+\psi_{\vec{k}}\right) }\left[ \frac{\left(
1-\kappa_{i}^{2}\right) \left( 1+\kappa_{e}^{2}\right) }{\left(
1+\kappa_{i}^{2}\right) \kappa_{e}\nu_{i}}\right. \nonumber\\
& \left. +\frac{\left( 1-\kappa_{e}^{2}\right) \left( 1+\kappa_{i
^{2}\right) }{\left( 1+\kappa_{e}^{2}\right) \kappa_{i}\nu_{e}}+\left(
\frac{\kappa_{e}}{\nu_{e}}+\frac{\kappa_{i}}{\nu_{i}}\right) \frac
{k_{\parallel}^{2}}{k_{\perp}^{2}}\right] , \label{gamma_FB
\end{align
\begin{align}
& \gamma_{\mathrm{GD}}=\frac{(1+\kappa_{e}^{2})(1+\kappa_{i}^{2})(\vec
{k}\cdot\vec{U}_{0})}{\kappa_{e}\kappa_{i}(\kappa_{e}+\kappa_{i})(1+\psi
_{k})^{2}k_{\perp}^{2}}\nonumber\\
& \times\left\{ (\kappa_{e}-\kappa_{i})\left( \vec{k}_{\parallel}\cdot
\frac{\nabla_{\parallel}n_{0}}{n_{0}}-\vec{k}_{\perp}\cdot\frac{\nabla_{\perp
}n_{0}}{n_{0}}\ \frac{k_{\parallel}^{2}}{k_{\perp}^{2}}\right) \right.
\nonumber\\
& -\left. \left[ 1+(1+\kappa_{i}\kappa_{e})\frac{k_{\parallel}^{2
}{k_{\perp}^{2}}\right] (\vec{k}\times\hat{b})\cdot\frac{\nabla_{\perp}n_{0
}{n_{0}}\right\} , \label{Gamma_GD
\end{align
\begin{equation}
\gamma_{T}=\frac{k_{\perp}^{2}\left[ 1+(1+\kappa_{e}^{2})k_{\parallel
^{2}/k_{\perp}^{2}\right] \left[ 1+(1+\kappa_{i}^{2})k_{\parallel
^{2}/k_{\perp}^{2}\right] (T_{e}+T_{i})}{(\kappa_{e}+\kappa_{i})(1+\psi
_{\vec{k}})eB}, \label{gamma_T
\end{equation}
and the first-order shifted frequencies $\Omega_{\vec{k}1}^{(e,i)}$ are given
by Eqs.~(36) and (39) from \citet{Dimant:Magnetosphere2011_Budget}.
For magnetized electrons at a sufficiently high altitude where
$\kappa_{e}\gg1$ and
$\psi_{\perp}\equiv(\kappa_{e}\kappa_{i})^{-1}\ll
\nu_{e}/\nu_{i}\simeq10$ hold together (see below), presuming
$|k_{\parallel}|/k_{\perp
}\lesssim\kappa_{e}^{-1}\lll\kappa_{i}^{-1}$, we obtain
\begin{equation}
\Omega_{\vec{k}1}^{(i)}\approx\frac{\vec{k}\cdot\vec{U}_{0}}{1+\psi_{\vec{k}
},\qquad\Omega_{\vec{k}1}^{(e)}\approx-\psi_{\vec{k}}(1+\kappa_{i}^{2
)\Omega_{\vec{k}1}^{(i)} \label{Omegi_reduced
\end{equation}
so tha
\begin{equation}
\gamma_{\mathrm{FB}}\approx\frac{\psi_{\vec{k}}\left( 1-\kappa_{i
^{2}\right) (\Omega_{\vec{k}1}^{(i)})^{2}}{\left( 1+\psi_{\vec{k}}\right)
\nu_{i}},\qquad\gamma_{T}\approx\frac{\psi_{\vec{k}}k_{\perp}^{2}C_{s}^{2
}{(1+\psi_{\vec{k}})\nu_{i}}, \label{gamma_FB_reduced
\end{equation
\begin{equation}
\gamma_{\mathrm{GD}}=\frac{(1+\kappa_{i}^{2})\Omega_{\vec{k}1}^{(i)}
{\kappa_{i}(1+\psi_{\vec{k}})k_{\perp}^{2}}\left[ \kappa_{e}\vec
{k}_{\parallel}\cdot\frac{\nabla_{\parallel}n_{0}}{n_{0}}-(\vec{k}\times
\hat{b})\cdot\frac{\nabla_{\perp}n_{0}}{n_{0}}\right] ,
\label{gamma_GD_reduced
\end{equation}
where $C_{s}^{2}\equiv(T_{e}+T_{i})/m_{i}$.
Equation~(\ref{gamma_FB_reduced}) shows that for fluid
particles ion inertia is a destabilizing factor
($\gamma_{\mathrm{FB}}>0$) for all altitudes below the
magnetization boundary defined by $\kappa_{i}=1$. For higher
altitudes with $\kappa_{i}>1$, ion inertia is a stabilizing
factor, $\gamma_{\mathrm{FB}}\leq0$. The physical nature of
this peculiar feature has been discussed in
\citet{DimOppen2004:ionthermal1}. Note that the multipliers
$\left( 1+\psi_{\vec{k}}\right) $ in
Eqs.~(\ref{Omegi_reduced}) to (\ref{gamma_GD_reduced}) differ
from the corresponding multipliers in the well-known
expressions \citep{Fejer:Theory84} with the traditional
definition of $\psi$ by a term
$\psi_{\perp}\kappa_{i}^{2}=\kappa_{i}/\kappa_{e}\equiv\theta
_{0}^{2}\simeq1.8\times10^{-4}$, that can be neglected.
Now we discuss some of the conventionally neglected terms in
the linear dispersion relation. The second and third terms in
the square bracket of Eq.~(\ref{beta_FB_final}) have never been
taken into account. In the case under consideration, the ratio
of the second term to the first one is
$\nu_{i}/(\kappa_{i}\kappa_{e}\nu_{e})=\psi_{\perp}(\nu_{i}/\nu_{e})$.
In the lower ionosphere, we have $\nu_{i}/\nu_{e}\simeq0.1$, so
that the electron inertia in the perpendicular to $\vec{B}_{0}$
plane can be neglected provided $\psi_{\perp}\ll10$. This
condition is fulfilled well above the 90~km altitude, i.e., in
essentially within the entire altitude range of a possible FB
instability development. At lower altitudes (the upper
\textit{D} region), the electron inertia can become an
additional stabilizing factor for the electron thermal driven
instability \citep{Dimant:Physical97}.
For the third term in the square brackets of Eq.~(\ref{beta_FB_final}), the
ratio of its two sub-terms is $(\kappa_{e}/\nu_{e})/(\kappa_{i}/\nu
_{i})=(m_{i}/m_{e})(\nu_{i}/\nu_{e})^{2}\simeq550$, so that in the parallel to
$\vec{B}_{0}$ direction it is the electron inertia term $\propto\kappa_{e}/\nu
_{e}$ that largely dominates. Then the ratio of the third term to the first
one becomes $(\nu_{e}/\nu_{i})k_{\parallel}^{2}/k_{\perp}^{2}$ $\simeq
10(k_{\parallel}/k_{\perp})$. It is of order unity or larger for the wave
aspect angles, $\theta\equiv\arctan(k_{\parallel}/k_{\perp})\gtrsim20^{\circ
$. Thus, particle inertia in the parallel to $\vec{B}_{0}$
direction, which is largely due to electrons, can only affect
waves whose wavevectors lie well off the perpendicular plane,
usually outside the angular domain of linear instabilities.
Such waves are heavily damped and hardly play a role in any
linear or non-linear processes. This means that for waves in
the lower ionosphere the last term in the square bracket of
Eq.~(\ref{beta_FB_final}) is probably never of importance.
\begin{acknowledgments}
This work was supported by National Science Foundation
Ionospheric Physics Grants No.~ATM-0442075, ATM-0819914 and
ATM-1007789. The authors thank Dr. V.~G.~Merkin for useful
discussions of conductance issues in global MHD models.
\end{acknowledgments}
|
\subsection{Relation to the vertex splitting model} We now briefly introduce the vertex splitting model and show how the ag--model can be seen as a special case. The parameters of the vertex splitting model are given by a set of non--negative weights $w_{i,j}$, with $1 \leq i,j,i+j-2 \leq D$, where $D\geq 2$ (or $D=\infty$) is a fixed number which denotes the maximum vertex degree in the trees. These weights are referred to as {\it partitioning weights} and the so--called {\it splitting weights} are defined by
\begin{equation}
w_i = \frac{i}{2} \sum_{j=1}^{i+1} w_{j,i+2-j},\qquad 1 \leq i \leq D.
\end{equation}
Starting from a fixed finite planar tree, in each discrete time step a new edge is added as follows.
\begin {list}{\labelitemi}{\leftmargin=2em}
\item [(a)] Select a given vertex $v$ of degree $i$ with relative probability $w_i$.
\item [(b)] Randomly partition the edges which contain $v$ into two disjoint sets of adjacent edges: $V$ of size $k-1$ and $V'$ of size $i-(k-1)$, with probability $w_{k,i+2-k}/w_i$. For a given $k$, all such partitionings are taken to be equally likely.
\item [(c)] Move all edges in $V'$ from $v$ to a new vertex $v'$ and join $v$ and $v'$ by a new edge.
\end {list}
We allow a small generalization of the above growth rule: single out a vertex of degree one in the initial tree and call it the root and modify (a) in such a way that the root is selected with relative probability $w_r \geq 0$. Each time the root is split we define the new vertex of degree one to be the root. The vertex splitting model has very general growth rules and it includes many other models of random trees as special cases or limiting cases, see \cite{vs,phdsos} for more detailed discussion.
The ag--model can be recovered from the vertex splitting model by assigning the weight $w_r = \alpha/2$ to splitting the root and choosing the nonzero partitioning weights as follows
\begin {eqnarray} \nonumber
w_{2,2} &=& \alpha, \\ \nonumber
w_{k,2} &=& \frac{\alpha}{2}, \quad 3\leq k \leq D, \\ \label{partw}
w_{k+1,1} &=& \frac{\overline{w}_{k}}{k}, \quad 2 \leq k \leq D.
\end {eqnarray}
The splitting weights are then
\begin {equation} \label{splittingweights}
w_k = \left(\frac{\alpha}{2} + a\right) k + 1 - 2a - \alpha, \quad 2 \leq k \leq D
\end {equation}
and $w_k = 0$ if $k>D$. Note that in the case $D=\infty$ and $\gamma < \alpha/2$, the weight $w_1$ is negative. We will however include this case in the ag--model since the total weight of any transition is still positive.
A similar relationship between the $\alpha\gamma$--model and the vertex splitting model was discussed in \cite{vs}. However, in that case one needs to take $w_{3,1} = \infty$ which means that comparison of results in the two models is not necessarily reliable. The ag--model is therefore more interesting as a special case and due to its simplicity, yet non--triviality, it serves as a good testing ground for non--rigorous results obtained in the vertex splitting model.
\subsection{Outline} The paper is organized as follows. In Section \ref{s:rpt} we define rooted planar trees and introduce a convenient notation for representing random trees. Thereafter we give a proper definition of the ag--model which was described informally above. In Section \ref{s:mb} we show that the model has the Markov branching property and we calculate its first split distribution. In Section \ref{s:cfvm} we show, using methods from \cite{siggi}, that the finite volume probability measures generated by the random growth operation, converge to a measure on the set of infinite trees. Furthermore, we characterize the infinite volume measure. In Section \ref{s:hd} we calculate the annealed Hausdorff dimension with respect to the infinite volume measure and in a certain special case, we calculate the almost sure Hausdorff and spectral dimensions. The results we obtain, support certain scaling assumptions which were made in the vertex splitting model. We conclude by commenting on the distribution of the degrees of vertices in the trees and correlations between degrees of neighbouring vertices by recalling results from \cite{vs}. In order to improve readability, proofs of theorems and lemmas are in most cases collected in Appendix B.
\section {Random planar trees} \label{s:rpt}
In this section we begin by defining the set of rooted, planar trees and endow it with a metric. Then we define a convenient notation for representing random trees and introduce the model which will be studied in the paper.
Start with a tree graph $\tau$ which has vertices of finite or countably infinite degree and at least one vertex of degree one. By convention we define the root $r$ of $\tau$ to be a vertex of degree one and we label the unique nearest neighbour of the root by $(1)$. The rest of the vertices are labeled in the following recursive way. The children of a given vertex in the tree (apart from $r$) with label $(\ell)$ are labeled with sequences $(\ell,1), (\ell, 2),\ldots$, see Fig.~\ref{f:planartree}. A rooted planar tree is a tree $\tau$ along with such a lexicographical labeling. From here on, we will always work with rooted, planar trees unless otherwise stated and will simply refer to them as trees. We denote the set of trees with $n$ edges by $\mathcal{T}_n$ and the set of all trees, finite and infinite, by $\mathcal{T}$.
\begin{figure} [!h]
\centerline{\scalebox{0.8}{\includegraphics{planar.eps}}}
\caption{Left: An example of a rooted, planar tree $\tau$ and a left subtree $\tau_0$ (boxed in gray). Right: The graph ball $B_3(\tau)$ and the left ball $L_3(\tau)$ (boxed in gray). The root is indicated by a circled vertex.} \label{f:planartree}
\end{figure}
A tree $\tau_0$ is said to be a {\it left subtree} of $\tau$ if it is a connected subtree of $\tau$ which contains $r$ and has the properties that if it contains a vertex with label $(\ell,k)$ then it contains all vertices with labels $(\ell,i)$ with $i\leq k$, see Fig.~\ref{f:planartree}. Let $B_R(\tau)$ be the graph ball of radius $R$ centered on the root of $\tau$. We define the {\it left ball} of radius $R$, $L_R(\tau)$, as the maximal left subtree of $B_R(\tau)$ with vertices of degree no greater than $R$, see Fig.~\ref{f:planartree}. A metric $d$ is defined on $\mathcal{T}$ by
\begin {equation}
d(\tau,\tau') = \inf\left\{\frac{1}{R}~\Bigg| ~ L_R(\tau) = L_R(\tau')\right\}.
\end {equation}
The metric $d$ was first introduced in \cite{ngtrees} and we refer to this paper for some properties of the metric space $(\mathcal{T},d)$.
Define the root joining operation $\ast$ in the following way. Given trees $\tau_1, \tau_2, \ldots, \tau_{k}$, $k\geq 1$, let $\tau = \emptyset \ast \tau_1\ast \tau_2 \ast \cdots \ast \tau_k$ be the tree obtained by (I) identifying the roots of $\tau_1, \tau_2,\ldots,\tau_k$ and labeling them by (1), (II) replacing the first element '$1$' of each label in $\tau_j$ by '$1,j$', $1\leq j \leq k$, and (III) connecting a new root $r$ to the vertex (1). If $k>1$ we may omit the $\emptyset$ symbol, see Fig.~\ref{F_split_0}. Note that in general
\begin {equation*}
\tau_1\ast\tau_2 \ast \cdots \ast \tau_k \neq \tau_{\sigma(1)}\ast\tau_{\sigma(2)} \ast \cdots \ast \tau_{\sigma(k)}
\end {equation*}
for a permutation $\sigma$ of $\{1,2,\ldots,k\}$.
\begin{figure} [!h]
\centerline{\scalebox{0.6}{\includegraphics{joining.eps}}}
\caption{The root joining operation.} \label{F_split_0}
\end{figure}
Let $\pi_n$ be a probability distribution on $\mathcal{T}_n$. We define a random tree $T_n$ by the canonical probability generating function
\begin {equation} \label{rt}
T_n = \sum_{\tau \in \mathcal{T}_n} \pi_n(\tau) \tau.
\end {equation}
The above sum of trees and multiplication of trees by a scalar are formal and provide a convenient way of storing information on the probability measure $\pi_n$.
\subsection{The ag--model}
Using the notation introduced above, we now define the ag--model which was described informally in the introduction. Let
\begin {equation}
W(n) = n-a.
\end {equation}
We introduce a growth operation $I$ in the following recursive way. Let $s$ be the single edge tree and define $I(s) = \emptyset \ast s$. For a tree $\tau = \emptyset \ast \tau_1 \ast \cdots \ast \tau_{k-1}$ define the random tree
\begin {eqnarray} \nonumber
I(\tau) &=& \frac{1}{W(|\tau|)} \Bigg( \sum_{i=1}^{k-1} W(|\tau_i|) ~\emptyset \ast \tau_1 \ast \cdots \ast I(\tau_i) \ast \cdots \ast \tau_{k-1} \\ \nonumber
&& +~ \alpha ~\emptyset \ast \tau + \frac{\overline{w}_k}{
k} \sum_{i=1}^k \tau_1 \ast \cdots \ast \tau_{i-1}\ast s \ast \tau_{i} \ast \cdots \ast \tau_{k-1} \Bigg) \\
\end {eqnarray}
where $|\tau|$ denotes the number of edges in $\tau$. The growth operation $I$ is equivalent to the growth rule which was described informally in Fig.~\ref{f:growth} (left) in the introduction. The ag--model is defined recursively as the random tree $P_{n}$ which satisfies $P_{1} = s$ and
\begin{equation} \label{themodel}
P_{n} = I(P_{n-1}).
\end {equation}
We denote the probability measure on $\mathcal{T}_n$, generated by this growth process by $\nu_{n}$.
\section {Markov branching} \label{s:mb}
A sequence of random trees $(T_n)_{n \geq 1}$ is said to satisfy a Markov branching property, or to be Markovian self--similar, if there exist functions $q_k(n_1,\ldots,n_{k-1})$, $k\geq 2$ such that for all $n \geq 2$
\begin {equation} \label{mbproperty}
T_{n} = \sum_{k=2}^\infty \sum_{n_1+\cdots+n_{k-1} = n-1} q_k(n_1,\ldots,n_{k-1})~\emptyset \ast T_{n_1} \ast \cdots \ast T_{n_{k-1}}.
\end {equation}The functions $q_k(n_1,\ldots,n_{k-1})$ are referred to as the {\it first split distribution} of $(T_n)_{n\geq 1}$. We use the convention that $q_k(n_1,\ldots,n_{k-1}) = 0$ if any of the arguments $n_1,\ldots,n_{k-1}$ equals zero.
\begin {proposition} \label{p:MB1}
The random trees $(P_{n})_{n\geq 1}$, defined by (\ref{themodel}), have the Markov branching property with a first split distribution which satisfies $q_2(1) = 1$,
\begin {equation}
q_2(n) = \frac{1}{W(n)}\left(W(n-1) q_2(n-1) +\alpha\right),
\end {equation}
\begin {eqnarray} \nonumber
q_k(n_1,\ldots,n_{k-1}) &=& \frac{1}{W(n)} \sum_{i=1}^{k-1}\Bigg(W(n_i-1) q_k(n_1,\ldots,n_i-1,\ldots,n_{k-1}) \\ \nonumber
&& + ~ \frac{\overline{w}_{k-1}}{k-1} \delta_{n_i,1} q_{k-1}(n_1,\ldots,n_{i-1},n_{i+1},\ldots,n_{k-1})\Bigg) \\ \label{qrec2}
\end {eqnarray}
for $3 \leq k \leq D$ and
\begin {equation}
q_{D+1}(n_1,\ldots,n_D) = 0
\end {equation}
where $n_1 + \cdots + n_{k-1} = n$.
\end {proposition}
\begin{proof}
We use induction on $n$. $P_n$ clearly satisfies (\ref{mbproperty}) for $n=2$. Assume it satisfies (\ref{mbproperty}) for some $n$. Then
\begin {eqnarray*}
P_{n+1} &=& \sum_{k=2}^\infty \sum_{n_1+\cdots+n_{k-1} = n-1} q_{k}(n_1,\ldots,n_{k-1}) \frac{1}{W(n)} \Bigg({\alpha}~ \emptyset \ast(\emptyset\ast P_{n_1}\ast \cdots \ast P_{n_{k-1}})\\
&& +~ \frac{\overline{w}_k}{k} \sum_{i=1}^k P_{n_{1}} \ast \cdots \ast P_{n_{i-1}}\ast s \ast P_{n_{i}} \ast \cdots \ast P_{n_{k-1}} \\
&& +~\sum_{i=1}^{k-1} {W(n_i)} ~\emptyset \ast P_{n_1} \ast \cdots \ast I(P_{n_i}) \ast \cdots \ast P_{n_{k-1}}\Bigg)\\
&=& \frac{1}{W(n)}\left(\alpha + W(n-1)q_2(n-1)\right) \emptyset\ast P_n \\
&&~+ \sum_{k=3}^\infty \sum_{n_1+\cdots+n_{k-1} = n}\frac{1}{W(n)}\sum_{i=1}^{k-1} \Bigg(\frac{\overline{w}_{k-1}}{k-1} \delta_{n_i,1}q_{k-1}(n_1,\ldots,n_{i-1},n_{i+1},\ldots,n_{k-1}) \\
&& +~ W(n_i-1) q_{k}(n_1,\ldots,n_i-1,\ldots,n_{k-1})\Bigg) \emptyset\ast P_{n_1} \ast \cdots \ast P_{n_{k-1}}.\\
\end {eqnarray*}
This shows that (\ref{mbproperty}) also holds for $n+1$ and we conclude that it holds for all $n\geq 2$.
\end {proof}
The recursions for the first split distribution in Proposition \ref{p:MB1} can be solved with straightforward methods. We state the result in the following proposition which can easily be proved by induction. The method for finding the solution is described in Appendix A.
\begin {proposition}
The first split distribution of the sequence $(P_n)_{n\geq1}$ is given by
\begin {eqnarray} \nonumber
q_k(n_1,\ldots,n_{k-1}) &=& \frac{ \Gamma\left(k-2+\frac{1-\alpha}{a}\right)}{\Gamma\left(\frac{1-\alpha}{a}\right)\Gamma(k)} \frac{\G{1-a}\G{n+1}}{a\Gamma\left(n+1-a\right)}\prod_{i=1}^{k-1} \frac{a\Gamma(n_i-a)}{\G{1-a}\G{n_i+1}} \\
&& \times ~\left(1-a - \alpha + \alpha \sum_{i=1}^{k-1}\frac{n_i}{n+1-n_i}\right) \nonumber \\ \label{qfuncgengen}
\end {eqnarray}
where $n_1 + \cdots + n_{k-1} = n$.
\end {proposition}
We will repeatedly use the following standard, easily derived identities when we work with the above first split distribution \cite{abram}
\begin {equation} \label{sumofgamma}
\sum_{n=1}^\infty \frac{c\Gamma(n-c)}{\G{1-c}\G{n+1}} z^{n} = 1-(1-z)^c
\end {equation}
and
\begin {equation} \label{ordergamma}
\frac{\Gamma(n-c)}{\G{n+1}} = n^{-c-1}\left(1+O\left(n^{-1}\right)\right).
\end {equation}
\section{Convergence of the finite volume measures} \label{s:cfvm}
In this section we show that the measures $\nu_n$ generated by the growth process converge weakly to a measure $\nu$ on the set of infinite trees. By weak convergence we mean that for all bounded functions $f$ which are continuous in the topology generated by the metric $d$
\begin {equation}
\int_{\mathcal{T}} f(\tau)d\nu_{n} \longrightarrow \int_{\mathcal{T}} f(\tau)d\nu, \quad\quad \text{as $n \longrightarrow \infty$.}
\end {equation}
We will call an infinite non--backtracking path from the root, a {\it spine}. Let $\tau$ be a tree with exactly one spine and let $v$ be a vertex on the spine ($v\neq r$) with degree $k$. We call the $k-2$ finite subtrees of $\tau$ which are attached to the vertex $v$ {\it outgrowths} from the spine.
\begin {theorem} \label{th:conv}
Let $\alpha > 0$. The measures $\nu_{n}$, viewed as probability measures on $\mathcal{T}$, converge weakly, as $n \longrightarrow \infty$, to a probability measure $\nu$ which is concentrated on the set of trees that have exactly one spine. The degrees of the vertices on the spine are independently distributed by
\begin {equation} \label{phiinfgen}
\phi(k) = \frac{\alpha \G{\frac{1+a}{a}}\G{k-2+\frac{1-\alpha}{a}}}{\G{\frac{1-\alpha}{a}}\G{k-1+\frac{1}{a}}}, \qquad k \geq 2.
\end {equation}
The outgrowths from the spine are finite with probability one and outgrowths from different vertices are independently distributed. If a vertex $v$ on the spine has degree $k$ and $\tau_1,\ldots,\tau_{L}$ are the outgrowths from $v$ to the left of the spine (in that order) and $\tau_{L+1},\ldots,\tau_{k-2}$ are the outgrowths from $v$ to the right of the spine (in that order), then their joint distribution is
\begin {equation} \label{muinfgen}
\mu_k(\tau_1,\ldots,\tau_{k-2}) = \frac{\Gamma\left(k-1+\frac{1}{a}\right)}{\Gamma\left(\frac{1+a}{a}\right)\Gamma(k-1)}\frac{1}{1+m}\prod_{i=1}^{k-2} \frac{a \Gamma(|\tau_i|-a)}{\G{1-a}\Gamma(|\tau_i|+1)}\nu_{|\tau_i|}(\tau_i),
\end {equation}
where $m = |\tau_1| + \cdots + |\tau_{k-2}| \geq k-2$ and $k\geq 3$.
\end {theorem} \noindent
A proof to the above theorem is given on page \pageref{pr:conv} in Appendix B.
We point out that the distributions $\mu_k$ are independent of how many of the outgrowths are to the left or to the right of the spine. For an ordered sequence of $k-2$ outgrowths, there are $k-1$ different ways to arrange them around the spine.
Below, we comment on some special cases. When $D < \infty$, $\phi(k) = 0$ for $k > D$ as it should be. In the case $D=3$ and $\alpha = 1/2$ the trees are generic, i.e.~ $\nu_n$ is a critical Galton--Watson process conditioned to have $n$ edges. This follows from the fact that the first split distributions can, in this case, be written as
\begin {equation}
q_2(n) = \frac{1}{2}\frac{Z_n}{Z_{n+1}} \qquad \text{and} \qquad q_3(n_1,n_2) = \frac{1}{4}\frac{Z_{n_1} Z_{n_2}}{Z_{n_1+n_2+1}}
\end {equation}
with
\begin {equation}
Z_n = \frac{\G{n+\frac{1}{2}}}{\sqrt{\pi}\G{n+2}}.
\end {equation}
$Z_n$ can be interpreted as a finite volume partition function corresponding to branching weights $w_1 = w_3 = 1/4$ and $w_2 = 1/2$, see e.g.~\cite{sdgt}. Furthermore, this is the only special case in which we obtain generic trees. This can be seen from the fact that when $D>3$, outgrowths from the same vertex on the spine are dependent.
When $D=\infty$ and $\gamma = 1$,
\begin {equation} \label{eq:degexp}
\phi(k) = \alpha (1-\alpha)^{k-2}
\end {equation}
and it falls off exponentially in $k$. When $D=\infty$ and $\gamma < 1$ we find that for large $k$
\begin {equation} \label{eq:degpower}
\phi(k) = \frac{\alpha \G{\frac{2-\gamma}{1-\gamma}}}{\G{\frac{1-\alpha}{1-\gamma}}} k^{-1-\frac{\alpha}{1-\gamma}}\left(1+O\left(k^{-1}\right)\right)
\end {equation}
i.e.~it falls off with a power law in $k$. From the last formula, we see that when $\alpha \leq 1-\gamma$, the expected value of the degree of a vertex on the spine is infinite. A simple and interesting special case arises when $\gamma = 0$ in which case the outgrowths from the spine are single leaves. Such graphs have been referred to as {\it caterpillars} in the literature. The degrees of the vertices on the spine are distributed independently by
\begin {equation} \label{phiinfcat}
\phi(k) = \frac{\alpha \Gamma\left(k-1-\alpha\right)}{\Gamma\left(k\right)\Gamma\left(1-\alpha\right)}
\end {equation}
and they have an infinite expected value for all values of $\alpha$. These caterpillars are a special case of 'caterpillars at a phase transition' in the equilibrium statistical mechanical model studied in \cite{jonsson:2009}. We will consider this special case in more detail in the next section.
\section{The Hausdorff dimension} \label{s:hd}
The Hausdorff dimension is a notion of dimension of graphs and is defined in terms of how the volume of the graph ball $B_R$ scales with its radius $R$. The Hausdorff dimension of a graph $G$ is defined as
\begin {equation} \label{dhdef}
d_H = \lim_{R\rightarrow\infty} \frac{\log(|B_R(G)|)}{\log(R)}
\end {equation}
provided that the limit exists. This definition is only interesting on an infinite graph. On the hyper--cubic lattice $\mathbb{Z}^d$ it holds that $d_H = d$ but in general $d_H$ is not an integer. This dimension has been studied by physicists, especially in the quantum gravity literature, see e.g.~\cite{book} and should not be confused with the usual notion of Hausdorff dimension in a metric space, although there are some similarities.
The Hausdorff dimension can be defined in different ways for random graphs. If the graphs are distributed by $\nu$ then they might first of all have, $\nu$--almost surely, a Hausdorff dimension $d_H$ as defined above. Secondly, we define the annealed Hausdorff dimension as
\begin {equation} \label{dhdefexp}
\bar{d}_H = \lim_{R\rightarrow\infty} \frac{\log(\langle|B_R(G)|\rangle_\nu)}{\log(R)}
\end {equation}
where $\langle \cdot \rangle_\nu$ denotes expected value with respect to $\nu$.
There is another notion of dimensionality which applies when one considers a sequence of finite volume measures $(\nu_n)_{n\geq 0}$ on a set of graphs. It is usually defined in terms of how the average value of some typical distance in the graph (the maximum distance between vertices, the mean distance of vertices from the root, etc.) scales in relation to the volume of the graph $n$ as it grows. This dimension has also been referred to as the Hausdorff dimension in the physics literature but to avoid confusion we will refer to it here as the {\it fractal dimension} and denote it by $d_f$. To give a more precise definition, we adopt the one from \cite{vs} which is as follows: Define the radius of a finite tree $T$ by
\begin {equation}
R_T = \frac{1}{2|T|} \sum_{v \in V(T)} d_T(r,v) \sigma_T(v)
\end {equation}
where $V(T)$ is the vertex set of $T$, $r$ is the root, $d_T$ is the graph metric and $\sigma_T(v)$ denotes the degree of $v$. The fractal dimension is defined as
\begin {equation} \label{eq:fractal}
d_f = \lim_{n\rightarrow\infty} \frac{\log(n)}{\log(\langle R_T\rangle_{\nu_n})}.
\end {equation}
If $\nu_n$ converge to a measure $\nu$ concentrated on infinite graphs, $d_f$ has been observed to be equal to $d_H$ (or $\bar{d}_H$) in many situations, a simple example is the uniform tree and modifications of it, see e.g.~\cite{sdgt}. It is however straightforward to find a counterexample where $d_H \neq d_f$ and it is not entirely clear which conditions guarantee equality. We will comment on this relation in the ag-model below.
We will now calculate the annealed Hausdorff dimension of the trees distributed by $\nu$ from Theorem \ref{th:conv}.
\begin {theorem}\label{th:annhaus}
Let $\max\{0,a\} < \alpha \leq 1$. The random trees, distributed by $\nu$ described in Theorem \ref{th:conv}, have an annealed Hausdorff dimension
\begin {equation}
\bar{d}_H = \frac{1}{\alpha}.
\end {equation}
\end {theorem}
To prove Theorem \ref{th:annhaus}, we need to analyse the large $R$ behaviour of $\langle |B_R| \rangle_{\nu}$. In order to simplify the notation we let $(\emptyset)$ be the empty tree and define $\mu_2((\emptyset)) = 1$. We then extend the probability distributions $\mu_k$, $k \geq 2$, to probability distributions on
$ \bigcup_{k=2}^\infty \mathcal{T}^{k-2}$ and define
\begin {equation} \label{mudef}
\mu = \sum_{k=2}^\infty \phi(k) \mu_k.
\end {equation}
Since the outgrowths from different vertices on the spine are i.i.d.~it is clearly sufficient to show that
\begin {equation}
\Big\langle \sum_{i} |B_R(\tau_i)| \Big\rangle_{\mu} = R^{1/\alpha - 1} (1+o(1))
\end {equation}
as $R\longrightarrow \infty$. This follows from the Lemma below which is proved on page \pageref{pr:annhaus} in Appendix B.
\begin {lemma} \label{l:annhaus}
For $\max\{0,a\} < \alpha \leq 1$,
\begin {eqnarray*}
\Big\langle \sum_{i} |B_R(\tau_i)| \Big\rangle_{\mu} &=& \frac{(R + \frac{\alpha - a}{\alpha})\G{\frac{\alpha-a}{\alpha}} \G{R+\frac{1-a}{\alpha}}}{\G{\frac{1-a}{\alpha}}\G{R+\frac{2\alpha -a}{\alpha}}}-1.
\end {eqnarray*}
\end {lemma}
Note that $\bar{d}_H = \infty$ when $D = \infty$ and $\alpha \leq 1-\gamma$ ($\alpha \leq \max\{0,a\}$) since then the expected value of degrees of vertices on the spine is infinite. However, the $\nu$--almost sure Hausdorff dimension might still be finite. We confirm this in the case $D=\infty$ and $\gamma = 0$, when the trees are caterpillars.
\begin {theorem} \label{th:ashaus}
Let $0 < \alpha \leq 1$ and $\gamma = 0$. Then
\begin {equation}
d_H = \frac{1}{\alpha}
\end {equation}
$\nu$--almost surely.
\end {theorem} \noindent
The proof is given on page \pageref{pr:ashaus} in Appendix B.
\subsection{Comparison to the fractal dimension}
In the original paper on the vertex splitting model \cite{vs} it was shown that the expected value of the radius of a tree can be written as
\begin {equation} \label{eq:radius}
\langle R_T \rangle_{\nu_n} = \frac{n+1}{2n}\sum_{n_2=0}^n (2n_2+1)\sum_{k} \tilde{q}_k(n-n_2,n_2)
\end {equation}
where $\tilde{q}_k(n_1,n_2)$ is the probability that a uniformly chosen vertex $v$ has degree $k$ and that the volume of the subtree attached to $v$ containing the root is $n_1$ and that the other subtrees attached to $v$ have a total volume $n_2$. Furthermore, in the case of linear splitting weights $w_i = Ai+B$, $\tilde{q}_k(n_1,n_2)$ was shown to be a solution of a system of linear recursion equations determined by the growth rules of the vertex splitting model, see \cite[Section 3]{vs}. These recursion equations could not be solved explicitly but it was assumed that the following scaling holds
\begin {equation} \label{eq:2pscaling}
\tilde{q}_k(n_1,n_2) \sim (n_1+n_2)^{-2+\lambda}\omega_k(n_1/(n_1+n_2))
\end {equation}
for some $\lambda$ and ``scaling functions`` $\omega_k$. The linear recursions were thus reduced to an eigenvalue equation for $\lambda$
\begin {equation} \label{eq:eig}
\mathbf{C} \mathbf{x} = \lambda \mathbf{x}
\end {equation}
where $\lambda$ is the Perron--Frobenius eigenvalue of the $\binom{D}{2} \times \binom{D}{2}$ matrix $\mathbf{C}$ indexed by a pair of two indices $ki$, $2 \leq k \leq D$, $2 \leq i < k$, and given by the matrix elements
\begin {equation} \label{eigeq}
C_{ki,jn} = \frac{1}{w_2}(w_{k,j+2-k}((j-i)\delta_{i,n} + i \delta_{n,j-k+i}) - w_k \delta_{k,j}\delta_{i,n}).
\end {equation}
Comparing (\ref{eq:radius}) and (\ref{eq:2pscaling}) to (\ref{eq:fractal}) allows one to find the fractal dimension
\begin {equation} \label{eq:fracresult}
d_f = \lambda^{-1}.
\end {equation}
The scaling assumption (\ref{eq:2pscaling}) was not proven but the results (\ref{eq:eig}-\ref{eq:fracresult}) were supported by simulations in the case $D=3$.
It is interesting to compare $d_f$, corresponding to the weights (\ref{partw}) of the ag--model, to the values of $\bar{d}_H$ obtained in Theorem \ref{th:annhaus}. It is straightforward to solve (\ref{eq:eig}) for small values of $D$ and find that $d_f = \bar{d}_H = 1/\alpha$. Furthermore, we have calculated $d_f$ in the case $D=\infty$ and $\gamma = 1$ by solving (\ref{eq:eig}) numerically. We used a cutoff $D=30$ on the system which is expected to closely approximate the case $D=\infty$, since the vertex degree distribution is believed to fall of exponentially in this case, cf.~(\ref{ddist}). The results are shown in Fig.~\ref{f:hausinf}. The agreement we find, supports the validity of the scaling assumption (\ref{eq:2pscaling}) to a very high maximum degree $D$.
\begin{figure} [!h]
\centerline{\scalebox{0.3}{\includegraphics{hausinf.eps}}}
\caption{Comparison of $d_f$ (gray squares) and $\bar{d}_H = 1/\alpha$ (solid line) in the case $D=\infty$ and $\gamma = 1$. Using the weights (\ref{partw}), we calculated $d_f$ numerically from (\ref{eq:eig}-\ref{eq:fracresult}) for $\alpha = i/100,~1\leq i \leq 100$, using a cutoff $D=30$ on the system.} \label{f:hausinf}
\end{figure}
\subsection{The spectral dimension}
We conclude this section by mentioning another notion of dimension of graphs called the {\it spectral dimension}. It is defined in terms of how the return probability of a random walker on the graph decays with time $t$. More precisely, for a tree $\tau$ let $p_\tau(t)$ be the probability that a simple random walk which leaves the root at time $t=0$ is back at the root at time $t$. The spectral dimension of $\tau$ is defined as
\begin {equation}
d_s = -\lim_{t\rightarrow\infty}\frac{2\log(p_\tau(t))}{\log(t)}
\end {equation}
provided the limit exist. The spectral dimension can take any value greater than one and does not necessarily agree with the Hausdorff dimension. We refer to \cite{barlow:2005,sdgt,combs,brushes} for discussion of the spectral dimension of several types of random graphs.
It would be interesting to calculate the spectral dimension of the trees distributed by $\nu$. For now we only have results in the case when the trees are caterpillars.
\begin {theorem} \label{th:spec}
Let $0 < \alpha \leq 1$ and $\gamma = 0$. If $d_s$ exists then
\begin {equation}
d_s = \dfrac{2}{1+\alpha}
\end {equation}
$\nu$-- almost surely
\end {theorem} \noindent
The theorem is proved on page \pageref{pr:spec} in Appendix B.
\section{Vertex degree distribution and correlations}
In this section we use results from the vertex splitting model \cite{vs} to calculate the vertex degree distribution and correlation between the degrees of neighbouring vertices in the ag--model. Not all results in this section are rigorous and we will comment on this point below.
Let $X_{i,n}$ be the number of vertices of degree $i$ in a random tree with $n$ edges and define the vertex degree densities
\begin {equation}
\rho_i = \lim_{n\rightarrow\infty} \frac{\mathbb{E}(X_{i,n})}{n}.
\end {equation}
It was shown in \cite[Section 2]{vs} that the densities in the vertex splitting model satisfy the linear equation
\begin{equation} \label{mf}
\rho_i = -\frac{w_i}{w_2}\rho_i + \sum_{j\geq k-1} \frac{j w_{i,j+2-i}}{w_2}\rho_j
\end{equation}
assuming that the splitting weights are linear $w_i = Ai+B$ and under certain technical conditions on the partitioning weights. The splitting weights are linear in the ag--model, cf.~(\ref{splittingweights}), and one can check that for small $D$ the technical conditions needed on the partitioning weights are fulfilled. However, it is not certain whether (\ref{mf}) holds for general $D$ and one would need further analysis to verify that. It is straightforward to solve (\ref{mf}) for the weights (\ref{partw}) and thus we find that the vertex degree densities in the ag--model are given by
\begin {equation}
\rho_1 = \frac{1-\alpha}{2-a-\alpha}
\end {equation}
and
\begin {equation} \label{ddist}
\rho_k = \frac{(1-a)\G{\frac{a+2-\alpha}{a}}\G{k-2+\frac{1-\alpha}{a}}}{(2-a-\alpha)(2-\alpha)\G{\frac{1-\alpha}{a}}\G{k-1+\frac{2-\alpha}{a}}}, \qquad {k\geq 2}
\end {equation}
provided that (\ref{mf}) holds. By sending $\alpha$ to zero we find that these results agree with results previously obtained in the preferential attachment model \cite{albert0,rapoport}. Also note, that in the case $D=\infty$, $\rho_k$ has in general a power law behaviour except when $\gamma = 1$ ($a=0$) in which case it falls of exponentially with rate $(1-\alpha)/(2-\alpha)$. This resembles properties of the degree distribution of the vertices on the spine, cf.~(\ref{eq:degexp}) and (\ref{eq:degpower}).
Let $X_{ij,n}$ be the number of edges with endpoints of degree $i$ and $j$ in a random tree with $n$ edges, using the convention that the vertex of degree $i$ is the one closer to the root. Define the density
\begin {equation} \label{limitcorr}
\rho_{ij} = \lim_{n\rightarrow\infty}\frac{\mathbb{E}(X_{i,j,n})}{n}.
\end {equation}
It was shown in \cite{vs} that these densities in the vertex splitting model satisfy
\begin {eqnarray} \nonumber
\rho_{jk} &=& -\frac{w_j+w_k}{w_2}\rho_{jk} + (j-1)\frac{w_{j,k}}{w_2}\rho_{j+k-2} + (j-1)\sum_{i\geq j-1} \frac{w_{j,i+2-j}}{w_2} \rho_{ik}\\
&& +~ (k-1)\sum_{i\geq k-1} \frac{w_{k,i+2-k}}{w_2}\rho_{ji} \label{rhojk}
\end {eqnarray}
assuming that the limit (\ref{limitcorr}) exists. The densities give us information about correlation between vertex degrees of neighbouring vertices. It can be measured with a correlation coefficient
\begin {equation}
r = \frac{\sum_{j,k}(j-1)(k-1)(\bar{\rho}_{jk}-\bar{\rho}_j\bar{\rho}_k)}{\sum_{j}(j-1)^2\bar{\rho}_j - \left(\sum_{j}(j-1)\bar{\rho}_j\right)^2}
\end {equation}
where
\begin {equation}
\bar{\rho}_k = \frac{k\rho_k}{\sum_i i\rho_k} \qquad \text{and} \qquad \bar{\rho}_{ij} = \frac{\rho_{ij}+\rho_{ji}}{2}.
\end {equation}
The coefficient $r$ takes values between $-1$ and $1$. If $r < 0$ the graph is said to show disassortative mixing and vertices with high degree prefer to be neigbours of vertices with low degree. If $r > 0$ the graphs are said to show assortative mixing and vertices with high degree prefer to be neighbours of vertices with high degree, see e.g.~\cite{newman}.
We will conlude this section by calculating $r$ for two choices of parameters in the ag--model, namely $D=3$ and $D=\infty$, $\gamma = 1$. We consider (\ref{rhojk}) with the weights given in (\ref{partw}). In the case $D=3$, (\ref{rhojk}) can be explicitly solved and we find that
\begin {equation} \label{r1}
r = -\frac{\alpha(4-3\alpha)}{(5-3\alpha)(3-\alpha)}, \qquad (D=3).
\end {equation}
This can of course be repeated for small values of $D$. However, when $D$ is large or infinite it is more difficult to solve (\ref{rhojk}) explicitly. Instead, we study the generating function
\begin {equation}
S(x,y) = \sum_{j,k\geq 2} \bar{\rho}_{jk} x^{j-1} y^{k-1}
\end {equation}
and use the fact that
\begin {equation}
\partial_x\partial_y S(1,1) = \sum_{j,k}(j-1)(k-1)\bar{\rho}_{jk}
\end {equation}
to calculate $r$. Equation (\ref{rhojk}) becomes a linear, first order partial differential equation in terms of the generating function $S(x,y)$. It can in principle be solved for a general set of parameters, however we only comment on the case $D=\infty$, $\gamma = 1$. In that case, the coefficients of the derivative terms in the PDE are zero and we get an ordinary equation for $S(x,y)$. The solution is
\begin {equation}
S(x,y) = \frac{\sigma (x^2 f(y)+y^2 g(x))}{1- \sigma x - \sigma y} + x f(y) + y g(x) + \rho_{22} x y
\end {equation}
where
\begin {equation}
f(y) = \sum_{k=3}^\infty \rho_{2k} y^{k-1} \quad \text{and} \quad g(x) = \sum_{j=3}^\infty \rho_{j2} x^{j-1},
\end {equation}
\begin {eqnarray} \nonumber
\rho_{2k} &=& \frac{2(1-3\alpha + \alpha^2)}{(1-\alpha)(2-\alpha)^2} \sigma^k+\frac{\alpha}{(1-\alpha)^2} \eta^k,\qquad k\geq 2, \\ \nonumber
\rho_{j2} &=& -\frac{2\alpha^4-15\alpha^3+38\alpha^2-37\alpha+8+(2\alpha^2-6\alpha +4)j}{(1-\alpha)^2(2-\alpha)^2} \sigma^j \\ \nonumber
&& +~ \frac{2\alpha^2-6\alpha +2 + \alpha j}{(1-\alpha)^2} \eta^j, \qquad j\geq 3,
\end {eqnarray}
with
\begin {equation}
\eta = \frac{1-\alpha}{2-\alpha} \quad \text{and} \quad \sigma = \frac{1-\alpha}{3-\alpha}.
\end {equation}
From these expressions we find that
\begin {equation} \label{r2}
r = -\frac{\alpha}{2(1+\alpha)} \qquad (D=\infty, \gamma = 1).
\end {equation}
We plot the solutions (\ref{r1}) and (\ref{r2}) together in Fig.~\ref{f:plot}.
\begin{figure} [!h]
\centerline{\scalebox{0.20}{\includegraphics{ass.eps}}}
\caption{Comparison of Equations (\ref{r1}) ($D=3$, black) and (\ref{r2}) ($D=\infty$, $\gamma=1$, gray).} \label{f:plot}
\end{figure}
The two curves are similar in both cases. If $\alpha = 0$ then $r=0$ which agrees with results which have previously been obtained for the preferential attachment model \cite{newman}. In this case the vertices which are close to the root are 'old' in the sense that once they reach the maximum degree (which they eventually do with probability one) they do not change again. Thus, a lot of vertices of high degree become neigbours. When $\alpha$ is increased above zero, a repulsions is introduced between these vertices, the value of $r$ decreases and the trees show disassortative mixing. When $\alpha$ goes to 1, the trees approach the same non--random graph, a spine with no outgrowths. As a consequence, the value of $r$ approaches the same value in both cases.
\section{Conclusions}
We introduced the ag--model, a special case of the vertex splitting model which has the Markov branching property. For particular choices of parameters it reduces to models of generic trees \cite{sdgt}, preferential attachment \cite{albert0} and non--generic caterpillars \cite{jonsson:2009}. It was proved that the finite volume measures generated by the growth rules converge to a measure which is concentrated on the set of trees with exactly one spine and the limiting measure was described explicitly. The same has been done before in Ford's $\alpha$--model \cite{siggi} and a special case of the $\alpha \gamma$--model \cite{phdsos}. Extension of these convergence results to the vertex splitting model is a work in progress.
There is another notion of convergence of random trees, referred to as the {\it scaling limit}, see e.g.~\cite{aldousoriginal,continuum}. This means, roughly, that a random tree $T_n$ viewed as a metric space with the graph metric $d_{\text{gr}}$ suitably scaled, converges weakly, in the Gromov--Hausdorff topology, to a continuum random tree. In a recent paper on Markov branching trees \cite{markovss}, Haas and Miermont proved that under certain natural conditions on the first split distributions the scaling limit of the trees is a self--similar fragmentation tree, in the Gromov--Hausdorff-Prokhorov topology. We expect that this theory applies to the model studied in this paper and it would be interesting to confirm that. Moreover, it is an interesting and challenging problem to generalize the results on the scaling limit to the vertex splitting model when Markov branching is absent.
The annealed Hausdorff dimension, with respect to the infinite volume measure of the ag--model, was calculated for a certain range of the parameters. The results partly support scaling assumptions which were made when calculating the fractal dimension in the vertex splitting model \cite{vs}. In the special case of growing caterpillar graphs we calculated, almost surely, the Hausdorff and spectral dimension. It turns out that the dimensions are related by the formula
\begin {equation}
d_s = \frac{2 d_H}{1+d_H}.
\end {equation}
This equation holds in general for tree models which satisfy a certain uniformity condition and under the assumption that vertex degrees are uniformly bounded from the above \cite{barlow:2005}. We expect this relation to hold in the ag--model and it would be desirable to check whether it holds in the vertex splitting model.
It is possible to study other interesting observables in the ag--model such as the vertex degree distribution and correlations between degrees of neighbouring vertices in large trees. It would be interesting to give a rigorous proof of (\ref{ddist}) and even to get stronger convergence results for the random variables $X_{i,n}$. This can presumably be done, at least for some range of the parameters, using results on generalized Pólya urns \cite{svante}. Furthermore, it would be interesting to confirm the validity of (\ref{mf}) for as general set of parameters as possible. Similar results about the convergence of $X_{ij,n}$ are desirable.
A natural question is whether the ag--model is the only special case of the vertex splitting model which has the Markov branching property. As was noted in the introduction, the $\alpha \gamma$--model has the Markov branching property but it is not strictly a special case of the vertex splitting model, rather a limiting case. Since the vertex splitting model has local and isotropic growth rules, one might also ask whether there exists some other notion of self--similarity which could be used to handle the general case. An understanding of this could be a key element towards a solution of the most general case.
\begin {ack}
I am deeply indebted to Thordur Jonsson and François David for helpful discussions and comments.
\end {ack}
|
\section{Introduction}
The unexpectedly large forward-backward asymmetry in the production of \ensuremath{{t\bar t}}\
pairs at the Tevatron as observed by CDF $A^\ensuremath{{t\bar t}} = 0.193 \pm
0.069$~\cite{Aaltonen:2008hc} and D\O\ $A^{\ensuremath{{t\bar t}}\ } = 0.24 \pm
0.14$~\cite{Abazov:2007qb} in 2008 generated a lot of interest, because it is
significantly higher than the Standard Model (SM) prediction, $A^{\ensuremath{{t\bar t}}\, \rm
(SM)} \approx 0.06$~\cite{Antunano:2007da, Bowen:2005ap, Kuhn:1998kw,
Almeida:2008ug}. One reason for being excited about this measurement is that the
top quark is a very sensitive probe of putative new physics at the TeV scale,
because of its large mass and coupling to the Higgs. Therefore, one might expect
signs of new physics to first show up in top physics. This hope has received a
boost by the recent CDF analysis, which showed that the asymmetry arises from
\ensuremath{{t\bar t}}\ events with high invariant masses~\cite{Aaltonen:2011kc}
\begin{eqnarray}\label{CDFasym}
A^\ensuremath{{t\bar t}}(m_\ensuremath{{t\bar t}} > 450\,{\rm GeV}) &=& 0.475 \pm 0.114 \,, \nonumber \\
A^\ensuremath{{t\bar t}}(m_\ensuremath{{t\bar t}} < 450\,{\rm GeV}) &=& -0.116 \pm 0.153 \,.
\end{eqnarray}
The updated D\O\ result, only available integrated over \ensuremath{m_{t\bar t}}, and
uncorrected for effects from reconstruction or selection, $A^\ensuremath{{t\bar t}} = 0.08 \pm
0.04$~\cite{D0afb}, is consistent with the integrated CDF result, $A^\ensuremath{{t\bar t}} =
0.158 \pm 0.075$~\cite{Aaltonen:2011kc}. So is the recent $A^\ensuremath{{t\bar t}} = 0.417 \pm
0.157$ measurement~\cite{CDFdilepton} in the dilepton channel, in which the raw
asymmetries, binned as in Eq.~(\ref{CDFasym}), also support the same trend. The
physics responsible for this anomaly may be related to CDF's high $p_T$ excess
in a boosted top search~\cite{CDFboostedtop}. The large asymmetry at high
masses points towards tree-level exchange of a new heavy particle with strong
couplings to first and third generation quarks~\cite{Jung:2009jz,
Frampton:2009rk, Shu:2009xf, Arhrib:2009hu, Dorsner:2009mq, Burdman:2010gr,
Cheung:2011qa, Bai:2011ed, Shelton:2011hq, Blum:2011up, Grinstein:2011yv}. For
fits of four-fermion operators to the asymmetry data
see~\cite{Degrande:2010kt,Jung:2009pi}.
In absence of flavor symmetries, new states at the TeV scale with strong
couplings to quarks are severely constrained by the agreement of a vast amount
of flavor physics data with the SM (meson-anti-meson mixing, $CP$ violation,
rare decays). We are therefore motivated to look for an explanation of the
\ensuremath{{t\bar t}}\ asymmetry from new states whose couplings (and masses) preserve the full
flavor symmetries of the Standard Model quarks along the lines of
Refs.~\cite{Bauer:2009cc, Arnold:2009ay, Grinstein:2011yv}.
To do so, we classify the new particles not only by their spin and gauge
charges, but also by their quantum numbers under the flavor symmetries $SU(3)_Q
\times SU(3)_U \times SU(3)_D$. Here $SU(3)_Q$ is the set of transformations
which rotate the three generations of left-handed quark doublets, $Q$, and
$SU(3)_{U/D}$ transformations rotate the right-handed quark singlets, $U/D$. For
simplicity, and because this leads to the nicest model, we focus on the case
where the new states couple only to right-handed up-type
quarks.\footnote{Coupling to up-type quarks is preferred because it accesses the
large up-quark parton distribution function, and even $SU(3)_Q$ symmetric
couplings to the quark doublets give rise to new flavor violation proportional
to CKM matrix elements.} Depending on whether the coupling is to two quarks or
to a quark and an anti-quark, the new states have quantum numbers of a
``diquark" with baryon number 2/3 or a ``noquark'' with baryon number 0. Under
$SU(3)_U$ flavor the new particles must transform in one of the irreducible
representations contained in
\begin{equation}\label{options}
\mbox{diquark:}\ \ \ensuremath{{\mathbf{3}}} \otimes \ensuremath{{\mathbf{3}}} = \ensuremath{\mathbf{\ov 3}} \oplus \ensuremath{\mathbf{6}}\,, \quad
\mbox{noquark:}\ \ \ensuremath{{\mathbf{3}}} \otimes \ensuremath{\mathbf{\ov 3}} = \ensuremath{\mathbf{1}} \oplus \ensuremath{\mathbf{8}}\,. \
\end{equation}
With regards to generating an asymmetry, the diquark models are nice because
diquarks contribute to \ensuremath{{t\bar t}}\ production in the $u$-channel (see
Fig.~\ref{fig:diag}). This new source of top quarks is peaked in the forward
direction and can easily produce a large asymmetry. The ``noquarks'' in the
flavor singlet representation are closely related to the extensively discussed
axigluons and do not provide a very good fit to the asymmetry data. The main
problem is that they are $s$-channel resonances coupling to up quarks and to top
quarks. They would give rise to features in the invariant mass distribution $d
\sigma_\ensuremath{{t\bar t}}/d m_\ensuremath{{t\bar t}}$ of \ensuremath{{t\bar t}}\ pairs and also of dijets at the Tevatron
and the LHC~\cite{Bai:2011ed}. The case of the \ensuremath{\mathbf{8}}\ of flavor is more
interesting as it contributes to \ensuremath{{t\bar t}}\ production in the $t$-channel and the
$s$-channel. It is possible to find good fits to both the asymmetry as well as
the total \ensuremath{{t\bar t}}\ cross section in this case for either light ($M_8\sim
300$~GeV) or heavy ($M_8 \sim 1000$~GeV) new states~\cite{Grinstein:2011yv}.
Data on dijet resonances from the Tevatron~\cite{Aaltonen:2008dn, CDFchidist,
Abazov:2009mh} and SPS~\cite{Arnison:1986np, Alitti:1990aa} rule out flavor
universal couplings in the light case. In addition, light spin one particles are
associated with gauge symmetries. This would imply at least an approximate gauge
symmetry of flavor, which would need to be broken by the mechanism generating
the up-type Yukawa couplings. The heavy state requires very large couplings to
generate a large enough asymmetry and can be shown to violate recent bounds on
dijets from the LHC~\cite{Khachatryan:2011as}.\,\footnote{The dijet bounds can
be evaded with large flavor breaking in the third generation
couplings~\cite{Grinstein:2011yv}.}
We will not consider the \ensuremath{\mathbf{1}}\ or \ensuremath{\mathbf{8}}\ any further.
\begin{figure}[tb]
\includegraphics[width=.42\columnwidth]{diquark} \hfil\hfil
\includegraphics[width=.42\columnwidth]{noquark}
\caption{Feynman diagrams for the new contribution to $u\bar u\to \ensuremath{{t\bar t}}$ from a
diquark (left) and a noquark (right). The flavor symmetry implies an
additional $t$-channel diagram for the noquark~\ensuremath{\mathbf{8}}\ (not pictured).}
\label{fig:diag}
\end{figure}
Concentrating on renormalizable interactions, the \ensuremath{\mathbf{\ov 3}}\ and \ensuremath{\mathbf{6}}\ must be
complex scalar fields and their couplings to quarks can be written as
\begin{equation}\label{coupling}
\lambda\, U^{a\alpha}\, U^{b\beta} (S^*)^{r\rho}\, T_{rab}\, T_{\rho\alpha\beta}\,.
\end{equation}
Flavor indices $a,b$ on the quarks run from 1 to 3 and the flavor index on $S$
runs over $r=1\ldots3$ or $r=1\ldots6$. We also explicitly displayed the color
indices $\alpha,\beta=1\ldots3$ on the quarks
and $\rho=1\ldots3$ or $\rho=1\ldots6$ on $S$.
Note that the color and flavor quantum numbers of $S$ are correlated,
because the two identical quark fields must be symmetric under interchange of
color and flavor indices. The scalar \ensuremath{\mathbf{\ov 3}}\ is a $(\overline 3,1)_{4/3}$
under the Standard Model gauge group $(SU(3)_{c}, SU(2)_{w})_{U(1)}$ whereas the
\ensuremath{\mathbf{6}}\ is a $(6,1)_{4/3}$. The invariant tensors $T$ are therefore the same for
color and flavor. $T_{rab}$ for the \ensuremath{\mathbf{\ov 3}}\ is antisymmetric in $a$ and $b$
whereas for the \ensuremath{\mathbf{6}}\ it is symmetric. We emphasize that in either case these
couplings preserve the $SU(3)_U$ flavor symmetry and do not contribute to flavor
violating processes.
An obvious but important consequence of the flavor symmetry is that it relates
processes involving quarks in different generations.
In particular, we observe an important distinction between the \ensuremath{\mathbf{\ov 3}}\ and
the \ensuremath{\mathbf{6}}. Equation~(\ref{coupling}) for the \ensuremath{\mathbf{6}}\ contains a coupling of two up
quarks to $S$. This leads to an $s$-channel resonance in $uu\to uu$
scattering, which accesses the largest high energy parton luminosities at the
LHC and is already severely constrained by recent dijet analyses from
CMS~\cite{Khachatryan:2010jd, Khachatryan:2011as} and
ATLAS~\cite{Collaboration:2010eza}. As we will show below, a measurement of
dijet angular distributions at CMS~\cite{Khachatryan:2011as} already rules out
the entire range of couplings and masses for the \ensuremath{\mathbf{6}}\ which generates an
appreciable \ensuremath{{t\bar t}}\ asymmetry. The \ensuremath{\mathbf{\ov 3}}\ is antisymmetric in its couplings
to quarks so that the coupling to two up quarks vanishes.
This leaves the scalar triplet as our only flavor preserving candidate for
explaining the \ensuremath{{t\bar t}}\ forward-backward asymmetry at the Tevatron. In the
following Sections we compare the predictions of this model to all relevant data
\begin{enumerate}\vspace*{-4pt}\itemsep 0pt
\item The binned \ensuremath{{t\bar t}}\ asymmetry in
Eq.~(\ref{CDFasym})~\cite{Aaltonen:2011kc};
\item The $p\bar p\to \ensuremath{{t\bar t}}$ total cross section, $\sigma_\ensuremath{{t\bar t}} = (7.5 \pm
0.48)$\,pb~\cite{CDFsigma};
\item Measurement of ${\rm d}\sigma_\ensuremath{{t\bar t}}/{\rm d}\ensuremath{m_{t\bar t}}$~\cite{Aaltonen:2009iz};
\item The recent CMS dijet analysis~\cite{Khachatryan:2011as}.
\end{enumerate}\vspace*{-4pt}
We find that unlike all the other models the triplet scalar is currently
unconstrained by dijet data. However, there is some tension between the large
\ensuremath{{t\bar t}}\ asymmetry required at high invariant masses and the shape of the
measured differential cross section as a function of the invariant mass of the
\ensuremath{{t\bar t}}\ pair. In particular, a 40\% asymmetry requires a new physics
contribution which is very asymmetric and comparable in size to the \ensuremath{{t\bar t}}\
cross section from QCD, but which does not significantly change the shape of the
cross section. We are not aware of any model in the literature which completely
accomplishes this.
The best fit to the \ensuremath{{t\bar t}}\ asymmetry and cross section in our model is obtained
for triplet masses in the range 500--800~GeV with relatively large couplings,
$\lambda \sim 1.5-3.0$ (for even larger couplings our perturbative calculations
quickly become suspect). It is therefore easily within reach of the LHC, and
both single and pair production of these states should be possible with large
cross sections. The most promising processes are single production with top
quarks $u g \rightarrow \bar{t} S_{tu} \rightarrow \ensuremath{{t\bar t}} j$ and pair production
$gg,\, \ensuremath{{u\bar u}} \rightarrow S S^* \rightarrow jjjj,\, \ensuremath{{t\bar t}} jj$. In the context
of flavor-symmetric models it would be particularly interesting if one could
measure the forward-backward asymmetries in dijet events with pairs of high
$p_T$ charm or bottom quark jets at the
Tevatron~\cite{Bai:2011ed,Strassler:2011vr}. Our model predicts an asymmetry for
charm quarks similar to that for top quarks and vanishing new physics
contribution to the bottom quark asymmetry.
The remainder of the paper is organized as follows. In Section~\ref{sec:model}
we define the \ensuremath{\mathbf{\ov 3}}\ and the \ensuremath{\mathbf{6}}\ models, compute the \ensuremath{{t\bar t}}\ cross section
and asymmetry in each and determine the preferred region in parameter space by
fitting to the Tevatron \ensuremath{{t\bar t}}\ data. We then compute the predicted dijet rates
at the Tevatron and LHC and show that the preferred region of the \ensuremath{\mathbf{\ov 3}}\
model is still allowed whereas the \ensuremath{\mathbf{6}}\ model is ruled out by the CMS dijet
measurement. In Section~\ref{sec:lhc} observable predictions for the LHC from
the diquark of the \ensuremath{\mathbf{\ov 3}}\ model are explored.
\section{The \ensuremath{\mathbf{\ov 3}}\ and the \ensuremath{\mathbf{6}}}
\label{sec:model}
The two models contain a new scalar in the \ensuremath{\mathbf{\ov 3}}\ or \ensuremath{\mathbf{6}}\ representation of
flavor $SU(3)_U$. The interaction Lagrangian is
\begin{equation}\label{3def}
{\cal L} = \lambda\, U^{a\alpha}\, U^{b\beta} (S^*)^{r\rho}\,
T_{rab}\, T_{\rho\alpha \beta} \, + {\rm h.c.}\,,
\end{equation}
where Latin (Greek) indices denote flavor (color). The fields $U$ are the
right-handed up-type quark singlets with SM charges $(3,1)_{2/3}$, their Lorentz
indices are contracted with $i\sigma_2$. The invariant tensor for the \ensuremath{\mathbf{\ov 3}}\
is $T_{rab} = \epsilon_{rab}/\sqrt{2}$. For the \ensuremath{\mathbf{6}}\ the tensor $T_{rab}$ can
be decomposed into six $3\times 3$ real symmetric matrices, $T_r$. In computing
amplitudes from exchange of these particles we will use the identities
\begin{eqnarray}
(T_3)_{rab}\, (T_{3})^{rcd} &=&
\frac12 \big(\delta^c_a \delta^d_b - \delta^c_b \delta^d_a\big)\,, \nonumber\\
(T_6)_{rab}\, (T_6)^{rcd} &=&
\frac12 \big(\delta^c_a \delta^d_b + \delta^c_b \delta^d_a\big)\,,
\end{eqnarray}
which also fix the normalization of our invariant tensors. Here $T^{rcd} \equiv
T_{rcd}^* = T_{rcd}$, because we defined the tensors to be real. Each model has
only two parameters, the coupling $\lambda$, and the mass $m_S$. Our definition
of the coupling $\lambda$ differs by a factor of $\sqrt{2}$ in normalization
compared to the coupling $y$ of Ref.~\cite{Shu:2009xf}, so that
$\lambda=y/\sqrt{2}$.
Of course, other viable models may be obtained by including additional free
parameters. In the context of minimal flavor violation one might be motivated to
consider insertions of $Y_UY_U^\dagger$. These would have very small effects on
light quarks but could change the top quark couplings significantly. We have
refrained from doing so in the interest of simplicity and because we expect
insertions of $Y_UY_U^\dagger$ generated from loops to be small.
The process $u(p_1)\, \bar u(p_2) \to t(k_1)\, \bar t(k_2)$ is given in the SM
by $s$-channel gluon exchange. The $S$ interaction mediates a $u$-channel
contribution (Fig.~\ref{fig:diag}). Including both contributions, the
differential partonic cross section is~\cite{Shu:2009xf, Arhrib:2009hu,
Dorsner:2009mq}
\begin{eqnarray}\label{dcostheta}
\frac{{\rm d}\sigma_\ensuremath{{t\bar t}}}{{\rm d}\cos\theta} &=&
\frac14\, \frac19\, \frac\beta{2\pi s}
\bigg[g_s^4\, \frac{t_t^2 + u_t^2 + 2 s\, m_t^2}{s^2} \nonumber\\
&&{} + g_s^2\, \lambda^2\, C_0\, \frac{u_t^2+s\, m_t^2}{s\, u_S}
+ \lambda^4\, C_2\, \frac{u_t^2} {u_S^2} \bigg] .
\end{eqnarray}
Here $\theta$ is the scattering angle between the outgoing top and the incoming
quark in the partonic center-of-mass frame, and $\beta \equiv \sqrt{1 - 4m^2_t /
s}$. The Mandelstam variables are $s \equiv (p_1 + p_2)^2$, $t \equiv (p_1 -
k_1)^2$, $u \equiv (p_1 - k_2)^2$, and we denoted $t_X = t - m_X^2$ and $u_X = u
- m_X^2$ ($X=t,S$). Finally, the color factors are $C_0=1$ and $C_2=3/4$ in the
case of the \ensuremath{\mathbf{\ov 3}}, and $C_0=-1$ and $C_2=3/2$ for the \ensuremath{\mathbf{6}}.\,\footnote{Here
we note a typographical error in the sign of $C_0$ for the case of the \ensuremath{\mathbf{6}}\ in
Ref.~\cite{Shu:2009xf}, which was corrected in Ref.~\cite{Arhrib:2009hu}.}
\subsection{\boldmath Fitting the \ensuremath{{t\bar t}}\ asymmetry at the Tevatron}
Our strategy is to fit the models to the \ensuremath{{t\bar t}}\ related data from the Tevatron,
and then explore whether the resulting parameter space is consistent with other
experiments. We do not perform a $\chi^2$ fit to the data, because the required
correlations are not available. One can get a reasonable understanding of the
``goodness of fit'' of the models by plotting the constraints from various
experiments as functions of the models' two parameters.
\begin{figure}[tb]
\includegraphics[width=1.0\columnwidth]{threeplots}
\caption{Regions in $m_S$ vs.\ $\lambda$ parameter space for the \ensuremath{\mathbf{\ov 3}}\
model which are allowed by the total cross section contraint. The red (lighter)
and blue (darker) shaded regions correspond to $\sigma_\ensuremath{{t\bar t}} -
\sigma_\ensuremath{{t\bar t}}^{\rm (SM)} = (0.0 \pm 0.7)$\,pb and $(1.0 \pm 0.7)$\,pb,
respectively. The contours in the left [right] plot show the new physics
contribution to the \ensuremath{{t\bar t}}\ forward-backward asymmetry $A^{\ensuremath{{t\bar t}}}_{\rm
NP}(\ensuremath{m_{t\bar t}} > 450$~GeV) [$A^{\ensuremath{{t\bar t}}}_{\rm NP}(\ensuremath{m_{t\bar t}} < 450$~GeV)]. The total
asymmetry in each bin is obtained by adding the Standard Model contributions,
$A^{\ensuremath{{t\bar t}}}_{\rm SM}=0.09$ (left) and $0.04$ (right)~\cite{Aaltonen:2011kc}.}
\label{fig:threeplots}
\end{figure}
In agreement with previous work we find that the well-measured total cross
section, $\sigma(p\bar p \to \ensuremath{{t\bar t}}) = (7.5 \pm 0.48)$\,pb~\cite{CDFsigma},
provides the most important constraint on the parameter space. There is
currently some debate in the literature about the precise value of the
theoretical prediction. State of the art
NLO+NNL~\cite{Moch:2008qy,Cacciari:2008zb,Kidonakis:2008mu} calculations quote
about a 10\% uncertainty, and fit the data well. On the other hand, recent
calculations resumming threshold logs obtain lower values, around
6.5\,pb~\cite{Ahrens:2010zv}. Given this uncertainty in the predictions, we
choose a conservative approach. We plot the allowed regions corresponding to the
central values of each of the theory predictions, with approximately 10\% total
uncertainties, $\pm 0.7$\,pb. Thus, the NLO+NNL calculations allow the new
physics contribution to the \ensuremath{{t\bar t}}\ cross section to account for $\sigma_\ensuremath{{t\bar t}}
- \sigma_\ensuremath{{t\bar t}}^{\rm (SM)} = (0.0 \pm 0.7)$\,pb, and yield the red (lighter)
shaded allowed regions in Figure~\ref{fig:threeplots} for the \ensuremath{\mathbf{\ov 3}}\ model.%
\footnote{In our calculations of the QCD and new physics contributions we
applied a K-factor of $1.3$. We use the CTEQ-5L parton distribution functions~\cite{Lai:1999wy} implemented in Mathematica, and checked that MSTW 2008~\cite{Martin:2009iq} gives compatible results.}
Using the threshold resummed predictions, there is additional room for new
physics contributions to the \ensuremath{{t\bar t}}\ cross section, and the blue (darker) shaded
regions show $\sigma_\ensuremath{{t\bar t}} - \sigma_\ensuremath{{t\bar t}}^{\rm (SM)} = (1.0 \pm 0.7)$\,pb. In
the former case there are two allowed regions in parameter space. The less
interesting region is near the Standard Model and has small Yukawa couplings and
therefore small effects on the forward-backward asymmetry. The regions of
interest correspond to narrow bands in parameter space with larger Yukawa
couplings. This region is consistent with the total cross section constraint
because of a cancellation: the new physics squared contribution to the cross
section cancels against the interference with the QCD contribution. Since the
total cross section is the most precise of the measurements, the allowed
parameter space is largely defined by these narrow bands.
Overlaid on the same plot are contours of constant predicted \ensuremath{{t\bar t}}\
forward-backward asymmetry in the high invariant mass bin, $\ensuremath{m_{t\bar t}} > 450$~GeV.
One sees that our model can generate parton level \ensuremath{{t\bar t}}\ asymmetries between
20\% and 30\% from the new physics alone. The standard model contributes an
additional asymmetry of 0.09 in this bin~\cite{Aaltonen:2011kc}. To a reasonable
approximation the two contributions can simply be added, and the combined
asymmetry overlaps the $1\,\sigma$ preferred region of the CDF measurement
$A^\ensuremath{{t\bar t}}(m_\ensuremath{{t\bar t}} > 450\,{\rm GeV}) = 0.475 \pm 0.10 \pm 0.05$. In the right
panel of Fig.~\ref{fig:threeplots} we show contours of constant forward-backward
asymmetry at low invariant masses, $\ensuremath{m_{t\bar t}} <450$~GeV, overlaid with the same
regions allowed by the total cross section constraint. One sees that even though
a large asymmetry of 20--30\% is generated at high invariant masses, the
asymmetry at low invariant masses is always less than 10\% in the region of the
parameter space allowed by the total cross section constraint. Based on these
``eyeball'' fits we define two benchmark points which provide reasonable fits to
the CDF asymmetry and total cross section,
\begin{eqnarray}\label{studypoints}
&& \mbox{(1) ``low mass":}\quad\ m_S = 500\,\mathrm{GeV}, \quad\ \lambda = 2.3\,,
\qquad\nonumber\\
&& \mbox{(2) ``high mass":}\quad m_S = 750\,\mathrm{GeV}, \quad\ \lambda = 3.0\,.
\end{eqnarray}
The same fit result for the \ensuremath{\mathbf{6}}\ model is shown in Fig.~\ref{fig:sixplots}.
This model cannot accommodate as high \ensuremath{{t\bar t}}\ forward-backward asymmetries as
the \ensuremath{\mathbf{\ov 3}}. Moreover, as discussed below, the dijet constraints already rule
it out. For definiteness we also define two benchmark points for this model with
$(m_S=1200\ {\rm GeV},\ \lambda=1.9)$ and $(m_S=500\ {\rm GeV},\ \lambda=0.8)$.
The two points generate high invariant mass ($\ensuremath{m_{t\bar t}} >$ 450 GeV) asymmetries
of 12\% and 6\%, respectively.
\begin{figure}[tb]
\includegraphics[width=1.0\columnwidth]{sixplots}
\caption{The contour plots as in Fig.~\ref{fig:threeplots}, but for the \ensuremath{\mathbf{6}}\
model.}
\label{fig:sixplots}
\end{figure}
CDF also provided a binning of the \ensuremath{{t\bar t}}\ asymmetry by the rapidity difference
between the top anti-top pair. There is a large asymmetry at large rapidity
difference and a small asymmetry at small rapidity difference. We find that in
our model this binning does not generate new constraints on the parameter space.
Our model is consistent with the rapidity binned data within the sizable
$1\,\sigma$ errors quoted in Ref.~\cite{Aaltonen:2011kc}.
In Table~\ref{tab:studypoints} we summarize the predictions of the two benchmark
points in comparison to the measurements. Our predictions for the asymmetry
include both new physics and standard model
contributions~\cite{Aaltonen:2011kc}.
\begin{table}[b]
\begin{tabular}{c|c|c c}
Observable & Measurement & ~Point\,(1)~ & Point\,(2) \\
\hline\hline
$A^\ensuremath{{t\bar t}}(m_\ensuremath{{t\bar t}} > 450\,{\rm GeV})$~ & $0.475 \pm 0.114$ & 0.30 & 0.36 \\
$A^\ensuremath{{t\bar t}}(m_\ensuremath{{t\bar t}} < 450\,{\rm GeV})$ & ~$-0.116 \pm 0.153$~ & 0.10 & 0.07 \\
\hline
$A^\ensuremath{{t\bar t}}(|\Delta y| \geq 1)$ & $0.611 \pm 0.256$ & 0.42 & 0.46 \\
$A^\ensuremath{{t\bar t}}(|\Delta y| < 1)$ & $0.026 \pm 0.118$ & 0.12 & 0.12 \\
\hline
$\sigma_\ensuremath{{t\bar t}} - \sigma_\ensuremath{{t\bar t}}^{\rm (SM)}$ & \mbox{see the text}
& 0.7\,pb & 0.5\,pb \\
\hline\hline
\end{tabular}
\caption{Comparison of the CDF binned asymmetry
measurements~\cite{Aaltonen:2011kc} with the benchmark points in
Eq.~(\ref{studypoints}) for the \ensuremath{\mathbf{\ov 3}}\ model.}
\label{tab:studypoints}
\end{table}
Finally, the shape of the \ensuremath{{t\bar t}}\ cross section, ${\rm d}\sigma_\ensuremath{{t\bar t}} / {\rm d}\ensuremath{m_{t\bar t}}$,
has also been measured~\cite{Aaltonen:2009iz}. The SM prediction for this
spectrum is also known at NLO+NNL~\cite{Moch:2008qy, Cacciari:2008zb,
Kidonakis:2008mu}. However, the theoretical uncertainties are larger than for
the total cross section, especially at large \ensuremath{m_{t\bar t}}~\cite{Ahrens:2010zv}. We
do not perform a fit to the spectrum, and only compare the prediction of our
model for the cross section in a high invariant mass bin, 700 GeV$< \ensuremath{m_{t\bar t}} <$
800 GeV, following~\cite{Blum:2011up}. This bin is fairly far from the bulk of
the \ensuremath{{t\bar t}}\ data, and therefore tests a different region of the \ensuremath{{t\bar t}}\ spectrum
than the total cross section. In addition, the cross section at the highest
invariant masses is expected to be the most sensitive to the new physics
contributions. We find that there is significant tension between the measured
cross section and the model prediction for $\sigma_\ensuremath{{t\bar t}} - \sigma_\ensuremath{{t\bar t}}^{\rm
(SM)}$ in this bin, with the latter being about twice the SM prediction (a
similar excess is found in other models in the literature). Given that both
theoretical and experimental uncertainties are substantial for the tail of the
\ensuremath{{t\bar t}}\ spectrum, we set this issue aside and explore what new information can
be obtained from LHC experiments in the context of this model.
\subsection{Dijet constraints}
We next study the dijet constraints on the \ensuremath{\mathbf{\ov 3}}\ model and contrast them
with the corresponding constraints for the \ensuremath{\mathbf{6}}. Since the coupling of the
\ensuremath{\mathbf{\ov 3}}\ in flavor space is $\epsilon_{rab} U^a U^b$ it does not mediate
$uu\to uu$ scattering. This is fortunate because a scalar $s$-channel resonance
of the leading $uu$ parton luminosity with coupling $\lambda > 1$ can be ruled
out for masses $\sim$~0.4--3~TeV with the recent CMS analysis of dijet angular
distributions~\cite{Khachatryan:2011as}. The \ensuremath{\mathbf{\ov 3}}\ model does predict an
$s$-channel dijet resonance in $uc \to uc$ scattering. As we will see the
sensitivity of the CMS analysis is close to what would be required to discover
it. The model also gives rise to dijets from $u$-channel $\ensuremath{{u\bar u}}\to \ensuremath{{c\bar c}}$
processes both at the LHC and at the Tevatron. However, this process is less
sensitive than $\ensuremath{{u\bar u}}\to \ensuremath{{t\bar t}}$ or $uc \to uc$.
The CMS collaboration measured the differential dijet cross section ${\rm d} \sigma /
{\rm d}\chi$, where $\chi = (1+|\cos\theta|) /(1-|\cos\theta|)$ is defined such that
it makes the QCD prediction for ${\rm d} \sigma / {\rm d}\chi$ flat at small $\theta$. In
our models we predict
\begin{eqnarray}\label{dsigmadchi}
\frac{{\rm d}\sigma}{{\rm d}\chi} &=&
\frac{1}{36\pi s (1+\chi)^2} \bigg[ g_s^4 \left( \frac{s^2+u^2}{t^2}
+ \frac{s^2+t^2}{u^2} -\delta\, \frac{2s^2}{3ut} \right) \nonumber\\
&&{} + D_0\, g_s^2\, \lambda^2\left( \frac{s}{t} + \frac{s}{u} \right)
\frac{s(s-m_S^2)}{(s-m_S^2)^2 + m_S^2\, \Gamma_S^2} \nonumber\\
&&{} + 3 D_2\, \lambda^4 \frac{s^2}{(s-m_S^2)^2 + m_S^2\, \Gamma_S^2} \bigg]\,.
\end{eqnarray}
For $u u \rightarrow u u$ in the \ensuremath{\mathbf{6}}\ model we have $\delta = 1$, $D_0=-2$, and
$D_2=2$, while for $uc \rightarrow uc$ in the \ensuremath{\mathbf{\ov 3}}\ model $\delta = 0$,
$D_0=1$, and $D_2=1/2$. In both models the width of the resonance is $\Gamma_S =
m_S\, \lambda^2/(8\pi)$. Note that $\chi$ does not distinguish between $\pm
\cos\theta$, and consequently when computing ${\rm d} \sigma / {\rm d}\chi$ for processes
where the final state is composed of distinguishable particles, we summed the
contributions of both signs of $\cos\theta$.
The CMS analysis contains 9 different dijet invariant mass regions, from
250--350~GeV to above 2.2~TeV. For optimal sensitivity to $s$-channel new
physics beyond the standard model, one looks for a rise in the number of events
in the most central bin, $1<\chi<2$. QCD predicts an approximately flat
distribution over all $\chi$ bins with a small rise in the central bin. The
measured cross section in each bin is normalized to the total dijet cross
section sumed over all $\chi$ bins. The remaining uncertainty in the central
$\chi$ bin is estimated by CMS to be about 10\%~\cite{Khachatryan:2011as}.
\begin{figure}[t]
\includegraphics[width=\columnwidth,clip]{chitriplet}
\caption{Normalized dijet cross sections in $\delta\chi=1$ bins for $1<\chi<6$.
The blue (darker) histogram shows the QCD prediction, while the green (lighter)
histogram shows the normalized sum of QCD and the new physics $uc\to uc$
contribution to the dijet rate in the \ensuremath{\mathbf{\ov 3}}\ model. The left and right
plots show $500\,\mathrm{GeV} < m_{jj} < 650\,\mathrm{GeV}$ and $850\,\mathrm{GeV} < m_{jj} <
1100\,\mathrm{GeV}$, respectively. These are the mass ranges for which the largest
excess above QCD is predicted for the two benchmark points (1) and (2) in
Eq.~(\ref{studypoints}). In either case the excess in the most central bin is
about 10\% which is comparable to the uncertainties quoted
in~\cite{Khachatryan:2011as}.}
\label{fig:dijet3}
\end{figure}
The scalar resonance in $uc\to uc$ scattering in the \ensuremath{\mathbf{\ov 3}}\ model does
predict an increased dijet rate at small $\chi$, most strongly for dijet
invariant masses near $m_S$. In Fig.~\ref{fig:dijet3} we show the predicted
normalized dijet rate $(1 / \sigma_{\rm dijet})\,({\rm d}\sigma_{\rm dijet} /
{\rm d}\chi)$ for our model compared with QCD. For each of the two benchmark points
we show the first five $\chi$ bins for dijet invariant masses near $m_S$. There
is a discernible rise in the cross section above QCD, but the systematic
uncertainties shown in Ref.~\cite{Khachatryan:2011as} are of the same size or
larger than the new physics effect in Fig.~\ref{fig:dijet3}. However, increased
statistics combined with data driven background subtraction (at higher and lower
$m_{\rm dijet}$) may be sensitive to this $uc$ resonance in the future. If the
data starts to show signs of a resonance, it would be very interesting to employ
(even limited) charm tagging to confirm the presence of charm quarks.
\begin{figure}[t]
\includegraphics[width=\columnwidth,clip]{chisextet}
\caption{Same as Fig.~\ref{fig:dijet3}, but for $uu\to uu$ in
the \ensuremath{\mathbf{6}}\ model. The left plot shows the contribution for $500\,\mathrm{GeV} < m_{jj}
< 650\,\mathrm{GeV}$, while the right plot shows that for $1100\,\mathrm{GeV} < m_{jj} <
1400\,\mathrm{GeV}$.}
\label{fig:dijet6}
\end{figure}
For comparison, we also show the predicted dijet $\chi$ spectrum for the \ensuremath{\mathbf{6}}\
model in Fig.~\ref{fig:dijet6}. Here the resonance is in $uu \to uu$ scattering,
and consequently the new physics signal is very large. We see that the heavy
benchmark point for the \ensuremath{\mathbf{6}}\ is ruled out by a large margin. The light
benchmark point predicts a rise of about 30\% beyond QCD which is also ruled
out.
\subsection{(Un)Naturalness: SUSY and the Landau pole}
Fundamental scalar particles such as our triplet suffer from a naturalness
problem, as their masses are quadratically sensitive to ultraviolet scales.
Therefore, one should think of this theory as the low energy limit of a more
complete theory in which the scalar arises as a composite. Alternatively, the
scalar mass may be made natural with supersymmetry. It is straightforward to
supersymmetrize our model and we briefly describe the resulting model here.
The minimal supersymmetric version of our model is the MSSM with one extra
chiral superfield $\overline S$ with identical gauge and flavor quantum numbers
as our scalar, and another chiral superfield $S$ with the opposite gauge and
flavor quantum numbers. The Yukawa coupling of Eq.~(\ref{3def}) is lifted to a
superpotential term $W = \lambda\, UU\!S$ where now $U$ and $S$ denote the full
superfields. Flavor and gauge indices are contracted as in the
non-supersymmetric theory. The mass of the scalar gets contributions from
supersymmetry breaking and supersymmetry preserving terms. If it primarily
arises from supersymmetry breaking one expects the $S+\overline S$ fermions to
be light. These fermions are R-parity odd, they can be produced in pairs at the
LHC or appear in cascade decays of squarks, giving events with high jet
multiplicities and missing energy. The resulting signatures are similar to the
ones of a light gluino.
A separate issue the reader may worry about is that the relatively large Yukawa
couplings invalidate perturbation theory and lead to a Landau pole not far from
the mass of the triplet. The leading term in the beta function for~$\lambda$~is
\begin{equation}
16 \pi^2\, \frac{{\rm d} \lambda}{{\rm d}(\ln \mu)} = 4\, \lambda^3 \,.
\end{equation}
We see that there are no large multiplicity factors associated with color and
flavor and the loop expansion parameter is approximately $\lambda^2/(4\pi^2)$,
which is perturbative for the couplings of interest. The solution to the
renormalization group equation is $1/\lambda^2(\mu) = \ln(\Lambda/\mu)/(2
\pi^2)$, where $\Lambda$ is the Landau pole.
For example, $\lambda(m_S)=2.3$ and
$m_S=500$~GeV gives $\Lambda \sim 21$~TeV. (The other benchmark point in
Eq.~(\ref{studypoints}), $\lambda(m_S)=3$ and $m_S=750$~GeV, gives $\Lambda \sim
7$~TeV.) In either case the Landau pole is far enough that dimension-6 operators
suppressed by $16\pi^2/\Lambda^2$ are much smaller than the $S$ exchange
diagrams.
\section{\boldmath LHC signals: $uc$ resonance and $pp\to \ensuremath{{t\bar t}} j$}
\label{sec:lhc}
While \ensuremath{{t\bar t}}\ production at the Tevatron is dominated by $\ensuremath{{q\bar q}} \to \ensuremath{{t\bar t}}$,
the total $pp\to\ensuremath{{t\bar t}}$ cross section at the LHC is dominated by $gg\to\ensuremath{{t\bar t}}$,
which does not exhibit a forward-backward asymmetry. By measuring \ensuremath{{t\bar t}}\ pairs
at higher average rapidity or higher invariant mass, one can enhance the \ensuremath{{q\bar q}}\
initial state, and thus possibly check the Tevatron observation. Another
possibility is to measure the difference of the $t$ and $\bar t$ production
rates at high rapidity (maybe at LHCb), where one could be sensitive to the
asymmetry without reconstructing both the $t$ and the $\bar t$ particles in the
same event.
Since neither of these measurements are straightforward, it is worthwhile to
explore possible other signatures. One exciting possibility is that
the colored scalars discussed could manifest themselves in future
higher sensitivity dijet analyses. The constraints are already quite
powerful for dijet resonances of valence quarks (for example
ruling out the \ensuremath{\mathbf{6}}). The \ensuremath{\mathbf{\ov 3}}\ appears as an $s$-channel
resonance only in $uc \to uc$ scattering and in the
$u$-channel in $\ensuremath{{u\bar u}}\to \ensuremath{{c\bar c}}$.
Depending on the choice of parameters in our model either of these
processes may be observable with increasing statistics, and would
be a spectacular discovery, especially if combined with even some
limited charm tagging.
\begin{figure}[tb]
\includegraphics[width=.5\columnwidth]{dog}
\caption{The flavor conserving diquark contribution to $\ensuremath{{t\bar t}} j$ production at
the LHC. Two other diagrams with the gluon attached to the $S$ and to the
$\bar t$ are not shown.}
\label{fig:puppy}
\end{figure}
Another promising signature, discussed recently in Ref.~\cite{Gresham:2011dg}
(in the context of flavor violating models), is to look for $ug \to S_{ut} \bar
t \to \ensuremath{{t\bar t}} u$ production (see Fig.~\ref{fig:puppy}). Since in our preferred
scenario $m_S \gg m_t$, the $u$-quark jet from the decay of $S$ would have high
$p_T^{\rm jet}$, where the standard model background (known at
NLO~\cite{Dittmaier:2008uj}) is suppressed. The $\ensuremath{{t\bar t}} j$ cross section from
our model would be quite large. For example, at the 7\,TeV LHC the new physics
contribution is 2\,pb for $m_S = 600\,\mathrm{GeV}$ and $\lambda=1/\sqrt{2}$ (and
7.7\,pb for $m_S = 400\,\mathrm{GeV}$, $\lambda=1/\sqrt{2}$)~\cite{Gresham:2011dg}, and
for arbitrary coupling it is enhanced by $2\,\lambda^2$. Comparing with recent
predictions for the $\ensuremath{{t\bar t}} j$ distributions at the 7\,TeV LHC~\cite{alioli,
Kardos:2011qa}, we find that simply cutting on $p_T$ of the hardest jet, our
signal is somewhat smaller than the SM background. The experimental sensitivity
can probably be optimized and substantially enhanced by measuring the $\ensuremath{{t\bar t}} j$
rate as a function of both $p_T^{\rm jet}$ and \ensuremath{m_{t\bar t}}\ (since for the SM
background, but not for the signal coming from $\bar t S$ decay, $p_T^{\rm jet}$
and \ensuremath{m_{t\bar t}}\ are anti-correlated), and choosing suitable cuts on both
variables.
In addition, pair production of $S S^*$ gives rise to a \ensuremath{{t\bar t}}\ensuremath{{u\bar u}}\ (or $4j$)
final state where the unflavored jets have large $p_T^{\rm jet}$. The rate is
smaller than $\bar t S$ production, however the second hard jet is advantageous
for rejecting SM backgrounds. Especially when the LHC gets to higher energy, the
sensitivity in this channel may become competitive with the $\ensuremath{{t\bar t}} j$ signal.
We conclude that the most promising signals of the \ensuremath{\mathbf{\ov 3}}\ at the LHC
are either the $s$-channel resonance contribution whose $uc \to uc$ origin might
become established even with very limited charm tagging, or, especially, $\ensuremath{{t\bar t}}
j$ production at high $p_T^{\rm jet}$ (maybe combined with an \ensuremath{m_{t\bar t}}\ cut).
Without a much more detailed analysis than the present study, it is not possible
to determine how the sensitivities of these two searches will compare when
the actual experimental analyses are carried out. We hope that both can be
pursued by ATLAS and CMS, as they are also interesting searches for models
beyond those motivated by the Tevatron \ensuremath{{t\bar t}}\ forward-backward asymmetry
discussed in this paper.
{\it Note added:} The predicted \ensuremath{{t\bar t}}\ cross section rises in our model compared to the SM for high invariant masses. For example, the cross section for \ensuremath{m_{t\bar t}}\ above 1 TeV at the 7 TeV LHC is larger than the QCD cross section by a factor of 2 to 3 in the preferred region of parameter space. Therefore future
measurements (combined with improved theory predictions) should discover or rule
out our model. Note that the cross section above 1 TeV due to the exchange of a
lighter particle cannot be approximated with a higher dimensional operator.
Therefore the results of \cite{Delaunay:2011gv,AguilarSaavedra:2011vw} do not
apply to our model.
\begin{acknowledgements}
We thank Nima Arkani-Hamed, Andy Cohen, Liam Fitzpatrick, Fabio Maltoni, Gilad
Perez, David Shih, Brock Tweedie, and Tomer Volansky for helpful conversations.
This work was supported in part by the Director, Office of Science, Office of
High Energy Physics of the U.S.\ Department of Energy under contract
DE-AC02-05CH11231 (ZL) and DE-FG02-01ER-40676 (MS and GMT).
\end{acknowledgements}
|
\section{Introduction}
\label{s_introduction}
One of the most successful dynamical systems ideas in the study
of turbulent flows has been the \textit{Proper Orthogonal Decomposition
(POD)} \cite{HLB96,Sir87abc}.
POD has been used to generate {\em reduced-order models (ROMs)} for the prediction and
control of structure dominated turbulent flows
\cite{AHLS88,bergmann2009enablers,buffoni2006low,NAMTT03,Pod01}.
The idea is straightforward:
Instead of using billions of local finite element basis functions equally distributed
in space, POD uses only a few (usually $\mathcal{O}(10)$) global basis functions
that represent the most energetic structures in the system.
Thus, the computational cost in a {\em direct numerical simulation (DNS)} of a complex flow
can be reduced by orders of magnitude when POD is employed.
Despite its widespread use (hundreds of papers being published every year),
POD has several well-documented drawbacks.
In this report, we address one of them, namely the numerical instability of a
straightforward POD Galerkin procedure applied to a complex flow
\cite{aubry1993preserving}.
To address this issue, we draw inspiration from the methodologies used in
the numerical stabilization of finite element discretization of convection-dominated
flows.
Specifically, we employ the \textit{variational multiscale (VMS)} approach used
by Layton in \cite{layton2002connection}, which adds artificial viscosity only to
the smallest resolved scales.
We also note that an approach similar to that used in \cite{layton2002connection}
was proposed by Guermond in \cite{guermond1999stabilization,guermond1999stabilisation}.
We emphasize that the VMS philosophy is particularly appropriate to the POD setting,
in which the hierarchy of small and large structures appears naturally.
Indeed, the POD modes are listed in decreasing order of their kinetic energy content.
We also note that, although a VMS-POD approach was announced in
\cite{borggaard2008reduced,borggaard2011artificial}
and another VMS-POD approach was used in \cite{bergmann2009enablers},
to the authors' knowledge this is the first time that
the VMS formulation in \cite{layton2002connection}
has been applied in a POD setting.
In this report, the new VMS-POD model is analyzed and tested in the numerical
approximation of a {\em convection-dominated convection-diffusion} problem
\begin{eqnarray}
\label {cdcd}
\left\{
\begin{array}{ll}
u_t
- \varepsilon \, \Delta u
+ {\bf b} \cdot \nabla u
+ g u
= f
& \qquad \text{ in } (0, T] \times \Omega \, , \\
u(x,0) = u_0(x)
& \qquad \text{ in } \Omega \, , \\
u(x, t) = 0
& \qquad \text{ on } (0, T]\times \partial \Omega \, ,
\end{array}
\right.
\end{eqnarray}
where $\varepsilon \ll 1$ is the diffusion parameter,
${\bf b}$ with $\| {\bf b} \| = \mathcal{O}(1)$ the given convective field,
$g$ the reaction coefficient,
$f$ the forcing term,
$\Omega \subset \mathbbm{R}^2$ the computational domain,
$t \in [0,T]$, with $T$ the final time, and
$u_0(\cdot)$ the initial condition.
Without loss of generality, we assume in what follows that the boundary conditions
are homogeneous Dirichlet.
We emphasize that the new VMS-POD model targets turbulent flows described by
the {\em Navier-Stokes equations (NSE)}.
We chose the mathematical setting in \eqref{cdcd}, however, because it
is simple, yet relevant to our ultimate goal (since $\varepsilon \ll \| {\bf b} \|$).
Of course, once we fully understand the behavior of the new VMS-POD model in
this simplified setting, we will analyze and apply it in the NSE setting.
The rest of the paper is organized as follows.
In Section \ref{s_vms_pod}, we briefly describe the POD methodology
and introduce the new VMS-POD model.
The error analysis for the finite element discretization of the new
model is presented in Section \ref{s_estimates}.
The new methodology is tested numerically in Section \ref{s_numerics}
for a problem displaying shock-like phenomena.
Finally, Section \ref{s_conclusions} presents the conclusions and future
research directions.
\section{Variational Multiscale Proper Orthogonal Decomposition}
\label{s_vms_pod}
\subsection{Proper Orthogonal Decomposition}
\label{ss_pod}
In this section, we briefly describe the POD.
For a detailed presentation, the reader is referred to \cite{HLB96,KV99,Sir87abc}.
Let $X$ be a real Hilbert space endowed with the inner product $(\cdot,\cdot)_X$,
and $u(\cdot,t)\in X, t\in[0,T]$ be the state variable of a dynamical system.
Given the time instances $t_1, \ldots, t_N \in [0,T]$, we consider the ensemble of
snapshots
\begin{eqnarray}
V := \mbox{span}\left\{ u(\cdot,t_1), \ldots, u(\cdot,t_N) \right\},
\label{snapshots}
\end{eqnarray}
with dim $V = d$.
The POD seeks a low-dimensional basis $\{ \varphi_1, \ldots, \varphi_r \}$,
with $r \ll d$, which optimally approximates the input collection.
Specifically, the POD basis satisfies
\begin{eqnarray}
\min \frac{1}{N} \sum_{i=1}^N \left\| u(\cdot,t_i) -
\sum_{j=1}^r \bigl( u(\cdot,t_i) \, , \, \varphi_j(\cdot) \bigr)_X \, \varphi_j(\cdot)
\right\|_X^2 \, ,
\label{pod_min}
\end{eqnarray}
subject to the conditions that $(\varphi_i,\varphi_j)_X = \delta_{ij}, \ 1 \leq i, j \leq r$.
In order to solve \eqref{pod_min}, we consider the eigenvalue problem
\begin{eqnarray}
K \, v = \lambda \, v \, ,
\label{pod_eigenvalue}
\end{eqnarray}
where $K \in \mathbbm{R}^{N \times N}$, with
$\displaystyle K_{ij} = \frac{1}{N} \, \bigl( u(\cdot,t_j) , u(\cdot,t_i) \bigr)_X \,$,
is the snapshot correlation matrix,
$\lambda_1 \geq \lambda_2 \geq \ldots \geq \lambda_d >0$ are the positive eigenvalues, and
$v_k, \, k = 1, \ldots, d$ are the associated eigenvectors.
It can then be shown (see, e.g., \cite{HLB96,KV99}), that the solution
of \eqref{pod_min} is given by
\begin{eqnarray}
\varphi_{k}(\cdot) = \frac{1}{\sqrt{\lambda_k}} \, \sum_{j=1}^{N} (v_k)_j \, u(\cdot , t_j),
\quad 1 \leq k \leq r,
\label{pod_basis_formula}
\end{eqnarray}
where $(v_k)_j$ is the $j$-th component of the eigenvector $v_k$.
It can also be shown that the following error formula holds:
\begin{eqnarray}
\frac{1}{N} \sum_{i=1}^N \left\| u(\cdot,t_i) -
\sum_{j=1}^r \bigl( u(\cdot,t_i) \, , \, \varphi_j(\cdot) \bigr)_X \, \varphi_j(\cdot)
\right\|_X^2
= \sum_{j=r+1}^{d} \lambda_j \, .
\label{pod_error_formula}
\end{eqnarray}
In what follows, we will use the notation
$X^r = \mbox{span}\{ \varphi_1, \varphi_2, \ldots, \varphi_r \} \, .$
Although $X$ can be any real Hilbert space, in what follows we consider
$X := H_0^1(\Omega)$.
In the form it has been presented so far, POD seems to be only
a data compression technique.
Indeed, equation \eqref{pod_min} simply says that the POD basis is the
best possible approximation of order $r$ of the given data set.
In order to make POD a predictive tool, one couples the POD with the
Galerkin procedure.
This, in turn, yields a ROM, i.e., a dynamical
system that represents the evolution in time of the Galerkin truncation.
We now briefly present the derivation of this ROM, we highlight one of its
main drawbacks and we propose a method to address this deficiency.
The POD-Galerkin truncation is the approximation $u_r \in X^r$ of $u$:
\begin{eqnarray}
u_r(x, t)
:= \sum_{j=1}^{r} a_j(t) \, \varphi_j(x) .
\label{pod_g_truncation}
\end{eqnarray}
Plugging \eqref{pod_g_truncation} into \eqref{cdcd} and multiplying by test
functions in $X^r \subset X$ yields the {\em POD-Galerkin (POD-G)} model
\begin{eqnarray}
\hspace*{0.8cm} (u_{r, t} , v_r)
+ \varepsilon (\nabla u_r , \nabla v_r)
+ ({\bf b} \cdot \nabla u_r , v_r)
+ (g \, u_r, v_r)
= (f , v_r)
\quad \forall \, v_r \in X^r .
\label{pod_g}
\end{eqnarray}
The main advantage of the POD-G model \eqref{pod_g} over a
straightforward {\em finite element (FE)} discretization of \eqref{cdcd} is clear -
the computational cost of the former is dramatically lower than that of the latter.
There are, however, several well-documented disadvantages of \eqref{pod_g},
such as its numerical instability in convection-dominated flows \cite{sirisup2004spectral}.
To address this issue, we draw inspiration from the methodologies used in
numerical stabilization of finite element discretizations of such flows.
\subsection{Variational Multiscale}
\label{ss_vms}
The \textit{Variational Multiscale (VMS)} method introduced by Hughes and his group
\cite{Hug95,HMJ00,HMOW01,HOM01} has been successful in the numerical stabilization
of turbulent flows
\cite{gravemeier2006consistent,gravemeier2004three,john2005finite,john2008finite_a,john2008finite_b,john2006two}.
The idea in VMS is straightforward:
Instead of adding artificial viscosity to all resolved scales, in VMS artificial viscosity is
only added to the smallest resolved scales.
Thus, the small scale oscillations are eliminated without polluting the large scale
components of the approximation.
The VMS method has been extensively developed, various numerical methods
being used.
The finite element discretization of the resulting VMS model has evolved in several
directions:
Hughes and his group proposed a VMS formulation for the NSE
in which a Smagorinsky model \cite{BIL05,Sma63} was added only to the smallest resolved scales
\cite{Hug95,HMJ00,HMOW01,HOM01}.
A different type of VMS approach, based on the residual of the NSE, was proposed
in Bazilevs et al. \cite{bazilevs2007variational}.
One of the earliest VMS ideas for convection-dominated convection-diffusion equations
was proposed by Guermond in \cite{guermond1999stabilization,guermond1999stabilisation}.
In this VMS formulation, the smallest scales were modeled by using finite element
spaces enriched with bubble functions.
Layton proposed in \cite{layton2002connection} a VMS approach similar to that of Guermond.
In this VMS approach, however, the smallest resolved scales were modeled by projection
on a coarser mesh.
The VMS approach proposed in \cite{layton2002connection} was extended to the NSE
in a sequence of papers by John and Kaya
\cite{john2005finite,john2008finite_a,john2008finite_b,john2006two}.
The variational formulation used by the FE methodology fits very well with the VMS
approach.
The definition of the smallest resolved scales, however, often poses many challenges
to the FE method.
Indeed, one needs to enrich the FE spaces with bubble functions \cite{guermond1999stabilization,guermond1999stabilisation},
consider hierarchical FE bases \cite{HMJ00}, or use a projection on a coarser mesh
\cite{layton2002connection}.
\subsection{The VMS-POD Model}
\label{ss_model}
POD represents the perfect setting for the VMS methodology, since the hierarchy of
the basis is already present.
Indeed, the POD basis functions are already listed in descending order of their kinetic
energy content.
Based on this observation, we next propose a VMS based POD model.
To this end, we consider the following spaces:
$
X := H_0^1(\Omega), \,
X^h \subset H_0^1(\Omega), \,
X^r := \text{ span} \{ \varphi_1, \varphi_2, \ldots, \varphi_r \}, \,
X^R := \text{ span} \{ \varphi_1, \varphi_2, \ldots, \varphi_R \}, \, \text{ where } R < r,
\, \text {and } \,
L^R,
$
where $L^R$ will be defined later.
Note that
$
X^R \subseteq X^r \subset X^h \subset X.
$
We also consider $P_R : L^2(\Omega) \longrightarrow L^R$, the orthogonal projection of
$L^2(\Omega)$ on $L^R$, defined by
\begin{eqnarray}
( u - P_R u , v_R ) = 0
\qquad \forall \, v_R \in L^R .
\label{def_pr}
\end{eqnarray}
Let also $P_R^{'} := \mathbbm{I} - P_R$.
We are now ready to define the \textit{Variational Multiscale Proper Orthogonal Decomposition
(VMS-POD) model}:
\begin{eqnarray}
\begin{split}
(u_{r, t} , v_r)
& + \varepsilon \, (\nabla u_r , \nabla v_r)
+ \alpha \, (P_R^{'} \nabla u_r , P_R^{'} \nabla v_r) \\
& + ({\bf b} \cdot \nabla u_r , v_r)
+ (g u_r, v_r)
= (f , v_r)
\quad \forall \, v_r \in X^r .
\label{vms_pod}
\end{split}
\end{eqnarray}
The third term on the LHS of \eqref{vms_pod} represents the artificial viscosity
that is added only to the smallest resolved scales of the gradient.
We note that, although a VMS-POD approach was announced in
\cite{borggaard2008reduced,borggaard2011artificial}
and another one was used in \cite{bergmann2009enablers}, to the authors' knowledge
this is the first time that
the VMS formulation in \cite{layton2002connection} is applied in a POD setting.
In the next two sections, we will first estimate the error made in the finite element
discretization of the new VMS-POD model \eqref{vms_pod} and then use
it in a numerical test.
\section{Error Estimates}
\label{s_estimates}
In this section, we prove estimates for the average error
$\displaystyle \frac{1}{N+1} \, \sum_{n=0}^{N} \| u^n - u_r^n \|$,
where the approximation $u^n$ is the solution of \eqref{cdcd_weak}
(the weak form of \eqref{cdcd})
and
$u_r^n$ is the solution of \eqref{vms_pod_fe}
(the FE discretization of the VMS-POD model \eqref{vms_pod}).
To this end, we follow the approach in \cite{heitmann2007subgridscale}
(see also \cite{kaya2004numerical}).
We emphasize, however, that our presentation is different in that it has to
include several results pertaining to the POD setting.
To this end, we use some of the developments in \cite{luo2009finite}
(see also \cite{HPS05,KV01,LCNY08,schu2010reduced}).
We start by introducing some notation and we list several results that will
be used throughout this section.
For clarity of notation, we will denote by $C$ a generic
constant that can depend on all the parameters in the system, except on
$d$ (the number of POD modes retained in the Galerkin truncation),
$N$ (the number of snapshots),
$r$ (the number of POD modes used in the POD-G model \eqref{pod_g}),
$R$ (the number of POD modes used in the projection operator in the
VMS-POD model \eqref{vms_pod}),
$h$ (the mesh-size in the FE discretization),
$\alpha$ (the artificial viscosity coefficient), and
$\varepsilon$ (the diffusion coefficient).
Of particular interest is the independence of the generic constant $C$ from
$\varepsilon$.
Indeed, we will prove estimates that are {\em uniform with respect to $\varepsilon$},
which is important when convection-dominated flows (such as the NSE) are considered.
We introduce the bilinear forms
$
b(u,v)
:= ({\bf b} \cdot \nabla u , v)
+ (g \,u , v) , \
a(u, v)
:= \varepsilon (\nabla u, \nabla v)
+ b(u,v) , \
$
and
$
A(u,v)
:= a(u, v)
+ \alpha \, (P_R^{'} \nabla u, P_R^{'} \nabla v) .
$
We also consider the weighted norm
$\| u \|_{a, b, \alpha}^2
:= a \, \| u \|^2
+ b \, \| \nabla u \|^2
+ \alpha \, \| P_R^{'} \nabla u \|^2 .
$
We now make the following assumption, which is used in proving the well-posedness
of the weak formulation of \eqref{cdcd}.
\begin{assumption}[Coercivity and Continuity]
\begin{eqnarray}
g
- \frac{1}{2} \, \nabla \cdot {\bf b}
\geq \beta
> 0
\quad \text{and} \quad
\max \{ \| g \| , \| {\bf b} \| \}
= \gamma > 0 .
\end{eqnarray}
\label{assumption_coercivity}
\end{assumption}
For the FE discretization of \eqref{cdcd}, we consider a family
of finite dimensional subspaces $X^h$ of $X = H_0^1(\Omega)$, such that,
for all $v \in H^{m+1} \cap X$, the following assumption is satisfied.
\begin{assumption}[Approximability]
\begin{eqnarray}
\hspace*{0.6cm} \inf_{v_h \in X^h}
\left\{
\| v - v_h \|
+ h \, \| \nabla v - \nabla v^h \|
\right\}
\leq C \, h^{m+1} \, \| v \|_{m+1}
\qquad
1 \leq m \leq k ,
\label{assumption_approximability_1}
\end{eqnarray}
\label{assumption_approximability}
\end{assumption}
\noindent where $k$ is the order of accuracy of $\{ X^h \}$.
We also assume that the finite element spaces $\{ X^h \}$ satisfy the
following inverse estimate.
\begin{assumption}[FE Inverse Estimate]
\begin{eqnarray}
&& \| \nabla v_{h} \|
\leq C \, h^{-1} \ \| v_{h} \|
\qquad \forall \, v_h \in X^{h} .
\end{eqnarray}
\label{assumption_inverse_fem}
\end{assumption}
A similar inverse estimate for POD is proven in \cite{KV01}.
For completeness, we present it below.
We also include a new estimate and present its proof.
\begin{lemma}[POD Inverse Estimate]
Let $M_{r} \in \mathbbm{R}^{r \times r}$ with $M_{i j} = (\varphi_j , \varphi_i)$ be the POD mass matrix,
$H_{r} \in \mathbbm{R}^{r \times r}$ with $H_{i j} = (\nabla \varphi_j , \nabla \varphi_i)$ be the POD stiffness matrix,
$S_{r} \in \mathbbm{R}^{r \times r}$ with $S_{i j} = (\varphi_j , \varphi_i)_{H^1}$ be the POD mass matrix in the $H^1$-norm,
and $\| \cdot \|_2$ denote the matrix 2-norm.
Then, for all $v_r \in X^r$, the following estimates hold.
\begin{eqnarray}
\| v_r \|_{L^2}
&\leq& \sqrt{\| M_r \|_2 \, \| S_r^{-1} \|_2 } \, \| v_r \|_{H^1} \, ,
\label{lemma_inverse_pod_1} \\
\| v_r \|_{H^1}
&\leq& \sqrt{\| S_r \|_2 \, \| M_r^{-1} \|_2 } \, \| v_r \|_{L^2} \, ,
\label{lemma_inverse_pod_2} \\
\| \nabla v_r \|_{L^2}
&\leq& \sqrt{\| H_r \|_2 \, \| M_r^{-1} \|_2 } \, \| v_r \|_{L^2} \, .
\label{lemma_inverse_pod_3}
\end{eqnarray}
\label{lemma_inverse_pod}
\end{lemma}
\begin{proof}
The proof of estimates \eqref{lemma_inverse_pod_1} and \eqref{lemma_inverse_pod_2}
was given in \cite{KV01} (see Lemma 2 and Remark 2).
The proof of \eqref{lemma_inverse_pod_3} follows along the same lines:
Let $v_r = \sum_{i=1}^{r} x_j \varphi_j$ and $\bx = (x_1, \ldots, x_r)^T$.
From the definition of $H_r$, it follows that $\| \nabla v_r \|_{L^2}^2 = \bx^T \, H_r \, \bx$.
Since $H_r$ is symmetric, its matrix 2-norm is equal to its Rayleigh quotient \cite{demmel1997applied}:
$\| H_r \|_2 = \max\limits_{\bx \neq {\bf 0}} \frac{\bx^T \, H_r \, \bx}{\bx^T \, \bx}$.
Thus, we get:
\begin{eqnarray}
\| \nabla v_r \|_{L^2}^2 = \bx^T \, H_r \, \bx \leq \| H_r \|_2 \, \bx^T \, \bx \, .
\label{lemma_inverse_pod_3_1}
\end{eqnarray}
Furthermore, since $M_r^{-1}$ is also symmetric, we get
$\by^T \, M_r^{-1} \, \by \leq \| M_r^{-1} \|_2 \, \by^T \, \by$ for all vectors $\by \in \mathbbm{R}^r$.
We also note that, since $M_r$ is symmetric positive definite, we can use its Cholesky
decomposition $M_r = L_r \, L_r^T$, where $L_r$ is a lower triangular nonsingular
matrix \cite{demmel1997applied}.
Thus, letting $\by = L_r \, \bx$, we get:
\begin{eqnarray}
\| M_r^{-1} \|_2
\geq \frac{\by^T \, M_r^{-1} \, \by}{\by^T \, \by}
= \frac{\bx^T \, L_r^T \, (L_r^{-1})^T \, L_r^{-1} \, L_r \, \bx}{\bx^T \, L^T \, L \, \bx}
= \frac{\bx^T \, \bx}{\bx^T \, M_r \, \bx} \, .
\label{lemma_inverse_pod_3_2}
\end{eqnarray}
Inequalities \eqref{lemma_inverse_pod_3_1} and \eqref{lemma_inverse_pod_3_2}
imply the following inequality, which proves \eqref{lemma_inverse_pod_3}:
$
\| \nabla v_r \|_{L^2}^2
\leq \| H_r \|_2 \, \| M_r^{-1} \|_2 \, \bx^T \, M_r \, \bx
= \| H_r \|_2 \, \| M_r^{-1} \|_2 \, \| v_r \|^2_{L^2} \, .
$
\end{proof}
\begin{remark}
We note that, in our setting, \eqref{lemma_inverse_pod_3} can be improved.
Indeed, since $S_r$ is the identity matrix when $X = H_0^1$, we get:
\begin{eqnarray}
\| \nabla v_r \|_{L^2}
\leq \| v_r \|_{H^1}
\leq \sqrt{\| S_r \|_2 \, \| M_r^{-1} \|_2 } \, \| v_r \|_{L^2}
= \sqrt{\| M_r^{-1} \|_2 } \, \| v_r \|_{L^2} \, .
\label{remark_inverse_pod_1}
\end{eqnarray}
We note, however, that in general \eqref{remark_inverse_pod_1} might not hold.
\label{remark_inverse_pod}
\end{remark}
To prove optimal error estimates in time, we follow \cite{KV01} and include the finite difference
quotients $\overline{\partial} u(t_n) = \frac{u(t_n) - u(t_{n-1})}{\Delta t}$, where $n = 1, \ldots, N$, in the set
of snapshots $V := \mbox{span}\left\{ u(t_0), \ldots, u(t_N), \overline{\partial} u(t_1), \ldots, \overline{\partial} u(t_N) \right\}$.
As pointed out in \cite{KV01}, the error formula \eqref{pod_error_formula} becomes:
\begin{eqnarray}
\begin{split}
& \hspace*{0.55cm} \frac{1}{2 N + 1} \, \sum_{i=0}^N
\left\|
u(\cdot,t_i) - \sum_{j=1}^r \bigl( u(\cdot,t_i) \, , \, \varphi_j(\cdot) \bigr)_X \, \varphi_j(\cdot)
\right\|_X^2 \\
& + \ \frac{1}{2 N + 1} \, \sum_{i=1}^N
\left\|
\overline{\partial} u(\cdot,t_i) - \sum_{j=1}^r \bigl( \overline{\partial} u(\cdot,t_i) \, , \, \varphi_j(\cdot) \bigr)_X \, \varphi_j(\cdot)
\right\|_X^2
= \sum_{j=r+1}^{d} \lambda_j \, .
\end{split}
\label{pod_error_formula_new}
\end{eqnarray}
\medskip
After these preliminaries, we are ready to derive the error estimates.
The weak form of \eqref{cdcd} reads:
\begin{eqnarray}
(u_t, v)
+ a(u,v)
= (f ,v)
\qquad \forall \, v \in X \, .
\label{cdcd_weak}
\end{eqnarray}
The VMS-POD model for \eqref{cdcd_weak} with a backward Euler time
discretization reads:
Find $u_r^n \in X^r$ such that:
\begin{eqnarray}
\frac{1}{\Delta t} \, (u_r^{n+1} -u_r^n, v_r)
+ A(u_r^{n+1},v_r)
= (f^{n+1} ,v_r)
\qquad \forall \, v_r \in X^r \, .
\label{vms_pod_fe}
\end{eqnarray}
The following stability result for $u_r^n$ holds:
\begin{theorem}
The solution $u_r^n$ of \eqref{vms_pod_fe} satisfies the following bound
\begin{eqnarray}
\| u_r^n \|
\leq \| u_r^0 \|
+ \Delta t \, \sum_{n=0}^{N-1} \| f^{n+1} \| .
\label{theorem_stability_1}
\end{eqnarray}
\label{theorem_stability}
\end{theorem}
\begin{proof}
Choosing $v_r := u_r^{n+1}$ in \eqref{vms_pod_fe}, we get:
\begin{eqnarray}
\frac{1}{\Delta t} \, (u_r^{n+1} -u_r^n, u_r^{n+1})
+ A(u_r^{n+1}, u_r^{n+1})
= (f^{n+1} , u_r^{n+1}) \, .
\label{theorem_stability_2}
\end{eqnarray}
By applying the Cauchy-Schwarz inequality on both sides of \eqref{theorem_stability_2}
and simplifying by $\| u_r^{n+1} \|$, we get:
\begin{eqnarray}
\| u_r^{n+1} \|
- \| u_r^n \|
\leq \Delta t \, \| f^{n+1} \| .
\label{theorem_stability_3}
\end{eqnarray}
Summing from $0$ to $N-1$ the inequality in \eqref{theorem_stability_3}, we get
\eqref{theorem_stability_1}.
\end{proof}
In order to prove an estimate for $\| u^n - u_r^n \|$,
we will first consider the {\it Ritz projection} $w_r \in X^r$ of $u \in X$:
\begin{eqnarray}
A(u - w_r , v_r)
= 0
\qquad \forall \, v_r \in X^r .
\label{ritz}
\end{eqnarray}
The existence and uniqueness of $w_r$ follow from Lax-Milgram lemma.
We now prove an estimate for $u^n - w_r^n$, the error in the Ritz projection.
\begin{lemma}
The Ritz projection $w_r^n$ of $u^n$ satisfies the following error estimate:
\begin{eqnarray}
\hspace*{1.0cm} \frac{1}{N} \, \sum_{n=1}^{N} \| u^n - w_r^n \|
&\leq& C \,
\Biggl\{
\left(
1
+ {\sqrt{\|M_r^{-1}\|_2}}
+ \alpha^{-1}
\right)^{1/2} \label{lemma_ritz_0} \\
&& \hspace*{1.0cm} \left(
h^{m+1} \, \frac{1}{N} \, \sum_{n=1}^{N} \| u^n \|_{m+1}
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j}
\right) \nonumber \\
&& \hspace*{0.1cm} + \sqrt{\varepsilon + \alpha} \,
\left(
h^{m} \, \frac{1}{N} \, \sum_{n=1}^{N} \| u^n \|_{m+1}
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j}
\right)
\Biggr\} . \nonumber
\end{eqnarray}
\label{lemma_ritz}
\end{lemma}
\begin{proof}
Setting $u := u^n$ in \eqref{ritz}, we get:
\begin{eqnarray}
A(u^n - w_r^n , v_r)
= 0
\qquad \forall \, v_r \in X^r .
\label{lemma_ritz_1}
\end{eqnarray}
We decompose the error $u^n - w_r^n$ as
$
u^n - w_r^n
:= \left( u^n - I_{h, r}(u^n) \right)
- \left( w_r^n - I_{h, r}(u^n) \right)
= \eta^n - \phi_r^n ,
$
where $I_{h, r}(u^n)$ is the interpolant of $u^n$ in the space $X^r$.
By the triangle inequality, we have:
\begin{eqnarray}
\frac{1}{N} \, \sum_{n=1}^{N} \| u^n - w_r^n \|
\leq \frac{1}{N} \, \sum_{n=1}^{N} \| \eta^n \|
+ \frac{1}{N} \, \sum_{n=1}^{N} \| \phi_r^n \|.
\label{lemma_ritz_3}
\end{eqnarray}
We start by estimating $\| \eta^n \|$.
We note that $I_{h, r}(u^n)$ consists of two parts:
We first consider $u_h^n$, the FE solution of \eqref{cdcd}, which yielded the
ensemble of snapshots $V$ defined in \eqref{snapshots}.
Then, we interpolate $u_h^n$ in $X^r$, which yields $I_{h, r}(u^n)$.
Note that this is different from \cite{heitmann2007subgridscale}, where
only the first part was present (see $(8)$ in \cite{heitmann2007subgridscale}).
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} \frac{1}{N} \, \sum_{n=1}^{N} \| \eta^n \|
&= \frac{1}{N} \, \sum_{n=1}^{N} \| u^n - I_{h, r}(u^n) \| \\
&\leq \frac{1}{N} \, \sum_{n=1}^{N} \| u^n - u_h^n \|
+ \frac{1}{N} \, \sum_{n=1}^{N} \| u_h^n - I_{h, r}(u^n) \| \, .
\label{lemma_ritz_4}
\end{split}
\end{eqnarray}
Using Assumption \ref{assumption_approximability}, it is easily
shown \cite{quarteroni1994numerical} that:
\begin{eqnarray}
\frac{1}{N} \, \sum_{n=1}^{N} \| u^n - u_h^n \|
\leq C \, h^{m+1} \, \frac{1}{N} \, \sum_{n=1}^{N} \| u^n \|_{m+1 }.
\label{lemma_ritz_4b}
\end{eqnarray}
Picking $\displaystyle I_{h, r}(u^n) := \sum_{j=1}^{r} (u^n , \varphi_j)_X \, \varphi_j$
in the last term on the RHS of \eqref{lemma_ritz_4} and then using
\eqref{pod_error_formula_new}, we get:
\begin{eqnarray}
\frac{1}{N} \, \sum_{n=1}^{N} \| u_h^n - I_{h, r}(u^n) \|
\leq \sqrt{\sum_{j = r+1}^{d} \lambda_j} .
\label{lemma_ritz_4c}
\end{eqnarray}
Note that we consider that the time instances $t_n = n \, \Delta t$ in the time discretization
\eqref{vms_pod_fe} are the same as the time instances at which the snapshots were taken.
If this is not the case, one should use a Taylor series approach (see $(4.8)$ in \cite{luo2009finite}).
Plugging \eqref{lemma_ritz_4b} and \eqref{lemma_ritz_4c} in \eqref{lemma_ritz_4},
we get:
\begin{eqnarray}
\frac{1}{N} \, \sum_{n=1}^{N} \| \eta^n \|
\leq C \, h^{m+1} \, \frac{1}{N} \, \sum_{n=1}^{N} \| u^n \|_{m+1 }
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j} .
\label{lemma_ritz_4cd}
\end{eqnarray}
Similarly, using that $X = H_0^1$ in \eqref{pod_error_formula_new}, we get:
\begin{eqnarray}
\frac{1}{N} \, \sum_{n=1}^{N} \| \nabla \eta^n \|
\leq C \, h^{m} \, \frac{1}{N} \, \sum_{n=1}^{N} \| u^n \|_{m+1 }
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j} .
\label{lemma_ritz_4d}
\end{eqnarray}
Equation \eqref{lemma_ritz_1} implies:
\begin{eqnarray}
A(u^n - w_r^n, v_r)
= A(\eta^n - \phi_r^n, v_r)
= 0 .
\label{lemma_ritz_5}
\end{eqnarray}
Choosing $v_r = \phi_r^n$ in \eqref{lemma_ritz_5} implies:
\begin{eqnarray}
A(\phi_r^n, \phi_r^n)
= A(\eta^n, \phi_r^n) .
\label{lemma_ritz_6}
\end{eqnarray}
We decompose the bilinear form $A$ into its symmetric and skew-symmetric parts:
$A := A_{s} + A_{ss}$,
where
$A_{s}(u, v)
:= \alpha \, (P_R^{'} \nabla u , P_R^{'} \nabla v)
+ \varepsilon \, (\nabla u, \nabla v)
+ \left( \left( g - \frac{1}{2} \nabla \cdot {\bf b} \right) \, u , v \right)$, and
$A_{ss}(u, v)
:= \left( {\bf b} \cdot \nabla u + \newline \frac{1}{2} \left( \nabla \cdot {\bf b} \right) \, u , v \right) \, $.
Equation \eqref{lemma_ritz_6} implies:
\begin{eqnarray}
A_{s}(\phi_r^n, \phi_r^n)
+ \cancelto{0}{A_{ss}(\phi_r^n, \phi_r^n)}
= A_{s}(\eta^n, \phi_r^n)
+ A_{ss}(\eta^n, \phi_r^n) \, .
\label{lemma_ritz_8}
\end{eqnarray}
Assumption \ref{assumption_coercivity} implies that
$A_{s}(\phi_r^n, \phi_r^n) \geq C \, \| \phi_r^n \|_{1,\varepsilon,\alpha}^2$.
Thus, using the Cauchy-Schwarz and Young's inequalities,
\eqref{lemma_ritz_8} becomes:
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} & C \, \| \phi_r^n \|_{1,\varepsilon,\alpha}^2
\leq A_{s}(\phi_r^n, \phi_r^n)^{1/2} \, A_{s}(\eta^n , \eta^n)^{1/2}
+ A_{ss}(\eta^n, \phi_r^n) \\
& \leq \frac{1}{2} \, A_{s}(\phi_r^n, \phi_r^n)
+ \frac{1}{2} \, A_{s}(\eta^n , \eta^n)
+ ({\bf b} \eta^n , \nabla \phi_r^n)
+ \frac{1}{2} \, \left( \left( \nabla \cdot {\bf b} \right) \eta^n , \phi_r^n \right) \, .
\label{lemma_ritz_9}
\end{split}
\end{eqnarray}
Rearranging and using Assumption \ref{assumption_coercivity}, \eqref{lemma_ritz_9} becomes:
\begin{eqnarray}
\hspace*{1.0cm} C \, \| \phi_r^n \|_{1, \varepsilon, \alpha}^2
\leq C \,
\biggl(
|A_{s}(\eta^n, \eta^n)|
+ |({\bf b} \eta^n , \nabla \phi_r^n)|
+ |\left( \left( \nabla \cdot {\bf b} \right) \eta^n , \phi_r^n \right)|
\biggr) \, .
\label{lemma_ritz_10}
\end{eqnarray}
We now estimate each term on the RHS of \eqref{lemma_ritz_10}.
\begin{eqnarray}
\begin{split}
|A_{s}(\eta^n, \eta^n)|
& = \varepsilon \, \| \nabla \eta^n \|^2
+ \left( \left( g - \frac{1}{2} \nabla \cdot {\bf b} \right) \, \eta^n , \eta^n \right) \\
& + \alpha \, \| P_R^{'} \nabla \eta^n\|^2
\leq C \, \| \eta^n \|_{1, \varepsilon, \alpha}^2 \, .
\label{lemma_ritz_11}
\end{split}
\end{eqnarray}
To estimate the second term on the RHS of \eqref{lemma_ritz_10}, we first
note that $\| P_R \| \leq 1$ (since $P_R$ is $L^2$-projection) and use the inverse
estimate \eqref{lemma_inverse_pod_2} in Lemma \ref{lemma_inverse_pod} to
obtain:
\begin{eqnarray}
\| P_R (\nabla \phi_r^n) \|
\leq \| \nabla \phi_r^n \|
\leq \| \phi_r^n \|_{H^1}
\leq \sqrt{\| M_r^{-1} \|_2} \, \| \phi_r^n \| \, .
\label{lemma_ritz_11.5}
\end{eqnarray}
Using that $(P_R u , P_R^{'} v) = 0 \ \forall \, u, v$, the Cauchy-Schwarz
and Young's inequalities, and the inverse estimate \eqref{remark_inverse_pod_1},
we then get:
\begin{eqnarray}
\hspace*{1.0cm} |({\bf b} \eta^n , \nabla \phi_r^n)|
&\leq& |( P_R ({\bf b} \eta^n) , P_R (\nabla \phi_r^n) )|
+ |( P_R^{'} ({\bf b} \eta^n) , P_R^{'} (\nabla \phi_r^n) )| \label{lemma_ritz_12} \\
&\leq& \| P_R ({\bf b} \eta^n) \| \, \| P_R (\nabla \phi_r^n) \|
+ \| P_R^{'} ({\bf b} \eta^n) \| \, \| P_R^{'} (\nabla \phi_r^n) \| \nonumber \\
&\leq& C \, \sqrt{\| M_r^{-1} \|_2} \, \| P_R ({\bf b} \eta^n) \| \, \| \phi_r^n \|
+ \| P_R^{'} ({\bf b} \eta^n) \| \, \| P_R^{'} (\nabla \phi_r^n) \| \nonumber \\
&\leq&
\left(
\frac{1}{\beta} \, C \, \| M_r^{-1} \|_2 \, \| P_R ({\bf b} \eta^n) \|^2
+ \frac{\beta}{4} \, \| \phi_r^n \|^2
\right) \nonumber \\
&+&
\left(
\frac{1}{2 \, \alpha} \, \| P_R^{'} ({\bf b} \eta^n) \|^2
+ \frac{\alpha}{2} \, \| P_R^{'} (\nabla \phi_r^n) \|^2
\right). \nonumber
\end{eqnarray}
We note that this is exactly why we need the inverse estimate in
in Lemma \ref{lemma_inverse_pod}: to absorb $\| \phi_r^n \|^2$ in the LHS
of \eqref{lemma_ritz_10}.
If we had used $\| \nabla \phi_r^n \|^2$ instead, then we would have had to absorb it
in $\varepsilon \, \| \nabla \phi_r^n \|^2$ on the LHS, and so the RHS would have
depended on $\varepsilon$.
Finally, by using the Cauchy-Schwarz and Young's inequalities, the third
term on the RHS of \eqref{lemma_ritz_10} can be estimated as follows:
\begin{eqnarray}
|\left( \left( \nabla \cdot {\bf b} \right) \eta^n , \phi_r^n \right)|
\leq C \, \| \eta^n \| \, \| \phi_r^n \|
\leq C \,
\left(
\frac{1}{\beta} \, \| \eta^n \|^2
+\frac{\beta}{4} \, \| \phi_r^n \|^2
\right) \, .
\label{lemma_ritz_13}
\end{eqnarray}
Collecting estimates
\eqref{lemma_ritz_10},
\eqref{lemma_ritz_11},
\eqref{lemma_ritz_12} and
\eqref{lemma_ritz_13},
we get:
\begin{eqnarray}
\begin{split}
\| \phi_r^n \|_{1, \varepsilon, \alpha}^2
\leq C \biggl(
\| \eta^n \|_{1, \varepsilon, \alpha}^2
& + \frac{1}{\beta} \, \| M_r^{-1} \|_2 \, \| P_R ({\bf b} \eta^n) \|^2 \\
& + \frac{1}{2 \, \alpha} \, \| P_R^{'} ({\bf b} \eta^n) \|^2
+ \frac{1}{\beta} \, \| \eta^n \|^2
\biggr) \, .
\label{lemma_ritz_14}
\end{split}
\end{eqnarray}
The last term on the RHS of \eqref{lemma_ritz_14}, can be absorbed in
$C \, \| \eta^n \|_{1, \varepsilon, \alpha}^2$.
Since $\| P_R \| \leq 1$ ($P_R$ is $L^2$-projection) and $\| {\bf b} \| \leq \gamma$
(by Assumption \ref{assumption_coercivity}), we get:
\begin{eqnarray}
\frac{1}{\beta} \, \| M_r^{-1} \|_2 \, \| P_R ({\bf b} \eta^n) \|^2
\leq C \, \| M_r^{-1} \|_2 \, \| \eta^n \|^2 \, .
\label{lemma_ritz_14_a}
\end{eqnarray}
Since $\| P_R \| \leq 1$ ($P_R$ is $L^2$-projection) and $\| {\bf b} \| \leq \gamma$
(by Assumption \ref{assumption_coercivity}), we get:
\begin{eqnarray}
\frac{1}{2 \, \alpha} \, \| P_R^{'} ({\bf b} \eta^n) \|^2
\leq \frac{C}{\alpha} \, \| \eta^n \|^2 \, .
\label{lemma_ritz_14_b}
\end{eqnarray}
Thus, using \eqref{lemma_ritz_14_a} and \eqref{lemma_ritz_14_b} in
\eqref{lemma_ritz_14}, we get:
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} \| \phi_r^n \|_{1, \varepsilon, \alpha}^2
\leq C \biggl(
\| \eta^n \|^2
+ \varepsilon \, \| \nabla \eta^n \|^2
& + \alpha \, \| P_R^{'} (\nabla \eta^n) \|^2
+ C \, \| M_r^{-1} \|_2 \, \| \eta^n \|^2 \\
& + \frac{1}{2 \, \alpha} \, \| P_R^{'} ({\bf b} \eta^n) \|^2
+ \frac{C}{\alpha} \, \| \eta^n \|^2
\biggr) \, .
\label{lemma_ritz_14_c}
\end{split}
\end{eqnarray}
Since $P_R$ is $L^2$-projection, $\| P_R^{'} \| \leq 1$, and thus the second
term on the RHS of \eqref{lemma_ritz_14_c} can be bounded as follows:
$\alpha \, \| P_R^{'} (\nabla \eta^n) \|^2 \leq \alpha \, \| \nabla \eta^n \|^2$.
Summing in \eqref{lemma_ritz_14_c}, we get:
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} \frac{1}{N} \, \sum_{n=1}^{N} \| \phi_r^n \|_{1, \varepsilon, \alpha}^2
& \leq C \,
\left(
1
+ \|M_r^{-1}\|_2
+ \alpha^{-1}
\right) \,
\frac{1}{N} \, \sum_{n=1}^{N} \| \eta^n \|^2 \\
& + (\varepsilon + \alpha) \,
\frac{1}{N} \, \sum_{n=1}^{N} \| \nabla \eta^n \|^2 \, .
\label{lemma_ritz_14_d}
\end{split}
\end{eqnarray}
Using \eqref{lemma_ritz_4cd} and \eqref{lemma_ritz_4d} in
\eqref{lemma_ritz_14_d}, we get:
\begin{eqnarray}
\hspace*{1.2cm} \frac{1}{N} \, \sum_{n=1}^{N} \| \phi_r^n \|
&\leq& C \,
\Biggl\{
\left(
1
+ {\|M_r^{-1}\|_2}
+ \alpha^{-1}
\right)^{1/2} \label{lemma_ritz_14_e} \\
&& \hspace*{1.0cm}
\left(
h^{m+1} \, \frac{1}{N} \, \sum_{n=1}^{N} \| u^n \|_{m+1}
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j}
\right) \nonumber \\
&& \hspace*{1.0cm}
+ \sqrt{\varepsilon + \alpha} \,
\left(
h^{m} \, \frac{1}{N} \, \sum_{n=1}^{N} \| u^n \|_{m+1}
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j}
\right)
\Biggr\} . \nonumber
\end{eqnarray}
Using \eqref{lemma_ritz_3}, \eqref{lemma_ritz_4cd}, and \eqref{lemma_ritz_14_e}, we get
\eqref{lemma_ritz_0}.
\end{proof}
\begin{corollary}
The Ritz projection $w_r^n$ of $u^n$ satisfies the following error estimate:
\begin{eqnarray}
\hspace*{1.0cm} \| (u^n - w_r^n)_t \|
&\leq& C \,
\Biggl\{
\left(
1
+ {\|M_r^{-1}\|_2}
+ \alpha^{-1}
\right)^{1/2} \\
&& \hspace*{1.0cm}
\left(
h^{m+1} \| u_t \|_{L^2(H^{m+1})}
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j}
\right) \nonumber \\
&& \hspace*{1.0cm}
+ \sqrt{\varepsilon + \alpha} \,
\left(
h^{m} \, \| u_t \|_{L^2(H^{m+1})}
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j}
\right)
\Biggr\} \nonumber .
\label{corollary_ritz_t_0}
\end{eqnarray}
\label{corollary_ritz_t}
\end{corollary}
\begin{proof}
The proof follows along the same lines as the proof of Lemma \ref{lemma_ritz},
with the error $u^n - w_r^n$ replaced by $(u^n - w_r^n)_t = (\eta_n - \phi_r^n)_t$.
Note that it is exactly at this point that we use the fact that the finite difference
quotients $\overline{\partial} u(t_n)$ are included in the set of snapshots (see also Remark 1
in \cite{KV01}).
Indeed, as in the proof of Lemma \ref{lemma_ritz}, the error $(u^n - w_r^n)_t$ is split
into two parts:
$\displaystyle
(u^n - w_r^n)_t
:= \left( u^n_t - I_{h, r}(u^n_t) \right)
- \left( (w_t)_r^n - I_{h, r}(u^n_t) \right)
= \eta_t^n - (\phi_t)_r^n .
$
As in \eqref{lemma_ritz_4}--\eqref{lemma_ritz_4cd}, $\eta_t^n$ can be estimated as
follows.
\begin{eqnarray}
\begin{split}
\| \eta_t^n \|
& \leq \| u_t^n - u_{h , t}^n \|
+ \| u_{h , t}^n - I_{h, r}(u^n_t) \| \\
& \leq C \, \left(
h^{m+1} \| u_t \|_{m+1}
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j}
\right) \, ,
\label{corollary_ritz_t_1}
\end{split}
\end{eqnarray}
where in the last inequality in \eqref{corollary_ritz_t_1} we used \eqref{pod_error_formula_new}.
\end{proof}
We are now ready to prove the main result of this section.
\begin{theorem}
Assume that
\begin{eqnarray}
L^R = \nabla X^R = \text{span} \{ \nabla \varphi_1, \ldots, \nabla \varphi_R \} \, .
\label{theorem_error_0}
\end{eqnarray}
Then the following error estimate holds:
\begin{eqnarray}
&& \hspace*{0.5cm} \frac{1}{N+1} \, \sum_{n=0}^{N} \| u^n - u_r^n \|
\leq C \,
\Biggl\{
\left(
1
+ \|M_r^{-1}\|_2
+ \alpha^{-1}
\right)^{1/2} \label{theorem_error_1} \\
&& \hspace*{1.0cm} \left(
h^{m+1} \,
\frac{1}{N} \, \sum_{n=1}^{N} \left( \| u^n \|_{m+1} + \| u_t \|_{L^2(H^{m+1})} \right)
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j}
\right) \nonumber \\
&& \hspace*{1.0cm}
+ \sqrt{\varepsilon + \alpha} \,
\left(
h^{m} \, \frac{1}{N} \, \sum_{n=1}^{N} \left( \| u^n \|_{m+1} + \| u_t \|_{L^2(H^{m+1})} \right)
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j}
\right)
\nonumber \\
&& \hspace*{1.0cm}
+ \| u^0 - u_r^0 \|
+ \Delta t \, \| u_{tt} \|_{L^2(L^2)} \nonumber \\
&& \hspace*{1.0cm}
+ \sqrt{\alpha} \,
\left(
h^{m} \, \frac{1}{N} \, \sum_{n=1}^{N} \| u^n \|_{m+1}
+ \sqrt{\sum_{j = R+1}^{d} \lambda_j}
\right) \,
\Biggr\} . \nonumber
\end{eqnarray}
\label{theorem_error}
\end{theorem}
\begin{proof}
We evaluate \eqref{cdcd_weak} at $t_{n+1}$, we let $v = v_r$, and then we add and subtract
$\displaystyle \left( \frac{u^{n+1} - u^n}{\Delta t}, v_r\right)$:
\begin{eqnarray}
\begin{split}
\left(
u^{n+1}_t
- \frac{u^{n+1} - u^n}{\Delta t} , v_r
\right)
& + \left( \frac{u^{n+1} - u^n}{\Delta t}, v_r \right) \\
& + a(u^{n+1}, v_r)
= (f^{n+1},v_r) .
\label{theorem_error_1.5}
\end{split}
\end{eqnarray}
Subtracting \eqref{vms_pod_fe} from \eqref{theorem_error_1.5}, we obtain the error equation:
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} & \left(
u^{n+1}_t
- \frac{u^{n+1} - u^n}{\Delta t} , v_r
\right)
+ \left(
\frac{u^{n+1} - u_r^{n+1}}{\Delta t} , v_r
\right)
- \left(
\frac{u^n - u_r^n}{\Delta t} , v_r
\right) \\
& \hspace*{3.5cm} + A(u^{n+1} - u_r^{n+1} , v_r)
+ (a-A)(u^{n+1} , v_r)
= 0 .
\label{theorem_error_2}
\end{split}
\end{eqnarray}
We now decompose the error as
$
u^n - u_r^n
= \bigl( u^n - w_r^n \bigr)
- \bigl( u_r^n - w_r^n \bigr)
= \eta^n - \phi_r^n ,
$
which, by the triangle inequality, implies:
\begin{eqnarray}
\| u^n - u_r^n \|
\leq \| \eta^n \| +
\| \phi_r^n \| .
\label{theorem_error_3b}
\end{eqnarray}
We note that $\| \eta^n \|$ has already been bounded in Lemma \ref{lemma_ritz}.
Thus, in order to estimate the error, we only need to estimate $\| \phi_r^n \|$.
The error equation \eqref{theorem_error_2} can be written as:
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} && \left(
u^{n+1}_t
- \frac{u^{n+1} - u^n}{\Delta t} , v_r
\right)
+ \left(
\frac{\eta^{n+1} - \eta^n}{\Delta t} , v_r
\right)
- \left(
\frac{\phi_r^{n+1} - \phi_r^n}{\Delta t} , v_r
\right) \\
&& + A(\eta^{n+1} - \phi_r^{n+1} , v_r)
+ (a-A)(u^{n+1} , v_r)
= 0 .
\label{theorem_error_4}
\end{split}
\end{eqnarray}
We pick $v_r := \phi_r^{n+1}$ in \eqref{theorem_error_4}, we note that,
since $\phi_r^{n+1} \in X^r$, $A(\eta^{n+1} , \phi_r^{n+1}) = 0$, and we get:
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} & A(\phi_r^{n+1} , \phi_r^{n+1})
+ \frac{1}{\Delta t} \, (\phi_r^{n+1} - \phi_r^n , \phi_r^{n+1})
= \frac{1}{\Delta t} \, (\eta^{n+1} - \eta^n , \phi_r^{n+1}) \\
& \hspace*{4.0cm} + ( r^n , \phi_r^{n+1} )
+ (a-A)(u^{n+1} , \phi_r^{n+1}) ,
\label{theorem_error_5}
\end{split}
\end{eqnarray}
where $\displaystyle r^n = u^{n+1}_t - \frac{u^{n+1} - u^n}{\Delta t}$.
We now start estimating all the terms in \eqref{theorem_error_5}.
The terms on the LHS of \eqref{theorem_error_5} are estimated
as follows:
\begin{eqnarray}
&& A(\phi_r^{n+1} , \phi_r^{n+1})
\geq \beta \, \| \phi_r^{n+1} \|^2
+ \varepsilon \, \| \nabla \phi_r^{n+1} \|^2
+ \alpha \, \| P_R^{'} \nabla \phi_r^{n+1} \|^2 .
\label{theorem_error_6} \\
&& \frac{1}{\Delta t} \, (\phi_r^{n+1} - \phi_r^n , \phi_r^{n+1})
\geq \frac{1}{\Delta t} \, \left(\| \phi_r^{n+1} \|^2 - \| \phi_r^n \| \, \| \phi_r^{n+1} \| \right) .
\end{eqnarray}
Now we estimate the RHS of \eqref{theorem_error_5}
by using the Cauchy-Schwarz and Young's inequalities:
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} & \left(
\frac{1}{\Delta t} \,
(\eta^{n+1} - \eta^n)
+ r^n , \phi_r^{n+1}
\right)
\leq \left\| \frac{1}{\Delta t} \, (\eta^{n+1} - \eta^n) + r^n\right\| \, \| \phi_r^{n+1} \| \\
& \hspace*{3.0cm} \leq \frac{1}{2 \, \beta} \, \left\| \frac{1}{\Delta t} \, (\eta^{n+1} - \eta^n) + r^n\right\|^2
+ \frac{\beta}{2} \, \| \phi_r^{n+1} \|^2 . \label{theorem_error_7} \\
\end{split}
\end{eqnarray}
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} & (a-A)(u^{n+1} , \phi_r^{n+1})
= - \alpha \, (P_R^{'} \nabla u^{n+1} , P_R^{'} \nabla \phi_r^{n+1}) \\
& \leq \alpha \, \| P_R^{'} \nabla u^{n+1} \| \, \| P_R^{'} \nabla \phi_r^{n+1} \|
\leq \frac{\alpha}{2} \, \| P_R^{'} \nabla u^{n+1} \|^2
+ \frac{\alpha}{2} \, \| P_R^{'} \nabla \phi_r^{n+1} \|^2 .
\label{theorem_error_8}
\end{split}
\end{eqnarray}
Using \eqref{theorem_error_6}-\eqref{theorem_error_8} and absorbing
RHS terms into LHS terms, \eqref{theorem_error_5} now reads:
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} & \frac{1}{\Delta t} \, (\| \phi_r^{n+1} \|^2 - \| \phi_r^n \| \, \| \phi_r^{n+1} \|)
+ \frac{\beta}{2} \, \| \phi_r^{n+1} \|^2
+ \varepsilon \, \| \nabla \phi_r^{n+1} \|^2 \\
& + \frac{\alpha}{2} \, \| P_R^{'} \nabla \phi_r^{n+1} \|^2
\leq \frac{1}{2 \, \beta} \,
\left\| \frac{1}{\Delta t} \,(\eta^{n+1} - \eta^n) + r^n\right\|^2
+ \frac{\alpha}{2} \, \| P_R^{'} \nabla u^{n+1} \|^2 .
\label{theorem_error_9}
\end{split}
\end{eqnarray}
By using Young's inequality, the first term on the LHS of \eqref{theorem_error_9}
can be estimated as follows:
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} \| \phi_r^{n+1} \|^2 - \| \phi_r^n \| \, \| \phi_r^{n+1} \|
& \geq \| \phi_r^{n+1} \|^2
- \frac{1}{2} \, \| \phi_r^n \|^2
- \frac{1}{2} \, \| \phi_r^{n+1} \|^2 \\
& = \frac{1}{2} \, \| \phi_r^{n+1} \|^2
-\frac{1}{2} \, \| \phi_r^n \|^2 .
\label{theorem_error_10}
\end{split}
\end{eqnarray}
Using \eqref{theorem_error_10} in \eqref{theorem_error_9} and multiplying by
$2 \, \Delta t$, we get:
\begin{eqnarray}
&& \hspace*{1.0cm} \| \phi_r^{n+1} \|^2
-\| \phi_r^n \|^2
+ \Delta t \, \| \phi_r^{n+1} \|_{1, \varepsilon, \alpha}^2
\label{theorem_error_11} \\
&& \hspace*{2.0cm} \leq C \, \left( \Delta t \, \left\| \frac{1}{\Delta t} \, (\eta^{n+1} - \eta^n) + r^n \right\|^2
+ \alpha \, \Delta t \, \| P_R^{'} \nabla u^{n+1} \|^2 \right) \nonumber \\
&& \hspace*{2.0cm} \leq C \,
\left( \Delta t \, \left\| \frac{1}{\Delta t} \, (\eta^{n+1} - \eta^n) \right\|^2 + \Delta t \, \| r^n \|^2
+ \alpha \, \Delta t \, \| P_R^{'} \nabla u^{n+1} \|^2 \right) . \nonumber
\end{eqnarray}
Summing from $n=0$ to $n=N-1$ in \eqref{theorem_error_11}, we get:
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} & \max_{0 \leq n \leq N} \| \phi^n_r \|^2
+ \sum_{n=0}^{N-1} \Delta t \, \| \phi_r^{n+1} \|_{1, \varepsilon, \alpha}^2
\leq C \, \Biggl(
\Delta t \, \sum_{n=0}^{N-1} \left\| \frac{1}{\Delta t} \, (\eta^{n+1} - \eta^n) \right\|^2 \\
& \hspace*{2.0cm} + \| \phi^0_r \|^2
+ \Delta t \, \sum_{n=0}^{N-1} \left\| r^n \right\|^2
+ \, \alpha \, \Delta t \, \sum_{n=0}^{N-1} \| P_R^{'} \nabla u^{n+1} \|^2 \Biggr) .
\label{theorem_error_12}
\end{split}
\end{eqnarray}
Proceeding as in \cite{thomee2006galerkin} (see also \cite{heitmann2007subgridscale}),
we estimate the first term on the RHS of \eqref{theorem_error_12} as follows.
We start by writing:
\begin{eqnarray}
\eta^{n+1} - \eta^n
= \int_{t_{n}}^{t_{n+1}} \eta_t \, dt \, .
\label{theorem_error_12a}
\end{eqnarray}
Taking the $L^2$-norm in \eqref{theorem_error_12a} and applying the Cauchy-Schwarz
inequality, we get:
\begin{eqnarray}
\begin{split}
\hspace*{1.0cm} \| \eta^{n+1} - \eta^n \|
& \leq \int_{t_{n}}^{t_{n+1}} 1 \, \| \eta_t \| \, dt
\leq \left( \int_{t_{n}}^{t_{n+1}} 1^2 \, dt \right)^{1/2} \,
\left( \int_{t_{n}}^{t_{n+1}} \| \eta_t \|^2 \, dt \right)^{1/2} \\
&\leq (\Delta t)^{1/2} \, \left( \int_{t_{n}}^{t_{n+1}} \| \eta_t \|^2 \, dt \right)^{1/2} \, ,
\label{theorem_error_12b}
\end{split}
\end{eqnarray}
which implies
$\displaystyle
\Delta t \, \left\| \frac{1}{\Delta t} \, (\eta^{n+1} - \eta^n) \right\|^2
\leq \left( \int_{t_{n}}^{t_{n+1}} \| \eta_t \|^2 \, dt \right)^{1/2}
$.
Summing from $n=0$ to $n=N-1$, we get
$\displaystyle
\Delta t\, \sum_{n=0}^{N-1} \left\| \frac{1}{\Delta t} \, (\eta^{n+1} - \eta^n) \right\|^2
\leq \| \eta_t \|_{L^2(L^2)},
$
which was bound in Corollary \ref{corollary_ritz_t}.
We thus obtain:
\begin{eqnarray}
&& \Delta t\, \sum_{n=0}^{N-1} \left\| \frac{1}{\Delta t} \, (\eta^{n+1} - \eta^n) \right\|^2
\leq C \,
\Biggl\{
\left(
1
+ {\|M_r^{-1}\|_2}
+ \alpha^{-1}
\right)^{1/2} \label{theorem_error_12c} \\
&& \hspace*{4.0cm} \left(
h^{m+1} \| u_t \|_{L^2(H^{m+1})}
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j}
\right) \nonumber \\
&& \hspace*{3.0cm} + \sqrt{\varepsilon + \alpha} \,
\left(
h^{m} \, \| u_t \|_{L^2(H^{m+1})}
+ \sqrt{\sum_{j = r+1}^{d} \lambda_j}
\right)
\Biggr\} . \nonumber
\end{eqnarray}
To estimate the third term on the RHS of \eqref{theorem_error_12}, we use a Taylor series
expansion of $u^n$ around $u^{n+1}$:
\begin{eqnarray}
u^n
= u^{n+1}
- u^{n+1}_t \, \Delta t
+ \int_{t_{n}}^{t_{n+1}} u_{tt}(s) \, (t_n - s) \, ds \, .
\label{theorem_error_12d}
\end{eqnarray}
Taking the $L^2$-norm in \eqref{theorem_error_12d} and applying the Cauchy-Schwarz
inequality, we get
$\displaystyle
\| r^n \|
\leq \int_{t_{n}}^{t_{n+1}} 1 \, \| u_{tt} \| \, ds
\leq (\Delta t)^{1/2} \, \| u_t \|_{L^2(L^2)} \, .
$
Summing from $n=0$ to $n=N-1$, we get:
\begin{eqnarray}
\Delta t\, \sum_{n=0}^{N-1} \left\| r^n \right\|^2
\leq \Delta t^2 \, \| u_{tt} \|^{2}_{L^2(L^2)} \, .
\label{theorem_error_14}
\end{eqnarray}
To estimate the last term on the RHS of \eqref{theorem_error_12}, we
use the fact that $L^R = \nabla X^R$ (assumption \eqref{theorem_error_0}).
We emphasize that this is the {\it only instance} in the proof where the assumption
$L^R = \nabla X^R$ is used.
Thus, we get:
\begin{eqnarray}
&& \alpha \, \Delta t \, \sum_{n=0}^{N-1} \| P_R^{'} \nabla u^{n+1} \|^2
= \alpha \, \Delta t \, \sum_{n=0}^{N-1} \| \nabla u^{n+1} - P_R \nabla u^{n+1}\|^2 \label{theorem_error_15} \\
&& \hspace*{2.5cm} \stackrel{\eqref{theorem_error_0}}{\leq} C \, \alpha \, \frac{1}{N} \, \sum_{n=0}^{N-1}
\inf_{v_R \in X^R} \| \nabla u^{n+1} - \nabla v_R \|^2
\nonumber \\
&& \hspace*{2.5cm} \stackrel{\eqref{pod_error_formula}, \eqref{assumption_approximability_1}}{\leq}
C \, \alpha \,
\left(
h^{m} \, \frac{1}{N} \, \sum_{n=1}^{N} \| u^n \|_{m+1 }
+ \sqrt{\sum_{j = R+1}^{d} \lambda_j}
\right)^2 \, . \nonumber
\end{eqnarray}
Using \eqref{theorem_error_12c}, \eqref{theorem_error_14}, and \eqref{theorem_error_15} in
\eqref{theorem_error_12},
the obvious inequality
$\displaystyle \max_{0 \leq n \leq N} \| \phi_r^n \|
\geq \frac{1}{N+1} \, \sum_{n=0}^{N} \| \phi_r^n \|$,
inequality \eqref{theorem_error_3b}, and the
estimates in Lemma \ref{lemma_ritz}, we obtain the error estimate \eqref{theorem_error_1}.
\end{proof}
\section{Numerical Results}
\label{s_numerics}
The goal of this section is twofold:
(i) to show that the new VMS-POD model \eqref{vms_pod} is significantly more
stable numerically than the standard POD-G model \eqref{pod_g}; and
(ii) to illustrate numerically the theoretical error estimate \eqref{theorem_error_1}.
We also use Theorem \ref{theorem_error} to provide theoretical guidance in
choosing an optimal value for the artificial viscosity coefficient $\alpha$ and use this algorithm
within our numerical framework.
Finally, we show that the VMS-POD model \eqref{vms_pod} displays a relatively
low sensitivity with respect to changes in the diffusion coefficient $\varepsilon$.
Thus, we provide numerical support for the theoretical estimate \eqref{theorem_error_1},
which is {\em uniform} with respect to $\varepsilon$.
The mathematical model used for all the numerical tests in this section is the
convection-dominated convection-diffusion equation \eqref{cdcd} with the following
parameter choices:
spatial domain $\Omega=[0, 1]\times[0, 1]$,
time interval $[0, T]=[0, 1]$,
diffusion coefficient $\varepsilon = 1\times 10^{-4}$,
convection field ${\bf b} = [\cos\frac{\pi}{3}, \sin\frac{\pi}{3}]^{T}$, and
reaction coefficient $g=1$.
The forcing term $f$ and initial condition $u_0(x)$ are chosen to satisfy the exact
solution $u(x,y,t) = 0.5\sin(\pi x)\sin(\pi y) \left[ \tanh\left(\frac{x+y-t-0.5}{0.04}\right) +1 \right]$, which is
similar to that used in \cite{guermond1999stabilization}.
As in the theoretical developments in Section \ref{s_estimates}, in this section we
employ the finite element method for spatial discretization and the backward Euler
method for temporal discretization of all models investigated.
All computations are carried out on a PC with 3.2 GHz Intel Xeon Quad-core processor.
We start by comparing the VMS-POD model \eqref{vms_pod} to the standard POD-G
model \eqref{pod_g}.
To generate the POD basis, we first run a DNS
with the following parameters: piecewise quadratic finite elements, uniform triangular mesh
with mesh-size $h = 0.01$, and time-step $\Delta t = 10^{-4}$.
A mesh refinement study indicates that DNS mesh resolution is achieved.
The average DNS error is
$\frac{1}{N+1}{\sum\limits_{n=0}^{N}\|u^n-u_h^n\|} = 2.04\times10^{-4}$, where
$N=1000$, and $u^n$ and $u^n_h$ are the exact solution and the finite element
solution at $t=n\Delta t$, respectively.
The CPU time of the DNS is $9.42 \times 10^4\, s$.
Since the forcing term is time-dependent, the global load vectors are stored for later
use in all the ROMs.
The POD modes are generated in $H^1$ by the method of snapshots;
the rank of the data set is $104$.
For both POD-ROMs (POD-G and VMS-POD), we use the same number of POD basis
functions: $r = 40$.
\begin{table}[ht]
\centering
\caption{Average errors for the POD-G model \eqref{pod_g} with different values of $r$.
Note that the POD-G model yields poor results.}
\label{table_PODG}
\begin{tabular}{|c|c|c|c|c|}
\hline
$r$&20&40&60&80\\
\hline
$\frac{1}{N+1}{\sum\limits_{n=0}^{N}\|u^n-u_r^n\|}$&$1.25\times 10^{-1}$&$1.11\times 10^{-1}$&$9.28\times 10^{-2}$&$8.20\times 10^{-2}$\\
\hline
\end{tabular}
\end{table}
We first test the POD-G model \eqref{pod_g}.
The CPU time for the POD-G model is $96.4\, s$, which is three orders of magnitude
lower than that of a brute force DNS.
The numerical solution at $t = 1$ is shown in Figure \ref{Comp_models} for both the
DNS (top) and the POD-G model (middle).
It is clear from this figure that, although the first $40$ POD modes capture $99.99\%$
of the system's kinetic energy, the POD-G model yields poor quality results and
displays strong numerical oscillations.
This is confirmed by the POD-G model's high average error $\frac{1}{N+1}{\sum\limits_{n=0}^{N}\|u^n-u_r^n\|} =1.11\times 10^{-1}$,
where $u_r^n$ is the POD-G model's solution at $t=n\Delta t$.
Indeed, the POD-G model's average error is almost three orders of magnitude higher
than the average error of the DNS.
The average errors for different values of $r$ listed in Table \ref{table_PODG} show that
increasing the number of POD modes ($r$) does not decrease significantly the average
error.
It is thus clear that the straightforward POD-G model, although computationally efficient,
is highly inaccurate.
Next, we investigate the VMS-POD model \eqref{vms_pod}.
We make the following parameter choices: $R=20$ and $\alpha=4.29\times 10^{-2}$.
The motivation for this choice is given later in this section.
The CPU time for the VMS-POD model \eqref{vms_pod} is $106.2\,s$, which is close
to the CPU time of the POD-G model \eqref{pod_g}.
The numerical solution at $t = 1$ for the VMS-POD model is shown in Figure
\ref{Comp_models} (bottom).
It is clear from this figure that the VMS-POD model is much more accurate than the
POD-G model.
Indeed, the VMS-POD model results are much closer to the DNS results than the
POD-G model results, since the numerical oscillations displayed by the latter are
dramatically decreased.
This is confirmed by the VMS-POD model's average error $\frac{1}{N+1}{\sum\limits_{n=0}^{N}\|u^n-u_r^n\|} = 4.48\times 10^{-3}$,
where $u_r^n$ is the VMS-POD solution at $t=n\Delta t$; this error is more than $20$
times lower than the error of the POD-G model.
\begin{figure}[h]
\centering
\caption{Numerical solution at $t = 1$:
DNS (top), POD-G model \eqref{pod_g} (middle), and VMS-POD model \eqref{vms_pod}
(bottom).
Note that the VMS-POD model is much more accurate than the POD-G model, decreasing
the unphysical oscillations of the latter.
The CPU times for both the VMS-POD and POD-G models are three orders of magnitude
lower than the CPU time for the DNS.\vspace*{0.1cm}
}\label{Comp_models}
\includegraphics[width=0.6\textwidth]{Opt12-g1-DNS.pdf}
\includegraphics[width=0.6\textwidth]{Opt12-g1-PODG-r40.pdf}
\includegraphics[width=0.6\textwidth]{Opt12-g1-PODVMS-r40R20.pdf}
\end{figure}
In conclusion, the VMS-POD model \eqref{vms_pod} dramatically decreases the error
of the POD-G model \eqref{pod_g} by adding numerical stabilization, while keeping the
same level of computational efficiency.
\begin{table}[ht]
\centering
\caption{VMS-POD model's average error $e = \frac{1}{N+1}{\sum\limits_{n=0}^{N}\|u^n-u_r^n\|}$ and its $e_3$ component
for different values of $R$.
}\label{Table_g1_R}
\begin{tabular}{|c|c|c|}
\hline
{$R$}&{$e_3$}&{$\frac{1}{N+1}{\sum\limits_{n=0}^{N}\|u^n-u_r^n\|}$} \\
\hline
1& 1.29$\times 10^{-1}$& 2.55$\times 10^{-2}$\\
4& 9.34$\times 10^{-2}$& 1.78$\times 10^{-2}$\\
7& 6.69$\times 10^{-2}$& 1.37$\times 10^{-2}$\\
10& 4.68$\times 10^{-2}$& 9.80$\times 10^{-3}$\\
13& 3.20$\times 10^{-2}$& 6.99$\times 10^{-3}$\\
\hline
\end{tabular}
\end{table}
\smallskip
We now turn our attention to the second major goal of this section - the numerical
illustration of the theoretical error estimate \eqref{theorem_error_1}.
Specifically, we investigate whether the asymptotic behavior of the RHS of estimate
\eqref{theorem_error_1} with respect to $R$ is reflected in the numerical results.
We focus on the asymptotic behavior with respect to $R$ since this is the main
parameter introduced by the VMS formulation; the asymptotic behavior with respect
to $r$ was investigated in \cite{borggaard2011artificial}, whereas the asymptotic behavior
with respect to $h$ and $\Delta t$ is standard \cite{BS94,thomee2006galerkin}.
To investigate the asymptotic behavior with respect to $R$, we have to ensure that
$\sqrt{\alpha}\sqrt{\sum\limits_{j=R+1}^d\lambda_j}$ (the only term that depends on $R$) dominates all the other
terms on the RHS of \eqref{theorem_error_1}.
To this end, we start collecting all the terms that depend on the exact solution $u$
and we include them in the generic constant $C$.
Next, we assume that the POD interpolation error in the initial condition $\| u^0 - u_r^0\|$
is negligible.
We also assume that the time-step is small enough to neglect
$\Delta t \, \| u_{t t} \|_{L^2(L^2)}$.
With these assumptions, the error estimate \eqref{theorem_error_1} can now be written
as $e \leq C \left( e_1 + e_2 + e_3 \right)$, where $e$ is the VMS-POD model's average
error, $C$ a generic constant independent of $r, R, h, \Delta t$ and $\alpha$,
$e_1 = \|M_r^{-1}\|_2^{\frac{1}{2}}h^{m+1}$, $e_2=\|M_r^{-1}\|_2^{\frac{1}{2}}\sqrt{\sum\limits_{j=r+1}^d\lambda_j}$, and
$e_3=\sqrt{\alpha}\sqrt{\sum\limits_{j=R+1}^d\lambda_j}$.
To ensure that $e_3$ dominates the other terms, we choose $r = 100$ and consider
relatively low values for $R$.
This choice for $r$, which is not optimal for practical computations, ensures, however,
that $e_3$ dominates $e_2$.
We also note that, when $h$ is small, $e_3$ dominates $e_1$ too.
Thus, to investigate the asymptotic behavior with respect to $R$ of the RHS of
\eqref{theorem_error_1}, we fix $\alpha =5\times 10^{-3}$, vary $R$ from $1$ to $14$,
and monitor the changes in $e_3$.
We restrict $R$ to this parameter range to ensure that $\sqrt{\sum\limits_{j=R+1}^d\lambda_j}$ (and thus $e_3$)
dominates $e_2$ and $e_1$.
Table \ref{Table_g1_R} lists the VMS-POD model's average error $e = \frac{1}{N+1}{\sum\limits_{n=0}^{N}\|u^n-u_r^n\|}$ and its
$e_3$ component for different values of $R$.
We emphasize that, in this case, $e_3$ dominates the other two error components
$e_1 = 3.81\times 10^{-3}$ and $e_2 = 2.87\times 10^{-3}$.
To see whether the theoretical linear dependency predicted by the theoretical error
estimate \eqref{theorem_error_1} is recovered in the numerical results in Table
\ref{Table_g1_R}, we utilize a linear regression analysis in Figure \ref{Figure_g1_R}.
This plot shows that the rate of convergence of $e$ with respect to $e_3$ is $0.9$,
which is close to the theoretical value of $1$ predicted by \eqref{theorem_error_1}.
We believe that this slight discrepancy is due to the fact that the mesh-size $h = 0.01$
that we have employed in this numerical investigation is not small enough for our
asymptotic study.
Summarizing the results above, we conclude that the theoretical error estimate in
\eqref{theorem_error_1} is recovered asymptotically (with respect to $R$) in our
numerical experiments.
\begin{figure}[h]
\centering
\caption{Linear regression of VMS-POD model's average error with respect to $e_3$.
The convergence rate is $0.9$, which is close to the theoretical value of $1$ predicted
by \eqref{theorem_error_1}.}
\label{Figure_g1_R}
\includegraphics[width=0.7\textwidth]{Opt12-g1-r100-regression-R1-14.pdf}
\end{figure}
\smallskip
Next, we use Theorem \ref{theorem_error} to provide theoretical guidance in choosing
an optimal value for the artificial viscosity coefficient $\alpha$.
The main challenge is that the theoretical error estimate \eqref{theorem_error_1} is
{\em asymptotic} with respect to $h, \Delta t$ and $r$, while in practical computations
we are using small, yet non-negligible values for these parameters.
Furthermore, the generic constant $C$ is problem-dependent and can play a significant
role in practical computations.
Notwithstanding these hurdles, we choose a value for $\alpha$ that minimizes the RHS
of \eqref{theorem_error_1}:
$\widetilde{\alpha}=\frac{h^{m+1}+\sqrt{\sum\limits_{j=r+1}^d\lambda_j}}{2h^m+\sqrt{\sum\limits_{j=r+1}^d\lambda_j}+\sqrt{\sum\limits_{j=R+1}^d\lambda_j}}$.
In the derivation of this formula, we made the same assumptions as those made in the
numerical investigation of the asymptotic behavior of the VMS-POD model's error and
we again considered that \eqref{theorem_error_1} can be written as
$e \leq C \left( e_1 + e_2 + e_3 \right)$.
We note that, if $\sqrt{\sum\limits_{j=r+1}^d\lambda_j}<<\sqrt{\sum\limits_{j=R+1}^d\lambda_j}$ and $h^{m}<<\sqrt{\sum\limits_{j=R+1}^d\lambda_j}$, then $\widetilde{\alpha}$
becomes too small in practical computations and the VMS-POD model becomes
similar to the inaccurate POD-G model.
To circumvent this, we use in our numerical tests a ``clipping" procedure by setting
$\alpha^{*} = \max\left\{\widetilde{\alpha},\frac{h}{2}\right\}$.
Table \ref{comp_sugA} lists the VMS-POD model's average error $e = \frac{1}{N+1}{\sum\limits_{n=0}^{N}\|u^n-u_r^n\|}$ for the
following values of $r, R$ and $\alpha$: $r = 20, 40$ and $60$; $R$ from $5$ to $r-5$
in increments of $5$; and $\alpha = 0.01 \, \alpha^{*}, \, \alpha^{*}$, and
$100 \, \alpha^{*}$.
Note that the VMS-POD model consistently performs best for $\alpha = \alpha^{*}$.
The only two slight deviations from this rule are for $r = 60$ ($R = 20$ and $R = 30$);
we again believe that this is due to the mesh-size $h = 0.01$, which is not small
enough for the asymptotic regime in Theorem \ref{theorem_error}.
\begin{table}[htb]
\centering
\caption{VMS-POD model's average error $e = \frac{1}{N+1}{\sum\limits_{n=0}^{N}\|u^n-u_r^n\|}$ for different values of $r$ and $R$,
and $\alpha = 0.01 \, \alpha^{*}, \, \alpha^{*}$, and $100 \, \alpha^{*}$.
Note that the VMS-POD model consistently performs best for $\alpha = \alpha^{*}$.}
\label{comp_sugA}
{\small
\begin{tabular}{|c|c|cc|cc|cc|}
\hline
{$r$}&{$R$}&{$0.01\alpha^{*}$}&{$e$}&{$\alpha^{*}$}&{$e$}&{$100\alpha^{*}$}&{$e$}\\
\hline
\multirow{3}{*}{20}&5& $1.2\times10^{-3}$&$1.0\times10^{-1}$&$1.2\times10^{-1}$
&$5.8\times10^{-2}$&$1.2\times10^{1}$&$7.8\times 10^{-2}$\\
{}&10&$2.0\times10^{-3}$&$9.5\times10^{-2}$&$2.0\times10^{-1}$
& $2.4\times10^{-2}$&$2.04\times10^{1}$&$2.6\times10^{-2}$\\
{}&15&$3.3\times10^{-3}$&$8.2\times10^{-2}$&$3.3\times10^{-1}$
& $2.0\times10^{-2}$&$3.3\times10^{1}$&$2.5\times10^{-2}$\\
\hline
\multirow{5}{*}{40}&5&$6.4\times10^{-5}$&$1.09\times10^{-1}$&$6.4\times10^{-3}$&$3.0\times10^{-2}$&$6.4\times10^{-1}$ &$7.2 \times10^{-2}$\\
{}&10&$1.1\times10^{-4}$&$1.0\times10^{-1}$&$1.1\times10^{-2}$&$1.8\times10^{-2}$&$1.1\times10^{0}$ &$2.5\times10^{-2}$\\
{}&20&$4.2\times10^{-4}$&$9.7\times10^{-2}$&$4.2\times10^{-2}$&$4.4\times10^{-3}$&$4.2\times10^{0}$ &$4.1\times10^{-3}$\\
{}&30&$1.7\times10^{-3}$&$6.8\times10^{-2}$&$1.7\times10^{-1}$&$8.1\times10^{-3}$&$1.7\times10^{1}$ &$1.0\times10^{-2}$\\
{}&35&$3.0\times10^{-3}$&$4.9\times10^{-2}$&$3.0\times10^{-1}$&$2.1\times10^{-2}$&$3.0\times10^{1}$ &$2.4\times10^{-2}$\\
\hline
\multirow{3}{*}
{60}&5&{$5.0\times10^{-5}$}&{$8.7\times10^{-2}$}&$5.0\times10^{-3}$&$1.8\times10^{-2}$&{$5.0\times10^{-1}$}&{$7.0\times10^{-2}$}\\
{}&10&{$5.0\times10^{-5}$}&{$8.7\times10^{-2}$}&$5.0\times10^{-3}$&$1.3\times10^{-2}$&{$5.0\times10^{-1}$}&{$2.4\times10^{-2}$}\\
{}&20&{$5.0\times10^{-5}$}&{$8.7\times10^{-2}$}&$5.0\times10^{-3}$&$1.0\times10^{-2}$&{$5.0\times10^{-1}$}&{$3.9\times10^{-3}$}\\
{}&30&{$1.2\times10^{-4}$}&{$8.0\times10^{-2}$}&$1.2\times10^{-2}$&$4.4\times10^{-3}$&{$1.2\times10^{0}$}&{$7.4\times10^{-4}$}\\
{}&40&{$5.4\times10^{-4}$}&{$5.5\times10^{-2}$}&$5.4\times10^{-2}$&$1.2\times10^{-3}$&{$5.4\times10^{0}$}&{$2.4\times10^{-3}$}\\
{}&50&{$1.8\times10^{-3}$}&{$2.6\times10^{-2}$}&$1.8\times10^{-1}$&$1.3\times10^{-2}$&{$1.8\times10^{1}$}&{$1.4\times10^{-2}$}\\
{}&55&{$2.9\times10^{-3}$}&{$2.2\times10^{-2}$}&$2.9\times10^{-1}$&$1.1\times10^{-2}$&{$2.9\times10^{1}$}&{$1.2\times10^{-2}$}\\
\hline
\end{tabular}
}
\end{table}
\smallskip
Finally, we investigate numerically the VMS-POD model's sensitivity with respect
to changes in the diffusion coefficient $\varepsilon$.
To this end, we run the VMS-POD model \eqref{vms_pod} with the same parameters
as above ($r = 40, \, R = 20$ and $\alpha=\alpha^{*}$) for different values of the
diffusion coefficient: $\varepsilon = 10^{-2}, 10^{-4}$ and $10^{-6}$.
Table \ref{table_error_epsilons} lists the average errors for DNS, POD-G and
VMS-POD models for different values of $\varepsilon$.
It is clear from this table that the POD-G model's average error is significantly higher
than the error of the DNS.
The VMS-POD model, however, performs well for all values of $\varepsilon$ and displays
a low sensitivity with respect to changes in the diffusion coefficient.
Thus, we provide numerical support for the theoretical estimate \eqref{theorem_error_1},
which is {\em uniform} with respect to $\varepsilon$.
\begin{table}[htb]
\centering
\caption{Average errors of DNS, POD-G and VMS-POD models for different values of the
diffusion coefficient $\varepsilon$.
The POD-G model performs poorly.
The VMS-POD model performs well and displays low sensitivity with respect to changes in
$\varepsilon$.}
\label{table_error_epsilons}
{\small
\begin{tabular}{|c|c|c|cc|}
\hline
\multirow{2}{*}{$\varepsilon$}&{${DNS}$}&{POD-G}&\multicolumn{2}{c|}{VMS-POD}\\
\cline{2-5}
{}&{$\frac{1}{N+1}\sum\limits_{n=0}^{N}\|u^n-u_h^N\|$}&{$\frac{1}{N+1}{\sum\limits_{n=0}^{N}\|u^n-u_r^n\|}$}&$\alpha$&$\frac{1}{N+1}{\sum\limits_{n=0}^{N}\|u^n-u_r^n\|}$\\
\hline
{$10^{-2}$}&{$1.10\times10^{-4}$}&{$1.10\times10^{-2}$}&{$4.05\times10^{-2}$}&{$4.27\times10^{-3}$}\\
{$10^{-4}$}&{$2.04\times10^{-4}$}&{$1.11\times10^{-1}$}&{$4.29\times10^{-2}$}&{$4.48\times10^{-3}$}\\
{$10^{-6}$}&{$1.88\times10^{-4}$}&{$1.17\times10^{-1}$}&{$9.65\times10^{-2}$}&$4.05\times10^{-3}$\\
{$10^{-8}$}&{$2.46\times10^{-4}$}&{$1.17\times10^{-1}$}&{$1.01\times10^{-1}$}&{$4.05\times10^{-3}$}\\
\hline
\end{tabular}
}
\end{table}
\section{Conclusions}
\label{s_conclusions}
We presented a new VMS closure modeling strategy for the numerical stabilization
of POD-ROMs of convection-dominated equations.
The new POD-ROM, denoted VMS-POD, utilizes an artificial viscosity term to add
numerical stabilization to the model.
Following the guiding principle of the VMS methodology, we only add artificial
viscosity to the small resolved scales.
Thus, no artificial viscosity is used for the large resolved scales.
The POD setting represents an ideal framework for the VMS approach, since
the POD modes are listed in descending order of their kinetic energy content.
A thorough numerical analysis for the finite element discretization of the new
VMS-POD model was presented.
The numerical tests showed the increased numerical stability of the new VMS-POD
model and illustrated the theoretical error estimates.
We also employed the theoretical error estimates to provide guidance in choosing
the artificial viscosity coefficient in practical computations.
We emphasize that the theoretical error estimates were uniform with respect to
$\varepsilon$, the diffusion coefficient.
The numerical tests confirmed the theoretical results:
The average error of the VMS-POD model showed a low sensitivity with respect
to changes in $\varepsilon$.
Although the new VMS-POD model targets general convection-domainted problems,
it was analyzed theoretically and tested numerically by using the convection-dominated
convection-diffusion equations.
We chose this simplified mathematical and numerical setting as a first step in a thorough
investigation of the new VMS-POD model.
Next, we will utilize the new VMS-POD model in the numerical simulation of turbulent
flows, such as $3D$ flow past a circular cylinder \cite{wang2011two}.
We also note that, to our knowledge, this is the first time that the VMS formulation used
in \cite{layton2002connection} for the numerical stabilization of finite element discretizations
has been used in a POD setting.
We will investigate in a future study the alternative VMS formulation proposed in
\cite{guermond1999stabilization} and compare it with the VMS-POD model that we
introduced in this report.
\bibliographystyle{plain}
|
\section{Introduction}\label{sec:introduction}
Data loss recovery -- for instance, for content distribution applications or for distributed storage systems -- is
widely addressed using erasure codes that operate at the transport or the application layer of the communication
system. These codes, referred to as upper-layer (UL) codes, extend source data packets with repair (redundant) packets,
which are used to recover the lost data at the receiver. They are generally proposed in conjunction with physical layer
codes, in order to maximize the reliability of the transmission system, especially in case of intermittent connectivity
or deep fading of the signal for short periods.
In such situations, the physical layer FEC fails and we can either ask for retransmission (only if a return channel
exists, and penalizing in broadcast/multicast scenarios) or use UL-FEC. Hence, the use of UL-FEC codes is of critical
importance in broadcast communication systems in general, and satellite communications in particular.
Low Density Parity Check (LDPC) codes constitute a very broad class of FEC codes, distinguished by the fact that they
are defined by sparse parity-check matrices, and can be iteratively decoded in linear time with respect to their
block-length. Invented by Gallager in early 60's \cite{gall-monograph}, but considered impractical to implement, these
codes have been neglected for more that three decades, and ``rediscovered'' in the late 90's \cite{MacK}. Nowadays, a
large body of knowledge has been acquired (analysis, optimization, construction); LDPC codes are known to be capacity
approaching codes for a large class of channels \cite{Rich-Shok-Urba}, and became synonymous with modern coding.
However, this capacity approaching property holds in the asymptotic limit of the code length, and codes optimized from
this asymptotic perspective may suffer significant performance degradation at practical lengths. Actually, the
asymptotic optimization, performed by using density-evolution methods \cite{Rich-Urba}, yields an {\em irregularity
profile}, which specifies the distribution of node-degrees in the bipartite (Tanner) graph \cite{Tann} associated with
the code. It is assumed that the girth\footnote{Length of a shortest cycle.} of the bipartite graph goes to infinity
with the code-length. Hence, optimized irregularity profiles can be used to construct codes that are ``long enough''
(at least few thousand bits) to avoid short cycles, although they must be ``short enough'' to be practical.
One of the most widely-used method for constructing finite length codes is the Progressive Edge Growth (PEG) algorithm
\cite{PEG}. It constructs bipartite graphs with large girth, by establishing edges progressively: the graph grows in an
edge-by-edge manner, optimizing each local girth. There is an {\em underlying edge order} within the PEG, corresponding
to the order in which edges are established in the graph. In general, edges are progressively established starting with
those incident to symbol-nodes of degree-$2$ and ending with those incident to symbol-nodes of maximum degree. However,
any other order with respect to the symbol-node degrees would also be possible. Besides, for a given symbol-node
degree, edges can be established in a {\em node-by-node} manner (all edges incident to some symbol node are established
before moving to the next symbol-node), or in a {\em degree-by-degree} manner (a first edge is established for each
symbol-node, then a second edge is established for each symbol-node, and so on until all the symbol-nodes reach the
given degree). Although this order may significantly impact the performance of the constructed code, it is rather
difficult to formalize and has practically not been investigated in the literature. There are however several papers
that aim to enhance the PEG construction by optimizing some objective function, as for instance minimizing the number
of cycles created \cite{Venkiah08}, or minimizing the approximate cycle extrinsic (ACE) message degree
\cite{Tian04:ACE}, \cite{Xiao04}.
In this paper we consider the PEG algorithm together with a {\em scheduling distribution}, which will be referred to as
scheduled-PEG, or SPEG for short. Within the SPEG algorithm, symbol-nodes are divided into subsets, each subset
containing symbol-nodes of same degree. Edges incident to the symbol-nodes of a subset are established in a
degree-by-degree manner, before moving to the next subset. The scheduling distribution specifies the fraction of nodes
within each subset. Our purpose is to find a scheduling distribution that yields the best performance in terms of
decoding overhead (performance metric widely used for UL-FEC). We rigorously formulate this optimization problem, and
we show that it can be addressed by using genetic optimization algorithms.
The paper is organized as follows. Section \ref{sec:Defi_Nots} gives a brief overview of the basic theory and
definitions related to LDPC codes, their iterative decoding, and the associated performance metrics over the BEC. The
construction of finite length LDPC codes is addressed in Section \ref{sec:Constr_FL}. The proposed Scheduled-PEG
algorithm is also introduced in this section. Section \ref{sec:Opt_SPEG} focuses on the optimization of the Scheduled
PEG algorithm and presents simulation results. Finally, Section \ref{sec:Conclusion} concludes the paper.
\section{LDPC Codes and Performance Metric over the Erasure Channel}\label{sec:ldpc_codes}\label{sec:Defi_Nots}
\subsection{Binary and non-binary LDPC codes}
In this paper we consider both binary and non-binary LDPC codes defined over some finite field $\f_q$, with $q=2^p$
\cite{Davey-MacKey}. If $p=1$, the code is binary. We fix, once for all, a vector space isomorphism:
\begin{equation}
\f_q \tilde{\longrightarrow} \f_2^p \label{eq:vector_isomors_def}
\end{equation}
Elements of $\f_q$ will be called symbols. We say that the binary sequence $(x_0,..,x_{p-1}) \in \f_2^p$ is the
\textit{binary image} of the symbol $X \in \f_q$, iff they correspond to each other by the above isomorphism.
An LDPC code is a linear code defined by a sparse parity-check matrix $H\in{\mathbb{M}}_{m,n}(\f_q)$. Alternatively, it
can be represented by a bipartite (Tanner) graph\footnote{By abusing language, throughout the paper, the term ``code''
will be used with the meaning of ``bipartite graph''.} \cite{Tann}, containing $n$ symbol-nodes and $m$
constraint-nodes associated respectively with the $n$ columns and $m$ rows of $H$. A symbol-node and a constraint-node
are connected by an edge if and only if the corresponding entry of $H$ is non-zero, in which case the edge is assumed
to be ``labeled'' by the non-zero entry.
Symbol-nodes take values in $\f_q$, and a constraint-node is said to be verified if the linear combination of neighbor
symbols (with coefficients given by the corresponding edge labels) is equal to zero. A non-binary word
$(X_1,\dots,X_n)$ is a codeword if it verifies all the constraint nodes of the graph.
The degree of a node is by definition the number of edges incident to that node (number of non-zero entries on the
corresponding row/column of $H$). A code is called {\em $(d_s, d_c)$-regular} if all symbol-nodes are of degree $d_s$
and all constraint-nodes are of degree $d_c$; otherwise it is called {\em irregular}.
Let $\Lambda_d$ and $\Gamma_d$ denote respectively the fractions of symbol and constraint nodes of degree-$d$. Let also
$\lambda_d$ and $\rho_d$ be the fractions of edges connected respectively to symbol and constraint nodes of degree-$d$.
The degree distribution polynomials, from the node and the edge perspective, are defined by:
$$\begin{array}{r@{\ }lr@{\ }ll}
\Lambda(x) = & \displaystyle\sum_d \Lambda_d x^{d}, & \ \Gamma(x) = & \displaystyle\sum_d \Gamma_d x^{d} & \mbox{\footnotesize(node-persp.)} \\
\lambda(x) = & \displaystyle\sum_d \lambda_d x^{d-1}, & \ \rho(x) = & \displaystyle\sum_d \rho_d x^{d-1} & \mbox{\footnotesize(edge-persp.)}
\end{array}$$
The designed code rate, denoted by $r$, is by definition:
$$r = 1 - \displaystyle\frac{\int_{0}^{1}\rho(x)\text{d}x}{\int_{0}^{1}\lambda(x)\text{d}x} = 1 - \displaystyle\frac{\Lambda '(1)}{\Gamma'(1)}$$
If the parity-check matrix is of rank $m$, then the (non-binary) code dimension is equal to $k = n-m$, and $r$ is equal
to the code rate, that is $r = \frac{k}{n}$.
We also denote by $K$ and $N$ the {\em binary code dimension} and the {\em binary code length}, respectively. Hence, $K
= kp$ and $N = np$.
\subsection{Iterative erasure decoding}
For binary LDPC codes over the BEC, the belief-propagation (BP) decoding translates into a simple technique of
recovering the erased bits, by iteratively searching for check-nodes with only one erased neighbor bit-node
\cite{zyablov}. Indeed, if a check-node is connected to only one erased bit-node, the value of the latter can be
recovered as the XOR of all the other neighbor bit-nodes. This value is then injected into the decoder, and the
decoding continues by searching for some other check-node having a single erased neighbor bit-node. We remark that the
erasure decoding can be performed on-the-fly: decoding starts as soon as the first bit is received and each new
received bit is injected on-the-fly into the decoder. The decoding will stop by itself if all the bits have been
recovered, or when it ``gets stuck'' because any check-node is connected to at least two erased bit-nodes (such a
configuration is called a {\em stopping set}).
The above considerations can also be generalized in the case of non-binary codes. It is important to note that we
consider {\em non-binary LDPC codes over a binary erasure channel}, which means that the channel erases bits from the
binary image of the transmitted codeword. Hence, a coded symbol can be completely erased (all the bits of its binary
image are erased), completely received (no bit of its binary image is erased), or partially erased/received (some bits
of its binary image are erased, some others are received). At the receiver part, the received bits are used to
(partially) reconstruct the corresponding symbols of the transmitted codeword. Thus, for each symbol node we can
determine the set of {\em eligible symbols}, i.e. symbols whose binary images match the received bits. Such a set
contains only one symbol (namely the transmitted one) if the corresponding symbol-node is completely received. These
sets are then iteratively updated, according to the linear constraints between symbol-nodes \cite{savin2008nbl}.
Alternatively, non-binary LDPC codes can be decoded by using their extended binary image~\cite{savin:blte}.
\subsection{Performance metrics for finite-length codes}\label{subsec:perf_metrics}
A performance metric that is often associated with on-the-fly decoding is the {\em decoding inefficiency}, defined as
the ratio between the number of received bits when decoding completes and the number of information bits
\cite{Byers98}. More precisely, we assume that the encoded bit-stream is permuted according to some random permutation
$\pi$. The permuted bit-stream is sequentially delivered to the decoder, which performs erasure decoding
on-the-fly\footnote{Such a random reception corresponds to a randomly interleaved erasure channel, which allows us to
dispense with a specific loss model.}: each bit is injected into the decoder in the appropriate position and erasure
decoding is performed until either decoding completes or it gets stuck. If decoding gets stuck, we inject the next bit
from the permuted bit-stream. We denote by $k_\pi$ the number of bits from the permuted bit-stream that have been
injected into the decoder when the erasure decoding completes; the value $k_\pi-k$ is referred to as {\em reception
overhead}. The decoding inefficiency is defined as $\inef(\pi) = \frac{k_\pi}{k}$ . It is a random variable, whose
value depends on the (random) permutation $\pi$. Note also that $\inef(\pi)\in\left[1, \frac{1}{r}\right]$, where $r$
is the code rate. The average decoding inefficiency is defined as the expected value of $\inef$; that is:
$$ \inefm = {\text{E}}\left[\inef\right] = \frac{1}{n!}\sum_{\pi}\inef(\pi)$$
Note that $\inefm = 1$ if and only if the code is MDS (its minimum distance is equal to $n-k+1$). More accurate
statistics about the decoding inefficiency are provided by the {\em probability of decoding failure}, defined as the
complementary cumulative distribution function (CCDF) of $\inef$:
$$F(x) = \Pr[ \inef(\pi) > x ], \ \ x\in\left[1, \frac{1}{r}\right]$$
Hence, $F(x)$ is the probability of decoding failure assuming that the number of bits received from the channel is
equal to $kx$, or equivalently, the reception overhead is equal to $k(x-1)$. Indeed, the $kx$ bits received from the
channel are the first $kx$ bits of the encoded bit-stream permuted by some permutation $\pi$. Hence, a decoding failure
occurs if and only if $\inef(\pi) > x$.
From the above definition, it also follows that:
$$\int_{1}^{\frac{1}{r}} F(x) \ \mbox{d}x = \inefm - 1$$
Therefore, if $\code_1$ and $\code_2$ are two codes such that $\inefm(\code_1) < \inefm(\code_2)$, then $\code_1$
presents a better performance in the waterfall region of $F$, but this might happen at the expense of a higher error
floor\footnote{The waterfall is the region in which the failure probability decreases very quickly as $x$ increases.
However, there might be a point after which the curve does not fall as quickly as before, in other words, there is a
region in which performance flattens. This region is called the error floor region.}.
Finally, we note that the probability of decoding failure can also be expressed in relation with the fraction of erased
bits (rather than the fraction of received bits); in this case it will be referred to as {\em Frame Error Rate} (FER).
The {\em Bit Error Rate} (BER) will denote the probability of a bit being erased (after the decoding process), assuming
that a certain fraction of bits have been received.
\subsection{Asymptotic performance}
Unlike the finite-length performance, the asymptotic performance does not refer to the performance of a given code, but
to the performance of a given family or {\em ensemble of codes}. Such an ensemble contains arbitrary-length codes that
share the same properties in terms of distributions of node-degrees in the associated bipartite graph.
Let $E(\lambda,\rho)$ denote the ensemble of LDPC codes of arbitrary length $n>0$, with edge perspective
degree-distributions polynomials $\lambda$ and $\rho$. When $n$ goes to infinity, (almost) all the codes behave
alike, and they exhibit a threshold phenomenon, separating the region where reliable transmission is possible from that
where it is not \cite{Rich-Urba}. Assume that an arbitrary code $\code_n \in E(\lambda,\rho)$, of length $n$, is used
over the BEC, and let $p_e$ denote the channel erasure probability. The threshold of the ensemble $E(\lambda,\rho)$ is
defined as the supremum value of $p_e$ (i.e. the worst channel condition) that allows transmission with an arbitrary
small error probability, assuming that $n$ goes to infinity. Let us denote this threshold value by
$p_{\mbox{\scriptsize th}}(\lambda,\rho)$. The threshold value is necessarily less than the channel capacity, that is
$p_{\mbox{\scriptsize th}}(\lambda,\rho) \leq 1-r$, where $r$ is the (asymptotic) code rate of the ensemble
$E(\lambda,\rho)$. Roughly speaking, this means that if an encoded sequence of length $n$ is transmitted over the
channel, it can be successfully decoded iff the fraction of erased bits is less than $p_e < p_{\mbox{\scriptsize
th}}(\lambda,\rho)$, with $p_e \rightarrow p_{\mbox{\scriptsize th}}(\lambda,\rho)$ as $n\rightarrow +\infty$. It is
assumed here that the girth of the graph goes to infinity with $n$, which actually happens for almost all the codes in
$E(\lambda,\rho)$. It follows that the decoding inefficiency, which can be expressed as $\frac{(1-p_e)N}{K} =
\frac{1-p_e}{r}$, also goes to a threshold value:
$$\ainef = \frac{1-\pth}{r},$$
which will be referred to as {\em inefficiency threshold}.
Given an ensemble $E(\lambda,\rho)$, its threshold value can be efficiently computed by tracking the fraction of
erased messages passed during the belief propagation decoding. This method is called density evolution (the name is due
to the fact that over more general channels, we have to track the message densities). For more details on density
evolution we refer to \cite{Rich-Urba} for binary codes, and \cite{rathi:det} and \cite{savin2008nbl} for non-binary
LDPC codes. The introduction of irregular codes, as well as the asymptotic optimization based on the density evolution
method, made possible the construction of capacity approaching ensembles of LDPC codes \cite{Rich-Shok-Urba}.
\section{Finite length LDPC codes construction}\label{sec:Constr_FL}
As discussed in the above section, the asymptotic threshold can be approached by long codes, which do not contain short
cycles. Short cycles may also harm the performance of short (finite-length) codes, as they can result in short stopping
sets. Hence, the PEG algorithm has been proposed, and is widely used, for constructing bipartite graphs with large
girth, in a best effort sense, by progressively establishing edges between symbol and check nodes in an edge-by-edge
manner~\cite{PEG}.
\subsection{Progressive Edge Growth algorithm}
A bipartite (Tanner) graph is denoted as $(S, C, E)$, where $S = \left\{s_1,s_2,...,s_{n}\right\}$ is the set of
symbol-nodes, $C=\left\{c_1,c_2,...,c_{m}\right\}$ is the set of constraint nodes and $E \subseteq S \times C$ is the
set of edges. An edge $\eij \in E$ corresponds to a non-zero entry $\hij$ of the parity check matrix $H$. We also
denote by $D_S$ the ``target sequence'' of symbol-node degrees, which is assumed to be sorted in non-decreasing order:
$$ D_S = \left\{ d_{s_1}, d_{s_2},..., d_{s_{n}} \left| d_{s_1} \leq d_{s_2} \leq ... \leq d_{s_{n}} \right.\right\}$$
where $d_{s_j}$ is the degree of symbol node $s_j$.
When the PEG algorithm starts, the set of edges is empty, $E = \emptyset$. Edges will be progressively added to $E$, as
explained shortly. Given a symbol node $s_j$, we denote by $\underline{C}_{s_j}$ the set of constraint-nodes whose
distance to $s_j$ is maximum, in the current settings; that is, given the current set of edges. The distance between
two nodes is the length of the shortest path connecting them. If there is no path between $s_j$ and some constraint
node $c_j$, the distance between them is set to $+\infty$. Hence, if $\ee_{s_j} = \emptyset$, the distance from $s_j$
to any constraint-node is $+\infty$ and $\underline{C}_{s_j} = C$, the set of all the constraint-nodes. If $\ee_{s_j}
\neq \emptyset$, $\underline{C}_{s_j}$ can be determined by expanding a subgraph from symbol node $s_j$ up to the
maximal depth (see \cite{PEG}).
Finally, we use $c_i \longleftarrow \left\{\underline{C}_{s_j} \mid \mbox{min deg}\right\}$ to denote a random
constraint node $c_i \in \underline{C}_{s_j}$, having the lowest degree (given the current set of edges of the graph).
The PEG algorithm can be summarized as follows:
\begin{algorithm}
\caption{Progressive Edge Growth algorithm}
\label{algo:PEG_ORG}
\begin{algorithmic}
\FOR{$j$ = 1 \TO $n$}
\FOR{$k$ = 1 \TO $d_{s_j}$}
\STATE Determine $\underline{C}_{s_j}$, given the current $E$;
\STATE $c_i \longleftarrow \left\{\underline{C}_{s_j} \mid \mbox{min deg}\right\}$;
\STATE Add edge $(s_j, c_i)$ to $E$;
\ENDFOR
\ENDFOR
\end{algorithmic}
\end{algorithm}
We note that edges are established in a {\em node-by-node} manner, meaning that all edges incident to some symbol-node
are established before moving to the next symbol-node. We have further assumed that symbol nodes are sorted in
increasing order with respect to their degrees. Finally, we observe that the constraint-node degree distribution of the
constructed Tanner graph is almost uniform, {\em i.e.} all constraint nodes have only one or at most two consecutive
degrees \cite{PEG}.
Let $\peg$ denote the ensemble of all Tanner graphs constructed by using the PEG algorithm. The average inefficiency
ratio over all graphs in $\peg$ is defined as:
$$\inefpeg\inputdata = \text{E}\left[ \,\inefm(\code) \mid \code \in \peg \,\right]$$
When parameters $\inputdata$ are implied, it will be simply denoted by $\inefpeg$. The corresponding standard deviation
is denoted by $\stdpeg$.
\subsection{Modified Progressive Edge Growth algorithm}
The PEG graphs have large girth with respect to random graphs. Consequently, PEG graphs have low error floor in
comparison with random graphs. However, the error floor can be further lowered, by using a modification of the PEG
algorithm, called ModPEG.
Let $\vsd$ denote the set of symbol nodes of degree $d$, where $1 \leq d \leq \dm$ and $\dm$ denotes the maximum
symbol-node degree. Let $n_d$ denote number of symbol nodes within $\vsd$, hence $\sum_d n_d = n$.
Within the ModPEG algorithm, for each $\vsd$, edges are established in a {\em degree-by-degree} manner: a first edge is
established for each symbol-node, then a second edge is established for each symbol-node, and so on until all the
symbol-nodes in $\vsd$ reach the required degree $d$.
\begin{algorithm}
\caption{Modified Progressive Edge Growth algorithm}
\label{algo:PEG_VS}
\begin{algorithmic}
\FOR{$d$ = 1 \TO $\dm$}
\FOR{$k$ = 1 \TO $d$}
\FOR{$s_j \in \vsd$}
\STATE Determine $\underline{C}_{s_j}$, given the current $E$;
\STATE $c_i \longleftarrow \left\{\underline{C}_{s_j} \mid \mbox{min deg}\right\}$;
\STATE Add edge $(s_j, c_i)$ to $E$;
\ENDFOR
\ENDFOR
\ENDFOR
\end{algorithmic}
\end{algorithm}
Let $\mpeg$ denote the ensemble of all Tanner graphs constructed by using the ModPEG algorithm. The average
inefficiency ratio over all graphs in $\mpeg$ is defined as:
$$\inefmpeg\inputdata = \text{E}\left[ \,\inefm(\code) \mid \code \in \mpeg \,\right]$$
When parameters $\inputdata$ are implied, it will be simply denoted by $\inefmpeg$. The corresponding standard
deviation is denoted by $\stdmpeg$.
\begin{table*}[!t]
\centering \caption{Optimized irregular LDPC codes, with rate $1/2$} \label{tab:opt_irr_distr}
\begin{tabular}{||c|l||r|r||r|r||r|r||}
\hline
Alphabet & Degree distributions & $\pth$ & $\ainef$ & $\inefpeg$ & $\stdpeg$ & $\inefmpeg$ & $\stdmpeg$\\
\hline
$\f_2$ & $\begin{array}{l} \Lambda(x) = 0.5489 x^2 + 0.2505 x^3 + 0.1608 x^7 + 0.0398 x^{30} \\ \Gamma(x) = 0.6609 x^8 + 0.3391 x^9 \end{array}$ & 0.4955 & 1.009 & 1.0829 & 1.475e-04 & 1.0876 & 1.202e-04\\
\hline
$\f_4$ & $\begin{array}{l} \Lambda(x) = 0.7140 x^2 + 0.2173 x^4 + 0.0687 x^{12} \\ \Gamma(x) = 0.7586 x^6 + 0.2414 x^7 \end{array}$ & 0.4926 & 1.0148 & 1.0604 & 2.903e-04 & 1.0685 & 2.117e-04 \\
\hline
$\f_8$ & $\begin{array}{l} \Lambda(x) = 0.7857 x^2 + 0.0529 x^3 + 0.1153 x^5 + 0.0461 x^{12} \\ \Gamma(x) = 0.2797 x ^5 + 0.7203 x ^6 \end{array}$ & 0.4931 & 1.0138 & 1.0655 & 3.188e-04 & 1.0726 & 2.8e-04 \\
\hline
$\f_{16}$ & $\begin{array}{l} \Lambda(x) = 0.8460 x^2 + 0.1056 x^5 + 0.0252 x^8 + 0.0232 x^{18} \\ \Gamma(x) = 0.3221 x^5 + 0.6779 x^6 \end{array}$ & 0.4945 & 1.011 & 1.0706 & 6.81e-04 & 1.0834 & 3.536e-04 \\
\hline
\end{tabular}
\end{table*}
\begin{figure*}[!t]
\centering
\subfloat[$\f_2$]{\label{fig:peg_avg_f2}\includegraphics[width=\columnwidth]{peg_f2.eps}}
\subfloat[$\f_4$]{\label{fig:peg_avg_f4}\includegraphics[width=\columnwidth]{peg_f4.eps}} \\
\subfloat[$\f_8$]{\label{fig:peg_avg_f8}\includegraphics[width=\columnwidth]{peg_f8.eps}}
\subfloat[$\f_{16}$]{\label{fig:peg_avg_f16}\includegraphics[width=\columnwidth]{peg_f16.eps}}
\caption{Average inefficiency ratios of ensemble Tanner graphs constructed by using optimized Scheduled PEG algorithm in comparison with original PEG and modified PEG algorithms for $K = 5000$ information bits}
\label{fig:peg_avg}
\end{figure*}
\begin{expl}
\label{expl:fin_perf_vs_asymp_perf} We consider irregular LDPC codes of rate $1/2$, defined over the $\f_{2}, \f_{4},
\f_{8}, \f_{16}$. The node-perspective symbol-node degree distributions and the asymptotic performance are shown in
Table \ref{tab:opt_irr_distr} (the constraint-node degree distributions are considered to be almost uniform). These
codes have been optimized by using density evolution methods. The binary code can be found in \cite{Shokrollahi2000b},
while non-binary codes have been optimized within this work. For each degree distribution, $100$ Tanner graphs have
been constructed, by both PEG and ModPEG algorithms, for codes with binary dimension $K = 5000$. The corresponding
average inefficiency ratios are also shown in Table \ref{tab:opt_irr_distr}. The details of these inefficiency ratios
are shown in Figure \ref{fig:peg_avg}, where it can be seen that using $\approx 20$ graphs is sufficient for obtaining
good estimates of $\inefpeg$ and $\inefmpeg$ values.
We also observe that, in all the four cases, the average inefficiency ratios $\inefpeg$ and $\inefmpeg$ are much larger
than $\ainef$. Also, the average inefficiency ratios corresponding to ModPEG are larger than those corresponding to
the original PEG. The reason is that the ModPEG algorithm improves the error floor region but also worsens the
waterfall region, as it can be seen in Figure \ref{fig:peg_ber_expl}: there are eight BER curves, corresponding to
eight LDPC codes drawn from the corresponding PEG and ModPEG ensembles of graphs.
\end{expl}
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{peg_ber.eps
\caption{BER performance of Tanner graphs constructed by using PEG and ModPEG algorithms, for irregular
LDPC codes shown in Table \ref{tab:opt_irr_distr}, with with $K = 5000$ bits.
\label{fig:peg_ber_expl
\end{figure}
\subsection{Scheduled Progressive Edge Growth algorithm}
Both PEG and ModPEG algorithms construct Tanner graph with large girth and low error floor. For optimized irregular
LDPC codes, however, the average inefficiency of finite length graphs is far away from the predicted (asymptotical)
threshold (even for codes with $K = 5000$ and $N = 10000$ bits, cf. Example \ref{expl:fin_perf_vs_asymp_perf}).
The Scheduled Progressive Edge Growth (SPEG) algorithm, proposed in this section, aims to improve the average
inefficiency of irregular LDPC codes.
We fix an integer $T \geq 1$. We consider a collection of disjoint symbol-nodes subsets $\vsdt\subseteq S$, indexed by
$t\in\{1,\dots,T\}$ and $d\in\{1,\dots\dm\}$, such that:
\begin{itemize}
\item $\vsdt \subseteq S_d$ ($\vsdt$ contains only symbol-nodes of degree $d$)
\item $\vs = \displaystyle \bigcup_{d=1}^{\dm} \bigcup_{t=1}^{T} \vsdt$.
\end{itemize}
We denote by $\ndt$ the number of symbol-nodes in $\vsdt$. It follows that $S_d = \cup_{t=1}^{T}\vsdt$, $n_d =
\displaystyle \sum_{i=0}^{t-1}\ndt$, and $n = \displaystyle \sum_{d=1}^{\dm} \sum_{i=0}^{t-1}\ndt$
The scheduled PEG algorithm works as follows: at time $t$, it connects progressively symbol-nodes within the subsets
$\vsdt$, for $d=1,\dots,\dm$. For each subset $\vsdt$, edges are established in a {\em degree-by-degree} manner.
\begin{algorithm}
\caption{Scheduled Progressive Edge Growth algorithm}
\label{algo:SPEG}
\begin{algorithmic}
\FOR{$t$ = 1 \TO $T$}
\FOR{$d$ = 1 \TO $\dm$}
\FOR{$k$ = 1 \TO $d$}
\FOR{$s_j \in \vsdt$}
\STATE Determine $\underline{C}_{s_j}$, given the current $E$;
\STATE $c_i \longleftarrow \left\{\underline{C}_{s_j} \mid \mbox{min deg}\right\}$;
\STATE Add edge $(s_j, c_i)$ to $E$;
\ENDFOR
\ENDFOR
\ENDFOR
\ENDFOR
\end{algorithmic}
\end{algorithm}
We denote by $\speg$ the ensemble of all Tanner graphs constructed by using the SPEG algorithm. The average
inefficiency ratio over all graphs in $\speg$ is defined as:
$$\inefspeg\!\inputdataSPEG = \text{E}\!\left[ \inefm(\code) \mid \code \in \speg \right]$$
When parameters $\inputdataSPEG$ are implied, it will be simply denoted by $\inefspeg$. The corresponding standard
deviation is denoted by $\stdspeg$.
In the simple case of $T = 1$, we have $S_d^{(1)} = S_d$, and the SPEG algorithm is equivalent to the ModPEG
algorithm. On the other hand, if $T = n$, each subset $\vsdt$ contains one single symbol-node, and the SPEG algorithm
is equivalent to the original PEG algorithm. For a general $T$ value, the SPEG algorithm is in between these two
extreme cases. It allows to explore the ensemble of LDPC codes with {\em fixed code-length and degree distributions}
and, as explained shortly, is aimed at finding codes with very small average inefficiency. As observed in Section
\ref{subsec:perf_metrics}, a lower average inefficiency corresponds to better performance of the code in the waterfall
region. However, generally there is a tradeoff between waterfall and error floor regions. Hence, to prevent excessive
degradation in the error floor region, the edges within each $\vsdt$ subset are established in a {\em degree-by-degree}
manner.
The use of the SPEG algorithm can potentially improve the average inefficiency of the constructed LDPC code, as shown
in the following example.
\begin{expl}\label{expl:speg_potential}
We consider non-binary LDPC codes with rate $1/2$, defined over $\f_{16}$, whose node-perspective degree distribution
polynomials are given in Table \ref{tab:opt_irr_distr}. The asymptotic threshold is equal to $\pth = 0.4945$,
corresponding to an asymptotic inefficiency $\ainef = 1.011$.
We want to construct a code with binary dimension $K = 5000$. Hence, the Tanner graph contains $n = 2500$ symbol-nodes
and $m = 1250$ constraint-nodes. We use the SPEG algorithm with $T = 3$, such that $n_d^{(t_1)} = n_d^{(t_2)}$, for any
$1\leq t_1,t_2 \leq T$. This means that each $S_d$ is partitioned into three subsets $\vsdt, t=1,2,3$, of same
cardinality.
As in Example \ref{expl:fin_perf_vs_asymp_perf}, we estimated the average inefficiency ratio of the ensemble $\speg$ by
simulating $100$ SPEG-codes. We obtained $\inefspeg = 1.0536$, which has to be compared with $\inefpeg = 1.0706$ and
$\inefmpeg = 1.0834$ from Table \ref{tab:opt_irr_distr}.
\end{expl}
This example shows that an appropriate choice of the subsets $\left\{\vsdt\right\}$ may improve the average
inefficiency of the constructed code. In the following section we propose a method that allows to optimize this choice,
by minimizing the average inefficiency of the corresponding ensemble of SPEG-codes.
\section{Optimized Scheduled-PEG construction}\label{sec:Opt_SPEG}
\subsection{Optimization algorithm}
The main idea behind the SPEG algorithm is that different choices of the {\em scheduling subsets}
$\left\{\vsdt\right\}$ might lead to codes with different performance. Our purpose is to find scheduling subsets that
minimize the average inefficiency of the corresponding ensemble of SPEG-codes. In order to properly formulate this
optimization problem, we have to take into consideration only the ``profile'' of the scheduling subsets, which consists
of the fractions of nodes within each subset. Precisely, let $\fsdt$ denote the fraction of symbol-nodes contained in
$\vsdt$; hence:
\begin{enumerate}
\renewcommand{\theenumi}{(\arabic{enumi}}
\item $\fsdi = \displaystyle\frac{\ndi}{n}$
\item $\displaystyle \sum_{t=1}^{T}\fsdt = \Lambda_d$
\item $\displaystyle \sum_{d=1}^{\dm}\sum_{t=1}^{T}\fsdt = \sum_{d=1}^{\dm}\Lambda_d=1$
\end{enumerate}
A family of parameters $\left\{\fsdt\right\}_{\mbox{\!\!\scriptsize$\begin{array}{l} 1\leq t \leq T \\
1 \leq d \leq \dm\end{array}$\!\!\!\!\!\!\!\!\!}}$ satisfying the condition (2) before is called {\em scheduling
distribution}.
When parameters $(n, m, \left\{\fsdt\right\})$ are fixed, we consider that the SPEG algorithm starts by randomly
choosing a family of scheduling subsets $\left\{\vsdt\right\}$, according to the given scheduling distribution, and
then it constructs a (random) Tanner graph as explained in the above section. We denote by $\spegopt$ the corresponding
ensemble of SPEG Tanner graphs, and we define its average inefficiency by:
$$\inefspeg\!\inputdataSPEGopt = \text{E}\!\left[ \inefm(\code) \mid \code \in \spegopt \right]$$
When parameters $n$ and $m$ are fixed, it will be simply denoted by $\inefspeg\left(\left\{\fsdt\right\}\right)$. This
represents the objective function of our optimization problem. Although it cannot be computed analytically,
$\inefspeg\left(\left\{\fsdt\right\}\right)$ can be efficiently estimated by simulating a finite number of codes $\code
\in \spegopt$.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{DiffEvolAlgo.eps
\caption{Flowchart of the differential evolution optimization algorithm
\label{fig:speg_flowchart
\end{figure}
Since our objective function cannot be expressed analy\-tically, we address the optimization problem by using generic
population-based metaheuristic optimization algorithms. More precisely, the {\em differential evolution} algorithm
\cite{DiffEvol} is used, which optimizes $\inefspeg\left(\left\{\fsdt\right\}\right)$ by iteratively trying to improve
the best current solution $\left\{f_{b,d}^{(t)}\right\}$. The flowchart of the optimization algorithm is illustrated in
Figure \ref{fig:speg_flowchart}.
The algorithm maintains a population of candidate solutions, which is randomly initialized. At each iteration, the
current population is evaluated, such as to find the best current solution. Then, a population of new candidate
solutions is generated by combining existing candidates (mutation and recombination), and then keeping whichever
candidate solution has the best score. In this way the objective function is treated as a black box that provides a
measure of quality of the candidate solutions.
\subsection{Simulation results}
We consider four ensembles of irregular LDPC codes of rate $1/2$, defined over the $\f_{2}, \f_{4}, \f_{8}, \f_{16}$,
whose node-perspective symbol-degree distributions are shown in Table \ref{tab:opt_irr_distr} (the constraint-node
degree distributions are considered to be almost uniform).
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{ineff_ratio.eps
\caption{Average inefficiency ratios of the ensembles $\peg$ and $\spegopt$
\label{fig:ineff_ratio
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{speg_ber_fer.eps
\caption{Bit Erasure Rates of three (among the best) codes from the ensembles $\peg$, $\mpeg$, and $\spegopt$
\label{fig:speg_ber_fer
\end{figure}
\begin{table*}[!t]
\centering \caption{Optimized scheduling distributions and corresponding inefficiency ratios, for irregular LDPC codes
of rate $1/2$, defined over $\f_2$, $\f_4$, $\f_8$ and $\f_{16}$, with binary dimension $K=5000$} \label{tab:opt_simul}
\begin{tabular}{|c|l|r|r||r|r||r|}
\hline
Alphabet & Optimized Scheduling & $\stdspeg$ & $\inefspeg$ & $\inefpeg$ & $\inefmpeg$ & $\inefrand$\\
\hline \hline
$\f_2$ & $\begin{array}{@{\!\!}|c|c|c|c|c|@{\!\!}}
t & f_{2}^{(t)} & f_{3}^{(t)} & f_{7}^{(t)} & f_{30}^{(t)} \\
\hline
1 & 0.2939 & 0.0690 & 0 & 0.0071 \\
\cline{1-1}
2 & 0.2523 & 0.1797 & 0.0787 & 0.0223 \\
\cline{1-1}
3 & 0.0028 & 0.0018 & 0.0820 & 0.0104 \\
\end{array}$
& 7.7e-04 & 1.0326 & 1.0829 & 1.0876 & 1.3367 \\
\hline \hline
$\f_4$ & $\displaystyle\begin{array}{@{\!\!}|c|c|c|c|@{\!\!}}
t & f_{2}^{(t)} & f_{4}^{(t)} & f_{12}^{(t)} \\
\hline
1 & 0.1876 & 0.0099 & 0.0040 \\
\cline{1-1}
2 & 0.2088 & 0 & 0.0021 \\
\cline{1-1}
3 & 0.3175 & 0.2073 & 0.0626 \\
\end{array}$
& 3.301e-04 & 1.0384 & 1.0604 & 1.0685 & 1.3129 \\
\hline \hline
$\f_8$ & $\begin{array}{@{\!\!}|c|c|c|c|c|@{\!\!}}
t & f_{2}^{(t)} & f_{3}^{(t)} & f_{5}^{(t)} & f_{12}^{(t)} \\
\hline
1 & 0.1120 & 0.0028 & 0.0020 & 0.0065 \\
\cline{1-1}
2 & 0.1844 & 0.0084 & 0.0036 & 0.0026 \\
\cline{1-1}
3 & 0.4893 & 0.0417 & 0.1097 & 0.0371 \\
\end{array}$
& 3.049e-04 & 1.0327 & 1.0655 & 1.0726 & 1.2256 \\
\hline \hline
$\f_{16}$ & $\begin{array}{@{\!\!}|c|c|c|c|c|@{\!\!}}
t & f_{2}^{(t)} & f_{5}^{(t)} & f_{8}^{(t)} & f_{18}^{(t)} \\
\hline
1 & 0.4118 & 0.0473 & 0 & 0.0004 \\
\cline{1-1}
2 & 0.1662 & 0.0004 & 0.0002 & 0.0071 \\
\cline{1-1}
3 & 0.2680 & 0.0579 & 0.0023 & 0.0157 \\
\end{array}$
& 6.407e-04 & 1.0302 & 1.0706 & 1.0834 & 1.1905 \\
\hline
\end{tabular}
\end{table*}
We fixed the parameter $T=3$ and, for each of the above ensembles, we used the differential evolution algorithm to find
a scheduling distribution that minimizes the average inefficiency of the corresponding SPEG codes. We considered codes
with binary dimension $K = 5000$; Hence, the binary code-length is $N=10000$ and the non-binary code-length (number of
symbol-nodes of the graph) is $n=N/p$, where $p\in\{1,2,3,4\}$. The optimized scheduling distributions
$\left\{\fsdt\right\}$ and the corresponding inefficiencies $\inefspeg\left(\left\{\fsdt\right\}\right)$ are shown in
Table \ref{tab:opt_simul}. For comparison purposes, we have also displayed in Table \ref{tab:opt_simul} the average
inefficiencies of the corresponding PEG, ModPEG, and random\footnote{Random codes with given $n$, $m$, and node-degree
distribution polynomials.} ensembles. We can see that the SPEG algorithm significantly improves the average
inefficiency ratios in comparison with those of the PEG and ModPEG algorithms.
Moreover, we constructed several finite length codes over $\f_{16}$, using the same scheduling distribution that has
been optimized for $K=5000$. The average inefficiency of these codes is plotted in Figure \ref{fig:ineff_ratio}. It can
be observed that they significantly outperform codes constructed by the original PEG algorithm.
For $K=5000$, Figure \ref{fig:speg_ber_fer} displays the bit error rates of one PEG code, one ModPEG code, and one SPEG
code with optimized scheduling distribution (these codes are chosen among the best codes of the corresponding
ensembles). We can see that the SPEG algorithm significantly improves the waterfall region, at the expense of a
slightly higher error floor.
\section{Conclusion}\label{sec:Conclusion}
The proposed Scheduled-PEG algorithm allows the enhancement of the classical PEG algorithm, by the introduction of a
scheduling distribution that specifies the order in which edges are established in the graph. The scheduling
distribution provides a way for exploring the ensemble of LDPC codes with fixed code-length and degree distributions,
and is aimed at finding codes with very small average inefficiency.
We showed that the SPEG algorithm can be successfully combined with genetic optimization algorithms, which
significantly improves the average inefficiency of the constructed LDPC codes over the classical-PEG construction. In
terms of error rate curves, this translates into a significant improvement of the waterfall region.
Finally, we remark that the optimization of the scheduling distribution makes use of the specific channel model
(through the use of the decoding inefficiency). Hence, LDPC codes constructed by using the SPEG algorithm together with
an optimized scheduling distribution are channel dependent. However, the proposed algorithm could be generalized for
more general channel models (e.g. by optimizing with respect to a target FER, or to the area under the FER curve.).
\bibliographystyle{./Bibliography/IEEEbib}
\footnotesize
|
\section{Introduction}
The phenomenon of flavor oscillation plays an important role in the physics of neutral meson and neutrino systems. In particular, flavor oscillation provides the only means to measure the extremely small mass and decay rate splittings among the neutral mesons, and also provides convincing evidence for the existence of non-zero neutrino masses. The theoretical descriptions of flavor oscillation fall into several categories, including the basic plane wave Pontecorvo formalism \cite{Pontecorvo:1968, Bilenky:1978lm}, intermediate \cite{Kayser:1981qm,Giunti:1991wd,Giunti:1992rw,Lipkin:1995tn,Kiers:1995zj,Grossman:1996eh,Lipkin:1999nb,Giunti:2000kw,Giunti:2002xg,Beuthe:2002ej,Bilenky:2011pk} and external \cite{Rich:1993qm, Giunti:1993se, Grimus:1996av, Giunti:1997sk, Campagne:1997no, Grimus:1998ft, Ioannisian:1999no, Stockinger:2000if} wavepacket approaches and quantum field theoretic results \cite{Okun:1982sr, Alfinito:1995sn, Blasone:1999ef, Binger:1999nj, Blasone:2001du, Blasone:2002jv, Li:2006qt, Cohen:2009dn, Keister:2010rq, Dvornikov:2010dc}. Some detailed reviews of these approaches, their underlying assumptions, results and difficulties can be found in Refs. \cite{PDG:2010,Bilenky:1978lm, Beuthe:2001rc, Akhmedov:2010ms} (and references therein). To be very brief: In the first, one assumes the flavor states are unitary combinations of plane-wave mass eigenstates that follow spacetime worldlines, and one must carefully define the proper times of the mass eigenstates in order to obtain the well-known Pontecorvo oscillation formula. The intermediate wavepacket approach treats the oscillating degrees of freedom as a linear combination of one-particle states, while the external wavepacket approach treats the oscillating particles as quantum fields, whose propagator is convolved with wavepackets at the source and detector.
A large amount of Literature has been devoted to deriving, studying and comparing oscillation formulae within these different approaches. Particular care has been taken to include important effects such as measurement uncertainties, coherence effects, the finite size of the detector and source, all of which together lead to somewhat complicated formulae. Our goal in this paper is less ambitious: Using a quantum field theoretic approach, we present a simple formalism of oscillation based entirely on the properties of the \emph{spatially} Fourier transformed propagator. We call this the spatial two-point function. The resulting oscillation formulae are particularly elegant, and precisely reproduce both the Pontecorvo neutrino result and (CP violating) neutral meson mixing results (see e.g. Ref. \cite{Branco:1999cp,Bigi:2009cp}) in appropriate parameter regimes.
To construct this formalism, we assume that the oscillation experiment measures the exchanged energy $E$ and source-detector displacement $\bm{L}$ to infinite precision, along with flavor at both the source and detector. This is possible because $E$, $\bm{L}$ and flavor are commuting observables, so that an amplitude which depends exclusively on these quantities is well-defined. The key idea is that the spatial two-point function $\Delta(E,\bm{L})$ in the flavor basis is a well-defined amplitude which encodes flavor oscillation over a displacement $\bm{L}$ at energy $E$. We therefore assume that the experiment amplitude is proportional to $\Delta(E,\bm{L})$ and explore the resulting oscillation formulae. The advantages of this description are: There is no ambiguity in the choice of reference frame - all computations are done in the lab frame and one never needs to introduce proper times into the formalism; the oscillation probabilities can be computed exactly; and one obtains formulae whose physical meaning can be easily discerned in various limits. Since we do neglect several real physical effects mentioned above --- in particular the physics of the source and detector are neglected --- the limits of the applicability of our theoretical description to actual oscillation experiments should be carefully examined. Nonetheless, we believe this approach provides an instructive, leading order description of the physics of oscillation in real experiments.
In terms of the previous Literature on this subject, our approach is best categorized as a special case of the above-mentioned external wavepacket formalism with stationary states \cite{Grimus:1996av}. However, to our knowledge, the oscillation physics contained just in the two-point function has not been thoroughly investigated and the resulting general oscillation formulae for unstable particles obtained by our approach have not been previously presented. One exception is Ref. \cite{Okun:1982sr}, whose amplitudes for stable fermions agree with our results for the special case of stable particles.
This paper is structured as follows. In Section \ref{sec:F} we present the oscillation formalism. In Section \ref{sec:EOP} the exact spatial two-point function $\Delta(E,\bm{L})$ for unstable fields is presented, and the exact oscillation probabilities and integrated oscillation probabilities are computed. In Section \ref{sec:R} we examine our results in several different parameter regimes and recover both the neutral meson-mixing results and Pontecorvo neutrino oscillation results in appropriate limits.
\section{Formalism}
\label{sec:F}
\subsection{Experiment Amplitude}
\label{sec:EA}
Our starting point is to consider an experiment which involves the propagation between a source and a detector of a set of fields $\{\phi^\alpha\}$, which are allowed to mix. As usual, $\alpha$ is an experimentally measurable label called the flavor, which is henceforth always denoted by a Greek index. The set $\{\phi^\alpha\}$ is called the flavor field basis.
In this paper we assume the $\alpha \to \beta$ oscillation experiment measures the exchanged energy $E$ and source-detector displacement $\bm{L}$ to infinite precision in the lab frame. The amplitude for the experiment must then have the form
\begin{equation}
\label{eqn:FFA}
\mathcal{M} = \mathcal{M}_{\alpha\beta}(E,\bm{L})~.
\end{equation}
This is a well-defined amplitude since $E$, $\bm{L}$ and flavor are commuting observables. Note that as a consequence of the infinitely precise $E$ and $\bm{L}$ measurement neither the time of travel nor the three-momentum between the source and detector is well-defined, because these observables do not commute with $E$ and $\bm{L}$ respectively. In other words the initial and final states of this amplitude must be energy-spatial eigenstates, rather than momentum-time eigenstates.
The key idea of this paper rests on the observation that the time Fourier-transformed time-ordered exact two-point function --- the spatial two-point function --- defined by
\begin{equation}
\Delta_{\alpha\beta}(E,\bm{L}) \equiv \int dt \Big\langle T\Big\{\phi^{\beta}(t,\bm{L})\phi^{\alpha\dagger}(0,\bm{0})\Big\}\Big\rangle e^{iEt}~,
\end{equation}
is the field theoretic object which encodes the oscillation of flavor $\alpha \to \beta$ over a displacement $\bm{L}$ with energy $E$. (As usual $\langle T\{\phi^{\beta}(x)\phi^{\alpha\dagger}(y)\}\rangle \equiv \Delta_{\alpha\beta}(x-y)$ is a function of $x-y$ due to translation invariance.) It is therefore natural to write
\begin{equation}
\label{eqn:DEA}
\mathcal{M}_{\alpha\beta}(E,\bm{L}) = \mathcal{A}^\alpha_{\textrm{S}}\mathcal{A}^\beta_{\textrm{D}}\Delta_{\alpha\beta}(E,\bm{L})~,
\end{equation}
(no sum over $\alpha$, $\beta$) where $\mathcal{A}^{\alpha}_{\textrm{S},\textrm{D}}$ encode the physics of the source and detector, which we have assumed factorizes out of the amplitude\footnote{The general criteria under which such a factorization may be possible in real oscillation experiments has been examined previously in detail (see e.g. Ref. \cite{Akhmedov:2010ms}). Foregoing such a discussion, our intent here is that the physics of the source and detector can be neglected up to their ability to distinguish flavor.}. Assuming $\mathcal{A}^{\alpha}_{\textrm{S},\textrm{D}}$ are known, the implication of Eq. (\ref{eqn:DEA}) is that $|\Delta_{\alpha\beta}(E,\bm{L})|^2$ is a measurable quantity, from which we may proceed to construct oscillation probabilities.
So far we have not specified the spin of $\phi^\alpha$. As is well-known, if $\phi^\alpha$ are massive they must create spin-$j$ particles, with $j$ a half-integer, which have $2j+1$ spin degrees of freedom. We assume that these spin degrees of freedom decouple, so that we need only consider scalar propagators henceforth.
\subsection{Oscillation Probability}
Having written down the amplitude for the experiment, we now define the flavor oscillation probability via
\begin{equation}
\label{eqn:OPD}
P_{\alpha \to \beta}(E,\bm{L}) \equiv \frac{\big|\Delta_{\alpha\beta}(E,\bm{L})\big|^2}{\underset{\gamma}{\sum}\big|\Delta_{\alpha\gamma}(E,\bm{L})\big|^2}~.
\end{equation}
Here $P_{\alpha\to\beta}$ forms a well-defined probability distribution, since $P_{\alpha\beta} \ge 0$ and $\sum_\beta P_{\alpha\beta} = 1$.
In some experiments, measurement of $|\Delta_{\alpha\beta}|^2$ at a precise $\bm{L}$ is replaced by a volume-averaged measurement,
\begin{equation}
\mathcal{A}^\textrm{I}_{\alpha\beta}(E) \equiv \int d^3\!\bm{L} \big|\Delta_{\alpha\beta}(E,\bm{L})\big|^2~.
\end{equation}
This is equivalent to the time-averaged amplitudes measured in e.g. meson mixing experiments, in which the initial and final flavor states are determined by tagging via decay products (see e.g. Ref. \cite{Branco:1999cp,Bigi:2009cp}). We can correspondingly define an integrated oscillation probability
\begin{equation}
\label{eqn:IOPD}
P_{\alpha \to \beta}^{\textrm{I}}(E) \equiv \frac{\mathcal{A}^\textrm{I}_{\alpha\beta}(E)}{\underset{\gamma}{\sum}\mathcal{A}^\textrm{I}_{\alpha\gamma}(E)} =\frac{ \int d^3\!\bm{L}\big|\Delta_{\alpha\beta}(E,\bm{L})\big|^2}{\underset{\gamma}{\sum}\int d^3\!\bm{L}\big|\Delta_{\alpha\gamma}(E,\bm{L})\big|^2}~.
\end{equation}
This is also a well-defined probability distribution.
\subsection{Propagator and 1PI Basis}
So far in this paper we have formulated a description of flavor oscillation in terms of just the exact quantum amplitude $\Delta(E,\bm{L})$, which is equivalently defined as the spatial Fourier transform of the exact propagator, $\Delta(p^2)$. Explicitly,
\begin{equation}
\label{eqn:EDSTP}
\Delta_{\alpha\beta}(E,\bm{L}) = \int \frac{d^3\bm{p}}{(2\pi)^3}\Delta_{\alpha\beta}(p^2)e^{i\bm{p}\cdot\bm{L}}~.
\end{equation}
Applying external field methods to the path-integral formulation of quantum field theory, it is a well-known result (see e.g. \cite{Weinberg:1996vol2,Zinn-Justin:2002}) that for a set of $N$ fields $\{\phi^\alpha\}$ the exact two-point function is the inverse of the exact two-point one-particle-irreducible (1PI) function: $\Delta_{\alpha\beta}(x-y) = -\Pi^{-1}_{\alpha\beta}(x-y)$. The Fourier transform of this result is
\begin{equation}
\label{eqn:EP}
\Delta_{\alpha\beta}(p^2) = \bigg[\frac{i}{p^2\bm{1} - M^2(p^2)}\bigg]_{\alpha\beta}~.
\end{equation}
Henceforth we shall call the $N\times N$ matrix of functions $M^2(p^2)$ the exact two-point 1PI function.
In general, one cannot compute the exact propagator $\Delta(p^2)$ exactly for all $p^2$. However, the combination of Eqs (\ref{eqn:EDSTP}) and (\ref{eqn:EP}) suggests that the exact spatial two-point function is sensitive only to the pole structure of $\Delta(p^2)$. As we shall see below, with suitable assumptions this pole structure depends only on physical masses and rest frame decay rates, permitting us to construct exact oscillation probabilities in terms of just these measureable quantities, $E$, $\bm{L}$, and a mixing matrix, despite our incomplete knowledge of the exact propagator.
Now, the exact propagator (\ref{eqn:EP}) is generally not diagonal in flavor space --- there would be no oscillation if this were the case --- but the analytic structure of $\Delta(p^2)$ is greatly simplified if the exact propagator can be diagonalized. Ultimately, we want to be able to write
\begin{equation}
\label{eqn:GDU}
\Delta_{\alpha\beta}(p^2) = U^{\alpha j} (U^{-1})^{j\beta}\Delta_{j}(p^2)~,\qquad \Delta_j(p^2) \equiv \frac{i}{p^2 - M_j^2(p^2)}~,
\end{equation}
so that
\begin{equation}
\label{eqn:CBD}
\Delta_{\alpha\beta}(E,\bm{L}) = U^{\alpha j}(U^{-1})^{j\beta}\Delta_j(E,\bm{L})~,\qquad \Delta_j(E,\bm{L}) \equiv \int \frac{d^3p}{(2\pi)^3} \frac{ie^{i\bm{p}\cdot \bm{L}}}{p^2 - M_j^2(p^2)}~.
\end{equation}
In Eq. (\ref{eqn:GDU}), $U$ is the constant and possibly unitary matrix that diagonalizes $\Delta(p^2)$ (equivalently $M^2(p^2)$), and $M^2_j(p^2)$ are the $N$ eigenvalues of $M^2(p^2)$. Below we'll see that the $M_j(p^2)$ determine the physical masses and rest frame decay rates of the particles propagating in $\Delta_{\alpha\beta}(p^2)$.
In constrast to the usual diagonalization of the classical Lagrangian mass terms, diagonalization of the exact propagator may be non-trivial. In Appendix \ref{sec:AB} we discuss the details of the diagonalization of $\Delta(p^2)$, the properties of $U$ and how this exact quantum formalism both relates to and differs from the usual classical mixing matrix formalism. For our purposes here, we assume $\Delta(p^2)$ is diagonalizable in the manner of Eq. (\ref{eqn:GDU}). Unless otherwise stated, we also assume $U$ is unitary. An immediate consequence of unitarity is that spatial two-point function can now be written as
\begin{equation}
\Delta_{\alpha\beta}(E,\bm{L}) = U^{\alpha j}U^{\beta j *}\Delta_j(E,\bm{L})~.
\end{equation}
Let us now define the 1PI basis. This basis is a generalization of the mass basis derived in the classical formalism, that may accommodate both unstable particles and a description of CP violation for two flavors. In particular, if $U$ is constant (but not necessarily unitary), then there exists a well-defined second basis of fields $\{\phi^j\}$, henceforth denoted by a Latin index, which are defined via the linear transformations
\begin{equation}
\label{eqn:UT}
\phi^{\alpha\dagger} = U^{\alpha j}\phi^{j\dagger}~.
\end{equation}
Note that $\phi^\dagger$ creates a particle state, while $\phi$ creates an anti-particle state: We have chosen this definition of basis change by $U$ in order that it coincides with the usual definition in terms of one-particle quantum states. We call $\phi^j$ the 1PI basis for the following reason. If $U$ is unitary, then observe that not only $M^2(p^2)$ but also the two-point function is diagonal, i.e. $\langle T \{\phi^{i}(x)\phi^{j\dagger}(y)\} \rangle = \delta_{ij}\Delta_j(x-y)$. This implies that $M^2_j(p^2)$ is the 1PI function for $\phi^j$, whence the name. In contrast, if $U$ is not unitary, then even though $M^2(p^2)$ is still diagonalized by $U$, we have $\langle T \{\phi^{i}(x)\phi^{j\dagger}(y)\} \rangle \not= \delta_{ij}\Delta_j(x-y)$. Therefore $M_j^2(p^2)$ is no longer the 1PI function for $\phi^j$. Nonetheless, we shall always refer to the field basis defined by Eq. (\ref{eqn:UT}) to be the 1PI basis, and often call $\phi^{j\dagger}$ the 1PI states.
Tying the diagonalization of the exact propagator and the definition of 1PI basis together, we can now explain why we have taken care to consider the case of non-unitary $U$: We do so to accommodate CP-violating two-flavor neutral meson oscillations (for three or more flavors, even unitary $U$ may have a CP-violating phase), which we consider in Sec. \ref{sec:NUM}. The idea is that the Hamiltonian for such a system is diagonalized by a constant non-unitary matrix \cite{Bigi:2009cp,Branco:1999cp}, so we therefore expect $U$ to be non-unitary too. In this context the flavor field basis (1PI basis) then corresponds to the CP conjugate states (evolution eigenstates). One deduces $U$ is a particular constant $2\times 2$ non-unitary matrix, from which we can immediately derive the usual oscillation formulae.
\subsection{Exact Propagator Analytic Structure}
Let us finally examine the analytic structure of $\Delta_j(E,\bm{L})$, which is explicitly
\begin{equation}
\label{eqn:1PIES}
\Delta_j(E,\bm{L}) = \int \frac{d^3p}{(2\pi)^3} \frac{ie^{i\bm{p}\cdot \bm{L}}}{p^2 - M_j^2(p^2)}~.
\end{equation}
If $\phi^j$ is unstable, then the propagator $\Delta_j(p^2)$ will have a unique Breit-Wigner or resonance pole, which by convention is a simple pole located at
\begin{equation}
\label{eqn:GEP}
p^2 = m_j^2 -im_j\Gamma_j~.
\end{equation}
Here $m_j$ is the physical mass and $\Gamma_j \ge 0$ is the rest frame decay rate. The non-zero imaginary part for this pole enforces the usual Feynman pole prescription and associated time-ordering, so that we need not add the usual $i\epsilon$ convergence term in the denominator of Eq. (\ref{eqn:1PIES}), provided we assume $\Gamma_j \not= 0$. Consequently, taking the $\Gamma_j \to 0^+$ limit, which corresponds to $\phi^j$ being stable, can only be performed after all integrations and other limits are evaluated.
By definition there are no higher order poles in $\Delta_j(p^2)$ and the residue at the pole (\ref{eqn:GEP}) is unity: Eq. (\ref{eqn:GEP}) and this latter condition are equivalent to $M^2_j(m_j^2 - im_j\Gamma_j) = m_j^2 - im_j\Gamma_j$ and $M^{2\prime}_j(m_j^2 - im_j\Gamma_j) = 0$ respectively.
\section{Exact Oscillation Probability}
\label{sec:EOP}
\subsection{Spatial Two-Point Function}
Computation of the oscillation probabilities (\ref{eqn:OPD}) and (\ref{eqn:IOPD}) boils down to computing the spatial two-point function $\Delta_{j}(E,\bm{L})$. As shown in Appendix \ref{app:CD}, the integral (\ref{eqn:1PIES}) can be performed exactly, with final result (\ref{eqn:AEPR})
\begin{equation}
\Delta_{j}(E,L) = \frac{i}{4\pi L}\exp\Bigg\{\frac{i}{\sqrt{2}}\bigg[\sqrt{R_j^2 + A_j^2} + R_j\bigg]^{1/2}L - \frac{1}{\sqrt{2}}\bigg[\sqrt{R_j^2 + A_j^2} - R_j\bigg]^{1/2}L\Bigg\}~.\label{eqn:EPR}
\end{equation}
in which
\begin{align}
R_j & \equiv E^2 - m_j^2~,\notag\\
A_j & \equiv m_j\Gamma_j~. \label{eqn:DAR}
\end{align}
Note that the exact result in Eq. (\ref{eqn:EPR}) is independent of the orientation of $\bm{L}$~.
\subsection{Exact Probabilities}
We may now compute the exact oscillation probability via application of Eqs. (\ref{eqn:OPD}), (\ref{eqn:CBD}) and (\ref{eqn:EPR}), and the exact integrated oscillation probability via Eqs. (\ref{eqn:IOPD}), (\ref{eqn:CBD}) and (\ref{eqn:EPR}). It is convenient to define the wavenumber and characteristic inverse decay lengths
\begin{align}
\omega_j & \equiv \frac{1}{\sqrt{2}}\bigg[\sqrt{R_j^2 + A_j^2} + R_j\bigg]^{1/2}~,\notag\\
\zeta_j & \equiv \frac{1}{\sqrt{2}}\bigg[\sqrt{R_j^2 + A_j^2} - R_j\bigg]^{1/2}~,\label{eqn:DOZ}
\end{align}
along with
\begin{equation}
\label{eqn:DDOZ}
\Delta \omega_{jk} \equiv \omega_j - \omega_k~, \qquad \Delta \zeta_{jk} \equiv \zeta_j - \zeta_k~, \qquad \bar{\zeta}_{jk} \equiv \zeta_j + \zeta_k~.
\end{equation}
We call $\Delta \omega_{jk}$ the oscillation wavenumber.
Exploiting the unitarity of $U$, one finds the exact oscillation probability
\begin{align}
P_{\alpha\to\beta}(E,L)
& = \Bigg\{\underset{j}{\sum} |U^{\alpha j}|^2|U^{\beta j}|^2e^{-2\zeta_{j}L} + 2\underset{j<k}{\sum} \mbox{Re} \Big[U^{\alpha j}U^{\beta j*}U^{\alpha k *}U^{\beta k}
e^{i\Delta\omega_{jk}L}e^{-\bar{\zeta}_{jk}L}\Big]\Bigg\}\notag\\
& \quad \times \bigg[\underset{j}{\sum} |U^{\alpha j}|^2e^{-2\zeta_{j}L}\bigg]^{-1}~.\label{eqn:GOPF}
\end{align}
For $\zeta_j \to 0^+$, this has the exact form of the Pontecorvo oscillation formula \cite{Pontecorvo:1968,PDG:2010}. We will show below that within a certain parameter regime, $\Delta \omega_{jk} \simeq (m_k^2 -m_j^2)/2E$, recovering the usual result. The exact integrated oscillation probability is similarly
\begin{align}
P_{\alpha\to\beta}^{\textrm{I}}(E)
& = \Bigg\{\underset{j}{\sum}|U^{\alpha j}|^2|U^{\beta j}|^2/2\zeta_j + 2\underset{j<k}{\sum} \mbox{Re}\bigg[U^{\alpha j}U^{\beta j*}U^{\alpha k *}U^{\beta k}\frac{\bar{\zeta}_{jk} + i\Delta\omega_{jk}}{\bar{\zeta}_{jk}^2 + \Delta\omega_{jk}^2}\bigg]\Bigg\}\notag\\
& \quad \times \bigg[\underset{j}{\sum} |U^{\alpha j}|^2/2\zeta_j\bigg]^{-1}~. \label{eqn:GIOPF}
\end{align}
(The $\Delta_j$ normalization $i/4\pi L$ plays an important role in computing the integrals in Eq. (\ref{eqn:IOPD}).) Note that for both Eqs. (\ref{eqn:GOPF}) and (\ref{eqn:GIOPF}) we have not assumed CP conservation.
\subsection{Two-Flavor Formulae}
It is particularly illuminating to present the oscillation probability and integrated oscillation probability for the case that there are just two flavors. In this case we can choose $U$ to be real, orthogonal: The only physical parameter is the mixing angle and there is no CP violation. We adopt the convention for two flavors that $\alpha = +,-$ and $j = 1,2$. One obtains oscillation probability
\begin{align}
P_{\alpha \to \beta}(E,L)
& = \bigg\{|U^{\alpha 1}|^2|U^{\beta 1}|^2\bigg\}\bigg[|U^{\alpha 1}|^2 + |U^{\alpha 2}|^2e^{-2\Delta \zeta_{21} L}\bigg]^{-1}\notag\\
& + \bigg\{|U^{\alpha 2}|^2|U^{\beta 2}|^2\bigg\}\bigg[|U^{\alpha 1}|^2e^{+2\Delta \zeta_{21} L} + |U^{\alpha 2}|^2\bigg]^{-1}\notag\\
& + \bigg\{2U^{\alpha 1} U^{\beta 1 } U^{\alpha 2 } U^{\beta 2} \cos(\Delta\omega_{12} L)\bigg\}\bigg[|U^{\alpha 1}|^2e^{+\Delta \zeta_{21} L} + |U^{\alpha 2}|^2e^{-\Delta \zeta_{21} L}\bigg]^{-1}~. \label{eqn:ODPE}
\end{align}
Assuming without loss of generality that $\Delta \zeta_{21} >0$, then for $\Delta \zeta_{21}L \gg 1$, the first term is asymptotically constant, the second decays to zero while the third term produces a damped oscillation decaying to zero, with oscillation wavenumber $\Delta \omega_{12}$. In particular, for $\Delta \zeta_{21}L \gg 1$,
\begin{equation}
\label{eqn:LDUPO}
P_{\alpha \to \beta}(E,L) \simeq |U^{\beta 1}|^2~.
\end{equation}
If we adopt the notation that $(\phi^\beta)^\dagger$ creates $|\beta\rangle$ and $(\phi^j)^\dagger$ creates $|j\rangle$, then the right side of Eq. (\ref{eqn:LDUPO}) is nothing but $|\langle \beta|1\rangle|^2$. This is the probability of the $\beta$ flavor state being measured as the $j=1$ 1PI state, which has the longer decay length $1/\zeta_1 > 1/\zeta_2$. This behavior is familiar to that found in the $K$ neutral meson system: Since the $K_L$ eigenstate has a much longer decay length than the $K_S$, then the $K_L$ will be exponentially more abundant at large distances from the source compared to the $K_S$ state. As a result, at large distances there is no more oscillation and the oscillation probabilities $K \to K$ or $\overline{K} \to K$ both collapse to $|\langle K|K_L\rangle|^2$. This is exactly the behavior in Eq. (\ref{eqn:LDUPO}).
Before presenting the two-flavor integrated oscillation probability, for convenience we first define
\begin{equation}
\label{eqn:DXY}
x \equiv \frac{\Delta\omega_{12}}{\bar{\zeta}_{12}}~,\qquad y \equiv \frac{\Delta\zeta_{21}}{\bar{\zeta}_{12}}~.
\end{equation}
In Sec. \ref{sec:SMS} below we shall verify that $x$ and $y$ reduce to their usual definitions $x \simeq \Delta m /\bar{\Gamma}$ and $y \simeq \Delta \Gamma/2 \bar{\Gamma}$ within a certain regime of the parameters $E$, $m_{1,2}$ and $\Gamma_{1,2}$. With the definitions (\ref{eqn:DXY}), Eq. (\ref{eqn:GIOPF}) reduces to
\begin{align}
P_{\alpha\to\beta}^{\textrm{I}}(E)
& = \Bigg\{|U^{\alpha 1}|^2|U^{\beta 1}|^2(1+y) + |U^{\alpha 2}|^2|U^{\beta 2}|^2(1-y) + 2U^{\alpha 1}U^{\beta 1}U^{\alpha 2}U^{\beta 2}\bigg[\frac{1-y^2}{1+x^2}\bigg]\Bigg\}\notag\\
& \quad \times \bigg[1 + y\big(|U^{\alpha 1}|^2 - |U^{\alpha 2}|^2\big)\bigg]^{-1}~.\label{eqn:TFIOP}
\end{align}
\subsection{CP Violation and Non-Unitary Diagonalization for Two Flavors}
\label{sec:NUM}
The above oscillation probabilities (\ref{eqn:GOPF}) and (\ref{eqn:GIOPF}) (or (\ref{eqn:OPD}) and (\ref{eqn:IOPD})) may be generalized to the case that $U$ is constant and non-unitary, which is applicable to the study of CP violation in two-flavor neutral meson mixing. In this context, we identify the flavor fields $(\phi^{\pm})^{\dagger}$ as the creation operators of the CP conjugate states $|P^0\rangle$ and $|\overline{P^0}\rangle$, while the 1PI basis $(\phi^{1,2})^{\dagger}$ create the evolution eigenstates $|P_{L,H}\rangle$ respectively. Comparing Eq. (\ref{eqn:UT}) with the usual notation for $|P^0\rangle$ and $|\overline{P^0}\rangle$ in terms of $|P_{L,H}\rangle$ (assuming CPT symmetry)
\begin{align}
|P^0\rangle & = \frac{1}{2p}\big(|P_L\rangle + |P_H\rangle\big)~,\notag\\
|\overline{P^0}\rangle & = \frac{1}{2q}\big(|P_L\rangle - |P_H\rangle\big)~,\label{eqn:POPLH}
\end{align}
then leads to the identification
\begin{equation}
\label{eqn:UCP}
U = \frac{1}{2pq}\bordermatrix{ & \mbox{\tiny{1}} & \mbox{\tiny{2}} \cr \mbox{\tiny{$+$}} & q & q\cr \mbox{\tiny{$-$}} & p & -p\cr}~,
\end{equation}
which is non-unitary for $|p/q| \not = 1$. As $|p/q| \not=1$ is sufficient for CP violation in Eq. (\ref{eqn:POPLH}), the consequence of the identification (\ref{eqn:UCP}) is that CP violation in two-flavor mixing is manifested as non-unitary diagonalization of the two-point function. Note also that if $U$ is non-unitary then the 1PI states are no longer orthogonal, as expected for the evolution eigenstates $|P_{L,H}\rangle$. That is $\langle T \phi^{i}(x)\phi^{j\dagger}(y) \rangle \not= \delta_{ij}\Delta_j(x-y)$. From Eq. (\ref{eqn:GDU}) we have for non-unitary $U$
\begin{equation}
\label{eqn:CBDNU}
\Delta_{\alpha\beta}(E,L) = U^{\alpha j}(U^{-1})^{j\beta}\Delta_{j}(E,L) ~,
\end{equation}
and from Eqs. (\ref{eqn:EPR}) and (\ref{eqn:DOZ}) one then finds exact amplitudes
\begin{align}
\big|\Delta_{\pm\pm}(E,L)\big|^2 & = \frac{1}{32\pi^2 L^2}e^{-\bar{\zeta}_{12}L}\Big[\cosh(\Delta\zeta_{12}L) + \cos(\Delta \omega_{12} L)\Big]~,\notag\\
\big|\Delta_{+-}(E,L)\big|^2 & = \frac{|q|^2}{|p|^2}\frac{1}{32\pi^2 L^2}e^{-\bar{\zeta}_{12}L}\Big[\cosh(\Delta\zeta_{12}L) - \cos(\Delta \omega_{12} L)\Big] \notag\\
& = \frac{|q|^4}{|p|^4}\big|\Delta_{-+}(E,L)\big|^2 ~.\label{eqn:ECPA}
\end{align}
These strongly resemble the amplitudes found from the usual meson mixing quantum mechanical analysis (see below), except that here a spatial dependence has replaced the usual time dependence and $\Delta\omega_{jk}$, $\Delta\zeta_{jk}$ and $\bar{\zeta}_{jk}$ have the form presented in Eq. (\ref{eqn:DDOZ}). One finds exact integrated oscillation probabilities
\begin{align}
P^{\textrm{I}}_{+\to +} & = \frac{2 + x^2 - y^2}{2 + x^2 - y^2 + \big|q/p\big|^2(x^2 + y^2)}~,\notag\\
P^{\textrm{I}}_{+\to -} & = \frac{x^2 + y^2}{x^2 + y^2 + \big|p/q\big|^2(2 + x^2 - y^2)}~,\label{eqn:EIPPQ}
\end{align}
and one can also contemplate measuring the ratio of the amplitudes for oscillation into either flavor state
\begin{align}
F &\equiv \frac{\mathcal{A}^{\textrm{I}}_{+-}(E)}{\mathcal{A}^{\textrm{I}}_{++}(E)} = \frac{\int{\rm
d}L|\Delta_{+-}(E,L)|^2}{\int{\rm d}L|\Delta_{++}(E,L)|^2} \notag\\
& = \bigg|\frac{q}{p}\bigg|^2\frac{x^2 + y^2}{2 + x^2 - y^2}~.\label{eqn:EARPQ}
\end{align}
Let us compare the exact results in Eqs. (\ref{eqn:EIPPQ}) and (\ref{eqn:EARPQ}) with the analogous formulae obtained via the usual quantum mechanical treatment of neutral meson oscillations. Following Refs. \cite{Branco:1999cp,Bigi:2009cp} we can write down the time evolution for an initial pure $|P^0\rangle$ and $|\overline{P^0}\rangle$ state
\begin{align}
|P^0(t)\rangle=g_+(t)|P^0\rangle+\frac{q}{p}~g_-(t)|\overline{P^0}\rangle\notag\\
|\overline{P^0}(t)\rangle=\frac{p}{q}~g_-(t)|P^0\rangle+g_+(t)|\overline{P^0}\rangle~,\label{eqn:QM}
\end{align}
where
\begin{equation}
|g_\pm(t)|^2=\frac{e^{-\bar{\Gamma} t}}{2}\Big[\cosh(\Delta\Gamma t/2)\pm\cos(\Delta m t)\Big]~.\label{eqn:EVO}
\end{equation}
In the standard notation
\begin{equation}
x=\Delta m/\bar{\Gamma}~,\qquad y=\Delta\Gamma/2\bar{\Gamma}~,
\end{equation}
where $\bar{\Gamma}=(\Gamma_1+\Gamma_2)/2$, the two formulae in Eqs. (\ref{eqn:EIPPQ}) should be compared with
\begin{equation}
\frac{\int_0^\infty{\rm d}t~|\langle P^0|P^0(t)\rangle|^2}{\int_0^\infty{\rm d}t~\Big[|\langle P^0|
P^0(t)\rangle|^2+|\langle \overline{P^0}| P^0(t)\rangle|^2\Big]}=\frac{2+x^2-y^2}{2+x^2-y^2+|q/
p|^2(x^2+y^2)}
\end{equation}
and
\begin{equation}
\frac{\int_0^\infty{\rm d}t~|\langle \overline{P^0}|P^0(t)\rangle|^2}{\int_0^\infty{\rm
d}t~\Big[|\langle P^0|P^0(t)\rangle|^2+|\langle \overline{P^0}| P^0(t)\rangle|^2\Big]}=
\frac{x^2+y^2}{x^2+y^2+|p/q|^2(2+x^2-y^2)}~.
\end{equation}
Finally Eq. (\ref{eqn:EARPQ}) should be compared to
\begin{equation}
\frac{\int_0^\infty{\rm d}t~|\langle \overline{P^0}|P^0(t)\rangle|^2}{\int_0^\infty{\rm
d}t~|\langle P^0|P^0(t)\rangle|^2}=\left|\frac{q}{p}\right|^2\frac{x^2+y^2}{2+x^2-y^2}~.
\end{equation}
Our exact results are in perfect agreement with those of the usual analysis, except that in Eqs. (\ref{eqn:EIPPQ}) and (\ref{eqn:EARPQ}), the parameters $x$ and $y$ have the more general definitions encoded in Eqs. (\ref{eqn:DAR}), (\ref{eqn:DOZ}) and (\ref{eqn:DXY}). Once again, in Sec. \ref{sec:SMS} below we shall verify that $x$ and $y$ reduce to their usual definitions $x \simeq \Delta m /\bar{\Gamma}$ and $y \simeq \Delta \Gamma/2 \bar{\Gamma}$ within a certain regime of the parameters $E$, $m_{1,2}$ and $\Gamma_{1,2}$.
Before proceeding, please note that the oscillation probabilities (\ref{eqn:GOPF}), (\ref{eqn:GIOPF}), (\ref{eqn:ODPE}), (\ref{eqn:TFIOP}) and (\ref{eqn:ECPA}) presented in this section are a function of only $\Delta\omega_{jk}$, $\Delta\zeta_{jk}$, $\bar{\zeta}_{jk}$ or $\zeta_j$ (of the latter three variables, only two are independent). Consequently, specifying just $\omega_j$ and $\zeta_j$ is sufficient to specify the oscillation probabilities. This shall be our practice throughout the remainder of this paper.
\section{Regimes}
\label{sec:R}
The results presented in Sec. \ref{sec:EOP} are elegant and concise, but their physical interpretation is not obvious. However, in various regimes of the parameters $E$, $m_j$ and $\Gamma_j$, our exact results for the wavenumber and characteristic inverse decay lengths $\omega$ and $\zeta$ reduce to simpler expressions with clear physical meanings. In this section we explore several different regimes of physical interest, and show that in certain regimes our results reproduce the well-known neutrino and neutral meson oscillation formulae.
\subsection{Particle Regime}
\label{sec:PR}
The first regime of interest is the case
\begin{equation}
\label{eqn:DPR}
R_j \gg A_j~,~~\mbox{i.e.}~~E^2 -m_j^2 \gg m_j\Gamma_j~.
\end{equation}
It is straightforward to expand Eq. (\ref{eqn:DOZ}) about $A_j = 0$, and to leading order in $A_j/R_j$ one obtains the wavenumber and characteristic inverse oscillation lengths
\begin{equation}
\label{eqn:PROZ}
\omega_j \simeq \sqrt{E^2 -m_j^2}~,\qquad\zeta_j \simeq \frac{m_j\Gamma_j}{2\sqrt{E^2 - m_j^2}}~.
\end{equation}
The oscillation probabilities (\ref{eqn:GOPF}) and (\ref{eqn:GIOPF}) follow immediately from this and Eqs. (\ref{eqn:DDOZ}), as do their two-flavor versions (\ref{eqn:ODPE}) and (\ref{eqn:TFIOP}). In particular, in this regime we have to leading order in $A_j/R_j$
\begin{equation}
\label{eqn:PRXY}
x \simeq 2\frac{\sqrt{R_1} - \sqrt{R_2}}{A_1/\sqrt{R_1} + A_2/\sqrt{R_2}}~,\qquad y \simeq \frac{A_1/\sqrt{R_1} - A_2/\sqrt{R_2}}{A_1/\sqrt{R_1} + A_2/\sqrt{R_2}}~.
\end{equation}
In this regime, the spatial two-point function for a 1PI state
\begin{equation}
\label{eqn:PRSP}
\Delta_j(E,L) \simeq \frac{i}{4\pi L} \exp\Bigg\{i\sqrt{E^2 -m_j^2}L - \frac{m_j\Gamma_j}{2\sqrt{E^2-m_j^2}}L\Bigg\}~.
\end{equation}
For $\Gamma_j \to 0^+$, this looks precisely like the propagator of an on-shell one-particle state with momentum $p = (E^2 -m_j^2)^{1/2}$. We therefore call the regime (\ref{eqn:DPR}) the particle regime. The resemblance of the amplitude (\ref{eqn:PRSP}) to that of a particle suggests that we should obtain both the neutrino and meson mixing oscillation formula within the particle regime. We explicitly verify this in the next two sections.
The physical meaning of the spatial two-point function perhaps becomes more clear if we define analogous Lorentz factors and proper time
\begin{equation}
\label{eqn:DBPR}
\gamma_j = E/m_j~, \qquad \beta_j = \sqrt{1 - m_j^2/E^2}~,\qquad\tau_j = L/\gamma_j\beta_j~.
\end{equation}
Substituting these into Eq. (\ref{eqn:PRSP}) we obtain the spatial two-point function in terms of $\tau_j$ instead of $L$, which we can interpret as a `rest frame' propagator. Explicitly,
\begin{equation}
\Delta_j(\tau_j) \simeq \frac{i}{4\pi L}\exp\bigg\{i m_j(\gamma^2_j-1)\tau_j - \frac{\Gamma_j\tau_j}{2}\bigg\}~.
\end{equation}
The second term in the exponential looks like the usual rest frame decay of an unstable particle, and in particular it is clear that $\Gamma_j$ can be interpreted as the rest frame decay rate. The first term looks like the usual proper time evolution of a particle, except for the $\gamma^2-1$ factor. This factor arises because $pL$ is not a Lorentz invariant, but rather $E\gamma\tau - pL = m\tau$ is. It is a consequence of the experiment measure $E$ rather than the time of transit between the source and detector.
Let us now proceed to verify that the usual neutrino and neutral meson mixing oscillation formulae are obtained in this particle regime.
\subsection{Small Mass Splitting: Neutral Meson Oscillation}
\label{sec:SMS}
In all known neutral meson systems, the mass difference between the two mass eigenstates is extremely small in comparison with their masses. For the $K$, $D$, $B_d$ and $B_s$ neutral meson systems one finds \cite{PDG:2010}
\begin{equation}
\label{eqn:MSD}
\bigg(\frac{\Delta m}{m}\bigg)_K \sim \bigg(\frac{\Delta m}{m}\bigg)_D \sim10^{-14}~,\qquad \bigg(\frac{\Delta m}{m}\bigg)_{B_d} \sim 10^{-13}~,\qquad \bigg(\frac{\Delta m}{m}\bigg)_{B_s} \sim 10^{-12}~.
\end{equation}
It seems then, that the appropriate regime for neutral meson oscillation is the particle regime with the additional constraint that the mass splitting is small. We define the mean mass $m$ and mass splitting $\Delta m$ via $m_1 = m + \Delta m/2$ and $m_2 = m - \Delta m/2$, so the small mass splitting limit is $\Delta m/m \ll 1$. We also define $y_0 \equiv \Delta \Gamma/2\bar{\Gamma}$, in which $\bar{\Gamma} \equiv (\Gamma_1 + \Gamma_2)/2$ and $\Delta \Gamma \equiv \Gamma_2 - \Gamma_1$.
Expanding the particle regime expressions (\ref{eqn:PRXY}) for $x$ and $y$ in the small mass splitting limit is complicated by the fact that neither $x$ nor $y$ can be expressed as function of $\Delta m/m$ alone. However, one may show that
\begin{align}
x & \simeq \frac{\Delta m}{\bar{\Gamma}}\bigg[1 + \frac{y_0}{2\beta^2}\frac{\Delta m}{m} + \sum_{p=2}^\infty \frac{X_p(y_0,m,E)}{2^p\beta^{2p}}\bigg(\frac{\Delta m}{m}\bigg)^p\bigg]\notag\\
y & \simeq \frac{\Delta \Gamma}{2\bar{\Gamma}}\bigg[1 + \frac{1- y_0^2}{2\beta^2y_0}\frac{\Delta m}{m} + \sum_{p=2}^\infty \frac{Y_p(y_0,m,E)}{2^p\beta^{2p}}\bigg(\frac{\Delta m}{m}\bigg)^p\bigg]~,\label{eqn:XYSM}
\end{align}
for which $X_p$ and $Y_p$ are rational functions of $y_0$, $m$ and $E$. The parameter $\beta$ is defined as in Eqs. (\ref{eqn:DBPR}), but for mass $m$. In general, any of $X_p$, $Y_p$ or $1/\beta$ could be arbitrarily large for some configuration of the parameters $y_0$, $m$ and $E$, so the expansions (\ref{eqn:XYSM}) are not always well-controlled power series in $\Delta m/m$. However, it's plausible that the expansions are well-controlled in the parameter space regimes, and respective very small mass splittings (\ref{eqn:MSD}), relevant to the neutral meson systems. In such regimes, we then obtain for sufficiently small mass splittings
\begin{equation}
x \simeq \frac{\Delta m}{\bar{\Gamma}}~,\qquad y \simeq \frac{\Delta \Gamma}{2\bar{\Gamma}}~.
\end{equation}
These are precisely the usual definitions for the parameters $x$ and $y$ in the neutral meson mixing formalism. Moreover, in terms of the `proper time' $\tau$ - defined for $m$ in Eqs. (\ref{eqn:DBPR}) - the same expansion in small mass splitting renders the CP violating amplitudes (\ref{eqn:ECPA})
\begin{align}
\big|\Delta_{++}(E,L)\big|^2 &\simeq \frac{1}{32\pi^2 L^2}\exp(-\bar{\Gamma}\tau)\Big[\cosh(\Delta\Gamma \tau/2) + \cos(\Delta m \tau)\Big]~,\notag\\
\big|\Delta_{+-}(E,L)\big|^2 & \simeq \frac{1}{32\pi^2 L^2}\frac{|q|^2}{|p|^2}\exp(-\bar{\Gamma}\tau)\Big[\cosh(\Delta\Gamma \tau/2) - \cos(\Delta m \tau)\Big]~.
\end{align}
Up to a normalization of $1/16\pi^2L^2$, these are exactly the time evolution amplitudes found within the usual meson mixing analysis \cite{Branco:1999cp,Bigi:2009cp} as is evident in Eq. (\ref{eqn:EVO}). The physical interpretation of $\tau$ is the proper time elapsed in the rest frame of a classical particle with mass $m$ and lab frame energy $E$ that traverses a distance $L$.
Since we have already verified that, in terms of $x$ and $y$, our integrated oscillation probabilities match those found in the usual treatment, we have therefore recovered the well-known meson mixing amplitudes and time-integrated probabilities from the structure of the two-point function alone. Our analysis, however, also implies that the usual meson mixing results are valid \emph{only} within the small mass splitting particle regime. Outside this regime the more general results of Sec. \ref{sec:NUM} will apply.
An immediate question is whether the regimes of validity of our derivation and the standard quantum mechanical one disagree. The standard derivation is performed in terms of time evolution, and requires a \emph{common} proper time for the mass eigenstates \cite{Branco:1999cp,Bigi:2009cp}, which are one-particle states. If the energy of the oscillation experiment is fixed, as we assume throughout this paper, then this assumption is equivalent to assuming $\Delta m \to 0$. So the standard derivation is similarly applicable only in the small mass splitting limit.
\subsection{(Ultrarelativistic) Stable Particle Regime: Neutrino Oscillation}
For a neutrino oscillation experiment we expect the neutrinos to be ultrarelativistic and stable in the lab frame. The ultrarelativistic stable limit of the particle regime corresponds to $E \gg m_j$ and $\Gamma_j \to 0^+$ for all $j$. Expanding in this limit, the wavenumbers and characteristic inverse decay lengths (\ref{eqn:PROZ}) become, to leading order in $m_j/E$,
\begin{equation}
\label{eqn:SPROZ}
\omega_j \simeq E - \frac{m_j^2}{2E}~,\qquad \zeta_j = 0~,
\end{equation}
so that
\begin{equation}
\label{eqn:SPR}
\Delta_j(E,L) \simeq \frac{i}{4\pi L}\exp\Bigg\{iEL - i\frac{m_j^2}{2E}L\Bigg\}~.
\end{equation}
Applying Eqs. (\ref{eqn:GOPF}) and the unitarity of $U$ leads immediately to
\begin{align}
P_{\alpha\to\beta}(E,L)
& \simeq \sum_j|U^{\alpha j}|^2|U^{\beta j}|^2 + 2\sum_{j<k} \mbox{Re} \bigg[ U^{\alpha j} U^{\beta j *} U^{\alpha k *} U^{\beta k} \exp\bigg\{ -i\frac{\Delta m_{jk}^2}{2E}L\bigg\}\bigg]\notag\\
& = \delta_{\alpha\beta} + 2\sum_{j<k} \mbox{Im} \Big[ U^{\alpha j} U^{\beta j *} U^{\alpha k *} U^{\beta k}\Big]\sin\bigg(\frac{\Delta m_{jk}^2}{2E}L\bigg)\notag\\
&\quad - 4\sum_{j<k} \mbox{Re} \Big[ U^{\alpha j} U^{\beta j *} U^{\alpha k *} U^{\beta k}\Big]\sin^2\bigg(\frac{\Delta m_{jk}^2}{4E}L\bigg)~,\label{eqn:NOF}
\end{align}
where $\Delta m_{jk}^2 \equiv m_j^2 - m_k^2$. This is precisely the Pontecorvo neutrino oscillation formula. Hence we have derived the neutrino oscillation formula in a purely quantum field theoretic formalism, involving just the structure of the spatial two-point function. Comparing Eqs. (\ref{eqn:PROZ}) and (\ref{eqn:SPROZ}), it is straightforward to generalize this result to just the stable particle regime $E > m_j$, $\Gamma_j \to 0^+$ via the replacement in Eq. (\ref{eqn:NOF})
\begin{equation}
-\frac{\Delta m_{jk}^2}{2E} \rightarrow \sqrt{E^2 -m_j^2} - \sqrt{E^2 - m_k^2}~.
\end{equation}
\subsection{(Deep) Virtual Regime}
Having verified that our exacts results reduce to the expected results for both the neutral meson and neutrino systems, let us now exploit the generality of Eqs. (\ref{eqn:EPR}), (\ref{eqn:DOZ}) and (\ref{eqn:GOPF}) to push $E$, $m_j$ and $\Gamma_j$ into non-standard, though physically relevant, regimes of parameter space. So far we have only considered regimes for which $E>m_j$, so let us now consider the case
\begin{equation}
-R_j \gg A_j~,~~\mbox{i.e.}~~m_j^2 - E^2 \gg m_j\Gamma_j~,
\end{equation}
which we call the virtual regime for reasons outlined below. If also $E \ll m_j$, then we call this the deep virtual regime. As we will investigate in Sec. \ref{sec:MR}
below, the virtual regime is particularly interesting if one 1PI state is very heavy compared to another.
In the virtual regime, the wavenumber and characteristic inverse decay lengths become to leading order in $A_j/|R_j|$
\begin{equation}
\label{eqn:DTROZ}
\omega_j \simeq \frac{m_j\Gamma_j}{2\sqrt{m_j^2 - E^2}}~,\qquad \zeta_j \simeq \sqrt{m_j^2 - E^2}~.
\end{equation}
Again, the oscillation probabilities (\ref{eqn:GOPF}) and (\ref{eqn:GIOPF}) follow immediately from this and Eqs. (\ref{eqn:DDOZ}). Within this regime, two flavors with a sufficiently small mass splitting have integrated oscillation probability described by the parameters
\begin{equation}
x \simeq \frac{\Delta \Gamma}{2\bar{\Gamma}}~,\qquad y \simeq \frac{\Delta m}{\bar{\Gamma}}~.
\end{equation}
These are, of course, just a swap of the usual parameters one sees in the particle regime.
In the virtual regime Eq. (\ref{eqn:EPR}) becomes
\begin{equation}
\label{eqn:EXP}
\Delta_j(E,L) \simeq \frac{i}{4\pi L}\exp\Bigg\{i\frac{m_j\Gamma_j}{2\sqrt{m_j^2-E^2}}L - \sqrt{m_j^2-E^2}L\Bigg\}~.
\end{equation}
The spatial two-point function no longer looks like that of a one-particle state. This is especially clear in the stable virtual case $m_j> E$ and $\Gamma_j \to 0 ^+$, for which $\Delta_j(E,L)$ is just an exponential decay: This was noticed previously in Ref. \cite{Okun:1982sr}. Note also that for the unstable case, the wavenumber is determined by the decay rate, rather than by a momentum $(E^2 -m_j^2)^{1/2}$, and vice versa for the characteristic inverse decay length.
Let us briefly comment on the physical interpretations of the virtual regime results, from which we derive its name. As explained in Sec. \ref{sec:EA}, the spatial two-point function $\Delta_j(E,L)$ does not encode the propagation of just a single particle with a definite momentum. Rather, as suggested by Eq. (\ref{eqn:1PIES}), we may think of the spatial two-point function as the continuous sum of a set of propagators, each corresponding to the propagation of a momentum eigenstate. The condition $E < m_j$ then implies all these momentum eigenstates must be off-shell, so in this case $\Delta_j(E,L)$ includes no on-shell propagating particles. That is they are virtual particles, whence the regime name. Alternatively, $E<m_j$ is analogous to the usual quantum mechanical tunnelling condition, with the mass acting as the potential barrier. From either point of view, we emphasize that we should expect $\Delta(E,L)$ to be exponentially suppressed in the stable case, precisely as we see in Eq. (\ref{eqn:EXP}).
\subsection{Mixed Regime}
\label{sec:MR}
It is interesting to consider the case that different 1PI states occupy different regimes. For example, we could consider a two-flavor oscillation in the case that one 1PI state is in the particle regime, while the other is in the virtual regime. We call such a case the mixed regime. This scenario doesn't occur for the neutral meson or neutrino systems because the mass splitting between 1PI states is very small compared to $E$. However, the large mass hierarchy of the quark sector combined with the possibility of quark oscillations \cite{Grossman:2011,Pilaftsis:1997}, provides a natural setting in which we may contemplate a mixed regime oscillation.
For concreteness, let us suppose there is a fourth quark doublet $(t',b')$, with masses much larger than the top quark mass. The existence of a fourth quark family is strongly constrained by the electroweak weak precision measurements \cite{Hong:2001,Kribs:2007,Hung:2008}, but nonetheless phenomenologically still perfectly viable (see Ref. \cite{PDG:2010} for mass bounds). A quark oscillation experiment could then involve a top quark decaying to a final state $t \to X_\beta$ via intermediate $b$ or $b'$ down-type quarks. The generic diagrammatic form of such an experiment is shown in Fig. \ref{fig:QE}.
\begin{figure}[t]
\begin{picture}(261,128) (0,0)
\SetWidth{0}
\SetColor{Black}
\Line[dash,dashsize=4,arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](172,95)(217,109)
\Line[dash,dashsize=4,arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](172,86)(215,73)
\Line[dash,dashsize=4,arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](64,82)(222,28)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](68,86)(162,91)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](1,113)(49,89)
\GOval(221,64)(63,14)(0){0.882}
\GOval(59,86)(12,12)(0){0.882}
\GOval(167,90)(12,12)(0){0.882}
\Text(29,104)[lb]{\Black{$t$}}
\Text(57,83)[lb]{\Black{S}}
\Text(164,87)[lb]{\Black{D}}
\Text(217,58)[lb]{\Black{$X_\beta$}}
\Text(100,95)[lb]{\Black{$b,b'$}}
\end{picture}
\caption{Quark oscillation experiment $t \to X_\beta$.}
\label{fig:QE}
\end{figure}
Let us adopt the following notation. The flavor of the down-type quarks is determined by their up-type partner, so we denote the down-type flavor quarks by $b_t$ and $b_{t^\prime}$. That is, $\alpha = t,t'$. Correspondingly the 1PI states are denoted $b$ and $b^\prime$, so $j=b,b'$. The idea here is that the top quark produces the flavor quark $b_t$ at the source vertex, S, while the generic final state $X_\beta$ in the detector can tag the flavor at vertex D. In order to describe the physics of this experiment using our formalism, and for simplicity, we also assume the following:
\begin{enumerate}[label=\roman{enumi})]
\item The amplitude of the experiment is described by Eq. (\ref{eqn:DEA}).
\item We neglect the presence of the other two down-type quarks $d$ and $s$, and consider an effective two-flavor mixing between the third and fourth quark generations. Consequently, the final state $X_\beta$ only measures flavors $\beta = t,t'$.
\item The $b$ is in the stable particle regime ($E > m_b$ and $\Gamma_b \to 0^+$), while $b'$ in the stable virtual regime ($E<m_{b'}$ and $\Gamma_{b'} \to 0^+$).
\item The energy, $E$, exchanged between S and D can be precisely measured.
\item The 2$\times$2 mixing matrix $U$, which diagonalizes the 1PI function, is unitary.
\end{enumerate}
The extent to which these assumptions are applicable to an actual quark oscillation experiment is questionable. Our intent is merely to demonstrate that with such assumptions, we can perhaps gain insight into the physics of quark oscillations by use of our formalism.
With these assumptions, we have wavenumber and characteristic inverse decay lengths
\begin{equation}
\omega_b = \sqrt{E^2 - m_b^2}~,\qquad \omega_{b'} = 0~,\qquad \zeta_b =0~,\qquad \zeta _{b'} = \sqrt{m_{b'}^2 - E^2}~,
\end{equation}
so that the oscillation wavenumbers
\begin{equation}
\Delta\omega_{bb'} = \sqrt{E^2 - m_b^2}~,\qquad\Delta\zeta_{b'b} = \sqrt{m_{b'}^2 - E^2}~.
\end{equation}
The oscillation probabilities follow immediately from Eqs. (\ref{eqn:ODPE}) and (\ref{eqn:TFIOP}), while the corresponding spatial two-point functions
\begin{align}
\Delta_b(E,L) & = \frac{i}{4\pi L}\exp\bigg\{i\sqrt{E^2-m_b^2}L\bigg\}~,\notag\\
\Delta_{b'}(E,L) & = \frac{i}{4\pi L}\exp\bigg\{-\sqrt{m^2_{b'} - E^2}L\bigg\}~.
\end{align}
In particular, note that the $b'$ two-point function is exponentially suppressed, as we expect for a virtual particle. Further, the integrated probability has parameters
\begin{equation}
x = \sqrt{\frac{E^2 - m_b^2}{m_{b'}^2 - E^2}}~,\qquad y = 1~,
\end{equation}
so that in the two-flavor integrated oscillation probability (\ref{eqn:TFIOP}) there is no longer any interference term --- and hence no oscillation --- between the two mass eigenstates. From Eq. (\ref{eqn:GIOPF}) one finds integrated oscillation probability
\begin{equation}
P_{t\to\beta}^{\textrm{I}}(E) = 2 \frac{|U^{t b}|^2|U^{\beta b}|^2}{1 + |U^{tb}|^2 - |U^{t b'}|^2}~.
\end{equation}
(Note that convergence of the integrated amplitude with $\zeta_b = 0$ is ensured by the usual $i\epsilon$ term, which is taken to zero after integration. Equivalently, since $m_b\Gamma_b$ acts as the $\epsilon$ throughout this paper, the stable limit is determined by taking $\Gamma_b \to 0^+$ after integration over $L$.) As expected, the probability is controlled purely by the mixing of the flavor states with the particle-like 1PI state $b$ which is in the stable particle regime.
\subsection{Threshold Regime}
One last regime of interest, which to the knowledge of the authors has not been previously discussed, is the case
\begin{equation}
|R_j| \ll A_j~,~~\mbox{i.e.}~~|E^2 -m_j^2| \ll m_j\Gamma_j~.
\end{equation}
We call this the threshold regime, since $E \simeq m_j$. To zeroth order in $|R_j|/A_j$, the wavenumber and characteristic inverse lengths reduce to
\begin{equation}
\omega_j = \zeta_j \simeq \sqrt{\frac{m_j\Gamma_j}{2}}~,
\end{equation}
and the oscillation probabilities follow as usual.
This time (\ref{eqn:EPR}) becomes
\begin{equation}
\Delta_{j}(E,L) \simeq \frac{i}{4\pi L}\exp\Bigg\{ (i-1)\sqrt{\frac{m_j\Gamma_j}{2}}L\Bigg\}~.
\end{equation}
Here, curiously, the inverse decay length and wavenumber both depend on the geometric mean of the decay rate and mass, and coincide. We are unaware of an intuitive physical reason why they should coincide at threshold. There is, however, a limited particle analog to this behavior. If we were to interpret $\omega_j$ as the momentum, as we did in the particle regime, then we would have
\begin{equation}
p_j^2\simeq m_j^2-m_j\Gamma_j/2~.
\end{equation}
That is, the 1PI states can be thought of as virtual particles slightly perturbed from the mass shell if $\Gamma_j \ll m_j$.
In the threshold regime, a small mass splitting for two flavors results in
\begin{equation}
x = y \simeq \frac{\sqrt{\Gamma_1} - \sqrt{\Gamma_2}}{\sqrt{\Gamma_1} + \sqrt{\Gamma_2}}
\end{equation}
while if also the decay rates have a small splitting, $\Gamma_{1,2} = \Gamma \mp \Delta \Gamma$, $\Delta \Gamma \ll \Gamma$, then
\begin{equation}
x = y \simeq -\Delta m/m - \Delta \Gamma/\Gamma \ll 1~.
\end{equation}
A well-motivated example of oscillation in which the threshold regime is applicable to both 1PI states is unknown to the authors. Despite this, we do wish to emphasize that the generality of Eqs. (\ref{eqn:GOPF}) and (\ref{eqn:GIOPF}) permits exploration of parameter regimes in which a quantum mechanical treatment might be unfeasible.
\section{Conclusions}
In this paper we have used only the structure of the spatial two-point function $\Delta_{\alpha\beta}(E,L)$ to derive general flavor oscillation probability formulae for unstable fields. We have not only shown that this structure reproduces the usual Pontecorvo neutrino oscillation formulae and time-integrated (CP violating) neutral meson mixing formulae, but we have also found generalized exact expressions with natural physical interpretations in several different parameter regimes. Our results for the stable particle and stable virtual regimes agree with the results of Ref. \cite{Okun:1982sr} for stable fermions. However, our exact oscillation probabilities for unstable fields and the analysis of the unstable particle, threshold and virtual regimes has not been previously presented.
The advantages of the formalism we have employed in this paper are several. The exact computability and integrability of $\Delta_{\alpha\beta}(E,L)$ permitted us to obtain exact, elegant probability oscillation formulae. Moreover, the choice of reference frame throughout this paper is the unambiguous laboratory frame: There is no need in our approach to contemplate mass eigenstate rest frames and proper times. To the extent that complicating effects such as coherence, finite detector and source size, non-trivial source and detector physics, and measurement uncertainty can be neglected, our results provide an instructive leading order description of the physics of flavor oscillation, that is valid over the entire $E,m,\Gamma$ parameter space.
In terms of future work, keeping in mind the large existing Literature on this subject, perhaps the most interesting avenue left to explore is the analogous formalism for flavor oscillation in matter, that is, the Mikheyev-Smirnov-Wolfenstein effect.
\acknowledgments
The authors thank Yuval Grossman, Jo\~ ao P. Silva and Philip Tanedo for helpful discussions. This work is supported by NSF grant number
PHY-0757868.
|
\section{Introduction}
This paper gives a new interpretation of the virtual braid group in terms of a tensor category $SC$ with generating morphisms $\mu_{ij}$ where this symbol denotes an abstract connecting string between
strands $i$ and $j$ in a diagram that otherwise is an identity braid on $n$ strands. These $\mu_{ij}$
satisfy the algebraic Yang-Baxter equation and they generate, in this interpretation, the pure virtual
braid group. The other generating morphisms of this category are elements $v_{i}$ that are
depicted as virtual crossings between strings $i$ and $i +1.$ The generators $v_{i}$ have all the relations for transpositions generating the symmetric group. An $n$-strand diagram that is a product of
these generators is regarded as a morphism from $[n]$ to $[n]$ where the symbol $[n]$ is regarded as
an ordered row of $n$ points that constitute the top or the bottom of a diagram involving $n$ strands.
The virtual braid group on $n$ strands is isomorphic to the group of morphisms in the String Category $SC$ from $[n]$ to $[n].$ Given that one studies the algebraic Yang-Baxter equation, it is natural to study the compositions of algebraic braiding operators placed in two out of the $n$ tensor lines and to let the permutation group of the tensor lines act on this algebra as the group generated by the virtual crossings.
This construction is in sharp contrast to the role of the virtual crossings in the original form of the virtual knot theory.
\smallbreak
Figure~\ref{musigma} illustrates most of the issues.
At the top of the figure we have illustrated the pure virtual braid $ \mu = \sigma v$ on two strands.
The permutation associated with $\mu$ is the identity, as each strand returns to its original position.
The braiding element $\sigma$ has been composed with the virtual crossing $v$, which acts as a permutation of the two strands. With these conventions in place we find that
$\mu$ satisfies the {\it algebraic Yang-Baxter equation}
$$\mu_{12} \mu_{13} \mu_{23} = \mu_{23} \mu_{13} \mu_{12}$$ and this is equivalent to the statement that $\sigma$ satisfies the {\it braiding relation} $$\sigma_{1}\sigma_{2}\sigma_{1} = \sigma_{2}\sigma_{1}\sigma_{2}.$$ This relationship is well-known and it is fundamental to the construction of representations of the Artin braid group and to the construction of quantum link invariants (see
\cite{Ohtsuki} for an account of these matters). In the present paper we will detail this relationship once again, and we shall see that it leads to alternative ways to understand the concept of virtual braiding and to generalizations of the formulation of quantum invariants of knots and links to quantum invariants of virtual knots and links (taken up to rotational equivalence described below).
\smallbreak
Here a notational issue
leads to a mathematical concept. View Figure~\ref{musigma} and notice how we have diagrammed
the algebraic Yang-Baxter relation. An element $\mu_{ij}$ is shown as a graphical connection between
vertical lines labeled $i$ and $j$ respectively. The vertical lines represent different factors in a tensor product in the usual interpretation where $\mu \in \mathcal{A} \otimes \mathcal{A}$ where $ \mathcal{A}$ is an algebra that carries a solution to the algebraic Yang-Baxter equation. We call the graphical edge representing $\mu_{ij}$ a {\it string connection} between the strands $i$ and $j.$
{\it The string connection is a topological model for a logical connection in the mathematics.}
The string going from vertical line $1$ to vertical line $3$ represents $\mu_{13},$ and it has nothing to do with strand $2$ except as in the plane the strand $2$ happens to come between strands
$1$ and $3.$ This means that in our diagram the graphical edge for $\mu_{13}$ intersects the vertical strand $2.$ This intersection is {\it virtual} in the sense that it is just an artifact of the planar drawing.
There is no conceptual connection between $\mu_{13}$ and strand $2$.
\smallbreak
We see that virtuality in the sense of artifactual coinicidence of topological entities will be a necessity in depicting logical connection as topological connection. For this reason, the string diagrammatics that we have adopted for the
algebraic Yang-Baxter equation can be taken as a starting point for the development of the virtual braid group. In this paper, we have started with the usual virtual braid group and reformulated it in this algebraic context.
The attentive reader will see that one could start with the formalism of the algebraic Yang-Baxter equation, construct the appropriate categories and first arrive at the pure virtual braid group and then at the virtual braid group. All of these constructions come from the concept of making topological models for logical connections in mathematical structures.
\smallbreak
The Artin braid group $B_{n}$ is motivated by a combination of topological considerations and the desire for a group structure that is very close to the structure of the symmetric group $S_{n}.$
The virtual braid group $VB_{n}$ is motivated at first by a natural extension of
the Artin braid group in the context of virtual knot theory. The virtual crossings appear as artifacts of the presentation of virtual knots in the plane where those knots acquire extra crossings that are not really part of the essential structure of the virtual knot. We add virtual crossings to
the Artin braid group and follow the principles of virtual knot theory for handling them. These virtual crossings appear crucially in the virtual braid group, and turn into the generators of the symmetric group embedded in the virtual braid group. Thus we arrive at the action of the symmetric group in either case,
but with different motivations. Seen from the categorical view, the virtual crossings are interpreted as generators of the symmetric group whose action is added to the algebraic structure of the pure virtual braid group, and they become part of the embedded symmetry of the structure of the virtual braid group. The pure virtual braid group is seen to be a natural monoidal category generated by formal elements satisfying the algebraic Yang-Baxter equation. The virtual braid group is then an extension of the pure virtual braid group by the symmetric group.
It has nothing to do with the plane and nothing to do with virtual crossings. It is a natural group associated with the structure of algebraic braiding. This is our motivation for constructing the category $SC.$
\smallbreak
Here is a quick technical description of our category.
We define a strict monoidal category
$SC$ that is
freely generated by one object $*$ and three morphisms
$\mu: * \otimes * \longrightarrow * \otimes *,$ $\mu': * \otimes * \longrightarrow * \otimes *,$
and $v:* \otimes * \longrightarrow * \otimes *.$ This basic structure, subjected to appropriate relations
can be understood via morphisms $\mu_{ij}$ defined in terms of the generating morphisms, where the symbol
$\mu_{ij}$ can is interpreted as a connection between strands $i$ and $j$ in a diagram that otherwise is an identity on $n$ strands. The $\mu_{ij}$ satisfy the algebraic Yang-Baxter equation in the sense that for $i < j< k$,
$\mu_{ij}\mu_{ik}\mu_{jk} = \mu_{jk}\mu_{ik}\mu_{ij}.$
The other basic morphisms of this category are elements $v_{i}$ that can be
depicted as virtual crossings between strings $i$ and $i +1.$ The $v_{i}$ are obtained from $v$ by tensoring with identity morphisms
$* \longrightarrow *.$ The $v_{i}$ generate the symmetric group $S_{n}.$ The $\mu_{ij}$ are obtained from $\mu$ by the action of the symmetric group that is generated by the $v_{i}.$ Composition with an individual $v_{i}$ makes a transposition of indices on the $\mu_{kl},$ generating all of them from the basic $\mu$ and $\mu'.$ An $n$-strand diagram that is a product of
basic morphisms is a morphism from $[n]$ to $[n]$ where the symbol $[n]$ is
an ordered row of $n$ points that constitute the top or the bottom of a diagram involving $n$ strands.
Here $[n] = * \otimes * \cdots * \otimes *$ for
a tensor product of $n$ $*$'s. In Figure~\ref{musigma} we illustrate the diagrammatic interpretation of
$\mu$ and the fundamental relation of $\mu$ and $v$ with an elementary braiding element $\sigma.$
The relation is $\mu = \sigma v.$ The virtual braid $\sigma v$ is pure in the sense that its associated
permutation is the identity.
\smallbreak
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{F0.eps}
\caption{ Algebraic Yang-Baxter Equation and Braiding Relation }
\label{musigma}
\end{center}
\end{figure}
The category we describe is a natural structure for an algebraist interested in exploring formal properties of the algebraic Yang-Baxter equation, and it is directly related to more topological points of view about virtual links and virtual braids. In fact, a closely related category, under differrent motivation, was constructed in \cite{KRH} where the intent was to construct a category that would be naturally associated with a Hopf algebra on the one hand, and would receive topological tangles, knots and links under a functor from the tangle category to the Hopf algebra category. The present category, giving the structure of the virtual braid group, is a subcategory of that category associated with a general Hopf algebra. We explain this relationship in detail in Section 6 of the present paper. See also Remark 10 of \cite{Bellingeri} and references therein for another earlier observation of the relationship of the algebraic Yang-Baxter equation with the pure virtual braid group.
\smallbreak
We now describe exactly the structure of the paper.
We develop our model for the virtual braid group by first recalling, in Section 2, its usual definition motivated by virtual knot theory. We then proceed to reformulate the virtual braid group in terms of the above mentioned generators. By the time we reach Theorem 1, we have reformulated the virtual
braid group in terms of the new generators. We then use this approach to give a presentation of the
pure virtual braid group in Theorem 2.
More precisely, in Section 2 we give a presentation for the virtual braid group in terms of our stringy model. We start by describing the usual presentation of the virtual braid group in terms of classical
braid generators and virtual generators that act as permutations between pairs of adjacent strands in the braid, and relations among them (see Figures~ \ref{genrs} -- \ref{forbiddenpic}). Elementary connecting strings (see Figure~\ref{elstrings}) are defined as elementary pure virtual braids --
products of braid generators and virtual generators as in Figure~\ref{musigma}. We then generalize the notion of connecting
string and show that it has the formal diagrammatic property of being stretched and contracted as shown in Figure~\ref{longstrings}. This property makes the string a topological model for a logical connection as we have advertised earlier in this introduction. With these constructions we then rewrite presentations for the virtual braid group and, in Section 3, show how the connection with strings generates the pure virtual braid group with a set of relations that correspond to the algebraic Yang-Baxter equation. See Theorem~2.
\smallbreak
In Section 4 we construct the String Category discussed in this introduction and we show that
the virtual braid group on $n$ strands is isomorphic to the group of morphisms in the String Category $SC$ from $[n]$ to $[n]$ (see Theorem 3). In Section 5 we detail the relationship with the algebraic Yang-Baxter equation and show how to use solutions of the algebraic Yang-Baxter equation to obtain representations of the pure virtual braid group and virtual braid group.
In Section 6 we discuss a generalization of the virtual
braid group to the virtual tangle category.
We show in this section how our work on the structure of the virtual braid group fits into the structure of the virtual tangle category. The virtual tangle category can be used for obtaining invariants of knots and links via Hopf algebras. The invariants we obtain are invariants of {\em rotational virtual knots and links}
where the term rotational means that we do not allow the use of the first virtual Reidemeister move.
See Figure~\ref{vmoves}.
For the virtual tangle category, the rules for regular isotopy of rotational virtuals are
shown in Figure~\ref{regtang}. This is a most convenient category for working with virtual knots and links, and every quantum link invariant
for classical knots and links extends to an invariant for rotational virtual knots and links. In this section we show how a generalization
of the string connectors defined previously in the paper enables the construction of quantum virtual link
invariants associated with Hopf algebras. The paper ends with two subsections on Hopf algebras. The concept of a quasi-triangular Hopf algebra creates an algebraic context for
solutions to the algebraic Yang-Baxter equation. This algebraic context gives rise to categories and relationships with knot theory and virtual knot theory that connect directly with the contents of our investigation.
\section {A Stringy Presentation for the Virtual Braid Group}
\subsection{The virtual braid group}
Let's begin with a presentation for the virtual braid group.
The set of isotopy classes of virtual braids on $n$ strands forms a group, the {\it virtual braid
group} denoted $VB_n$, that was introduced in \cite{VKT}. The group operation is the usual braid
multiplication (form $bb'$ by attaching the bottom strand ends of $b$ to the top strand ends of $b'$).
$VB_n$ is generated by the usual braid generators
$\sigma_{1},\ldots, \sigma_{n-1}$ and by the virtual generators $v_{1},\ldots, v_{n-1},$
where each virtual crossing $v_{i}$ has the form of the braid generator $\sigma_{i}$ with the
crossing replaced by a virtual crossing. See Figure~\ref{genrs} for illustrations. Recall that in virtual
crossings we do not distinguish between under and over crossing. Thus, $VB_n$ is an extension of
the classical braid group $B_n$ by the symmetric group $S_n$, whereby $v_i$ corresponds to the elementary transposition $(i,i+1)$.
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{F1.eps}
\caption{ The generators of $VB_n$ }
\label{genrs}
\end{center}
\end{figure}
\smallbreak
Among themselves the braid generators satisfy the usual {\it braiding relations}:
$$ \begin{array}{lcccl}
\text{(B1)} & \sigma_{i}\sigma_{i+1}\sigma_{i} & = & \sigma_{i+1}\sigma_{i} \sigma_{i+1}, & \\
\text{(B2)} & \sigma_i \sigma_j & = & \sigma_j \sigma_i, & \mbox{for} \ j\neq i\pm 1. \\
\end{array}$$
Among themselves, the virtual generators are a presentation for the symmetric group $S_{n}$, so
they satisfy the following {\it virtual relations:}
$$ \begin{array}{lcccl}
\text{(S1)} & v_{i}v_{i+1}v_{i} & = & v_{i+1}v_{i} v_{i+1}, & \\
\text{(S2)} & v_i v_j & = & v_j v_i, & \mbox{for} \ j\neq i\pm 1, \\
\text{(S3)} & {v_{i}}^2 & = & 1. & \\
\end{array}$$
The {\it mixed relations} between virtual generators and braiding generators are as follows:
$$ \begin{array}{lcccl}
\text{(M1)} & v_{i} \sigma_{i+1} v_{i} & = & v_{i+1} \sigma_i v_{i+1}, & \\
\text{(M2)} & \sigma_i v_j & = & v_j \sigma_i, & \mbox{for} \ j\neq i\pm 1. \\
\end{array}$$
To summarize, the virtual braid group $VB_{n}$ has the following presentation \cite{VKT}.
\begin{equation}\label{brpresi}
VB_{n} = \left< \begin{array}{ll} \begin{array}{l}
\sigma_1, \ldots ,\sigma_{n-1}, \\
v_1, \ldots, v_{n-1} \\
\end{array} &
\left|
\begin{array}{l} (B1), (B2), \\
(S1), (S2), (S3), \\
(M1), (M2)
\end{array} \right. \end{array} \right>
\end{equation}
It is worth noting at this point that the virtual braid group $VB_n$ does not embed in the classical braid group $B_n$, since the virtual braid group contains torsion elements (the $v_{i}$ have order two) and it is well--known that $B_n$ does not. But the classical braid group embeds in the virtual braid group just as classical knots embed in virtual knots. This fact may be most easily deduced from \cite{KUP}, and can also be seen from \cite{M} and \cite{FRR}. For reference to previous work on virtual knots and braids the reader should consult
\cite{CS1,DK,GPV,HR,HRK,VKT,SVKT,DVK,KADOKAMI,Kamada,KiSa,KUP,M,Satoh,TURAEV,V1,V2,KL3,KL4} and references therein. For work on welded braids and welded knots, see \cite{FRR,Kamada,KL3,KL4}.
For Markov--type theorems for virtual braids (and welded braids), giving sets of moves on virtual
braids that generate the same equivalence classes as the oriented virtual link types of their closures, see \cite{Kamada} and \cite{KL4}. Such
theorems are important for understanding the structure and classification of virtual knots and
links.
\smallbreak
The second mixed relation in the presentation of the virtual braid group will be called the {\it local detour move} and it is illustrated in Figure~\ref{locdetour}. The following relations are also local detour moves for virtual
braids and they are easy consequences of the above.
\begin{equation}\label{}
\begin{array}{ccc}
v_{i}v_{i+1} {\sigma_i}^{\pm 1} & = & {\sigma_{i+1}}^{\pm 1}v_{i}v_{i+1}, \\
{\sigma_{i}}^{\pm 1} v_{i+1}v_{i} & = & v_{i+1}v_{i} {\sigma_{i+1}}^{\pm 1}. \\
\end{array}
\end{equation}
This set of relations taken together define the basic local isotopies for virtual braids. Each relation is a braided version of a local virtual link isotopy. The local detour move is written equivalently:
\begin{equation}\label{sigmadet}
\sigma_{i+1} = v_i v_{i+1} \sigma_i v_{i+1} v_i.
\end{equation}
Notice that Eq.~\ref{sigmadet} is the braid detour move of the $i$th strand around
the crossing between the $(i+1)$-st and the
$(i+2)$-nd strand (see first two illustrations in Figure~\ref{detsigma}) and it provides an inductive way of expressing
all braiding generators in terms of the first braiding generator $\sigma_{1}$ and the virtual
generators $v_1,\ldots,v_{n-1}$ (see first and last illustrations in Figure~\ref{detsigma}), that is:
\begin{equation}\label{}
\sigma_{j} = (v_{j-1}\ldots v_2v_1)\,(v_{j}\ldots v_3v_2)\,\sigma_1\, (v_2v_3\ldots
v_{j})\,(v_1v_2\ldots v_{j-1}).
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[width=3.5cm]{F2.eps}
\caption{ The local detour}
\label{locdetour}
\end{center}
\end{figure}
\smallbreak
\begin{figure}
\begin{center}
\includegraphics[width=11cm]{F3.eps}
\caption{ Detouring the crossing $\sigma_{i+1}$}
\label{detsigma}
\end{center}
\end{figure}
\noindent In \cite{KL3} we derive the following reduced presentation for
$VB_{n}$:
\begin{equation}\label{brreduced}
VB_{n} = \left< \begin{array}{ll}
\begin{array}{c}
\sigma_1, \\
v_1, \ldots ,v_{n-1} \\
\end{array} &
\left|
\begin{array}{l}(S1), (S2), (S3) \\
\sigma_1 v_j=v_j \sigma_1, \ \ \ \mbox{for} \ j>2 \\
v_1 \sigma_1 v_1\,v_2 \sigma_1 v_2\, v_1 \sigma_1 v_1 =
v_2 \sigma_1 v_2\, v_1 \sigma_1 v_1 \, v_2 \sigma_1 v_2 \\
\sigma_1\, v_2 v_3 v_1 v_2 \sigma_1 v_2 v_1 v_3 v_2 =
v_2 v_3 v_1 v_2 \sigma_1 v_2 v_1 v_3 v_2 \, \sigma_1 \\
\end{array} \right. \end{array} \right>
\end{equation}
The local detour move gives rise to a generalized {\it detour move}, by which any box in the braid can be detoured to any position in the braid, see Figure~\ref{boxdet}.
\smallbreak
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{F4.eps}
\caption{ Detouring a box}
\label{boxdet}
\end{center}
\end{figure}
Finally, it is worth recalling that in virtual knot theory there are ``forbidden moves" involving two real crossings and one virtual. More precisely, there are two types of
forbidden moves: One with an over arc, denoted $F_1$ and another with an under arc, denoted $F_2$. See \cite{VKT} for explanations and interpretations. Variants of the forbidden moves are illustrated in Figure~\ref{forbiddenpic}. So, relations of the types:
\begin{equation}\label{forbidden}
\sigma_{i} v_{i+1} \sigma_{i}^{-1} = \sigma_{i+1}^{-1} v_{i} \sigma_{i+1} \quad (F1) \qquad \mbox{and} \qquad \sigma_{i}^{-1} v_{i+1} \sigma_{i} = \sigma_{i+1} v_{i} \sigma_{i+1}^{-1} \quad (F2)
\end{equation}
are not valid in virtual knot theory.
\begin{figure}
\begin{center}
\includegraphics[width=10cm]{F5.eps}
\caption{The forbidden moves}
\label{forbiddenpic}
\end{center}
\end{figure}
\subsection{}
We now wish to describe a new set of generators and relations for the virtual braid group that makes it particularly easy to describe the pure virtual braid group, $VP_{n}$. In order to accomplish
this aim, we introduce the following elements of $VP_{n}$, for $i=1,\ldots,n-1$.
\begin{equation}\label{mui}
\mu_{i,i+1} := \sigma_{i} v_{i}
\end{equation}
We indicate $\mu_{i,i+1}$ by a connecting string between the $i$-th and $(i+1)$-st strands with a dark vertex on the $i$-th strand, a dark vertex on the $(i+1)$-st strand, and an arrow from left to right. View Figure~\ref{elstrings}. The inverses $\mu_{i,i+1}^{-1} = v_{i}\sigma_{i}^{-1} $
have same directional arrows but are indicated by using white vertices. By detouring it to the leftmost position of the braid, we can write $\mu_{i,i+1}$ in terms of $\mu_{12}$ conjugated by a virtual word:
\begin{equation}\label{muimu12}
\mu_{i,i+1} = (v_{i-1}\ldots v_2v_1)(v_{i}\ldots v_3v_2) \mu_{12} (v_2 v_3 \ldots v_{i}) (v_1 v_2\ldots v_{i-1}).
\end{equation}
We also introduce the elements
\begin{equation}\label{muibar}
\mu_{i+1,i} := v_{i} \sigma_{i} = v_{i} \mu_{i,i+1}v_{i}
\end{equation}
We indicate $\mu_{i+1,i}$ by a connecting string between the $i$-th and $(i+1)$-st strands, with a dark vertex
on the $i$-th strand, a dark vertex on the $(i+1)$-st strand, and an arrow from right to left (reversing the direction from $\mu_{i,i+1}$), view Figure~\ref{elstrings}. An illustration of Eq.~\ref{muibar} (see top of Figure~\ref{relstrings}) explains the reversing of the direction of the arrow in the graphical interpretation of $\mu_{i+1,i}$. The inverses $\mu_{i+1,i}^{-1} = \sigma_{i}^{-1} v_{i}$ have same directional arrows but are indicated by using white vertices. An analogous equation to Eq.~\ref{muimu12} holds:
\begin{equation}\label{muimu12bar}
\mu_{i+1,i} = (v_{i-1}\ldots v_2v_1)(v_{i}\ldots v_3v_2) \mu_{21} (v_2 v_3 \ldots v_{i}) (v_1 v_2\ldots v_{i-1})
\end{equation}
\begin{defn} \rm
The pure virtual braids $\mu_{i,i+1}, \mu_{i+1,i}$ and their inverses shall be called {\it elementary connecting strings}.
\end{defn}
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{F6.eps}
\caption{ The elementary connecting strings $\mu_{i,i+1}$, $\mu_{i+1,i}$ and their inverses}
\label{elstrings}
\end{center}
\end{figure}
\smallbreak
From Eqs.~\ref{mui} and \ref{muibar} follow directly the relations:
\begin{equation}\label{muimuibar}
v_i \mu_{i+1,i} = \mu_{i,i+1} v_i \ \ \mbox{and} \ \ \mu_{i+1,i}^{-1} v_i = v_i \mu_{i,i+1}^{-1},
\end{equation}
also illustrated in Figure~\ref{relstrings}.
\begin{figure}
\begin{center}
\includegraphics[width=5.2cm]{F7.eps}
\caption{ Relations between the elementary connecting strings}
\label{relstrings}
\end{center}
\end{figure}
\smallbreak
Further, we generalize the notion of a connecting string and define, for $i<j$, the element $\mu_{ij}$ (a connecting string from strand $i$ to strand $j$) by the formula
\begin{equation}\label{muij}
\mu_{ij} := v_{j-1} v_{j-2} \ldots v_{i+1} \, \mu_{i,i+1} \, v_{i+1} \ldots v_{j-2} v_{j-1}.
\end{equation}
In a diagram $\mu_{ij}$ is denoted by a connecting string from strand $i$ to strand $j$, with dark vertices on these two strands and an arrow pointing from left to right, view Figure~\ref{longstrings}.
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{F8.eps}
\caption{ Connecting strings}
\label{longstrings}
\end{center}
\end{figure}
\smallbreak
We also generalize, for $i<j$, the elements $\mu_{i+1,i}$ to the elements:
\begin{equation}\label{muijbar}
\mu_{ji} := t_{ij}\, \mu_{ij}\, t_{ij}
\end{equation}
where $t_{ij} = v_{i} v_{i+1} \ldots v_{j} \ldots v_{i+1} v_{i}$ is the element of $S_{n}$ (generated by the $v_{i}$'s) that interchanges strands $i$ and $j$, leaving all other strands fixed.
We denote $\mu_{ji}$ by a connecting string from strand $i$ to strand $j$, with dark vertices, and an arrow pointing from right to left. Figure~\ref{exchange} illustrates the example $\mu_{31} = v_{2}v_{1}v_{2}\mu_{13}v_{2}v_{1}v_{2}$. It is easily verified that
\begin{equation}\label{muijbar2}
\mu_{ji} = v_{j-1} v_{j-2} \ldots v_{i+1} \, \mu_{i+1,i} \, v_{i+1} \ldots v_{j-2} v_{j-1}
\end{equation}
\smallbreak
\begin{figure}
\begin{center}
\includegraphics[width=4.5cm]{F9.eps}
\caption{ The exchange of labels between $\mu_{ij}$ and $\mu_{ji}$}
\label{exchange}
\end{center}
\end{figure}
\noindent The inverses of the elements $\mu_{ij}$ and $\mu_{ji}$ have same directional arrows respectively, but white dotted vertices.
\begin{defn}\rm
The elements $\mu_{ij}, \ \mu_{ji}$ and their inverses shall be called {\it connecting strings}.
\end{defn}
With the above conventions we can speak of connecting strings $\mu_{rs}$ for any $r,s$. It is important to have the elements $\mu_{ji}$ when $j > i$, but in the algebra they are all defined in terms of the $\mu_{ij}$. The importance of having the elements $\mu_{ji}$ will become clear when we restrict to the pure virtual braid group.
\begin{rem} \rm In the definition of $\mu_{ij}$ if we consider $\mu_{i,i+1}$ as a virtual box inside the virtual braid we can use the (generalized) detour moves to bring it to any position, as Figure~\ref{longstrings} illustrates. This means that the contraction of $\mu_{ij}$ to $\mu_{i,i+1}$ may be pulled anywhere between the $i$-th and the $j$-th strands. By the same reasoning the contraction of $\mu_{ji}$ to $\mu_{i+1,i}$ may be also pulled anywhere between the $i$-th and the $j$-th strands.
\end{rem}
\subsection{}
We shall next give some relations satisfied by the connecting strings. Before that we need the following remark.
\begin{rem}\label{action} \rm
The symmetric group $S_n$ clearly acts on $VB_n$ by conjugation. By their definition (Eqs. \ref{mui}, \ref{muibar}, \ref{muij}, \ref{muijbar}, \ref{muijbar2}), this action on connecting strings translates into permuting their indices, that is, a permutation $\tau \in S_n$ acting on $\mu_{rs}$ will change it to $\mu_{\tau (r),\tau (s)}$. This means that $S_n$ acts by conjugation also on the subgroup of $VB_n$ generated by the $\mu_{ij}$'s. Moreover, by Eqs.~\ref{muimu12}, \ref{muibar}, all connecting strings may be obtained by the action of $S_n$ on $\mu_{12}$. For $\sigma \in S_{n}$ we regard $\sigma$ both as a product of the elements $v_{i}$ and as a permutation of the set $\{1,2,3, \ldots, n \}.$
Further, any relation in $VB_n$ transforms into a valid relation after acting on it an element of $S_n$. In particular, a commuting relation between connecting strings will be transformed to a new commuting relation between connecting strings.
\end{rem}
\begin{lem}\label{stringyslides}
The following relations hold in $VB_n$ for all $i$.
\begin{enumerate}
\item $v_{i} \mu_{i,i+1} = \mu_{i+1,i} v_{i}$ \ , \ \ $v_{i} \mu_{i+1,i} = \mu_{i,i+1} v_{i}$
\item $v_{i+1} \mu_{i,i+1} = \mu_{i,i+2} v_{i+1}$ \ , \ \ $
v_{i+1} \mu_{i+1,i} = \mu_{i+2,i} v_{i+1}$
\item $v_{i-1} \mu_{i,i+1} = \mu_{i-1,i+1} v_{i-1}$ \ , \ \ $
v_{i-1} \mu_{i+1,i} = \mu_{i+1,i-1} v_{i-1}$
\item $v_{j} \mu_{i,i+1} = \mu_{i,i+1} v_{j}$ \ , \ \ $
v_{j} \mu_{i+1,i} = \mu_{i+1,i} v_{j}, \ \ j\neq i-1,\, i,\, i+1$.\\
\end{enumerate}
The above local relations generalize to similar ones involving different indices. Relations 1 are generalized by Eq.~\ref{muijbar}, reflecting the mutual reversing of $\mu_{ij}$ and $\mu_{ji}$, recall Figures~\ref{relstrings} and \ref{exchange}. Relations 2 and 3 are the {\rm local slide moves}, as illustrated in Figure~\ref{slides}, and they generalize to the {\rm slide moves} coming from the defining equations: $\mu_{i+1,k}=v_i \mu_{ik}v_i$ for any $k<i$ or $k>i+1$. Relations 4 and their generalizations: $v_{j} \mu_{ik} = \mu_{ik} v_{j}$ for any $k\neq i$ and $j\neq i-1,\, i,\, k-1,\, k$, are all commuting relations. All these relations result from the action of any $\tau \in S_n$ on $\mu_{12}$:
\begin{equation}\label{mixed}
\text{{\boldmath $\tau^{-1} \, \mu_{12} \, \tau =\mu_{\tau (1),\tau (2)}$} }.
\end{equation}
\end{lem}
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{F10.eps}
\caption{ Slide moves}
\label{slides}
\end{center}
\end{figure}
\begin{proof}
All relations 1,2 and 3 follow directly from the definitions of the elements $\mu_{ij}$ and
$\mu_{ji}$. For example, $v_{i+1} \mu_{i,i+1} = \mu_{i,i+2} v_{i+1}$ is equivalent to the defining relation $\mu_{i,i+2} = v_{i+1} \mu_{i,i+1} v_{i+1}$. Figure~\ref{slidepf} illustrates the proof of a local slide move. Relations 4 follow immediately from the commuting relations (S2) and (M2) of $VB_n$. The generalizations of all types of moves follow from the local ones after using detour moves. Finally, the derivation of all relations from the action of $S_n$ on $\mu_{12}$ is explained in Remark~\ref{action} and, more precisely, by the Eqs.~\ref{muimu12}, \ref{muij}, \ref{muimu12bar}, \ref{muijbar2}.
\end{proof}
\begin{figure}
\begin{center}
\includegraphics[width=4.5cm]{F11.eps}
\caption{ Proving a local slide move}
\label{slidepf}
\end{center}
\end{figure}
\begin{lem}\label{stringycom}
The following commuting relations among connecting strings hold in $VB_n$.
\begin{enumerate}
\item $ \text{{\boldmath $\mu_{12} \mu_{34} = \mu_{34} \mu_{12}$}}$
\item $\mu_{14} \mu_{23} = \mu_{23} \mu_{14} \ \ \ \ \ (\text{action by } (324))$
\item $\mu_{13} \mu_{24} = \mu_{24} \mu_{13} \ \ \ \ \ (\text{action by } (23))$
\end{enumerate}
The above local relations generalize to commuting relations of the form:
\begin{equation}\label{com}
\mu_{ij}\mu_{kl} = \mu_{kl}\mu_{ij}, \ \ \{i,j\} \cap \{k,l\} = \emptyset.
\end{equation}
All the above commuting relations result from relation~1 by actions of permutations (indicated for relations~2, 3 to the right of each relation). Moreover, for any choice of four strands there are exactly $24$ such commuting relations that preserve the four strands.
\end{lem}
\begin{proof}
Relation 1 clearly rest on the virtual braid commuting relations (B2) and (M2). We shall show how relation 2 reduces to relation 1. In the proof we underline in each step the generators of $VB_n$ on which virtual braid relations are applied.
\vspace{.1in}
\noindent $\begin{array}{rcl}
\mu_{i,i+3} \mu_{i+1,i+2} & = & v_{i+2} v_{i+1} \mu_{i,i+1} \underline{v_{i+1} v_{i+2} \mu_{i+1,i+2}} \\
\ & \stackrel{detour}{=} & v_{i+2} v_{i+1} \underline{\mu_{i,i+1} \mu_{i+2,i+3}} v_{i+1} v_{i+2} \\
\ & \stackrel{(1)}{=} & \underline{v_{i+2} v_{i+1} \mu_{i+2,i+3}} \mu_{i,i+1} v_{i+1} v_{i+2} \\
\ & \stackrel{detour}{=} & \mu_{i+1,i+2} v_{i+2} v_{i+1} \mu_{i,i+1} v_{i+1} v_{i+2} \\
\ & = & \mu_{i+1,i+2} \mu_{i,i+3}. \\
\end{array}$
\vspace{.1in}
\noindent Figure~\ref{commute} illustrates how relation 3 also reduces to relation 1. Notice now that relations 2 and 3 can be derived from relation 1 by conjugation by the permutations $(324)$ and $(23)$ respectively. Let us see how this works specifically for relation~2: the indices of relation~1 against the indices of relation~2 induce the permutation $(324) = v_2v_3$. This means that conjugating relation~1 by the word $v_2v_3$ will yield relation~2.
Notice also that there are $24$ commuting relations in total involving the strands $1,2,3,4$ and indices in any order. Likewise for any choice of four strands. The derivation of all relations from the action of $S_n$ on relation~1 is clear from Remark~\ref{action}.
\end{proof}
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{F12.eps}
\caption{ A local commuting relation}
\label{commute}
\end{center}
\end{figure}
\begin{lem}\label{stringybr}
The following {\rm stringy braid relations} hold in $VB_n$.
\begin{enumerate}
\item $\text{{\boldmath $\mu_{12} \mu_{13} \mu_{23} = \mu_{23} \mu_{13} \mu_{12}$}}$
\item $\mu_{21} \mu_{23} \mu_{13} = \mu_{13} \mu_{23} \mu_{21} \ \ \ \ \ (\text{action by } (12))$
\item $\mu_{13} \mu_{12} \mu_{32} = \mu_{32} \mu_{12} \mu_{13} \ \ \ \ \ (\text{action by } (23))$
\item $\mu_{32} \mu_{31} \mu_{21} =\mu_{21} \mu_{31} \mu_{32} \ \ \ \ \ (\text{action by } (13))$
\item $\mu_{23} \mu_{21} \mu_{31} = \mu_{31} \mu_{21} \mu_{23} \ \ \ \ \ (\text{action by } (123))$
\item $\mu_{31} \mu_{32} \mu_{12} = \mu_{12} \mu_{32} \mu_{31} \ \ \ \ \ (\text{action by } (132))$
\end{enumerate}
The above relations generalize to three-term relations of the form:
\begin{equation}\label{strbr}
\mu_{ij}\mu_{ik}\mu_{jk} = \mu_{jk}\mu_{ik}\mu_{ij}, \ \ \ \text{for any distinct } i,j,k.
\end{equation}
All six relations stated above result from the action on relation 1 by permutations of $S_n$, which only permute the indices $\{1,2,3 \}$. These permutations are indicated to the right of each relation. Moreover, for any choice of three strands there are exactly six relations analogous to the above, which all result from relation~1 by actions of appropriate permutations that preserve the three strands each time.
\end{lem}
\begin{figure}
\begin{center}
\includegraphics[width=5cm]{F13.eps}
\caption{ The stringy braid relation}
\label{sbr}
\end{center}
\end{figure}
\begin{proof}
Figure~\ref{sbr} illustrates relation 1. Relation~1 rests on the braid relations (B1) of $VB_n$. Indeed, let us prove one relation of this type. See also Figure~\ref{sybpf} for a pictorial proof.
\vspace{.1in}
\noindent $\begin{array}{rcl}
\mu_{i+1,i+2} \mu_{i,i+2} \mu_{i,i+1} & = & (\sigma_{i+1} \underline{v_{i+1}) (v_{i+1}} \sigma_{i} \underline{v_{i} v_{i+1}) (\sigma_{i}} v_{i}) \\
\ & \stackrel{(S3,M1)}{=} & \underline{\sigma_{i+1} \sigma_{i} \sigma_{i+1}} \underline{v_{i} v_{i+1} v_{i}} \\
\ & \stackrel{(B1,S1)}{=} & \sigma_{i} \sigma_{i+1} \underline{\sigma_{i} v_{i+1} v_{i}} v_{i+1} \\
\ & \stackrel{(M1,S3)}{=} & \sigma_{i} \sigma_{i+1} \underline{v_{i+1} v_{i} v_{i+1}} v_{i+1} \sigma_{i+1} v_{i+1} \\
\ & \stackrel{(S1)}{=}& \sigma_{i} \underline{\sigma_{i+1} v_{i} v_{i+1}} v_{i} v_{i+1} \sigma_{i+1} v_{i+1} \\
\ & \stackrel{(M1)}{=} & (\sigma_{i} v_{i}) (v_{i+1} \sigma_{i} v_{i} v_{i+1}) (\sigma_{i+1} v_{i+1}) \\
\ & = & \mu_{i,i+1} \mu_{i,i+2} \mu_{i+1,i+2}. \\
\end{array}$
\vspace{.1in}
The other five stated relations follow from relation~1. Indeed, substituting the $\mu_{ji}$'s from Eqs.~\ref{muibar} and \ref{muijbar}, and drawing the two sides of a relation we notice that there is always a region where, by the slide relations, all three connecting strings become consecutive without any of them having to be reversed, thus enabling application of the first relation. This diagrammatic argument confirms the fact that all six relations are derived from the first one by the action of appropriate elements of $S_n$. Let us see how this works specifically for relation~5: the indices of relation~1 against the indices of relation~5 induce the permutation $(123) = v_2v_1$. This means that conjugating relation~1 by the word $v_2v_1$ will yield relation~5.
Finally, the derivation of all stringy braid relations from the action of $S_n$ on relation~1 is clear from Remark~\ref{action}.
\end{proof}
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{F14.eps}
\caption{ Proof of the stringy braid relation}
\label{sybpf}
\end{center}
\end{figure}
Another remark is now due.
\begin{rem}\rm The forbidden moves of virtual knot theory are naturally forbidden also in the stringy category. For example, the forbidden relations $F1, F2$ of Eq.~\ref{forbidden} translate into the following corresponding {\it forbidden stringy relations} $SF1, SF2$:
\begin{equation}\label{strforbidden}
\mu_{i,i+2} \mu_{i+1,i+2} = \mu_{i+1,i+2} \mu_{i,i+2} \quad (SF1) \qquad \mbox{and} \qquad \mu_{i,i+2} \mu_{i,i+1} = \mu_{i,i+1} \mu_{i,i+2} \quad (SF2)
\end{equation}
which, together with all similar relations arising from conjugating the above by permutations, are not valid in the stringy category. See Figure~\ref{sforbidden} for illustrations.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{F15.eps}
\caption{ Stringy forbidden moves}
\label{sforbidden}
\end{center}
\end{figure}
\end{rem}
\subsection{The stringy presentation}
We will now define an abstract stringy presentation for $VB_n$ that starts from the concept of connecting string and recaptures the virtual braid group.
By Eq.~\ref{mui} we have
\begin{equation}\label{sigmai}
\sigma_{i} = \mu_{i,i+1}v_{i}
\end{equation}
so, the connecting strings $\mu_{ij}$ can be taken as an alternate set of generators of the virtual braid group,
along with the virtual generators $v_{i}$. The relations in this new presentation consist in the results we proved above in Lemmas~\ref{stringyslides}, \ref{stringycom}, \ref{stringybr} describing the
interaction of connecting strings with virtual crossings, the commutation properties of connecting strings, the stringy braiding relations, and the usual relations $(S1), (S2), (S3)$ in the symmetric group $S_{n}$. For the work
below, recall that we have defined the element $t_{ij} = v_{i} v_{i+1} \ldots v_{j} \ldots v_{i+1} v_{i}$
that corresponds to the transposition $(ij)$ in $S_{n}$.
In any presentation of a group $G$ containing the elements $\{ v_1, \ldots, v_{n-1}\}$ and the relations $(S1), (S2), (S3)$ among them,
we have an action of the symmetric group $S_{n}$ on the group $G$ defined by conjugation by an element $\tau$ in $S_{n}$, expressed in terms of the $v_{i}$:
\[
g^{\tau} = \tau g \tau^{-1}
\]
for $g$ in $G$. In particular, we can consider $t_{ij}\, g \,t_{ij}$
as the action by the transposition $t_{ij}$ on an element $g$ of $G$.
We will use this action to define a stringy model of the virtual braid group.
\begin{defn}\rm Let $VS_{n}$ denote the following stringy group presentation.
\begin{equation}\label{stringypres}
VS_{n} = \left< \begin{array}{ll} \begin{array}{l}
\mu_{ij} ,\quad 1 \le i \ne j \le n, \\
v_1, \ldots, v_{n-1} \\
\end{array} &
\left|
\begin{array}{l}
\tau \mu_{ij} \tau^{-1} = \mu_{\tau (i), \tau (j)}, \quad \tau \in S_{n} \\
\mu_{12}\mu_{13}\mu_{23} = \mu_{23}\mu_{13}\mu_{12} \\
\mu_{12}\mu_{34} = \mu_{34}\mu_{12} \\
(S1), (S2), (S3) \\
\end{array} \right. \end{array} \right>
\end{equation}
\end{defn}
We can now state the following theorem.
\begin{thm} The stringy group $VS_{n}$ is isomorphic to the virtual braid group
$VB_{n}.$
\end{thm}
\begin{proof} First we define a homomorphism $F: VB_{n} \longrightarrow VS_{n}$ by
$F(v_{i}) = v_{i}$ and $F(\sigma_{i}) = \mu_{i,i+1} v_{i}$, and extend the map to be a
homomorphism on words in the generators of the virtual braid group. In order to show that this map is well-defined, we must show that it preserves the relations in the virtual braid group.
Since $F(v_{i}) = v_{i}$, the relations among the $v_{i}$ with themselves are preserved
identically. The commuting relations in the braid group are
$\sigma_{i} \sigma_{j} = \sigma_{j} \sigma_{i}$ when $|i-j| > 2$. Thus we must show that
$$
\mu_{i,i+1}v_{i} \mu_{j,j+1}v_{j} = \mu_{j,j+1}v_{j}\mu_{i,i+1}v_{i}.
$$
But this follows immediately from relations~4 of Lemma~\ref{stringyslides} and from Lemma~\ref{stringycom}. The mixed commuting relations $(M2)$ follow also directly from relations $(S2)$ and relations~4 of Lemma~\ref{stringyslides}. This completes the verification that the commuting relations in the virtual braid group are compatible with $F$.
The detour moves $(M2)$ in the virtual braid group go under $F$ to the slide relations of Lemma~\ref{stringyslides}. We illustrate this in Figure~\ref{sdetour}.
It remains to prove that the braiding relations $(B1)$
carry over to $VS_{n}$ under $F$. Indeed:
$$F(\sigma_{i}\sigma_{i+1}\sigma_{i}) = \mu_{i,i+1}v_{i}\mu_{i+1,i+2}v_{i+1}\mu_{i,i+1}v_{i}$$
$$\stackrel{Lemma~\ref{stringyslides}}{=} \mu_{i,i+1}\mu_{i,i+2}\mu_{i+1,i+2}v_{i}v_{i+1}v_{i},$$
while
$$F(\sigma_{i+1}\sigma_{i} \sigma_{i+1}) = \mu_{i+1,i+2}v_{i+1}\mu_{i,i+1}v_{i}\mu_{i+1,i+2}v_{i+1}$$
$$\stackrel{Lemma~\ref{stringyslides}}{=} \mu_{i+1,i+2}\mu_{i,i+2}\mu_{i,i+1}v_{i+1}v_{i}v_{i+1}.$$
and the two expressions are equal from Lemma~\ref{stringybr} and relations~$(S1)$. This completes the proof that the mapping $F$ is a well-defined homomorphism of groups.
We now define an inverse mapping $G:VS_{n} \longrightarrow VB_{n}$ by
$G(v_{i}) = v_{i}$ and $G(\mu_{i, i+1}) = \sigma_{i} v_{i}$. At this stage we have two pieces of work to accomplish: We must extend $G$ to all of $VB_{n}$ and we must show that $G$ is well-defined and that it preserves the relations in the group presentation. This will be done in the next paragraphs.
First of all, we have the $VS_{n}$ relations:
$$ \tau^{-1}\, \mu_{ij} \, \tau = \mu_{\tau (i), \tau (j)}$$
for all $\tau$ in $S_{n}$. In particular, this means that if $\tau (1) = i$ and
$\tau (2) = j$, then
$$\mu_{ij} = \tau^{-1}\, \mu_{12} \, \tau.$$
Thus we can define
$$G(\mu_{ij}) = \tau^{-1} G(\mu_{12}) \tau = \tau^{-1} \sigma_{1} v_{1} \tau.$$
It is easy to see that this is well-defined by noting that if $\lambda$ is another
permutation such that $\lambda(1) = i$ and $\lambda(2) = j$, then $\lambda = \tau \gamma$
where $\gamma$ is a permutation that fixes $1$ and $2$. But such a permutation commutes
with $\sigma_{1} v_{1}$ as is easy to see in the virtual braid group. Hence $\lambda$ can replace
$\tau$ in the formula for $G(\mu_{ij})$ with no change. We leave it as an exercise for the reader
to check that our definition of $G(\mu_{i,i+1})$ in the previous paragraph agrees with the present definition. This completes the definition of the map $G$. We now need to see that it respects the
other relations in $VB_{n}.$
We must show that
$$G(\mu_{12}\mu_{34}) = G(\mu_{34}\mu_{12}).$$
Just note that
$$ G(\mu_{12}\mu_{34}) = \sigma_{1}v_{1} \sigma_{3}v_{3} = \sigma_{3 }v_{3} \sigma_{1}v_{1} = G(\mu_{34}\mu_{12}),$$
by the commuting relations in the virtual braid group.
Finally, we must prove
$$G( \mu_{12}\mu_{13}\mu_{23}) = G(\mu_{23}\mu_{13}\mu_{12}).$$
Note that $\mu_{13} = v_{2}\mu_{12}v_{2}$, so we must prove that in the
virtual braid group,
$$\sigma_{1}v_{1} v_{2} \sigma_{1} v_{1} v_{2} \sigma_{2} v_{2} = \sigma_{2}v_{2} v_{2} \sigma_{1} v_{1} v_{2} \sigma_{2} v_{2}.$$
Figure~\ref{sybpf} illustrates how this identity follows via braiding and detour moves.
We have verified that the mapping $G$ is well-defined and, by definition,
the compositions $F\circ G$ and $G\circ F$ are the identity on $VS_{n}$ and
$VB_{n}$. Therefore $VS_{n}$ and $VB_{n}$ are isomorphic groups. This completes
the proof of the Theorem.
\end{proof}
\begin{figure}
\begin{center}
\includegraphics[width=4.0cm]{F16.eps}
\caption{ The detour moves correspond to the slide moves in the stringy category }
\label{sdetour}
\end{center}
\end{figure}
Finally, we also give below a reduced presentation for $VB_n$, which derives immediately from (\ref{brreduced}).
\begin{prop}
The following is a reduced stringy presentation for $VB_{n}$:
\begin{equation}\label{strreduced}
VB_{n} = \left< \begin{array}{ll}
\begin{array}{c}
\mu_{12}, \\
v_1, \ldots ,v_{n-1} \\
\end{array} &
\left|
\begin{array}{l}
\mu_{12} v_j=v_j \mu_{12}, \ \ \ \mbox{for} \ j>2 \\
\mu_{12} \, v_2 \mu_{12} v_2 \, v_1v_2 \mu_{12} v_2 v_1 =
v_1v_2 \mu_{12} v_2 v_1 \, v_2 \mu_{12} v_2 \, \mu_{12} \\
\mu_{12}\, v_2 v_3 v_1 v_2 \mu_{12} v_2 v_1 v_3 v_2 =
v_2 v_3 v_1 v_2 \mu_{12} v_2 v_1 v_3 v_2 \, \mu_{12} \\
(S1), (S2), (S3) \\
\end{array} \right. \end{array} \right>
\end{equation}
\end{prop}
\noindent The second relation is the stringy braid relation~1 of Lemma~\ref{stringybr} and the third relation is the commuting relation~1 of Lemma~\ref{stringycom}.
\section {The Pure Virtual Braid Group}
\subsection{ A presentation for the pure virtual braid group}
From presentation Eq.~\ref{brpresi} of $VB_{n}$ we have a surjective homomorphism $$\pi : VB_{n} \longrightarrow S_{n}$$ defined by
$$\pi(\sigma_{i}) = \pi(v_{i}) =v_{i}.$$
For a virtual braid $b$, we refer to $\pi(b)$ as the {\it permutation associated with the virtual braid $b$}, and we define the {\it pure virtual braid group} $VP_{n}$ to be the kernel of the homomorphism $\pi$. Hence, $VP_{n}$ is a normal subgroup of $VB_{n}$ of index $n!$. So, $VP_{n} \cdot S_n = VB_{n}$. Moreover, $VP_n \cap S_n = \{ id \}$. Hence, $VB_{n} = VP_{n} \rtimes S_n$. Equivalently, we have the exact sequence
\[ 1 \longrightarrow VP_{n} \longrightarrow VB_{n} \longrightarrow S_n \longrightarrow 1.\]
A presentation for $VP_n$ can be now derived immediately from the stringy presentation of $VB_{n}$ as an application of the Reidemeister-Schreier process \cite{Fox,MKS,Stillwell}. To see this, we first need the following.
\begin{lem} The subgroup $VP_{n}$ of \ $VB_{n}$ is generated by the elements $\mu_{ij}$ for all $i \neq j$.
\end{lem}
\begin{proof} Indeed, by Eqs.~\ref{mui} and \ref{muibar}, $\sigma_i = \mu_{i,i+1}v_{i} = v_{i} \mu_{i+1,i}$. So, any element $b \in VB_{n}$ can be written as a product in the $\mu_{ij}$'s and the $v_{k}$'s. Furthermore, by the slide relations of Lemma~\ref{stringyslides}, all $\mu_{ij}$'s can pass to the top of the braid, leaving at the bottom a word $\tau$ in the $v_{k}$'s, such that $\tau = \pi(b)$. Thus, if $b \in VP_{n}$ then $\tau$ must be the identity permutation. This completes the proof of the Lemma.
\end{proof}
We can now give a stringy presentation of $VP_{n}$.
\begin{thm} The following is a presentation for the pure virtual braid group.
\begin{equation}\label{purepres}
VP_{n} = \left< \begin{array}{ll}
\begin{array}{c}
\mu_{rs}, \quad r\neq s \\
\end{array} &
\left|
\begin{array}{l}
\mu_{ij}\mu_{ik}\mu_{jk} = \mu_{jk}\mu_{ik}\mu_{ij}, \quad \text{for all distinct } i,j,k \\
\mu_{ij}\mu_{kl} = \mu_{kl}\mu_{ij}, \quad \{i,j\} \cap \{k,l\} = \emptyset \\
\end{array} \right. \end{array} \right>
\end{equation}
\end{thm}
\begin{proof} Having reformulated the presentation of the virtual braid group, the proof is now a direct application of the
Reidemeister-Schreier technique. The relations in $VP_{n}$ arise as conjugations of the relations in $VB_{n}$ by coset representatives of $VP_{n}$ in $VB_{n}$, which are the elements of $S_n$. The relations $(S1), (S2), (S3)$ describe $S_{n}$ and are used for choosing the coset representatives. We now describe the process from the point of view of covering spaces.
We have $VP_{n} \subset VB_{n}$ as a normal subgroup with the subgroup $S_{n}$ acting
on it by conjugation. $VP_{n}$ is the fundamental group of the covering space $E$ of a cell complex $B$ with fundamental group $VB_{n}$, where $E$ has
group of deck transformations $S_{n}.$ Since the elements of the symmetric group lift to paths in the covering space, the relations $\tau \mu_{ij} \tau^{-1} = \mu_{\tau (i), \tau (j)}$ serve to describe the action of the symmetric group on the loops in the covering space (these loops are the lifts of the elements $\mu_{ij}$).
We choose basic relations in $VP_{n}$ to be the lifts at a specific basepoint of the braiding relation
$ \mu_{12}\mu_{13}\mu_{23} = \mu_{23}\mu_{13}\mu_{12}$
and the commuting relation $\mu_{12}\mu_{34} = \mu_{34}\mu_{12}$. All other relations are obtained from these by the action of $S_{n}$, and all relations constitute the two orbits of the basic relations under this action. For example the
relations
$$\mu_{ij}\mu_{ik}\mu_{jk} = \mu_{jk}\mu_{ik}\mu_{ij}$$
constitute the orbit under the action of $S_{n}$ on the single basic braiding relation
$$\mu_{12}\mu_{13}\mu_{23} = \mu_{23}\mu_{13}\mu_{12}.$$
The same pattern applies to the commuting relations.
This gives the statement of the Theorem and completes the proof.
\end{proof}
\subsection {Semi-Direct Product Structure}
The virtual braid group and the pure virtual braid group can be described in terms of semi-direct products of groups, just as is begun
in the paper by Bardakov \cite{Bardakov} and continued in \cite{Paris}. In this section we remark that these decompositions are based on the following algebra:
The Yang-Baxter relation has the generic form $$\mu_{i,i+1}\mu_{i,i+2}\mu_{i+1,i+2} = \mu_{i+1,i+2}\mu_{i,i+2}\mu_{i,+1}$$
which is abstractly in the form $$ABC = CBA$$ and can be rewritten in the form $B^{-1}ABC = B^{-1}CBA$ or
$$A^{B} = C^{B}AC^{-1}.$$ This allows one to rewrite some of the Yang-Baxter relations in terms of the conjugation action of the group
on itself, and this is the key to the structural work pioneered by Bardakov.
\section{A String Category for the Virtual Braid Group}
In this section we summarize our results by pointing out that the string connectors and the virtual crossings can be regarded as generators of a category whose algebraic structure yields the virtual
braid group and the pure virtual braid group. There are many relations in the definition of
this category. These relations all act to make the string connection a topological model of a logical connection between strands in this tensor category. The specific topological interpretations of all these relations have been discussed in the preceding sections of this paper.
\smallbreak
We define a strict monoidal category with generating morphisms
$\mu_{ij}$ where this symbol is interpreted as an abstract string or connection between
strands $i$ and $j$ in a diagram that otherwise is an identity braid on $n$ strands just as defined in the previous sections. The other generators of this category are morphisms $v_{i}$ that are
interpreted as virtual crossings between strings $i$ and $i +1.$ The generators $v_{i}$ have all the relations for transpositions generating the symmetric group. Compositions of these elements generate
the morphisms of the category. The relations among these morphisms are exactly the relations
described for the $v_{k}$ and the $\mu_{ij}$ in the previous sections. We will now define this category
using a minimal number of generators.
\begin{defn} \label{sc} \rm
Consider the strict monoidal category freely generated by one object $*$ and three morphisms
$$\mu: * \otimes * \longrightarrow * \otimes *,$$ $$\mu': * \otimes * \longrightarrow * \otimes *,$$
and $$v:* \otimes * \longrightarrow * \otimes *.$$ Let $\mu_{12} = \mu \otimes id_{*},$
$\mu_{21} = \mu' \otimes id_{*},$
$v_{1} = v \otimes id_{*}, v_{2} = id_{*} \otimes v.$ Here we express these elements in
three strands (tensor factors). For an arbitrary number of tensor factors, we write
$$v_{i} = id_{*} \otimes \cdots \otimes id_{*} \otimes v \otimes id_{*} \otimes \cdots \otimes id_{*}$$ where $v$ occurs in the $i$-th place in this tensor product. More generally, it is understood that
$$\mu_{12} = \mu \otimes id_{*} \otimes \cdots \otimes id_{*}$$
and that $$\mu_{21} = \mu' \otimes id_{*} \otimes \cdots \otimes id_{*}$$
for an arbitrary number of tensor factors.
\smallbreak
\noindent For each natural number $n$, the symbols
$$[n]= * \otimes * \otimes \cdots \otimes *$$ with $n$ $*$'s are the objects in the category. One can regard $[n]$ as an ordered row of $n$ points that constitute the top or the bottom of a diagram involving $n$ strands.
\smallbreak
Now quotient this category by the following relations (compare with the reduced presentation of the virtual
braid group in Proposition 1).
\begin{enumerate}
\item $\mu \mu' = id_{* \otimes *} = \mu' \mu ,$
\item $v v = id_{*},$
\item $\mu_{12} v_j=v_j \mu_{12}, \ \ \ \mbox{for} \ j>2 , $
\item $ \mu_{12} \, v_2 \mu_{12} v_2 \, v_1v_2 \mu_{12} v_2 v_1 =
v_1v_2 \mu_{12} v_2 v_1 \, v_2 \mu_{12} v_2 \, \mu_{12} , $
\item $\mu_{12}\, v_2 v_3 v_1 v_2 \mu_{12} v_2 v_1 v_3 v_2 =
v_2 v_3 v_1 v_2 \mu_{12} v_2 v_1 v_3 v_2 \, \mu_{12}, $
\item $v_{i} v_{i+1} v_{i} = v_{i+1} v_{i} v_{i+1}, $
\item $v_i v_j = v_j v_i, \mbox{for} \ j\neq i\pm 1.$
\end{enumerate}
This quotient is called the {\em String Category} and denoted $SC.$ The category $SC$ is still strict
monoidal.
\end{defn}
To recapture the connecting string morphisms $\mu_{ij}$ in the String Category context, we follow the formalism of the previous sections. Define
$$\mu_{i,i+1} = id_{*} \otimes \cdots \otimes id_{*} \otimes \mu \otimes id_{*} \otimes \cdots \otimes id_{*}$$ where
$\mu$ occurs in the $i$ and $i+1$ places in the tensor product and define
$$\mu_{i+1,i} = id_{*} \otimes \cdots \otimes id_{*} \otimes \mu' \otimes id_{*} \otimes \cdots \otimes id_{*}$$ where
$\mu'$ occurs in the $i$ and $i+1$ places in the tensor product.
Define, for $i<j$, the element $\mu_{ij}$ by the formula
\begin{equation}\label{catmuij}
\mu_{ij} = v_{j-1} v_{j-2} \ldots v_{i+1} \, \mu_{i,i+1} \, v_{i+1} \ldots v_{j-2} v_{j-1}.
\end{equation}
and define
\begin{equation}\label{catmuji}
\mu_{ji} = v_{j-1} v_{j-2} \ldots v_{i+1} \, \mu_{i+1,i} \, v_{i+1} \ldots v_{j-2} v_{j-1}.
\end{equation}
\begin{rem} \rm
In this notation, relation $(4)$ in Definition \ref{sc} becomes the algebraic Yang-Baxter equation
$$\mu_{12} \mu_{13} \mu_{23} = \mu_{23} \mu_{13} \mu_{12},$$ and relation $(5)$ becomes the
commuting relation $$\mu_{12} \mu_{34} = \mu_{34} \mu_{12}.$$
Then one has, as consequences, the general algebraic Yang-Baxter equation and commuting relations, as we have described them in earlier sections of the paper:
$$\mu_{ij}\mu_{ik}\mu_{jk} = \mu_{jk}\mu_{ik}\mu_{ij}, \quad \text{for all distinct } i,j,k$$
and
$$\mu_{ij}\mu_{kl} = \mu_{kl}\mu_{ij}, \quad \{i,j\} \cap \{k,l\} = \emptyset.$$
Diagrammatically, $\mu_{ij}$ consists in
$n$ parallel strands with a string connector between the $i$-th and $j$-th strands directed from
$i$ to $j.$ Similarly, $v_{i}$ corresponds to a diagram of $n$ strands where there is a virtual crossing
between the $i$-th and $(i+1)$-st strands. An $n$-strand diagram that is a product of
these generators is regarded as a morphism from $[n]$ to $[n]$ for $n$ any natural number. We interpret $\mu_{ij}$ and $v_{i}$ diagrammatically
according to the conventions previously established in this paper.
\end{rem}
The morphisms $v_{i}$ effect the action of the symmetric group and the category models the
virtual braid group in the following precise sense.
\begin{thm}
The virtual braid group on
$n$ strands is isomorphic to the group of morphisms from $[n]$ to $[n]$ in the String Category.
\end{thm}
\begin{proof}
By Proposition 1, for any positive integer $n,$ the group of endomorphisms of the object
$ [n] = *^{\otimes n}$ is isomorphic to $VB_{n}.$
\end{proof}
The point of this categorical formulation of the virtual braid groups is that we see how these groups form a natural extension of the symmetric groups by formal elements that satisfy the algebraic Yang-Baxter equation.
The category we desribe is a natural structure for an algebraist interested in exploring formal properties of the algebraic Yang-Baxter equation. It should be remarked that the relationship between the relations
in the virtual pure braid group and the algebraic Yang-Baxter equation was also pointed out in
\cite{BER}. See also \cite{Bellingeri} Remark 10. We have taken this observation further to point out that the virtual braid group is a direct result of forming a convenient category associated with the algebraic Yang-Baxter equation.
\smallbreak
For the reader who would like to take the String Category $SC$ as a starting point for the theory of virtual
braids, here is a description of how to read our figures for that purpose. Figure~\ref{genrs} illustrates the
permutation generators $v_{i}$ for the String Category. The braiding elements $\sigma_{i}$ will be
defined in terms of the string generators. Elementary connecting strings are given in Figure~\ref{elstrings}.
It is implicit in Figure~\ref{elstrings} how to define the braiding elements $\sigma_{i}$ by composing
string generators with permutations (virtual crossings). See also Figure~\ref{relstrings}, which illustrates basic relationships among
string generators, permutations and braiding operators. Figure~\ref{longstrings} illustrates the general connecting strings and their relations with the permutation operators. In particular, Figure~\ref{longstrings} shows how any string
connection can be written in terms of a basic string generator and a product of permutations.
Figure~\ref{exchange}
illustrates how $\mu_{ij}$ and $\mu_{ji}$ are related diagrammatically. Figures~\ref{slides},~\ref{slidepf} and~\ref{commute} show the basic slide relations between string connections and permutations.
Figure~\ref{sbr} illustrates the
algebraic Yang-Baxter relation as it occurs for the string connectors.
\section{Representations of the Virtual and Pure Virtual Braid Groups}
\subsection{}
Let $A$ be an algebra over a ground ring $k$. Let $\rho \in A \otimes A$ be an element of the tensor product of $A$ with itself. Then $\rho$ has the form given by the following equation
\begin{equation}
\rho = \sum_{i = 1}^{N} e_{i} \otimes e^{i}
\end{equation}
where $e_{i}$ and $e^{j}$ are elements of the algebra $A$.
We will write this sum symbolically as
\begin{equation}
\rho = \sum e \otimes e'
\end{equation}
where it is understood that this is short-hand for the above specific summation.
We then define, for $i < j$, $\rho_{ij} \in A^{\otimes n}$ by the equation
\begin{equation}
\rho_{ij} = \sum 1_{A} \otimes \cdots \otimes 1_{A} \otimes e \otimes 1_{A} \otimes \cdots \otimes 1_{A} \otimes e' \otimes 1_{A} \otimes \cdots \otimes 1_{A}
\end{equation}
where the $e$ occurs in the $i$-th tensor factor and the $e'$ occurs in the $j$-th tensor factor.
With $i < j$ we also define $\rho_{ji}$ by reversing the roles of $e$ and $e'$ as shown in the next equation
\begin{equation}
\rho_{ij} = \sum 1_{A} \otimes \cdots \otimes 1_{A} \otimes e' \otimes 1_{A} \otimes \cdots \otimes 1_{A} \otimes e \otimes 1_{A} \otimes \cdots \otimes 1_{A}
\end{equation}
where $e'$ occurs in the $i$-th tensor factor and $e$ occurs in the $j$-th tensor factor.
We say that $\rho$ is a {\it solution to the algebraic Yang-Baxter equation} if it satisfies, in
$A^{\otimes n}$ for $n \geq 3,$ the equation
\begin{equation} \label{ybe}
\rho_{12} \rho_{13} \rho_{23} = \rho_{23} \rho_{13} \rho_{12}.
\end{equation}
It is immediately obvious that if $\rho$ satisfies the algebraic Yang-Baxter equation, then, for any
pairwise distinct $i, j, k$ we have
\begin{equation}
\rho_{ij} \rho_{ik} \rho_{jk} = \rho_{jk} \rho_{ik} \rho_{ij}.
\end{equation}
This gives all possible versions of the algebraic Yang-Baxter equation occuring in the tensor product
$A^{\otimes n}.$
\smallbreak
The following proposition is an immediate consequence of our presentation for the pure virtual braid group.
\begin{prop} Let $VP_{n}$ denote the pure virtual braid group with generators $\mu_{ij}$ and relations as given in
Theorem~2 of Section~3. Let $A$ be an algebra with an invertible algebraic solution to the Yang-Baxter equation denoted by
$\rho \in A \otimes A$ as described above. Define $$rep: VP_{n} \longrightarrow A^{\otimes n}$$ by the equation
\[
rep(\mu_{ij}) = \rho_{ij}.
\]
Then $rep$ extends to a representation of the the virtual braid group to the tensor algebra $A^{\otimes n}$.
\end{prop}
\begin{proof}
It follows at once from the definitions of the $\rho_{ij}$ that $\rho_{ij}\rho_{kl} = \rho_{kl}\rho_{ij}$ whenever the sets
$\{i,j\}$ and $\{k,l\}$ are disjoint. Thus, we have shown that the $\rho_{ij}$ satisfy all the relations in the pure virtual braid group.
This completes the proof of the Proposition.
\end{proof}
Next, we show how to obtain representations of the full virtual braid group. To this purpose, consider the algebra $Aut(A^{\otimes n})$ of
linear automorphisms of $A^{\otimes n}$ as a module over $A$. Assume that we are given an invertible
solution to the algebraic Yang-Baxter equation,
$\rho \in A \otimes A$, and define $\tilde{\rho}_{ij}:A^{\otimes n} \longrightarrow A^{\otimes n} $ by the equation
$\tilde{\rho}_{ij}( \alpha ) = \rho_{ij} \alpha$ where $\alpha \in A^{\otimes n}.$ Since $\rho$ is invertible, $\tilde{\rho}_{ij} \in Aut(A^{\otimes n})$. Let $P_{ij} : A^{\otimes n} \longrightarrow A^{\otimes n}$ be the mapping that interchanges the $i$-th and $j$-th tensor
factors. Note that $P_{ij} \in Aut(A^{\otimes n}).$ We let $P_{i}$ denote $P_{i,i+1}.$ We now define
$$Rep: VB_{n} \longrightarrow Aut(A^{\otimes n})$$ by
the equations
$$Rep(\mu_{ij}) = \tilde{\rho}_{ij} \ \ \ \mbox{and} \ \ \ Rep(v_{i}) = P_{i}.$$
The next proposition is a consequence of presentation~(\ref{stringypres}) for the virtual braid group.
\begin{prop} The mapping $Rep:VB_{n} \longrightarrow Aut(A^{\otimes n})$, defined above, is a representation of the virtual braid group to a subgroup of $Aut(A^{\otimes n}).$
\end{prop}
\begin{proof} It is clear that the elements $P_{i}$ obey all the relations in the symmetric group $S_{n}$.
By presentation~(\ref{stringypres}) it remains to show that letting $\lambda = Rep(\tau)$ where $\tau$ is an element of
$S_{n}$, the relations
$$\lambda \rho_{ij} \lambda^{-1} = \tilde{\rho}_{\tau(i), \tau(j)}, \ \ \ \tau \in S_{n}$$
are satisfied in $Aut(A^{\otimes n}).$
Since $\rho_{ij}$ is defined via the placement of the $e$ and $e'$ factors in the summation for $\rho$ on the $i$-th and $j$-th strands, these
relations are immediate. This completes the proof of the proposition.
\end{proof}
\begin{rem} \rm The method we have described for constructing a representation of the
virtual braid group from an algebraic solution to the Yang-Baxter equation generalizes the well-known
construction of a representation of the classical Artin braid group from a solution to the Yang-Baxter equation in braided form. In the usual method for constructing the classical representation, one composes the algebraic solution with a permutation, obtaining a solution to the braiding equation
$(B1)$. This composition is the same as our relation
$$
\sigma_{i} = \mu_{i,i+1} v_{i}
$$
between the braiding element $\sigma_{i}$ and the stringy generator
$\mu_{i,i+1}$ for the pure virtual braid group. Without the concept of virtuality, the direct relationship
of the algebraic Yang-Baxter equation with the braid groups would not be apparent. We see that, from an algebraic point of view, the virtual braid group is an entirely natural construction. It is the universal
algebraic structure related to viewing solutions to the algebraic Yang-Baxter equation inside tensor products of algebras and endowing these tensor products with the natural permutation action of the symmetric group.
\end{rem}
Solutions to the algebraic Yang-Baxter equation are usually thought of as deformations of the identity mapping on a two-fold tensor product $A \otimes A.$ We think of a braiding operator as a deformation of
a transposition, and so one goes between the algebraic and braided versions of such operators by
composition with a transposition.
\smallbreak
The Artin braid group $B_{n}$ is motivated by a combination of topological considerations and the desire for a group structure that is very close to the structure of the symmetric group $S_{n}$.
We have seen that the virtual braid group $VB_{n}$ is motivated at first by a natural extension of
the Artin braid group in the context of virtual knot theory, but now we see a different motivation for
the virtual braid group. Given that one studies the algebraic Yang-Baxter equation in the
context of tensor powers of an algebra $A$, it is thoroughly natural to study the compositions
of algebraic braiding operators placed in two out of the $n$ tensor lines (the stringy generators) and to
let the permutation group of the tensor lines act on this algebra. As we have seen in (\ref{stringypres}), this
is precisely the virtual braid group. Viewed in this way, the virtual braid group has nothing to do with the plane and nothing to do with virtual crossings. It is a natural group associated with the structure of
algebraic braiding.
\subsection{A Representation Category for the String Category}
We now give a categorical interpretation of virtual knot theory and the virtual braid group in terms
of representation modules associated to an algebra $A$ over a commutative ring $k$ with an algebraic Yang-Baxter element $\rho$ as above.
Let $End(A^{\otimes n})$ denote the linear endomorphisms of $A^{\otimes n}$ as a module over $A$.
View $End(A^{\otimes n})$ as the set of morphisms in a category $Mod^{n}_{k}$ with $A^{\otimes n}$ as the single object. We single out the following morphisms in this category:
\begin{enumerate}
\item $\alpha_{1} \otimes \alpha_{2} \otimes \cdots \otimes \alpha_{n} \in A^{\otimes n}$ acting
on $A^{\otimes n}$ by left multiplication,
\item the elements of the symmetric group $S_{n}$, generated by transpositions of adjacent
tensor factors.
\end{enumerate}
In making the representation of $VB_{n}$ we have used the
stringy generators $\mu_{ij}$ and mapped them to sums of morphisms of the first type above.
The virtual braid group $VB_{n}$ described via (\ref{stringypres}), can be viewed as a category with one
object and generators $\mu_{ij}$ and $v_{k}.$ We let $Mod_{k}$ denote the category that is obtained
by taking all of the categories $Mod^{n}_{k}$ together with objects $A^{\otimes n}$ for each natural number $n$ and morphisms from all of the $End(A^{\otimes n}).$
\begin{rem} \rm
Of course any associative algebra can be seen as
a single object category with morphisms the elements of the algebra. But here we have a pictorial
representation of the morphisms as stringy braid diagrams. These diagrams, which capture the pure virtual braid group so far, can be generalized
by taking the transpositions of the form $P_{i, i+1}$ via a diagram of lines
$i$ and $i+1$ crossing through one another to form virtual crossings $v_{i}.$
Seen from the categorical view that we have developed in these last sections, the virtual crossings are interpreted as generators of the symmetric group whose action is added naturally to the algebraic structure of the pure virtual braid group. By bringing in this action, we expand the pure virtual braid group to the virtual braid group. The virtual crossings have thus become part of the embedded symmetry of the structure of the virtual braid group. This is in sharp contrast to the role of the virtual crossings in the original form of the virtual knot theory. There the virtual crossings appear
as artifacts of the presentation of virtual knots in the plane where those knots acquire extra crossings
that are not really part of the essential structure of the virtual knot. Nevertheless, these same crossings
appear crucially in the virtual braid group, and turn into the generators of the symmetric group embedded in the virtual braid group. With the use of the full set of $\mu_{ij}$ in (\ref{stringypres}) the detour moves and other remnants of the virtual crossings as artifacts have completely disappeared into the permutation action. We will continue the categorical discussion for the virtual braid group, after first discussing certain aspects of knot theory and the tangle categories.
\end{rem}
We can now state a general representation theorem.
\smallbreak
\begin{thm} Any monoidal functor
$$F: SC \longrightarrow Mod_{k}$$
gives rise to a representation of $VB_{n}:$
$$f \in End_{SC}([n]) \simeq VB_{n} \longmapsto F(f) \in End_{k}(A^{\otimes n})$$
where $A = F(*).$
\end{thm}
\begin{proof} The proof follows from the previous discussion.
\end{proof}
The representations of $VB_{n}$ that we have here derived can be interpreted as follows.
\begin{thm} Let $\rho \in A \otimes A$ be a solution of the algebraic Yang-Baxter equation, where
$A$ is an algebra over a commutative ring $k.$ One can then define a monoidal functor
$$F_{A}:SC \longrightarrow Mod_{k}$$ by setting $F_{A}(*) = A$, $F_{A}(\mu) = \tilde{\rho}$,
and $F_{A}(v_{i}) = P$, where the endomorphisms $\tilde{\rho}$ and $P$ of $A \otimes A$ are
given by $$\tilde{\rho}(x \otimes y) = \rho(x \otimes y)$$ and $$P(x \otimes y) = y \otimes x$$
for all $x, y \in A.$
\end{thm}
\begin{proof} The proof follows from the previous discussion.
\end{proof}
\subsection {Virtual Hecke Algebra}
>From the point of view of the theory of braids the Hecke algebra $H_{n}(q)$ is a quotient of the
group ring ${\Bbb Z}[q,q^{-1}][B_{n}]$ of the Artin braid group by the ideal generated by the quadratic expressions
\begin{equation} \label{HeckeQuad}
\sigma_{i}^2 - z\sigma_{i} - 1
\end{equation}
for $i = 1, 2, \ldots n-1,$ where $z = q - q^{-1}.$
This corresponds to the identity $\sigma_{i} - \sigma_{i}^{-1} = z1$,
which is sometimes regarded diagrammatically as a skein identity for calculating knot polynomials.
By the same token, we define the {\it virtual Hecke algebra} $VH_{n}(q)$ to be the quotient of the group ring
${\Bbb Z}[q,q^{-1}][VB_{n}]$ by the ideal generated by Eqs.~\ref{HeckeQuad}.
\smallbreak
There are difficulties in extending structure theorems for the Hecke algebra to corresponding structure
theorems for the virtual Hecke algebra, such as finding normal forms, studying the representation theory
and constructing Markov traces.
Yet, some matters of representations do generalize directly.
In particular, let $W$ be a module over ${\Bbb Z}[q, q^{-1}]$
and let $I:W \longrightarrow W$ be the identity operator.
If $R:W \otimes W \longrightarrow W \otimes W$ is a solution to the Yang-Baxter equation satisfying
$$R^{2} = zR + I,$$ then one has a corresponding representation
$T: VH_{n}(q) \longrightarrow Aut(W^{\otimes n})$. This representation is specified as follows.
\begin{equation}
T(\sigma_{i} ) = \sum 1 \otimes \cdots \otimes 1 \otimes R \otimes 1 \otimes \cdots \otimes 1
\end{equation}
where $R$ operates on the $i$-th and $(i+1)$-st tensor factors, and
\begin{equation}
T(v_{i}) = \sum 1 \otimes \cdots \otimes 1 \otimes P \otimes 1 \otimes \cdots \otimes 1
\end{equation}
where $P$ acts by permuting the $i$-th and $(i+1)$-st tensor factors. It is easy to see that this gives a representation of the virtual Hecke algebra.
\smallbreak
One can hope that the presence of such representations would shed light on the existence of
a generalization of the Ocneanu trace \cite{Jones} on the Hecke algebra to a corresponding trace
and link invariant using the virtual Hecke algebra. At this point there is an issue about the nature of the
generalization. One can aim for a trace on the virtual Hecke algebra that is compatible with the
Markov Theorem for virtual knots and links as formulated in \cite{Kamada,KL4}. This means that the trace must be compatible with both classical and virtual stabilization. This is a trace that
is difficult to achieve. A simpler trace is possible by working in rotational virtual knot theory where virtual stabilization is not allowed \cite{VKT}. See the next section for a discussion of unoriented quantum invariants for rotational virtuals.
We will report on the relation of this approach with the Markov Theorem for virtual knots and links in a separate paper.
\smallbreak
Another line of investigation is suggested by translating the basic Hecke algebra relation into the language of stringy connections. We have $\sigma = \mu v$ for the abstract relation between a braiding generator, a connector and a virtual element. Thus, the virtual Hecke relation
$\sigma^{2} = z \sigma + 1$
becomes $$\mu v \mu = z \mu + v,$$ and it is possible to work in the presentation (20) of the virtual
braid group to find a structure theory for the virtual Hecke algebra.
\section {Rotational Virtual Links, Quantum Algebras, Hopf algebras and the Tangle Category}
This section will show how the ideas and methods of this paper fit together with representations
of quantum algebras (to be defined below) and Hopf algebras and invariants of virtual links.
We begin with a quick review of the theory of virtual links (in relation to virtual braids), and we construct
the virtual tangle category. This category is a natural generalization of the virtual braid group. A functor from the virtual tangle category to an algebraic category will form a generalization of the representations of virtual braid groups that we have discussed in the previous section. This functor
is related to (rotational) invariants of virtual knots and links. It is not hard to see that the construction given in this section defines a category (for arbitrary Hopf algebras) that generalizes the String Category given earlier in this paper. The category that we define here contains virtual crossings, special elements that satisfy the algebraic Yang-Baxter equation and also cup and cap operators. The subcategory
without the cup and cap operators and without any (symbolic) algebra elements except those involved with the algebraic Yang-Baxter operators is isomorphic to the String Category.
\smallbreak
A word to the reader about this section: In one sense this section is a review of known material in the form that Kauffman and Radford \cite{KRH} have shaped the theory of quantum invariants of knots and three-manifolds via finite-dimensional Hopf algebras. On the other hand, this theory is generalized here to invariants of rotational virtual knots and links. This generalization is new, and it is directly related to the structure of the virtual braid group as described in the earlier part of this paper. We have given a complete sketch of this generalization.
The reader should take the word {\it sketch} seriously and concentrate on the sequence of diagrams that depict the ingredients of the theory. Taking this point of view, the reader can see that the appearance of the algebraic Yang-Baxter element in our diagrams (See Figure~\ref{ApplyFunctor}) is aided by using a connecting string exactly analogous to the connecting string in the earlier part of the paper. The generalization follows by taking the functorial image of the virtual tangle category defined in this section.
\smallbreak
\subsection{Virtual Diagrams}
We begin with Figure~\ref{vmoves}. This figure illustrates the moves on virtual knot and link
diagrams that serve to define the theory of virtual knots and links. Two knot or link diagrams
with virtual and classical crossings are said to be {\it virtually isotopic} if one can be obtained from
the other by a finite sequence of these moves. In the figure the moves are divided into
type A, B and C moves. Moves of type A are the classical Reidemeister moves. These
are essentially the same as corresponding moves in the Artin braid group except for the boxed
move involving a loop in the diagram. The move involving this loop is usually called the
{\it first Reidemeister move.} When we forbid the first Reidemeister move, the equivalence relation is
called {\it regular isotopy}. The moves of type B are purely virtual and (except for the move involving
a virtual loop) correspond to the properties of virtual crossings in the virtual braid group. We call the
equivalence relation that forbids both the virtual loop move and the classical loop move
{\it virtual regular isotopy}. Finally, we have moves of type C. These are the local detour moves, and
they correspond to the mixed moves in the virtual braid group.
\smallbreak
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=10cm]{F17.eps}
\end{tabular}
\caption{Virtual moves}
\label{vmoves}
\end{center}
\end{figure}
In this section we will work with virtual knots and links up to virtual regular isotopy. In addition to the
usual kinds of virtual phenomena, we will see some extra features in looking at this equivalence relation.
Two virtual knot or link diagrams are said to be {\it rotationally equivalent} if they are equivalent under virtual regular isotopy. {\it Rotational virtual knot theory}
is the study of the rotational equivalence classes of virtual knot and link diagrams. Studied under this equivalence relation, virtual knot and link diagrams are called {\it rotational virtuals.} We shall say that a virtual knot or link is {\it rotationally knotted} or {\it rotationally linked} if it is not equivalent to an unknot or an unlink under virtual regular isotopy.
View Figure~\ref{rotateknot} and Figure~\ref{rotatelink}. In the first figure we illustrate a rotational
virtual knot, and in the second we show a rotational virtual link. Both the knot and the link are kept from
being trivial by the presence of flat loops as discussed above. There is much more to say about rotational virtuals, and we refer the reader to \cite{VKT} for some steps in this direction.
\smallbreak
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=2.5cm]{F18.eps}
\end{tabular}
\caption{A rotational virtual knot}
\label{rotateknot}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=3.3cm]{F19.eps}
\end{tabular}
\caption{A rotational virtual link}
\label{rotatelink}
\end{center}
\end{figure}
\subsection{The Virtual Tangle Category}
The advantage in studying virtual knots up to virtual regular isotopy is that all so-called quantum link
invariants generalize to invariants of virtual regular isotopy. This means that virtual regular isotopy is
a natural equivalence relation for studying topology associated with solutions to the Yang-Baxter
equation.
\smallbreak
Here we create a context by defining the {\it Virtual Tangle Category}, $VTC,$ as indicated
in Figure~\ref{regtang}.
The tangle category is generated by the morphisms
shown in the box at the top of this figure. These generators are: a single identity line, right-handed and left-handed
crossings, a cap and a cup, a virtual crossing. The objects in the tangle category consist in the set of
$[n]$'s where $n = 0, 1,2, \ldots.$ For a morphism $[n] \longrightarrow [m]$, the numbers $n$ and $m$ denote, respectively, the number of free arcs at the bottom and at the top of the diagram that represents the morphism. The morphisms are like braids except that they can (due to the presence of the cups and caps) have different numbers of free ends at the top and the bottom of their diagrams.
\smallbreak
The sense in which the elementary morphisms (line, cup, cap, crossings) generate the tangle category is composition as shown in Figure~\ref{tangprod}. For composition, the segments are matched so that the number of lower free ends on each segment is equal to the number of upper free ends on the segment below it.
The Figure~\ref{tangprod} shows a virtual trefoil as a morphism from
$[0]$ to $[0]$ in the category.
The tensor product of morphisms is the horizontal juxtaposition of their
diagrams. Each of the seven horizontal segments of the figure represents one of the elementary morphisms tensored with the identity line. Consequently there is a well-defined composition of all of the segments and this composition is a morphism $[0] \longrightarrow [0]$ that represents the knot.
\smallbreak
The basic equivalences of morphisms are
shown in Figure~\ref{regtang}. Note that $II, III, V$ are formally equivalent to the rules for unoriented
virtual braids. The zero-th move is a cancellation of consecutive maxima and minima, and the move $IV$
is a swing move in both virtual and classical relations of crossings to maxima and minima. It should be
clear that the tangle category is a generalization of the virtual braid group with a natural inclusion of unoriented virtual braids as special tangles in the category. Standard braid closure and the plat closure of braids
have natural definitions as tangle operations. Any virtual knot or link can be
represented in the tangle category as a morphism from $[0]$ to $[0],$ and one can prove that {\it two virtual links are
virtually regularly isotopic if and only if their tangle representatives are equivalent in the tangle category.}
None of the rules for equivalence in the tangle category involve either a classical loop or a
virtual loop. This means that the virtual tangle category is a natural home for the theory of rotational
virtual knots and links.
\smallbreak
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=8cm]{F20.eps}
\end{tabular}
\caption{Regular isotopy with respect to the vertical direction}
\label{regtang}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=6cm]{F21.eps}
\end{tabular}
\caption{Virtual trefoil as a morphism in the tangle category}
\label{tangprod}
\end{center}
\end{figure}
\subsection{Quantum Algebra and Category}
Now we shift to a category associated with an algebra that is directly related to our representations
of the virtual braid group. We take the following definition \cite{KP, KRH}: A {\it quantum algebra} $A$
is an algebra over a commutative ground ring $k$ with an invertible mapping $s: A \longrightarrow A$
that is an {\it antipode}, that is $s(ab) = s(b)s(a)$ for all $a$ and $b$ in $A,$ and there is an element
$\rho \in A \otimes A$ satisfying the algebraic Yang-Baxter equation as in Equation~\ref{ybe}:
$$\rho_{12} \rho_{13} \rho_{23} = \rho_{23} \rho_{13} \rho_{12}.$$
We further assume that $\rho$ is invertible and that
$$ \rho^{-1} = (1_{A} \otimes s)\circ \rho = (s \otimes 1_{A})\circ \rho.$$
The multiplication in the algebra is usually denoted by $m: A\otimes A \longrightarrow A$ and is
assumed to be associative. It is also assumed that the algebra has a multiplicative unit element.
The defining properties of a quantum algebra are part of the properties of a Hopf algebra, but a Hopf algebra has a comultiplication $\Delta: A \longrightarrow A \otimes A$
that is a homomorphism of algebras, plus a list of further relations, including a fundamental relationship between the multiplication, the comultiplication and the antipode. In the interests of simplicity, we shall restrict ourselves to quantum algebras here, but most of the remarks that follow apply to Hopf algebras, and particularly quasi-triangular Hopf algebras. Information on Hopf algebras is included at the end of this section. See \cite{KRH} for more about these connections.
\smallbreak
We construct a category $Cat(A)$ associated with a quantum algebra $A$. This category is a very close relative to the virtual tangle category. $Cat(A)$ differs from the tangle category in that it has only virtual crossings, and
there are labeled vertical lines that carry elements of the algebra $A.$ See Figure~\ref{CatMorph}.
Each such labeled line is a morphism in the category. The virtual crossing is a generating morphism as are the cups, caps and labeled lines. The objects in this category are the same
entities $[n]$ as in the tangle category. This category is identical in its framework to the tangle category but the crossings are not
present and lines labeled with algebra are present.
Given $a,b \in A$ we compose the morphisms corresponding to $a$ and $b$ by taking a line labeled
$ab$ to be their composition. In other words, if $\langle x \rangle$ denotes the morphism in
$Cat(A)$ associated with $x \in A$, then
$$\langle a \rangle \circ \langle b \rangle = \langle ab \rangle.$$
As for the additive structure in the algebra, we extend the category to an additive category by formally adding the generating morphisms (virtual crossings, cups, caps and algebra line segments). In Figure~\ref{CatMorph} we illustrate the composition of
such morphisms and we illustrate a number of other defining features of the category $Cat(A).$
\smallbreak
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=8cm]{F22.eps}
\end{tabular}
\caption{ Morphisms in $Cat(A)$}
\label{CatMorph}
\end{center}
\end{figure}
In the same figure we illustrate how the tensor product of elements $a \otimes b$ is represented by parallel vertical
lines with $a$ labeling the left line and $b$ labeling the right line. We indicate that the virtual
crossing acts as a permutation in relation to the tensor product of algebra morphisms. That is, we
illustrate that $$\langle a \rangle \otimes \langle b \rangle \circ P = P \circ \langle b \rangle \otimes \langle a \rangle.$$ Here $P$ denotes the virtual crossing of two segments, and is regarded as a morphism $P: V \otimes V \longrightarrow V \otimes V$ (see remark below). Since the lines interchange, we expect $P$ to behave as the permutation of the two tensor factors.
\smallbreak
In Figure~\ref{CatMorph} we show the notation
$V$ for the object $[1]$ in this category and we use $V \otimes V = [2]$, $V \otimes V \otimes V= [3]$
and so on for all the natural number objects in the category. We write
$[0] = k$, identifying the ground ring with the ``empty object" $[0].$ It is then axiomatic that
$k \otimes V = V \otimes k = V.$ Morphisms are indicated both diagrammatically and in terms of arrows
and objects in this figure. Finally, the figure indicates the arrow and object forms of the cup and the
cap, and crucial axioms relating the antipode with the cup and the cap.
A cap is regarded as a morphism from $V \otimes V$ to $k$, while
a cup is regarded as a morphism form $k$ to $V \otimes V.$
The basic property of the cup and the cap is the {\em Antipode Property: if
one ``slides" a decoration across the maximum or minimum in a counterclockwise
turn, then the antipode $s$ of the algebra is applied to the decoration.}
In categorical terms this property says $$Cap \circ (\langle 1 \rangle \otimes a) = Cap \circ
(\langle sa \rangle \otimes 1 )$$ and $$(\langle a \rangle \otimes 1) \circ Cup = (1 \otimes \langle sa \rangle ) \circ Cup.$$ Here $1$ denotes the identity morphism for $[0]$. These properties and other naturality properties of the cups and the
caps are illustrated in Figure~\ref{CatMorph} and Figure~\ref{antipode}. The naturality properties of the flat diagrams in this category include regular homotopy of immersions (for diagrams without algebra decorations), as illustrated in these figures.
\smallbreak
In Figure~\ref{antipode} we see how the antipode property of the cups and caps leads to a
diagrammatic interpretation of the antipode. In the figure we see that the antipode $s(a)$ is represented by composing with a cap and a cup on either side of the morphism for $a$. In terms of the composition
of morphisms this diagram becomes
$$\langle sa \rangle = (Cap \otimes 1) \circ (1 \otimes \langle a \rangle \otimes 1)\circ(1 \otimes Cup).$$
Similarly, we have
$$\langle s^{-1}a \rangle = (1 \otimes Cap) \circ(1 \otimes \langle a \rangle \otimes 1)\circ( Cup \otimes 1).$$
This, in turn, leads to the
interpretation of the flat curl as an element $G$ in $A$ such that
$s^{2}(a) = GaG^{-1}$ for all $a$ in $A.$ $G$ is a flat curl diagram interpreted
as a morphism in the category. We see that, formally, it is natural to interpret
$G$ as an element of $A$. In a so-called {\em ribbon Hopf
algebra} there is such an element already in the algebra. In the general case it
is natural to extend the algebra to contain such an element.
\smallbreak
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=8cm]{F23.eps}
\end{tabular}
\caption{Diagrammatics of the antipode}
\label{antipode}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=8cm]{F24.eps}
\end{tabular}
\caption{Formal trace}
\label{trace}
\end{center}
\end{figure}
\subsection{The Basic Functor and the Rotational Trace}
We are now in a position to describe a functor $F$ from the virtual tangle category $VTC$
to $Cat(A).$ (Recall that the virtual tangle category is defined for virtual link diagrams without
decorations. It has the same objects as $Cat(A).$)
$$F:VTC \longrightarrow Cat(A)$$
The functor $F$ decorates each
positive crossing of the tangle (with respect to the vertical - see Figure~\ref{Functor})
with the Yang-Baxter element (given by the quantum algebra $A$)
$\rho = \Sigma e \otimes e^{'}$ and each negative crossing (with respect to the
vertical) with $\rho^{-1} = \Sigma s(e) \otimes e^{'}$. The form of the
decoration is indicated in Figure~\ref{Functor}. Since we have labelled the negative crossing with the
inverse Yang-Baxter element, it follows at once that the two crossings are mapped to inverse elements in the category of the algebra. {\it This association is a direct generalization of our mapping
of the virtual braid group to the stringy connector presentation.}
\smallbreak
We now point out the structure of the image of a knot, link or tangle under this functor.
The key point about this functor is that, because quantum algebra elements can be
moved around the diagram, we can concentrate all the image algebra in one place.
Because the flat curls are identified with either $G$ or $G^{-1}$, we can use
regular homotopy of immersions to bring the image under $F$ of each component of a virtual link diagram to the form of a circle with a single concentrated decoration (involving a sum over many
products) and a reduced pattern of flat curls that can be encoded as a power of the special
element $G.$ Once the underlying curve of a link component is converted to a loop with
total turn zero, as in Figure~\ref{trace}, then we can think of such a loop, with algebra labeling the
loop, as a representative for a formal trace of that algebra and call it $TR(X)$ as in the figure.
In the figure we illustrate that for such a labeling $$TR(ab) = TR(ba),$$ thus one can take a product of
algebra elements on a zero-rotation loop up to cyclic order of the product. In situations where we
choose a representation of the algebra or in the case of finite dimensional Hopf algebras where one
can use right integrals \cite{KRH}, there are ways to make actual evaluations of such traces. Here we
use them formally to indicate the result of concentrating the algebra on the loop.
\smallbreak
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=8cm]{F25.eps}
\end{tabular}
\caption{The functor $F: VTC \longrightarrow Cat(A)$}
\label{Functor}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=8cm]{F26.eps}
\end{tabular}
\caption{Inverse and antipode}
\label{twist}
\end{center}
\end{figure}
One further comment is in order about the antipode. In Figure~\ref{twist} we show that our axiomatic assumption about the antipode (the sliding rule around maxima and minima) actually demands that
the inverse of $\rho$ is $(s \otimes 1_{A})\circ \rho = (1_{A} \otimes s )\circ \rho$. This follows by
examining the form of the inverse of the positive crossing in the tangle category by turning that crossing
to produce an identity between the positive crossing and the negative crossing twisted with additional
maxima and minima. This relationship shows that if we set the functor $F$ on a right-handed crossing as we have done, then the way it maps the inverse crossing is forced and that this inverse corresponds to the inverse of $\rho$ in the quantum algebra. Thus the quantum algebra formula for the inverse of
$\rho$ is forced by the topology.
\smallbreak
In Figure~\ref{ApplyFunctor} we illustrate the entire functorial process for the virtual trefoil of Figure~\ref{tangprod}. The virtual trefoil is denoted by $K$, and we find that $F(K)$ reduces to a zero-rotation circle with the
inscription $ e' s(f) s^{2}(e) s^{3}(f') G^{2}. $ We can, therefore, write the equation
$$F(K) = TR[e' s(f) s^{2}(e) s^{3}(f') G^{2}].$$
Another way to think about this trace expression is to regard it as a Gauss code for the knot that has extra structure. {\it The chords in the Gauss diagram are the string connectors of the beginning of this paper, generalized to the algebra category $Cat(A).$} The powers of the antipode and the power of $G$ keep track of rotational features in the diagram as it lives in the tangle category up to regular isotopy.
We now see that the mapping of the virtual braid group to the braid group generated by permutations
and string connectors has been generalized to the functor $F$ taking the virtual tangle category
to the abstract category of a quantum algebra. We regard this generalization as an appropriate context for thinking about virtual knots, links and braids.
\smallbreak
The category $Cat(A)$ of a quantum algebra $A$ can be generalized to an abstract category
with labels, virtual crossings, and with stringy connections that satisfy the algebraic Yang-Baxter equation. Each such stringy connection has a left label $e$ or $s(e)$ and a right label $e'.$ We retain the formalism of the antipode
as a formal replacement for adjoining a label with a cup and a cap. The resulting {\it abstract algebra category} will be denoted by $\overline{Cat(A)}$. Since we take this category with no further relations,
the functor $ \overline{F}:VTC \longrightarrow \overline{Cat(A)}$ is an equivalence of categories. This functor is
the direct analog of our reformulation of the virtual braid group in terms of stringy connectors.
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=9cm]{F28.eps}
\end{tabular}
\caption{The functor $F: T \longrightarrow Cat(A)$ applied to a virtual trefoil}
\label{ApplyFunctor}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=6cm]{F29.eps}
\end{tabular}
\caption{Virtual braid and closure}
\label{vbclosure}
\end{center}
\end{figure}
\subsection{Virtual Braids and Their Closures}
The functor $F:VTC \longrightarrow Cat(A)$ defined in the last subsection can be restricted to
the category of virtual (unoriented) braids that we will denote here by $VB.$ If the reader then examines
the result of this functor he will see that the image of a virtual crossing is a virtual crossing, and the image of a braid generator is a string connection (expressed in terms of $Cat(A).$). If we use the
corresponding functor
$$ \overline{F}:VB \hookrightarrow VTC \longrightarrow \overline{Cat(A)},$$ then the image
$ \overline{F}(VB)$ is an abstract category of
string connections and permutations that is (up to orientation) identical with our String Category $SC$
studied throughout this paper. This remark brings us full circle and shows how the String Category
fits in the context of the quantum link invariants discussed in this part of the paper. In particular,
view the bottom part of Figure~\ref{vbclosure} where we have illustrated the image under
$ \overline{F}$ of a particular virtual braid. Each classical crossing in the braid is replaced by a string connector followed by a virtual crossing. The string connector is interpreted as in the abstract Hopf algebra category, but in the braid image there is no other structure than the connectors and the virtual crossings. This shows how the braid lands in a subcategory that is isomorphic with our main category
$SC.$
\smallbreak
Now recall that one can move from virtual braids to virtual knots and links by taking the {\it braid closure.}
The closure $\overline{b}$ of a braid $b$ is obtained by attaching planar disjoint arcs from the outputs of a braid to its inputs as illustrated in Figure~\ref{vbclosure}. The result of the closure is a virtual knot or link. In particular, this means that we can express rotational quantum link invariants by applying
$F$ to the closure of virtual braid and then taking the trace $TR$ described in the last section.
alternatively, one can regard the invariant as $A$-valued where $A$ is the quantum algebra that supports the functor $F.$ Altogether, this subsection and the examples in Figure~\ref{vbclosure}
indicate the close relationship of the different constructions that have been outlined in this paper and
how the structure of the virtual braid group is intimately related to quantum link invariants for rotational
virtual links.
\subsection{Hopf Algebras and Kirby Calculus}
In Figure~\ref{kirby} we illustrate how one can use this concentration of algebra on the loop in the context of a Hopf algebra that has a right integral. The right integral is a function
$\lambda: A \longrightarrow k$ satisfying $$\lambda(x) 1_{A} = \Sigma \lambda(x_{1})x_{2}$$
where the coproduct in the Hopf algebra has formula $\Delta(x) = \Sigma x_{1} \otimes x_{2}$.
Here we point out how the use of the coproduct corresponds to doubling the lines in the diagram, and that if one were to associate the function
$\lambda$ with a circle with rotation number one, then the resulting link evaluation will be invariant under the so-called Kirby move.
The Kirby move replaces two link components with new ones by doubling one component and connecting one of the components of the double with the other component. Under our functor from the virtual tangle category to the category for the Hopf algebra, a knot goes to a circle with algebra concentrated at $x.$ The doubling of the knot goes to concentric circles labeled with the coproduct
$\Delta(x) = \Sigma x_{1} \otimes x_{2}.$ Figure~\ref{kirby} shows how invariance under the handle-slide in the tangle category corresponds the integral equation
$$\lambda(x) y = \Sigma \lambda(x_{1})x_{2} y.$$
It turns out that
classical framed links $L$ have an associated compact oriented three manifold $M(L)$ and that
two links related by Kirby moves have homeomorphic three-manifolds. Thus the evaluation of links using the right integral yields invariants of three-manifolds. Generalizations to virtual three-manifolds are under investigation \cite{DK1}. We only sketch this point of view here, and refer the reader to \cite{KRH}.
\smallbreak
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=8cm]{F27.eps}
\end{tabular}
\caption{The Kirby move}
\label{kirby}
\end{center}
\end{figure}
\subsection {Hopf Algebra}
This section is added for reference about Hopf algebras. Quasitriangular Hopf algebras are an
important special case of the quantum algebras discussed in this section.
\smallbreak
Recall that a {\it Hopf algebra} \cite{Sweedler} is a bialgebra $A$ over a commutative ring $k$ that has an associative multiplication $m:A \otimes A \longrightarrow A,$
and a coassociative comultiplication, and is equipped with a counit, a unit
and an antipode. The ring $k$ is usually taken to be a field. The associative law for the multiplication $m$ is expressed by the equation $$m(m \otimes1_{A}) = m(1_{A} \otimes m)$$ where $1_{A}$ denotes the identity map on A.
\smallbreak
The coproduct
$\Delta :A \longrightarrow A \otimes A$ is an algebra homomorphism and is
coassociative in the sense that
$$(\Delta \otimes 1_{A})\Delta = (1_{A} \otimes \Delta) \Delta.$$
The {\it unit} is a mapping from $k$ to $A$ taking $1_k$ in $k$ to $1_A$ in $A$ and, thereby, defining an action of $k$ on $A.$
It will be convenient to just identify the $1_k$ in $k$ and the $1_A$ in $A$, and to ignore
the name of the map that gives the unit.
\smallbreak
The counit is an algebra mapping from $A$ to $k$ denoted by $\epsilon :A \longrightarrow k.$ The following formula for the counit dualize the structure inherent in the unit:
$$(\epsilon \otimes 1_{A}) \Delta = 1_{A} = (1_{A} \otimes \epsilon) \Delta.$$
It is convenient to write formally
$$\Delta (x) = \sum x_{1} \otimes x_{2} \in A \otimes A$$
to indicate the decomposition of the coproduct of $x$ into a sum of first and second factors in the two-fold tensor product of $A$ with itself. We shall often drop the summation sign and write
$$\Delta (x) = x_{1} \otimes x_{2}.$$
The antipode is a mapping $s:A \longrightarrow A$ satisfying the equations
$$m(1_{A} \otimes s) \Delta (x) = \epsilon (x)1_{A} \quad \mbox{and} \quad m(s \otimes 1_{A}) \Delta (x)= \epsilon (x)1_{A}.$$
It is a consequence of this definition that $s(xy) = s(y)s(x)$ for all $x$ and $y$ in A. \vspace{3mm}
A {\it quasitriangular Hopf algebra} \cite{Drinfeld} is a Hopf algebra $A$ with an element
$\rho \in A \otimes A$ satisfying
the following conditions:
\vspace{3mm}
\noindent
(1) $\rho \Delta = \Delta' \rho$ where $\Delta'$ is the composition
of $\Delta$ with the map on
$A \otimes A$ that switches the two factors. \vspace{3mm}
\noindent
(2) $$\rho_{13} \rho_{12} = (1_{A} \otimes \Delta) \rho,$$ $$\rho_{13} \rho_{23} = (\Delta \otimes 1_{A})\rho.$$
The symbol $\rho_{ij}$ denotes the placement of the first and second tensor factors of $\rho$ in the $i$ and $j$ places in a triple tensor product. For example, if $\rho = \sum e \otimes e'$ then $$\rho_{13} = \sum e \otimes 1_{A} \otimes e'.$$
Conditions (1) and (2) above imply that $\rho$ has an inverse and that
$$ \rho^{-1} = (1_{A} \otimes s^{-1}) \rho = (s \otimes 1_{A}) \rho.$$
It follows easily from the axioms of the quasitriangular Hopf algebra that $\rho$ satisfies the Yang-Baxter equation
$$\rho_{12} \rho_{13} \rho_{23} = \rho_{23} \rho_{13} \rho_{12}.$$
A less obvious fact about quasitriangular Hopf algebras is that there exists an element $u$ such that $u$ is invertible and $s^{2}(x) = uxu^{-1}$ for all $x$ in $A.$ In fact, we may take $u = \sum s(e')e$ where $\rho = \sum e \otimes e'.$ This result, originally due to Drinfeld \cite{Drinfeld}, follows from the diagrammatic categorical context of this paper. \vspace{3mm}
An element $G$ in a Hopf algebra is said to be {\em grouplike} if $\Delta (G) = G \otimes G$ and $\epsilon (G)=1$ (from which it follows that $G$ is invertible and $s(G) = G^{-1}$). A quasitriangular Hopf algebra is said to be a {\em ribbon Hopf algebra} \cite{RTG,KRH} if there exists a grouplike element $G$ such that (with $u$ as in the previous paragraph) $v = G^{-1}u$ is in
the center of $A$ and $s(u) = G^{-1}uG^{-1}$. We call $G$ a special grouplike element of $A.$
\vspace{3mm}
Since $v=G^{-1}u$ is central, $vx=xv$ for all $x$ in $A.$ Therefore $G^{-1}ux = xG^{-1}u.$ We know that $s^{2}(x) = uxu^{-1}.$ Thus $s^{2}(x) =GxG^{-1}$ for all $x$ in $A.$ Similarly, $s(v) = s(G^{-1}u) = s(u)s(G^{-1})=G^{-1}uG^{-1}G =G^{-1}u=v.$ Thus, the square of the
antipode is represented as conjugation by the special grouplike element in a ribbon Hopf algebra, and the central element $v=G^{-1}u$ is invariant under the antipode.
\smallbreak
This completes the summary of Hopf algebra properties that are relevant to the last section of the paper.
|
\section{Introduction}
A compact star is one of the most stable forms of matter in the universe.
The only instability that threatens its existence is collapse into
a black hole, triggered by unstable radial oscillation modes that
push the star below its Schwarzschild radius. However, there are other
unstable oscillation modes, the so called r-modes \cite{Papaloizou:1978zz,Lindblom:1999yk},
which damp the rotation of the star by the emission of gravitational
radiation \cite{Andersson:1997xt}. In slowly-rotating stars r-modes
themselves are damped by bulk and shear viscosity, but at high rotation
frequencies they are unstable and grow exponentially. In a companion
paper \cite{Alford:2010fd} we analyzed these instability regions
in detail. The main result of that study was that these regions vary
greatly between qualitatively different classes of stars containing
distinct phases of strongly interacting matter, but are extremely
insensitive to unknown quantitative details of the equation of state
and the transport properties within a given class. Therefore, a proper
understanding of the r-mode dynamics could in the future provide robust
signatures for the presence of exotic phases in compact stars.
Since exponentially growing r-modes will destroy the star if their
growth is not stopped by some non-linear mechanism, the fact that
fast spinning compact stars are observed suggests that such a non-linear
damping mechanism is indeed present. More importantly, even if stopped
at a finite amplitude, r-modes still strongly emit gravitational waves
and could provide an extremely efficient mechanism for the spin-down
of compact stars \cite{Owen:1998xg} and an interesting signal for
terrestrial gravitational wave detectors. Spin-down due to r-modes
could explain the observed absence of fast-spinning young stars despite
the fact that their creation during a supernova could naturally lead
to a fast spinning remnant. For spin-down via r-modes the size of
the saturation amplitude is crucial. If the amplitude $\alpha$ is
too low, it takes too long to spin down the star; if the amplitude
is too large, $\alpha\!>\! O\left(1\right)$ , the r-mode would disrupt
the star's structure, and even before this point the r-mode could
be destroyed. If the r-mode amplitude saturates at an intermediate
value, a fast spin-down is possible. Previously, various mechanisms
for the large-amplitude behavior of r-modes have been suggested \cite{Lindblom:2000az}.
They include the coupling between different modes \cite{Bondarescu:2007jw,Arras:2002dw},
the decay into daughter modes and the eventual transformation of the
r-mode energy into differential rotation \cite{Gressman:2002zy,Lin:2004wx},
friction between different layers of the star, and surface effects
in the star\textquoteleft{}s crust \cite{Bildsten:2000ApJ...529L..33B}.
Here we study an alternative mechanism that does not involve such
complicated non-linear dynamical or structural effects. It is present
already in a standard hydrodynamical description and exploits the
fact that at large amplitudes the damping due to bulk viscosity increases
dramatically with the amplitude \cite{Alford:2010gw,Reisenegger:2003pd,Madsen:1992sx,Shovkovy:2010xk}.
In this suprathermal regime, where the deviation from chemical equilibrium
$\mu_{\Delta}$ fulfills $\mu_{\Delta}\!\gtrsim\! T$, the viscous
damping could overcome the initial gravitational instability and saturate
the r-mode. However, as shown in \cite{Alford:2010gw}, the bulk viscosity
has a maximum as a function of the amplitude and decreases again at
even larger amplitudes. If the amplitude exceeds this critical value
then the r-mode growth cannot be stopped by viscous damping and other
non-linear dynamic effects \cite{Lindblom:2000az,Gressman:2002zy,Lin:2004wx,Arras:2002dw,Bondarescu:2007jw}
are required to saturate it. Nevertheless, we find that over a significant
region of the parameter space the suprathermal enhancement is indeed
sufficient to saturate the r-mode at a finite amplitude and the r-mode
can then efficiently spin down the star. This is in contrast to certain
non-linear hydrodynamical effects where the r-mode could completely
decay \cite{Lin:2004wx} and would not be able to cause an appreciable
spin-down of the star.
For r-modes with amplitudes sufficiently below the maximum, we give
general analytic expressions for the suprathermal damping time valid
for various forms of dense matter. For r-modes with arbitrary amplitudes,
where an analytic evaluation is not possible, we give a general expression
for the bulk viscosity damping time that includes the complete parameter
dependence required for the analysis of the star's evolution, encoded
in a two-parameter function that can be numerically computed and tabulated
for different star models. This offers an explicit framework for the
consistent inclusion of the r-mode saturation into a star evolution
analysis and supersedes previously necessary model assumptions \cite{Owen:1998xg}.
We will analyze the same star models as in our recent companion paper
\cite{Alford:2010fd} and thereby extend this study to the suprathermal
regime. These include neutron stars, hybrid stars and strange stars.
In addition, motivated by the recent observation of a $2M_{\odot}$
star \cite{Demorest:2010bx,Ozel:2010bz} we only study models that
also yield heavy stars. Moreover, similar to the analytic results
in \cite{Alford:2010fd}, we give approximate analytic expressions
for the maximum saturation amplitude that exhibit the detailed parameter
dependence on the equation of state and the transport properties of
dense matter. In addition to the standard fundamental $m\!=\!2$ r-mode
we also study the saturation of higher multipoles and find that they
saturate at significantly lower amplitudes. In this work we concentrate
on the effects of suprathermal bulk viscosity on the damping times
\cite{Alford:2010ep} and the suitably extended concept of the instability
regions, and defer a comprehensive analysis of the star evolution
to future work.
\section{Damping of star oscillations}
In this section we discuss the prerequisites needed for the analysis
of compact star oscillations and their damping. In particular we discuss
the large amplitude enhancement of the bulk viscosity which will provide
a mechanism for the saturation of r-modes. This topic had been studied
in \cite{Alford:2010gw} and we refer the reader to this work for
further details. The analysis further requires the stable equilibrium
configuration of the star and the r-mode profile. These have been
discussed in more detail in our companion article on the damping of
small amplitude r-modes \cite{Alford:2010fd} and we will here only
briefly recall these results.
\subsection{Suprathermal viscosity}
The bulk viscosity describes the local dissipation of energy in a
fluid element in one cycle of compression and rarefaction, driven
by some oscillation mode of the star. The integration over the whole
star yields the corresponding total energy dissipation of an oscillation
mode as will be discussed in section \ref{sec:Damping-time-scales}.
Recently the bulk viscosity of large-amplitude oscillations has been
studied in detail in \cite{Alford:2010gw}, which builds on the classic
work \cite{Madsen:1992sx} and provides general expressions valid
for various forms of matter and arbitrary equations of state.
Bulk viscosity is generally induced by slow weak-interaction processes
whose rate takes the parametric form
\begin{equation}
\Gamma^{(\leftrightarrow)}=-\tilde{\Gamma}T^{\delta}\mu_{\Delta}\Biggl(1+\sum_{j=1}^{N}\chi_{j}\left(\frac{\mu_{\Delta}^{2}}{T^{2}}\right)^{j}\Biggr)\,,\label{eq:gamma-parametrization}\end{equation}
where $T$ is the temperature and $\mu_{\Delta}$ represents the
quantity that is driven out of equilibrium due to the oscillations
and its re-equilibration leads to the bulk viscosity. The latter is
given by the difference of the sums of chemical potentials of the
particles in the initial and final state of the relevant weak process.
The $\chi_{j}$ are coefficients that characterize the non-linear
(large amplitude) contribution to the re-equilibration rate. The series
terminates at a finite order $N$ which is determined by the number
and type of particles in the initial and final states of the dominant
re-equilibration process and is connected to the temperature dependence
via $\delta\!=\!2N$.
As has been shown in \cite{Alford:2010gw} the bulk viscosity can
be written in a general form that expresses its full underlying parameter
dependence in terms of the coefficients $d$ of the driving term and
$f$ of the feedback term
\[
d\equiv\frac{C}{T}\frac{\Delta n}{\bar{n}}\:,\quad f\equiv\frac{B\tilde{\Gamma}T^{\delta}}{\omega}\,,\]
in the differential equation that determines the oscillation $\mu_{\Delta}$.
These expressions depend on the oscillation frequency $\omega$, the
conserved number density fluctuation $\Delta n/\bar{n}$, as well
as the susceptibilities
\begin{equation}
C\equiv\bar{n}\left.\frac{\partial\mu_{\Delta}}{\partial n}\right|_{x}\quad,\quad B\equiv\frac{1}{\bar{n}}\left.\frac{\partial\mu_{\Delta}}{\partial x}\right|_{n}\,,\label{eq:susceptibilities}\end{equation}
with respect to the density $n$ and the fraction $x$ of a particular
particle that is driven out of equilibrium. The general result can
be parameterized in the form \cite{Alford:2010gw}
\begin{equation}
\zeta=\zeta_{max}^{<}\,{\cal I}\left(d,f\right)\,,\label{eq:general-viscosity}\end{equation}
where the global maximum of the viscosity
\[
\zeta_{max}^{<}=\frac{C^{2}}{2\omega B}\]
is taken at
\begin{equation}
d=0\:,\negmedspace f=1\quad\Rightarrow\quad T_{max}=\left(\frac{\omega}{\tilde{\Gamma}B}\right)^{\frac{1}{\delta}}\label{eq:maximum-temperature}\end{equation}
and the non-trivial dimensionless integral
\begin{equation}
{\cal I}(d,f)\equiv\frac{2}{\pi Td}\int_{0}^{2\pi}\mu_{\Delta}(\varphi;d,f)\cos(\varphi)d\varphi\label{eq:viscosity-integral}\end{equation}
has to be tabulated for each considered form of matter with a particular
dominant weak re-equilibration process. Since both $d$ and $f$ depend
on the temperature and the required parameter regions are therefore
non-trivial, see \cite{Alford:2010ht}, this is most conveniently
done in terms of the new variable $\tilde{d}\!\equiv\! d\, f^{1/\delta}$
for the function $\tilde{{\cal I}}\left(\tilde{d},f\right)\!=\!{\cal I}(d,f)$.
This form of the bulk viscosity is crucial for the study of damping
times below, which involve an integration over the whole star, and
therefore requires the complete density and amplitude dependence of
the viscosity in addition to the temperature and frequency dependence
which are used in standard analyses.
In general the bulk viscosity features three distinct characteristic
regions. Almost all previous analyses of r-mode damping have been
limited to the \emph{subthermal} regime $\mu_{\Delta}\!\ll\! T$,
where $\mu_{\Delta}$ is linear in $d\!\sim\!\Delta n$, so the viscosity
is independent of the amplitude and has the analytic form
\begin{equation}
\zeta^{<}=\zeta_{max}^{<}\frac{2f}{1+f^{2}}=\frac{C^{2}\tilde{\Gamma}T^{\delta}}{\omega^{2}+(B\tilde{\Gamma}T^{\delta})^{2}}\label{eq:sub-viscosity}\end{equation}
Yet, since the r-mode is unstable and rises exponentially it eventually
reaches the \emph{suprathermal regime} $\mu_{\Delta}\!\gtrsim\! T$
which will be studied below. This regime is divided into two qualitatively
different parts. In an intermediate regime $\mu_{\Delta}$ is still
linear in $d$ but the viscosity strongly rises. For $f\!\ll\!1$
this part allows an analytic solution given by \begin{align}
\zeta^{\sim} & =\zeta_{max}^{<}f\sum_{j=0}^{N}\frac{\left(2j+1\right)!!\chi_{j}}{2^{j-1}\left(j+1\right)!}d^{2j}\label{eq:madsen-approximation}\\
& =\frac{C^{2}\tilde{\Gamma}T^{\delta}}{\omega^{2}}\left(1+\sum_{j=1}^{N}\frac{\left(2j+1\right)!!\chi_{j}}{2^{j}\left(j+1\right)!}\biggl(\frac{C}{T}\frac{\Delta n}{\bar{n}}\biggr)^{\!2j}\right)\nonumber \end{align}
which can be combined with the subthermal result to give an approximate
analytic solution for both regions $\zeta^{\lesssim}\!=\!\zeta^{<}\!+\!\theta(T\!-\! T_{max})\zeta^{\sim}$.
At even higher amplitudes the rise of $\mu_{\Delta}$ becomes weaker
due to non-linear saturation effects so that the viscosity has a maximum
and decreases again. In the asymptotic limit $\mu_{\Delta}\!\gg\! T$,
the viscosity scales as\begin{equation}
\zeta\sim\left(\frac{\Delta n}{\bar{n}}\right)^{-\frac{2N}{2N+1}}\,.\label{eq:supra-limit}\end{equation}
As discussed in appendix \ref{sec:strange-matter-approximation},
for the special case of strange quark matter an approximate analytic
result is possible that includes this large amplitude regime.
In contrast to the bulk viscosity, the shear viscosity of dense matter
is independent of the frequency and amplitude of an external oscillation
and over certain temperature ranges its dependence on temperature
is approximately a simple power law. Shear viscosity becomes large
at low temperatures and therefore it is the dominant process for damping
of the r-modes of cooler stars. Correspondingly we can be parameterize
the shear viscosity as
\begin{equation}
\eta=\tilde{\eta}T^{-\sigma}\label{eq:shear-parametrization}\end{equation}
by simply factoring out the temperature dependence with exponent $\sigma$.
\subsection{Star models}
We study in this work the same model examples of compact stars as
in our companion paper \cite{Alford:2010fd}. These include neutron
stars, strange stars \cite{Witten:1984rs}, and hybrid stars, and
we consider in each case a star with a standard mass of $1.4\, M_{\odot}$
and a heavy star with a mass of $2\, M_{\odot}$. For the neutron
stars we use nuclear matter obeying the APR equation of state \cite{Akmal:1998cf}
and as a low density extension of the APR data we use \cite{Baym:1971pw,Negele:1971vb}.
In order to apply our general results to other equations of state,
away from chemical equilibrium we use the simple quadratic parameterization
in terms of the symmetry energy employed in \cite{Lattimer:1991ib}.
With the exception of ultra-heavy neutron stars the APR equation of
state allows only modified Urca processes \cite{Friman:1978zq,Sawyer:1989dp,Reisenegger:1994be}
which are the dominant microscopic processes that induce bulk viscosity.
Direct Urca processes \cite{Haensel:1992zz,Reisenegger:1994be} are
only possible in the core of a neutron star close to the mass limit
which we study in addition. The dominant contribution to the shear
viscosity in hadronic matter results from non-Fermi liquid enhanced
lepton-lepton scattering \cite{Shternin:2008es}.
For the strange stars and the quark core of the hybrid stars we use
quark matter obeying a simple equation of state in terms of parameters
$c$, $m_{s}$ and $B$,
\begin{align}
p_{par} & =\frac{1-c}{4\pi^{2}}\left(\mu_{d}^{4}+\mu_{u}^{4}+\mu_{s}^{4}\right)-\frac{3m_{s}^{2}\mu_{s}^{2}}{4\pi^{2}}\label{eq:quark-eos-model}\\
& \quad+\frac{3m_{s}^{4}}{32\pi^{2}}\left(3+4\log\!\left(\frac{2\mu_{s}}{m_{s}}\right)\right)-{\cal B}+\frac{\mu_{e}^{4}}{12\pi^{2}}\,.\nonumber \end{align}
This is a generalization of the parameterization employed in \cite{Alford:2004pf}
from which the equilibrium pressure is obtained via the conditions
of weak equilibrium $\mu_{s}\!=\!\mu_{d}\!=\!\mu_{u}\!+\!\mu_{e}$
and electrical neutrality. The shear viscosity in quark matter arises
from non-Fermi-liquid-enhanced quark-quark scattering \cite{Heiselberg:1993cr}
and the bulk viscosity from non-leptonic flavor-changing weak processes
\cite{Heiselberg:1992bd,Madsen:1992sx}. For our hybrid stars we make
the assumption of local electrical neutrality, excluding the possibility
of a mixed phase and its wealth of geometric structures.
The equilibrium star configuration is then determined by the general-relativistic
Tolman-Oppenheimer-Volkov (TOV) equations \cite{Tolman:1939jz}. Characteristics
of the examples that we consider are given in table \ref{tab:star-models}.
The susceptibilities $B$ and $C$ of the considered forms of matter
are given in table \ref{tab:strong-parameters}. Finally, the reequilibration
parameters $\tilde{\Gamma}$, $\delta$ and $\chi$ in the parameterization
of the weak rate eq.\,(\ref{eq:gamma-parametrization}) and the parameters
of the shear viscosity eq.\,(\ref{eq:shear-parametrization}) are
given for the considered forms of dense matter in table \ref{tab:weak-parameters}.
Note that there are higher non-linearities in the weak rate of hadronic
compared to quark matter, which in accordance with eq.\,(\ref{eq:madsen-approximation})
yields a steeper rise of the viscosity with amplitude so that suprathermal
effects are even more important in hadronic matter than in quark matter.
\begin{table}[H]
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& $M\left[M_{\odot}\right]$ & $M_{core}\left[M_{\odot}\right]$ & $R\left[km\right]$ & $n_{c}\left[n_{0}\right]$ & $\left\langle n\right\rangle \left[n_{0}\right]$ & $\Omega_{K}\left[kHz\right]$\tabularnewline
\hline
NS & $1.4$ & $\left(1.39\right)$ & $11.5$ & $3.43$ & $1.58$ & $6.02$\tabularnewline
\hline
& $2.0$ & $\left(1.99\right)$ & $11.0$ & $4.91$ & $2.46$ & $7.68$\tabularnewline
\hline
& $2.21$ & $0.85$ & $10.0$ & $7.17$ & $3.37$ & $9.31$\tabularnewline
\hline
SS & $1.4$ & $-$ & $11.3$ & $2.62$ & $1.91$ & $6.17$\tabularnewline
\hline
& $2.0$ & $-$ & $11.6$ & $4.95$ & $2.43$ & $7.09$\tabularnewline
\hline
HS & $1.4$ & $1.06$ & $12.7$ & $2.32$ & $1.17$ & $5.16$\tabularnewline
\hline
& $2.0$ & $1.81$ & $12.2$ & $4.89$ & $1.84$ & $6.62$\tabularnewline
\hline
\end{tabular}
\caption{\label{tab:star-models}Properties of the considered models of neutron
stars (NS), strange stars (SS) and hybrid stars (HS). We show the
mass of the star $M$, the mass of the core $M_{core}$, the radius
$R$, the baryon density at the center of the star $n_{c}$ given
in units of nuclear saturation density $n_{0}$, the average density
$\left\langle n\right\rangle $ and the Kepler frequency $\Omega_{K}$.
The neutron stars were obtained by solving the relativistic TOV equations
for catalyzed neutron matter using the APR equation of state \cite{Akmal:1998cf}
with low density extension \cite{Baym:1971pw,Negele:1971vb} and the
strange stars with a quark gas bag model with $c=0$, $m_{s}=150\,{\rm MeV}$
and a bag parameter $B=\left(138\,{\rm {\rm MeV}}\right)^{4}$. Large mass
hybrid stars are only found when strong interaction corrections are
considered, cf. \cite{Alford:2004pf}, and we find a $2\, M_{\odot}$
star for $c=0.4$, $m_{s}=140\,{\rm MeV}$, $B=\left(137\,{\rm MeV}\right)^{4}$.}
\end{table}
\begin{table*}
\begin{tabular}{|c|c|c|c|}
\hline
& $A$ & $B$ & $C$\tabularnewline
\hline
hadronic matter & $m_{N}\left(\frac{\partial p}{\partial n}\right)^{-1}$ & $\frac{8S}{n}\!+\negthinspace\frac{\pi^{2}}{\left(4\left(1\!-\!2x\right)S\right)^{2}}$ & $4\!\left(1\!-\!2x\right)\!\left(n\frac{\partial S}{\partial n}\!-\!\frac{S}{3}\right)$\tabularnewline
\hline
hadronic gas & $\frac{3m_{N}^{2}}{\left(3\pi^{2}n\right)^{2/3}}$ & $\frac{4m_{N}^{2}}{3\left(3\pi^{2}\right)^{1/3}n^{4/3}}$ & $\frac{\left(3\pi^{2}n\right)^{2/3}}{6m_{N}}$\tabularnewline
\hline
quark matter (gas: $c=0$) & $3\!+\!\frac{m_{s}^{2}}{\left(1\!-\! c\right)\mu_{q}^{2}}$ & $\frac{2\pi^{2}}{3(1-c)\mu_{q}^{2}}\left(1\!+\!\frac{m_{s}^{2}}{12(1-c)\mu_{q}^{2}}\right)$ & $-\frac{m_{s}^{2}}{3(1-c)\mu_{q}}$\tabularnewline
\hline
\end{tabular}
\caption{\label{tab:strong-parameters}Strong interaction parameters, defined
in eqs.\,(\ref{eq:susceptibilities}) and (\ref{eq:A-parameter}),
describing the response of the particular form of matter. In the case
of interacting hadronic matter a quadratic ansatz in the proton fraction
$x$ parameterized by the symmetry energy $S$ is employed. The expressions
for a hadron and quark gas are given to leading order in $n/m_{N}^{3}$
respectively next to leading order in $m_{s}/\mu$.}
\end{table*}
\begin{table*}
\begin{minipage}[t]{0.6\textwidth}%
\begin{tabular}{|c||c|c|c|c|c|}
\hline
Weak process & $\tilde{\Gamma}\,\bigl[\mathrm{MeV}^{(3-\delta)}\bigr]$ & $\delta$ & $\chi_{1}$ & $\chi_{2}$ & $\chi_{3}$\tabularnewline
\hline
\hline
quark non-leptonic & $6.59\!\times\!10^{-12}\,\Bigl(\frac{\mu_{q}}{300\,\mathrm{MeV}}\Bigr)^{5}$ & $2$ & $\frac{1}{4\pi^{2}}$ & $0$ & $0$\tabularnewline
\hline
hadronic direct Urca & $5.24\!\times\!10^{-15}\left(\!\frac{x\, n}{n_{0}}\!\right)^{\!1/3}$ & $4$ & $\frac{10}{17\pi^{2}}$ & $\frac{1}{17\pi^{4}}$ & $0$\tabularnewline
\hline
hadronic modified Urca & $4.68\!\times\!10^{-19}\left(\!\frac{x\, n}{n_{0}}\!\right)^{\!1/3}$ & $6$ & $\frac{189}{367\pi^{2}}$ & $\frac{21}{367\pi^{4}}$ & $\frac{3}{1835\pi^{6}}$\tabularnewline
\hline
\end{tabular}%
\end{minipage}%
\begin{minipage}[t]{0.4\textwidth}%
\begin{tabular}{|c||c|c|}
\hline
Strong/EM process & $\tilde{\eta}\,\bigl[\mathrm{MeV}^{(3+\sigma)}\bigr]$ & $\sigma$\tabularnewline
\hline
\hline
quark scattering & $1.98\!\times\!10^{9}\alpha_{s}^{-\frac{5}{3}}\left(\frac{\mu_{q}}{300\,\mathrm{MeV}}\right)\negthickspace^{\frac{14}{3}}$ & $\frac{5}{3}$\tabularnewline
\hline
leptonic scattering & $1.40\!\times\!10^{12}\left(\frac{x\, n}{n_{0}}\right)\negthickspace^{\frac{14}{9}}$ & $\frac{5}{3}$\tabularnewline
\hline
nn-scattering & $5.46\!\times\!10^{9}\left(\frac{n}{m_{N}n_{0}}\right)\negthickspace^{\frac{9}{4}}$ & $2$\tabularnewline
\hline
\end{tabular}%
\end{minipage}
\caption{\label{tab:weak-parameters}\emph{Left panel:} Parameters of the general
parameterization of the weak rate eq.\,(\ref{eq:gamma-parametrization})
for different processes of particular forms of matter which determine
the damping due to bulk viscosity. The coefficients $\chi_{i}$ parameterize
the non-linear dependence on the chemical potential fluctuation $\mu_{\Delta}$
arising in the suprathermal regime of the viscosity which is relevant
for large amplitude r-modes. \emph{Right panel:} Parameters arising
in the parameterization eq.~(\ref{eq:shear-parametrization}) of
the shear viscosity for different strong and electromagnetic interaction
processes. The leptonic and quark scatterings arise from a non-Fermi
liquid enhancement due to unscreened magnetic interactions. }
\end{table*}
\subsection{R-mode profile}
R-modes are normal modes of rotating stars and are obtained from a
linear perturbation analysis around the static star configurations
discussed above. We consider a star rotating with angular frequency
$\Omega$. To simplify this analysis it is performed in a Newtonian
approximation and in a slow rotation expansion in $\Omega^{2}/\left(\pi G_{N}\bar{\rho}\right)$,
where $\bar{\rho}$ is the average density of the unperturbed star.
The analysis of the damping due to bulk viscosity strictly requires
an expansion to next to leading order. The oscillation frequency depends
on the frame. For an r-mode with angular dependence $Y_{m+1}^{m}$
the frequency in a frame rotating with the unperturbed star $\omega_{r}$,
which is relevant for microscopic quantities like the viscosity, and
the corresponding frequency $\omega_{i}$, observed by an observer
in an inertial frame, are given by
\begin{equation}
\omega_{r}\equiv\omega=\kappa(\Omega)\Omega\quad,\quad\omega_{i}=\omega_{r}-m\Omega\label{eq:frequency-connection}\end{equation}
where the function $\kappa$ has an analog expansion and reads to
lowest order $\kappa_{0}\!=\!2/\left(m\!+\!1\right)$. For dissipation
via bulk viscosity the relevant quantity is the density compression
due to the r-mode with dimensionless amplitude $\alpha$. It vanishes
to leading order in the slow rotation expansion and reads at next
to leading order \cite{Lindblom:1999yk,Lindblom:1998wf}
\begin{align}
\left|\frac{\Delta n}{\bar{n}}\right| & \approx\sqrt{\frac{4m}{\left(m+1\right)^{3}\left(2m+3\right)}}\frac{2}{\left(m+1\right)\kappa(\Omega)}\alpha AR^{2}\Omega^{2}\label{eq:r-mode-profile}\\
& \quad\cdot\left(\left(\left(\frac{r}{R}\right)^{m+1}+\delta\Phi_{0}\right)\left|Y_{m+1}^{m}(\theta,\phi)\right|+\cdots\right)\nonumber \end{align}
where the ellipsis denotes further contributions involving next to
leading order corrections of the potentials that are explicitly given
in \cite{Lindblom:1999yk}. In this expression $A$ denotes the inverse
squared speed of sound
\begin{equation}
A\equiv\left.\frac{\partial\rho}{\partial p}\right|_{0}\label{eq:A-parameter}\end{equation}
evaluated at equilibrium. There are different conventions for the
amplitude $\alpha$ in the literature and we follow the convention
for the amplitude given in \cite{Lindblom:1998wf} but take into account
the corrections to the latter result in \cite{Lindblom:1999yk}. This
convention is usually used in the literature %
\footnote{We note that we had previously used the alternative $\alpha$-definition
\cite{Lindblom:1999yk} in the proceedings article \cite{Alford:2010ep}
and a preprint version of the present article, which led to lower
values for the saturation amplitudes, see also Appendix \ref{sec:R-mode-profile}.%
} and in this case the above expression breaks down for $\alpha\!>\! O\left(1\right)$.
For more details on the r-mode expression see Appendix \ref{sec:R-mode-profile}.
In general one must numerically solve a differential equation to obtain
$\delta\Phi_{0}$, which is the leading order correction to the gravitational
potential \cite{Alford:2010fd}. However, in the special case of a
star with a constant density profile $\delta\Phi_{0}$ can be shown
to be subleading compared to the first term in the inner parentheses
of eq.\,(\ref{eq:r-mode-profile}). For our numerical analysis below
we will make the approximation to neglect the additional second order
corrections in the slow rotation expansion given by the ellipsis in
eq. (\ref{eq:r-mode-profile}), which amounts to replacing the Lagrangian
perturbation by the Eulerian perturbation, but we include the second
order corrections to the frequency eq.\,(\ref{eq:frequency-connection}).
General analytic results showed that this is a good approximation
for the computation of the small amplitude instability regions \cite{Alford:2010fd}.
Here we will give semi-analytic results for the saturation amplitudes,
valid beyond leading order, that show the influence of the second
order terms.
We will see below that, in obtaining a precise assessment of the damping
time of large amplitude r-modes, the radial dependence of the density
perturbation plays a vital role. The radius enters eq.\,(\ref{eq:r-mode-profile})
explicitly and also via the density dependence of the inverse squared
speed of sound $A$ (fig.\,\ref{fig:dE/dp-density-dependence}) and
the radial density dependence of the star (fig.\,2 in \cite{Alford:2010fd}).
The radial variation of the density is moderate in a strange star,
but much more pronounced in neutron stars where the r-mode amplitude
grows strongly in the outer regions of the star.
\begin{figure}
\includegraphics{dE-dp}
\caption{\label{fig:dE/dp-density-dependence}The density dependence of the
inverse squared speed of sound $A\equiv d\rho/dp$ (which enters the
r-mode profile multiplicatively) for the different forms of matter
in table\,\ref{tab:strong-parameters} as well as generic polytropic
models. The solid line represents interacting APR matter, the dashed
line a hadron gas and the dotted line shows the result for a quark
gas. The structure at intermediate densities in the APR curve arises
from phase transitions and the use of finite differences to compute
the derivative, but due to the mild contribution of the denser inner
regions of the star to the damping these, as well as the known problem
that the APR equation of state becomes acausal at high density, have
no influence on our results below.}
\end{figure}
\section{Damping time scales\label{sec:Damping-time-scales}}
\subsection{General expressions}
The amplitude of the r-mode oscillations evolves with time dependence
$\exp(i\omega t-t/\tau)$. We can decompose the decay rate $1/\tau$
as \begin{equation}
\frac{1}{\tau(\Omega)}=\frac{1}{\tau_{G}(\Omega)}+\frac{1}{\tau_{B}(\Omega)}+\frac{1}{\tau_{S}(\Omega)}\end{equation}
Where $\tau_{G}$, $\tau_{B}$ and $\tau_{S}$ are gravitational
radiation, bulk viscosity and shear viscosity time scales, respectively.
The time scale of the r-mode growth due to gravitational wave emission
is given by \cite{Lindblom:1998wf}
\begin{equation}
\frac{1}{\tau_{G}}=-\frac{32\pi\left(m-1\right)^{2m}}{\left(\left(2m+1\right)!!\right)^{2}}\left(\frac{m+2}{m+1}\right)^{\!2m+2}\tilde{J_{m}}GMR^{2m}\Omega^{2m+2}\label{eq:gravitational-time}\end{equation}
and the damping time due to the shear viscosity can be written as
\cite{Alford:2010fd}
\begin{equation}
\frac{1}{\tau_{S}}=\sum_{l}\frac{\left(m-1\right)\left(2m+1\right)\tilde{S}_{m}^{\left(l\right)}\Lambda_{{\rm QCD}}^{3+\sigma}R}{\tilde{J}_{m}MT^{\sigma}}\label{eq:shear-viscosity-damping-time}\end{equation}
with the radial integral constants
\begin{align}
\tilde{J_{m}} & \equiv\frac{1}{MR^{2m}}\int_{0}^{R}\rho\left(r\right)r^{2m+2}dr\label{eq:J-tilde}\\
\tilde{S}_{m}^{\left(l\right)} & \equiv\frac{1}{R^{2m+1}\Lambda_{{\rm QCD}}^{3+\sigma}}\int_{R_{i}^{\left(l\right)}}^{R_{o}^{\left(l\right)}}\tilde{\eta}r^{2m}dr\label{eq:S-tilde}\end{align}
where $\rho$ is the energy density and $\Lambda_{{\rm QCD}}$ a generic
QCD scale introduced to make the constant dimensionless. When several
shells with inner radii $R_{i}^{\left(l\right)}$and outer radii $R_{o}^{\left(l\right)}$
and distinct phases and/or transport coefficients are present in a
compact star, the damping time integral over the star decomposes into
a sum of contributions of the individual shells $l$. The bulk viscosity
damping time is given by \cite{Lindblom:1999yk}
\begin{align}
\frac{1}{\tau_{B}}= & \frac{\kappa^{2}}{\alpha^{2}\tilde{J}_{m}MR^{2}}\int d^{3}x\Bigl|\frac{\Delta\rho}{\bar{\rho}}\Bigr|^{2}\zeta\Bigl(\Bigl|\frac{\Delta\rho}{\bar{\rho}}\Bigr|^{2}\Bigr)\label{eq:bulk-damping-time}\end{align}
Using the general expression for the bulk viscosity eqs.\,(\ref{eq:general-viscosity})
and (\ref{eq:viscosity-integral}) and expressing the fluctuation
in the conserved energy density $\Delta\rho/\bar{\rho}$ by the same
fluctuation of the baryon density $\Delta n/\bar{n}$, gives for the
damping time the general expression
\begin{align}
\frac{1}{\tau_{B}} & =\frac{4\pi\Omega^{3}}{(m+1)^{2}\tilde{J}_{m}MR^{2}\kappa}\sum_{l}{\cal T}^{\left(l\right)}(a,b)\,.\label{eq:general-bulk-viscosity-time}\end{align}
in terms of integrals over the individual shells. Defining the reduced
density oscillation
\begin{equation}
\Bigl(\frac{\Delta n}{\bar{n}}\Bigr)_{\mathrm{red.}}\equiv\frac{\left(m+1\right)\kappa(\Omega)}{2\alpha\Omega^{2}}\frac{\Delta n}{\bar{n}}\end{equation}
we can write
\begin{align}
& {\cal T}^{\left(l\right)}(a,b)\equiv\int_{R_{i}^{\left(l\right)}}^{R_{o}^{\left(l\right)}}\! dr\, r^{2}\int\! d\theta\sin\theta\left|\Bigl(\frac{\Delta n}{\bar{n}}\Bigr)_{\mathrm{red.}}(r,\theta)\right|^{2}\frac{C(r)^{2}}{B(r)}\nonumber \\
& \cdot\tilde{{\cal I}}\left(aC(r)B(r)^{1/\delta}\tilde{\Gamma}(r)^{1/\delta}\left|\Bigl(\frac{\Delta n}{\bar{n}}\Bigr)_{\mathrm{red.}}(r,\theta)\right|,\, bB(r)\tilde{\Gamma}(r)\right)\label{eq:T-function}\end{align}
depending on only two independent parameters
\begin{align*}
a & \equiv\frac{\kappa_{0}\alpha\Omega^{2}}{\kappa\omega^{1/\delta}}=\frac{2\alpha\Omega^{\left(2\delta-1\right)/\delta}}{\left(m+1\right)\kappa^{\left(\delta+1\right)/\delta}}\\
b & \equiv\frac{T^{\delta}}{\omega}=\frac{T^{\delta}}{\kappa\Omega}\end{align*}
Note that in eq.\,(\ref{eq:T-function}) all local quantities can
have different functional forms in different shells, as given in tables
\ref{tab:strong-parameters} and \ref{tab:weak-parameters}, but to
make the expression readable we do not show the explicit suffixes
$(l)$. Whereas strange stars are basically homogeneous and consist
of a single phase, the crust of neutron and hybrid stars is extremely
inhomogeneous and complicated. Although there are no free protons
and thereby no Urca processes, the ultra-heavy nuclei present in the
inner crust as well as the clusters in potential pasta phases still
feature analogous beta-processes. Since oscillations likewise push
the system out of beta-equilibrium an analogous suprathermal enhancement
of the bulk viscosity contribution from these phases is expected.
However, there are to our knowledge no results for the bulk viscosity
in the inner crust, yet \cite{Chamel:2008ca}. Therefore we will in
our numeric results given below neglect the contribution from the
crust and only include the contribution from the core. The core does
not have a sharply-defined boundary: we chose it conventionally to
be at baryon density $n_{0}/4$ corresponding to the lowest point
in the APR table, but check the dependence on this choice.
\subsection{Approximate limits of the bulk viscosity damping time}
In the subthermal regime $\mu_{\Delta}\!\ll\! T$ the bulk viscosity
eq.\,(\ref{eq:sub-viscosity}) is independent of the r-mode amplitude,
so the angular integral in eq.\,(\ref{eq:bulk-damping-time}) is
trivial. The damping time in the subthermal regime is then given by
\begin{equation}
\frac{1}{\tau_{B}^{<}}=\frac{16m}{(2m+3)(m+1)^{5}\kappa}\frac{R^{5}\Omega^{3}}{\tilde{J}_{m}M}\sum_{l}{\cal T}_{m}^{<\left(l\right)}\Bigl(\frac{T^{\delta}}{\kappa\Omega}\Bigr)\label{eq:subthermal-damping-time}\end{equation}
in terms of the one dimensional radial integral\[
{\cal T}_{m}^{<\left(l\right)}(b)\equiv\frac{b}{R^{3}}\int_{R_{i}^{\left(l\right)}}^{R_{o}^{\left(l\right)}}dr\, r^{2}\frac{A^{2}C^{2}\tilde{\Gamma}}{1+\tilde{\Gamma}^{2}B^{2}b^{2}}\left(\left(\frac{r}{R}\right)^{m+1}+\delta\Phi_{0}\right)^{2}\]
This expression was used to study the small amplitude instability
regions in \cite{Alford:2010fd}. Here we want to study the large-amplitude
saturation and therefore it is useful to obtain an analytic expression
that includes the large-amplitude enhancement of the bulk viscosity.
In the intermediate, linear regime and for $f\!\ll\!1$ the general
analytic approximation for the bulk viscosity eq.\,(\ref{eq:madsen-approximation})
is valid. Since this local condition has to be fulfilled everywhere
in the star, the global parameters $a$ and $b$ must be smaller than
certain bounds that are determined by the particular properties of
the considered stars. We recall from \cite{Alford:2010gw} that the
approximation is particularly useful for hadronic matter with modified
Urca processes where it covers almost the entire range of physical
local density amplitudes at millisecond frequencies. A plot of the
regions of validity of the individual analytic approximations of the
bulk viscosity for different forms of matter is given in \cite{Alford:2010ht}.
Analogous to the low temperature/high frequency approximation in the
subthermal regime eq. (\ref{eq:subthermal-damping-time}), in the
intermediate, linear regime an explicit evaluation is possible. With
the analytic expression for the bulk viscosity eq. (\ref{eq:madsen-approximation})
the angular integrals over the spherical harmonics arising in the
r-mode profile take the form
\[
\int\!\! d\Omega_{\theta\phi}\left|Y_{m+1}^{m}(\theta,\phi)\right|^{2n}\!=\!\frac{4\pi\left(2n-1\right)!!(m\, n)!}{(2(m+1)n+1)!!}\Bigl(\frac{\left(2m+3\right)!!}{4\pi\, m!}\Bigr)^{n}\]
and this yields a result that apart from the evaluation of the remaining
radial integrals is analytic (see also \cite{Reisenegger:2003pd,Bonacic:2003th})
\begin{widetext}
\begin{equation}
\frac{1}{\tau_{B}^{\sim}}=\frac{16\left(2m+1\right)!!\Lambda_{QCD}^{9-\delta}R^{5}T^{\delta}\Omega^{2}}{(m+1)^{5}(m-1)!\kappa^{2}\tilde{J}_{m}\Lambda_{EW}^{4}M}\sum_{j=0}^{N}\frac{((2j+1)!!)^{2}(m(j+1))!\chi_{j}\tilde{V}_{m,j}}{(j+1)!\left(2(m+1)(j+1)+1\right)!!}\left(\frac{2m\left(2m+1\right)!!}{\pi\left(m+1\right)^{5}m!\kappa^{2}}\frac{\Lambda_{QCD}^{2}R^{4}\alpha^{2}\Omega^{4}}{T^{2}}\right)^{j}\label{eq:linear-bulk-vicosity-time}\end{equation}
\end{widetext}The dependence on all local parameters, like the equation
of state, the weak rate, the density dependence of the particular
star and its r-mode profile is contained in a few dimensionless radial
integral constants ($\Lambda_{EW}$ is a typical electroweak scale)
\begin{align}
& \tilde{V}_{m,j}\equiv\frac{\Lambda_{EW}^{4}}{\Lambda_{QCD}^{7-\delta+2(j+1)}R^{3}}\label{eq:V-tilde}\\
& \cdot\int_{0}^{{\cal R}}\!\! dr\, r^{2}\tilde{\Gamma}\!\left(r\right)\!\biggl(A\!\left(r\right)C\!\left(r\right)\biggl(\left(\frac{r}{R}\right)^{m+1}\!+\!\delta\Phi_{0}\!\left(r\right)\biggr)\biggr)^{2(j+1)}\nonumber \end{align}
but the dependence on the parameters of the r-mode evolution $\Omega$,
$\alpha$ and $T$ is entirely explicit in eq.\,(\ref{eq:linear-bulk-vicosity-time}).
The $j\!=\!0$ term in eq.\,(\ref{eq:linear-bulk-vicosity-time})
is precisely the approximate subthermal result eq.\,(\ref{eq:subthermal-damping-time})
in the considered approximation. The constants $\tilde{V}_{j}\!\equiv\!\tilde{V}_{2,j}$
for the fundamental r-mode are given for several stars in table \ref{tab:parameters-amplitude-values}.
Although these parameters can vary significantly for different stars,
it is quite striking that as far as the bulk viscosity is concerned,
the complex details of the individual stars are encoded in a few constants.
Note in particular that the parametric form eq.\,(\ref{eq:linear-bulk-vicosity-time})
remains valid for the full second order r-mode expression and only
the constants eq.\,(\ref{eq:V-tilde}) are changed. At sufficiently
large amplitudes the largest power in the sum in eq.\,(\ref{eq:linear-bulk-vicosity-time})
dominates and due to the connection $\delta\!=\!2N$ the bulk viscosity
damping time becomes temperature independent in this approximation.
Note also that the integrals $\tilde{V}_{m,N}$, as well as the general
expression eq.\,(\ref{eq:T-function}) feature an extremely pronounced
radial dependence, both due to the explicit radial dependence and
the radial dependence of the inverse squared speed of sound $A$,
that strongly weights the outer parts of the star.
\begin{table*}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
star model & shell & $\tilde{J}$ & $\tilde{S}$ & $\tilde{V}_{0}$ & $\tilde{V}_{1}$ & $\tilde{V}_{2}$ & $\tilde{V}_{3}$ & $\alpha_{sat}\left(\Omega_{K}\right)$\tabularnewline
\hline
NS $1.4\, M_{\odot}$ & core & $1.81\times10^{-2}$ & $7.68\times10^{-5}$ & $1.31\times10^{-3}$ & $4.24\times10^{-3}$ & $2.02\times10^{-2}$ & $0.105$ & $3.68$\tabularnewline
\cline{1-1} \cline{4-9}
NS $1.4\, M_{\odot}$ gas & & & $4.32\times10^{-6}$ & $1.28\times10^{-4}$ & $5.52\times10^{-5}$ & $3.88\times10^{-5}$ & $3.03\times10^{-5}$ & $14.3$\tabularnewline
\cline{1-1} \cline{3-9}
NS $2.0\, M_{\odot}$ & & $2.05\times10^{-2}$ & $2.25\times10^{-4}$ & $1.16\times10^{-3}$ & $4.92\times10^{-3}$ & $3.25\times10^{-2}$ & $0.238$ & $2.52$\tabularnewline
\hline
NS $2.21\, M_{\odot}$ & d.U. core & $2.02\times10^{-2}$ & $5.05\times10^{-4}$ & $1.16\times10^{-8}$ & $7.24\times10^{-12}$ & $5.39\times10^{-15}$ & $-$ & $-$\tabularnewline
\cline{2-2} \cline{5-9}
& m.U. core & & & $9.34\times10^{-4}$ & $4.42\times10^{-3}$ & $3.39\times10^{-2}$ & $0.288$ & $2.60$\tabularnewline
\hline
SS eq.\,(\ref{eq:quark-eos-model}) & all & $\frac{3}{28\pi}$ & $\frac{\hat{\eta}\mu_{q}^{14/3}}{5\Lambda_{QCD}^{14/3}\alpha_{s}^{5/3}}$ & $\frac{\Lambda_{EW}^{4}\hat{\Gamma}m_{s}^{4}\mu_{q}^{3}}{9\Lambda_{QCD}^{7}(1-c)^{2}}$ & $\frac{\Lambda_{EW}^{4}\hat{\Gamma}m_{s}^{8}\mu_{q}}{15\Lambda_{QCD}^{9}(1-c)^{4}}$ & $-$ & $-$ & eq.\,(\ref{eq:analytic-quark-sat-amp})\tabularnewline
\cline{1-1} \cline{3-9}
SS $1.4\, M_{\odot}$ & & $3.08\times10^{-2}$ & $3.49\times10^{-6}$ & $3.53\times10^{-10}$ & $1.24\times10^{-12}$ & $-$ & $-$ & $1.16$\tabularnewline
\cline{1-1} \cline{3-9}
SS $2.0\, M_{\odot}$ & & $2.65\times10^{-2}$ & $4.45\times10^{-6}$ & $3.58\times10^{-10}$ & $9.70\times10^{-13}$ & $-$ & $-$ & $1.56$\tabularnewline
\hline
HS $1.4\, M_{\odot}$ & quark core & $1.70\times10^{-2}$ & $3.11\times10^{-6}$ & $1.38\times10^{-10}$ & $1.75\times10^{-13}$ & $-$ & $-$ & $2.25$\tabularnewline
\cline{2-2} \cline{4-9}
& hadr. core & & $9.71\times10^{-7}$ & $1.39\times10^{-3}$ & $4.70\times10^{-3}$ & $2.23\times10^{-2}$ & $0.116$ & $3.66$\tabularnewline
\hline
HS $2.0\, M_{\odot}$ & quark core & $2.00\times10^{-2}$ & $5.25\times10^{-6}$ & $3.76\times10^{-10}$ & $7.75\times10^{-13}$ & $-$ & $-$ & $1.59$\tabularnewline
\cline{2-2} \cline{4-9}
& hadr. core & & $5.24\times10^{-6}$ & $1.07\times10^{-3}$ & $4.12\times10^{-3}$ & $2.31\times10^{-2}$ & $0.134$ & $2.94$\tabularnewline
\hline
\end{tabular}
\caption{\label{tab:parameters-amplitude-values}Radial integral parameters
and static saturation amplitude of a $m\!=\!2$ r-mode for the stars
considered in this work. The constant $\tilde{J}$, $\tilde{S}$ and
$\tilde{V_{i}}$ are given by eqs.\,(\ref{eq:J-tilde}), (\ref{eq:S-tilde})
and (\ref{eq:V-tilde}), respectively, using the generic normalization
scales $\Lambda_{QCD}=1$ GeV and $\Lambda_{EW}=100$ GeV. Note that
the subthermal parameter $\tilde{V}_{0}$ corresponds to $\tilde{V}$
in \cite{Alford:2010fd} where the subscript was omitted for simplicity
and the strange star expressions are given to leading order in $m_{s}/\mu$.}
\end{table*}
\subsection{Results for the damping times}
Using the expressions for the microscopic parameters given in tables
\ref{tab:strong-parameters} and \ref{tab:weak-parameters} in the
general expressions eqs. (\ref{eq:gravitational-time}), (\ref{eq:shear-viscosity-damping-time})
and (\ref{eq:general-bulk-viscosity-time}) we obtain the gravitational
and viscosity time scales as a function of their dependent macroscopic
parameters. These are given in fig.\,\ref{fig:damping-times} for
the cases of a neutron star with damping due to modified Urca reactions
(left panel) and a strange star (right panel) by the solid lines as
a function of temperature and for different amplitudes ranging from
top to bottom from the subthermal result at infinitesimal amplitude
to the extreme case $\alpha\!=\!10.$
At sufficiently low temperature the strong increase of the bulk viscosity
with the (local) amplitude $\Delta n/\bar{n}$ damps r-modes with
large (global) dimensionless amplitude $\alpha$ at significantly
shorter time scales. As found before from the analytic expression
eq.\,(\ref{eq:linear-bulk-vicosity-time}), given by the dotted curves,
the damping time is temperature independent in this low temperature
and intermediate amplitude regime. In contrast, due to the generic
form of the bulk viscosity, featuring a universal maximum, the damping
of large amplitude r-modes is not enhanced at high temperatures. As
a direct consequence of the subthermal maximum of the bulk viscosity
\cite{Alford:2010gw}, the corresponding {}``resonant'' temperature
where the damping time is minimal is at roughly $10^{9}$ K for strange
stars and $10^{11}$ K for neutron stars. Correspondingly, r-modes
are entirely unstable at high temperatures. However, due to the strong
suprathermal enhancement at low temperatures the damping undercuts
the gravitational time scale at sufficiently large amplitude so that
the r-mode growth will slow down and eventually saturate. The corresponding
amplitudes are strikingly very similar for the two different classes
of stars, as will be discussed in more detail below. At very large
amplitudes $\alpha\!\sim\! O\left(10\right)$ the damping times decrease
again at all temperatures as a consequence of the behavior of the
bulk viscosity \cite{Alford:2010gw}.
The dot-dashed curves in the neutron star plot on the left panel of
fig.\,\ref{fig:damping-times} show the damping time if the crust
is assumed to start already at the higher density $n_{0}/2$ instead
of $n_{0}/4$ so that only the contribution from the correspondingly
smaller core is taken into account%
\footnote{The range $n_{0}/4$ to $n_{0}/2$ should provide an estimate for
the uncertainty of this boundary. For instance in \cite{Chamel:2008ca}
an intermediate value of $n_{0}/3$ is given.%
}. Although the damping times are larger, as expected, the amplitude
at which the viscous damping can saturate the mode is not drastically
changed, so that our results given below remain qualitatively unchanged
in this case. Actually, when the damping from the crust would be properly
taken into account this should rather enhance the damping and decrease
the r-mode amplitude compared to those obtained in this work. In the
case of the strange star on the right panel of fig.\,\ref{fig:damping-times}
the dashed curves also show the analytic approximation discussed in
appendix \ref{sec:strange-matter-approximation}, where the star is
assumed to be homogeneous. As can be seen, the corresponding expression
eq. (\ref{eq:homogenous-quark-damping-time}) gives an approximate
estimate for the damping time at all temperatures and for amplitudes
up to the maximum of the viscosity, and only fails at higher amplitudes,
where the bulk viscosity cannot saturate the r-mode anymore and where
it is thereby not physically relevant. The deviations compared to
the numeric result stem mainly from the fact that the density in the
strange star is not entirely constant (fig.\,2 in \cite{Alford:2010fd}).
Previous neutron star analyses \cite{Reisenegger:2003pd,Bonacic:2003th}
have employed an r-mode profile that does not feature the stronger
additional radial dependence due to the low density enhancement of
the inverse squared speed of sound $A$, shown in fig.\,\ref{fig:dE/dp-density-dependence},
for a realistic equation of state. E.g. the r-mode profile given in
eq.\,(6.6) of \cite{Lindblom:2001hd} features roughly a generic
$r^{3}$-dependence. The strong r-dependence in our present treatment,
however, strongly amplifies the damping in the outer regions of the
star. Therefore, the contribution of the crust to the viscous damping
should be relevant and would further decrease the saturation amplitude.
The current restriction of our analysis to the core presents therefore
an upper bound for the saturation amplitude obtained when the damping
of the whole star is considered. The second order effects in contrast
increase the small amplitude instability region \cite{Lindblom:1999yk}
and can correspondingly be expected to likewise increase the saturation
amplitude. A more thorough treatment of all these effects in the future
is clearly desirable.
\begin{figure*}
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.85]{damping-times-APR-1. 4}%
\end{minipage}%
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.85]{damping-times-Bag-1. 4}%
\end{minipage}
\caption{\textcolor{green}{\label{fig:damping-times}}The relevant r-mode time
scales for $1.4\, M_{\odot}$ stars rotating at their Kepler frequency.
\emph{Left panel: }Neutron star. \emph{Right panel: }Strange star.
The dotted horizontal line presents the time scale $\tau_{G}$ associated
to the growth of the mode due to gravitational wave emission. The
dashed rising curve shows the damping time $\tau_{S}$ due to shear
viscosity. The damping time $\tau_{B}$ due to bulk viscosity is given
for different dimensionless r-mode amplitudes $\alpha=0$, $0.01$,
$0.1$, $1$ and $10$ by the solid curves. The thin dotted curves
correspond to the analytic linear approximation eq. (\ref{eq:linear-bulk-vicosity-time})
and are below the shown plot range for the largest amplitude. The
thin dot-dashed curves on the left panel show the change when only
a smaller core (ranging to a density of $n_{0}/2$ instead of $n_{0}/4$)
is taken into account. The thin dashed curves on the right panel represent
the approximate analytic expression eq. (\ref{eq:homogenous-quark-damping-time})
given in the appendix which is not valid above the maximum of the
bulk viscosity and therefore not shown for the large amplitude results.}
\end{figure*}
\section{Saturation amplitudes}
Because of the strong decrease of the viscous damping time due to
the suprathermal enhancement the damping can dominate at sufficiently
large amplitudes. In this case the definition of the instability regions
have to be extended. The latter are standardly defined in the subthermal
regime and are independent of the amplitude. One could extend this
concept by the definition of amplitude dependent instability regions
which would shrink with increasing amplitude. However, since the amplitude
can neither be inferred from observation nor is it a parameter that
can be dialed, but is rather determined dynamically, we refrain from
this possibility and rather introduce the concept of a \emph{static
saturation amplitude}. The latter is defined by the amplitude at which
the r-mode would saturate at fixed temperature and frequency and is
given by the solution of the equation
\begin{equation}
\frac{1}{\tau_{G}(\Omega)}+\sum_{l}\left(\frac{1}{\tau_{S}^{\left(l\right)}(T)}+\frac{1}{\tau_{B}^{\left(l\right)}(T,\Omega,\alpha_{sat})}\right)=0\label{eq:saturation-condition}\end{equation}
where $l$ runs over the contributions from the different shells of
the star. The boundaries of the above mentioned amplitude dependent
instability regions are by definition simply the contour lines $\alpha_{sat}(T,\Omega)\!=$const.
and the boundary of the classic instability region, in particular,
corresponds to $\alpha_{sat}=0$. In case several solutions of eq.\,(\ref{eq:saturation-condition})
exist, only the smallest one is physical and if no solution exists
then viscous damping alone cannot saturate the r-mode according to
this definition. Actually, at the same time the r-mode grows, the
star generally also cools or reheats and spins down respectively up
so that the above amplitudes do not have to be reached. In particular
the star could leave the parameter regions where a saturation according
to the above criterion is not possible before the r-mode can actually
explode.
\subsection{Analytic approximation}
Similar to the analytic expression for the boundary of the instability
region \cite{Alford:2010fd} an analytic expression for the static
saturation amplitude can be obtained. The analytic linear approximation
$\tau_{B}^{\sim}$ applies as long as the r-mode amplitude is small
enough so that the bulk viscosity is sufficiently below its maximum.
When the r-mode amplitude is at the same time large enough that the
highest power in eq. (\ref{eq:linear-bulk-vicosity-time}) dominates,
the damping time becomes temperature independent. Both conditions
are met sufficiently far inside the instability region so that the
damping time simplifies to
\begin{widetext}
\begin{align*}
\frac{1}{\tau_{B}^{\sim}} & \xrightarrow[j=N=\delta/2]{}\frac{2^{4+N}m^{N}\left(\left(2m+1\right)!!\right)^{N+1}\left(\left(2N+1\right)!!\right)^{2}(m\left(N+1\right))!\chi_{N}\tilde{V}_{m,N}}{\pi^{N}\left(m+1\right)^{5\left(N+1\right)}\left(m!\right)^{N}\left(m-1\right)!\left(N+1\right)!\left(2\left(m+1\right)\left(N+1\right)+1\right)!!\kappa^{2N+2}\tilde{J}_{m}}\frac{\Lambda_{{\rm QCD}}^{9}R^{5+4N}\alpha^{2N}\Omega^{4N+2}}{\Lambda_{EW}^{4}M}\end{align*}
At saturation this has to match the gravitational time scale $1/\tau_{B}+1/\tau_{G}=0$
which yields the general result
\begin{align}
\alpha_{sat}= & \left(\frac{\pi^{1+\delta/2}\left(m-1\right)^{2m}\left(m+1\right)^{3+5\delta/2-2m}\left(m+2\right)^{2+2m}\left(\left(m-1\right)!\right)^{1+\delta/2}\left(\frac{\delta}{2}+1\right)!\left(2\left(m+1\right)\left(\frac{\delta}{2}+1\right)+1\right)!!\kappa^{\delta+2}}{2^{\delta/2-1}\left(\left(2m+1\right)!!\right)^{3+\frac{\delta}{2}}\left(\left(\delta+1\right)!!\right)^{2}(m\left(\frac{\delta}{2}+1\right))!\chi_{\frac{\delta}{2}}}\right)^{1/\delta}\nonumber \\
& \;\cdot\frac{\tilde{J}_{m}^{2/\delta}\Lambda_{EW}^{4/\delta}G^{1/\delta}M^{2/\delta}}{\tilde{V}_{m,\frac{\delta}{2}}^{1/\delta}\Lambda_{QCD}^{9/\delta}R^{2+\left(5-2m\right)/\delta}\Omega^{2-2m/\delta}}\label{eq:analytic-saturation-amplitude}\end{align}
\end{widetext}where $\kappa$ is defined by eq. (\ref{eq:frequency-connection}).
In the cases of strange stars with non-leptonic processes $\delta=2$
and hadronic matter with modified Urca processes $\delta=6$ this
gives for the $m=2$ r-mode
\begin{align}
\alpha_{sat}^{\left(SS\right)} & \approx5.56\cdot10^{-5}\frac{\tilde{J}}{\tilde{V}_{1}^{1/2}}\frac{M_{1.4}}{R_{10}^{5/2}}\approx1.61\frac{\left(1-c\right)^{2}M_{1.4}}{m_{150}^{4}\mu_{300}^{1/2}R_{10}^{5/2}}\label{eq:analytic-quark-sat-amp}\\
\alpha_{sat}^{\left(NS\right)} & \approx10.8\frac{\tilde{J}^{1/3}}{\tilde{V}_{3}^{1/6}}\frac{M_{1.4}^{1/3}}{R_{10}^{13/6}\Omega_{ms}^{4/3}}\nonumber \end{align}
where $\tilde{J}\!\equiv\!\tilde{J}_{2}$ and $\tilde{V}_{i}\!\equiv\!\tilde{V}_{2,i}$
are given for the normalization scales used in table \ref{tab:parameters-amplitude-values}.
Here $m_{150}$, $\mu_{300}$, $M_{1.4}$, $R_{10}$ and $\Omega_{ms}$
are the effective strange quark mass in units of 150 MeV, the quark
chemical potential in units of 300 MeV, the stars mass in units of
$1.4\, M_{\odot}$, the radius in units of $10$ km and the angular
velocity in units of $2\pi$ kHz corresponding to a millisecond pulsar,
respectively. For strange stars the above expression for the intermediate
amplitude bulk viscosity damping time has the same frequency dependence
as the gravitational time scale eq. (\ref{eq:gravitational-time}),
so $\alpha_{sat}$ is basically constant throughout the instability
region. However, it rises with decreasing frequency for neutron stars
where the frequency dependence of the bulk viscosity is weaker. In
contrast to the analytic expressions for the extrema of the instability
region given in \cite{Alford:2010fd}, which are very insensitive
to the microscopic transport parameters, the saturation amplitude
is more sensitive to the suprathermal bulk viscosity parameter $\tilde{V}$.
Whereas the saturation amplitude of neutron stars still depends on
$\tilde{V}$ rather mildly due to the power $1/6$, the saturation
amplitude for strange stars obtained from the generic equation of
state eq.\,(\ref{eq:quark-eos-model}) decreases with the {}``interaction
parameter'' $c$ and even more strongly with the effective strange
quark mass $m_{s}$.
\subsection{Numeric solution}
Let us now discuss the numerical solution for the saturation amplitude.
In the following plots figs.\,\ref{fig:saturation-amplitude-APR}
to \ref{fig:instability-regions-multipoles} the static saturation
amplitude is shown as a function of temperature and amplitude and
they feature generally 3 distinct regions. The light (blue) surface
shows the saturation amplitude where the r-mode growth is stopped
by suprathermal damping. Due to the characteristic behavior of the
bulk viscosity \cite{Alford:2010gw} which does not feature an amplitude
enhancement for temperatures above the temperature $T_{max}$ eq.
(\ref{eq:maximum-temperature}), the r-mode is not damped at all by
viscous effects in the high temperature regime as denoted by the dark
(red) area on the right hand side. In the flat (green) region surrounding
the instability region the r-mode is entirely stable and already damped
by the shear or the subthermal bulk viscosity so that $\alpha_{sat}\!=\!0$.
\begin{figure*}
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.72]{had-sat-amp-1. 4}%
\end{minipage}%
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.72]{had-sat-amp-2. 0}%
\end{minipage}
\caption{\label{fig:saturation-amplitude-APR}The static saturation amplitude,
at which the r-mode growth is stopped by suprathermal viscous damping
for the APR neutron stars. \emph{Left panel:} $1.4\, M_{\odot}$ .
\emph{Right panel:} $2.0\, M_{\odot}$. The light (green) shaded area
denotes the stable region where the r-mode is damped away. At large
frequencies a plateau with amplitudes $O\!\left(1\right)$ is reached.
In the dark (red) region at high temperatures the r-mode is entirely
unstable and cannot be saturated by viscous effects.}
\end{figure*}
The left panel of fig.\,\ref{fig:saturation-amplitude-APR} shows
the static saturation amplitude for the $m=2$ r-mode of a $1.4\, M_{\odot}$
neutron star. Although this might be hard to see in certain regions
of the plot, the saturation amplitude rises steeply, within a narrow
interval, from zero at the boundary of the instability region towards
its interior. At large frequencies it reaches a plateau value that
is nearly independent of the temperature as predicted by the analytic
expression eq.\,(\ref{eq:analytic-saturation-amplitude}). As described
by the latter expression, inside the instability region the saturation
amplitude rises strongly with decreasing frequency and since the instability
region shrinks in width and eventually ends at low frequencies where
the amplitude vanishes and the mode is damped, it features a peak-like
structure. The maximum static saturation amplitude reached at the
peak is in this case unphysically large whereas the plateau value
at the Kepler frequency is still roughly $3.5$. The suprathermal
bulk viscosity can therefore in principle saturate r-modes within
the lower part of the instability region. However, in the present
case, where only the damping of the core is taken into account, the
static saturation amplitudes are at the limit where a standard r-mode
analysis is valid. Moreover these amplitudes are so far larger than
those of alternative saturation mechanisms \cite{Bondarescu:2007jw,Arras:2002dw,Gressman:2002zy,Lin:2004wx}.
It is interesting to mention once more, though, that the extreme radial
dependence of the r-mode profile eq. (\ref{eq:r-mode-profile}) strongly
weights the outer regions of the star due to power law dependences
of the inverse bulk viscosity damping time eq. (\ref{eq:bulk-damping-time})
with exponents $O\left(30\right)$ for neutron and hybrid stars which
is further enhanced by the density dependence of the inverse speed
of sound. The contribution of the crust could thereby be decisive
to obtain a realistic estimate of the impact of the non-linear viscosity.
In this context it is also important that a similar enhancement of
the bulk viscosity has been found for superfluid matter \cite{Alford:2011df}.
As noted in \cite{Alford:2010fd} there is a second instability region
at high temperatures above $10^{11}$ K and as argued above the suprathermal
bulk viscosity cannot saturate the r-mode in this high temperature
regime. It is an interesting question if the r-mode can become large
during this initial part of a star's evolution, and if so whether
the r-mode is saturated at sufficiently small values by other non-linear
mechanism or if the r-mode growth is not stopped before it reaches
the regime where the structural stability of the star is at stake.
In the latter case this instability phase might extend the violent
supernova stage and actively shape the remnant by additional mass
shedding and thereby determine its initial size and angular momentum.
We will discuss these points further in the conclusion.
\begin{figure*}
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.72]{bag-sat-amp-1. 4}%
\end{minipage}%
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.72]{bag-sat-amp-2. 0}%
\end{minipage}
\caption{\label{fig:saturation-amplitude-Bag}The saturation amplitude for
the considered strange stars. \emph{Left panel:} $1.4\, M_{\odot}$
. \emph{Right panel:} $2.0\, M_{\odot}$ . In the latter case the
suprathermal viscosity cannot stop the r-mode instability at frequencies
larger than the maximum frequency of the stability window (where the
saturation amplitude diverges) - for the considered star slightly
below the Kepler frequency - as well as in the high temperature part
of the instability region. The saturation amplitudes of the plateau
in the lower part of the instability region are of the same order
as in the hadronic case shown in fig.\,\ref{fig:saturation-amplitude-APR}.}
\end{figure*}
The saturation amplitude for the $1.4\, M_{\odot}$ strange star is
shown on the left panel of fig.\,\ref{fig:saturation-amplitude-Bag}.
Since the maximum of the stability window is above the Kepler frequency
there are in this plot two separate parts of the instability region.
As predicted by eq.\,(\ref{eq:analytic-saturation-amplitude}) the
plateau value of the saturation amplitude in the lower part is approximately
temperature and frequency independent. Strikingly it is of similar
size as the saturation amplitude for the $1.4\, M_{\odot}$ neutron
star at its Kepler frequency. The high temperature part of the instability
region where the viscosity again cannot saturate the r-mode extends
in this case down to lower temperatures than for neutron stars.
\begin{figure*}
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.72]{hyb-sat-amp-1. 4}%
\end{minipage}%
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.72]{hyb-sat-amp-2. 0}%
\end{minipage}
\caption{\label{fig:saturation-amplitude-Hybrid}The saturation amplitude for
the considered hybrid stars. \emph{Left panel:} $1.4\, M_{\odot}$
. \emph{Right panel:} $2.0\, M_{\odot}$ . The saturation in the low
temperature part of the instability region is mostly established by
the bulk viscosity of the quark core, whereas the saturation in the
mid temperature part comes mainly from the hadronic shell.}
\end{figure*}
The left panel of fig. \ref{fig:saturation-amplitude-Hybrid} shows
the corresponding plot for the $1.4\, M_{\odot}$ hybrid star. As
found previously in \cite{Alford:2010fd} the instability region has
here three parts that are separated by two stability windows arising
from the resonant behavior of the bulk viscosities in the different
shells. Due to their parametrically different temperature dependence
the bulk viscosity of the quark shell dominates at low temperature,
whereas the bulk viscosity of the hadronic shell dominates at high
temperatures. Correspondingly the saturation amplitude in the low
temperature part of the instability region shows the qualitative behavior
found for strange matter whereas the intermediate temperature part
shows the qualitative behavior found for hadronic matter. Since the
region where the peak in fig.\,\ref{fig:saturation-amplitude-APR}
is located is {}``cut out'' by the stability window, the remaining
peak of the hadronic intermediate part of the instability region in
fig.\,\ref{fig:saturation-amplitude-Hybrid} reaches only a much
lower amplitude.
The result for the heavy $2.0\, M_{\odot}$ neutron star is given
on the right panel of fig.\,\ref{fig:saturation-amplitude-APR}.
As had been found previously in \cite{Alford:2010fd} the instability
region is larger for such heavy stars. The figure shows in addition
that saturation occurs at a somewhat higher amplitude. The $2.0\, M_{\odot}$
strange star is given on the right panel of fig.\,\ref{fig:saturation-amplitude-Bag}.
In this case the maximum of the stability window is below the Kepler
frequency. Similar to the high temperature behavior discussed before,
the r-mode cannot be damped by viscous effects above this maximum.
It is interesting to recall from \cite{Alford:2010fd} that in the
case of quark matter an approximate analytic expression for the location
of the maximum of the stability window exists
\begin{align}
\Omega_{max}^{\left(SS\right)} & \approx\frac{0.434m_{s}^{4/3}R^{1/3}}{\left(1-c\right)^{1/3}G^{1/3}M^{2/3}}\\
T_{max}^{\left(SS\right)} & \approx\frac{0.210\left(1-c\right)^{1/3}m_{s}^{2/3}R^{1/6}}{\hat{\Gamma}^{1/2}G^{1/6}\mu_{q}^{3/2}M^{1/3}}\end{align}
where $\hat{\Gamma}\equiv\tilde{\Gamma}/\mu_{q}^{5}$. This shows
that in addition to a large star mass, a small effective strange quark
mass in the quark matter equation of state eq.\,(\ref{eq:quark-eos-model})
increases the total instability region both at high frequency and
high temperature. In contrast, for the heavy $2.0\, M_{\odot}$ hybrid
star shown on the right panel of fig.\,\ref{fig:saturation-amplitude-Hybrid}
such a total instability region does not arise since although the
quark core cannot saturate the r-mode, the hadronic shell alone still
provides sufficient damping to do so. In summary r-modes in massive
stars are more unstable than in light stars since both their small
amplitude instability regions are larger and they are less efficiently
saturated by the large amplitude enhancement of the bulk viscosity.
\begin{figure*}
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.72]{had-sat-amp-max.png}%
\end{minipage}%
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.72]{had-sat-amp-max-dir.png}%
\end{minipage}
\caption{\label{fig:saturation-amplitude-APR-dir}The static saturation amplitude,
at which the r-mode growth is stopped by suprathermal viscous damping
for the APR neutron stars at the maximum mass $2.21\, M_{\odot}$,
where direct Urca processes become allowed. \emph{Left panel:} direct
Urca is only allowed in a small inner core region, see Table \ref{tab:star-models}.
\emph{Right panel:} the same model when direct Urca is artificially
turned on in the entire core.}
\end{figure*}
The left panel of fig.\,\ref{fig:saturation-amplitude-APR-dir} shows
the static saturation amplitude for a neutron star with $2.21\, M_{\odot}$
which is the maximum mass allowed by the APR equation of state. In
this case direct Urca reactions are possible in a small inner core
region of mass $0.85\, M_{\odot}$. As had already been observed in
\cite{Alford:2010fd}, direct Urca reactions only slightly alter the
instability boundary by a small notch at its right hand side. Since
suprathermal damping from outer layers dominates due to the strong
radial dependence of the r-mode the static saturation amplitude is
likewise only slightly reduced by the small direct Urca core. However,
because the size of the inner direct Urca core depends strongly on
the equation of state and there are equations of state where the direct
Urca core is considerably larger, we show on the right panel of fig.\,\ref{fig:saturation-amplitude-APR-dir}
for comparison the (unphysical) case that the direct Urca reactions
are artificially switched on in the entire core. This represents an
upper limit for the possible effect of direct Urca reactions and shows
that in this extreme case the static saturation amplitude at large
frequency is reduced and the increase towards lower frequencies is
considerably weakened according to the $1/\Omega$-behavior predicted
by eq. (\ref{eq:analytic-saturation-amplitude}).
\begin{figure*}
\begin{tabular}{cc}
\includegraphics[scale=0.65]{had-sat-amp-1. 4-mode-2} & \includegraphics[scale=0.625]{had-sat-amp-1. 4-mode-3}\tabularnewline
\includegraphics[scale=0.64]{had-sat-amp-1. 4-mode-4} & \includegraphics[scale=0.65]{had-sat-amp-1. 4-mode-5}\tabularnewline
\end{tabular}
\caption{\label{fig:instability-regions-multipoles}Saturation amplitudes for
the first four multipole r-modes of the $1.4\, M_{\odot}$ neutron
star (top, left: $m=2$; top, right: $m=3$; bottom, left: $m=4$;
bottom right: $m=5$). The results are obtained in the linear approximation
eq. (\ref{eq:linear-bulk-vicosity-time}).}
\end{figure*}
In contrast to the previous results that evaluated the damping time
eq.\,(\ref{eq:bulk-damping-time}) numerically the top, left panel
of fig.\ref{fig:instability-regions-multipoles} employs the approximate
analytic expression eq.\,(\ref{eq:linear-bulk-vicosity-time}) for
the $m=2$ mode of the $1.4\, M_{\odot}$ neutron star. Comparing
it to the numerical result in fig.\,\ref{fig:saturation-amplitude-APR}
shows that the corrections are very small and because the maximum
of the bulk viscosity of hadronic matter with modified Urca reactions
is reached only for large amplitudes, eq.\,(\ref{eq:linear-bulk-vicosity-time})
provides a very good approximation in this case. In contrast, the
use of the linear approximation which neglects the large amplitude
decrease of the bulk viscosity, strongly overestimates the damping
for the case of strange stars and misses the previously discussed
total instability region at high frequency in fig.\,\ref{fig:saturation-amplitude-Bag}.
Fig.\,\ref{fig:instability-regions-multipoles} also shows the saturation
amplitudes of different multipole r-modes, given for the first four
multipoles $m=2$ to $5$ of the $1.4\, M_{\odot}$ neutron star .
The higher multipoles saturate at lower amplitudes than the $m=2$
and therefore the use of the linear approximate is well justified
in this case. Interestingly, although the right segments of the lower
part of the instability boundary of these higher order r-modes had
recently been shown to be very similar to that of the fundamental
$m=2$ mode \cite{Alford:2010fd}, fig.\,\ref{fig:instability-regions-multipoles}
shows that although the peak value of the saturation amplitude of
these modes decreases, the value at the Kepler frequency stays nearly
constant. Therefore, these higher multipoles could be relevant for
the spin-down evolution since the spin-down torque due to gravitational
wave emission depends strongly on the amplitude \cite{Owen:1998xg};
for sufficiently small amplitude modes this dependence is quadratic.
So if the suprathermal damping is responsible for the r-mode saturation,
the restriction to the lowest order mode, that had been employed in
all present analyses, should present only a first approximation.
\begin{figure*}
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.85]{sat-amps-ms}%
\end{minipage}%
\begin{minipage}[t]{0.5\textwidth}%
\includegraphics[scale=0.85]{sat-amps-4ms}%
\end{minipage}
\caption{\label{fig:sat-amps-ms}Comparison of the saturation amplitudes for
the different $1.4\, M_{\odot}$ stars. \emph{Left panel:} Stars spinning
with a period of $1$ ms. \emph{Right panel:} Same for stars rotating
with a period of $4$ ms. Shown are the considered neutron star (solid),
the hybrid star (dashed) and the strange star (dotted). The thick
curves present the numerical results and the thin horizontal segments
denote the analytic values obtained from eq.\,(\ref{eq:analytic-saturation-amplitude}). }
\end{figure*}
The saturation amplitudes of the different $1.4\, M_{\odot}$ stars
are finally compared with each other and the analytic approximation
eq.\,(\ref{eq:linear-bulk-vicosity-time}) in fig.\,\ref{fig:sat-amps-ms}.
Surprisingly, all stars feature saturation amplitudes of the same
order of magnitude for millisecond pulsars, despite their very different
microscopic and structural aspects. As noted before, for larger oscillation
periods hadronic stars and to some extent also hybrid stars feature
considerably larger saturation amplitudes than strange stars due to
the parametrically different frequency dependence, see eq.\,(\ref{eq:analytic-saturation-amplitude}).
The analytic approximation yields in most cases a reasonable approximation
to the full results with errors below the $10\%$ level. In general
the analytic result overestimates the actual amplitude since it only
describes the result far away from the boundaries and boundary effects
play a role. In contrast at large frequencies the analytic approximation
underestimates the saturation amplitude since the considered frequencies
are already close to the critical values, where the saturation amplitude
diverges, see fig.\,(\ref{fig:saturation-amplitude-Bag}). Nevertheless,
the analytic approximation provides an important and reliable estimate
for the order of magnitude of the static saturation amplitude which,
as will be discussed in more detail below, provides an upper limit
for saturation amplitudes taken in dynamical star evolutions.
\section{Conclusions}
Using the recent general results for the bulk viscosity that include
its non-linear behavior at large amplitudes we have derived expressions
for the r-mode damping time that show that in the regime below the
resonant temperature of the bulk viscosity, large amplitude r-modes
are damped on considerably shorter time scales than low amplitude
oscillations. In contrast, the universal maximum of the bulk viscosity
found in \cite{Alford:2010gw} implies that at very high temperatures
and frequencies r-modes cannot be damped at all by viscous effects
since there is no enhancement in the suprathermal limit. We find that
for most stars considered in this work the corresponding critical
frequency is above the Kepler frequency. On the other hand the r-modes
of all considered stars are unstable at temperatures that are expected
to be present when a proto-neutron star is created. At lower temperatures
our results lead to an extension of the concept of the instability
region of an r-mode since the latter is only initially unstable at
small amplitudes but the suprathermal viscous damping can saturate
the r-mode growth at finite amplitudes. We find that well within the
instability region the static saturation amplitude $\alpha_{sat}$
defined in the text is temperature independent and takes values $O\!\left(1\right)$
at milli-second frequencies for all considered stars. This is incidentally
the order of magnitude that had been assumed in early r-mode analyses
\cite{Owen:1998xg}. Yet, the static values obtained here represent
only an upper limit for the actual amplitude reached in the dynamic
evolution. Our numeric results are confirmed by approximate analytic
expressions which reveal the dependence of these results on the various
underlying parameters. We also studied higher multipoles and find
that although the first few multipoles have instability regions that
are sizable, they feature similar saturation amplitudes as the fundamental
$m\!=\!2$ mode for millisecond pulsars and could thereby be relevant.
It is interesting to compare our saturation mechanism and the obtained
results for the saturation amplitudes with previously proposed mechanisms.
In general when there are different competing saturation mechanisms,
the one with the smallest saturation amplitude should dominate and
effectively saturate the mode. Explicit numerical analyses of the
general relativistic hydrodynamical equations \cite{Lindblom:2000az,Gressman:2002zy,Lin:2004wx}
would present the ideal way to study the saturation and star evolution.
Whereas some of these studies find saturation only at large amplitudes,
in others the r-mode can be completely destroyed by the decay into
daughter modes once it exceeds amplitudes $O\left(10^{-2}\right)$
\cite{Lin:2004wx}, see also \cite{Brink:2004bg}. However, the numerical
complexity limits these analyses so far to unphysically large values
of the radiation reaction force that are orders of magnitude above
the physical value and it is not clear to what extent the obtained
results can be extrapolated to the physical case. Another proposed
saturation mechanism relies on the non-linear coupling of different
oscillation modes \cite{Arras:2002dw,Bondarescu:2007jw,Bondarescu:2008qx,Brink:2004bg}.
These analyses find that this mechanism could saturate r-modes at
amplitudes as low as $O\left(10^{-5}\right)$. Due to the considerable
difficulties of a complete description of such a mode coupling mechanism,
these analyses have to rely on model systems of generic coupled oscillators
without a detailed connection to the complicated coupling of collective
star oscillations. In summary, within the present approximation to
neglect the neutron star crust, competing mechanisms will very likely
dominate and saturate the r-mode at lower values than the suprathermal
enhancement of the viscosity. However, these mechanisms still involve
simplifications and uncertainties. Our novel saturation mechanism,
in contrast, relies on standard viscous effects and microscopic physics
that is quantitatively well understood.
Let us now discuss the implications of our results for the spin-down
of compact stars. In the supernova formation process where a much
larger star contracts to a very compact object that takes over the
angular momentum it seems plausible that fast rotating proto-neutron
stars could be formed which spin with frequencies close to the Kepler
limit. According to our results r-mode oscillations are unstable in
this initial hot stage $T\!\gtrsim\!10^{10}K$ and cannot be saturated
by viscous effects for all considered forms of dense matter. Generically,
the cooling is very fast in this regime so that the evolution could
leave this instability region before large amplitude r-modes develop
or spin down the star. The star will then cool until it reaches the
lower instability zone and the r-mode develops. According to fig.\,\ref{fig:sat-amps-ms}
in this regime the r-mode can be saturated by viscous damping. For
strange stars such a saturation does not seem to be required at all
since the instability region is in this case located at comparably
low temperatures \cite{Madsen:1998qb} where the cooling becomes slow
and the star either quickly spins down \cite{Andersson:2001ev}, or
when reheating effects are considered it reheats again \cite{Drago:2007iy},
and leaves the instability region before the amplitude becomes large.
In this case the evolution wiggles around the instability line thereby
spinning down the star, but this can take billions of years due to
the strong reheating.
In contrast in the case of neutron and hybrid stars without strangeness,
the instability region is reached at large temperatures where cooling
is still fast and reheating effects are moderate, so that the evolution
quickly penetrates the instability region and a saturation mechanism
is indeed required to stop the r-mode growth \cite{Owen:1998xg}.
Since the static saturation amplitude increases continuously at the
boundary of the instability region the discussed \emph{static} value
does not have to be reached but a \emph{dynamic} equilibrium could
be established at a lower saturation amplitude that is reached once
the r-mode is sufficiently large that the spindown becomes efficient.
Due to this the viscous saturation could dominate competing saturation
mechanisms. Once the r-mode is saturated, the question is which one
of two competing processes, cooling or spin-down, is faster. Since
the cooling is slowing down at lower temperatures it is likely that
the spin-down wins and the evolution leaves the instability region
near its lower boundary. In this case no young compact stars with
frequencies larger than a tenth of the Kepler frequency would be possible
which is in good agreement with observations. An answer to the above
questions requires a detailed study of the combined spin-down and
cooling evolution of the star which will be presented elsewhere.
Strikingly our results suggest even another possibility for the spindown
of young stars that would be even faster and more violent. The core
bounce during the supernova process should excite rather large amplitude
oscillation modes in the forming compact core. Since r-modes are unstable
in this regime \cite{Alford:2010fd} these will grow further. Because
of the initial high temperatures, neutrinos are trapped inside the
proto-neutron stars for roughly a minute \cite{Prakash:1996xs}. Since
a neutron star crust, that could provide an efficient damping mechanism
\cite{Bildsten:2000ApJ...529L..33B}, is not formed at this point
and as our results show viscous effects cannot stop the r-mode growth,
the amplitude could indeed become large if other non-linear saturation
mechanisms likewise cannot operate efficiently in this turbulent environment.
In this case the loss of angular momentum could proceed not by gravitational
wave emission but by actual mass shedding and thereby effectively
as an extension of the supernova explosion that is driven by r-modes.
Since such a violent spindown should be fast the star could end up
at the lower boundary of the high temperature instability region before
the star becomes transparent to neutrinos and the cooling process
starts. Clearly, in this initial stage, which cannot rigorously be
separated from the aftermath of the supernova explosion, the dynamics
is highly non-linear and our simple r-mode analysis might not directly
apply. Whether such a mechanism is feasible will therefore require
further study, but this mechanism would naturally explain the observed
absence of fast, young pulsars independent of their internal composition
and it is striking that the frequencies of the high temperature instability
boundary also seem to agree well with fastest pulsars that are young
enough that they cannot be spun up by accretion \cite{Manchester:2004bp}.
Finally, r-modes should also be relevant for old accreting stars in
binary systems that are spun up and could enter the instability region
at low temperatures from below \cite{Reisenegger:2003cq}. As discussed
in \cite{Alford:2010fd}, strange and hybrid stars feature stability
windows at low temperatures where the r-mode is absent, so that such
stars could accelerate to frequencies close to the Kepler frequency.
In contrast for neutron stars there is no stability window at low
temperatures so that an accreting star would enter the unstable regime
already at low frequencies. Recall that the saturation amplitude of
neutron stars due to bulk viscosity has a characteristic form with
a pronounced peak close to the minimum of the instability region.
In case the r-mode is saturated by suprathermal bulk viscosity, the
steep rise of the amplitude close to the maximum should spin down
the star quickly and so that it cannot penetrate deep into the instability
region. This means that such stars should cluster close to the boundary
which might be a signature once more observational data for the temperature
of compact stars becomes available.
\begin{acknowledgments}
We thank Nils Andersson, Greg Comer, Brynmor Haskell, Prashant Jaikumar,
Andreas Reisenegger, Andrew Steiner and Ira Wasserman for helpful
discussions. This research was supported in part by the Offices of
Nuclear Physics and High Energy Physics of the U.S. Department of
Energy under contracts \#DE-FG02-91ER40628, \#DE-FG02-05ER41375.
\end{acknowledgments}
|
\section*{\contentsname}%
\@starttoc{toc}%
}
\catcode`@=1
\newcommand{\proj}[1]{\mbox{$|#1\rangle \!\langle #1 |$}}
\def\begin{equation}{\begin{equation}}
\def\end{equation}{\end{equation}}
\def\begin{eqnarray}{\begin{eqnarray}}
\def\end{eqnarray}{\end{eqnarray}}
\def\ffrac#1#2{\textstyle{#1\over#2}\displaystyle}
\def{\rm Tr}\,{{\rm Tr}\,}
\def\mathcal H{\mathcal H}
\def\epsilon{\epsilon}
\def\sigma{\sigma}
\def\alpha{\alpha}
\def\lambda{\lambda}
\def{\l_{\rm max}}{{\lambda_{\rm max}}}
\def\mathcal L{\mathcal L}
\newcommand{{ x}}{{ x}}
\newcommand{{\tau}}{{\tau}}
\def\lt#1{\left#1}
\def\rt#1{\right#1}
\def\t#1{\tilde{#1}}
\newcommand{{\mathbf{R}}}{{\mathbf{R}}}
\newcommand{{\mathbf{C}}}{{\mathbf{C}}}
\newcommand{{\cal M}}{{\cal M}}
\newcommand{\langle}{\langle}
\newcommand{\rangle}{\rangle}
\newcommand{{\cal T}}{{\cal T}}
\begin{document}
\title[Entanglement of 2 disjoint intervals in $c=1$ theories]
{Entanglement entropy of two disjoint intervals in $c=1$ theories}
\author{Vincenzo Alba$^1$, Luca Tagliacozzo$^2$, Pasquale Calabrese $^3$}
\address{
$^1$ Max Planck Institute for the Physics of Complex Systems, N\"othnitzer Str. 38, 01187 Dresden, Germany,\\
$^2$ School of Mathematics and Physics, The University of Queensland,
Australia,\\
ICFO, Insitut de Ciencias Fotonicas, 08860 Castelldefels (Barcelona) Spain}
\address{$^3$ Dipartimento di Fisica dell'Universit\`a di Pisa and INFN,
Pisa, Italy. }
\date{\today}
\begin{abstract}
We study the scaling of the R\'enyi entanglement entropy of two disjoint blocks of critical lattice models described by conformal
field theories with central charge $c=1$.
We provide the analytic conformal field theory result for the second order R\'enyi entropy for a free boson compactified on
an orbifold describing the scaling limit of the Ashkin-Teller (AT) model on the self-dual line.
We have checked this prediction in cluster Monte Carlo simulations of the classical two dimensional AT model.
We have also performed extensive numerical simulations of the
anisotropic Heisenberg quantum spin-chain with tree-tensor network techniques
that allowed to obtain the reduced density matrices of disjoint blocks of the
spin-chain and to check the correctness of the predictions for R\'enyi and
entanglement entropies from conformal field theory.
In order to match these predictions, we have extrapolated the numerical results by properly taking into account
the corrections induced by the finite length of the blocks to the leading scaling behavior.
\end{abstract}
\maketitle
\section{Introduction}
Let us imagine to divide the Hilbert space ${\cal H}$ of a given quantum system into two parts ${\cal H}_A$ and ${\cal H}_B$ such that
${\cal H}={\cal H}_A\otimes {\cal H}_B$.
When the system is in a pure state $|\Psi\rangle$, the bipartite entanglement between
A and its complement B,
can be measured in terms of the R\'enyi entropies \cite{Renyi}
\begin{equation}
S_A^{(n)}=\frac1{1-n}\log{\rm Tr}\,\rho_A^n\,,
\label{renyidef}
\end{equation}
where $\rho_A={\rm Tr}_B\,\rho$ is the reduced density matrix of
the subsystem A, and $\rho=|\Psi\rangle\langle\Psi|$ is the
density matrix of the whole system.
The knowledge of $S_A^{(n)}$ as a function of $n$ identifies univocally the
full spectrum of non-zero eigenvalues of $\rho_A$ \cite{cl-08},
and provides complementary information about the
entanglement to the one obtained from the von Neumann entanglement
entropy $S_A^{(1)}$.
Furthermore, the scaling of $S_A^{(n)}$ with the size of A in the ground-state
of a one-dimensional system is more suited than $S_A^{(1)}$ to
understand if a faithful representation of the state in term of a
matrix product state can be or cannot be obtained with polynomial
resources in the length of the chain \cite{mps,cv-09}.
For a one-dimensional critical system whose scaling limit is described by a conformal
field theory (CFT), in the case when A is an interval of length $\ell$
embedded in an infinite system, the asymptotic large $\ell$ behavior of the quantities determining the R\'enyi
entropies is \cite{Holzhey,cc-04,Vidal,cc-rev}
\begin{equation}\fl \label{Renyi:asymp}
{\rm Tr}\,\rho_A^{n}
\simeq c_n \left(\frac{\ell}{a}\right)^{c(n-1/n)/6}\,,\qquad \Rightarrow
S_A^{(n)}\simeq\frac{c}6 \left(1+\frac1n\right)\log \frac{\ell}a +c'_n\,,
\end{equation}
where $c$ is the central charge of the underlying CFT and $a$ the inverse of an ultraviolet cutoff (e.g. the lattice spacing).
The prefactors $c_n$ (and so the additive constants $c'_n$) are non universal
constants (that however satisfy universal relations \cite{fcm-10}).
The central charge is an ubiquitous and fundamental feature of a conformal field theory
\cite{c-lec}, but it does not always identify the universality class of the theory.
A relevant class of relativistic massless quantum field theories are the $c=1$ models, which describe many
physical systems of experimental and theoretical interest.
The one-dimensional Bose gas with repulsive interaction, the (anisotropic) Heisenberg spin chains, the Ashkin-Teller model
and many others
are all described (in their gapless phases) by $c=1$ theories.
These are all free-bosonic field theories where the boson field satisfies different periodicity constraints, i.e. it is
compactified on a specific target space.
The two most notable examples are the compactification on a circle (corresponding to the Luttinger liquid field theory)
and on a $Z_2$ orbifold (corresponding to the Ashkin-Teller model \cite{z-87,book,dvv-87}).
The critical exponents
depend in a continuous way on the compactification radius of the bosonic field.
A survey of the CFTs compactified on a circle or on a $Z_2$ orbifold is given in Fig. \ref{fig14},
in a standard representation \cite{book,dvv-87}.
The horizontal axis is the compactification radius on the circle $r_{\rm circle}$, while
the vertical axis represents the value of the $Z_2$ orbifold compactification radius $r_{\rm orb}$.
The two axes cross in a single point, meaning that the theories at $r_{\rm circle}=\sqrt2$ and at $r_{\rm orb}=1/\sqrt2$
are the same. (The graph is not a cartesian plot, i.e. it has no meaning to have one $r_{\rm circle}$ and one $r_{\rm orb}$ at
the same time.)
For some values of $r_{\rm circle}$ and $r_{\rm orb}$, we report statistical mechanical models and/or field theories to
which they correspond.
In the following we will consider the Ashkin-Teller model that on the self-dual line is described
by $ r_{\rm orb}\in[\sqrt{2/3},\sqrt{2}]$ and the XXZ spin chain in zero magnetic field that is described by $ r_{\rm circle}\in[0,1/\sqrt2]$.
We mention that different compactifications have been studied \cite{g-88}, but they correspond to more
exotic statistical mechanical models and will not be considered here.
\begin{figure}[t]
\includegraphics[width=0.9\textwidth]{fig14.png}
\caption{Survey of $c=1$ theories corresponding to a free boson
compactified on a circle (horizontal axis) and on an orbifold (vertical axis) as reported e.g. in Refs. \cite{book}.
For some values of $r_{\rm circle}$ and $r_{\rm orb}$, the corresponding statistical mechanical models are reported.
The XXZ spin chain in zero magnetic field lies on the horizontal axis in the interval $r_{\rm circle}\in[0,1/\sqrt2]$.
The self-dual line of the Ashkin-Teller model lies on the vertical axis in the interval $ r_{\rm orb}\in[\sqrt{2/3},\sqrt{2}]$.}
\label{fig14}
\end{figure}
According to Eq. (\ref{Renyi:asymp}), the central charge of the CFT can be extracted
from the scaling of both the R\'enyi and von Neumann entropies.
In the last years, this idea has overcome the previously available techniques of determining $c$, e.g. by measuring the
finite size corrections to the ground state energy of a spin chain \cite{ecorr}.
However, the dependence of the scaling of the entropies of a single
block only on the central charge prevents to extract from them other
important parameter of the model such as the compactification radius.
It has been shown that instead the entanglement entropies of disjoint intervals are sensitive to the full operator content of the CFT
and in particular they depend on the compactification radius and on the symmetries of the target space.
Thus they encode complementary information about the underlying
conformal field theory of a given critical quantum/statistical system
to the knowledge of the central charge present in the scaling of the single block entropies.
(Oppositely in 2D systems with conformal invariant wave-function, the entanglement entropy of a single region depends on the
compactification radius \cite{2d}.)
This observation boosted an intense theoretical activity aimed at determining R\`enyi entropies of disjoint intervals both
analytically and numerically \cite{fps-08,cg-08,cct-09,ch-04,ffip-08,kl-08,rt-06,atc-09,ip-09,fc-10,h-10,fc-10b,c-10,cct-11}.
A part of this paper is dedicated to
consolidate some of the results already provided in other works where
they either have been studied only on very small chains, with the
impossibility of properly taking into account the severe finite size
corrections \cite{fps-08} or have been tested in the specific cases of spin
chains equivalent to free fermionic models \cite{atc-09,fc-10}.
An important point to recall when dealing with more than one
interval is that
the R\'enyi entropies in Eq. (\ref{renyidef}) measure only the entanglement of the
disjoint intervals with the rest of the system.
They do {\it not} measure the entanglement of one interval with respect to the other, that instead requires the
definition of more complicated quantities because $A_1\cup A_2$ is in a mixed state (see e.g. Refs. \cite{Neg} for a
discussion of this and examples).
Furthermore, it must be mentioned that some results about the entanglement of two disjoint intervals are at the basis of
a recent proposal to "measure" the entanglement entropy \cite{c-11}.
\subsection{Summary of some CFT results for the entanglement of two disjoint intervals}
We consider the case of two disjoint intervals
$A=A_1\cup A_2=[u_1,v_1]\cup [u_2,v_2]$.
By global conformal invariance, in the thermodynamic limit, ${\rm Tr}\, \rho_A^n$ can be written as
\begin{equation}\fl
{\rm Tr}\, \rho_A^n
=c_n^2 \left(\frac{|u_1-u_2||v_1-v_2|}{|u_1-v_1||u_2-v_2||u_1-v_2||u_2-v_1|}
\right)^{\frac{c}6(n-1/n)} F_{n}(x)\,,
\label{Fn}
\end{equation}
where $x$ is the four-point ratio (for real $u_j$ and $v_j$, $x$ is real)
\begin{equation}
x=\frac{(u_1-v_1)(u_2-v_2)}{(u_1-u_2)(v_1-v_2)}\,.
\label{4pR}
\end{equation}
The function $F_n(x)$ is a universal function (after being normalized such that $F_n(0)=1$) that encodes all the information about
the operator spectrum of the CFT and in particular about the compactification radius.
$c_n$ is the same non-universal constant appearing in Eq. (\ref{Renyi:asymp}).
Furukawa, Pasquier, and Shiraishi \cite{fps-08} calculated
$F_2(x)$ for a free boson compactified on a circle of radius $r_{\rm circle}$
\begin{equation}
F_{2}(x)=
\frac{\theta_3 (\eta \tau) \theta_3 (\tau/\eta )}{ [\theta_3 (\tau)]^{2}},
\label{F2}
\end{equation}
where $\theta_\nu$ are Jacobi theta functions and the (pure-imaginary) $\tau$ is given by
\begin{equation}
x= \left[\frac{\theta_2(\tau)}{\theta_3(\tau)}\right]^4\,,\qquad
\tau(x)=i\,\frac{ _2 F_1(1/2,1/2;1;1-x)}{ _2 F_1(1/2,1/2;1;x)}\,.
\label{mapping}
\end{equation}
$\eta$ is a universal critical exponent related to the compactification radius $\eta= 2r_{\rm circle}^2$.
\footnote{Because of the symmetry $\eta\to1/\eta$ or $r_{\rm circle}\to 1/2 r_{\rm circle}$ for any conformal property
one could also define $\eta= 1/2r_{\rm circle}^2$ as sometimes done in the literature.
However, corrections to scaling are not symmetric in $\eta\to1/\eta$ and this is often source of confusion.
A lot of care should be used when referring to one or another notation.}
This has been extended to general integers $n\geq 2$ in Ref. \cite{cct-09}
\begin{equation}
F_n(x)=
\frac{\Theta\big(0|\eta\Gamma\big)\,\Theta\big(0|\Gamma/\eta\big)}{
[\Theta\big(0|\Gamma\big)]^2}\,,
\label{Fnv}
\end{equation}
where $\Gamma$ is an $(n-1)\times(n-1)$ matrix with elements \cite{cct-09}
\begin{equation}
\Gamma_{rs} =
\frac{2i}{n} \sum_{k\,=\,1}^{n-1}
\sin\left(\pi\frac{k}{n}\right)\beta_{k/n}\cos\left[2\pi\frac{k}{n}(r-s)
\right]\,,
\label{Gammadef}
\end{equation}
and
\begin{equation}
\beta_y=\frac{\, _2 F_1(y,1-y;1;1-x)}{\, _2 F_1(y,1-y;1;x)}\,.
\label{betadef}
\end{equation}
$\eta$ is the same as above, while $\Theta$ is the Riemann-Siegel theta function
\begin{equation}
\label{theta Riemann def}
\Theta(0|\Gamma)\,\equiv\,
\sum_{m \,\in\,\mathbf{Z}^{\alpha-1}}
\exp\big[\,i\pi\,m^{\rm t}\cdot \Gamma \cdot m\big]\,.
\end{equation}
The analytic continuation of Eq. (\ref{Fnv}) to real $n$ for general values of
$\eta$ and $x$ (to obtain the von Neumann entanglement entropy) is still an open problem, but results for
$x\ll1$ and $\eta\ll1$ are analytically known \cite{cct-09,cct-11}.
The function $F_n(x)$ is known exactly for arbitrary integral $n$ also for the critical Ising field theory \cite{cct-11}.
However, in the following we will need it only at $n=2$ (i.e. $F_2(x)$) for which it assumes the simple form \cite{atc-09}
\begin{equation}\fl
F_2^{\rm Is}(x)=\frac1{\sqrt{2}}\Bigg[
\left(\frac{(1 + \sqrt{x}) (1 + \sqrt{1 - x})}2\right)^{1/2} + x^{1/4} + ((1 - x) x)^{1/4} + (1 - x)^{1/4} \Bigg]^{1/2}.
\label{CFTF2}
\end{equation}
In Ref. \cite{cct-11}, it has been proved that in any CFT the function $F_n(x)$ admits the small $x$ expansion
\begin{equation}
F_n(x)=1+ \left(\frac{x}{4n^2}\right)^{ \alpha}s_2(n)+ \left(\frac{x}{4n^2}\right)^{2\alpha}s_4(n)+\dots \,,
\label{Fexpintro}
\end{equation}
where $\alpha$ is the lowest scaling dimension of the theory.
The functions $s_j(n)$ are calculable from a modification of the short-distance expansion \cite{cct-11}, and
in particular it has been found \cite{cct-11}
\begin{equation}
s_2(n)={\cal N}\, \frac{n}2
\sum_{j=1}^{n-1}
\frac1{\left[ \sin\left(\pi\frac{j}{n}\right) \right]^{2\alpha}}\,,
\label{s2cft}
\end{equation}
where the integer ${\cal N}$ counts the number of inequivalent correlation functions giving the same contribution.
This expansion has been tested against the exact results for the free compactified boson (Ising model)
with $\alpha=\min[\eta,1/\eta]$ ($\alpha=1/4$) and ${\cal N}=2$ (${\cal N}=1$).
All the results we reported so far are valid for an infinite system.
Numerical simulations are instead performed for finite, but large, system sizes.
According to CFT \cite{cc-rev}, we obtain the correct result for a chain of finite length $L$ by
replacing all distances $u_{ij}$ with the {\it chord distance} $L/\pi \sin(\pi u_{ij}/L)$ (but different finite size forms exist
for excited states \cite{abs-11}).
In particular the single interval entanglement is \cite{cc-04}
\begin{equation}
\label{SnFS}
{\rm Tr}\,\rho_A^{n}
\simeq c_n \left[\frac{L}{\pi a} \sin\left(\frac{\pi \ell}{L}\right)\right]^{-c(n-1/n)/6}\,,
\end{equation}
and for two intervals, in the case the two subsystems $A_{1}$ and $A_{2}$ have the same length $\ell$ and are placed at
distance $r$, the four-point ratio $x$ is
\begin{equation}
x=\left(\frac{\sin\pi\ell/L}{\sin\pi (\ell+r)/L}\right)^2\,.
\label{xFS}
\end{equation}
\subsection{Organization of the paper}
In this paper we provide accurate numerical tests for the functions $F_n(x)$ in truly interacting lattice models
described by a CFT with $c=1$.
In Sec. \ref{sec2} we derive the CFT prediction for the function $F_2(x)$ of a free boson compactified on an orbifold
describing, among the other things, the self-dual line of the AT model when $r_{\rm orb}\in[\sqrt{2/3},\sqrt{2}]$.
In order to check this result, we needed to develop a classical Monte Carlo algorithm in Sec. \ref{ATmc} based on the ideas
introduced in Ref. \cite{cg-08}.
This algorithm is used in Sec. \ref{ATres} to determine $F_2(x)$ for several points on the self-dual line.
We also consider the XXZ spin-chain in zero magnetic field to test the correctness of Eq. (\ref{Fnv}).
In order to extend the results of Ref. \cite{fps-08} to longer
chains, we have used a tree tensor network algorithm that has allowed us
to study chains of length up to $L=128$ with periodic boundary conditions.
In this way, we have been able to perform a detailed finite size
analysis that was difficult solely with the data from exact diagonalization reported in Ref. \cite{fps-08}.
The analysis also shows that only through the knowledge of the unusual corrections to
the leading scaling behavior \cite{ccen-10,ce-10,cc-10,ccp-10,xa-11,fc-10} we are able to perform a quantitative test of Eq. (\ref{Fnv}).
The tree tensor network algorithm is described in Sec. \ref{ttn:sec}, while the numerical results are presented in Sec. \ref{XXZ:sec}.
The various sections are independent one from each other,
so that readers interested only in some results should have an easy access to them without reading the whole paper.
\section{$n=2$ R\`enyi entanglement entropy for two intervals in the Ashkin-Teller model}
\label{sec2}
In a quantum field theory ${\rm Tr}\,\rho_A^n$ for integer $n$ is proportional to the partition function on an $n$-sheeted Riemann surface
with branch cuts along the subsystem $A$, i.e. ${\rm Tr}\,\rho_A^n=Z_n(A)/Z_1^n$ where $Z_n(A)$
is the partition function of the field theory on a conifold where $n$ copies of the manifold ${\cal R}={\rm system}\times R^1$
are coupled along branch cuts along each connected piece of $A$ at a time-slice $t=0$ \cite{cc-rev,cc-05p}.
Specializing to CFT, for a single interval on the infinite line, this equivalence leads to Eq. (\ref{Renyi:asymp}) \cite{cc-04},
whose analytic continuation to non-integer $n$ is straightforward.
When the subsystem $A$ consists of $N$ disjoint intervals (always in an infinite system),
the $n$-sheeted Riemann surface ${\cal R}_{n,N}$ has genus $(n-1)(N-1)$ and cannot be mapped to the complex plane so that
the CFT calculations become more complicated.
However, for two intervals ($N=2$), when for a given theory the partition function on a generic Riemann surface of genus
$g$ with arbitrary {\it period matrix} is known, ${\rm Tr}\,\rho_A^n$ can be easily deduced exploiting the results
of Refs. \cite{cct-09,cct-11}.
In fact, a by-product of the calculation for the free boson \cite{cct-09} is that the $(n-1)\times (n-1)$ period matrix is always given
by Eq. (\ref{Gammadef}).
Although derived for a free boson, the period matrix is a pure geometrical object and it is only related to the structure of the world-sheet
${\cal R}_{n,2}$ and so it is the same for any theory.
This property has been used in Ref. \cite{cct-11} to obtain $F_n(x)$ for the Ising universality class for any $n$, in agreement with
previously known numerical results \cite{fc-10}.
When also $n=2$, the surface ${\cal R}_{2,2}$ is topologically equivalent to a torus for which the partition function is known
for most of the CFT.
The torus modular parameter $\tau$ is related to the four-point ratio by Eq. (\ref{mapping}).
Thus, the function $F_2(x)$ is proportional to the torus partition function where $\tau$ is given by Eq. (\ref{mapping}) and with
the proportionality constant fixed by requiring $F_2(0)=1$.
This way of calculating $S^{(2)}_A$ is much easier than the general one for $S^{(n)}_A$ \cite{Dixon,cct-09} and indeed it has
been used to obtain the first results both for the free compactified boson \cite{fps-08} and for the Ising model \cite{atc-09}.
For a conformal free bosonic theory with action
\begin{eqnarray}
S=\frac{1}{2\pi}\int d z d \bar{z} \,\partial\phi\bar{\partial}\phi\,,
\label{fba}
\end{eqnarray}
the torus partition functions are known exactly both for circle and orbifold compactification \cite{tori,s-87,book}.
We now recall some well-known facts in order to fix the notations and derive the function $F_2(x)$ for the Ashkin-Teller
model.
The bosonic field $\phi$ is said to be compactified on a circle of radius $r_{\rm circle}$ when $\phi=\phi+2\pi r_{\rm circle}$.
The torus partition function (and the one on the $n$-sheeted Riemann surface) should be derived with this constraint.
It is a standard CFT exercise to calculate the resulting torus partition function \cite{tori,book}
\begin{equation}
Z_{\rm circle}(\eta)= \frac{\theta_3(\eta\tau) \theta_3(\tau/\eta)}{|\eta_D(\tau)|^2}\,,
\end{equation}
where $\eta_D(\tau)$ is the Dedekind eta function and $\eta=2r_{\rm circle}^2$.
Using Eq. (\ref{mapping}) and some properties of the elliptic functions, Eq. (\ref{F2}) for $F_2(x)$ follows \cite{fps-08}.
When specialized at $\eta=1/2$ (or $\eta=2$), $F_2(x)$ has the simple form
\begin{equation}
F_2^{\rm XX}(x)= \sqrt{(1+x^{1/2})(1+(1-x)^{1/2})/2}\,,
\label{F2XX}
\end{equation}
that describes the XX spin-chain (that is equivalent to free fermions via the non-local Jordan-Wigner transformation).
The concept of orbifold emerges naturally in the context of theories
whose Hilbert space admits some discrete symmetries.
Let us assume that $G$ is a discrete symmetry.
For the free bosonic theory, the simplest example is the one we are interested in, i.e. the $Z_2$ symmetry.
It acts on the point of the circle $S^1$ in the following way
\begin{eqnarray}
g:\phi\rightarrow -\phi\,.
\end{eqnarray}
For the partition function of a theory on the torus, we introduce the notation \cite{book}
\begin{equation}
\quad\begin{picture}(2,2)
\put(0,-4){\framebox(15,15)}
\put(3.5,-14){$\pm$}
\put(-9,1.5){$\pm$}
\end{picture}\quad\quad
\end{equation}
\vspace{3mm}
where the $\pm$ denotes the boundary conditions on the two directions on
the torus. The full partition function, given a finite discrete group $G$, is
\begin{eqnarray}
Z_{{\cal T}/G}=\frac{1}{|G|}\sum\limits_{g,h\in G}\quad
\begin{picture}(2,2)
\put(0,-4){\framebox(15,15)}
\put(3.5,-14){\it h}
\put(-9,1.5){\it g}
\end{picture}
\label{modout}
\end{eqnarray}
where $|G|$ denotes the number of elements in the group.
The generalization to higher genus Riemann surfaces is straightforward (but it is not so easy to obtain results,
see e.g. \cite{dvv-87,orb2}).
Now we specialize Eq. (\ref{modout}) to the case of the $Z_2$ symmetry.
Since the action (\ref{fba}) is invariant under $g:\phi\rightarrow -\phi$,
we have the torus partition function for the free boson on the orbifold \cite{tori,s-87,book}
\begin{equation}
Z_{orb}=\frac{1}{2}\bigg(\quad
\begin{picture}(2,2)
\put(0,-4){\framebox(15,15)}
\put(3.5,-14){$+$}
\put(-9,1.5){$+$}
\end{picture}\qquad +
\quad\begin{picture}(2,2)
\put(0,-4){\framebox(15,15)}
\put(3.5,-14){$+$}
\put(-9,1.5){$-$}
\end{picture}\quad\quad+
\quad\begin{picture}(2,2)
\put(0,-4){\framebox(15,15)}
\put(3.5,-14){$-$}
\put(-9,1.5){$+$}
\end{picture}\quad\quad+
\quad\begin{picture}(2,2)
\put(0,-4){\framebox(15,15)}
\put(3.5,-14){$-$}
\put(-9,1.5){$-$}
\end{picture}\quad\quad
\bigg)\,.
\end{equation}
\vspace{1mm}
Standard CFT calculations lead to the result \cite{book}
\begin{eqnarray}
Z_{\rm orb}(\eta)=\frac{1}{2}\bigg(Z_{\rm circle}(\eta)+\frac{|\theta_3\theta_4|}{\eta_D\bar{\eta}_D}+\frac{|\theta_2\theta_3|}{\eta_D\bar{\eta}_D}+
\frac{|\theta_2\theta_4|}{\eta_D\bar{\eta}_D}\bigg)\,,
\end{eqnarray}
where all the $\tau$ arguments in $\theta_\nu$ and $\eta_D$ are understood.
At the special point $\eta=1/2$ (or $\eta=2$) we get
\begin{eqnarray}\fl
Z_{\rm orb}(\eta=1/2)=\frac{1}{2}\bigg(\frac{|\theta_3|^2+|\theta_4|^2+|\theta_2|^2}{2|\eta_D|^2}+
\frac{|\theta_3\theta_4|}{\eta_D\bar{\eta}_D}+\frac{|\theta_2\theta_3|}{\eta_D\bar{\eta}_D}+
\frac{|\theta_2\theta_4|}{\eta_D\bar{\eta}_D}\bigg)={Z}_{\rm Ising}^2\,.
\label{atpft}
\end{eqnarray}
Thus, from the orbifold partition function, using the last identity and normalizing such that $F_2^{AT}(0)=1$,
we can write the funcion $F_2^{\rm AT}(x)$ as
\begin{equation}
F_2^{\rm AT}(x)=\frac{1}{2}\bigg(F_2(x)-F_2^{XX}(x)\bigg)+(F_2^{\rm Is}(x))^2\,,
\label{atf2}
\end{equation}
where $F_2(x)$ is given in Eq. (\ref{F2}),
$F_2^{XX}(x)$ is the same at $\eta=1/2$ (cf. Eq. (\ref{F2XX})) and $F_2^{\rm Is}(x)$ is the result for
Ising (cf. Eq. (\ref{CFTF2})).
As a consequence of the $\eta\leftrightarrow 1/\eta$ symmetry of $F_2(x)$, also $F_2^{AT}(x)$ displays the same invariance.
For small $x$, recalling that $F_2(x)- 1\sim x^{{\rm min}[\eta,\eta^{-1}]}$,
$F_2^{XX}- 1\sim x^{1/2}$ and $F_2^{\rm Is}- 1\sim x^{1/4}$, we have
\begin{equation}
F_2^{AT}(x)-1\sim
\cases{
x^{1/4}& for $\eta\ge 1/4$\,,\\
x^{{\rm min}[\eta,\eta^{-1}]} & for $\eta\le 1/4$\,.
}
\end{equation}
The critical Ashkin-Teller model lies in the interval $\sqrt{2/3}<r_{\rm orb}<\sqrt{2}$ and so
$4/3<\eta=2r_{\rm orb}^2< 4$.
Thus we have $F_2^{AT}(x)-1\sim x^{1/4}$ along the whole self-dual line.
$F_2^{\rm AT}(x)$ for various values of $\eta$ in the allowed range is reported in Fig. \ref{log_curve},
where the behavior for small $x$ is highlighted in the inset to show the constant $1/4$ exponent.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.8\textwidth]{curve.png}
\end{center}
\caption{$F_2(x)$ for the Ashkin-Teller model on the self-dual line for some values of $\eta$.
Inset: $F_2(x)-1$ in log-log scale to highlight the small $x$ behavior. The black-dashed line is $\sim x^{1/4}$. }
\label{log_curve}
\end{figure}
\section{The classical Ashkin-Teller model and the Monte Carlo simulation}
\label{ATmc}
The two dimensional Ashkin-Teller (AT) model on a square
lattice is defined by the
Hamiltonian
\begin{eqnarray}
H=J\sum\limits_{\langle ij\rangle}\sigma_i\sigma_j +
J'\sum\limits_{\langle ij\rangle}\tau_i\tau_j+K\sum
\limits_{\langle ij\rangle}\sigma_i\sigma_j\tau_i\tau_j\,,
\label{ash}
\end{eqnarray}
where $\sigma_i$ and $\tau_i$ are classical Ising variables (i.e. can assume only the values $\pm1$).
Also the product $\sigma\tau$ can be considered as an Ising variable.
The model has a rich phase diagram whose features are reported in full details in Baxter's book \cite{BB}.
We review in the following only the main features of this phase diagram.
Under any permutation of the variables $\sigma,\tau,\sigma\tau$ the
AT model is mapped onto itself.
At the level of the coupling constants, this implies that the model is invariant under any
permutation of $J,J',K$.
For $K=0$, the AT model corresponds to two decoupled Ising models in
$\sigma$ and $\tau$ variables.
For $K\rightarrow \infty$ it reduces to a single Ising model with coupling constant $J+J'$.
For $J=J'=K$ it corresponds to the four-state Potts model.
It is useful to restrict to the symmetric Ashkin-Teller model where $J=J'$
\begin{eqnarray}
H=J\sum\limits_{\langle ij\rangle}(\sigma_i\sigma_j +
\tau_i\tau_j)+K\sum
\limits_{\langle ij\rangle}\sigma_i\sigma_j\tau_i\tau_j\,.
\label{sat}
\end{eqnarray}
The full phase diagram is reported in Fig. \ref{phadia} (in units of the inverse temperature $\beta=1$).
The model corresponds to two decoupled critical Ising models at $K=0$
and $2J=\log(1+\sqrt{2})$. For $J=0$ it is equivalent to
a critical Ising model in the variable $\sigma\tau$ with critical points at
$2K_\pm=\pm\log(1+\sqrt{2})$.
For $K\rightarrow\infty$ there are two critical Ising points at
$2J=\pm\log(1+\sqrt{2})$.
On the diagonal $J=K$ the system
corresponds to a 4-state Potts model which is critical at $K=(\log 3)/4$.
The different kinds of orders appearing in the phase diagram are explained in the caption of Fig. \ref{phadia}.
All the continuous lines in Fig.~\ref{phadia} are {\it critical lines}.
The blue lines C-Is are in the Ising universality class. The line starting from AFIs belongs to the
antiferromagnetic Ising universality class.
On the red line ABC the system is critical and the critical exponents vary continuously \cite{cont,BB}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.9\textwidth]{phadia.png}
\end{center}
\caption{Phase diagram of the 2D symmetric Ashkin-Teller model defined by the Hamiltonian (\ref{sat}).
The red ABC line is the self dual line.
The point $B$ at $K=0$ corresponds to two uncoupled Ising models.
The point $C$ is the critical four-state Potts model at $K=J=(\log 3)/4$.
At $J=0$ there are two critical Ising points at $K=\pm(\log(1+\sqrt{2}))/2$, one (Is) ferromagnetic and
the other (AFIs) antiferromagnetic.
For $K\rightarrow\infty$ there is another critical Ising point at $J=(\log(1+\sqrt{2}))/2$.
All continuous lines are critical.
The blue lines $C-Is$ and the one starting at $AFIs$ are in the Ising universality class.
The red line is critical with continuously varying critical exponents.
The region denoted by I corresponds to a ferromagnetic phase for all the variables.
In the region II, $\sigma$, $\tau$, and $\sigma\tau$ are paramagnetic.
In the region III only $\sigma\tau$ is ferromagnetic and in region IV $\sigma\tau$ exhibits
antiferromagnetic order while $\sigma$ and $\tau$ are paramagnetic.}
\label{phadia}
\end{figure}
The AT model on a planar graph can be mapped to another AT model on the dual graph.
When specialized to the square lattice, the phase diagram is equivalent to its dual on the self-dual line:
\begin{eqnarray}
e^{-2K}=\sinh(2J)\,.
\end{eqnarray}
On this line, the symmetric AT model maps onto an homogeneous
six-vertex model which is exactly solvable \cite{BB}.
It follows that on the self-dual line the model is critical for $K\le(\log3)/4$
and its critical behavior is described by a CFT with $c=1$.
Along the self-dual line the critical exponents vary countinuously and are exactly known.
For later convenience it is useful to parametrize the self dual line by a new parameter $\Delta$
\begin{equation}
e^{4J}=\frac{\sqrt{2-2\Delta}+1}{\sqrt{2-2\Delta}-1}\,,\qquad\qquad
e^{4K}=1-2\Delta\,,
\end{equation}
with $-1<\Delta<1/2$. In terms of $\Delta$, the orbifold compactification radius is \cite{s-87}
\begin{equation}
\eta={2r_{\rm orb}^2}=\frac{4\arccos (-\Delta)}\pi=\frac{2}{K_L}\,,
\label{etaAT}
\end{equation}
where $K_L$ is the equivalent of the Luttinger liquid parameter for the AT model.
\subsection{Cluster representation and Monte Carlo simulation}
A Swendsen-Wang type cluster algorithm for the AT model has been proposed in Ref. \cite{dom} and then
re-derived in a simpler way by Salas and Sokal \cite{ss}.
Here we partly follow the derivation of Salas and Sokal and we restrict to the symmetric
AT Hamiltonian (\ref{sat}) and assume $J\ge|K|$.
Using the identities for Ising type variables
\begin{equation}
\sigma_i\sigma_j=2\delta_{\sigma_i\,\sigma_j}-1\,, \qquad
\tau_i\tau_j=2\delta_{\tau_i\,\tau_j}-1\,,
\end{equation}
we can rewrite Eq. (\ref{sat}) as
\begin{equation}\fl
-H=J\sum\limits_{\langle ij\rangle}(2\delta_{\sigma_i\,\sigma_j}+
2\delta_{\tau_i\,\tau_j}-2)+K\sum\limits_{\langle
ij\rangle}(2\delta_{\sigma_i\,\sigma_j}-1)(2\delta_{\tau_i\,\tau_j}-1)\,.
\end{equation}
For convenience we shift the interaction (\ref{sat}) by $-4J$.
In order to write the Boltzmann weight associated to a specific
configuration we use
$\exp(w\delta_{\sigma_i\,\sigma_j})=(\exp(w)-1)\delta_{\sigma_i\,\sigma_j}+1$
and the analogous identity for the $\tau$ variables.
The Boltzmann weight of a given link $\langle ij\rangle$ is then
\begin{eqnarray}
{\cal{W}}_{\langle ij\rangle}(\sigma_i,\sigma_j,\tau_i,\tau_j)&=&e^{-4J}+[e^{-2(J+K)}-e^{-4J}][\delta_{\sigma_i
\,\sigma_j}+\delta_{\tau_i\,\tau_j}]+\nonumber \\ &&+
[1-2e^{-2(J+K)}+e^{-4J}]\delta_{\sigma_i\,\sigma_j}\delta_{\tau_i\,\tau_j}\,.
\label{atw}
\end{eqnarray}
The key idea for the Swendsen-Wang algorithm is to introduce two new auxiliary Ising-type variables $m_{ij}$ and $n_{ij}$
living on the link $\langle ij\rangle$.
We redefine the Boltzmann weight on the link
$\langle ij\rangle$ as \cite{ss}
\begin{eqnarray}\fl
{\cal{W}}_{\langle ij\rangle}&&(\sigma_i,\sigma_j,\tau_i,\tau_j,m_{ij},n_{ij})=
e^{-4J}\delta_{m_{ij} 0}\delta_{n_{ij}0}+ \nonumber \\\fl&&
+[e^{-2(J+K)}-e^{-4J}][ \delta_{\sigma_i \sigma_j}\delta_{m_{ij}1}\delta_{n_{ij}0}+\delta_{\tau_i\tau_j}
\delta_{m_{ij}0}\delta_{n_{ij}1}]+ \nonumber\\ \fl&&+
[1-2e^{-2(J+K)}+e^{-4J}]\delta_{\sigma_i \sigma_j}\delta_{\tau_i \tau_j}
\delta_{m_{ij}1}\delta_{n_{ij}1} \,.
\label{atwc}
\end{eqnarray}
Summing over $m_{ij}$ and $n_{ij}$ we obtain the weight in Eq. (\ref{atw}).
Eq. (\ref{atwc}) has a graphical interpretation in terms of clusters.
In fact we can divide the links of the lattice in ``activated'' (if $m_{ij}=1$) or ``inactive'' (if $m_{ij}=0$).
The same considerations hold for the $n_{ij}$ variables.
Therefore, each link of the lattice can be activated by setting $m_{ij}=1$ or $n_{ij}=1$.
The active links connect different lattice sites forming clusters.
There are clusters referring to the $\sigma$ variables (called $\sigma$-clusters) and to the $\tau$ variables ($\tau$-clusters).
Isolated lattice sites are clusters as well.
Obviously, the lattice sites belonging to the $\sigma$-clusters ($\tau$-clusters) have the same value of $\sigma$ ($\tau$).
The partition function of the extended model defined by the weight (\ref{atwc}) can be written as
\begin{equation}
Z=\sum\limits_{\sigma,\tau=\pm 1}\sum\limits_{m,n=\pm 1}\prod\limits_{\langle ij\rangle}{\cal W}_{\langle
ij\rangle}(\sigma_i,\sigma_j,\tau_i,\tau_j,m_{ij},n_{ij})\,.
\end{equation}
We now proceed to the following definitions.
We divide all the links into three classes: we define $l_0$ the total
number of inactivated links; $l_1$ the total number of links
connecting sites which belong only to one type of clusters either a
$\sigma$-cluster or a $\tau$-cluster. We define $l_2$ the total
number of links on which $m$ and $n$ are both equal to $1$.
Furthermore we introduce the quantities
\begin{eqnarray}
B_0\equiv e^{-4J}\,,\\
B_1\equiv [e^{-2(J+K)}-e^{-4J}]\,,\\
B_2\equiv [1-2e^{-2(J+K)}+e^{-4J}]\,.
\end{eqnarray}
The following step is to perform the summation over $\sigma,\tau$ in Eq. (\ref{atwc}).
This is readily done, obtaining the final expression for the partition function
\begin{eqnarray}
Z=\sum\limits_{{\cal C}\{\tau,\sigma\}}B_0^{l_0}B_1^{l_1}B_2^{l_2}\,2^{C^{\sigma}+C^{\tau}} \,,
\label{atpf}
\end{eqnarray}
where we denoted with $C^{\sigma}$ the number of $\sigma$-clusters and
with $C^\tau$ the total number of $\tau$-clusters.
In the counting of $\tau$-clusters ($\sigma$-clusters) we included all the lattice
sites connected by a link on which $m_{ij}=1$ ($n_{ij}=1$).
Isolated sites (with respect to $m$ or $n$ or both) count as single clusters.
The links where $m_{ij}=1,n_{ij}=1$ contribute to both types of clusters.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.5\textwidth]{cluster.png}
\end{center}
\caption{A typical cluster configuration on a $12\times 12$ lattice.
Green lines are $\sigma$-clusters and red dashed lines are $\tau$-clusters.
Links in blue are double links. Periodic boundary conditions on both directions are used.}
\label{clu_ins}
\end{figure}
\subsection{Swendsen-Wang algorithm (the direct and embedded algorithms)}
We are now in position to write the Swendsen-Wang algorithm for the symmetric AT model.
The Monte-Carlo procedure can be divided in two steps.
In the first one, given a configuration for $(\sigma,\tau)$ variables, we construct a configuration of the
$(m,n)$ variables. In the second step we update the $(\sigma,\tau)$
variables at given $(m,n)$.
The details of the step one are
\begin{itemize}
\item if $\sigma_i=\sigma_j$ and $\tau_i=\tau_j$, we choose
$(m_{ij},n_{ij})$ with the following probabilities:
\begin{itemize}
\item $(m_{ij},n_{ij})=(1,1)$ with $p_1=1-2e^{-2(J+K)}+e^{-4J}$,
\item $(m_{ij},n_{ij})=(1,0)$ with $p_2=e^{-2(J+K)}+e^{-4J}$,
\item $(m_{ij},n_{ij})=(0,1)$ with $p_2=e^{-2(J+K)}+e^{-4J}$,
\item $(m_{ij},n_{ij})=(0,0)$ with $p_3=1-p_1-2p_2$,
\end{itemize}
\item if $\sigma_i=\sigma_j$ and $\tau_i=-\tau_j$, the probabilities are
\begin{itemize}
\item $(m_{ij},n_{ij})=(1,0)$ with $p_1=1-e^{-2(J-K)}$,
\item $(m_{ij},n_{ij})=(0,0)$ with $p_2=1-p_1$,
\end{itemize}
\item if $\sigma_i=-\sigma_j$ and $\tau_i=\tau_j$, the probabilities are
\begin{itemize}
\item $(m_{ij},n_{ij})=(1,0)$ with $p_1=1-e^{-2(J-K)}$,
\item $(m_{ij},n_{ij})=(0,0)$ with $p_2=1-p_1$,
\end{itemize}
\item if $\sigma_i=-\sigma_j$ and $\tau_i=-\tau_j$ we choose
$(m_{ij},n_{ij})=(0,0)$ with probability $1$.
\end{itemize}
In the step two, given the configuration of $(m,n)$ generated using the
rules above we build the connected $\sigma$-clusters and
$\tau$-clusters. The value of $\sigma$ ($\tau$) spins are
required to be equal within each $\sigma$-cluster ($\tau$-cluster).
We choose randomly the spin value in each cluster and independently of
the value assumed on the other clusters. This completes the update scheme.
(Note a typo in Ref. \cite{ss}: the minus sign in step 2 and 3 of the update is missing.)
In Ref. \cite{ss} also the so called embedded version of the cluster algorithm is introduced.
Its implementation is slightly easier compared to the direct algorithm.
In the embedded algorithm instead of treating both $\sigma$ and $\tau$ at the same time, one
deals with only one variable per time.
Let us consider the Boltzmann weight of a link $\langle ij\rangle$ at fixed configuration of $\tau$
\begin{eqnarray}
{\cal W}_{\langle
ij\rangle}(\sigma_i,\sigma_j,\tau_i,\tau_j)=e^{-2(J+K\tau_i\tau_j)}+
(1-e^{-2(J+K\tau_i\tau_j)})\delta_{\sigma_i\,\sigma_j} \,.
\end{eqnarray}
The model defined by this weight can be simulated with a standard Swendsen-Wang algorithm for the Ising model
using the effective coupling constant
\begin{eqnarray}
J^{eff}_{ij}=J+K\tau_i\tau_j\,.
\end{eqnarray}
This is no longer translation invariant, but this does not affect the effectiveness of the cluster algorithm for the Ising model as long as
$J^{eff}_{ij}\ge 0$.
The same reasoning applies to the case of fixed $\sigma$.
Thus, the embedded algorithm is made of two steps
\begin{itemize}
\item For a given configuration of $\tau$ variables, we apply a standard
Swendsen-Wang algorithm to $\sigma$ spins. The probability arising in
the update step is $p_{ij}=1-e^{-2(J+K\tau_i\tau_j)}$.
\item For a given configuration of $\sigma$ variables, we update $\tau$ with
the same algorithm and probability $p_{ij}=1-e^{-2(J+K\sigma_j\sigma_i)}$.
\end{itemize}
Direct and embedded algorithms are both extremely effective
procedures to sample the AT configurations.
However, very important for the following, Eq. (\ref{atpf}) for the
partition function does not hold anymore for a $n$-sheeted Riemann surface and we do not know whether it is possible
to write the embedded algorithm for this case.
\subsection{R\'enyi entanglement entropies via Monte Carlo simulation of a classical system.}
In this section we summarize the method introduced by Caraglio and
Gliozzi \cite{cg-08} to obtain the R\'enyi entropies via simulations of classical systems and we generalize it to the AT model.
The partition function $Z={\rm Tr}\, e^{-\beta H}$ of a $d$-dimensional quantum system at
inverse temperature $\beta$ can be written as an Euclidean path integral in $d+1$ dimensions \cite{cc-rev}.
Thus for the $n$-th power of the partition function one has
\begin{eqnarray}
Z^n=\int\prod\limits_{k=1}^n{\cal D}[\phi_k]e^{-\sum\limits_{k=1}^n S(\phi_k)}
\end{eqnarray}
where $\phi_k\equiv\phi_k(\vec x, \tau)$ is a field living on the $k$-th replica of the system and $S(\phi_k)$ is the euclidean
action ($\tau$ is the imaginary time.)
The actual form of the action is not important, but for the
sake of simplicity we restrict to the case of nearest-neighbor interactions
\begin{eqnarray}
S(\phi_k)=\sum\limits_{\langle ij\rangle}F(\phi_k(i),\phi_k(j))\,,
\end{eqnarray}
and the function $F$ is arbitrary.
We recall that ${\rm Tr}\,\rho_A^n$ can be obtained by considering the euclidean partition function over
a $n$-sheeted Riemann surface with branch cuts along the subsystem $A$ \cite{cc-rev}.
(This equivalence is also the basis of all quantum Monte Carlo methods to simulate the block
entanglement in any dimension \cite{qmc}.)
Caraglio and Gliozzi constructed this $n$-sheeted Riemann surface for the lattice model in the following way.
Let us consider a square lattice (for simplicity) and take the two points of its dual lattice surrounding $A$ (that in 1+1 dimension is just
an interval with two end-points).
The straight line joining them defines the cut that we call $\lambda$.
The length of $\lambda$ is equal to the length of $A$.
Let us consider $n$ independent copies of this lattice with a cut.
The $n$-sheeted Riemann lattice is defined by assuming that all the links of the $k$-th replica intersecting the cut
connect with the next replica $k+1(\textrm{mod}\,n)$.
To get the partition function over the $n$ sheeted Riemann surface we define the
corresponding coupled action
\begin{equation}\fl
S^n(\phi_k)=\sum\limits_{k=1}^n\sum\limits_{\langle ij\rangle\notin
\lambda}F(\phi_k(i),\phi_k(j))+ \sum\limits_{\langle ij\rangle\in
\lambda}F(\phi_k(i),\phi_{k+1 ({\rm mod}\,n)}(j)) \,.
\label{act}
\end{equation}
This definition can be used in any dimension, even though we will use here only $d=2$.
Finally, calling $Z_n(A)$ the partition function over the action (\ref{act}), ${\rm Tr}\,\rho^n_A$ is given by
\begin{eqnarray}
{\rm Tr}\,\rho^n_A=\frac{Z_n(A)}{Z^n}\,.
\end{eqnarray}
Following Ref. \cite{cg-08} we introduce the observable
\begin{eqnarray}
{\cal O}\equiv
e^{-S^n(\phi_1,\phi_2,\dots,\phi_n;\lambda)+\sum_{k=1}^n
S(\phi_k;\lambda)} \,,
\label{obs}
\end{eqnarray}
where $S^n$ and $S$ are the euclidean actions of the model defined on the $n$-sheeted lattice and on the $n$
independent lattices respectively.
The sum is restricted to links crossing the cut, as the presence of $\lambda$ in the arguments stresses.
It then follows
\begin{equation}
\langle{\cal O}\rangle_n\equiv\frac{Z_n(A)}{Z^n}={\rm Tr}\,\rho^n_A\,,
\end{equation}
where $\langle\cdot\rangle_n$ stands for the average taken
onto the uncoupled action $\sum_{k=1}^n S(\phi_k)$.
We can now discuss our improvement to the procedure highlighted so far.
The practical implementation of Eq. (\ref{obs}) to calculate ${\rm Tr}\,\rho_A^n$ is
plagued by severe limitations:
analyzing the Monte-Carlo evolution of the observable, one notices that it
shows a huge variance because it is defined by an exponential.
Direct application of Eq. (\ref{obs}) is possible then only for small lengths of the subsystem $A$.
In order to overcome this problem, let us consider the quantity ${Z_n(A)}/{Z^n}$ and imagine to divide the
subsystem in $L$ parts to have $A=A_1\cup A_2\dots \cup A_L$, with the
lengths of the various parts being arbitrary.
Moreover we define a set of subsystems $\hat{A}_i\equiv \cup_{k=1}^i A_i$.
Then it holds
\begin{eqnarray}
\frac{Z_n(A)}{Z^n}=\prod\limits_{i=0}^L\frac{Z_n(\hat{A}_{i+1})}{Z_n(\hat{A}_i)}\,.
\label{trick}
\end{eqnarray}
Eq. (\ref{trick}) is very useful because each term in the product can be simulated
effectively using a modified version of (\ref{obs}) if we choose
the length of $A_i$ to be small enough.
In fact, by definition, we have
\begin{eqnarray}
\langle{\cal O}(\hat{A}_i)\rangle_{{\cal
R}_n(\hat{A}_i)}\equiv\frac{Z_n(\hat{A}_{i+1})}{{\cal
Z}_n(\hat{A}_i)} \,,
\label{trick1}
\end{eqnarray}
where ${\cal O}(\hat{A}_i)$ is the modified observable
\begin{eqnarray}
{\cal O}(\hat{A}_i)\equiv\exp(-S^n(\hat{A}_{i+1})+S^n(\hat{A_i}))\,.
\end{eqnarray}
We stress that in Eq. (\ref{trick1}) the expectation value in the l.h.s
must be taken on the coupled action on the Riemann surface
with cut $\hat{A}_i$.
The disadvantage of Eq. (\ref{trick}) is that, to simulate large subsystems, one has to perform $L$
independent simulations and then build the observable taking the product of the results.
If the dimension of each piece $A_i$ is small this task requires a large computational effort.
Another important aspect is the estimation of the Monte Carlo error:
if each term in (\ref{trick}) is obtained independently, the error in the product is
\begin{eqnarray}
\frac{\sigma({\cal O})}{\overline{\cal O}}=\sqrt{\sum\limits_{i=0}^L\frac{\sigma^2({\cal O}({\hat A}_i))}{{\overline{{\cal O}({\hat A}_i)}}^2}}\,.
\label{error}
\end{eqnarray}
If the lengths of the intervals $A_i$ are all equal, then the
single terms of the summation in Eq. (\ref{error}) do not change
much and the total error should scale as $\sqrt{L}$.
Caraglio and Gliozzi \cite{cg-08} used another strategy to circumvent the problem with the observable in Eq. (\ref{obs}).
The trick was to consider the Fortuin-Kastelayn cluster expansion of the partition function of the Ising model.
The analogous for the AT model was reported in the previous section
\begin{eqnarray}
Z=\sum\limits_{{\cal
C}\{\sigma,\tau\}}B_0^{l_0}B_1^{l_1}B_2^{l_2}\,2^{C^{\sigma}+C^{\tau}} \,,
\end{eqnarray}
where ${\cal C}^{\sigma,\tau}$ are the $\sigma/\tau$-cluster configurations.
Going from $n$ independent sheets to the $n$-sheeted lattice, the type of
links and their total number do not change, but
the number of clusters does change, and so we get the cluster expression of observable (\ref{obs}) for the AT model
\begin{eqnarray}
{\cal O}(\hat{A_i})= 2^{[C_\sigma(\hat{A}_{i+1})+C_\tau(\hat{A}_{i+1})
- C_\sigma(\hat{A}_i)-C_\tau(\hat{A}_i)]}\,,
\label{cobs}
\end{eqnarray}
where $C_\sigma(\hat{A}_i)$ ($C_\tau(\hat{A}_i)$) denote the total
number of $\sigma$-clusters ($\tau$-clusters) on the Riemann surface
with cut $\hat{A}_i$.
Since the clusters are non local objects, they represent ``improved'' observables and the variance
for the Monte Carlo history of Eq. (\ref{cobs}) is much smaller than in the naive implementation.
\section{The entanglement entropy in the Ashkin-Teller model}
\label{ATres}
\begin{figure}[t]
\begin{center}
\includegraphics[width=.8\textwidth]{error.png}
\end{center}
\caption{${\rm Tr}\,\rho_A^2$ for a single interval of length $\ell$ in a finite system of length $L=120$.
Data have been obtained by Monte Carlo simulations using the embedded algorithm.
The orange points correspond to the SUSY model and the green ones to the $Z_4$ parafermions.
The black crosses at $\ell=10$ are data obtained using the direct algorithm.
Inset: behavior of the statistical error of ${\rm Tr}\,\rho_A^2$ vs $\ell$ for the SUSY model.
The blue-dashed line is the expected form $A+B\ell^{1/2}$.}
\label{example}
\end{figure}
\subsection{The single interval}
We first present the results for the Ashkin-Teller model for a single interval.
Although these results do not provide any new information about the model, they are fundamental checks
for the effectiveness of the Monte Carlo algorithms.
We performed simulations using both algorithms described in the previous section: the direct cluster algorithm and the embedded one.
When using the direct algorithm, measures are performed using the observable (\ref{cobs}),
while for the embedded algorithm we used the observable in Eq. (\ref{obs}).
In Fig. \ref{example} we report the results of the simulations of ${\rm Tr}\,\rho_A^2$
for the SUSY model ($r_{\rm orb}=\sqrt{3}/2$ in Fig. \ref{fig14})
and for the $Z_4$ parafermions ($r_{\rm orb}=\sqrt{3/2}$) both for $L=120$.
The orange and green points are obtained using the embedded algorithm.
To check the implementation of the cluster observable, we report at $\ell=10$ the data obtained
using the direct algorithm and Eq. (\ref{cobs}).
The perfect agreement between the two results confirms the correctness of both implementations.
Note that ${\rm Tr}\,\rho_A^2$ is a monotonous function of $\ell$, in contrast with the parity effects found for the XXZ spin chain
\cite{ccen-10,ce-10} that also corresponds to a vertex model \cite{BB}.
In the inset we show the behavior of the statistical error of the observable (\ref{obs}) in the SUSY case
as function of the subsystem length $\ell$.
It agrees with the prediction in Eq. (\ref{error}) and its absolute value is extremely small, smaller than the size of the points
in the main plot in Fig. \ref{example}.
Analogous results have been obtained for all the critical points on the self-dual line using both algorithms.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.8\textwidth]{teller_c2.png}
\caption{Plot of $c_2(L_c)$ as function of $L_c$ for different $\ell$ and $L$.
Three points on the self-dual line are reported: four-states Potts model, uncoupled Ising, and SUSY.
The dashed lines are fits to the function $c_2+BL_c^{-K_L}$ ($c_2+AL_c^{-K_L}+BL_c^{-2K_L}$ for the 4-states Potts model) where
$K_L$ is $1/2,1,4/3$ respectively for the four-states Potts model, Ising, and SUSY.
In the inset we report $c_n$ for $n=3,4$ for the SUSY point.
The dashed lines are fit to $A+BL_c^{2K_L/n}$, with $K_L=4/3$ fixed.}
\label{c2}
\end{center}
\end{figure}
The results for ${\rm Tr}\,\rho_A^2$ in a finite system are asymptotically described by the CFT prediction (\ref{SnFS}) with $n=2$ and $c=1$.
It is then natural to compute the ratio
\begin{equation}
c_2(L_c)= \frac{{\rm Tr}\,\rho_A^2}{(\frac{L}{\pi}\sin(\frac{\pi}{L}\ell))^{-1/4}}\,,
\label{c2L}
\end{equation}
that is expected to be asymptotically a function of the chord-length
$L_c=[\frac{L}{\pi}\sin(\frac{\pi}{L}\ell)]$.
This allows to extract the non-universal quantity $c_2$ and to check the form of the corrections to the scaling.
In Fig. \ref{c2} we report the results for $c_2(L_c)$
for the SUSY point, for the two uncoupled Ising models, and for the four states Potts model.
It is evident that for large $L_c$, $c_2(L_c)$ approaches a constant value around $0.5$.
This is a first confirmation of the CFT predictions on the self-dual line.
The previous results also provide a test for the theory of the corrections to the scaling to $S^{(n)}_A$.
It has been shown \cite{ccen-10,ce-10} that for gapless models described by a Luttinger liquid theory, the corrections to
the scaling have the form $\ell^{-2K_L/n}$ (or $L_c^{-2K_L/n}$ for finite systems)
where $K_L$ is the Luttinger parameter, related to the circle compactification radius $K_L=1/2\eta$.
On the basis of general CFT arguments \cite{cc-10}, it has been argued that this scenario is valid for any CFT
and so also for the AT model with $K_L$ replaced by the dimension of a proper operator.
It is then natural to expect that for the AT model this dimension is $K_L$ in Eq. (\ref{etaAT}), also on the basis of
the results for the Ising model \cite{ccen-10,ij-08}.
The dashed lines in Fig. \ref{c2} are fits of $c_2(L_c)$ with the function $c_2+A L_c^{-K_L}$.
The agreement is always very good, except for the four-state Potts model, for which
the exponent of the leading correction $K_L$ assumes the smallest value and so
subleading corrections enter (as elsewhere in similar circumstances, see e.g. \cite{ce-10}).
In fact, the fit with the function $c_2+A L_c^{-K_L}+B L_c^{-2K_L}$ is in perfect agreement with the data (but the presence
of another fit parameter makes this result not so robust).
This analysis confirms that $K_L$ is the right exponent governing the corrections to the scaling.
\begin{figure}[t]
\includegraphics[width=\textwidth]{teller_Susy_F2.png}
\caption{
$F_2^{\rm}(x)$ versus the four point ratio $x$ for the SUSY model.
The red points are extrapolations obtained using the finite-size ansatz (\ref{ansatz}).
The blue-dashed line is the $CFT$ prediction.
Inset: $F_2^{\rm lat}(x)$ vs $1/\ell^{-2/3}$ for the four values of $x$ used in the extrapolation ($x=0.134,0.25,0.5,0.587$).
The dashed lines are fits to finite-size ansatz (\ref{ansatz}). }
\label{Susy_F2}
\end{figure}
In the inset of Fig. \ref{c2} we also report the values of $c_n$ for $n=3,4$ as a function of $L_c$.
$c_n$ becomes smaller as $n$ increases as for the $XXZ$ \cite{ccen-10},
XX \cite{jk-04}, and Ising \cite{ij-08,ccd-08} spin-chains.
The dashed lines are fits to the expected scaling behavior $L_c^{-2K_L/n}$ of the corrections, that reproduce perfectly the data.
\subsection{The entanglement entropy of two disjoint intervals.}
In this section we investigate the entanglement entropy of two disjoint intervals and check the correctness of our prediction
(\ref{atf2}) for the AT model on the self-dual line.
As for all other cases studied so far numerically (i.e. Heisenberg \cite{fps-08}, Ising \cite{atc-09,fc-10}, and XY \cite{fc-10} chains),
strong scaling corrections affect the determination of the scaling function $F_n(x)$.
CFT predictions have been confirmed only using the general theory of corrections to the scaling \cite{ccen-10,ce-10,cc-10,ccp-10}.
In order to determine the function $F_n(x)$, we consider the ratio
\begin{equation}
F^{\rm lat}_n(x)= \frac{{\rm Tr}\, \rho_{A_1\cup A_2}^n}{{\rm Tr}\, \rho_{A_1}^n {\rm Tr}\, \rho_{A_2}^n} (1-x)^{c(n-1/n)/6 }\,,
\label{Flat}
\end{equation}
and, on the basis of the general CFT arguments \cite{cc-10}, we expect that the
the leading correction to scaling can be effectively taken into account by the scaling ansatz
\begin{eqnarray}
F_n^{\rm lat}(x)=F_n^{\rm CFT}(x)+\ell^{-2\omega/n}f_n(x)+\dots\,.
\label{ansatz}
\end{eqnarray}
For the Ising model it has been found $\omega=1/2$ \cite{atc-09,fc-10}.
Since for $\eta=2$ the AT Hamiltonian reduces to two uncoupled Ising models, one naively expects
$\omega=K_L/2$ along the whole self-dual critical line of the AT model.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.7\textwidth]{teller_x05_all.png}
\end{center}
\caption{$F_2^{\rm lat}(1/2)$ as function of $\eta^{-1}$ for different models (Ising, SUSY, $Z_4$ parafermions, and four-states Potts model).
The blue-dashed line is the CFT prediction.
The (colored) points close to the curve are extrapolations obtained with the finite-size scaling ansatz (\ref{ansatz}).
The black crosses are the Monte Carlo data used for the fits.
The block lengths used range from $\ell=5$ to $\ell=80$.}
\label{all}
\end{figure}
Hereafter we only consider ${\rm Tr}\,\rho^2_{A}$.
We start our analysis from the SUSY point that (assuming $\omega=K_L/2$) should have the smaller corrections to scaling.
In Fig. \ref{Susy_F2} we show Monte Carlo data at $\ell=10,20$ ($L=120$) for $F_2^{\rm lat}(x)$ plotted against the
four point ratio $x$ defined as in Eq. (\ref{xFS}).
We report with the blue dashed line the asymptotic CFT result (cf. Eq. (\ref{atf2})).
As in all other cases considered in the literature \cite{atc-09,fc-10}, the curves for $F_2(x)$ at $\ell=10,20$ are not
symmetric functions of $x\to1-x$, as instead the asymptotic CFT prediction must always be \cite{fps-08}.
This is due to the non-symmetrical finite-size corrections $f_2(x)$ in Eq. (\ref{ansatz}).
We extrapolate the result at $\ell\to\infty$ using the ansatz (\ref{ansatz}) and $\omega=2/3$.
The extrapolations are reported as red points in Fig. \ref{Susy_F2}.
There is a very good agreement between the extrapolations and the theoretical curve.
Since the correction exponent $\omega=2/3$ is rather large, and so the corrections small,
even small subsystems such as $\ell=10,20$ are enough to obtain a good extrapolation.
In the inset of Fig. \ref{Susy_F2} we report the Monte Carlo data for $F^{\rm lat}_2(x)$ against $\ell^{-2/3}$.
The linear behavior in this inset confirms the validity of the ansatz (\ref{ansatz}) and the reported
straight lines are the fits giving the extrapolations reported in the main panel.
\begin{figure}[t]
\includegraphics[width=.6\textwidth]{teller_fits.png}
\includegraphics[width=.6\textwidth]{teller_fits_2.png}
\caption{$F_2^{\rm CFT}(1/2)-F_2^{\rm lat}(1/2)$ versus $1/\ell$.
The dashed lines are fits to the finite-size scaling ansatz (\ref{ansatz}) fixing the value of $F_2^{\rm CFT}(1/2)$.
Left: the same plot in log-log scale. }
\label{fits}
\end{figure}
We also investigate other points on the self dual line, namely the $4$-states Potts model ($\eta=4$), the
parafermion $Z_4$ ($\eta=3$), the uncoupled Isings ($\eta=2$).
In Fig.~\ref{all} we report $F^{\rm lat}_2(x)-F_2^{\rm CFT}(x)$ at fixed $x=1/2$ versus $\eta^{-1}$ for all the mentioned models.
we report $x=1/2$ because it is the value of $x$ providing the most stable estimate, but also other values have been studied.
Indeed, on one hand, the computational cost of the simulations decreases going toward $x=1$ (the reason being evident from
the definition of $x$ for which smaller lattice sizes are needed).
On the other hand, scaling corrections become more severe in the region $x\sim 1$, as clear from the results for the SUSY model
in Fig. \ref{Susy_F2}.
Thus $x=1/2$ represents the best compromise between these two drawbacks.
The dashed curves in the left panel of Fig.~\ref{fits} are fits of the data with Eq. (\ref{ansatz}) obtained by fixing the value of
$F_2^{\rm CFT}(x)$ to its predicted value (cf. Eq. (\ref{atf2})).
There is a very good agreement with the full theoretical picture, confirming in particular the correctness of the exponent governing
the leading correction to the scaling.
For the $Z_4$ parafermions and for the four-state Potts model, we needed very large values of $\ell$ in order to
show the correct asymptotic behavior (the range of $\ell$ reported in the plot is in fact $5\le\ell\le80$).
This is made clearer in the right panel of Fig. \ref{fits} where the same data are shown in log-log scale.
In Fig. \ref{all} we reports the fits obtained by
fixing only the exponent of the corrections $\omega=K_L/2$ and leaving $F_2^{\rm CFT}(1/2)$ free.
For all considered values of $\eta$, the extrapolation of $F^{\rm lat}_2(1/2)$ to $\ell\to\infty$ is compatible (within error bars)
with the expected result $F_2^{\rm CFT}(1/2)$.
We finally study the correction amplitude $f_2(x)$ in Eq. (\ref{ansatz}).
This function is the main reason of the asymmetry in $x\to1-x$ for $F_2^{\rm lat}(x)$ and knowing its gross features
could greatly simplify future analyses.
For the Ising model, it has been found that $f_2(x)\sim x^{1/4}$ for small $x$, that is the same behavior of $F_2(x)-1$.
Since along the whole self-dual line $F_2(x)-1\sim x^{1/4}$, we would expect
\begin{equation}
f_2(x)\sim x^{1/4}\,.
\label{f2hyp}
\end{equation}
For the Ising model (i.e. $\eta=2$), this scenario has been already verified with high precision \cite{atc-09}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.7\textwidth]{sub_corr.png}
\end{center}
\caption{Monte Carlo data for $f_2(x)$ obtained as
$f_2(x)=(F^{\rm lat}_2(x)-F_2^{CFT}(x))\ell^{K_L/2}$ as function of $x$. We show data for
$\ell=10$ and various models (SUSY, $Z_4$ parafermions,
Ising model and the model corresponding to $\eta^{-1}=0.74$). The
blue-dashed lines are asymptotic fits to $Ax^{1/4}$.}
\label{sub}
\end{figure}
In Fig. \ref{sub} we report $f_2(x)$ obtained as $f_2(x)=(F_2^{\rm CFT}(x)-F_2^{\rm lat})\ell^{K_L/2}$
as function of $x$ (in logarithmic scale to highlight the small $x$ behavior).
All data correspond to $\ell=10$ and various values of $L$.
For the two largest values of $\eta$ ($Z_4$ parafermionic theory at $\eta=3$ and for the Ising model at $\eta=2$),
we observe an excellent agreement with our conjecture $f_2(x)\sim x^{1/4}$.
However decreasing the value of $\eta$, i.e. for the SUSY model at $\eta=3/2$ and for the model at $\eta^{-1}=0.74$,
the behavior of $f_2(x)$ is not as linear as before, especially for high value of $x$.
Nonetheless for $x<0.4$ the data confirm the behavior $x^{1/4}$.
Furthermore, it seems that for any $\eta\neq2$, subleading terms in the expansion for small $x$ appear
and they are vanishing only for the Ising model.
\section{The Tree Tensor Network}
\label{ttn:sec}
This section is divided into two parts.
First we explain in a self contained way how to extract the spectrum of the reduced density matrix of
some specific bipartitions of a pure state encoded in a Tree Tensor Network (TTN).
We only recall the basic definitions introduced in Ref. \cite{TTN} and refer the reader to the literature for complementary
works on the subject \cite{fnw-92,f-97,hieida,lcp-00,mrs-02,sdv-06,nagaj-08,silvi,Dur,Murg,Plenio,Gliozzi,gauge}.
Secondly we quickly recall how to use TTN to calculate the ground state of the anisotropic Heisenberg spin-chain.
\subsection{Tree Tensor network and reduced density matrices.}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.8\textwidth]{SmallTree.png}
\caption{Examples of TTN for a $N=4$ lattice and a $N=8$ lattice.}
\label{fig:SmallTree}
\end{center}
\end{figure}
We consider a one dimensional lattice $\mathcal{L}$ made of $N $ sites, where each site is described by a local Hilbert
space $\mathbf{V}$ of finite dimension $d$.
In this work the state is the ground state $|\Psi_{\mbox{\tiny GS}}\rangle$ of some local Hamiltonian $H$ defined on $\mathcal{L}$,
but in general it could be an arbitrary pure state $|{\Psi}\rangle \in \mathbf{V}^{\otimes N}$ defined on the lattice $\mathcal{L}$.
A generic state $|\Psi\rangle\in \mathbf{V}^{\otimes N}$ can always be expanded as
\begin{eqnarray}
|\Psi\rangle = \!\sum_{i_1=1}^d ~ \sum_{i_2=1}^d \cdots \sum_{i_N=1}^d T_{i_1i_2 \cdots i_N}
| i_1\rangle| i_2 \rangle \cdots | i_N \rangle,
\label{eq:local_expansion}
\end{eqnarray}
where the $d^{N}$ coefficients $T_{i_1i_2 \cdots i_N}$ are complex numbers and the vectors
$\{| 1_s \rangle, |2_s\rangle, \cdots, |d_s\rangle \}$ denote a local basis on the site $s\in \mathcal{L}$.
We refer to the index $i_s$ that labels a local basis for site $s$ ($i_s=1,\cdots,d$) as a \emph{physical} index.
In the case we are interested in, the tensor of coefficients $T_{i_1i_2 \cdots i_N}$ in Eq. (\ref{eq:local_expansion}) is the result
of the contraction of a TTN.
As shown in Fig. \ref{fig:SmallTree} for lattices of $N=4$ and $N=8$ sites, a TTN decomposition of $T_{i_1i_2 \cdots i_N}$ consists
of a collection of tensors $w$ that have both \emph{bond} indices and \emph{physical} indices.
The tensors are interconnected by the bond indices according to a tree pattern.
The $N$ physical indices correspond to the leaves of the tree.
Upon summing over all the bond indices, the TTN produces the $d^N$ complex coefficients $T_{i_1i_2 \cdots i_N}$ of
Eq. (\ref{eq:local_expansion}).
\begin{figure}[t]
\begin{center}
\includegraphics[width=8cm]{Isometry}
\caption{ (i) Diagrammatic representation of the two types of isometric tensors in the TTN for a $N=4$ lattice in Fig. \ref{fig:SmallTree}.
(ii) Graphical representation of the constraints in Eqs. (\ref{eq:const1}) and (\ref{eq:const3}) fulfilled by the isometric tensors. }
\label{fig:Isometric}
\end{center}
\end{figure}
The tensors in the TTN will be constrained to be \emph{isometric}, in the following sense.
As shown in Fig. \ref{fig:Isometric} for the $N=4$ lattice of Fig. \ref{fig:SmallTree}, each tensor $w$ in a TTN has at
most one upper leg/index $\alpha$ and two lower indices/legs $\beta_1, \beta_2$, so that its entries read
$(w)^{\alpha}_{\beta_1,\beta_2}$ (everything can be generalized to tensors with more upper and lower legs \cite{TTN}).
Then we impose that
\begin{equation}
\sum_{\beta_1 , \beta_2} (w)_{\beta_1 , \beta_2}^{\alpha}(w^{\dagger})^{\beta_1 , \beta_2}_{\alpha'} = \delta_{\alpha\alpha'}.
\label{eq:isometry}
\end{equation}
For clarity, throughout this paper we use diagrams to represent tensors networks as well as tensor manipulations.
For instance, the constraints for the tensors $w_1$ and $w_2$ of the TTN of Fig. \ref{fig:SmallTree} for a $N=4$ lattice, namely
\begin{eqnarray}
\sum_{\beta_1 \beta_2 } (w_1)_{\beta_1 \beta_2}^{\alpha} (w_1^{\dagger})^{\beta_1 \beta_2}_{\alpha'} &=& \delta_{\alpha\alpha'}, \label{eq:const1}\\
\sum_{\beta_1 \beta_2} (w_2)_{\beta_1 \beta_2}(w_2^{\dagger})^{\beta_1 \beta_2} &=& 1,\label{eq:const3}
\end{eqnarray}
are represented as the diagrams in Fig. \ref{fig:Isometric}(ii).
We refer to a tensor $w$ that fulfills Eq. (\ref{eq:isometry}) as an \emph{isometry}.
An intuitive interpretation of the use of a TTN to represent a state $|\Psi\rangle$ can be obtained in terms of a
coarse-graining transformation for the lattice $\mathcal{L}$.
Notice that the isometries $w$ in Fig. \ref{fig:SmallTree} are organized in layers.
The bond indices between two layers can be interpreted as defining the sites of an effective lattice.
In other words, the TTN defines a sequence of increasingly coarser lattices
$\{\mathcal{L}_0, \mathcal{L}_1, \cdots, \mathcal{L}_{T-1} \}$, where
$\mathcal{L}_0 \equiv \mathcal{L}$ and each site of lattice $\mathcal{L}_{\tau}$ is defined in terms of several sites of
$\mathcal{L}_{\tau-1}$ by means of an isometry $w_{\tau}$, see Fig. \ref{fig:CoarseGrain}.
In this picture, a site of the lattice $\mathcal{L}_{\tau}$ effectively corresponds to some number $n_{\tau}$ of sites of the original
lattice $\mathcal{L}_0$.
For instance, each of the two sites of $\mathcal{L}_{2}$ in Fig. \ref{fig:CoarseGrain} corresponds to $8$ sites of $\mathcal{L}_0$.
Similarly, each site of lattice $\mathcal{L}_{1}$ corresponds to $4$ sites of $\mathcal{L}_0$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=8cm]{CoarseGraining}
\caption{The isometric TTN of Fig. \ref{fig:SmallTree} for a $N=8$ lattice $\mathcal{L}_0$ with periodic boundary conditions
(the blue external circle) is associated with a coarse-graining transformation that generates a sequence of increasingly
coarse-grained lattices $\mathcal{L}_1$, $\mathcal{L}_2$ and $\mathcal{L}_3$ (the inner circles).
Notice that in this example we have added an extra index to the top isometry $w_{3}$, corresponding to the single site of an
extra top lattice $\mathcal{L}_3$, which we can use to encode in the TTN a whole subspace of $\mathbf{V}^{\otimes N}$
instead of a single state $|\Psi\rangle$.}
\label{fig:CoarseGrain}
\end{center}
\end{figure}
The use of isometric tensors, and the fact that each bond unambiguously defines two parts $(A:B)$ of the chain which are connected only through that bond as displayed in Fig. \ref{fig:bond}, implies that the rank of that bond in the TTN is given by the Schmidt rank $\chi(A:B)$ of the partition $(A:B)$ \cite{sdv-06}.
Thus the reduced density matrix $\rho_A$ for a set $A$ of sites of $\mathcal{L}$ is
\begin{equation}
\rho_A = \tr_{B} \proj{\Psi} = \sum_{\alpha} p_{\alpha} \proj{\Psi^{A}_{\alpha}},
\label{eq:rhoA}
\end{equation}
where
$p_{\alpha}$ are the eigenvalues of $\rho_A$.
It follows then the R\'enyi entanglement entropies $S_A^{(n)}$ are
\begin{equation}
S_A^{(n)}=\frac1{1-n}\log{\rm Tr}\,\rho_A^n=
\frac1{1-n} \log \sum_{\alpha} p_{\alpha}^n\,,
\end{equation}
and for $n=1$
\begin{equation}
S_A^{(1)}= -\tr(\rho_A \log \rho_A) = -\sum_{\alpha} p_{\alpha} \log p_{\alpha}.
\label{eq:entropy}
\end{equation}
\begin{figure}[t]
\begin{center}
\includegraphics[width=6cm]{bipartitions}
\caption{By erasing one of the indices in the TTN the spin chain is always divided in two parts $A$ and $B$ \cite{sdv-06}.
Here we show that in the case of the $N=8$ lattice of Fig. \ref{fig:SmallTree} there are three classes of indices, identified by
their position in the TTN.
i) physical bonds connect a single spin with the rest of the lattice,
ii) bond indices of the first layer connect a block of two adjacent spins to the rest of the lattice,
iii) bond indices of the third layer of the lattice connect four adjacent spins, to the other half.
This implies that the rank of the index is the Schmidt rank of the respective partition.
}
\label{fig:bond}
\end{center}
\end{figure}
In the following we denote the ranks of the tensor $w_{\tau}$, $\alpha, \beta_1, \beta_2$
as $, \chi^{\tau},\chi^{\tau -1}, \chi^{\tau -1}$. In general, they fulfill
\begin{equation}
\chi^{\tau} < (\chi^{\tau -1})^2,
\end{equation}
meaning that $w_{\tau}$ projects states in $\mathbf{V}^{\tau-1}\otimes \mathbf{V}^{\tau-1}$ into the smaller Hilbert
space $\mathbf{V}^{\tau}$.
For a critical chain, the logarithmic scaling of the entanglement entropy (cf. Eq. (\ref{Renyi:asymp}))
implies that the rank of the isometries should at least grow proportionally to the length of the block represented by the effective spins
\begin{equation}
\chi^{\tau} \propto n_{\tau},
\end{equation}
which means that while moving to higher layer of the tensor network the rank of the isometries increases.
This also implies that the leading cost of the computation is concentrated in contracting the first few layers of the TTN.
If $N=2^T$ and we describe a pure state (so that the rank of the $\alpha_{\tau}$ is one) the maximal rank of the tensors in the
TTN is
\begin{equation}
\chi=\max_{\tau} \chi ^{\tau}=\chi^{T-1}.
\end{equation}
In Ref. \cite{TTN} it has been shown that
i) a TTN description of the ground state of chain of length $N$ with periodic boundary conditions can be obtained numerically with
a cost of order $\mathcal{O}(\log N \chi^4)$.
ii) From the TTN it is also straightforward to compute the spectrum $\{p_{\alpha}\}$ of the reduced density matrix $\rho_A$
(cf. Eq. (\ref{eq:rhoA})) when $A$ is a block of contiguous sites corresponding to an effective site of
any of the coarse-grained lattices $\mathcal{L}_1, \cdots, \mathcal{L}_{T-1}$.
Fig. \ref{fig:SpectEval} illustrates the tensor network corresponding to $\rho_A$ for the case when $A$ is one half of the chain.
Many pairs of isometries are annihilated. In addition, the isometries contained within region $A$ can be removed since they
do not affect the spectrum of $\rho_A$.
From the spectrum $\{p_{\alpha}\}$, we can now obtain the R\'enyi entropies $S_A^{(n)}$.
The leading cost for computing the spectrum of the reduced density matrix $\rho_A$ for this class of bipartitions is due to
the contractions of the first layers of the TTN.
When the bipartition is such that $A$ is a quarter of the chain, this implies a cost proportional to
$\mathcal{O}(\chi'^3\chi^2)\le \chi' \chi^4 $, where $\chi'=\chi^{T-2}$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=9cm]{SpectEval}
\caption{Computation of the spectrum $\{p_{\alpha}\}$ of the reduced density matrix $\rho_A$ for a block $A$ that corresponds
to one of the coarse-grained sites.
(i) Tensor network corresponding to $\rho_A$ where $A$ is half of the lattice.
(ii) Tensor network left after several isometries are annihilated with their Hermitian conjugate.
(iii) since the spectrum of $\rho_A$ is not changed by the isometries acting on $A$, we can eliminate them and
we are left with a network consisting of only two tensors, which can now be contracted together.
The cost of this computation is
proportional to $\mathcal{O}(\chi'^3\chi^2)\le \mathcal{O}( \chi^5 )$.}
\label{fig:SpectEval}
\end{center}
\end{figure}
It is also possible to compute the reduced density matrix $\rho_A$ when $A$ is composed of two disjoint subintervals
$A_1$ and $A_2$,
where now each of the two intervals is a block of contiguous sites corresponding to an effective site of the coarse grained lattice.
The cost of this computation is again dominated by contracting the upper part of the tensor network, and the most expensive case is
obtained by considering $A$ as the collection of two $N/4$ spins blocks, separated by $N/4$ spins.
The tensor network corresponding to this $\rho_A$ is shown in Fig. \ref{fig:SpectEvaltwoBlocks}.
Also in this case many pairs of isometries are annihilated. The isometries contained within the composed region $A$ can also
be removed since they do not affect the spectrum of $\rho_A$.
The cost of contracting this tensor network is proportional to
$\max [\mathcal{O}(\chi^2 \chi'^4), \mathcal{O}(\chi^3 \chi'^2)] < \mathcal{O}(\chi^6)$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=9cm]{SpectEvalTwoBlocks}
\caption{Computation of the spectrum $\{p_{\alpha}\}$ of the reduced density matrix $\rho_A$ when $A$ corresponds to two
coarse-grained sites separated by one coarse grained site from both sides.
(i) Tensor network corresponding to $\rho_A$ where $A$ is a quarter of the lattice.
(ii) Tensor network left after several isometries are annihilated with their Hermitian conjugate.
(iii) Since the spectrum of $\rho_A$ is not changed by the isometries acting on $A$, we can eliminate them and we
are left with a network consisting of only few tensors, which can now be contracted together.
The cost of contracting this tensor network is proportional to
$\max[\mathcal{O}(\chi^2 \chi'^4),\mathcal{O}(\chi^3 \chi'^2)] < \mathcal{O}(\chi^6)$.}
\label{fig:SpectEvaltwoBlocks}
\end{center}
\end{figure}
\subsection{The TTN and the anisotropic Heisenberg spin-chain}
In the previous subsection we have shown how to extract the spectrum of the reduced density matrix for a single and a double
spin block from a TTN state.
In this manuscript we are interested in reduced density matrices calculated on the ground-state of the anisotropic
Heisenberg spin chain (XXZ model) in zero magnetic field, defined by the Hamiltonian
\begin{equation}
H=\sum_{j=1}^L [\sigma^x_j\sigma^x_{j+1}+\sigma^y_j \sigma^y_{j+1}+\Delta
\sigma^z_j\sigma^z_{j+1}]\, ,
\label{HXXZ}
\end{equation}
where $\sigma_j^\alpha$ are the Pauli matrices at the site $j$. Periodic boundary conditions are assumed.
We are interested in gapless conformal phases of the model, that is $-1<\Delta\leq 1$.
This phase is described by a free-bosonic CFT compactified on a circle with radius that depends on the parameter $\Delta$
\begin{equation}
\eta=2r_{\rm circle}^2=\frac1{2K_L}=\frac{\arccos (-\Delta)}\pi\,,
\label{etaDe}
\end{equation}
where $K_L$ is the Luttinger liquid parameter. \footnote{Notice similarities and differences between Eq. (\ref{etaDe}) and
its analogous for the AT model (\ref{etaAT}). The relation between $\eta$ and $r^2$ and the relation between $K_L$ and
$\Delta$ are the same for both XXZ spin-chain and AT model, but the relation between $\eta$ and $\Delta$ (or $K_L$ and $r$)
is different.
}
The sign convention in the Hamiltonian (\ref{HXXZ}) is such that the model is (anti)ferromagnetic for $\Delta<0$ ($\Delta>0$).
Hamiltonian (\ref{HXXZ}) is diagonalizable by means of Bethe ansatz.
However, obtaining the spectrum of the reduced density matrix from Bethe ansatz is still a major problem
and only results for small subsystems are known \cite{afc-09,ncc-09}.
For this reason we exploit variational TTN techniques to obtain the ground state.
Here we follow the variational procedure described in detail in Ref. \cite{TTN},
where the generic technique (consisting of assuming a tensor network
description of the ground state and minimize the energy variationally
improving the tensors one by one as described, i.e., in Ref. \cite{cv-09}) has been specialized and optimized for the case of a TTN.
We exploit translation invariance by using the same tensor at each layer of the TTN.
One could also improve the efficiency further by exploiting the $U(1)$ symmetry of the Hamiltonian (\ref{HXXZ}), i.e. the
rotations around the $z$ axis. However we did not make use of this symmetry here.
\section{The Block Entanglement of the Anisotropic Heisenberg spin-chain}
\label{XXZ:sec}
\begin{figure}[t]
\begin{center}
\includegraphics[width=.8\textwidth]{XXZ_single.png}
\end{center}
\caption{TTN data for the non universal constant $c_2(L_c)$ as function of the chord length $L_c$
for different values of $\Delta$. The dashed curves are fits
to the function $A+BL_c^{-K_L}$. The reported data have been
obtained with $L=128$ for $\Delta=0,0.1,0.6$ and $L=64$ for the
other values.}
\label{XXZsingle}
\end{figure}
In this section we report the TTN results for the R\`enyi entropies in the XXZ spin-chain for a single and a double interval.
As a main advantage compared to the classical Monte Carlo simulations performed for the AT model, with a single TTN simulation we
obtain the spectrum of the reduced density matrix and hence any R\`enyi entropy, including von Neumann $S_A^{(1)}$.
Oppositely with the Monte Carlo methods only R\`enyi entropies $S_A^{(n)}$ of integer order $n\geq2$
can be obtained and each of them requires an independent simulation.
\subsection{The single interval.}
We first present the TTN results for the single interval.
These have been already obtained with many numerical variational techniques \cite{ccen-10,lsca-06,osc,xa-11}
and are reported here only to test the accuracy of the TTN and to fix units/scales etc.
Using variational TTN, we find the ground-state of the XXZ Hamiltonian (\ref{HXXZ}) and from this we extract the
spectrum of the reduced density matrix of the single block, as explained in the previous section.
We then numerically obtain ${\rm Tr}\,\rho_A^n$.
The maximum size of the chain that we consider is $L=128$.
The subsystem lengths considered are $\ell=2,4,8,16,32$.
Notice that with the TTN method, using a binary tree as we are doing, we can effectively access only subsystems sizes of the form
$2^m$ with $m$ arbitrary integer, as it should be clear from the previous section.
In particular this limits the calculation to even values of $\ell$ and we can not study the parity effects reported
in Ref. \cite{ccen-10,ce-10}.
We considered different values of the anisotropy parameter $\Delta$, namely $\Delta=-0.3,-0.1,0,0.1,0.2,0.4,0.6,0.8,1$.
The TTN becomes less effective for values of $\Delta\leq-0.5$. This can be easily traced back to the smallness
of the finite-size gap that in the minimization process causes the algorithm to be stuck in meta-stable states when the
system size is large enough. This drawback could be cured by using larger values of $\chi$ (and so larger computational cost), but
as we shall see, the considered values of $\Delta$ suffice to draw a very general picture of the entanglement.
For the isotropic Heisenberg antiferromagnet at $\Delta=1$ we ignore the presence of logarithmic corrections
to the scaling \cite{lsca-06,cc-10}, that have a minimal effect for all our aims.
\begin{figure}
\begin{center}
\includegraphics[width=.8\textwidth]{XXZtr_2.png}
\end{center}
\caption{TTN data for $F^{\rm lat}_2(x)$ as function of $x$ for various sizes of the chain $L=16,32,64,128$, subsystem lengths
$\ell=4,8,16,32$, and $\Delta=-0.3,-0.1,01,0.6$.
Different values of $\Delta$ are distinguished by different colors, while different symbols denote different
values of $\ell$. The arrows denote the (asymptotically) increasing subsystem sizes $\ell$. }
\label{XXZtr_2}
\end{figure}
As for the AT model, we study the quantity $c_2(L_c)$ defined by the ratio in Eq. (\ref{c2L}).
The results are shown in Fig. \ref{XXZsingle} for all considered values of $\Delta$.
The scaling corrections are evident, especially for larger values of $\Delta$, as expected \cite{ccen-10}.
These corrections for ${\rm Tr}\rho_A^n$ are indeed of the form $L_c^{-2K_L/n}$ \cite{ccen-10} ($K_L$ is defined in Eq. (\ref{etaDe})).
The dashed lines reported in Fig.~\ref{XXZsingle} are fits to this form for $n=2$, showing the
agreement between TTN data and the fits.
We checked that all the TTN data agree with the ones obtained in Ref. \cite{ccen-10} using density matrix renormalization group.
The agreement is perfect and for this reason we refer to the above paper for a detailed study of ${\rm Tr}\,\rho_A^n$ for $n>2$.
\subsection{Double interval: the $n=2$ case.}
We now consider a subsystem made of two parts $A_1$ and $A_2$ of equal length $\ell$.
We start by studying the quantity ${\rm Tr}\,\rho^2_{A_1\cup A_2}$ for finite chains and extract the
universal function $F_2^{\rm CFT}(x)$ by proper extrapolation.
Since we only consider even $\ell$, corrections to the scaling are expected to be monotonic in $\ell$ also for $F_2(x)$,
oppositely to the case of arbitrary $\ell$ parity \cite{fps-08,fc-10}.
The CFT prediction for the function $F_2(x)$ for the XXZ chain is Eq. (\ref{F2}) with $\eta$ given by Eq. (\ref{etaDe}).
\begin{figure}[t]
\begin{center}
\includegraphics[width=.8\textwidth]{XXZfits.png}
\end{center}
\caption{TTN data for $F_2^{\rm lat}(1/2)-F_2^{\rm CFT}(1/2)$ as function of $1/\ell$ for various
$\Delta$. The dashed lines are fits to the function with the generalized finite-$\ell$ ansatz (\ref{ansatz2}).
}
\label{XXZfits}
\end{figure}
In Fig. \ref{XXZtr_2} we report TTN data for $F^{\rm lat}_2(x)$ (obtained with the ratio defined in Eq. (\ref{Flat})) as function of the cross
ratio $x$ for $\Delta=-0.3,-0.1,0.1,0.6$ and subsystem sizes
$\ell=4,8,16,32$.
The different values of $\Delta$ are denoted with different colors, while the different symbols stand for the various $\ell$.
On the same figure we also show the asymptotic $F_2^{\rm CFT}(x)$ as dashed lines.
It is evident that strong scaling corrections affect the data, as expected.
Colored arrows denote the direction of (asymptotically) increasing subsystem sizes.
Very surprisingly, while for $\Delta=-0.3,-0.1,0.1$ the asymptotic CFT result is approached from below,
for $\Delta=0.6$ it is approached from above.
Moreover, for $\Delta=0.6$ the behavior of the data is not monotonic.
This contrasts the results obtained for the AT model in the previous sections and the ones obtained for the XX and Ising
spin-chains \cite{fc-10}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.7\textwidth]{XXZtr_3.png}
\end{center}
\caption{TTN data for $F_3^{\rm lat}(x)$ as function of $x$ for various sizes of the chain,
$\Delta=-0.3,0.1,0.6,1$, and subsystem lengths $\ell=4,8,16,32$. We
denote with different symbols the values of $\ell$ and with different
colors the various $\Delta$. The dashed curves are the theoretical
results given by Eq. (\ref{Fnv}). The arrows denote the
(asymptotically) increasing subsystem sizes $\ell$.}
\label{XXZtr_3}
\end{figure}
In order to shed some light on this unexpected phenomenon, it is worth to look at $F_2^{\rm lat} (x)$ as functions of $\ell$ for fixed
values of $x$.
In Fig. \ref{XXZfits} we report one of these plots for $x=1/2$. Analogous figures are obtained for other values of $x$.
Corrections to the scaling are non-monotonic in the range $0.2\leq\Delta\leq0.7$.
This phenomenon can be understood if further corrections to the scaling are taken into account.
There are two corrections that can be responsible of this behavior.
On the one hand, corrections of the form $\ell^{-m K_L}$ (from $\ell^{-2mK_L/n}$ at $n=2$) for any integer $m$ are know to
be present \cite{ce-10},
on the other hand usual analytic corrections such as $\ell^{-1}$ are generically expected to exist for any quantity from general
scaling arguments.
Thus the most general finite-$\ell$ ansatz has the form
\begin{equation}
F_2^{\rm lat}(x)=F_2^{\rm CFT}(x)+\frac{f_2(x)}{\ell^{K_L}} +\frac{f_A(x)}{\ell}+\frac{f_B(x)}{\ell^{2K_L}}\dots\,,
\label{ansatz2}
\end{equation}
where the first correction is the {\it unusual} one employed also for the Ashkin-Teller model,
and the other two are the ones just discussed.
The effect of subleading corrections is enhanced by the fact the the amplitude functions $f_2(x)$ and
$f_A(x)$ or $f_B(x)$ have opposite signs determining the non-monotonic behavior.
Unfortunately, for values of $\Delta$ for which the effect of subleading corrections is more pronounced (i.e. $0.1\leq\Delta\leq0.6$),
we have $K_L<1<2K_L$, making difficult to disentangle corrections with close exponents.
Thus, in order to present analyses of a good quality, we ignore the last correction (i.e. we fix $f_B(x)=0$).
To check the proposed scenario, we performed the fit of the data in Fig. \ref{XXZfits} with the ansatz (\ref{ansatz2}) and $f_B(x)=0$.
The results of the fits are reported in the same figure, showing perfect agreement with the data for all the values of $\Delta$.
We repeated the same analysis for other values of $x$, finding the same quality of fits as for $x=1/2$.
However, we cannot exclude that corrections of the form $\ell^{-2K_L}$ have an important role.
\subsection{Double interval: the $n=3$ case.}
Now we report the same analysis performed for ${\rm Tr}\, \rho_A^2$ for the third moment of $\rho_A$, i.e. ${\rm Tr}\, \rho_A^3$.
Again we consider finite-size XXZ spin-chains and extract the universal function $F^{\rm CFT}_3(x)$ by finite-size analysis.
The expected CFT result is given for general $n$ by Eq. (\ref{Fnv}).
In Fig.~\ref{XXZtr_3} we show TTN data for $F^{\rm lat}_3(x)$ (obtained from Eq. (\ref{Flat}))
at $\Delta=-0.3,0.1,0.6,1$ and subsystem sizes up to $\ell=32$.
We also show the theoretical curves given by Eq. (\ref{Fnv}).
As for the $n=2$, the asymptotic universal curve is approached from below for
$\Delta\le 0.6$, and from above for $\Delta\geq 0.6$.
Furthermore, the behavior of the numerical data for $\Delta>0.6$ is non monotonic.
This suggests that the ansatz in Eq. (\ref{ansatz}) is not enough to describe accurately the TTN data and further corrections
to the scaling should be included as for ${\rm Tr}\, \rho_A^2$.
For $n=3$, the leading corrections to the scaling are described by the ansatz (\ref{ansatz}), i.e. the leading exponent is $2K_L/3$.
Thus, for the cases when subleading corrections are more important (i.e. for $\Delta\geq0.6$) the ordering of the exponents
is $2K_L/3<4K_L/3<1$ and so it is reasonable to ignore the analytic correction.
Thus we fit TTN data with the function
\begin{equation}
F^{\rm lat}_3(x)-F_3^{\rm CFT}(x)=f_3(x)\ell^{-2K_L/3}+f_B(x)\ell^{-4K_L/3}\,.
\label{f3fits}
\end{equation}
In Fig.~\ref{XXZtr_3x05} we report TTN data for
$F^{\rm lat}_3(x)-F_3^{\textrm{CFT}}(x)$ for $x=1/2$ and several values of $\Delta$.
The dashed lines are fits with the finite-size ansatz (\ref{f3fits}), that perfectly reproduce the data.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.7\textwidth]{XXZfits_tr3.png}
\end{center}
\caption{TTN data for $F^{\rm lat}_3(x)$ at fixed $x=1/2$ as function of $\ell^{-1}$ for $\ell=4,8,16,32$,
The considered values of $\Delta$ are $\Delta=-0.3,-0.1,01,0.6,1$. The dashed curves are fits with the ansatz (\ref{f3fits})}
\label{XXZtr_3x05}
\end{figure}
\subsection{Double interval: The von Neumann entropy.}
TTN gives access to the full spectrum of the reduced density matrix of $A_1\cup A_2$ and so to the entanglement
entropy $S_1^{(n)}$ as well.
In Fig. \ref{XXZvn} we report the function $F_{VN}^{\rm lat}(x)$ defined as
\begin{equation}
F_{VN}^{\rm lat}(x)=S_{A_1\cup A_2}^{(1)}-S_{A_1}^{(1)}-S_{A_2}^{(1)}-\frac13\log (1-x)\,,
\end{equation}
for $\Delta$ in the interval $[-0.3,1]$ for various $L$ up to 128 and subsystem sizes $\ell=2,4,8,16,32,64$.
We indicate with different symbols different values of $\Delta$, while the colors are for various sizes $\ell$.
As known from many other investigations on single and double intervals (quantum Ising spin chain, XY
model, XXZ) the von Neumann entropy does not show oscillations with the parity of the subsystem and
the corrections are much smaller, actually negligible from any practical porpouse.
Fig. \ref{XXZvn} confirms this observation for the two interval entanglement entropy for the XXZ spin-chain in a wide range of $\Delta$.
Indeed, at fixed value of $\Delta$ perfect data collapse is observed even for very small values of $\ell$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.8\textwidth]{XXZvn.png}
\end{center}
\caption{TTN data for the von Neumann entropy for various values of
$\Delta$ in the interval $[-0.3,1]$. We show with different symbols
the values of $\Delta$ while different colors stand for different
$\ell$ and lattice sizes.}
\label{XXZvn}
\end{figure}
Unfortunately, as already stated in the introduction, the CFT prediction for $F_{VN}(x)$ is unknown for general $x$ because
the analytic continuation of $F_n(x)$ to non-integer $n$ is not achievable.
However, an expression for the leading term of the small $x$ expansion of $F_{VN}(x)$ has been recently extracted \cite{cct-11}
from Eq. (\ref{s2cft})
\begin{equation}
F_{VN}(x)=\bigg(\frac{x}{4}\bigg)^\alpha\sqrt{\pi}\frac{\Gamma(\alpha+1)}{2\Gamma(\alpha+3/2)}\,,
\label{smallx}
\end{equation}
where $\alpha=\textrm{min}[\eta,1/\eta]$ and $\Gamma$ is the Euler function (not to be confused with the $\Gamma$
matrix in Eq. (\ref{Gammadef})).
In order to check the correctness of this formula, in Fig.~\ref{XXZvn_log} we report the same data for $F_{VN}(x)$
in a log-log scale to highlight the power-law behavior for small $x$.
We also report the small $x$ expected from Eq. (\ref{smallx}).
For $\Delta=-0.3$ the agreement is good, but it gets worse increasing $\Delta$.
The natural explanation is that the considered values of $x$ are not small enough
for the asymptotic Eq. (\ref{smallx}) to be valid.
We should then include further terms in the small $x$ expansion.
As explained in Ref. \cite{cct-11}, further coefficients in the expansion for small $x$ are difficult to obtain in general.
However, there is a term that is very easy to obtain and that (luckily enough) is responsible of the previous disagreement.
Indeed, as shown in Ref. \cite{cct-11} (cf. Eq. 70 and 71 there) the function $F_n(x)$ has always (i.e. independently of $\eta$)
a simple $O(x)$ contribution coming from the denominator in Eq. (\ref{Fnv}), i.e. $|\Theta(0|\Gamma)|^2=1+ x (n-1/n)/6$, that can be
easily analytically continued giving
\begin{equation}
F_{VN}(x)=\bigg(\frac{x}{4}\bigg)^\alpha\sqrt{\pi}\frac{\Gamma(\alpha+1)}{2\Gamma(\alpha+3/2)}-\frac{x}3+O(x^{2\alpha})\,.
\label{smallx2}
\end{equation}
Notice that the added term becomes more important when $\alpha$ is close to $1$, i.e. in the XXZ spin-chain when $\Delta$
approaches $1$.
In Fig. \ref{XXZvn_log} we also report the prediction (\ref{smallx2}) as dashed line, that is asymptotically in perfect agreement with the
numerical data for all values of $\Delta$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.8\textwidth]{vn1.png}
\end{center}
\caption{TTN data for the von Neumann entropy for
$\Delta=-0.3,0.1,0.6,0.8$ (the data for different $\Delta$ are denoted
with different symbols) in log-log scale. We used different colors to
indicate the different block sizes $\ell$ and lattice sizes $L$.
The continuous lines are the small $x$ behavior obtained from (\ref{smallx}).
The dashed lines are the small $x$ behavior where the $O(x)$ term has been added as in Eq. (\ref{smallx2}).}
\label{XXZvn_log}
\end{figure}
\section{Conclusions}
\label{concl}
In this manuscript we provided a number of results for the asymptotic scaling of the R\'enyi entanglement entropies in
strongly interacting lattice models described by CFTs with $c=1$.
Schematically our results can be summarized as follows.
\begin{itemize}
\item We provided the analytic CFT result for the scaling function $F_2(x)$ for $S_A^{(2)}$ in the case of a free boson compactified
on an orbifold describing, among the other things, the scaling limit of the Ashkin-Teller model on the self-dual line.
The final result is given in Eq. (\ref{atf2}).
\item We developed a cluster Monte Carlo algorithm for the two-dimensional Ashkin-Teller model (generalizing the
procedure of Caraglio and Gliozzi \cite{cg-08} for the Ising model) that gives the scaling functions of the R\'enyi entanglement
entropy (for integer $n$) of the corresponding one-dimensional quantum model.
With this algorithm, we calculated numerically the scaling function $F_2(x)$ of the AT model along the self-dual line
and we confirm the validity of the CFT prediction.
In order to obtain a quantitative agreement,
the corrections to scaling induced by the finite length of the blocks are properly taken into account.
\item We considered the XXZ spin chains by means of a tree tensor network (TTN) algorithm.
The low-energy excitations of model are described by a free boson compactified on a circle for which CFT predictions are already
available both for $n=2$ \cite{fps-08} and for general integer $n$ \cite{cct-09}.
Taking into account the corrections to the scaling, we confirm these predictions (that resisted until now to
quantitative tests) for $n=2,3$.
Furthermore, we provide numerical determinations of the scaling function of the von Neumann entropy (cf. Fig. \ref{XXZvn})
for which CFT predictions do not exist yet for general $x$. For small $x$ we confirm the recent prediction of Ref. \cite{cct-11} (cf.
Fig. \ref{XXZvn_log}).
\end{itemize}
The methods we employed (classical Monte Carlo with cluster observables and TTN) are very general techniques that can be
easily adapted to other models of physical interest. On the CFT side, it must be mentioned that a closed form for the functions
$F_n(x)$ at integer $n$ for a free boson compactified on an orbifold is not yet available,
but work in this direction is in progress \cite{ct-prep}.
\section*{Acknowledgments}
We thank John Cardy, Maurizio Fagotti, Erik Tonni, Ettore Vicari, and Guifre Vidal for useful discussions.
This work has been partly done when PC was guest of the Galileo Galilei Institute in Florence whose hospitality is kindly
acknowledged.
\section*{References}
|
\section{Introduction}
The Great Observatory All-sky LIRGs Survey (GOALS, Armus et al. 2009)
is a multi-wavelength study of the
most luminous infrared galaxies in the local Universe, selected from
the 60~$\mu$m flux limited IRAS Revised Bright Galaxy Sample (RBGS:
Sanders et al. 2003). As the brightest far-infrared selected galaxies
in the sky, these objects are the most amenable for study at all
wavelengths.
Luminous Infrared Galaxies (LIRGs: $L_{\rm ir} > 10^{11} L_{\odot}$
\footnote{$L_{\rm ir} \equiv L(8{-}1000\mu$m)}) have proven to be an
extremely important class of extragalactic objects. In the local
Universe they are more numerous than optically selected starburst and
Seyfert galaxies and quasi-stellar objects (QSOs) at comparable
bolometric luminosity. Strong interactions and mergers of gas-rich
spirals appear to be the trigger for the most luminous infrared
objects, which are fueled by a mixture of intense starbursts and AGN,
with the latter becoming more dominant with increasing $L_{\rm
ir}$. At the highest luminosities, ultraluminous infrared galaxies
(ULIRGs: $L_{\rm ir} > 10^{12} L_{\odot}$), may represent an important
stage in the formation of QSOs and powerful radio galaxies, and they
may also represent a primary stage in the formation of massive
ellipticals, the formation of globular clusters, and the metal
enrichment of the intergalactic medium (see Sanders \& Mirabel 1996,
for a more complete review and, e.g., Chapman et al 2005, Hopkins et al
2006, Veilleux et al 2009 for recent developments).
The X-ray survey data presented here (C-GOALS) is the X-ray component
of the GOALS multi-wavelength survey. This initial C-GOALS paper
presents data obtained by us and others with the Chandra X-ray
Observatory (Chandra, hereafter). Data obtained by us in Cycles $7+8$
have been combined with data from the Chandra Archive to produce the
first complete X-ray survey of the most luminous sources in the GOALS
sample. Previous X-ray investigations of LIRGs, either by Chandra, XMM-Newton
or Suzaku, have been presented in Ptak et al. (2003), Franceschini et
al (2003), Teng et al. (2005, 2009), and Grimes et al. (2005).
The C-GOALS sample and observations are described in \S2 and \S3,
respectively, with the major results (X-ray images, flux density spectra, radial surface
brightness profiles) presented in \S4. Derived properties from the
X-ray spectra, and a discussion of trends of X-ray properties with
infrared luminosity are presented in \S5. A Summary of our results is
given in \S6. Notes for each object are presented in the Appendix.
[Note: A consistent set of detailed images for all targets can be
found in the electronic edition of this paper.] The cosmology adopted
here is consistent with that adopted by Armus et al. (2009). Cosmological distances
were computed by first correcting for the 3-attractor flow model of
Mould et al. (2000) and adopting $H_0 = 70$ km s$^{-1}$Mpc$^{-1}$,
$\Omega_{\rm V} = 0.72$, and $\Omega_{\rm M} = 0.28$ based on the
5-year WMAP results (Hinshaw et al. 2009), as provided by the
NASA/IPAC Extragalactic Database (NED).
\section{The Sample}
\begin{figure}
\begin{center}
\centerline{\includegraphics[width=0.35\textwidth,angle=0]{f01.pdf}}
\caption{The distribution of infrared luminosity, $L_{\rm ir}
(8-1000 \mu$m) for the 44 objects in our C-GOALS sample with
log~$(L_{\rm ir}/L_{\odot})=11.73-12.57$.
For the infrared luminosity, we have chosen to use units of erg s$^{-1}$ (bottom axis) ,
rather than the unit of $L_{\odot}$ (top axis) normally used in infrared astronomy,
in order to be consistent with X-ray luminosity measurements.}
\end{center}
\end{figure}
The RBGS is a complete sample of 629 extragalactic objects with IRAS
60$\mu$m flux density greater than 5.24 Jy, covering the full sky above
Galactic latitude $\vert b\vert >5^{\circ }$. The GOALS sample contains 202 ``systems'' (77 of which
contain multiple galaxies), drawn from the RBGS with the luminosity
threshold $L_{\rm ir}\geq 10^{11}L_{\odot}$ (see Armus et al. 2009 for
further details).
The current C-GOALS sample represents the high luminosity part of the
GOALS sample, and is complete down to log~$(L_{\rm
ir}/L_{\odot})=11.73$ . The object with the largest IR luminosity,
log~$(L_{\rm ir}/L_{\odot})= 12.57$, is UGC 8058 (hereafter the more
common name, Mrk 231, is used in the text). Table 1 gives basic data for all 44
galaxies in the C-GOALS sample. The redshift range is $z = $0.010-0.088.
The median IR luminosity of the sample log~$(L_{\rm ir}/L_{\odot})=11.99$
corresponding to log~$L_{\rm ir} = 45.58$ (erg s$^{-1}$. The distribution of
infrared luminosity is shown in Fig. 1.
\section{Observations and data reduction}
Twenty-six objects were observed with Chandra in Cycle-8 (PI:
D. Sanders) with a uniform 15~ks exposure on each target. All the
observations were carried out in imaging mode with the ACIS-S detector
operated in VFAINT mode. The Chandra data for the remaining 18 targets
were taken from the Archive, including four galaxies in the Cycle 7
mini survey of LIRGs with double nuclei (UGC 4881, VV 705,
F08572+3915, and F14348--1447) (PI: S Komossa). The exposure times for
these 4 objects are also 15~ks each. For other archival data, the
exposure times vary from 10~ks to 154~ks. The observation log for the
44 galaxies is shown in Table 2. Where multiple Chandra observations
are found, an imaging observation with the ACIS-S is chosen. For
Mrk~231, four ACIS-S observations exist in the archive, and we use all
four to construct and analyze the X-ray images, while only the first
observation was used for the analysis of the radial surface brightness
profiles and the X-ray colour, since the source is not variable and
the single observation provides good quality data for this purpose.
In Table 2, the Galactic HI column density, estimated from the
Leiden/Argentine/Bonn (LAB) survey (Kalberla et al. 2005), is also
quoted. Some of the Cycle-8 targets lie at low Galactic latitudes
and thus have been little studied because of high Galactic
extinction. The Galactic HI absorption column exceeds $10^{21}$ cm$^{-2}$\
for six objects.
The data reduction was performed using the Chandra data analysis
package CIAO version 3.4.1, and HEASARC's FTOOLS.
\begin{table*}
\centering
\caption{The C-GOALS Sample.}
\begin{tabular}{rrlcccccl}
No. & IRAS Name & Optical ID & RA (NED) & Dec (NED) & $z$ & $D_{\rm L}$ & log~$(L_{\rm ir})$ & Other name\\
& {} & {} & (J2000) & (J2000) & (km/s) & (Mpc) & ($L_\odot$) & {}\\
{(1)} & {(2)} & {(3)} & {(4)} & {(5)} & {(6)} & {(7)} & {(8)} & {(9)}\\[5pt]
1 & F01364--1042 & IRAS F01364--1042 & 01h38m52.92s& -10d27m11.4s& 0.0483& 210.0& 11.85 & {}\\
2 & F04454--4838& ESO 203-IG1& 04h46m49.50s& -48d33m32.9s& 0.0529& 235.0& 11.86 & {}\\
3 & F05081+7936& VII Zw 31& 05h16m46.44s& +79d40m12.6s& 0.0537& 240.0& 11.99 &{} \\
4 & F05189--2524& IRAS F05189--2524& 05h21m01.47s& -25d21m45.4s& 0.0425& 187.0& 12.16 & {}\\
5 & F06259--4708& ESO 255-IG7& 06h27m22.45s& -47d10m48.7s& 0.0388& 173.0& 11.90 & {}\\
6 & 07251--0248& IRAS 07251--0248& 07h27m37.55s& -02d54m54.1s& 0.0875& 400.0& 12.39 & {}\\
7 & F08520--6850& ESO 60-IG16& 08h52m31.29s& -69d01m57.0s& 0.0463& 210.0& 11.82 & {}\\
8 & F08572+3915& IRAS F08572+3915& 09h00m25.39s& +39d03m54.4s& 0.0584& 264.0& 12.16 & {}\\
9 & 09022--3615 & IRAS 09022--3615& 09h04m12.70s& -36d27m01.1s& 0.0596& 271.0& 12.31 & {}\\
10 & F09111--1007& IRAS F09111--1007& 09h13m37.61s& -10d19m24.8s& 0.0541& 246.0& 12.06 & {}\\
11 & F09126+4432& UGC 4881& 09h15m55.11s& +44d19m54.1s& 0.0393& 178.0& 11.74 & Arp 55\\
12 & F09320+6134& UGC 5101& 09h35m51.65s& +61d21m11.3s& 0.0394& 177.0& 12.01 & {}\\
13 & F10038--3338& ESO 374-IG 032$^{\mathrm{a}}$ & 10h06m04.80s& -33d53m15.0s& 0.0341& 156.0& 11.78 & \\
14 & F10173+0828& IRAS F10173+0828& 10h20m00.21s& +08d13m33.8s& 0.0491& 224.0& 11.86& {}\\
15 & F10565+2448& IRAS F10565+2448& 10h59m18.14s& +24d32m34.3s& 0.0431& 197.0& 12.08 &{}\\
16 & F11257+5850& NGC 3690& 11h28m32.25s& +58d33m44.0s& 0.0104& 50.7& 11.93& Arp 299 \\
17 & F12112+0305& IRAS F12112+0305& 12h13m46.00s& +02d48m38.0s& 0.0733& 340.0& 12.36 & {}\\
18 & F12540+5708& UGC 8058 & 12h56m14.23s& +56d52m25.2s& 0.0422& 192.0& 12.57& Mrk 231 \\
19 & 13120--5453 & IRAS 13120--5453& 13h15m06.35s& -55d09m22.7s& 0.0308& 144.0& 12.32& {}\\
20 & F13136+6223& VV 250a& 13h15m35.06s& +62d07m28.6s& 0.0311& 142.0& 11.81& Arp 238 \\
21 & F13182+3424& UGC 8387& 13h20m35.34s& +34d08m22.2s& 0.0233& 110.0& 11.73& IC 883, Arp 193 \\
22 & F13428+5608& UGC 8696& 13h44m42.11s& +55d53m12.7s& 0.0378& 173.0& 12.21& Mrk 273\\
23 & F14348--1447& IRAS F14348--1447& 14h37m38.37s& -15d00m22.8s& 0.0827& 387.0& 12.39 & {}\\
24 & F14378--3651& IRAS F14378--3651& 14h40m59.01s& -37d04m32.0s& 0.0676& 315.0& 12.23&{} \\
25 & F14547+2449& VV 340a& 14h57m00.68s& +24d37m02.7s& 0.0337& 157.0& 11.74& Arp 302\\
26 & F15163+4255& VV 705& 15h18m06.28s& +42d44m41.2s& 0.0402& 183.0& 11.92& I Zw 107 \\
27 & F15250+3608& IRAS F15250+3608& 15h26m59.40s& +35d58m37.5s& 0.0552& 254.0& 12.00 &{}\\
28 & F15327+2340& UGC 9913& 15h34m57.12s& +23d30m11.5s& 0.0182& 87.9& 12.28& Arp 220 \\
29 & F16330--6820& ESO 69-IG6& 16h38m12.65s& -68d26m42.6s& 0.0464& 212.0& 11.98& {}\\
30 & F16504+0228& NGC 6240& 16h52m58.89s& +02d24m03.4s& 0.0245& 116.0& 11.93 & {}\\
31 & F17132+5313& IRAS F17132+5313& 17h14m20.00s& +53d10m30.0s& 0.0509& 232.0& 11.96 & {}\\
32 & F17207--0014& IRAS F17207--0014& 17h23m21.96s& -00d17m00.9s& 0.0428& 198.0& 12.46 &{}\\
33 & F18293--3413& IRAS F18293--3413& 18h32m41.13s& -34d11m27.5s& 0.0182& 86.0& 11.88&{} \\
34 & F19115--2124& ESO 593-IG8& 19h14m30.90s& -21d19m07.0s& 0.0487& 222.0& 11.93&{} \\
35 & F19297--0406& IRAS F19297--0406& 19h32m21.25s& -03d59m56.3s& 0.0857& 395.0& 12.45 &{}\\
36 & 19542+1110 & IRAS 19542+1110& 19h56m35.44s& +11d19m02.6s& 0.0650& 295.0& 12.12 &{}\\
37 & F20550+1655& CGCG 448-020& 20h57m23.90s& +17d07m39.0s& 0.0361& 161.0& 11.94& II Zw 96\\
38 & F20551--4250& ESO 286-IG19& 20h58m26.79s& -42d39m00.3s& 0.0430& 193.0& 12.06& {}\\
39 & 21101+5810 & IRAS 21101+5810& 21h11m30.40s& +58d23m03.2s& 0.0390& 174.0& 11.81& {}\\
40 & F22467--4906& ESO 239-IG2& 22h49m39.87s& -48d50m58.1s& 0.0430& 191.0& 11.84&{} \\
41 & F22491--1808& IRAS F22491--1808& 22h51m49.26s& -17d52m23.5s& 0.0778& 351.0& 12.20& {}\\
42 & F23128--5919& ESO 148-IG2& 23h15m46.78s& -59d03m15.6s& 0.0446& 199.0& 12.06&{} \\
43 & F23180--6929& ESO 77-IG14& 23h21m04.53s& -69d12m54.2s& 0.0416& 186.0& 11.76& {}\\
44 & F23365+3604& IRAS F23365+3604& 23h39m01.27s& +36d21m08.7s& 0.0645& 287.0& 12.20 &{}\\
\end{tabular}
\begin{list}{}{}
\item[$^{\mathrm{a}}$] When the IRAS Revised Bright Galaxy Sample (RBGS, Sanders et al. 2003)
was compiled, IRAS F10038--3338 was mistakenly cross-identified with the optical source
IC 2545. The proper optical counterpart is ESO 374-IG 032. (See the Essential Notes in NED.)
\item[Column (1):] Through number of the object, which are also used in other tables and Fig. 2.
\item[Column (2):] Original IRAS source, where an ``F" prefix indicates the Faint Source Catalog and no prefix indicates the Point Source Catalog.
\item[Column (3):] Optical cross-identification, where available from NED.
\item[Column (4):] The best available source right ascension (J2000) in NED as of October 2008.
\item[Column (5):] The best available source declination (J2000) in NED as of October 2008.
\item[Column (6):] The best available heliocentric redshift, in NED as of October 2008.
\item[Column (7):] The luminosity distance in megaparsecs derived by correrecting the heliocentric velocity for the 3-attractor flow model
of Mould et al. (2000) and adopting cosmological parameters
$H_0 = 70$~km~s$^{-1}$~Mpc$^{-2}$, $\Omega_{\rm V} = 0.72$, and $\Omega_{\rm M} = 0.28$ based on the
5-year WMAP results (Hinshaw et al. 2009), as provided by NED.
\item[Column (8):] The total infrared luminosity in $\rm log_{10}$ Solar units computed using the flux densities reported
in the RBGS and the lumiosity distances in column (7) using the formulae
$L_{\rm ir}/L_{\odot} = 4\pi (D_{\rm L [m]})^2~(F_{\rm ir}~[W~m^{-2}])/3.826\times 10^{26} [W m^{-2}]$,
where
$F_{\rm ir} = 1.8\times 10^{-14}\{13.48 f_{12\mu m}[Jy] + 5.16 f_{25\mu m}[Jy] + 2.58 f_{60\mu m}[Jy] + f_{100\mu m}[Jy]\} [W m^{-2}]$
(Sanders \& Mirabel 1996).
\item[Column (9):] Other conventionally used object name.
\end{list}
\end{table*}
\begin{table*}
\begin{center}
\caption{Chandra observation log.$^{\mathrm{a}}$}
\begin{tabular}{rlcccccc}
No. & Galaxy & Obs ID & Date & Mode & Exp. Time & 0.4-7 keV$^{\mathrm{b}}$ & $N_{\rm H,Gal}$$^{\mathrm{c}}$ \\
& & & & (ks) & (cts) & ($10^{20}$ cm$^{-2}$) \\[5pt]
\multicolumn{7}{c}{\bf Cycle 8 targets} \\
1 & F01364--1042 & 7801 & 2007 Sep 10 & VFAINT & 14.57 & $46.0\pm 7.3$ & 2.0 \\
2 & ESO 203-IG1 & 7802 & 2008 Jan 17 & VFAINT & 14.85 & 0 ($<3$) & 1.4 \\
3 & VII Zw 31 & 7887 & 2007 May 27 & VFAINT & 14.98 & $173.8\pm 13.3$ & 7.4 \\
5 & ESO 255-IG7 & 7803 & 2007 May 27 & VFAINT & 14.57 & $341.8\pm 18.8$ & 3.8 \\
6 & 07251--0248 & 7804 & 2006 Dec 01 & VFAINT & 15.43 & $12.7\pm 3.6$ & 14.6 \\
7 & ESO 60-IG16 & 7888 & 2007 May 31 & VFAINT & 14.68 & $122.6\pm 11.1$ & 5.2 \\
9 & 09022--3615 & 7805 & 2007 Sep 04 & VFAINT & 14.85 & $265.3\pm 16.5$ & 26.6 \\
10 & F09111--1007 & 7806 & 2007 Mar 20 & VFAINT & 14.63 & $118.7\pm 11.1$ & 4.6 \\
13 & ESO 374-IG32 & 7807 & 2007 Mar 07 & VFAINT & 14.36 & $75.0\pm 9.1$ & 8.8 \\
14 & F10173+0828 & 7808 & 2008 Jan 18 & VFAINT & 14.98 & $9.8\pm 3.2$ & 2.3 \\
19 & 13120--5453 & 7809 & 2006 Dec 01 & VFAINT & 14.67 & $300.7\pm 17.4$ & 21.3 \\
20 & VV 250 & 7810 & 2007 Aug 22 & VFAINT & 14.85 & $391.9\pm 20.1$ & 2.0 \\
21 & UGC 8387 & 7811 & 2007 Feb 19 & VFAINT & 14.07 & $251.4\pm 16.0$ & 1.0 \\
24 & F14378--3651 & 7889 & 2007 Jun 25 & VFAINT & 13.86 & $45.3\pm 6.8$ & 6.3 \\
25 & VV 340 & 7812 & 2006 Dec 17 & VFAINT & 14.86 & $331.4\pm 20.4$ & 3.3 \\
29 & ESO 69-IG6 & 7813 & 2007 Jun 21 & VFAINT & 14.54 & $330.1\pm 18.3$ & 9.1 \\
31 & F17132+5313 & 7814 & 2007 Apr 03 & VFAINT & 14.85 & $90.8\pm 10.0$ & 1.9 \\
33 & F18293-3413 & 7815 & 2007 Feb 25 & VFAINT & 14.04 & $444.6\pm 21.5$ & 9.7 \\
34 & ESO 593-IG8 & 7816 & 2007 Jun 09 & VFAINT & 14.97 & $158.1\pm 13.7$ & 8.1 \\
35 & F19297--0406 & 7980 & 2007 Jun 18 & VFAINT & 16.42 & $85.8\pm 9.7$ & 15.1 \\
36 & 19542+1110 & 7817 & 2007 Sep 10 & VFAINT & 14.98 & $324.3\pm 18.0$ & 14.0 \\
37 & CGCG 448-020 & 7818 & 2007 Sep 10 & VFAINT & 14.56 & $301.3\pm 18.2$ & 6.9 \\
39 & 21101+5810 & 7819 & 2007 Jul 01 & VFAINT & 14.85 & $21.7\pm 4.8$ & 37.2 \\
40 & ESO 239-IG2 & 7820 & 2007 Sep 10 & VFATIN & 14.57 & $151.5\pm 13.1$ & 0.9 \\
41 & F22491--1808 & 7821 & 2007 Jul 13 & VFAINT & 14.97 & $50.9\pm 7.3$ & 2.3 \\
43 & ESO 77-IG14 & 7822 & 2008 Jan 26 & VFAINT & 14.98 & $84.1\pm 9.3$ & 3.3 \\[5pt]
\multicolumn{7}{c}{\bf Archival data} \\
4 & F05189-2524 & 3432 & 2002 Jan 03 & FAINT & 14.86 & $2016.9\pm 45.0$ & 1.7 \\
8 & F08572+3915 & 6862 & 2006 Jan 26 & FAINT & 14.94 & $9.7\pm 3.2$ & 2.0 \\
11 & UGC 4881 & 6857 & 2006 Jan 12 & FAINT & 14.77 & $69.4\pm 8.4$ & 1.4 \\
12 & UGC 5101 & 2033 & 2001 May 28 & FAINT & 49.32 & $482.9\pm 22.3$ & 3.0 \\
15 & F10565+2448 & 4552 & 2003 Oct 23 & FAINT & 28.87 & $335.2\pm 18.8$ & 1.1 \\
16 & NGC 3690 & 6227 & 2005 Feb 14 & FAINT & 10.19 & $2526.0\pm 52.1$ & 0.9 \\
17 & F12112+0305 & 4110 & 2003 Apr 15 & VFAINT & 9.87 & $42.6\pm 6.6$ & 1.8 \\
18 & UGC 8058 & 1031$^{\mathrm{d}}$ & 2000 Oct 19 & FAINT & 39.25 & $2312.5\pm 66.2$ & 1.0 \\
18 & UGC 8058 & 4028 & 2003 Feb 03 & VFAINT & 39.68 & $2205.6\pm 59.7$ & 1.0 \\
18 & UGC 8058 & 4029 & 2003 Feb 11 & VFAINT & 38.63 & $2070.5\pm 57.3$ & 1.0 \\
18 & UGC 8058 & 4030 & 2003 Feb 20 & VFAINT & 36.01 & $1876.6\pm 52.0$ & 1.0 \\
22 & UGC 8696 & 809 & 2000 Apr 19 & VFAINT & 44.19 & $2054.0\pm 45.8$ & 0.9 \\
23 & F14348-1447 & 6861 & 2006 Mar 12 & FAINT & 14.72 & $75.9\pm 8.8$ & 7.5 \\
26 & VV 705 & 6858 & 2006 Sep 11 & FAINT & 14.47 & $157.8\pm 12.6$ & 1.8 \\
27 & F15250+3608 & 4112 & 2003 Aug 27 & VFAINT & 9.84 & $26.6\pm 5.3$ & 1.5 \\
28 & UGC 9913 & 869 & 2000 Jun 24 & FAINT & 56.49 & $1555.1\pm 47.1$ & 3.9 \\
30 & NGC 6240 & 1590 & 2001 Jul 29 & FAINT & 36.69 & $10010.7\pm 103.5$ & 4.9 \\
32 & F17207-0014 & 2035 & 2001 Oct 24 & FAINT & 48.53 & $476.6\pm 23.1$ & 9.7 \\
38 & ESO 286-IG19 & 2036 & 2001 Oct 31 & FAINT & 44.87 & $767.8\pm 28.3$ & 3.3 \\
42 & ESO 148-IG2 & 2037 & 2001 Sep 30 & FAINT & 49.31 & $1052.2\pm 34.9$ & 1.6 \\
44 & F23365+3604 & 4115 & 2003 Feb 03 & VFAINT & 10.10 & $28.8\pm 5.4$ & 9.6 \\
\end{tabular}
\begin{list}{}{}
\item[$^{\mathrm{a}}$] The 26 Cycle 8 Chandra observations (PI: D. Sanders) are listed first, followed
by archival data for the additional 18 objects with log~($L_{\rm ir}/L_\odot) > 11.73$. All observations
were obtained in imaging mode with the ACIS-S, and the targets were placed at the
nominal pointing position on the detector.
\item [$^{\mathrm{b}}$] The source counts are corrected for background and
measured in the 0.4-7 keV band. The counts from separate components
in a single objects are summed together.
\item [$^{\mathrm{c}}$] The Galactic absorption
column density is taken from the LAB HI map
by Kalberla et al. (2005).
\item [$^{\mathrm{d}}$] This observation was used to make the radial surface-brightness profile.
\end{list}
\end{center}
\end{table*}
\section{Results}
\begin{table*}
\setlength{\tabcolsep}{0.03in}
\begin{center}
\caption{X-ray spectral properties for the C-GOALS sample}
\begin{tabular}{rlccccccccc}
No. & Galaxy & $SX$ & $HX$ & {\sl HR} & $F_{\rm SX}$ & $F_{\rm HX}$ & $L_{\rm SX}$ & $L_{\rm HX}$ & {\sl SX/IR} & {\sl HX/IR} \\
& & (1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) \\
\multicolumn{11}{c}{\bf Cycle 8 Data} \\
1 & F01364--1042 & $2.14\pm 0.41$ & $0.86\pm 0.30$ & $-0.43\pm 0.18$ & 4.5e-15 & 2.0e-14 & 2.4e40 & 1.5e41 & -5.05 & -4.26 \\
2 & ESO 203-IG1 & $0.00\pm 0.20$ & $0.00\pm 0.20$ & --- & $<$1.5e-15 & $<$3.0e-15 & $<$5.6e39 & $<$6.3e40 & $<$-5.69 &$<$-5.14 \\
3 & VII Zw 31 & $9.49 \pm 0.80$ & $2.01 \pm 0.38$ & $-0.65\pm 0.09$ & 3.0e-14 & 3.5e-14 & 2.0e41 & 2.7e41 & -4.27 & -4.14 \\
5 & ESO 255-IG7 N & $12.40 \pm 0.93$ & $2.77 \pm 0.46$ & $-0.64\pm 0.08$ & 3.5e-14 & 2.9e-14 & 1.2e41 & 1.6e41 & -4.41 & -4.28 \\
5 & ESO 255-IG7 C & $4.77 \pm 0.57$ & $1.35 \pm 0.32$ & $-0.56\pm 0.12$ & 1.5e-14 & 2.1e-14 & 5.3e40 & 1.2e41 & -4.76 & -4.41 \\
5 & ESO 255-IG7 S & $1.53 \pm 0.33$ & $0.63 \pm 0.23$ & $-0.42\pm 0.20$ & 4.5e-15 & 6.8e-15 & 1.5e40 & 3.8e40 & -5.31 & -4.90 \\
6 & 07251--0248 & $0.77 \pm 0.22$ & $0.05 \pm 0.06$ & $-0.87\pm 0.38$ & 3.3e-15 & $<$3e-15 & 6.2e40 & $< $5.8e40 & -5.18 & $<-5.21$ \\
7 & ESO 60-IG16 & $3.24 \pm 0.47$ & $5.09 \pm 0.59$ & $+0.22\pm 0.09$ & 8.1e-15 & 1.0e-13 & 5.5e40 & 7.3e41 & -4.66 & -3.54 \\
9 & 09022--3615 & $10.45 \pm 0.84$ & $7.38 \pm 0.72$ & $-0.17\pm 0.06$ & 2.9e-14 & 1.4e-13 & 4.4e41 & 2.0e42 & -4.25 & -3.59 \\
10 & F09111--1007E & $5.27\pm 0.60$ & $0.97 \pm 0.27$ & $-0.69\pm 0.13$ & 1.6e-14 & 1.2e-14 & 1.2e41 & 9.8e40 & -4.57 & -4.65 \\
10 & F09111--1007W & $1.20 \pm 0.29$ & $0.55 \pm 0.21$ & $-0.37\pm 0.22$ & 3.6e-15 & 4.5e-15 & 2.1e40 & 3.8e40 & -5.32 & -5.06 \\
13 & ESO 374-IG32 & $4.02 \pm 0.55$ & $1.22 \pm 0.32$ & $-0.62\pm 0.13$ & 1.3e-14 & 1.7e-14 & 3.8e40 & 6.3e40 & -4.78 & -4.56 \\
14 & F10173+0828 & $0.59 \pm 0.21$ & $0.06 \pm 0.06$ & $-0.82\pm 0.42$ & 2.1e-15 & 1.0e-15 & 1.1e40 & 6.6e39 & -5.40 & -5.62 \\
19 & 13120--5453 & $11.63 \pm 0.89$ & $9.00 \pm 0.79$ & $-0.13\pm 0.06$ & 3.7e-14 & 1.4e-13 & 1.1e41 & 4.5e41 & -4.86 & -4.25 \\
20 & VV 250 E & $16.90 \pm 1.08$ & $6.94 \pm 1.75$ & $-0.42\pm 0.06$ & 5.3e-14 & 9.5e-14 & 1.2e41 & 3.1e41 & -4.32 & -3.90 \\
20 & VV 250 W & $1.78 \pm 0.36$ & $0.54 \pm 0.22$ & $-0.53\pm 0.20 $ & 6.0e-15 & 1.0e-15 & 1.3e40 & 2.5e39 & -5.28 & -6.00 \\
21 & UGC 8387 & $13.77 \pm 0.99$ & $3.98 \pm 0.54$ & $-0.55\pm 0.07$ & 4.4e-14 & 4.3e-14 & 5.4e40 & 6.4e40 & -4.58 & -4.51 \\
24 & F14378--3651 & $1.92 \pm 0.38$ & $1.34 \pm 0.31$ & $-0.18\pm 0.15$ & 6.7e-15 & 2.1e-14 & 8.0e40 & 3.4e41 & -4.91 & -4.28 \\
25 & VV 340 N & $16.58 \pm 1.08$ & $2.55 \pm 0.55$ & $-0.74\pm 0.08 $ & 4.8e-14 & 7.3e-14 & 1.3e41 & 2.9e41 & -4.21 & -3.86 \\
25 & VV 340 S & $2.35 \pm 0.41$ & $0.92 \pm 0.33$ & $-0.44\pm 0.18 $ & 6.6e-15 & 3.6e-15 & 1.7e40 & 1.2e40 & -5.09 & -5.25 \\
29 & ESO 69-IG6 (N) & $8.88 \pm 0.79$ & $1.78 \pm 0.37$ & $-0.67\pm 0.10$ & 2.8e-14 & 1.8e-14 & 1.4e41 & 1.0e41 & -4.42 & -4.56 \\
29 & ESO 69-IG6 (S) & $2.19 \pm 0.39$ & $9.97 \pm 0.83$ & $+0.64\pm 0.09$ & 8.1e-15 & 1.7e-13 & 3.5e40 & 1.2e42 & -5.02 & -3.49 \\
31 & F17132+5313 & $5.13 \pm 0.60$ & $0.76 \pm 0.33$ & $-0.74\pm 0.14$ & 1.3e-14 & 1.1e-14 & 8.3e40 & 8.6e40 & -4.63 & -4.61 \\
33 & F18293--3413 & $22.92 \pm 1.29$ &$8.66 \pm 0.82$ & $-0.45\pm 0.05$ & 6.9e-14 & 1.3e-13 & 5.7e40 & 9.4e40 & -4.71 & -4.49 \\
34 & ESO 593-IG8 & $9.01 \pm 0.81$ & $1.65 \pm 0.44$ & $-0.69\pm 0.11$ & 2.1e-14 & 2.5e-14 & 1.2e41 & 1.9e41 & -4.44 & -4.24 \\
35 & F19297--0406 & $3.85 \pm 0.49$ &$1.21 \pm 0.30$ & $-0.52\pm 0.13$ & 1.2e-14 & 8.2e-15 & 3.0e41 & 1.8e41 & -4.56 & -4.78 \\
36 & 19542+1110 & $4.80 \pm 0.57$ & $17.12 \pm 1.07$ & $+0.56\pm 0.06$ & 1.8e-14 & 2.9e-13 & 2.2e41 & 4.1e42 & -4.36 & -3.09 \\
37 & CGCG 448-020 & $15.91 \pm 1.07$ & $4.41 \pm 0.64$ & $-0.57\pm 0.07 $ & 4.9e-14 & 6.7e-14 & 1.4e41 & 2.8e41 & -4.38 & -4.08 \\
39 & 21101+5810 & $1.17 \pm 0.29$ & $0.33 \pm 0.17$ & $-0.56\pm 0.25$ & 3.5e-15 & 4.0e-15 & 1.9e40 & 2.0e40 & -5.12 & -5.09 \\
40 & ESO 239-IG2 & $9.14 \pm 0.83$ & $1.31 \pm 0.36$ & $-0.75\pm 0.11$ & 3.0e-14 & 1.9e-14 & 1.4e41 & 9.7e40 & -4.28 & -4.44 \\
41 & F22491--1808 & $3.19 \pm 0.47$ & $0.14 \pm 0.14$ & $-0.91\pm 0.20$ & 8.3e-15 & 3.0e-15 & 1.3e41 & 0.6e41 & -4.67 & -5.01 \\
43 & ESO 77-IG14 N & $2.65 \pm 0.42$ & $0.94 \pm 0.26$ & $-0.47\pm 0.15$ & 7.8e-15 & 1.7e-14 & 9.6e40 & 6.8e40 & -4.36 & -4.51 \\
43 & ESO 77-IG14 S & $1.85 \pm 0.35$ & $0.22 \pm 0.13$ & $-0.78\pm 0.23$ & 6.0e-15 & 6.1e-15 & 2.4e40 & 2.5e40 & -4.96 & -4.95 \\[5pt]
\multicolumn{11}{c}{\bf Archival Data} \\
4 & F05189-2524 & $27.15 \pm 1.35$ & $107.00 \pm 2.69$ & $+0.60\pm 0.03$ & 8.2e-14 & 2.0e-12 & 3.4e41 & 1.3e43 & -4.21 & -2.63 \\
8 & F08572+3915 & $0.12\pm 0.09$ & $0.45\pm 0.18$ & $+0.57\pm 0.41$ & 9.0e-15 & 2.5e-14 & 8.0e40 & 2.0e41 & -4.83 & -4.44 \\
11 & UGC 4881 E & $2.09 \pm 0.38$ & $0.32 \pm 0.15$ & $-0.73\pm 0.21$ & 6.7e-15 & 7.8e-15 & 2.5e40 & 4.1e40 & -4.82 & -4.61 \\
11 & UGC 4881 W & $2.13 \pm 0.38$ & $0.13 \pm 0.12$ & $-0.88\pm 0.24$ & 7.1e-15 & 3.0e-15 & 2.7e40 & 1.3e40 & -4.19 & -4.51 \\
12 & UGC 5101 & $6.34 \pm 0.36$ & $3.49 \pm 0.27$ & $-0.29\pm 0.05$ & 1.9e-14 & 9.1e-14 & 7.0e40 & 4.7e41 & -4.74 & -3.92 \\
15 & F10565+2448 & $9.37 \pm 0.58$ & $2.21 \pm 0.30$ & $-0.62\pm 0.07$ & 2.8e-14 & 2.8e-14 & 1.2e41 & 1.6e41 & -4.58 & -4.46 \\
16 & NGC 3690 T & $204.84 \pm 4.64$ & $41.40 \pm 2.17$ & $-0.66\pm 0.03$ & 6.2e-13 & 6.8e-13 & 1.9e41 & 2.6e41 & -4.23 & -4.10 \\
16 & NGC 3690 E & $54.46 \pm 2.33$ & $15.43 \pm 1.24$ & $-0.56\pm 0.04$ & 1.5e-13 & 2.6e-13 & 7.7e40 & 9.6e40 & -4.49 & -4.40 \\
16 & NGC 3690 W & $58.71 \pm 2.42$ & $15.04 \pm 1.23$ & $-0.59\pm 0.04$ & 1.6e-13 & 2.6e-13 & 8.0e40 & 9.6e40 & -4.02 & -3.94 \\
17 & F12112+0305 & $3.22 \pm 0.57$ & $1.09 \pm 0.34$ & $-0.49\pm 0.17$ & 9.8e-15 & 1.9e-14 & 1.3e41 & 4.0e41 & -4.83 & -4.34 \\
18 & UGC 8058 & $34.23 \pm 1.04$ & $19.65 \pm 0.74$ & $-0.27\pm 0.03$ & 9.5e-14 & 4.0e-13 & 3.9e41 & 3.0e42 & -4.56 & -3.67 \\
22 & UGC 8696 core & $29.02 \pm 0.82$ & $16.20 \pm 0.62$ & $-0.28\pm 0.02$ & 7.4e-14 & 4.2e-13 & 2.5e41 & 2.5e42 & -4.39 & -3.39 \\
22 & UGC 8696 tail & $7.09 \pm 0.50$ & $0.00 \pm 0.31$ & $-1.00\pm 0.12$ & 1.7e-14 & 2.5e-15 & 5.6e40 & 1.1e40 & --- & --- \\
23 & F14348-1447 T & $3.80 \pm 0.51$ & $1.32 \pm 0.31$ & $-0.48\pm 0.13$ & 1.3e-14 & 2.4e-14 & 2.9e41 & 7.5e41 & -4.51 & -4.09 \\
23 & F14348-1447 S & $1.15 \pm 0.28$ & $0.80 \pm 0.24$ & $-0.18\pm 0.19$ & 4.1e-15 & 1.4e-14 & 9.0e40 & 4.5e41 & -4.90 & -4.20 \\
26 & VV 705 N & $8.34 \pm 0.76$ & $1.39 \pm 0.32$ & $-0.71\pm 0.10$ & 2.6e-14 & 1.6e-14 & 9.7e40 & 6.8e40 & -4.46 & -4.62 \\
26 & VV 705 S & $1.03 \pm 0.27$ & $0.06 \pm 0.07$ & $-0.90\pm 0.34$ & 3.5e-15 & 2.1e-15 & 1.3e40 & 8.6e39 & --- & --- \\
27 & F15250+3608 & $2.78 \pm 0.54$ & $0.00 \pm 0.18$ & $-1.0\pm 0.30$ & 8.7e-15 & $<$4.2e-15 & 6.2e40 & $<$4.5e40 & -4.87 & $<-$5.11 \\
28 & UGC 9913 & $21.90 \pm 0.70$ & $4.68 \pm 0.46$ & $-0.65\pm 0.04$ & 2.7e-14 & 6.6e-14 & 2.0e40 & 6.8e40 & -5.56 & -5.03 \\
30 & NGC 6240 & $209.55 \pm 2.45$ & $59.57 \pm 1.34$ & $-0.56\pm 0.01$ & 6.1e-13 & 1.2e-12 & 8.1e41 & 2.1e42 & -3.19 & -2.97 \\
32 & F17207-0014 & $7.13 \pm 0.40$ & $2.70 \pm 0.26$ & $-0.45\pm 0.05$ & 1.9e014 & 3.9e-14 & 9.9e40 & 2.2e41 & -5.04 & -4.70 \\
38 & ESO 286-IG19 & $14.83 \pm 0.58$ & $1.98 \pm 0.24$ & $-0.77\pm 0.05$ & 4.3e-14 & 3.3e-14 & 2.1e41 & 2.1e41 & -4.32 & -4.32 \\
42 & ESO 148-IG2 Tot & $14.46 \pm 0.56$ & $6.56 \pm 0.43$ & $-0.38\pm 0.04$ & 4.0e-14 & 1.1e-14 & 1.9e41 & 8.3e41 & -4.36 & -3.72 \\
42 & ESO 148-IG2 N &$2.78 \pm 0.24$ & $0.41 \pm 0.11$ & $-0.74\pm 0.10$ & 7.6e-15 & 4.2e-15 &3.7e40 & 2.4e40 & -4.47 & -4.66 \\
42 & ESO 148-IG2 S & $4.44 \pm 0.30$ & $5.48 \pm 0.34$ & $+0.11\pm 0.05$ & 1.3e-14 & 9.0e-14 & 5.7e40 & 6.6e41 & -4.75 & -3.69\\
44 & F23365+3604 & $1.67 \pm 0.41$ & $1.07 \pm 0.33$ & $-0.22\pm 0.20$ & 5.7e-15 & 1.1e-14 & 5.3e40 & 1.6e41 & -5.06 & -4.58 \\
\end{tabular}
\begin{list}{}{}
\item[Column (1):] Background corrected count rate in the
0.5-2 keV band in unit of $10^{-3}$ ct s$^{-1}$
\item[Column (2):] Background
corrected count rate in the 2-8 keV band in unit of $10^{-3}$ ct
s$^{-1}$.
\item[Column (3):] X-ray colour, as defined by $HR = (H-S)/(H+S)$.
\item[Column (4):] Observed 0.5-2 keV band flux (erg\thinspace s$^{-1}$\thinspace cm$^{-2}$).
\item[Column (5):] Observed 2-7 keV band flux (erg\thinspace s$^{-1}$\thinspace cm$^{-2}$).
\item[Column (6):] 0.5-2 keV luminosity corrected for Galactic absorption (erg\thinspace s$^{-1}$).
\item[Column (7):] 2-10 keV luminosity corrected for Galactic absorption (erg\thinspace s$^{-1}$).
\item[Column (8):] Logarithmic luminosity ratio of the 0.5-2 keV and 8-1000 $\mu$m bands;
\item[Column (9):] Logarithmic luminosity ratio of the 2-10 keV and 8-1000 $\mu$m bands.
\end{list}
\end{center}
\end{table*}
Basic X-ray properties obtained from the Chandra data are presented in
Table 3. For each object, the ACIS-S count rates in the soft (0.5-2
keV) and hard (2-7 keV) bands, the X-ray colour, estimated X-ray
fluxes, X-ray luminosities in the two bands, and their logarithmic
ratio to the infrared luminosity, $L_{\rm ir} (8-1000\mu$m), are
listed. The count rates are computed for spatially distinctive
components within a single object when they are clearly resolved, as
opposed to the source counts in Table 2, which were collected from the
whole X-ray emitting region per object. Similarly, IR luminosity is
divided into respective components, estimated from IRAS HIRES processing
(Surace et al. 2004) and new estimates based on Spitzer MIPS
photometry (see Howell et al. 2010). The X-ray colour is defined as ${\sl HR} =
(H-S)/(H+S)$, often referred as the ``Hardness Ratio'', where $H$ and
$S$ are background-corrected counts in the 2-8 keV and 0.5-2 keV
bands, respectively.
We use the energy ranges, particularly for the hard X-ray band, that
are slightly different from one another, but optimized for different
purposes, which can sometimes be confusing. Here, we summarize the
choice of the energy ranges used in this paper. 1) 0.4-7 keV: This is
effectively the full ACIS-S bandpass, and used only for observed
counts; 2) 0.5-2 keV: The soft X-ray band is always this range; 3) 2-8
keV: The choice of this band is to match the conventional calculation
of hardness ratio, {\sl HR}, and used only for the hard band counts in
Table 3. The quoted counts are as observed, although the majority of
our sources have negligible counts above 7 keV where the background
would simply increases; 4) 2-7 keV: This is the nominal range for our
observed hard band flux, estimated from spectral fitting, and
optimized for the signal to noise ratio of the data for the majority
of our sources for the reason mentioned above; and 5) 2-10 keV: For
the hard band luminosity, we opted to use this extended range for a
purpose of comparison with results from other X-ray observatories,
since the 2-10 keV is the standard band for missions like XMM-Newton.
The images, spectra, colour and the selection of AGN based
on the colour, the correlation of X-ray and infrared luminosities and the
surface brightness distribution of individual objects are presented in
the subsections below.
\subsection{Images}
\begin{figure*}
\centerline{\includegraphics[width=0.95\textwidth,angle=0]{f02a.pdf}}
\end{figure*}
\begin{figure*}
\centerline{\includegraphics[width=0.95\textwidth,angle=0]{f02b.pdf}}
\end{figure*}
\begin{figure*}
\centerline{\includegraphics[width=0.95\textwidth,angle=0]{f02c.pdf}}
\caption{ X-ray and optical images of the 44 C-GOALS sample
galaxies. X-ray (0.4-7 keV) brightness contours (magenta) are overlaid on the
HST-ACS F814W image (grey scale) for each object. The orientation of
all images is north up and east to the left. Eleven contour levels
are defined in the fixed surface brightness range $4\times
10^{-5} - 7\times 10^{-3}$ [counts s$^{-1}$
arcsec$^{-2}$], which is divided into ten equal logarithmic
intervals. For majority of objects, these contour levels were
applied. For nine objects (F10173+0828, F17132+5313, F19297--0406,
CGCG 448-020, F01364--1042, IRAS 21101+5810, F10565+2448, UGC 9913,
and UGC 5101), a further lower interval was added to outline lower
surface brightness features. For four bright objects, (UGC 8696, NGC
3690, F05189--2524 and UGC 8058), custom contour levels were made for
describing their X-ray morphology -- these are listed in Table 4.}
\end{figure*}
Figure 2 shows the X-ray brightness contours overlaid on the HST-ACS
F814W (I-band) image (Evans et al., in prep) for all 44 targets. The contours are made
from a 0.4-7 keV image, smoothed using a circular Gaussian filter
with typical dispersion of 1 arcsec.
Note that the deep image of Mrk 231 reaches a significantly lower
brightness level, and thus the contour levels are drawn down to a
level which is a factor of 10 lower than for the other objects.
\begin{table}
\begin{center}
\caption{X-ray contours for bright objects}
\begin{tabular}{rlccc}
No. & Galaxy & Low & High & n \\
& & (1) & (2) & (3) \\[5pt]
4 & F05189--2524 & $3.0\times 10^{-5}$ & $4.2\times 10^{-3}$ & 10 \\
16 & NGC 3690 & $2.7\times 10^{-5}$ & $1.8\times 10^{-3}$ & 10 \\
18 & UGC 8058 & $8.5\times 10^{-7}$ & $3.7\times 10^{-4}$ & 11 \\
22 & UGC 8696 & $2.3\times 10^{-6}$ & $1.8\times 10^{-3}$ & 9 \\
\end{tabular}
\begin{list}{}{}
\item[Column (1)]The lowest contour (cts s$^{-1}$ arcsec$^{-2}$) in Fig. 2.
\item[Column (2)]The highest contour (cts s$^{-1}$ arcsec$^{-2}$) in Fig. 2.
\item[Coliumn (3)]The number of contour levels in Fig. 2.
\end{list}
\end{center}
\end{table}
The appearance of the X-ray images often differs dramatically between the
soft and hard bands. In addition to the full band (0.4-7 keV) image in
unsmoothed ($0.5^{\prime\prime}\times 0.5^{\prime\prime}$ pixel size)
and smoothed forms, images of the same region of the sky in the soft
(0.5-2 keV) and hard (2-7 keV) bands are shown. The same smoothing has
been applied to the 0.5-2 keV and 2-7 keV images. Linear scaling
is normally used for presentation, but for a few sources with
strongly peaked emission, a logarithmic scale was used to show faint
extended features. The scale bar represents an angular scale of 5
arcsec. The soft (0.5-2 keV) and hard
(2-7 keV) band radial profiles, described in the following section,
are also shown. An example for these figures is shown in Fig. 3 for UGC 8387.
Similar style multi-panel figures were made for all 44 C-GOALS targets and
these can be found in the online material.
\begin{figure*}
\centerline{\includegraphics[width=0.8\textwidth,angle=0]{f03.pdf}}
\caption{An example of X-ray images and the surface brightness
distributions for individual objects is presented. The object shown
here is UGC 8387 (= IC 883). Similar multi-panel figures for the
other 43 objects in our C-GOALS sample can be found in Appendix.
{\it (Upper left)}: The X-ray (0.4-7 keV) brightness contours
overlaid on the HST ACS I-band image. {\it (Upper right)}: The
radial surface brightness profiles in the 0.5-2 keV (open circles)
and 2-7 keV (filled squares) bands are shown. These profiles are the
azimuthal average and are measured from the peak of 2-7 keV emission
at 1 arcsec resolution. The Point Spread Function (PSF) computed at
3 keV is shown as a dotted histogram where the peak has been matched
to the observed 2-7 keV peak for comparison. The bottom four-panel
figure shows (left to right) the raw 0.4-7 keV image, the gaussian
smoothed version of the same image, the 0.5-2 keV smoothed image,
and the 2-7 keV smoothed image. The pixel size is $\approx
0.5^{\prime\prime} \times 0.5^{\prime\prime}$. The scale bar
indicates 5 arcsec.}
\end{figure*}
As seen clearly in UGC 8387, soft X-ray nebulae extending along the
minor axis of the galaxy, or in the direction displaced from the
optical light distribution, are often observed. These extended SX
nebulae are likely associated with a galactic-scale outflow from the
nuclear region, driven by either a starburst or AGN. Other examples
are F18293--3413, VV 340 N, UGC 9913 (hereafter Arp 220), NGC 6240 and
ESO 148-IG2.
\subsection{Flux density spectra}
X-ray spectra are traditionally shown as count rate spectra, i.e.,
data folded through the detector response. However, in this paper, in
the interest of comparing the X-ray spectra with other
multi-wavelength datasets in GOALS, the ACIS spectra have been
corrected for the detector response curve and further converted into
flux density units (Fig. 4). This correction may introduce some
uncertainty, particularly when a spectral bin is wide within which the
response varies rapidly, e.g., at the high energy end of the
bandpass. In spite of this caveat, we think that this presentation has
merit given that the spectral properties can be directly observed
without resorting to spectral fitting. For this purpose, the flux
density range for all spectra was also kept to be 2 orders of
magnitude except where the dynamic range of the spectra exceeds 2
decades. The flux density is in units of [$10^{-14}$ erg cm$^{-2}$
s$^{-1}$ keV$^{-1}$], and can readily be converted into units of [W
m$^{-2}$ Hz$^{-1}$] by multiplying by $4.17\times 10^{-35}$.
For some objects, data of spatially separate components are shown in Fig. 4, and when we
consider it appropriate, data for the total emission are shown to help
facilitate comparison with spectra taken from other X-ray
observatories at lower spatial resolution. Finally, we note that
these are not ``unfolded spectra'', which are sometimes presented in
the literature. Instead, they have been corrected solely for the
detector effective area while preserving the energy resolution of the
detector, and thus are independent of any spectral model which might
be fitted. [Note: The spectra shown in Fig. 4 are for display purposes only; all physical
quantities reported in this paper were obtained through conventional
spectral fitting to the count rate spectra with appropriate detector
responses.]
\begin{figure*}
\centerline{\includegraphics[width=\textwidth,angle=0]{f04a.pdf}}
\end{figure*}
\begin{figure*}
\centerline{\includegraphics[width=\textwidth,angle=0]{f04b.pdf}}
\end{figure*}
\begin{figure*}
\centerline{\includegraphics[width=\textwidth,angle=0]{f04c.pdf}}
\caption{The X-ray flux density spectra of the C-GOALS sources, obtained from the Chandra ACIS.
The energy scale is as observed. The flux density is in units of [$10^{-14}$ erg s$^{-1}$ cm$^{-2}$ keV$^{-1}$].
}
\end{figure*}
\subsection{X-ray colour and AGN selection}
\begin{figure}
\begin{center}
\hbox{
\includegraphics[width=0.48\textwidth,angle=0]{f05.pdf}
}
\caption{The X-ray colour ({\sl HR}) as a function of infrared
luminosity and its distribution. See the text for the definition of
{\sl HR}. While the {\sl HR} is derived for individual components in
a single object when the X-ray sources are well resolved into
multiple components (i.e., the nuclear separation is $\sim
4^{\prime\prime}$ or larger; see Table 3). The threshold for
selecting AGN, based on the {\sl HR} $> -0.3$, is indicated by the
dashed line. Objects in which a 6.4 keV Fe K$\alpha $ line detetion
has been reported are marked by open circles. Three known Compton
thick AGN with strong (6.4 keV) Fe K (NGC 3690 W, NGC 6240, VV340 N)
lie below the {\sl HR} threshold. The {\sl HR} distribution
histogram of {\sl HR} is attached on the right hand side. The median
is {\sl HR} $= -0.56$. }
\end{center}
\end{figure}
\begin{table}
\setlength{\tabcolsep}{0.05in}
\begin{center}
\caption{Classification of C-GOALS sources}
\begin{tabular}{rlccccl}
No. & Object & VO & YKS & $D_{\rm agn}$ & [Ne {\sc v}] & $X_{\rm agn}$\\%& MS \\
& & (1) & (2) & (3) & (4) & (5)\\[5pt]
\multicolumn{7}{c}{\bf X-ray selected AGN} \\
4 & F05189--2524 & S2 & S2 & 1 & Y & CL\\
7 & ESO 60-IG16 & -- & -- & -- & -- & C\\
8 & F08572+3915 NW & L & S2 & 0.6 & N & C\\
9 & 09022--3615 & -- & -- & -- & N & C\\
12 & UGC 5101 & L & S2 & 0.6 & Y & CL\\
16 & NGC 3690 W & H & -- & -- & -- & L\\
18 & UGC 8058 & S1 & S1 & 1 & N & CL\\
19 & 13120--5453 & -- & -- & -- & Y & C\\
22 & UGC 8696 & S2 & S2 & 0.8 & Y & CL\\
23 & F14348--1447 SW & L & cp & 0.7 & N & C \\
24 & F14378--3651 & S2 & -- & -- & N & C\\
25 & VV 340 N & L & cp & 0.5 & Y & L\\
30 & NGC 6240 & L & L & 0.8 & Y & L\\
36 & 19542+1110 & -- & -- & -- & -- & C\\
42 & ESO 148-IG2 & -- & -- & -- & -- & C\\
44 & F23365+3604 & L & cp & 0.3 & N & C\\[5pt
\multicolumn{7}{c}{\bf Others (HXQ sample)} \\
1 & F01364--1042 & L & L & 0.8 & -- & \\%& c\\
2 & ESO 203-IG1 & -- & -- & -- & -- & \\% & a\\
3 & VII Zw 31 & H & -- & -- & Y & \\%& e\\
5 & ESO 255-IG7 & -- & H & 0 & -- & \\%& a\\
6 & 07251--0248 & -- & -- & -- & -- & \\%& c\\
10 & F09111--1007 & -- & -- & -- & Y & \\%& a\\
11 & UGC 4881 SW/NE & H/H & cp/cp & 0.3/0.2 & -- &\\%& b\\
13 & ESO 374-IG32 & -- & cp & 0.3 & -- & \\%& b\\
14 & F10173+0828 & -- & -- & -- & -- & \\%& c\\
15 & F10565+2448 & H & cp & 0.2 & N & \\%& a\\
16 & NGC 3690 E & H & -- & -- & -- & \\%& b\\
17 & F12112+0305 & L & S2 & 0.7 & N & \\%& b\\
20 & VV 250 NW/SE & H/H & cp/cp & 0.4/0.3 & -- & \\%& a\\
21 & UGC 8387 & L & cp & 0.5 & Y & \\%& c\\
26 & VV 705 S/N & L/H & cp/cp & 0.5/0.3 & -- &\\%& b\\
27 & F15250+3608 & L & cp & 0.4 & N & \\%& d\\
28 & UGC 9913 & L & L & 0.7 & N & \\%& b\\
29 & ESO 69-IG6 & -- & -- & -- & -- & \\%& a\\
31 & F17132+5313 & H & H & 0 & Y & \\%& b\\
32 & F17207--0014 & H & H & 0 & N & \\%& d\\
33 & F18293--3413 & -- & -- & -- & -- & \\%& b\\
34 & ESO 593-IG008 S/N & L/H & S2/cp & 0.4/0.2 & Y & \\%& b\\
35 & F19297--0406 & -- & -- & -- & N & \\%& b\\
37 & CGCG 448-020 & H & H & 0 & N & \\%& b\\
38 & ESO 286-IG19 & H & H & 0 & -- & \\%& c\\
39 & 21101+5810 & -- & -- & -- & -- & \\%& a\\
40 & ESO 239-IG2 & -- & -- & -- & -- & \\%& c\\
41 & F22491--1808 & H & H & 0 & N & \\%& b\\
43 & ESO 77-IG14 & H & -- & -- & -- & \\%& a\\
\end{tabular}
\begin{list}{}{}
\item[Column (1):] Optical class based on Veilleux \& Osterbrock
(1987) diagram from Veilleux et al. (1995, 1999). H: star-forming;
S1: Seyfert 1; S2: Seyfert 2; L: LINER.
\item[Column (2):] The SDSS class from Yuan, Kewley \& Sanders
(2010). The symbols are the same as in Column (1), plus an additional
class of ``composite" objects; cp: composite of starburst and AGN.
\item[Column (3):] The AGN fraction $D_{\rm agn}$ derived from
[OI]/H$\alpha $ diagram (see Yuan et al. 2010 for details).
\item[Column (4):] Detection of [Ne V] 14.32 $\mu $m from Farrah et al.
2007. New results on CGCG 448-020 (Inami et al. 2010), VV340 N (Armus
et al. 2009), UGC 8387 (Modica et al. 2010), and other GOALS
objects (Petric et al. 2010) are also included.
\item[Column (5):] AGN selection criteria: C: X-ray colour ($HR \geq 0.3$); L: 6.4 keV Fe K line.
\end{list}
\end{center}
\end{table}
The X-ray colour, or hardness ratio, {\sl HR}, gives the relative strength of the X-ray emission
above and below 2 keV (in counts). Since strong emission above 2 keV
is often associated with an absorbed X-ray source with a column density,
$N_{\rm H}$, in the range of $10^{22}$-$10^{24}$cm$^{-2}$, which, in turn
indicates the presence of an obscured AGN, it often
serves as a crude probe of AGN.
As described in Iwasawa et al. (2009), AGN are selected as follows: The
primary criterion is an hard X-ray spectrum as assessed by the X-ray
colour, {\sl HR}. The values of {\sl HR} as a function of $L_{\rm ir}$
are plotted in Fig 5. The median value of {\sl HR} is
$-0.56$.
Objects with {\sl HR} $> -0.3$ are classified as an AGN. This threshold is
chosen because ULIRGs known to host AGN (e.g. Mrk 231, UGC 8696 (= Mrk
273), UGC 5101) cluster just above this value. All of the optically
identified AGN, i.e., Seyfert 1 and Seyfert 2 galaxies, in our sample
are selected by this criterion.
Some Compton-thick AGN are missed by a {\sl HR} $> -0.3$ selection
because of their weakness in the hard band given that only reflected radiation
is observed. The relative strength of the hard X-ray emission is
largely suppressed, giving a small value for {\sl HR}. Objects that show a
strong Fe~K line at 6.4~keV, a characteristic signature of a
Compton-thick AGN, are also classified as AGN (e.g. NGC 6240 --
Iwasawa \& Comastri 1998; Vignati et al. 1999; Ikebe et al. 2000;
Komossa et al. 2003; NGC 3690 West -- Della Ceca et al. 2002, Ballo et
al 2004; VV 340a -- Armus et al. 2009; UGC 5101 -- Imanishi et al.
2003). Note that high-ionization Fe K lines from FeXXV or FeXXVI at
6.7-7 keV that have been found in a few objects, (e.g., NGC 3690 East
-- Ballo et al. 2004; Arp 220 -- Iwasawa et al. 2005), are not
considered here as evidence for AGN since they can also originate from
hot gas produced by a starburst.
The selection by {\sl HR} and the 6.4 keV Fe K line finds 16 objects that
contain AGN. The remaining objects are characterized by relatively soft spectra,
defined by small {\sl HR} values or relative weakness of the hard
X-ray band emission (2-8 keV). The integral spectral properties of
this sample of 29 ``hard X-ray quiet'' (HXQ, as defined by small {\sl
HR}) objects are reported in a separate paper (Iwasawa et al. 2009),
in which the detection and origin of a high-ionization Fe~K line in
the integrated spectrum is discussed. The X-ray classifications of the
sample are given in Table 5, along with optical and SDSS spectral
types, the mid-IR [Ne V] $\lambda 14.3 \mu $m detection, and the X-ray
AGN selection criteria, which are met. We
note that ESO 286-IG19 is classified as AGN in Francschini et al
(2003), based on the XMM-Newton data (see Appendix A for more detail),
while neither our X-ray or mid-IR criteria selected this object as an
AGN. Although ESO 286-IG19 remains as a viable AGN candidate, the
discussion below assume that this object is a HXQ.
In general, AGN being powerful X-ray emitters, an X-ray observation is
a sensitive probe of AGN unless they are hidden behind
Compton thick obscuration. However, it is not necessarily true that
AGN selected by the X-ray technique are energetically important for
the bolometric output of LIRGs. Here we have made an attempt to access the
importance of AGN contribution to the IR luminosity of the objects
hosting the X-ray selected AGN, using the observed X-ray properties.
However the uncertainty in these estimate should be quite large, as
explained below, and they can be considered as a guide only.
There are two steps for estimating the bolometric luminosity of an AGN
from the observed X-ray luminosity: 1) Absorption correction which
recovers the instrinsic X-ray luminosity by removing the flux
suppression effect due to absorption; 2) Bolometric correction which
converts the absorption-corrected X-ray luminosity to the bolometric
luminosity, assuming a typical spectral energy distribution (SED) of
an unobscured AGN (e.g., Elvis et al. 1994). In our rough estimates, the
absorption correction factors of 3 for a Compton thin AGN and 100 for
a Compton thick AGN (i.e., 1\% of 2-10 keV emission from a hidden AGN
is visible) are adopted for the 2-10 keV band. The bolometric
correction is assumed to be $L_{\rm bol}/L_{2-10}=30$, where
$L_{2-10}$ is absorption-corrected 2-10 keV luminosity, which is
adopted from Marconi et al. (2004) for objects with $L_{\rm bol}\sim
10^{12}L_{\odot}$.
When applying the absorption correction, there is an ambiguity
inherent in the AGN selected by X-ray colour alone with a poor quality
spectrum. They show flat X-ray spectra (Fig. 4), as the {\sl HR}
analysis infers, but it is not trivial whether they are Compton thick
or thin objects without a good constraint on a Fe K line. As the
respective absorption correction factors differ significantly, the
classification of the obscuration type introduces a large uncertainty
in the AGN luminosity. Since only weak reflected light is observed
from a Compton thick AGN, we tentatively assume that the X-ray colour
selected AGN with log $(L_{\rm HX}/L_{\rm ir})<-4$ are Compton thick
AGN candidates\footnote{The hard X-ray to [OIII] $\lambda 5007$ ratio
is often used as a diagnostic for a Compton thick AGN (e.g., Bassani
et al 1999). However, this diagnostic may fail for dusty objects
like LIRGs because emission lines from the narrow-line regions are
likely suppressed as well.} and they are marked as such in Table
6. It is also noted that AGN with a reasonable quality spectrum
showing a low-energy cut-off due to moderate absorption, $N_{\rm
H}\sim 10^{22}-10^{23}$ cm$^{-2}$, [F05189--2524, ESO 60-IG16, Mrk
273, 19542+1110] all have larger values between $-3.4$ and $-2.5$ in
log $(L_{\rm HX}/L_{\rm ir})$.
The absorption correction for a Compton thick AGN is also rather
uncertain, since it strongly depends on the geometry of the obscuring
matter, which is not well known. While a uniform correction factor of
$\times 100$ is used here, the final error in $F_{\rm agn}$ can be
easily larger than 0.5 dex, as demonstrated by $F_{\rm agn} > 1$ for
NGC 6240. In any case, taking the face values, the median of log
$F_{\rm agn}$ is $-1$, and there are only two objects (Mrk 231 and NGC
6240) in which the AGN contribution exceeds 50\% of their infrared
luminosity. In summary, according to this simplistic estimate,
AGN-dominated objects would seem to be in the minority in our sample,
and the typical AGN contribution to the infrared luminosity would
appear to be $\sim 10$\%.
\begin{table}
\begin{center}
\caption{Estimates of AGN contribution to IR luminosity}
\begin{tabular}{rlccc}
No. & Galaxy & log (HX/IR) & CT & log $F_{\rm agn}$ \\
& & (1) & (2) & (3) \\[5pt]
4 & F05189--2524 & $-2.63$ & & $-0.65$ \\
7 & ESO 60-IG16 & $-2.54$ & & $-1.54$ \\
8 & F08572+3915 & $-4.44$ & $\ast $ & $-0.94$ \\
9 & 09022--3615 & $-3.59$ & & $-1.59$ \\
12 & UGC 5101 & $-3.92$ & $\circ $ & $-0.42$ \\
16 & NGC 3690 W & $-4.07$ & $\circ $ & $-0.57$ \\
18 & UGC 8058 & $-3.67$ & $\circ $ & $-0.17$ \\
19 & 13120--5453 & $-4.25$ & $\ast $ & $-0.75$ \\
22 & UGC 8689 & $-3.39$ & & $-1.39$ \\
23 & F14348--1447 S & $-4.90$ & $\ast $ & $-1.50$ \\
24 & F14378--3651 & $-4.28$ & $\ast $ & $-0.78$ \\
25 & VV 340 N & $-3.86$ & $\circ $ & $-0.36$ \\
30 & NGC 6240 & $-2.97$ & $\circ $ & $+0.53$ \\
36 & 19542+1110 & $-3.05$ & & $-1.09$ \\
42 & ESO 148-IG2 & $-3.72$ & & $-1.72$ \\
44 & F23365+3604 & $-4.58$ & $\ast $ & $-1.08$ \\
\end{tabular}
\begin{list}{}{}
\item[(1)] Logarithmic luminosity ratio of the observed 2-10 keV and
8-1000 $\mu$m bands, reproduced from Table 3.
\item[(2)] Compton thick AGN with Fe K detection are marked as $\circ
$. Candidate Compton thick AGN (see text for detail) are marked as
$\ast $.
\item[(3)] Logarithmic fraction of the X-ray estimate for AGN contribution
to the 8-1000$\mu$m luminosity.
\end{list}
\end{center}
\end{table}
The X-ray-selected AGN represent 37\% of the total sample
(16/44). This figure is comparable to that inferred from the mid-IR
diagnostics for the same IR luminosity range (Petric et al. 2010). When
the sample is divided into two luminosity bins, above and below the
median IR luminosity of log $(L_{\rm ir}/L_\odot) = 11.99$, 12 of these
16 AGN (75\%) are above the median, in agreement with previous studies
showing the fraction of AGN rising with increasing $L_{\rm ir}$ (e.g.,
Veilleux et al. 1995). If objects with [Ne V]$\lambda 14.3\mu $m
detection (Table 5), which is generally considered as evidence for
AGN, are included, the AGN fraction would increase up to 48\%. In
terms of merger stage, X-ray-selected AGN tend to be found more in
mergers of later stages. For example, 50\% (9/18) of final stage
mergers (nuclear separation $< 1$ kpc) are X-ray-seleted AGN, while
only 26\% (7/27) of earlier stage mergers are X-ray selected AGN. We
refer to Evans et al. (in prep.) for details on the galaxy morphology
and merger-stage classifications based on our HST imaging.
NGC 6240 is the only object in the sample where two nuclei both show
evidence for harboring a powerful AGN \footnote{Mrk 266 (Mazzarella et
al 2010) and Mrk 463 (Bianchi et al. 2008), two GOALS galaxies, but
with $L_{\rm ir}$ below our current C-GOALS sample threshold, are other
objects which are found to have X-ray evidence for AGN in both
nuclei.} out of 24 objects that have double or tripple nuclei. It
should be noted that we do not consider NGC 3690 to have a double AGN,
as the Fe XXV line found in the eastern galaxy is not taken as an AGN
signature. Other AGN are found either in a single nucleus (8 objects)
or in one member of a double nucleus system (7 objects).
\subsection{X-ray luminosities and correlation with $L_{\rm ir}$}
\begin{figure}
\begin{center}
\centerline{\includegraphics[width=0.4\textwidth,angle=0]{f06.pdf}}
\caption{X-ray luminosity distribution in the soft (0.5-2 keV; upper
panel) and hard (2-10 keV; lower panel) bands.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\centerline{\includegraphics[width=0.42\textwidth,angle=0]{f07a.pdf}}
\centerline{\includegraphics[width=0.42\textwidth,angle=0]{f07b.pdf}}
\caption{ Plots of X-ray (upper panel: 0.5-2 keV; lower panel: 2-10
keV) versus infrared luminosity. The X-ray luminosity has been
corrected only for Galactic absorption. The X-ray selected AGN are
marked with filled squares. The typical error bars are indicated. For
objects that are not detected, the 95 per cent upper limits are
indicated by arrows. When multiple components are resolved in a
single system, their luminosities are computed separately (Table
3). The decomposition of the infrared luminosity in individual
objects is described in Appendix A. There is a moderate correlation,
with a correlation coefficient of $\sim 0.6$ for both plots. }
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\centerline{\includegraphics[width=0.4\textwidth,angle=0]{f08.pdf}}
\caption{X-ray to infrared luminosity ratio distribution. }
\end{center}
\end{figure}
The X-ray fluxes were estimated based on a best-fitting spectral model
in the 0.5-2 keV and 2-7 keV bands (i.e., no correction of
absorption). For faint sources, the hard band data are often not good
enough to constrain the shape, in which case a power-law of $\Gamma =
2.1$ is assumed to derive the 2-7 keV flux (e.g., Ranalli et al.
2003). The X-ray luminosities are estimated by correcting only for
Galactic absorption. [Note: Significant absorption intrinsic to the
sources is likely present. However, X-ray spectra of our objects are
complex with multiple components, while most of the spectra do not
have sufficient quality to uniquely decompose and estimate absorption
of each component. For this reason, we opt to present the luminosities
as observed.] The hard band luminosity is for the 2-10 keV band,
which is estimated by extrapolating the 2-7 keV spectrum. For a few
objects, the luminosity values have been updated from those in Iwasawa
et al (2009). The distribution of the X-ray luminosity in the soft and
hard bands is shown in Fig. 6. The luminosity distribution in both
bands peaks at log ($L_{\rm X}) \approx 41.1$ erg\thinspace s$^{-1}$, but the hard
band luminosity is spread over a wider range.
The median logarithmic values for the soft and hard band luminosities
are 41.1erg\thinspace s$^{-1}$ and 41.3 erg\thinspace s$^{-1}$, respectively.
A moderate correlation between the IR and X-ray luminosities can be
seen in Fig. 7, with a typical spread over an order of magnitude. Note that
when multiple components are present in a single object, their
luminosities are plotted separately in Fig. 7. When integrated
luminosities in a single objects are used, as shown in Iwasawa et al.
(2009)\footnote{In Iwasawa et al. (2009), the IR luminosity
($8-1000\mu$m) is also replaced by the FIR luminosity ($40-400\mu$m)
to allow direct comparison with the previous studies on the X-ray --
FIR correlation.}, the correlation becomes less clear. In the soft X-ray
band, AGN are mixed in with the distribution of non-AGN (HXQ) objects,
while in the hard X-ray band, AGN tend to be the more luminous X-ray
sources, as expected.
The X-ray to IR luminosity ratio distribution shown in Fig. 8 has a
significant spread, which is caused by the scatter around the
correlation between the luminosities (e.g. Fig. 7) rather than any
non-linear correlation. Typical values are log ($L_{\rm SX}/L_{\rm
ir}) = -4.53\pm 0.34$ for the soft X-rays and log ($L_{\rm
HX}/L_{\rm ir}) = -4.40\pm 0.63$ for the hard X-rays (the
uncertainties are the dispersion of the distributions). When
X-ray-selected AGN are excluded, a linear relation between logarithmic
X-ray and IR luminosities (Fig. 7) is given by ${\rm log} L_{\rm SX} =
(-4.6\pm 0.1)+{\rm log} L_{\rm ir}$ for the soft X-ray band and ${\rm
log} L_{\rm HX} = (-4.5\pm 0.1)+{\rm log} L_{\rm ir}$ for the hard
X-ray band. These values are significantly lower than those found for
local, star-forming galaxies with low star formation rates, e.g., log
$L_{\rm X}/L_{\rm FIR}\sim -3.7$ (Ranalli et al 2003). Note that a
direct comparison would need a correction for the different infrared
bandpasses (details of the direct comparison can be found in Iwasawa
et al 2009).
\subsection{Radial surface brightness profiles}
\begin{figure}
\begin{center}
\centerline{\includegraphics[width=0.42\textwidth,angle=0]{f09.pdf}}
\caption{The soft X-ray emission radius, $R_{\rm max}$ (kpc) versus
soft X-ray luminosity. Filled symbols are of X-ray-selected AGN and
open symbols are others. The typical error bars are indicated. No
trend can be seen. The median value of $R_{\rm max}$ is 5.3 kpc. }
\end{center}
\end{figure}
\begin{table*}[H]
\begin{center}
\caption{Properties derived from the X-ray radial surface brightness profiles.$^{\mathrm{a}}$}
\begin{tabular}{rlcccccc}
No. & Galaxy & ${\sl AS}$ & $r_{\rm max}$ & $R_{\rm max}$ & $r_{\rm hp}$ & $R_{\rm hp}$ & Morph \\
& & (kpc arcsec$^{-1}$) & (arcsec) & (kpc) & (arcsec) & (kpc) & HX \\[5pt]
\multicolumn{8}{c}{\bf Cycle 8 Data} \\
1 & F01364--1042 & 0.93 & 6.5 & 6.04 & 2.0 & 1.86 & P \\
3 & II Zw 31 & 1.06 & 6.0 & 6.36 & 2.0 & 2.12 & R \\
5 & ESO 255-IG7 N & 0.76 & 7.0 & 5.32 & 2.5 & 1.90 & R \\
5 & ESO 255-IG7 C & 0.76 & 5.0 & 3.80 & 2.0 & 1.52 & A \\
5 & ESO 255-IG7 S & 0.76 & 7.0 & 5.32 & 2.5 & 1.90 & R \\
6 & 07251--0248 & 1.74 & 4.0 & 6.96 & 1.5 & 2.61 & -- \\
7 & ESO 60-IG16 & 0.93 & 7.0 & 6.51 & 2.0 & 1.86 & P \\
9 & 09022--2615 & 1.19 & 5.5 & 6.54 & 1.5 & 1.78 & R \\
10 & F09111--1007 E & 1.08 & 4.0 & 4.32 & 1.0 & 1.08 & A \\
10 & F09111--1007 W & 1.08 & 4.0 & 4.32 & 1.0 & 1.08 & A \\
13 & ESO 374-IG32 & 0.73 & 8.0 & 5.84 & 2.5 & 1.82 & R \\
14 & F10173+0828 & 0.99 & -- & -- & -- & -- & A \\
19 & 13120--5453 & 0.63 & 3.0 & 1.89 & 1.0 & 0.63 & P \\
20 & VV 250 E & 0.63 & 5.5 & 3.46 & 1.5 & 0.95 & R \\
20 & VV 250 W & 0.63 & 6.0 & 3.78 & 2.5 & 1.57 & A \\
21 & UGC 8387 & 0.50 & 9.0 & 4.50 & 3.0 & 1.50 & R \\
24 & F14378--3651 & 1.37 & 3.0 & 4.11 & 1.0 & 1.37 & P \\
25 & VV 340 N & 0.70 & 18.0 & 12.60 & 5.5 & 3.85 & R \\
25 & VV 340 S & 0.70 & $>12.0$ & $>8.40$ & 7.5 & 5.25 & -- \\
29 & ESO 69-IG6 (N) & 0.94 & 3.0 & 2.82 & 1.0 & 0.94 & R \\
31 & F17132+5313 & 1.02 & 7.5 & 7.65 & 2.5 & 2.55 & R \\
33 & F18293--3413 & 0.38 & 10.0 & 3.80 & 3.0 & 1.14 & R \\
34 & ESO 593-IG8 & 0.98 & 12.0 & 11.76 & 4.5 & 4.41 & R \\
35 & F19297--0406 & 1.72 & 6.0 & 10.32 & 2.5 & 4.30 & P \\
36 & 19542+1110 & 1.30 & 5.0 & 6.50 & 1.0 & 1.30 & P \\
37 & CGCG 448-020 & 0.72 & 10.0 & 7.20 & 3.0 & 2.16 & R \\
39 & 21101+5810 & 0.77 & 6.0 & 4.62 & 1.5 & 1.16 & P \\
40 & ESO 239-IG2 & 0.85 & 6.0 & 5.10 & 1.5 & 1.27 & P \\
41 & F22491--1808 & 1.53 & 5.5 & 8.42 & 2.0 & 3.06 & P \\
43 & ESO 77-IG14 N & 0.84 & 4.0 & 3.36 & 1.0 & 0.84 & A \\
43 & ESO 77-IG14 S & 0.84 & 3.5 & 2.94 & 2.0 & 1.68 & P \\[5pt]
\multicolumn{8}{c}{\bf Archival Data} \\
4 & F05189--2524 & 0.83 & 2.0 & 1.7 & 1.0 & 0.83 & P \\
8 & F08572+3915 & 1.13 & -- & -- & -- & -- & A \\
11 & UGC 4881 E & 0.79 & 2.5 & 2.0 & 1.0 & 0.79 & P \\
11 & UGC 4881 W & 0.79 & 8.0 & 7.0 & 3.5 & 3.1 & -- \\
12 & UGC 5101 & 0.79 & 6.0 & 4.7 & 1.5 & 1.2 & P \\
15 & F10565+2448 & 0.87 & 7.0 & 6.1 & 2.0 & 1.7 & P \\
16 & NGC 3690 E & 0.24 & $>13$ & $>3.1$ & -- & -- & R \\
16 & NGC 3690 W & 0.24 & $>13$ & $>3.1$ & -- & -- & R \\
17 & F12112+0305 & 1.42 & 8.0 & 11.3 & 4.0 & 5.7 & P \\
18 & UGC 8058 & 0.85 & 5.0 & 4.3 & 1.0 & 0.85 & P \\
22 & UGC 8696 & 0.77 & 9.0 & 6.9 & 2.5 & 1.9 & R \\
23 & F14348--1447 & 1.58 & 6.5 & 10.3 & 3.0 & 4.8 & P \\
26 & VV 705 & 0.82 & 4.5 & 3.7 & 1.5 & 1.2 & A \\
27 & F15250+3608 & 1.10 & 4.0 & 4.4 & 1.5 & 1.6 & -- \\
28 & UGC 9913 & 0.41 & 30 & 12.2 & 8.5 & 3.5 & R \\
30 & NGC 6240 & 0.53 & 15 & 8.0 & 5.5 & 2.9 & R \\
32 & F17207--0014 & 0.87 & 11.0 & 9.6 & 3.0 & 2.6 & R \\
38 & ESO 286-IG19 & 0.85 & 9.0 & 7.7 & 3.5 & 3.0 & R \\
42 & ESO 148-IG2 & 0.88 & 7.0 & 6.2 & 3.5 & 3.1 & R \\
44 & F23365+3604 & 1.22 & 3.0 & 3.7 & 1.0 & 1.2 & A \\
\end{tabular}
\begin{list}{}{}
\item[$^{\mathrm{a}}$] The radius where the soft X-ray brightness goes below
1\% of the peak brightness at $1^{\prime\prime}$ resolution is defined
as $r_{\rm max}$ (arcsec). The corresponding physical radius, $R_{\rm max}$ (kpc),
is also listed. The half power
radius, $r_{\rm hp}$ (arcsec), is where the cumulative soft X-ray counts exceed
half of the integrated counts within $r_{\rm max}$ at $1^{\prime\prime}$
resolution. The corresponding physical radius, $R_{\rm hp}$ (kpc), is also
given. Whether the 2-7 keV emission is resolved or not is
indicated by ``P" (point-like), ``R" (resolved), and ``A" (ambiguous).
\end{list}
\end{center}
\end{table*}
\clearpage
Radial surface brightness profiles are derived in the soft (0.5-2 keV)
and hard (2-7 keV) bands at a resolution of 1 arcsecond. In general,
the soft X-ray emission is spatially extended while the hard X-ray
emission is compact. A few exceptions are sources dominated by
absorbed hard X-ray emission as seen in objects like IRAS 19542+1110,
where both bands show equally compact emission.
With the $0.5^{\prime\prime}$ resolution of the Chandra ACIS, the
presence of extended soft X-ray emission is immediately clear by
visual inspection for all of the sources. To quantify the soft X-ray
extension, azimuthally-averaged surface brightness profiles are
produced in the soft (0.5-2 keV) and hard (2-7 keV) bands separately
for individual targets. An example of these surface brightness
profiles is shown in Fig. 3 for UGC 8387, and those for all the other
objects are presented in on-line material. The soft X-ray profiles
show a range of shapes -- from an exponetial profile to a more peaky,
power-law type profile. Since not all of the sources have sufficient
counts, the soft X-ray profiles are quantified by two characteristic
radii, rather than fitting a profile model. The maximum extension
radius, $r_{\rm max}$, and half power radius, $r_{\rm hp}$, are
defined as follows: The radius where the surface brightness falls to
1\% of the peak brightness (usually of the central bin of the surface
brightness profile) is defined as $r_{\rm max}$ in arcsec. This radius
gives a measure of the source extension, which has little dependence
on the depth of the image. The total source count, $C_{\rm max}$,
integrated within $r_{\rm max}$ is then derived . $C_{\rm max}$ can
differ from the total source count when the surface brightness, which
is lower than 1\% of the peak brightness, extends to large radii. The
radius where the accumulated source count reaches half of $C_{\rm
max}$ is defined as the half power radius, $r_{\rm hp}$, in
arcsec. These quantities give a measure of the compactness/broadness
of the source extension. Since both $r_{\rm max}$ and $r_{\rm hp}$ are
determined as the innermost radial bin which satisfies the required
condition, their uncertainty is always 0.5 arcsec. For a compact
source (e.g., $r_{\rm hp}\sim 1^{\prime\prime}$), this half power
radius is dominated by the point spread function (PSF), and is likely
to be over-estimated. The radii in physical units (kpc) corresponding
to $r_{\rm max}$ and $r_{\rm hp}$ are also given in Table 7.
The hard X-ray sources are compact and often point-like. We note
however that there are some objects in which hard X-ray emission is
resolved, e.g., Arp 220, and UGC 8387 (=IC 883). As a
guide, the PSF simulated at an energy of 3
keV and normalized to the central peak of the hard X-ray emission is
plotted along side the observed surface brightness profiles. In Table 7,
the hard X-ray morphology is classified into three categories -- R/A/P.
When more than 2 data points deviate by more than
$2\sigma $ from the PSF it is denoted as R (resolved). When the data
points agree with the PSF within the error bars, it is denoted as P (point-like).
Any other case is denoted as A (ambiguous).
Figure 9 shows a plot of $R_{\rm max}$ versus the soft X-ray luminosity.
No clear correlation is seen even if objects containing AGN are removed.
The variety of shapes observed in the soft X-ray radial profiles
could be related to the origin of the X-ray emission. Generally
speaking, a gravitationally bound, virialized system shows an
exponential profile while a power-law profile is expected from a
free-flowing outflow from a compact central source. However, given the
angular scale for our sample galaxies (0.24-1.74 kpc arcsec$^{-1}$),
any compact nuclear starbursts are likely to be contained
within the innermost $1^{\prime\prime}$ bin. There are
several objects showing a power-law type, peaky, soft X-ray profile, and
they are noted in the Notes on Individual Objects (Appendix A). Absorption can
modify the soft X-ray profile if it has a radial dependence, e.g.,
centrally concentrated absorption, which suppress the brightness of
the compact nuclear component, e.g. Arp 220 and VV 340 N.
\section{Discussion}
In \S~4.3, we used relatively crude spectral information based on the X-ray
colour, (i.e. {\sl HR}), to assess the presence of an AGN. Here we discuss further
characterizations of the ACIS spectra for sufficiently bright
objects. Before presenting the spectral analyses, a brief summary of the properties of
X-ray sources in LIRGs and their spectra are described
below.
The primary origin of the X-ray emission in LIRGs is considered to be
a starburst and/or AGN. Given the dusty nature of LIRGs, if an AGN is
present, it is likely to be an absorbed source. Such absorbed X-ray
sources can be selected by X-ray colour analysis, as shown in
\S~4.3. When the absorbing column density exceeds $10^{24}$ cm$^{-2}$
and the absorbed transmitted component moves out of the Chandra
bandpass, the identification of an AGN becomes difficult as it must
rely on the detection of the Fe~K line at 6.4 keV in faint reflected
emission (see the discussion in \S~4.3), and the data are often not of
sufficient depth to make a clear detection. Such heavy nuclear
obscuration could, however, occur in many LIRGs, and sometimes the
hard X-ray band is the only available window to check for signatures
of an AGN, e.g. as for NGC 4945 (e.g., Iwasawa et al 1993; Spoon et al 2000).
In a starburst, there are various sources of X-ray emission (e.g.,
Persic \& Rephaeli 2002) whose origin can be traced back to
massive stars. Individual supernovae and their remnants, stellar-wind
heated ISM in star clusters, and X-ray binaries can all be X-ray
sources. In particular, high-mass X-ray binaries (HMXBs) are
considered to be the dominant source for the X-ray emission above 2
keV. A good correlation between the 2-10 keV luminosity and the star
formation rate have been found for nearby star forming galaxies
(Ranalli et al. 2003; Grimm et al. 2003; Gilfanov et al. 2004) as well as
for late type galaxies at higher redshift (Lehmer et al. 2008). The
galactic-scale emission nebulae, traced by H$\alpha $ and the soft X-ray
emission in local starburst galaxies like M~82, suggested that these
extended nebulae are produced by the shock heated interstellar medium, swept up by a
starburst-driven outflow (see e.g., Veilleux, Cecil \& Bland-Hawthorn
2005 for a review on galactic outflow phenomena, and Tomisaka \&
Ikeuchi 1988; Suchkov et al. 1996; Tenorio-Tagle \& Munoz-Yunon 1998;
Strickland \& Stevens 2000 for simulations). As hypothesized by
Chevallier \& Clegg (1985), hot ($T\sim 10^8$ K), high-pressure gas,
produced via the collective kinetic energy from supernovae and massive
stellar winds, is considered to drive the outflow. Given the
temperature, this hot gas could be a hard X-ray source, but it is also
expected to be rarefied, and thus not very radiative. The presence of weak
emission from such gas has been suggested for M 82 (Griffiths et al. 2000;
Strickland \& Heckman 2007; Ranalli et al. 2008) and NGC 253 (Pietsch et
al 2001).
\subsection{X-ray spectra and derived properties}
Except for some sources which are clearly absorbed AGN, the 0.4-7 keV
Chandra spectra of the majority of the GOALS sources appear similar --
namely an emission-line dominated soft X-ray band (below 2 keV) and a
hard X-ray tail. The relative composition of these two components is
not widely different between objects, as the {\sl HR} distribution
(Fig. 5) shows. Objects that have a spectrum of this type always show
totally different X-ray morphologies in the soft and hard X-ray bands
(see the images included as online material), suggesting distinct
origins for hard and soft X-rays. Therefore, we will discuss the soft
and hard band spectra separately below\footnote{It is generally
possible to reproduce a Chandra full-band spectrum for those objects
whose spectra can be represented as the sum of a power-law and a
thermal emission spectrum with sub-keV temperature and solar
abundance. A good fit can often be found with the thermal spectrum
accounting only for the emission-line bump around 1 keV, leaving not
only the hard band ($\geq 2$ keV) emission but also the softest
emission below $\sim 0.7$ keV to the power-law. This spectral
decomposition would be invalid when the imaging data show a totally
different spatial distribution between these two bands (as seen in
our sample objects), because they cannot come from the same
component represented by the power-law (see Appendix B).}.
As mentioned above, the hard X-ray emission is generally attributed to
HMXBs, {\it but this may not always be true for our objects}. After
excluding objects with clear AGN signatures, the hard band spectra of
the remaining objects have relatively low signal to noise mainly due
to the relative weakness of the source, hence they have been referred
to (see \S 4.3) as HXQ galaxies in Iwasawa et al. (2009). Any spectral
feature like an Fe~K line is not evident in individual noisy
spectra. However, the stacked spectrum of the 29 HXQ galaxies in our
C-GOALS sample shows a strong high-ionization Fe K line (Fe XXV) at
6.7 keV with EW of $\simeq 0.9$ keV (Iwasawa et al. 2009). This means
that a non negligible fraction of the HXQ galaxies have strong Fe XXV
emission in their hard band spectra, but they are buried in the noise
in the individual spectra. Such a strong Fe XXV is not seen in the
spectrum of HMXBs in our Galaxy, except when they are in an eclipse
phase -- they normally show a much weaker Fe line at 6.4~keV
(e.g. White et al. 1983, Torrejon et al. 2010). This means that, {\it unlike
for local star-froming galaxies, HMXBs are not the primary source of
the hard X-ray band emission seen in the more luminous C-GOALS
objects}. Also the 2-10 keV luminosity of the HXQ galaxies deviate
from the extrapolated correlation line on the X-ray quiet side ($\sim
0.7$ dex on average, Iwasawa et al. 2009). Most objects used to derive
the X-ray vs. FIR correlation (e.g., Ranalli et al. 2003) have $L_{\rm
ir} < 10^{11}L_{\odot}$, while the lower bound of the IR luminosity
of our sample is $10^{11.73}L_{\odot}$. {\it This suggests that a
transition in the nature of the dominant hard X-ray source in LIRGs
occurs somewhere in the range log $(L_{\rm ir}/L_\odot) = $11.0-11.73} .
In the soft X-ray band, at the spectral resolution of a CCD
spectrometer like the ACIS, heavy blending of emission lines makes it
difficult to distinguish between photoionized gas irradiated by an AGN
and thermal gas induced from a starburst. However, it is genrally
assumed that thermal emission is responsible for the emission-line
dominated spectra of LIRGs in the soft X-ray band, and this appears
even to be true for an object like NGC 6240 in which AGN signatures
appear only in the hard band (Netzer et al. 2005). Thus we also utilize
the thermal emission spectrum to compare with the soft X-ray
spectra. However, we still suspect that there might be a possible
contribution from photoionized gas to the soft X-ray spectra of some
objects, as will be discussed below.
In applying a thermal emission model, one problem specific to our
objects is the assumed solar abundance. The hot gas induced by a
starburst is expected to be polluted heavily by ejecta of core
collapse supernovae, which is rich in $\alpha $ elements relative to
iron. Therefore the abundance pattern should deviate significantly
from the solar pattern, and this has been found to be the case for
starburst knots in nearby galaxies, e.g., The Antennae (Fabbiano et al.
2004). When the data quality is good, a standard thermal emission
spectrum with a solar abundance pattern as computed by, e.g., MEKAL,
indeed does not agree with some of the observed emission line
strengths, and a modified abundance pattern, rich in $\alpha $
elements, gives a better description. This may not be evident for poor
quality spectra. Since the data quality between our spectra varies, we
first measure a spectral slope and then fit the thermal emission model
with the abundance pattern of core collapse SNe for all of the soft
X-ray data. These spectral fits were performed to the count rate
spectra by comparing the model folded through the detector response,
using XSPEC (ver. 11).
\subsubsection{Spectral slopes}
As a first order characterization of the spectral shape, spectral
slopes in the soft (0.8-2 keV) and hard (3-7 keV) bands, derived from
power-law fitting are given in Table 8. The photon index, $\Gamma$, is the
slope parameter of a power-law model for describing a photon spectrum,
defined as $dN/dE\propto E^{-\Gamma }$ photons cm$^{-2}$ s$^{-1}$
keV$^{-1}$, and is related to the energy index $\alpha $ ($F_{\rm
E}\propto E^{-\alpha}$) with $\Gamma = \alpha +1$. In order to
obtain meaningful constraints, the slopes were determined for those
spectral data which have more than 50~cts in the energy range of
interest. In fitting the power-law, Galactic absorption is assumed.
In most objects, the soft X-ray emission is dominated by (largely
unresolved) emission lines, and the spectrum turns over below 0.8 keV
where the Fe~L complex becomes weak. The power-law fits for the soft X-ray
spectra are therefore mainly for the emission-line blend, not the
underlying continuum.
\begin{figure}
\begin{center}
\centerline{\includegraphics[width=0.4\textwidth,angle=0]{f10.pdf}}
\caption{Distribution of spectral slope (photon index) measured by a
power-law fit in the 0.8-2 keV and 3-7 keV bands. A correction for Galactic
absorption has been made.}
\end{center}
\end{figure}
\subsubsection{Prominent emission lines}
The Fe~K line is a well known diagnostic of heavily obscured AGN, and
it is clearly seen in some spectra of our sample galaxies. Another
prominent spectral line we observe is Si K at $\sim 1.8$ keV. The
strongest line is usually from Si XIII at 1.85 keV. This line is
relatively isolated in the soft X-ray range so that the line
properties are easier to measure than is the case for other emission
lines. It could also provide a clue to the origin of the X-ray
emission.
When a line is detected at $> 2\sigma $ after fitting with a Gaussian,
its line flux and equivalent width with respect to the neighboring
continuum are measured. Tables 9 and 10 lists measured values for the Si and Fe lines,
respectively. Fe~K features are found at 6.4~keV (line emission from
cold Fe) and/or at higher energies (Fe XXV or Fe XXVI). The detection
of these Fe~K lines have been reported previously, based on
XMM-Newton, ASCA, and Chandra observations (Imanishi et al. 2003, Xia
et al. 2002, Armus et al. 2009, Komossa et al. 2003; Della ceca et al.
2002, Ballo et al. 2004; Clements et al. 2002, Iwasawa et al. 2005;
Iwasawa \& Comastri 1998; Turner \& Kraemer 2003, Braito et al.
2004). One addition is F17207--0014 for which a possible
high-ionization Fe~K line is detected ($2\sigma$) for the first time.
\begin{figure}
\begin{center}
\centerline{\includegraphics[width=0.4\textwidth,angle=0]{f11.pdf}}
\caption{The distribution of temperature, $kT$, when the thermal
emission spectrum is fitted to the 0.4-2 keV data (see Table
11). The median value is $kT = 0.88$ keV. }
\end{center}
\end{figure}
\subsubsection{Metal abundance pattern}
As discussed above, the soft X-ray spectrum may be more complicated
than a standard thermal emission spectrum with a solar metal-abundance
pattern. For a more physically motivated characterization of the soft
X-ray emission than given by a simple power-law fit, we adopt a
thermal emission model with a Type II SN abundance pattern (e.g.,
Nomoto et al. 1997) to fit the 0.4-2 keV spectra. The nominal
metallicity ratio relative to Oxygen for various metals is assumed (as
given in Dupke \& Arnaud 2001) and varies together in the fit: (Mg,
Si)/O = 1, (Ne, S)/O = 0.67, (Ar, Ca, Ni)/O = 0.46, Fe/O = 0.27. We
note that this abandance pattern is one of those calculated for
core-collapse SNe and may not be very accurate, but it deviates
significantly from the solar pattern and will provide a good reference
against the observed data. The metallicity measurement is primarily
driven by the strength of the Fe~L complex around 1 keV at the
temperature assumed for the gas. The metallicity results are given
for the O abundance relative to solar ($Z_{\rm SNII}$ in Table 11) in
the interest of $\alpha $ elements, i.e., $Z_{\rm SNII}$ is an
``equivalent'' Oxygen abundance, which is mainly determined by the Fe
L feature and converted by the above Fe/O ratio, since given the
quality of the data, the Oxygen line (e.g., OVIII at 0.65 keV) itself
cannot give a strong constraint. Here the solar abundance is that of
Anders and Grevess (1989).
Emission lines of He-like Mg and Si are located at
energies where the ACIS response is good, and in some qood quality
data their metallicity can be measured independently (when they are
strong). When this is the case, their metallicity measurements are
given separately. In this case, Si and Mg are assumed to have the same
abundance.
This prescription described above usually gives better agreement with
the data than that achieved with a solar abundance pattern, as
demonstrated for the spectrum of ESO 286-IG19 in Fig. 12. When a solar
abundance pattern is assumed, the metallicity is found to be only
$0.21^{+0.05}_{-0.03} Z_{\odot}$, and clearly underestimates
emission-lines such as Si XIII, which would be better described with
$\sim 2 Z_{\odot}$ (Table 11). The temperature measurement is also
affected by the choice of abundance pattern. In the example of ESO
286-IG19, the temperature derived from the solar abundance pattern is
$kT = 0.82\pm 0.03$, while the SNII pattern gives $kT = 0.70\pm 0.04$
keV.
The bright X-ray core of Mrk 231 is dominated by the unresolved
nuclear component even in the soft X-ray band. Given that a single thermal
emission model is not able to give a reasonable fit, the spectral
parameters are not presented in Table 11.
The argument for a non-solar abundance pattern has so far been based
on a single-temperature fit. One caveat is that a multi-temperature
model could mimic the same effect (e.g., Buote \& Fabian 1998) in
some cases, since it is difficult to distinguish between the two
interpretations with spectral data at the CCD resolution of the ACIS.
High-resolution spectroscopy that could identify individual lines could
be used to determine the characteristic temperature of the gas. This would tell whether
the gas is in a multi-phase. At least for the very extended nebulae,
as seen in Mrk 231 and Mrk 273, which emit almost no emission above 2
keV but strong Mg XI and Si XIII, higher temperature gas is unlikely
to be present. We note $\alpha $-element rich gas would be produced
naturally if a starburst is responsible for the metal enrichment.
\subsubsection{Strong Si XIII emission lines and their origin}
Even when assuming the $\alpha $-element rich abundance pattern of
core-collapse SNe, the observed Si metallicity sometimes requires an
even larger value (see Table 11). This may simply be due to more Si
enriched gas than theoretically predicted. However, the excess Si line
intensity could also be due to an additional source of the line
emission, e.g., photoionized gas illuminated by a hidden AGN. In fact,
among the objects listed in Table 11, all of the AGN that are identified
by spectral hardness and/or the Fe~K diagnostics show at least an
excess of $\times 2$ in the Si metallicity relative to $Z_{\rm
SNII}$. Since the energies around Si XIII (1.84 keV) are where the
spectrum of the soft X-ray emitting thermal gas starts to decline
rapidly, an extra contribution from any AGN related component would stand
out relatively well.
Here, we examine the Si line strengths of starbursts and AGN in an
alternative way by using the Si XIII detected objects in Table 9. Since
the metallicity, $Z_{\rm SNII}$, is driven by the Fe L emission (\S
5.1.3), the enhancement of $Z_{\rm Si}$ is basically the relative
strength of the Si XIII line to the Fe L bump around 0.8-1 keV. As the
heavily blended Fe L emission is not resolved at the CCD spectral
resolution, the mean 0.8-1 keV intensity is used as a crude measure of
Fe L emission, assuming that Fe L dominates in this band. Thus, the
mean Si XIII / Fe L ratio can be assessed between starburst and
AGN.
The starburst sample consists of 7 objects, F10565+2448, NGC 3690 E,
Arp 220, F17207--0014, F18293--3413, CGCG 448-020 and ESO
286-IG19. Note that objects with mid-IR [Ne V] detection are not included (see
below). The AGN sample consists of 5 known Compton-thick AGN, UGC 5101, Mrk
231, 13120--5453, VV 340 N and NGC 6240.
When making a mean spectrum of each sample, a redshift correction
was made due to the fact that the energy-scale shifts between objects
are not negligible. The spectral data of individual objects are chosen
so that the rest-frame 0.50-2.18 keV range is covered and then they
are binned at (rest-frame) 21 eV intervals. These spectra are
corrected for the effective area in the same way as those spectra in
Fig. 4 (\S 4.2), which is necessary when the redshift correction for
the energy scale is made. Since the soft X-ray band is sensitive even
to variations in Galactic column ($N_{\rm H}$ ranges between $9\times
10^{19}$ cm$^{-2}$ and $2.1\times 10^{21}$ cm$^{-2}$ for the relevant
objects here), an absorption correction for the Galactic column is also
made. Individual spectra are then normalised to the 0.8-1 keV
intensity before computing an average from the individual members of each
sample.
The mean 0.5-2.2 keV spectra of the starburst and AGN samples are
shown in Fig. 13a and 13b, where the 0.8-1 keV intensity of both spectra
has been set to the same level. The Si line strength relative
to the Fe L emission can be readily compared. The Si line of the mean
AGN spectrum appears to be stronger than that of the starburst by a
factor of $2.0\pm 0.6$. Albeit that this is a relatively crude measure, the result is
consistent with an enhanced Si line in AGN spectra, as suggested by the
metallicity fitting (Table 11).
Furthermore, an additional mean spectrum was constructed using the same
procedure, for five objects which are not selected as AGN by the
X-ray criteria, but which show a detectable mid-IR [Ne V] line in their
Spitzer IRS spectra (Petric et al. 2010): VII Zw 31, F09111--1007, UGC
8387, F17132+5313 and ESO 593-IG8 (see Table 5). The mean spectrum of
these Ne V detected objects also suggests an enhancement of Si XIII;
although the data are still noisy, the enhancement relative to the
starburst sample is a factor of $1.8\pm 0.9$.
While it is premature to propose such enhanced Si XIII as an
alternative to the Fe K diagnostic for a hidden AGN, it is interesting
to note that both X-ray selected Compton thick AGN and mid-IR selected
AGN candidates appear to show strong Si XIII on average, even when the data
quality is not good in the Fe K band. A contribution from an AGN
photoionized component in addition to a thermal component can be a
feasible explanation, unless the photoionized gas region is
affected by internal absorption.
\begin{figure*}
\begin{center}
\hbox{\centerline{\includegraphics[width=0.38\textwidth,angle=0]{f12a.pdf}
\includegraphics[width=0.38\textwidth,angle=0]{f12b.pdf}}}
\caption{The Chandra ACIS 0.4-2 keV spectrum of ESO 286-IG19, fitted
with a thermal emission spectrum (MEKAL) with a solar abundance
pattern (left) and that of the core-collapse SNe (right), in which
OVIII (0.65 keV), NeIX (1.0 keV), MgXI (1.36 keV) and SiXIII (1.85
keV) are more pronounced. The $\chi^2$ values for the two fits are
53 and 37 for 41 and 40 degrees of freedom, respectively. }
\end{center}
\end{figure*}
\begin{table}
\begin{center}
\caption{X-ray spectral slopes.$^{\mathrm{a}}$}
\begin{tabular}{rlcc}
No. & Galaxy & $\Gamma_{\rm S}$ & $\Gamma_{\rm H}$ \\[5pt]
\multicolumn{4}{c}{\bf Cycle 8 targets} \\
3 & VII Zw 31 & $2.7^{+0.4}_{-0.4}$ & --- \\
5 & ESO 255-IG7 N & $2.0^{+0.7}_{-0.6}$ & --- \\
5 & ESO 255-IG7 C & $2.0^{+0.5}_{-0.4}$ & --- \\
7 & ESO 60-IG16 & $2.1^{+1.4}_{-1.3}$ & $0.7^{+0.3}_{-0.8}$ \\
9 & 09022-3615 & $2.3^{+0.6}_{-0.6}$ & $-0.1^{+0.8}_{-0.3}$ \\
10 & F09111-1007 E & $3.1^{+0.9}_{-0.8}$ & --- \\
19 & 13120-5453 & $1.0^{+0.4}_{-0.4}$ & $2.6^{+1.5}_{-0.9}$ \\
20 & VV 250 E & $1.7^{+0.4}_{-0.4}$ & $3.6^{+0.9}_{-0.8}$ \\
21 & UGC 8387 & $2.5^{+0.4}_{-0.4}$ & --- \\
25 & VV 340 N & $3.8^{+0.4}_{-0.4}$ & $-2.5^{+1.7}_{-0.5}$ \\
29 & ESO 69-IG6 N & $3.0^{+0.4}_{-0.4}$ & --- \\
31 & F17132+5313 & $2.4^{+1.4}_{-1.1}$ & --- \\
33 & F18293-3413 & $2.7^{+0.4}_{-0.3}$ & $2.0^{+0.4}_{-1.0}$ \\
34 & ESO 593-IG8 & $2.2^{+0.6}_{-0.6}$ & --- \\
35 & F19297-0406 & $3.6^{+1.5}_{-1.4}$ & --- \\
36 & 19542+1110 & $\sim -2.8$ & $1.3^{+0.2}_{-0.5}$ \\
37 & CGCG 448-020 & $2.3^{+0.3}_{-0.3}$ & $1.7^{+0.5}_{-0.4}$ \\
40 & ESO 239-IG2 & $2.7^{+0.4}_{-0.4}$ & --- \\[5pt]
\multicolumn{4}{c}{\bf Archival data} \\
4 & F05189-2524 & $\sim 0$ & $0.48^{+0.20}_{-0.08}$ \\
12 & UGC 5101 & $2.0^{+0.3}_{-0.3}$ & $\sim -0.7$ \\
15 & F10565+2448 & $2.5^{+0.3}_{-0.3}$ & --- \\
16 & NGC 3690 E & $1.9^{+0.2}_{-0.2}$ & $1.0^{+0.7}_{-0.7}$ \\
16 & NGC 3690 W & $1.9^{+0.2}_{-0.2}$ & $1.1^{+0.5}_{-0.5}$ \\
18 & UGC 8058 core & $1.9^{+0.2}_{-0.1}$ & $0.71^{+0.15}_{-0.06}$ \\
18 & UGC 8058 nebula & $5.2^{+0.3}_{-0.3}$ & --- \\
22 & UGC 8696 & $2.8^{+0.2}_{-0.2}$ & $-1.8^{+0.1}_{-0.3}$ \\
26 & VV 705 N & $2.4^{+0.6}_{-0.5}$ & --- \\
28 & UGC 9913 & $1.8^{+0.2}_{-0.2}$ & $0.8^{+0.5}_{-0.3}$ \\
32 & F17207--0014 & $2.0^{+0.3}_{-0.3}$ & $0.7^{+0.9}_{-0.4}$ \\
38 & ESO 286-IG19 & $3.8^{+0.2}_{-0.2}$ & $0.7^{+1.0}_{-2.0}$ \\
42 & ESO 148-IG2 N & $3.0^{+0.5}_{-0.4}$ & --- \\
42 & ESO 148-IG2 S & $1.3^{+0.3}_{-0.3}$ & $1.4^{+0.3}_{-0.4}$ \\
\end{tabular}
\begin{list}{}{}
\item[$^{\mathrm{a}}$] Photon indices derived by fitting a power-law to the 0.8-2 keV
($\Gamma_{\rm S}$) and 3-7 keV ($\Gamma_{\rm H}$) bands after correcting for Galactic
absorption. Measurements are given only when the
detected source counts exceed 50~cts in the respective bands in order
for the spectral slope to have meaningful constraints.
\end{list}
\end{center}
\end{table}
\begin{table}
\begin{center}
\caption{Si {\sc xiii} (1.84 keV) measurements. $^{\mathrm{a}}$}
\begin{tabular}{rlccc}
No. & Galaxy & $I_{\rm Si}$ & $EW_{\rm Si}$ & Note\\
& & $10^{-6}$ ph\thinspace cm$^{-2}$\thinspace s$^{-1}$ & keV & \\[5pt]
10 & F09111--1007 E & $0.74^{+4.5}_{-3.7}$ & 0.44 &\\
12 & UGC 5101 & $0.66^{+0.22}_{-0.30}$ & 0.23 &\\
15 & F10565+2448 & $6.1^{+3.4}_{-2.8}$ & 0.17 &\\
16 & NGC 3690 E & $2.3^{+1.1}_{-1.1}$ & 0.15 &\\
18 & UGC 8058 core & $0.60^{+0.26}_{-0.27}$ & 0.05 &\\
18 & UGC 8058 nebula & $0.30^{+0.15}_{-0.14}$ & 0.39 &\\
19 & 13120--5453 & $2.1^{+0.8}_{-0.8}$ & 0.20 &\\
21 & UGC 8387 & $1.5^{+0.7}_{-0.7}$ & 0.26 &\\
25 & VV 340 N & $1.1^{+0.4}_{-0.5}$ & 0.26 &\\
28 & UGC 9913 & $1.3^{+1.5}_{-0.5}$ & 0.23 & ${\mathrm b}$\\
29 & ESO 69-IG6 N & $1.3^{+0.6}_{-0.6}$ & 0.34 &\\
30 & NGC 6240 & $20.4^{+3.5}_{-4.0}$ & 0.30 & ${\mathrm b}$\\
32 & F17207--0014 & $0.50^{+0.42}_{-0.20}$ & 0.14 &\\
33 & F18293--3413 & $4.0^{+0.8}_{-1.0}$ & 0.42 &\\
34 & ESO 593-IG8 & $0.85^{+0.54}_{-0.43}$ & 0.26 &\\
37 & CGCG 448-020 & $1.1^{+0.7}_{-0.5}$ & 0.16 & ${\mathrm c}$\\
38 & ESO 286-IG19 & $0.59^{+0.28}_{-0.26}$ & 0.16 &\\
\end{tabular}
\end{center}
\begin{list}{}{}
\item[$^{\mathrm a}$] Detections with $2\sigma$ or higher significance are listed.
\item[$^{\mathrm b}$] This line intensity is of a blend of Si~{\sc xiii} and Si~{\sc xiv}
and is measured by fitting a broad Gaussian.
\item[$^{\mathrm c}$] This emission-line is not found at the expected
energy of Si {\sc xiii}, but at $1.60\pm 0.06$ keV.
\end{list}
\end{table}
\begin{table}
\begin{center}
\caption{Fe K line features.$^{\mathrm{a}}$}
\begin{tabular}{rlccc}
No. & Galaxy & $E$ & $I_{\rm FeK}$ & $EW_{\rm FeK}$ \\
& & keV & $10^{-6}$ ph s$^{-1}$ cm$^{-2}$ & keV \\[5pt]
\multicolumn{5}{c}{\bf Cold line} \\
4 & F05189--2524 & 6.4 & $6.0^{+4.0}_{-2.8}$ & 0.12 \\
12 & UGC 5101 & 6.4 & $0.68^{+0.55}_{-0.31}$ & 0.30 \\
18 & UGC 8058 & 6.4 & $1.0^{+0.3}_{-0.4}$ & 0.10 \\
22 & UGC 8696 & 6.4 & $6.7^{+2.7}_{-2.4}$ & 0.24 \\
25 & VV 340 N & 6.4 & $1.8^{+1.1}_{-0.9}$ & 1.2 \\
30 & NGC 6240 & 6.4 & $15.4^{+2.0}_{2.0}$ & 0.40 \\[5pt]
\multicolumn{5}{c}{\bf High-ionzation line} \\
16 & NGC 3690 E & $6.65^{+0.03}_{-0.02}$ & $7.4^{+2.1}_{-2.5}$ & --- \\
28 & UGC 9913 & $6.64^{+0.03}_{-0.03}$ & $0.92^{+0.51}_{-0.32}$ & 0.93 \\
30 & NGC 6240 & $6.65^{+0.02}_{-0.02}$ & $8.0^{+2.0}_{-1.6}$ & 0.16 \\
32 & F17207--0014 & $6.90^{+0.05}_{-0.05}$ & $0.82^{+0.73}_{-0.41}$ & 0.82 \\
\end{tabular}
\begin{list}{}{}
\item[$^{\mathrm a}$] Line detections with $2\sigma$ or higher significance are
listed. The cold Fe~K$\alpha$ at 6.4~keV and the high-ionization
Fe~K$\alpha$, Fe ~{\sc xxv} or Fe~{\sc xxvi} are listed separately. The
line centroid energy is measured in the rest frame in keV.
\end{list}
\end{center}
\end{table}
\begin{table}
\setlength{\tabcolsep}{0.05in}
\begin{center}
\caption{Temperature, absorption and metallicity.$^{\mathrm{a}}$}
\begin{tabular}{rlcccc}
No. & Galaxy & $kT$ & $N_{\rm H}$ & $Z_{\rm SNII}$ & $Z_{\rm Si}$ \\
& & keV & $10^{21} cm^{-2}$ & $Z_{\odot}$ & $Z_{\odot}$ \\[5pt]
\multicolumn{6}{c}{\bf Cycle 8 targets} \\
3 & VII Zw 31 & $1.0^{+0.1}_{-0.1}$ & --- & $4_{-2}$ & $8^{+20}_{-5}$ \\
5 & ESO 255-IG7 N & $1.8^{+0.8}_{-0.5}$ & --- & $0.4^{+1.6}_{-0.4}$ & --- \\
9 & 09022-3615 & $1.1^{+0.5}_{-0.2}$ & --- & $0.05^{+0.1}_{-0.05}$ & $0.4^{+0.2}_{-0.2}$ \\
10 & F09111-1007 E & $0.90^{+0.18}_{-0.13}$ & --- & $1.0^{+1.2}_{-0.5}$ & $4^{+7}_{-2}$ \\
13 & ESO 374-IG32 & $0.68^{+0.15}_{-0.12}$ & --- & $0.6^{+3}_{-0.4}$ & $1.7^{+3.7}_{-1.2}$ \\
19 & 13120-5453 & $0.82^{+0.26}_{-0.14}$ & --- & $0.7^{+0.8}_{-0.3}$ & --- \\
20 & VV 250 & $1.0^{+0.1}_{-0.1}$ & --- & $0.7^{+0.5}_{-0.2}$ & $2.6^{+1.9}_{-0.7}$ \\
21 & UGC 8387 & $1.0^{+0.1}_{-0.1}$ & --- & $2.2^{+2.3}_{-0.7}$ & $7^{+9}_{-3}$ \\
25 & VV 340 N & $0.83^{+0.08}_{-0.07}$ & --- & $1.3^{+0.8}_{-0.4}$ & $3.0^{+2.6}_{-1.1}$ \\
29 & ESO 69-IG6 N & $0.88^{+0.10}_{-0.10}$ & --- & $4.5^{+20}_{-1.9}$ & $10^{+18}_{-6}$ \\
31 & F17132+5313 & $0.59^{+0.10}_{-0.10}$ & --- & $0.7_{-0.4}$ & --- \\
33 & F18293-3413 & $0.61^{+0.08}_{-0.07}$ & $7.2^{+0.8}_{-1.0}$ & $1.4^{+4.9}_{-0.7}$ & --- \\
34 & ESO 593-IG8 & $0.95^{+0.16}_{-0.11}$ & --- & $0.53^{+0.62}_{-0.29}$ & $2.1^{+1.5}_{1.0}$ \\
35 & F19297-0406 & $1.1^{+0.3}_{-0.2}$ & --- & $5_{-4}$ & --- \\
37 & CGCG 448-020 & $0.66^{+0.23}_{-0.09}$ & $2.9^{+0.7}_{-1.0}$ & $0.09^{+2.3}_{-0.09}$ &--- \\
40 & ESO 239-IG2 & $0.76^{+0.07}_{-0.06}$ & --- & $1.2^{+5}_{-0.6}$ & $3.5^{+10}_{-1.8}$ \\[5pt]
\multicolumn{6}{c}{\bf Archival data} \\
12 & UGC 5101 & $0.89^{+0.08}_{-0.07}$ & --- & $0.9^{+0.2}_{-0.2}$ & $4.7^{+1.5}_{-1.4}$ \\
15 & F10565+2448 & $1.0^{+0.1}_{-0.1}$ & --- & $1.1^{+0.5}_{-0.3}$ & $3.0^{+1.6}_{-1.0}$ \\
16 & NGC 3690 & $0.66^{+0.04}_{-0.04}$ & $2.4^{+0.3}_{-0.3}$ & $0.30^{+0.05}_{-0.05}$ & $0.43^{+0.10}_{-0.10}$ \\
18 & UGC 8058 neb & $0.50^{+0.03}_{-0.03}$ & --- & $0.31^{+0.06}_{-0.06}$ & $0.96^{+0.43}_{-0.33}$ \\
22 & UGC 8696 & $0.88^{+0.04}_{-0.04}$ & --- & $0.52^{+0.08}_{-0.07}$ & $1.1^{+0.2}_{0.2}$ \\
22 & UGC 8696 neb & $0.58^{+0.04}_{-0.04}$ & --- & $1.3^{+0.8}_{-0.3}$ & $1.7^{+2.2}_{0.8}$ \\
23 & F14348--1447 & $0.92^{+0.19}_{-0.15}$ & --- & $1.2^{+10}_{-0.8}$ & --- \\
26 & VV 705 N & $0.94^{+0.12}_{-0.09}$ & --- & $2.8^{+3.7}_{-1.7}$ & $8^{+9}_{-4}$ \\
28 & UGC 9913 & $0.90^{+0.07}_{-0.06}$ & $2.9^{+0.4}_{-0.3}$ & $0.37^{+0.12}_{-0.10}$ & $0.70^{+0.28}_{-0.21}$ \\
30 & NGC 6240 & $0.91^{+0.03}_{-0.03}$ & $3.0^{+0.x}_{-0.x}$ & $0.24^{+0.03}_{-0.03}$ & $0.54^{+0.06}_{-0.05}$ \\
32 & F17207--0014 & $0.59^{+0.12}_{-0.10}$ & $5.2^{+1.0}_{-1.0}$ & $0.30^{+0.32}_{-0.15}$ & $0.59^{+0.25}_{-0.23}$ \\
38 & ESO 286-IG19 & $0.70^{+0.04}_{-0.04}$ & --- & $1.0^{+0.3}_{-0.3}$ & $2.0^{+0.6}_{-0.5}$ \\
42 & ESO 148-IG2 N & $0.84^{+0.13}_{-0.10}$ & --- & $2.7^{+10}_{-1.3}$ & $6^{+8}_{-3}$ \\
42 & ESO 148-IG2 S & $1.01^{+0.18}_{-0.18}$ & --- & $0.73^{+0.42}_{-0.27}$ & $3.2^{+1.8}_{-1.0}$ \\
\end{tabular}
\begin{list}{}{}
\item[$^{\mathrm{a}}$] The 0.4-2 keV data are fitted with a thermal
emission spectrum assuming the metal abundance pattern of
core-collapse supernovae (see text for details). When excess
absorption above the Galactic value is required, the best-fit column
density ($N_{\rm H}$) is listed. $Z_{\rm SNII}$ is the metallicity
for Oxygen, which is tied to the other elements with the assumed
abundance pattern of Type II SNe. This value is mainly driven by the
Fe metallicity, which is assumed to be 0.27 times that of the Oxygen
metallicity, and which fits the Fe~L hump around 1~keV. When excess
metallicity for Si (and Mg) is required, the Silicon metallicity is
fitted independently ($Z_{\rm Si}$).
\end{list}
\end{center}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=0.37\textwidth,angle=0]{f13.pdf}
\caption{The mean 0.5-2.2 keV spectra of a) seven starbursts; b) five
Compton-thick AGN; and c) five non X-ray AGN but with [Ne V]$\lambda
14.3\thinspace \mu$m detection. The energy scale is in the rest
frame. The individual spectra have been corrected for Galactic
absorption and the detector effective area and then normalized to
the 0.8-1 keV intensity before averaging within each sample. These
mean spectra are constructed in order to facilitate comparison of the strength of Si
XIII at 1.84 keV relative to the 0.8-1 keV band where Fe L emission
dominates between the different classes of objects (see text for
detail).}
\end{center}
\end{figure}
\subsection{Discrete X-ray emission from star clusters}
There are a few objects that show discrete X-ray sources separate from
the main body of the galaxy emission. Some of these discrete X-ray
sources have clear optical counterparts identified in our high
resolution HST-ACS images. The optical counterparts are star clusters
or dwarf galaxies in the tidal tails. In most cases, only a few X-ray
counts are detected in the soft band. Inferred X-ray luminosities are
on the order of $10^{39}$ erg\thinspace s$^{-1}$. They are much more powerful than the
super star clusters in our Galaxy, for which the primary X-ray
production mechanism is considered to be shock heating by stellar
winds. If the X-ray emission in the C-GOALS objects comes from a
single source in a star cluster, then this could be an example of an ultra-luminous X-ray
source (ULX). Prime examples are the southern source in UGC 8387,
and the eastern triple source in CGCG 448-020 (see
Fig. 2).
\section{Summary}
\begin{list}{}{}
\item[1.] We present Chanda-ACIS data for a complete sample of 44
objects which represent the high luminosity portion of the GOALS
sample. The objects in this C-GOALS sample have
log ($L_{\rm ir}/L_\odot) = 11.73-12.57$ with $z = 0.010-0.088$, and
represent the most luminous IR selected galaxies in the local
universe. X-rays are detected from 43 out of the 44 objects, and their
arcsec-resolution images, spectra, and radial brightness profiles
are presented.
\item[2.] Objects with a clear AGN X-ray signature represent 37\%
(16/44) of the total sample, and 75\% (12/16) of these AGN are found
in the higher IR luminosity half of the sample at log $(L_{\rm
ir}/L_\odot) > 12.0)$. The AGN fraction would increase to 48\% if
objects with [Ne V]$\lambda 14.3\mu $m detection are included. These
AGN, however, appear to be a relatively minor ($\sim 10$\%)
component in the total energy output of the host galaxies apart from
a few exceptions.
\item[3.] While most objects show evidence for strong
interactions/mergers, NGC 6240 remains the only object to clearly
show a double AGN in our X-ray spectra. Eight AGN are found in
single nucleus objects and an additional seven have X-ray AGN in one of
the double nuclei.
\item[4.] For objects without obvious X-ray AGN signatures, X-ray
luminosities are found to be lower than expeted from their IR
luminosities based on the correlation found for nearby star-forming
galaxies with lower star formation rates. The hard X-ray emission of
these objects does not always appear to be due to HMXBs, given that
the stacked spectrum shows strong Fe XXV, a signature of hot gas.
\item[5.] The extended soft X-ray emission found in many objects shows
a spectrum consistent with a thermal emission spectrum with
an abundance pattern with enhanced $\alpha $ elements relative to
iron, as expected for an interstellar medium enriched by core collapse
SNe produced in a starburst.
\item[6.] A comparison between the soft X-ray spectra of starburst
galaxies and galaxies containing Compton thick AGN shows that, on
average, the latter show stronger Si XIII emission at 1.85 keV which
may be due to a contribution of AGN photoionized gas besides thermal
emission from a starburst.
\end{list}
\begin{acknowledgements}
This research has made use of software packages CIAO provided by the
Chandra X-ray Center (CXC) and FTOOLS provided by NASAS's HEASARC.
This research has also made use of the NASA/IPAC Extragalactic
Database (NED), which is operated by the Jet Propulsion Laboratory,
California Institute of Technology, under contract with the National
Aeronautics and Space Administration. The Chandra data archive is
maintained by the Chandra X-ray Center at the Smithsonian
Astrophysical Observatory. DBS acknowledges support from Chandra
grants GO6-7098X and GO7-8108A. Vivian U also acknowledges support
from a NASA Jenkins Predoctoral Fellowship and a SAO
Harvard-Smithsonian Predoctoral Fellowship. We thank the referee for
helpful comments.
\end{acknowledgements}
|
\section{Introduction}
It has been argued that in three dimensions hyperbolic closed field lines (X-lines)\cite{1990ApJ...350..672L}
or magnetic nulls\cite{1988JGR....93.8583G,1990ApJ...350..672L} are required for magnetic reconnection. In two dimensions, reconnection has also typically been discussed in terms of nulls (zero guide field) or X-lines (nonzero guide field). See Ref.~\cite{NR:BirnPriestBook} and references therein.
In this strictest definition of reconnection, the argument is this:
in ideal magnetohydrodynamics (MHD) these structures can lead to singularities
which are then resolved by non-ideal effects, and in the absence of
such structures, no singularities occur. In the presence of small electrical resistivity $\eta$,
these singularities become current sheets of finite but small thickness dependent on $\eta$.
For our purposes here, we
will describe the non-ideal effects as arising from finite electrical
resistivity $\eta$, i.e.~we consider reconnection in resistive MHD.\footnote{However, there is no fundamental reason why these concepts cannot
apply to reconnection with other non-ideal effects, such as electron
inertia and hyper-resistivity.} In this
strict definition, the question of whether reconnection occurs is
in the context of the limit $\eta\rightarrow0$, or the Lundquist
number $S\rightarrow\infty$.
In a line-tied system of finite length $-L\leq z\leq L$ with
fields having $B_{z}>0$, all field lines enter at
$z=-L$ and exit at $z=L$. Therefore, closed field lines cannot form. (We will
speak only in terms of closed field lines; in order for nulls
to form in this geometry, all components of ${\bf B}$ would have to change sign in the region.)
So in the strict
sense described above, no magnetic reconnection can occur. However,
in Refs.~\cite{delzanno:032904,huang:042102} line-tied modes which are tearing modes
in cylindrical geometry [for equilibria with $B_{\theta 0}(r)$, $B_{z0}(r)$] with $L=\infty$
were studied. In these $L=\infty$ modes, behaving as $e^{im\theta+ikz}$,
there is a mode rational surface consisting of closed field lines
in the equilibrium, and the perturbed fields have hyperbolic
and elliptic closed field lines (islands). The conclusion of these papers
was that the modes still behave as tearing modes (with growth rates
$\gamma$ proportional to a fractional power of $\eta$) in a sufficiently
long line-tied system $L>L_{\rm crit}$. The condition $L>L_{\rm crit}$
is defined as the range over which the tearing width $w_{t}$ of the
mode is greater than the geometric width $w_{g}$, the width of the
mode associated with the line-tying boundary condition [$w_g\propto 1/L$]. For $L<L_{\rm crit}$,
when $w_{t}<w_{g}$, the modes no longer resemble tearing modes; they have
no recognizable tearing layer and $\gamma\propto\eta$, i.e.~they involve
global resistive diffusion rather than reconnection.
These results suggest that in many cases it may make more sense to
define reconnection for fixed but large $S$, rather than in the limit
$S\rightarrow\infty$. This is particularly true for numerical simulations with limited values of the Lundquist number $S$ and for laboratory experiments with relatively low temperatures. Three such experiments suited for reconnection studies with line-tying are RWM\cite{PhysRevLett.96.015004}, RSX\cite{furno:2324}, and LAPD\cite{4989217}. This point of view is consistent with that expressed in
Ref.~\cite{Priest:1995fk,1996A&A...308..643D}, in which the concept of a \emph{quasi-separatrix
layer} (QSL) was introduced for fields without hyperbolic closed field
lines (or nulls) as a resolution of controversies associated with generalized magnetic reconnection\footnote{The controversies involving GMR related to whether localized current had to be from reconnection or could be due to double layers, Pfirsch-Schluter currents, or other sources.}\cite{Hesse:1988rt,Schindler:1988yq}. In fact, the concept of a QSL was used to analyze LAPD in \cite{PhysRevLett.103.105002}. For such magnetic fields (with large but fixed $S$), the geometric aspects of the field lines can act to separate the field lines in a manner qualitatively similar to fields with hyperbolic lines. If this separation occurs in a thin enough region, the physics is basically identical to the behavior in the presence of a hyperbolic line.
However, note that while the condition $w_g <w_t$ may be satisfied for such laboratory systems and for simulations with relatively small $S$, for solar coronal or astrophysical applications this requirement may hold only for unrealistically long plasmas.
In Refs.~\cite{Priest:1995fk,1996A&A...308..643D,ESASpec.Publ.448:7151999,Titov20021087,Titov:2002fj,0004-637X-660-1-863,0004-637X-693-1-1029}, it was suggested that the most effective measure of a QSL is the \emph{squashing factor}, which measures the squeezing and stretching of field lines.
The method of slip-squashing factors\cite{0004-637X-693-1-1029} also takes resistivity (or other non-ideal effects) into account. It involves computing the squashing factor plus taking into account the field line slippage due to non-ideal effects. For fields which are determined by boundary motions, this slippage is measured by comparing the initial field line mapping with the field line mapping at a later time.
In Sec.~\ref{sec:diagnostics} of this paper we discuss the squashing factor $Q$ and
a second diagnostic, related to the potential $\phi_i$ used in Refs.~\cite{1990ApJ...350..672L,1991ApJ...366..577L}.
The latter function is the scalar potential required to give $\mathbf{E}\cdot\mathbf{B}=0$,
a consequence of Ohm's law in ideal MHD. We first compute $Q$ and
$\phi_i$ for two examples of magnetic fields, namely a doublet-like
field and a field given by an equilibrium $B_{\theta0}(r)$, $B_{z0}(r)$
plus a single tearing mode, both in a finite region $-L<z<L$. We also compute $\phi_r$, the scalar potential required to satisfy ${\bf E}\cdot \mathbf{B}=\eta {\bf j}\cdot \mathbf{B}$ in resistive MHD. In Sec.~3 we review the treatment of
line tied modes in the \emph{two-mode approximation}\cite{evstatiev:072902}, which applies for long plasmas. In Sec.~4 we study QSLs in
a model representing growing tearing modes with line-tied boundary
conditions in the two-mode approximation. We show the squashing factor
$Q$ as well as the potentials $\phi_i$ and $\phi_r$ for modes of various amplitudes, in the two cases $w_t>w_g$ and $w_t < w_g$.
In Sec.~5 we summarize and discuss our results. In the Appendix we
discuss other mathematical issues related to the squashing factor.
\section{Reconnection diagnostics}\label{sec:diagnostics}
In this section we describe two diagnostics that can be used to identify reconnection and QSLs in systems with finite resistivity. The first diagnostic is the {\em squashing factor}, defined in Refs.~\cite{ESASpec.Publ.448:7151999,Titov20021087,Titov:2002fj,0004-637X-660-1-863}, and the second is the potential difference $\Delta\phi$ which is required to satisfy either the ideal MHD relation ${\bf E}\cdot{\bf B}=0$ (see Ref.~\cite{1990ApJ...350..672L}) or the resistive relation ${\bf E}\cdot{\bf B}=\eta {\bf j}\cdot {\bf B}$. Each of these diagnostics require computing the magnetic fields lines which run the length of the system, from $z=-L$ to $z=L$.
\subsection{Squashing Factor}
The first diagnostic we consider is the squashing factor.
Integrating field lines from one end of the system to the other defines a coordinate mapping $M:{\bf x}\to {\bf X}$, where $\bf x$ is the starting point of the line in the $z=-L$ plane, and $\bf X$ is the ending point in the $z=L$ plane. (In the more general case, we map from the set with normal field $B_n<0$ to the set with $B_n>0$.) Geometrically, stretching and squashing of flux tubes indicates a potential for reconnection, and the Jacobian $J$ of the mapping $M$ gives us information about these processes. The overall expansion of flux tubes is not associated with reconnection, but rather is related to variation of the guide field strength $B_z$. Specifically, flux conservation implies $B_z (x_1,x_2,-L)dx_1 dx_2=B_z (X_1,X_2,L)dX_1 dX_2$, so that the Jacobian determinant $D\equiv \det(J)$ satisfies $D=B_z (x_1,x_2,-L)/B_z (X_1,X_2,L)$. The mapping $M$ takes a flux tube with a circular cross-section and ``squashes'' it so that its cross-section is elliptical. In order to measure the stretching and squashing while compensating for the expansion of flux tubes if $B_z \neq {\rm const.}$, a natural quantity to consider\cite{Priest:1995fk,1996A&A...308..643D,ESASpec.Publ.448:7151999,Titov20021087,Titov:2002fj,0004-637X-660-1-863,0004-637X-693-1-1029} is the aspect ratio of the elliptical cross section of the flux tube at $z=L$.\footnote{In the Appendix, we discuss the possibility of transforming $(x_1,x_2)$ and $(X_1,X_2)$ to canonical variables, for which $D=1$. }
\begin{figure}[htbp]
\begin{center}
\commentfig{
\includegraphics[width=5cm]{figure1.pdf}
}
\caption{Squashing of flux tube cross-section by the mapping $M$. The major and minor axis of the ellipse have lengths $\rho_{max}=\max(|\delta{\bf X}|)$ and $\rho_{min}=\min(|\delta{\bf X}|)$ (where $\rho_i$ are the singular values of the matrix $J$), and so the aspect ratio of the ellipse is $R={\rho_{max}}/{\rho_{min}}$. This ratio removes the effect of variations in $B_z$, which would only give an overall expansion or contraction of the flux tube.\label{fig:squash}}
\end{center}
\end{figure}
Consider two field lines whose initial points are separated by the tangent vector $\delta{\bf x}$. Their endpoints will be separated by $\delta{\bf X}$, and the length of $\delta{\bf X}$ compared to the length of $\delta{\bf x}$ will tell us about the stretching of a flux tube containing the two field lines. The lengths of these two vectors can be directly compared since the Jacobian matrix $J$ ($J_{ij}=\partial X_i / \partial x_j$) maps tangent vectors, i.e.~$\delta\mathbf{X}=J\delta\mathbf{x}$. Thus, their lengths are related by
\begin{eqnarray}\label{eq:lengths}
(\delta\mathbf{X},\delta\mathbf{X})=(J\delta\mathbf{x},J\delta\mathbf{x})=(\delta\mathbf{x},J^{T}J\delta\mathbf{x}).
\end{eqnarray}
The symmetric positive definite matrix $J^{T}J$ is the covariant metric tensor on the surface at
$z=-L$ derived from a Euclidean metric tensor on the surface at $z=L$, and its eigenvalues $\lambda_\pm$, obtained from the Rayleigh quotient $(\delta\mathbf{x},J^{T}J\delta\mathbf{x}) /(\delta\mathbf{x},\delta\mathbf{x})$, tell us how the shape of a flux tube changes.\footnote{The quantity ${\rm Tr}(J^TJ)$ was described as the norm of the displacement gradient tensor in \cite{Priest:1995fk}. It can also be written as $\vert \nabla X_1 \vert^2 + \vert \nabla X_2 \vert^2$ and is actually the square of the Frobenius norm of $J$.}
In fact, the ratio of the singular values\cite{NR:svd_evals} $\rho_\pm = \sqrt{\lambda_\pm}$ equals the aspect ratio of the ellipse. See Fig.~\ref{fig:squash}.
This ratio is easily shown to obey
\begin{eqnarray}\label{eq:squashing}
R=\frac{\rho_{max}}{\rho_{min}} = \frac{Q}{2} + \sqrt{ \left(\frac{Q}{2}\right)^2 - 1},\quad{\rm where}\quad Q \equiv T/D = {\rm Tr}(J^T J) / |\det(J)|.
\end{eqnarray}
This quantity $Q$, which also equals $R+1/R$, is called the {\em asymptotic squashing factor}\cite{0004-637X-660-1-863}, or just {\em squashing factor}.
In regions where $Q$ is large, $R=Q-1/Q+\cdots \approx Q$. It was argued in Refs.~\cite{Priest:1995fk,1996A&A...308..643D,ESASpec.Publ.448:7151999,Titov20021087,Titov:2002fj,0004-637X-660-1-863,0004-637X-693-1-1029} that these regions, where the stretching and squashing are large, are candidates for the occurrence of reconnection. Such regions of large $Q$ are called {\em quasi-separatrix layers}. We are interested in only these regions and, following Refs.~\cite{Priest:1995fk,1996A&A...308..643D,ESASpec.Publ.448:7151999,Titov20021087,Titov:2002fj,0004-637X-660-1-863,0004-637X-693-1-1029}, we will thus focus on $Q$ rather than $R$. In the next section, we give an example which illustrates how $Q$ becomes larger and more concentrated for larger $L$.
\subsection{Squashing Factor Example: Doublet Field}\label{sec:Q_example}
Consider the vector potential with $A_{x_1} = -B_0 x_2$, $A_{x_2} = 0$, and $A_z = \frac{x_2^2}{2} -\frac{x_1^2}{2} + \frac{x_1^4}{4}$. This defines the magnetic field
\begin{eqnarray}\label{eq:doublet}
B_{x_1} = {x_2}, \quad
B_{x_2} = {x_1} - x_1^3, \quad
B_z = B_0.
\end{eqnarray}
The equations for field lines are then
\begin{eqnarray}
{d{x_1}}/{dz} = {x_2}/B_0, \quad
{d{x_2}}/{dz} = ({x_1}-x_1^3)/B_0.
\end{eqnarray}
These equations are in canonical form, with Hamiltonian $H=A_z(x_1,x_2)/B_0$ and $D=1$ (since $B_z=B_0={\rm const.}$). The contours of $A_z$ for this field are shown in Fig.~\ref{fig:db_Q}a. Integrating the field lines from $z=-L$ to $z=L$ with a symplectic integrator
determines an exactly area preserving mapping $(x_1,x_2,-L) \to (X_1,X_2,L)$. The Jacobian matrix of this mapping is estimated numerically by tracing field lines for a grid of initial points ${\bf x}$, and taking the difference of the ending points ${\bf X}$. Since this model is two dimensional, we can change the length $L$ without modifying the model, as we do in the 3D models in Sec.~\ref{sec:line-tied}. The squashing factor $Q$ as a function of $(x_1,x_2)$ is shown in Fig.~\ref{fig:db_Q} for lengths $L=1$, $2$, and $2.5$. Note that for $L=2$, $Q$ is peaked in an elliptically shaped region near the part of the stable manifold near the X-line. As $L$ increases, $Q$ becomes much more peaked and concentrated near the stable manifold and further along the stable manifold. The squashing factor $Q$ is shown in Fig.~\ref{fig:db_Q}c for $L=2.5$ as a function of the variables at $z=L$, namely $(X_1,X_2)$. The large values of $Q$ in this figure are concentrated on the {\em unstable} manifold because this same quantity $Q(X_1,X_2)$ could also have been computed by initializing at $(X_1,X_2,L)$ at $z=L$ and integrating backwards to $(x_1,x_2,-L)$.
In Fig.~\ref{fig:db_Q_3D} we show a three dimensional plot of the contour where $Q=1000$ for $L=2$. The concentration along the stable manifold for $z=-L$ and along the unstable manifold for $z=L$ is evident. This surface has the form of a hyperbolic flux tube, which arises in models of reconnection in solar physics, and the study of QSLs.\cite{Titov:2002fj,APD:currentsandHFT,0004-637X-582-2-1172}
\begin{figure}[htbp]
\begin{center}
\commentfig{
\includegraphics[]{figure2a.pdf}
\includegraphics[]{figure2b.pdf}
\includegraphics[]{figure2c.pdf}
}
\caption{Squashing factor $Q(x_1,x_2)$ for the model of Eq.~(\ref{eq:doublet}), with $B_0=1$ and (a)~$L=1$, (b)~$L=2$. In (c), the squashing factor for $L=2.5$ is shown plotted as a function of ending point $\bf X$.
In (a), contours of $A_z$ are shown, with arrows indicating the stable and unstable manifolds of the X-line at the origin. In this 2D case, the stable and unstable manifolds coincide to form a separatrix.
The quasi-separatrix layer traces out the {\em stable} manifold when plotted as a function of initial point $\bf x$ [(a) and (b)] and the {\em unstable} manifold when plotted as a function of the ending point $\bf X$ in (c).
\label{fig:db_Q}}
\end{center}
\end{figure}
\mycomment
\begin{figure}[htbp]
\begin{center}
\commentfig{
\includegraphics[height=4cm]{Q5.pdf}
\includegraphics[height=4cm]{Q5_rev.pdf}
}
\caption{Squashing factor $Q$ for for the model of Eq.~(\ref{eq:doublet}), with $B_0=1$ and $L=2.5$, plotted as a function of (a) the initial point $\bf x$ of the field line and (b) the ending point $\bf X$ of the field line. The quasi-separatrix layer traces out the {\em stable} manifold and the {\em unstable} manifold of the X-line at the origin in (a) and (b), respectively.}
\label{fig:db_Q_rev}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\commentfig{
\includegraphics[height=4.7cm]{figure3.pdf}
}
\caption{3-dimensional visualization of the quasi-separatrix layer for the model of Eq.~(\ref{eq:doublet}). The surface is formed by the field lines which have $Q= 1000$ for $L=2$.
\label{fig:db_Q_3D}}
\end{center}
\end{figure}
\subsection{Scalar Potential Difference $\Delta \phi$}
The second diagnostic we consider in this paper is the potential difference between the two ends of the field lines. We calculate $\Delta \phi=\Delta \phi_i$ from the ideal MHD Ohm's law $\mathbf{E}+\mathbf{v}\times\mathbf{B}=0$ to give zero
parallel electric field $ E_{||}\propto\mathbf{E}\cdot\mathbf{B}$, and calculate $\Delta \phi_r$ similarly from the resistive MHD Ohm's law $\mathbf{E}+\mathbf{v}\times\mathbf{B}=\eta {\bf j}$. (Again, we consider resistive MHD just for concreteness.)
The first approach, for ideal MHD, was used to analyze three dimensional reconnection in
the presence of hyperbolic closed field lines or nulls in Ref.~\cite{1990ApJ...350..672L}.
In this paper, we compute both $\Delta \phi_i$ and $\Delta \phi_r$ by taking the magnetic field $\mathbf{B}(\mathbf{x},t)$ and the inductive electric field $-\partial\mathbf{A}/\partial t$ computed from resistive MHD. We solve
$\mathbf{E}\cdot\mathbf{B}=0$ and $\mathbf{E}\cdot\mathbf{B}=\eta {\bf j}\cdot {\bf B}$ along field lines, namely
\begin{equation}
\mathbf{B}\cdot\nabla\phi_i=-\mathbf{B}\cdot\frac{\partial\mathbf{A}}{\partial t};\,\,\,\,\,\,\,\,\,\, \mathbf{B}\cdot\nabla\phi_r=-\mathbf{B}\cdot\frac{\partial\mathbf{A}}{\partial t} -\eta {\bf j}\cdot {\bf B}, \label{eq:BdotGradPhi}
\end{equation}
for the scalar potentials $\phi_i$, $\phi_r$ and, for both, evaluate the difference between the potential at the two ends of the field line at $z=-L$ and $z=L$.
The potential $\phi_r$ is exactly the single-valued scalar potential related to the fields obtained in resistive MHD (in a specific gauge) and represents the electrostatic potential due to small (quasineutral) charge variations required to balance the inductive electric fields of the mode locally along the field lines, in the presence of the resistive term $\eta {\bf j}\cdot {\bf B}$.
In computing $\phi_i$, we analyze the fields generated by resistive MHD in the context of \emph{ideal} MHD, in order to determine when $\Delta \phi_r \approx \Delta \phi_i$. In most regions of the plasma, the potential $\phi_i$ is also a single-valued function, again due to movement of electrons along ${\bf B}$ to cancel the inductive electric field. In regions where $\eta {\bf j}\cdot {\bf B}$ is small, $\Delta \phi_r \approx \Delta \phi_i$. On the other hand, if there is a closed field line in the system
and if $\partial\mathbf{A}/\partial t$ integrates to a nonzero value along this
line, then mulit-valuedness of $\phi_i$ will occur at points on the line and singular behavior will occur on
points that go to the line (the {\em stable manifold} of the hyperbolic X-line), because the associated flux change through any
surface bounded by the line cannot be cancelled by a single-valued electrostatic potential.\footnote{Note that a gauge change $\mathbf{A}\rightarrow\mathbf{A}+\nabla\chi,\,\,\phi\rightarrow\phi-\partial\chi/\partial t$
with $\chi$ continuous does not affect the singular nature of $\phi_i$
as found from Eq.~(\ref{eq:BdotGradPhi}). Similarly, in integrating
Eq.~(\ref{eq:BdotGradPhi}) from $z=-L$ to $z=L$, a smooth initial value $\phi(r,\theta,z=-L)$ is irrelevant.}
In the line-tying geometry of this paper, we do not expect to encounter true singularities
in $\phi_i$. However, we do find that $\phi_i$, like the squashing
factor $Q$, can be large and strongly localized, which is indicative
of reconnection-like behavior, e.g.~current sheets. This kind of behavior, with peaked but nonsingular values of $\phi_i$, was seen and discussed in \cite{1991ApJ...366..577L}.
It is useful to compute both $\phi_i$ and $\phi_r$, since the difference $\Delta\phi_i - \Delta\phi_r$ is simply the integral of $\eta {\bf j}\cdot {\bf B} \propto \eta j_{||}=E_{||}$ along the field line. Parallel currents are associated with reconnection effects, and this method allows us to not only compute the parallel currents (through $\Delta\phi_i - \Delta\phi_r$), but it also gives us quantities ($\Delta \phi_i$ and $\Delta \phi_r$) with which to compare the parallel currents. If $\Delta \phi_r\approx\Delta \phi_i$, then the effect of the parallel currents are small compared to the inductive electric fields. In regions where $\Delta\phi_i - \Delta\phi_r$ is large compared to $\Delta\phi_i$, however, the currents -- and thus reconnection effects -- are important.
The difference $\Delta\phi_i -\Delta\phi_r$ is equal to the quasi potential $\Xi$ of Ref.~\cite{0004-637X-631-2-1227}. However, we find that comparing the two separate potentials $\Delta\phi_i$ and $\Delta\phi_r$ is more informative than considering only $\Xi = \Delta\phi_i -\Delta\phi_r$.
As diagnostics for reconnection, there is one fundamental difference between the
squashing factor $Q$ and the potentials $\phi_i$, $\phi_r$: The former involves
only the structure of the magnetic field at a specific time, whereas
the potentials $\phi_i$, $\phi_r$ obtained from Eq.~(\ref{eq:BdotGradPhi}) involve
the structure of the magnetic field \emph{and} the inductive electric
field $-\partial\mathbf{A}/\partial t$. That is, an area of large
and localized $Q$ only indicates the possibility of having reconnection,
whereas a large and localized value of $\Delta\phi_i$ shows flux changes
{\em and} field line topology indicative of reconnection.
\subsection{Ideal MHD Scalar Potential Example: Single tearing Mode}
We use the compressible zero-$\beta$ visco-resistive MHD equations:
\begin{eqnarray}
\partial_t \rho &=& - \nabla\cdot ( \rho{\bf v})\\
\rho \partial_t {\bf v} +\rho {\bf v} \cdot \nabla {\bf v} &=& {\bf j} \times {\bf B} + \nu\rho\nabla^2{\bf v}\\
{\bf E} + {\bf v}\times {\bf B} &=& \eta {\bf j} \label{eq:ohm}\\
\nabla\times{\bf E}&=&-\partial_t {\bf B}\\
\nabla\times{\bf B} &=& {\bf j}\\
\nabla\cdot{\bf B} &=& 0,
\end{eqnarray}
in a cylinder with $-L\leq z\leq L$ and $0\leq r\leq r_w$, with $r_w=2$.
The equilibrium used was the equilibrium of \cite{delzanno:032904}, for which the fastest growing mode in an infinite cylinder is a tearing mode rather than an ideal MHD mode. This force-free equilibrium is specified by its axial current density
$
j_{z0}(r)={2}/({1+r^{\kappa}})
$
with $\kappa=6$, and $B_{z0}(r)$ determined by force balance with an integration constant $B_{z0}(r=0)=5$, with $\rho_0 =1$ and ${\bf v}_0 = 0$. This equilibrium is ``tokamak-like'' in the sense of having an increasing profile of the field line pitch $\mu(r)\equiv rB_{z0}(r)/B_{\theta 0}(r)$ and $B_{z0} \gg B_{\theta 0}$.
The linearized MHD equations are
\begin{eqnarray}
\partial_t \tilde{\bf v} &=& \left[ \nabla \times \tilde{\bf B} - \lambda(r) \tilde{\bf B} \right] \times {\bf B}_0 + \nu \nabla^2 \tilde{\bf v} \label{eq:mhdlin1}\\
\partial_t \tilde{\bf B} &=& \nabla \times \left( \tilde{\bf v} \times {\bf B}_0 - \eta \nabla \times \tilde{\bf B} \right),\label{eq:mhdlin2}
\end{eqnarray}
with $\tilde{{\bf B}}=\tilde{{\bf B}}(r)e^{im\theta +ikz}$ and $\tilde{{\bf v}}=\tilde{{\bf v}}(r)e^{im\theta +ikz}$. These equations are solved\footnote{As in Ref.~\cite{richardson:112511}, a small amount of divergence cleaning is used.} as an eigenvalue problem for the growth rate $\gamma$, with $m=1$, and $k=-0.13507$, giving a mode rational surface (where ${\bf k}\cdot{\bf B}_0=0$) at $r_s=1.2576$. For computing the field line diagnostics, we have superimposed the equilibrium fields with the perturbed fields multiplied by a small amplitude $a$.\footnote{For even relatively small values of the amplitude $a$, nonlinear effects could be important, invalidating the assumption that we can superimpose the equilibrium and perturbed fields obtained by linear theory. For our purposes in this paper, we will keep $a$ small so that such errors are small.} The system length $L$ was taken to be $L=20\pi/k$. This length and the value of $k$ were chosen for comparison with the results from the two-mode approximation in the next section.
We have computed $\Delta\tilde{\phi}_i$ by integrating Eq.~(\ref{eq:BdotGradPhi}) along field lines of the field ${\bf B}_0 + a_1\tilde{\bf B}$ from $z=-L$ to $z=L$, using $a=1\times 10^{-3}$ (see Fig.~\ref{fig:1mode_phi}). This quantity is concentrated around the mode rational surface at $r_s=1.2576$, near both the X-line ($\theta\approx\pi/2$), and the elliptic (O) line ($\theta\approx 3\pi/2$). On both the hyperbolic and the elliptic closed field lines, $|\Delta \tilde{ \phi}_i|$ is large because the perturbation is constant along these lines. For larger amplitude $a=0.01$, the island opens up, and $|\Delta \tilde{ \phi}_i|$ becomes localized around the stable manifold of the X-line (see Figs.~\ref{fig:1mode_phi_cont}b,c).
We have also shown the surfaces of the helical flux $\chi(r,\theta+kz)=\chi_0(r)+\tilde{\chi}(r,\theta+kz)=mA_z-krA_{\theta}$ at $z=-L$. Surface of section points for periodicity $L_p=2\pi/k$ lie on surfaces of constant $\chi$.
\begin{figure}[htbp]
\begin{center}
\commentfig{
\includegraphics[]{figure4a.pdf}
\includegraphics[]{figure4b.pdf}
\includegraphics[]{figure4c.pdf}
}
\caption{(a) $\Delta \phi_i(r,\theta)$, and (b) $\Delta \phi_r(r,\theta)$, computed by integrating Eq.~(\ref{eq:BdotGradPhi}) from $z=-L $ to $z=L$. Mode amplitude $a=1\times 10^{-3}$. (c) Difference between $\Delta \phi_i$ and $\Delta \phi_r$, which is the cumulative parallel current integrated along the field lines. Contours are of the helical flux, which for one mode is identical to the surface of section. The white arrows indicate the direction of the stable manifold of the X-line near $(r,\theta)=(1.25,\pi/2)$.
\label{fig:1mode_phi}}
\label{default}
\end{center}
\end{figure}
The potential $\Delta \tilde{ \phi}_i$ can also be computed analytically. For infinitesimal $\tilde{\phi}_i$ and $\tilde{\mathbf{A}}$ we use $d\theta/dz=1/\mu(r)$ for the equilibrium fields to solve $\mathbf{B}_0\cdot\nabla\tilde{\phi}_i=-\gamma{\bf B}_0\cdot\tilde{\mathbf{A}}$. With the normalization chosen so that $\tilde{A}_\theta(r)$ and $\tilde{A}_r(r)$ are pure imaginary, the potential difference is\footnote{The $r$ dependence of $\tilde{\bf A}$, ${\bf B}_{0}$, and $\mu$ is suppressed here for clarity.}:
\begin{eqnarray}
\Delta\tilde{\phi}_i(r,\theta_0)&=&\frac{\gamma{\mathbf B}_0\cdot \tilde{{\mathbf A}}}{B_{z0}}\int_{-L}^L dz ~\sin(m\theta(z)+kz)
\label{eq:sinccalc1}\\
&=&-\frac{\gamma{\mathbf B}_0\cdot \tilde{{\mathbf A}}}{B_{z0}} \left[\frac{\cos(m\theta_0+m(z+L)/\mu+kz)}{k+m/\mu}\right]_{-L}^L \\
&=& 2L\frac{\gamma{\mathbf B}_0\cdot \tilde{{\mathbf A}}}{B_{z0}}\sin(m(\theta_0+L/\mu))\frac{\sin\left((k+m/\mu)L\right)}{\left(k+m/\mu\right)L}\,\label{eq:sinccalc2}
\end{eqnarray}
where $\theta_0=\theta$ at $z=-L$. The quantity $\left(m/\mu(r)+k\right)={\bf k}\cdot {\bf B}_0 /B_{z0}$, so $\Delta\tilde\phi_i$ is proportional to a sinc function with a peak at the mode rational surface $r=r_s$. The width of $\Delta\tilde{\phi}_i$ in $r$ is
\begin{equation}
w_m=\left|\frac{\pi \mu(r_s)^2}{mL\mu '(r_s)}\right|.\label{eq:SingleModeWidth}
\end{equation}
Note that the cosine factor in $\Delta \tilde\phi_i$ depends on both $\theta_0$ and $r$ [through $\mu(r)$], so its zero contours will be tilted in the $(r,\theta_0)$ plane.
For this calculation with infinitesimal mode amplitude, there is no information about the hyperbolic and elliptic lines, and in fact the island is assumed to have zero width. The width $w_m$ represents the decorrelation due to the shear in the magnetic field at the mode rational surface, which is proportional to $\mu '(r_s)$. This width is consistent with the observed radial width of $\Delta{\phi}_i$ in Fig.~\ref{fig:1mode_phi}a. Note that in Fig.~\ref{fig:1mode_phi}a, the similar prefactor $({\bf B}\cdot \tilde{\bf A})/B_{z}$ goes rapidly to zero outside the mode rational surface, so the amplitude of $\Delta\phi_i$ falls off more rapidly than might be expected from the sinc function alone. A plot of $\Delta\tilde\phi_i$ as given by Eq.~(\ref{eq:sinccalc2}) [including the prefactor $({{\mathbf B}_0\cdot \tilde{\mathbf A}})/{B_{z0}}$] is nearly indistinguishable from the plot in Fig.~\ref{fig:1mode_phi}a, which was computed by integrating Eq.~(\ref{eq:BdotGradPhi}a) numerically along field lines of ${\bf B}_0 + a\tilde{\bf B}$.
\begin{figure}[htbp]
\begin{center}
\commentfig{
\includegraphics[]{figure5a.pdf}
\includegraphics[]{figure5b.pdf}
\includegraphics[]{figure5c.pdf}
}
\caption{(a) Ideal and resistive forms of $\Delta \phi$ as in Fig.~\ref{fig:1mode_phi}a,b, but along the line $\theta=\pi/2$. Note the large peak in $\Delta \phi_i$ near the mode rational surface.
(b) $\Delta \phi_i(r,\theta)$ and (c) $\Delta \phi_i(r,\theta)-\Delta \phi_r(r,\theta)$, as in Figs.~\ref{fig:1mode_phi}a and \ref{fig:1mode_phi}c, but with mode amplitude $a=0.01$. Notice how the structure becomes concentrated near the stable manifold of the X-line.
}
\label{fig:1mode_phi_cont}
\end{center}
\end{figure}
Fig.~\ref{fig:1mode_phi}b shows $\Delta \phi_r$, computed from Eq.~(\ref{eq:BdotGradPhi}b). The oscillations below $r=1.2$ are the same as in Fig.~\ref{fig:1mode_phi}a, but the larger peaks near $r=r_s$ have disappeared, because $-\gamma {\bf B} \cdot \tilde{\bf A}$ balances $\eta \tilde{\bf j} \cdot {\bf B}$ near $r=r_s$. The difference $\Delta\phi_i - \Delta\phi_r$ is plotted in Fig.~\ref{fig:1mode_phi}c, showing only the voltage drop due to the resistive term near $r=r_s$.
In Fig.~\ref{fig:1mode_phi_cont}a we show $\Delta \phi_i$ and $\Delta \phi_r$ as a function of $r$ for $\theta=\pi/2$. The large peak in near $r=r_s$ is evident in $\Delta \phi_i$ but not in $\Delta \phi_r$. We show $\Delta \phi_i$ in Fig.~\ref{fig:1mode_phi_cont}b, as in Fig.~\ref{fig:1mode_phi}a but with mode amplitude $a=0.01$. In Fig.~\ref{fig:1mode_phi_cont}c we show the difference $\Delta\phi_i - \Delta\phi_r$ as in Fig.~\ref{fig:1mode_phi}c. We see that the minimum and maximum values of $\Delta \phi_i$, at the X-line and the O-line, respectively, are equal and opposite, but the values near the X-line are more concentrated along the stable manifold of the X-line. In Fig.~\ref{fig:case2_1mode_phi}, we show a plot as in Fig.~\ref{fig:1mode_phi_cont}a for $L=252$. There is very little difference between $\Delta \phi_i$ and $\Delta \phi_r$ for this case. As we will discuss in the next section, this is consistent with $w_m$ being larger than the tearing mode width $w_t$.
\begin{figure}[htbp]
\begin{center}
\commentfig{
\includegraphics[]{figure6_new.pdf}
}
\caption{Ideal and resistive forms of $\Delta \phi$ as in Fig.~\ref{fig:1mode_phi_cont}a, but with parameters $\nu$, $\eta$, $k_1$, and $L$ as in Case II (see Table \ref{tab:params}). While the difference between $\Delta \phi_i$ and $\Delta \phi_r$ is largest near the mode rational surface, it is still small compared to $\Delta \phi_i$. This is an indication that mode width $w_m$ (Eq.~\ref{eq:SingleModeWidth}) is larger than the tearing width $w_t$ for this mode, and thus we expect that the effects of the line tying will be important.}
\label{fig:case2_1mode_phi}
\end{center}
\end{figure}
\section{Line-tied tearing modes}\label{sec:line-tied}
For analyzing modes with line-tying, we use the method of Ref.~\cite{evstatiev:072902}. We expand the perturbed velocity and magnetic field components $\xi$ as sums over the spectrum:
\begin{equation}\label{eq:spectrum}
\xi(r,\theta,z,t) = e^{im\theta + \gamma t}
\sum_k a_k \xi_k (r) e^{ikz},
\end{equation}
These \emph{radial eigenfunctions} satisfy the linearized equations (\ref{eq:mhdlin1}) and (\ref{eq:mhdlin2}) and the wavenumbers $k$ and the expansion coefficients $a_k$ are chosen so that the solution satisfies the line-tied boundary conditions at $z=\pm L$. If $L\gg 1$, then keeping only two terms $k_1$, $k_2$ in the sum gives a good approximation to the line-tied solution. This is the \emph{two mode approximation}. It is valid for large $L$ because $k_2-k_1$ is small for large $L$ and therefore $\xi_{k_1}(r) \simeq \xi_{k_2}(r)$. See Ref.~\cite{evstatiev:072902}. The calculation proceeds as follows.
The line-tied boundary conditions at $z=\pm L$ correspond to having the tangential components of $\tilde{{\bf E}}$ and the normal component of $\tilde{{\bf B}}$ be zero at $z=\pm L$. (We describe these modes as line-tied, in spite of the fact that the tangential velocity at $z=\pm L$ is nonzero due to finite resistivity.) We proceed by treating $\gamma$ as fixed for now, and choose $k_1$ and $k_2$ such that in the infinite cylinder results $\gamma(k_1)=\gamma(k_2)$. Then the line-tying condition leads to $a_1 e^{ik_1 z}+a_2 e^{ik_2 z}=0$ at $z=\pm L$.
This implies that
$k_2(\gamma)-k_1(\gamma)=l \pi /L$, giving a dispersion relation for $\gamma$. Here $l$ is an integer, but we can take $l=1$ without loss of generality, and for $l=1$ we have $a_1=a_2\equiv a$. The length of our line-tied system is then $2L$, where $L = \pi/(k_2-k_1)$.
For this line-tied mode satisfying the two-mode approximation, we calculate the geometric width $w_g$ in the following manner\cite{delzanno:032904}. We find the two mode rational surfaces (${\bf k}\cdot {\bf B}_0=0$) corresponding to $k_1$ and $k_2$ by $m/\mu(r_1)+k_1=m/\mu(r_2)+k_2=0$. The geometric width is the distance between these mode rational surfaces:
\begin{equation}
w_g\equiv r_2-r_1\simeq\left|\frac{(k_2-k_1)\mu(r_1)^2}{m\mu'(r_1)}\right|.\label{eq:geomwidth}
\end{equation}
In Ref.~\cite{delzanno:032904} it was shown that this definition is still qualitatively correct even for smaller lengths $L$, for which the two-mode approximation is not valid. It is interesting to note that the single mode width $w_m$ in Eq.~(\ref{eq:SingleModeWidth}) equals $w_g$ if the line-tying relation $k_2-k_1=\pi/L$ is substituted. We shall return to this point in the next section.
For $a$ large enough, the islands corresponding to the two modes will overlap and magnetic field line chaos will occur.\footnote{As in the last section, we will allow superposition of states with relatively large $a$ even when nonlinear effects are important. For sure, having field line chaos is such a nonlinear effect.} In order to be able to construct a surface of section plot with fields having a common period in $z$, we do the following: We first assume that $k_2/k_1= n_2/n_1$ is a rational number. In this case the two modes then have a common length $L_p$ over which they are periodic,
$
L_p = n_1 L_1 = n_2 L_2,
$
where $L_i = 2\pi/|k_i|$.
Also, it can be seen from the line-tying relation above that
$
L = \frac{1}{2}L_p/|n_2-n_1|.
$
So, if we take $|n_2-n_1|=1$, then $L_p = 2L$. From this we conclude that the fields are actually periodic over the length of the line-tied system $z=-L \to L$. For this special case, the map from $z=-L$ to $z=L$ is a surface of section, and we can plot its return map to assess chaos in the periodic system.
Given the perturbed fields $\tilde{\bf B}$ from this two-mode calculation, we can find $\tilde{\bf A}$ in the gauge where $\tilde{A}_r =0$. For the $k_1$ component, we obtain
\begin{eqnarray}
\tilde A_{z1}(r) &=& -\int^r_0 \tilde B_{\theta1}(r') \,dr',\\
\tilde A_{\theta1}(r) &=& \frac{m}{rk_1}\tilde A_{z1}(r) + \frac{i}{k_1} \tilde B_{r1}(r),\label{eq:Ath}
\end{eqnarray}
and similarly for the $k_2$ component.
The perturbed vector potential for the two mode approximation is thus
\begin{eqnarray}
\tilde{\bf A}(r,\theta,z,t) =a e^{\gamma t+im\theta}\left(\tilde{\bf A}_1(r)e^{ik_1 z} + \tilde{\bf A}_2(r)e^{ik_2 z}\right).
\end{eqnarray}
The inductive electric field ${\tilde{\bf E}}_I = -\partial_t {{\bf A}}$ is then:
\begin{eqnarray}
{\tilde{\bf E}}_I = -\gamma a e^{\gamma t+im\theta}\left(\tilde{{\bf A}}_1(r)e^{ik_1 z} + \tilde{{\bf A}}_2(r)e^{ik_2 z}\right).
\end{eqnarray}
This is the inductive field used for computing the scalar potentials $\Delta\phi_i$ and $\Delta\phi_r$ [c.f.~Eq.~(\ref{eq:BdotGradPhi})]. This mode behaves in $z$ as $e^{ik_0z} \cos(\delta k z)$, where $k_0$ is $(k_1+k_2)/2$ and $\delta k$ is $k_2-k_0$. Therefore at $r=r_0\approx(r_1+r_2)/2$ the mode is nearly constant along the field lines for most of the length of the plasma.
\section{Reconnection diagnostics for line tied modes}
Given the two-mode approximation for line-tied modes, we use the perturbed fields $\tilde{\bf B}$ to compute the squashing factor $Q$ and the scalar potentials $\phi_i$ and $\phi_r$ of Eq.~(\ref{eq:BdotGradPhi}). Two cases are compared: first, the plasma parameters are chosen so that the two-mode approximation is fairly accurate, and the tearing width $w_t$ (estimated from the width of the perturbed current near the mode rational surface) is larger than the geometric width $w_g$. The second case has these limits reversed, with $w_t<w_g$; in this case, the two-mode approximation is not as accurate, but can be used as a simple model for understanding qualitative aspects of the line-tied mode.
\begin{table}[htdp]
\caption{Parameters for the two line-tied cases\label{tab:twomode} }
\begin{center}
\begin{tabular}{rccccccccccc}
&$\eta$ & $\nu$ & $n_2/n_1$ & $k_1$ & $k_2$ & $\gamma$
& $w_g$ & $w_t$ & $L_1$ & $L_2$ & $L$ \\
\hline\hline
Case I: & $10^{-4}$ &$10^{-3}$& 19/20 & -0.135 & -0.128 & 0.0233 &0.047 &0.2&46.5&49.0&465 \\
Case II: & $10^{-6}$ &$10^{-5}$& 11/12 & -0.149 &-0.137 & 0.0110 & 0.084&0.05& 42.1&45.9& 252 \\
\end{tabular}
\end{center}
\label{tab:params}
\end{table}%
\subsection{Case I: $w_t>w_g$}
The plasma parameters for this case are $\eta = 1\times10^{-4}$ and $\nu = 1\times10^{-3}$. The two modes have $k_2/k_1 = 19/20$, $L=465$, and the values of $k_i$, $\gamma$,
and $L_i$ are given in Table \ref{tab:twomode}. Two different mode amplitudes are shown for this case. In Fig.~\ref{fig:caseI.1}, the mode amplitude is $a=1\times10^{-3}$. The surface of section shows two sets of islands with very thin secondary resonances between them. The $\Delta\phi_i$ calculation (Fig.~\ref{fig:caseI.1}a) shows structure near the islands for the two modes. The fact noted in the last section, just after Eq.~(\ref{eq:geomwidth}), namely that $w_m = w_g$, implies that the regions of large $|\Delta \phi_i|$ from the two island chains overlap. This is true because $w_g$, as defined in Eq.~(\ref{eq:geomwidth}), is the radial displacement between the island chains. The values of $|\Delta \phi_i|$ are peaked near the secondary resonances with $r=r_0$, between $r=r_1$ and $r=r_2$, and aligned with the stable manifolds of the X-lines of the two island chains. On the other hand, $\Delta\phi_r$ (Fig.~\ref{fig:caseI.1}b) is much smaller in magnitude and not localized. The squashing factor is also localized near the islands (Fig.~\ref{fig:caseI.1}c). Figure \ref{fig:caseI.2} shows results for the same two modes, but with amplitude $a=5\times10^{-3}$. Note that the surface of section now shows a chaotic region linking the islands, and $\Delta\phi_i$ (Fig.~\ref{fig:caseI.2}a) and squashing factor $Q$ (Fig.~\ref{fig:caseI.2}c) are both now more localized in the chaotic region and aligned with the stable manifolds of the X-lines. The resistive potential $\Delta \phi_r$ is smaller in magnitude and shows little variation in the island region (Fig.~\ref{fig:caseI.2}b).
\begin{figure}[htbp]
\begin{center}
\commentfig{
\includegraphics[]{figure6a.pdf}
\includegraphics[]{figure6b.pdf}
\includegraphics[]{figure6c.pdf}
}
\caption{Reconnection diagnostics for the two mode approximation with mode amplitude $a=1\times 10^{-3}$. All quantities computed with field lines mapped from $z=-L \to L$.
(a) $\Delta \phi_i(r,\theta)$, (b) $\Delta \phi_r(r,\theta)$, and (c) squashing factor $Q(r,\theta)$. Black dots show the surface of section map, which shows two sets of islands, and a secondary resonance. Stable manifolds of the X-lines are indicated by the white arrows.
\label{fig:caseI.1}
}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\commentfig{
\includegraphics[]{figure7a.pdf}
\includegraphics[]{figure7b.pdf}
\includegraphics[]{figure7c.pdf}
}
\caption{Reconnection diagnostics for the two mode approximation with mode amplitude $a=5\times 10^{-3}$. All quantities computed with field lines mapped from $z=-L \to L$.
(a) $\Delta \phi_i(r,\theta)$, (b) $\Delta \phi_r(r,\theta)$, and (c) squashing factor $Q(r,\theta)$. Black dots show the surface of section map, showing field line chaos due to the overlap of the two sets of islands. The two X-lines are still near $\theta =\pi /2,r=r_1$ and $\theta=3 \pi /2, r=r_2$, respectively.
\label{fig:caseI.2}
}
\end{center}
\end{figure}
\subsection{Case II: $w_t<w_g$}
The plasma parameters for this case are $\eta = 1\times10^{-6}$ and $\nu = 1\times10^{-5}$. The two modes have $k_2/k_1 = 11/12$, $L=252$, and the values of $k_i$, $\gamma$, and $L_i$ are given in Table \ref{tab:twomode}.
For this shorter system, $\Delta \phi_i$ and $\Delta \phi_r$ are nearly equal throughout the plasma (Fig.~\ref{fig:caseII}a), meaning that the resistive voltage drop, proportional to $\Delta \phi_i - \Delta \phi_r$, is very small. This is an indication that the mode has lost its tearing character because of the shorter length of the plasma.
This effect is quantified in Ref.~\cite{delzanno:032904}, where it was shown that the scaling of the growth rate is tearing-like (with $\gamma\propto$ a fractional power of $\eta$) when $w_t>w_g$ as in the previous case, but the mode exhibits resistive diffusion (with $\gamma \propto \eta$) whenever $w_t < w_g$.
Even in the region where the current is strong (the tearing layer) the finite length is a more important factor than the influence of the current.
On the other hand, the squashing factor (Fig.~\ref{fig:caseII}b) is quite localized in the region of the two island chains, indicating the stretching character of the X-lines of the islands. In this case, the variation of $Q$ in this region suggests the potential for reconnection, whereas the potentials $\Delta \phi_i$ and $\Delta \phi_r$ show that for this case there is essentially no reconnection going on. In contrast with $Q$, the slip-squashing factor of Ref.~\cite{0004-637X-693-1-1029} measures the actual field line slippage due to non-ideal effects as well as stretching.
\begin{figure}[htbp]
\begin{center}
\commentfig{
\includegraphics[]{figure8a.pdf}
\hspace{2em}
\includegraphics[]{figure8b.pdf}
}
\caption{
Reconnection diagnostics for the case $w_t<w_g$. (a) $\Delta\phi_i$, $\Delta\phi_r$, and a measure of the parallel current integrated along field lines, given by $\Delta\phi_i-\Delta\phi_r$ ($a=10^{-3}$, $\theta=\pi/2$). In this case, the finite length of the system is more important than the parallel current, as can be seen in (a), since $\Delta\phi_i$ and $\Delta\phi_r$ are nearly the same.
In (b) is shown ${\rm log}_{10}(Q)$ for various mode amplitudes $a$. The peaks in $Q$ are quite evident for large $a$ but not noticeable for small $a$.
The importance of the finite length ($w_t < w_g$) is not obvious from $Q$.
\label{fig:caseII}
}
\end{center}
\end{figure}
\section{Summary and conclusions}
We have constructed a cylindrical model for fields with a growing
mode in the presence of line-tying at $z=\pm L$. The fields in this
model consist of an equilibrium plus perturbed
fields $\tilde{\mathbf{E}}$, $\tilde{\mathbf{B}}$ of an arbitrary
but small amplitude. The perturbed fields are computed using the two-mode approximation. The two infinite
cylinder modes, with distinct axial mode numbers $k_{1}$, $k_{2}$,
are both tearing modes, so the line-tied mode formed from them should
have tearing character\cite{delzanno:032904,huang:042102} (i.e.~effectively
involve reconnection) for large $L$. Also, it is in this limit that
the two-mode approximation is valid. Refs.~\cite{delzanno:032904,huang:042102} showed that for shorter plasmas, the modes no longer have tearing character. We have traced field lines and
computed the squashing factor $Q$ and the electrostatic potentials
$\Delta \phi_i$, $\Delta \phi_r$ obtained from Eq.~(\ref{eq:BdotGradPhi}), where the inductive
electric field is given in terms of the linear modes by $\tilde{\mathbf{E}}_I=-\gamma\tilde{\mathbf{A}}$
and $\gamma$ is the linear growth rate.
This model, with an equilibrium plus linear modes of arbitrary amplitude,
is accurate for very small mode amplitudes but clearly inaccurate
for large amplitudes, for which the nonlinear terms are important.
Quantitative results for large amplitude nonlinear states require
nonlinear resistive MHD simulations, which are outside the scope of the present investigation.
We discussed the point that the squashing factor or squashing ratio depends on only the geometry of the magnetic fields, whereas the electrostatic potentials $\Delta \phi_i$ and $\Delta \phi_r$ involve the magnetic field geometry plus the inductive electric fields. So, large and localized values of $Q$ indicate a magnetic structure with the potential for reconnection; large and localized values of $\Delta\phi_i$ relative to $\Delta\phi_r$ indicate electric and magnetic fields for which reconnection should be occurring. A comparison between our methods using $\Delta\phi_i$ and $\Delta\phi_r$ and the use of the slip-squashing factor of Ref.~\cite{0004-637X-693-1-1029} should be interesting, but is also outside the scope of the present study.
We found very different behavior for $\Delta \phi_i$, $\Delta \phi_r$ in Case I, the long plasma case ($w_g<w_t$) and Case II, the short plasma case ($w_g>w_t$). In Case I, $\Delta\phi_i$ is much more peaked in the tearing layer region than $\Delta \phi_r$, indicative that reconnection plays a large role in the line-tied mode in this range. In Case II, $\Delta \phi_i$ and $\Delta \phi_r$ are basically equal, indicating that the shorter length precludes reconnection. These results are consistent with the results of \cite{delzanno:032904,huang:042102}, where $w_g$ and $w_t$ have the same respective meanings as in this paper. In that earlier work, it was found that the growth rates $\gamma$ scale as the appropriate fractional power of $\eta$ for Case I and scale linearly with $\eta$ for Case II.
Both $Q$
and $\Delta\phi_i$ showed localized but nonsingular structures in the areas
in which the corresponding periodic system has hyperbolic closed field
lines or (for larger amplitude modes) chaotic behavior. On the other hand, the squashing factor shows similar behavior for Cases I and II. In particular, it shows almost no variation near the mode islands for small amplitude modes and localized behavior for larger mode amplitudes, with no clear signature of the different mode structure in Case I and Case II.
\section*{Acknowledgments}
We wish to thank V.~Titov, E.~Zweibel, and V.~Mirnov for valuable suggestions. This research was supported by the DOE Office of Science,
Fusion Energy Sciences and performed under the auspices of
the NNSA of the U.S. DOE by LANL, operated by LANS
LLC under Contract No. DE-AC52-06NA25396.
\section*{Appendix: Issues related to the squashing factor $Q$}
As discussed in Sec.~\ref{sec:diagnostics}, the squashing factor is found in the following
manner. First, let $J$ be the Jacobian matrix for the map $\mathbf{x}\rightarrow\mathbf{X}$ which takes field lines from one surface, where $B_n<0$, to the other surface, where $B_n>0$. These surfaces are ${z=-L}$ and ${z=L}$, respectively, in our example.
For the special case $B_{z0}={\rm const.}$, we have
$D=\det(J)=1$. The squashing factor is then simply the trace of $J^{T}J$, as given by Eq.~(\ref{eq:squashing}) with
$D=1$. For the more general flux preserving case
with $B_{z}$ not constant, the determinant is $D=B_{z}(x_{1},x_{2},-L) /B_{z}(X_{1},X_{2},L) \neq 1 $. It is tempting to change to canonical variables, i.e.~variables ${\bf u}$ and ${\bf U}$ in which the equations of motion are canonical and the map is area
preserving. This is always possible by the Darboux theorem, and it may appear at first blush that the situation is simplified because the determinant in these variables is unity.
However, for the physical problems we consider, there is a natural metric on the two surfaces in the original variables, obtained from the Euclidean metric in 3D, so lengths are written as $(\delta{\bf x},G \delta{\bf x})$. (For our example, the surfaces $z=-L$ and $z=L$ are flat, so this metric is the identity, $G=I$.) For general $G$, the relevant Rayleigh quotient is
\begin{equation}\label{eq:RayQuo-1}
Q_R\equiv\frac{(\delta{\bf x},J^T G J \delta{\bf x})}{(\delta{\bf x},G \delta{\bf x})},
\end{equation}
where $J_{ij}={\partial X_{i}}/{\partial x_{j}}$. The relevant eigenvalues are the eigenvalues of $G^{-1}J^T G J$.
The metric in the new canonical variables inherits the metric $(\delta{\bf u},\tilde{G} \delta{\bf u})=(\delta{\bf x},G \delta{\bf x})$. This implies that $\tilde{G}$ satisfies $M^T \tilde{G} M=G$, where $M$ is the tangent map from $\delta {\bf x}$ to $\delta{\bf u}$ (and from $\delta {\bf X}$ to $\delta{\bf U}$). The new Rayleigh quotient is
\begin{equation}\label{eq:RayQuo-2}
\tilde{Q}_R\equiv\frac{(\delta{\bf u},\tilde{J}^T \tilde{G} \tilde{J} \delta{\bf u})}{(\delta{\bf u},\tilde{G} \delta{\bf u})},
\end{equation}
which equals $Q_R$. Therefore, if we transform to canonical variables and inherit the metric from the (physically motivated) metric on the original surface, we do not gain any simplicity from the relation $D=1$.
This issue is related to the issue that arises for finite time Lyapunov exponents (FTLE), where the interval $-L<z<L$ is replaced by the time interval $0<t<T$, and ${\bf x}\rightarrow {\bf X}(t=T)$. If we change variables but continue to use the Euclidean metric, the FTLE
\begin{equation}
h(T)=\frac{1}{T}\rm{ln} \left[ \frac {(\delta \mathbf{X}(T),\delta \mathbf{X}(T)) }{(\delta \mathbf{x},\delta \mathbf{x})}\right]
\end{equation}
are not coordinate invariant. The infinite time (largest) Lyapunov exponent, defined as $h=\lim_{T\rightarrow \infty} h(T)$,
\emph{is} coordinate invariant. The more natural question to ask is whether there is a natural metric. If so, then the metric changes upon a change of variables ${\bf x}\rightarrow {\bf u}$ and ${\bf X}\rightarrow {\bf U}$ so the FTLE are invariant as above. If there is no natural metric, then the FTLE depend on the metric, i.e.~the Rayleigh quotient $(\delta \mathbf{x},J^T G J \delta \mathbf{x}) / (\delta \mathbf{x},G\delta \mathbf{x})$ depends on the metric $G$. The infinite time Lyapunov exponents, however, are independent of $G$.
A final issue relates to the angle between the magnetic field and the two surfaces. If ${\bf B}$ is far from normal to the surface, a circular flux tube projects to an elongated ellipse on the surface, and this is not related to reconnection. This issue was dealt with in Ref.~\cite{0004-637X-660-1-863}. For the model used in this paper, $B_{z0} \gg B_{\theta0}$ and the magnetic field is almost normal to the surfaces $z=-L$ and $z=L$, so that this issue is not important.
\bibliographystyle{unsrt}
|
\section{Introduction}
Throughout this paper, $(\ensuremath{\mathcal X},\|\cdot\|)$ is a reflexive real Banach
space with topological dual $(\ensuremath{\mathcal X}^*,\|\cdot\|_*)$, and the canonical
bilinear form on $\ensuremath{\mathcal X}\times\ensuremath{\mathcal X}^*$ is denoted by $\pair{\cdot}{\cdot}$.
The distance function to a set $C\subset\ensuremath{\mathcal X}$ is
$d_C\colon x\mapsto\inf_{y\in C}\|x-y\|$, the metric projector onto
$C$ is $P_C\colon x\mapsto\menge{y\in C}{\|x-y\|=d_C(x)}$, and the
polar cone of $C$ is $C^\ominus=\menge{x^*\in\ensuremath{\mathcal X}^*}
{(\forall x\in C)\;\pair{x}{x^*}\leq 0}$.
$\Gamma_0(\ensuremath{\mathcal X})$ is the class of lower semicontinuous convex
functions $\varphi\colon\ensuremath{\mathcal X}\to\ensuremath{\,\left]-\infty,+\infty\right]}$ such that
$\ensuremath{\operatorname{dom}}\varphi=\menge{x\in\ensuremath{\mathcal X}}{\varphi(x)<\ensuremath{{+\infty}}}\neq\ensuremath{{\varnothing}}$.
A classical tool in linear hilbertian analysis is the following
orthogonal decomposition principle.
\begin{proposition}
\label{p:1}
Suppose that $\ensuremath{\mathcal X}$ is a Hilbert space, let $V$ be a closed
vector subspace of $\ensuremath{\mathcal X}$ with orthogonal complement $V^\bot$, and
let $x\in\ensuremath{\mathcal X}$. Then the following hold.
\begin{enumerate}
\item
$\|x\|^2=d_V^2(x)+d_{V^\bot}^2(x)$.
\item
$x=P_Vx+P_{V^\bot}x$.
\item
$\pair{P_Vx}{P_{V^\perp}x}=0$.
\end{enumerate}
\end{proposition}
In 1962, Moreau proposed a nonlinear extension of this decomposition.
\begin{proposition} {\rm \cite{Mor62a}}
\label{p:2}
Suppose that $\ensuremath{\mathcal X}$ is a Hilbert space, let $K$ be a nonempty
closed convex cone in $\ensuremath{\mathcal X}$, and let $x\in\ensuremath{\mathcal X}$. Then the following
hold.
\begin{enumerate}
\item
\label{p:2ii}
$\|x\|^2=d_K^2(x)+d_{K^\ominus}^2(x)$.
\item
\label{p:2i}
$x=P_Kx+P_{K^\ominus}x$.
\item
\label{p:2iii}
$\pair{P_Kx}{P_{K^\ominus}x}=0$.
\end{enumerate}
\end{proposition}
Motivated by problems in unilateral mechanics, Moreau further
extended this result in \cite{Mor62b} (see also \cite{More65}).
To state Moreau's decomposition principle, we require some basic
notions from convex analysis \cite{Livre1,Zali02}. Let
$\varphi$ and $f$ be two functions in $\Gamma_0(\ensuremath{\mathcal X})$.
The conjugate of $\varphi$ is the function
$\varphi^*$ in $\Gamma_0(\ensuremath{\mathcal X}^*)$ defined by
\begin{equation}
\label{e:elnido2011-03-02e}
\varphi^*\colon\ensuremath{\mathcal X}^*\to\ensuremath{\,\left]-\infty,+\infty\right]}\colon x^*\mapsto
\sup_{x\in\ensuremath{\mathcal X}}\big(\pair{x}{x^*}-\varphi(x)\big).
\end{equation}
Moreover, the infimal convolution of $\varphi$ and $f$ is
the function
\begin{equation}
\label{e:elnido2011-03-02f}
\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} f\colon\ensuremath{\mathcal X}\to\ensuremath{\,\left[-\infty,+\infty\right]}\colon
x\mapsto\inf_{y\in\ensuremath{\mathcal X}}\big(\varphi(y)+f(x-y)\big).
\end{equation}
Now suppose that $\ensuremath{\mathcal X}$ is a Hilbert space and set
$q=(1/2)\|\cdot\|^2$. Then, for every $x\in\ensuremath{\mathcal X}$, there exists
a unique point $p\in\ensuremath{\mathcal X}$ such that
$(\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} q)(x)=\varphi(p)+q(x-p)$; this point is denoted
by $p=\ensuremath{{\operatorname{prox}}}_\varphi x$.
The operator $\ensuremath{{\operatorname{prox}}}_\varphi\colon\ensuremath{\mathcal X}\to\ensuremath{\mathcal X}$ thus defined is
called the proximity operator of $\varphi$.
\begin{proposition}{\rm \cite{Mor62b,More65}}
\label{p:3}
Suppose that $\ensuremath{\mathcal X}$ is a Hilbert space, let
$\varphi\in\Gamma_0(\ensuremath{\mathcal X})$, set $q=\|\cdot\|^2/2$, and let
$x\in\ensuremath{\mathcal X}$. Then the following hold.
\begin{enumerate}
\item
$q(x)=(\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} q)(x)+(\varphi^*\ensuremath{\mbox{\footnotesize$\,\square\,$}} q)(x)$.
\item
\label{p:3i}
$x=\ensuremath{{\operatorname{prox}}}_\varphi x+\ensuremath{{\operatorname{prox}}}_{\varphi^*}x$.
\item
$\pair{\ensuremath{{\operatorname{prox}}}_\varphi x}{\ensuremath{{\operatorname{prox}}}_{\varphi^*}x}
=\varphi\big(\ensuremath{{\operatorname{prox}}}_\varphi x\big)+
\varphi^*\big(\ensuremath{{\operatorname{prox}}}_{\varphi^*}x\big)$.
\end{enumerate}
\end{proposition}
Note that, if in Proposition~\ref{p:3} $\varphi$ is the indicator
function of a nonempty closed convex cone $K\subset\ensuremath{\mathcal X}$, i.e.,
$\varphi=\iota_K$ where
\begin{equation}
\label{e:iota}
(\forall x\in\ensuremath{\mathcal X})\quad\iota_K(x)=
\begin{cases}
0,&\text{if}\;\;x\in K;\\
\ensuremath{{+\infty}},&\text{if}\;\;x\notin K,
\end{cases}
\end{equation}
we recover Proposition~\ref{p:2}.
The above hilbertian nonlinear decomposition principles have
found many applications in optimization and in various other
areas of applied mathematics (see for instance
\cite{Brog06,Coll79,Svva10,Siop07,Smms05,Hans84,Jbhu89,Hiri10,%
Luce10,Rock06} and the references therein)
and attempts have been made to extend them to more general Banach
spaces. The main result in this direction is the following
generalization of Proposition~\ref{p:2}\ref{p:2i}\&\ref{p:2iii}
in uniformly convex and uniformly smooth Banach spaces
(see also \cite{Albe05,Huyh11,Scho08,Song04} for alternate proofs
and applications), where $\Pi_C$ denotes the generalized projector
onto a nonempty closed convex subset $C$ of $\ensuremath{\mathcal X}$ \cite{Albe96},
i.e., if $J$ denotes the duality mapping of $\ensuremath{\mathcal X}$,
\begin{equation}
\label{e:gproj}
(\forall x\in\ensuremath{\mathcal X})\quad\Pi_C
x=\underset{y\in C}{\operatorname{argmin}}
\big(\|x\|^2-2\pair{y}{Jx}+\|y\|^2\big).
\end{equation}
\begin{proposition}{\rm \cite{Albe00}}
\label{p:4}
Suppose that $\ensuremath{\mathcal X}$ is uniformly convex and uniformly smooth,
let $J\colon\ensuremath{\mathcal X}\to\ensuremath{\mathcal X}^*$ denote its duality mapping, which is
characterized by
\begin{equation}
\label{e:dualitymap}
(\forall x\in\ensuremath{\mathcal X})\quad\|x\|^2=\pair{x}{Jx}=\|Jx\|^2_*,
\end{equation}
let $K$ be a nonempty closed convex cone in $\ensuremath{\mathcal X}$, and let
$x\in\ensuremath{\mathcal X}$. Then the following hold.
\begin{enumerate}
\item
\label{p:4i}
$x=P_Kx+ J^{-1}\big(\Pi_{K^\ominus}(Jx)\big)$.
\item
\label{p:4iii}
$\pair{P_Kx}{\Pi_{K^\ominus}(Jx)}=0$.
\end{enumerate}
\end{proposition}
The objective of the present paper is to unify and extend the
above results. To this end, we first discuss in Section~\ref{sec:2}
suitable notions of proximity in Banach spaces. Based on these, we
propose our extension of Moreau's decomposition in
Section~\ref{sec:3}. A feature of our analysis is to rely heavily
on convex analytical tools, which allows us to derive our main
result with simpler proofs than those utilized in the above
special case.
\section{Proximity in Banach spaces}
\label{sec:2}
Let $\varphi\in\Gamma_0(\ensuremath{\mathcal X})$.
As seen in the Introduction, if $\ensuremath{\mathcal X}$ is a Hilbert space, Moreau's
proximity operator is defined by
\begin{equation}
\label{e:2011-03-12a}
(\forall x\in\ensuremath{\mathcal X})\quad\ensuremath{{\operatorname{prox}}}_\varphi x=\underset{y\in\ensuremath{\mathcal X}}
{\operatorname{argmin}}\Big(\varphi(y)+\frac12\|x-y\|^2\Big).
\end{equation}
In this section we discuss two extensions of this operator
in Banach spaces. We recall that $\varphi$ is coercive if
$\lim_{\|y\|\to\ensuremath{{+\infty}}}\varphi(y)=\ensuremath{{+\infty}}$ and supercoercive if
$\lim_{\|y\|\to\ensuremath{{+\infty}}}\varphi(y)/\|y\|=\ensuremath{{+\infty}}$.
As usual, the subdifferential operator of $\varphi$ is denoted by
$\partial\varphi$. Finally, the strong relative interior of a
convex set $C\subset\ensuremath{\mathcal X}$ is
\begin{equation}
\label{e:dsri}
\ensuremath{{\operatorname{sri}}\,} C=\Menge{x\in C}{\bigcup_{\lambda>0}\lambda(C-x)=\ensuremath{\overline{\operatorname{span}}\,} (C-x)}.
\end{equation}
We shall also require the following facts.
\begin{lemma}[\rm\cite{More64,Rock66}]
\label{l:7}
Let $f\in\Gamma_0(\ensuremath{\mathcal X})$ and let $x^*\in\ensuremath{\mathcal X}^*$. Then $f-x^*$ is
coercive if and only if $x^*\in\ensuremath{\operatorname{int}\operatorname{dom}}\, f^*$.
\end{lemma}
\begin{lemma}[\rm{\cite[Theorem~3.4]{Ccm101}}]
\label{l:8}
Let $f\in\Gamma_0(\ensuremath{\mathcal X})$ be supercoercive. Then
$\ensuremath{\operatorname{dom}} f^*=\ensuremath{\mathcal X}^*$.
\end{lemma}
\begin{lemma}[\rm\cite{Atto86}]
\label{l:9}
Let $f$ and $\varphi$ be functions in $\Gamma_0(\ensuremath{\mathcal X})$ such that
$0\in\ensuremath{{\operatorname{sri}}\,}(\ensuremath{\operatorname{dom}} f-\ensuremath{\operatorname{dom}}\varphi)$. Then the following hold.
\begin{enumerate}
\item
\label{l:9i}
$(\varphi+f)^*=\varphi^*\ensuremath{\mbox{\footnotesize$\,\square\,$}} f^*$ and the infimal convolution
is exact everywhere: $(\forall x^*\in\ensuremath{\mathcal X}^*)(\ensuremath{\exists\,} y^*\in\ensuremath{\mathcal X}^*)$
$(\varphi+f)^*(x^*)=\varphi^*(y^*)+f^*(x^*-y^*)$.
\item
\label{l:9ii}
$\partial(\varphi+f)=\partial\varphi+\partial f$.
\end{enumerate}
\end{lemma}
\subsection{Legendre functions}
We review the notion of a Legendre function, which was
introduced in Euclidean spaces in \cite{Rock70} and extended to
Banach spaces in \cite{Ccm101} (see also \cite{Borw01} for
further developments in the nonreflexive case).
\begin{definition}{\rm\cite[Definition~5.2]{Ccm101}}
\label{d:legendre}
Let $f\in\Gamma_0(\ensuremath{\mathcal X})$. Then $f$ is:
\begin{enumerate}
\item
essentially smooth, if $\partial f$ is both locally
bounded and single-valued on its domain;
\item
essentially strictly convex, if $(\partial f)^{-1}$ is locally
bounded on its domain and $f$ is strictly convex on every
convex subset of $\ensuremath{\operatorname{dom}}\partial f$;
\item
a Legendre function, if it is both essentially smooth and
essentially strictly convex.
\end{enumerate}
\end{definition}
Some key properties of Legendre functions are listed below.
\begin{lemma}
\label{l:5}
Let $f\in\Gamma_0(\ensuremath{\mathcal X})$ be a Legendre function.
Then the following hold.
\begin{enumerate}
\item
\label{l:5i}
$f^*$ is a Legendre function {\rm\cite[Corollary~5.5]{Ccm101}}.
\item
\label{l:5ii}
$\ensuremath{\operatorname{dom}}\partial f=\ensuremath{\operatorname{int}\operatorname{dom}f}\neq\ensuremath{{\varnothing}}$ and $f$ is G\^ateaux differentiable
on $\ensuremath{\operatorname{int}\operatorname{dom}f}$ {\rm\cite[Theorem~5.6]{Ccm101}}.
\item
\label{l:5iii}
$\nabla f\colon\ensuremath{\operatorname{int}\operatorname{dom}f}\to\ensuremath{\operatorname{int}\operatorname{dom}f^*}$ is bijective with inverse
$\nabla f^*\colon\ensuremath{\operatorname{int}\operatorname{dom}f^*}\to\ensuremath{\operatorname{int}\operatorname{dom}f}$ {\rm\cite[Theorem~5.10]{Ccm101}}.
\end{enumerate}
\end{lemma}
\subsection{$D$-proximity operators}
In this subsection we discuss a notion of proximity based on
Bregman distances investigated in \cite{Sico03} and which goes back
to \cite{Cens92,Tebo92}.
The first extension of \eqref{e:2011-03-12a} was investigated in
\cite{Sico03}. Let $f\in\Gamma_0(\ensuremath{\mathcal X})$ be a Legendre function.
The Bregman distance associated with $f$ is
\begin{equation}
\label{e:elnido2011-03-05}
\begin{aligned}
D_f\colon\ensuremath{\mathcal X}\times\ensuremath{\mathcal X}&\to\,[0,\ensuremath{{+\infty}}]\\
(y,x)&\mapsto
\begin{cases}
f(y)-f(x)-\pair{y-x}{\ensuremath{\nabla f}(x)},&\text{if}\;\;x\in\ensuremath{\operatorname{int}\operatorname{dom}f};\\
\ensuremath{{+\infty}},&\text{otherwise}.
\end{cases}
\end{aligned}
\end{equation}
For every $\varphi\in\Gamma_0(\ensuremath{\mathcal X})$, we define the function
$\varphi\ensuremath{\mbox{\Large$\,\diamond\,$}} f\colon\ensuremath{\mathcal X}\to\ensuremath{\,\left[-\infty,+\infty\right]}$ by
\begin{equation}
\label{e:inf-convo-breg}
(\forall x\in\ensuremath{\mathcal X})\quad
(\varphi\ensuremath{\mbox{\Large$\,\diamond\,$}} f)(x)=\inf_{y\in\ensuremath{\mathcal X}}\big(\varphi(y)+
D_f(y,x)\big).
\end{equation}
The following proposition refines and complements some results of
\cite[Section~3.4]{Sico03}.
\begin{proposition}
\label{p:2011-04-13}
Let $f\in\Gamma_0(\ensuremath{\mathcal X})$ be a Legendre function, let
$\varphi\in\Gamma_0(\ensuremath{\mathcal X})$ be such that
\begin{equation}
\label{e:2011-04-13a}
0\in\ensuremath{{\operatorname{sri}}\,}(\ensuremath{\operatorname{dom}} f-\ensuremath{\operatorname{dom}}\varphi),
\end{equation}
and let $x\in\ensuremath{\operatorname{int}\operatorname{dom}}\, f$.
Suppose that one of the following holds.
\begin{enumerate}
\item
\label{p:2011-04-13i}
$\ensuremath{\nabla f}(x)\in\ensuremath{\operatorname{int}}(\ensuremath{\operatorname{dom}} f^*+\ensuremath{\operatorname{dom}}\varphi^*)$.
\item
\label{p:2011-04-13ii}
$\ensuremath{\operatorname{int}\operatorname{dom}}\, f^*\subset\ensuremath{\operatorname{int}}(\ensuremath{\operatorname{dom}} f^*+\ensuremath{\operatorname{dom}}\varphi^*)$.
\item
\label{p:2011-04-13iii}
$f$ is supercoercive.
\item
\label{p:2011-04-13iv}
$\inf\varphi(\ensuremath{\mathcal X})>\ensuremath{{-\infty}}$.
\end{enumerate}
Then there exists a unique point $p\in\ensuremath{\mathcal X}$ such that
$(\varphi\ensuremath{\mbox{\Large$\,\diamond\,$}} f)(x)=\varphi(p)+D_f(p,x)$;
moreover, $p$ lies in $\ensuremath{\operatorname{dom}}\partial\varphi\cap\ensuremath{\operatorname{int}\operatorname{dom}f}$ and it is
characterized by the inclusion
\begin{equation}
\label{e:2011-04-13b}
\ensuremath{\nabla f}(x)-\ensuremath{\nabla f}(p)\in\partial\varphi(p).
\end{equation}
\end{proposition}
\begin{proof}
Set $f_x\colon\ensuremath{\mathcal X}\to\ensuremath{\,\left]-\infty,+\infty\right]}\colon y\mapsto f(y)-\pair{y}{\ensuremath{\nabla f}(x)}$.
Then the minimizers of $\varphi+D_f(\cdot,x)$ coincide with those
of $\varphi+f_x$ and our assumptions imply that
\begin{equation}
\label{e:2011-04-13e}
\varphi+f_x\in\Gamma_0(\ensuremath{\mathcal X}).
\end{equation}
Now let $p\in\ensuremath{\mathcal X}$.
It follows from \eqref{e:2011-04-13a}, Lemma~\ref{l:9}\ref{l:9ii},
and Lemma~\ref{l:5}\ref{l:5ii} that
\begin{eqnarray}
(\varphi\ensuremath{\mbox{\Large$\,\diamond\,$}} f)(x)=\varphi(p)+D_f(p,x)
&\Leftrightarrow&p\;\:\text{minimizes}\;\:\varphi+f_x
\nonumber\\
&\Leftrightarrow&0\in\partial\big(\varphi+f_x\big)(p)
\nonumber\\
&\Leftrightarrow&0\in\partial\varphi(p)+\partial f(p)-\ensuremath{\nabla f}(x)
\nonumber\\
&\Leftrightarrow&0\in\partial\varphi(p)+\ensuremath{\nabla f}(p)-\ensuremath{\nabla f}(x)
\nonumber\\
&\Leftrightarrow&\ensuremath{\nabla f}(x)-\ensuremath{\nabla f}(p)\in\partial\varphi(p)
\label{e:2011-04-13c}\\
&\Rightarrow&p\in\ensuremath{\operatorname{dom}}\partial\varphi\cap\ensuremath{\operatorname{int}\operatorname{dom}f}.
\label{e:2011-04-13d}
\end{eqnarray}
Hence, the minimizers of $\varphi+f_x$ are in $\ensuremath{\operatorname{int}\operatorname{dom}f}$. However,
since $f$ is essentially strictly convex, it is strictly convex
on $\ensuremath{\operatorname{int}\operatorname{dom}f}$ and so is therefore $\varphi+f_x$. This shows
that $\varphi+f_x$ admits at most one minimizer.
It remains to establish existence.
\ref{p:2011-04-13i}:
It follows from \eqref{e:2011-04-13e} that, to show existence, it
is enough to show that $\varphi+f_x$ is coercive
\cite[Theorem~2.5.1(ii)]{Zali02}. In view of Lemma~\ref{l:7}, this
is equivalent to showing that $\ensuremath{\nabla f}(x)\in\ensuremath{\operatorname{int}\operatorname{dom}}\,(f+\varphi)^*$.
However, it follows from \eqref{e:2011-04-13a} and
Lemma~\ref{l:9}\ref{l:9i} that
\begin{equation}
\label{e:2011-04-13f}
\ensuremath{\operatorname{int}\operatorname{dom}}\,(f+\varphi)^*
=\ensuremath{\operatorname{int}\operatorname{dom}}\,(f^*\ensuremath{\mbox{\footnotesize$\,\square\,$}}\varphi^*)
=\ensuremath{\operatorname{int}}(\ensuremath{\operatorname{dom}} f^*+\ensuremath{\operatorname{dom}}\varphi^*).
\end{equation}
\ref{p:2011-04-13ii}$\Rightarrow$\ref{p:2011-04-13i}:
Lemma~\ref{l:5}\ref{l:5iii}.
\ref{p:2011-04-13iii}$\Rightarrow$\ref{p:2011-04-13ii}:
By Lemma~\ref{l:8}, $\ensuremath{\operatorname{dom}} f^*=\ensuremath{\mathcal X}^*$ and, since
$\ensuremath{\operatorname{dom}}\varphi^*\neq\ensuremath{{\varnothing}}$,
$\ensuremath{\operatorname{int}\operatorname{dom}}\, f^*\subset\ensuremath{\operatorname{int}}(\ensuremath{\operatorname{dom}} f^*+\ensuremath{\operatorname{dom}}\varphi^*)$.
\ref{p:2011-04-13iv}$\Rightarrow$\ref{p:2011-04-13ii}:
We have
$\inf\varphi(\ensuremath{\mathcal X})>\ensuremath{{-\infty}}$ $\Rightarrow$
$\varphi^*(0)=-\inf\varphi(\ensuremath{\mathcal X})<\ensuremath{{+\infty}}$ $\Rightarrow$
$0\in\ensuremath{\operatorname{dom}}\varphi^*$. Hence,
$\ensuremath{\operatorname{int}\operatorname{dom}}\, f^*\subset\ensuremath{\operatorname{int}}(\ensuremath{\operatorname{dom}} f^*+\ensuremath{\operatorname{dom}}\varphi^*)$.
\end{proof}
In view of Proposition~\ref{p:2011-04-13} and
Lemma~\ref{l:5}\ref{l:5iii}, the following is well
defined.
\begin{definition}
\label{d:bprox}
Let $f\in\Gamma_0(\ensuremath{\mathcal X})$ be a Legendre function and let
$\varphi\in\Gamma_0(\ensuremath{\mathcal X})$ be such that
$0\in\ensuremath{{\operatorname{sri}}\,}(\ensuremath{\operatorname{dom}} f-\ensuremath{\operatorname{dom}}\varphi)$. Set
\begin{equation}
E=(\ensuremath{\operatorname{int}\operatorname{dom}f})\cap\big(\ensuremath{\nabla f}^*\big(\ensuremath{\operatorname{int}}(\ensuremath{\operatorname{dom}} f^*+\ensuremath{\operatorname{dom}}\varphi^*)\big)\big).
\end{equation}
The $D$-proximity (or Bregman proximity) operator of $\varphi$
relative to $f$ is
\begin{equation}
\label{e:avril2011a}
\begin{aligned}
\ensuremath{{\operatorname{bprox}}}_\varphi^f\colon E\to\ensuremath{\operatorname{int}\operatorname{dom}f}\colon
x\mapsto\underset{y\in\ensuremath{\mathcal X}}
{\operatorname{argmin}}\big(\varphi(y)+D_f(y,x)\big).
\end{aligned}
\end{equation}
\end{definition}
\begin{remark}
\label{r:2011-03-13a}
In connection with Definition~\ref{d:bprox}, let us make a couple
of observations.
\begin{enumerate}
\item
It follows from Proposition~\ref{p:2011-04-13} that,
if $\ensuremath{\operatorname{int}\operatorname{dom}}\, f^*\subset\ensuremath{\operatorname{int}}(\ensuremath{\operatorname{dom}}\varphi^*+\ensuremath{\operatorname{dom}} f^*)$ (in particular
if $f$ is supercoercive or if $\inf\varphi(\ensuremath{\mathcal X})>\ensuremath{{-\infty}}$), then
$\ensuremath{{\operatorname{bprox}}}_\varphi^f\colon\ensuremath{\operatorname{int}\operatorname{dom}f}\to\ensuremath{\operatorname{int}\operatorname{dom}f}$.
\item
\label{r:2011-03-13aii}
Suppose that $\ensuremath{\mathcal X}$ is hilbertian and that $f=\|\cdot\|^2/2$, and
let $\varphi\in\Gamma_0(\ensuremath{\mathcal X})$.
Then $\varphi\ensuremath{\mbox{\Large$\,\diamond\,$}} f=\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} f$ and
$\ensuremath{{\operatorname{bprox}}}_\varphi^f=\ensuremath{{\operatorname{prox}}}_\varphi$.
\end{enumerate}
\end{remark}
\subsection{Anisotropic proximity operators}
An alternative extension of the notion of proximity can be
obtained by replacing the function $\|\cdot\|^2/2$ in
\eqref{e:2011-03-12a} by a Legendre function $f$.
\begin{proposition}
\label{p:2011-04-14}
Let $f\in\Gamma_0(\ensuremath{\mathcal X})$ be a Legendre function, let
$\varphi\in\Gamma_0(\ensuremath{\mathcal X})$ be such that
\begin{equation}
\label{e:2011-04-14b}
0\in\ensuremath{{\operatorname{sri}}\,}(\ensuremath{\operatorname{dom}} f^*-\ensuremath{\operatorname{dom}}\varphi^*),
\end{equation}
and let $x\in\ensuremath{{\operatorname{sri}}\,}(\ensuremath{\operatorname{dom}} f+\ensuremath{\operatorname{dom}}\varphi)$.
Then there exists a unique point $p\in\ensuremath{\mathcal X}$ such that
$(\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} f)(x)=\varphi(p)+f(x-p)$; moreover, $p$ is
characterized by the inclusion
\begin{equation}
\label{e:2011-03-11}
\ensuremath{\nabla f}(x-p)\in\partial\varphi(p).
\end{equation}
\end{proposition}
\begin{proof}
Using \eqref{e:2011-04-14b} and Lemma~\ref{l:9}\ref{l:9i}, we obtain
\begin{equation}
(\varphi^*+f^*)^*=\varphi^{**}\ensuremath{\mbox{\footnotesize$\,\square\,$}} f^{**}=\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} f
\end{equation}
and the fact that the infimum in the infimal convolution is
attained everywhere.
On the other hand, since $x\in\ensuremath{{\operatorname{sri}}\,}(\ensuremath{\operatorname{dom}} f+\ensuremath{\operatorname{dom}}\varphi)$, we have
\begin{equation}
0\in\ensuremath{{\operatorname{sri}}\,}\big(\ensuremath{\operatorname{dom}}\varphi-(x-\ensuremath{\operatorname{dom}} f)\big)
=\ensuremath{{\operatorname{sri}}\,}\big(\ensuremath{\operatorname{dom}}\varphi-\ensuremath{\operatorname{dom}} f(x-\cdot)\big).
\end{equation}
Consequently, by Lemma~\ref{l:9}\ref{l:9ii},
\begin{equation}
\label{e:sum-rule}
\partial\big(\varphi+f(x-\cdot)\big)
=\partial\varphi+\partial f(x-\cdot).
\end{equation}
Now let $p\in\ensuremath{\mathcal X}$. It follows from \eqref{e:sum-rule} and
Lemma~\ref{l:5}\ref{l:5ii} that
\begin{eqnarray}
p\;\:\text{minimizes}\;\:\varphi+f(x-\cdot)
&\Leftrightarrow&0\in\partial\big(\varphi+f(x-\cdot)\big)(p)
\nonumber\\
&\Leftrightarrow&0\in\partial\varphi(p)-\partial f(x-p)
\nonumber\\
&\Leftrightarrow&0\in\partial\varphi(p)-\ensuremath{\nabla f}(x-p)
\nonumber\\
&\Leftrightarrow&\ensuremath{\nabla f}(x-p)\in\partial\varphi(p)
\label{e:2011-04-26a}\\
&\Rightarrow&x-p\in\ensuremath{\operatorname{int}\operatorname{dom}f}.
\label{e:2011-04-26b}
\end{eqnarray}
To show uniqueness, suppose that $p$ and $q$ are two distinct
minimizers of $\varphi+f(x-\cdot)$. Then
$(\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} f)(x)=\varphi(p)+f(x-p)=\varphi(q)+f(x-q)$ and,
by \eqref{e:2011-04-26b}, $x-p$ and $x-q$ lie in $\ensuremath{\operatorname{int}\operatorname{dom}f}$.
Now let $r=(1/2)(p+q)$ and suppose that $p\neq q$.
Lemma~\ref{l:5}\ref{l:5ii} asserts that $f$ is strictly
convex on the convex set $\ensuremath{\operatorname{int}\operatorname{dom}f}=\ensuremath{\operatorname{dom}}\partial f$. Therefore,
invoking the convexity of $\varphi$,
\begin{align}
(\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} f)(x)
&\leq\varphi(r)+f(x-r)\nonumber\\
&<\frac12\big(\varphi(p)+\varphi(q)\big)+
\frac12\big(f(x-p)+f(x-q)\big)\nonumber\\
&=(\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} f)(x),
\end{align}
which is impossible.
\end{proof}
Using Proposition~\ref{p:2011-04-14}, we can now introduce the
anisotropic proximity operator of $\varphi$.
\begin{definition}
\label{d:aprox}
Let $f\in\Gamma_0(\ensuremath{\mathcal X})$ be a Legendre function and let
$\varphi\in\Gamma_0(\ensuremath{\mathcal X})$ be such that
$0\in\ensuremath{{\operatorname{sri}}\,}(\ensuremath{\operatorname{dom}} f^*-\ensuremath{\operatorname{dom}}\varphi^*)$. Set
\begin{equation}
E=\ensuremath{{\operatorname{sri}}\,}(\ensuremath{\operatorname{dom}} f+\ensuremath{\operatorname{dom}}\varphi).
\end{equation}
The anisotropic proximity operator of $\varphi$
relative to $f$ is
\begin{equation}
\label{e:avril2011b}
\ensuremath{{\operatorname{aprox}}}_\varphi^f\colon E\to\ensuremath{\mathcal X}\colon
x\mapsto\underset{y\in\ensuremath{\mathcal X}}{\operatorname{argmin}}
\big(\varphi(y)+f(x-y)\big).
\end{equation}
\end{definition}
\begin{remark}
\label{r:2011-03-13b}
Suppose that $\ensuremath{\mathcal X}$ is hilbertian and that $f=\|\cdot\|^2/2$, and
let $\varphi\in\Gamma_0(\ensuremath{\mathcal X})$. Then
$\ensuremath{{\operatorname{aprox}}}_\varphi^f=\ensuremath{{\operatorname{prox}}}_\varphi$.
\end{remark}
\section{Main result}
\label{sec:3}
In the previous section we have described two extensions of the
classical proximity operator. Our main result is a generalization
of Moreau's decomposition (Proposition~\ref{p:3}) in Banach
spaces which involves a mix of these two extensions.
\begin{theorem}
\label{t:puerto-princesa2011-03-06}
Let $f\in\Gamma_0(\ensuremath{\mathcal X})$ be a Legendre function, let
$\varphi\in\Gamma_0(\ensuremath{\mathcal X})$ be such that
\begin{equation}
\label{e:2011-04-15b}
0\in\ensuremath{{\operatorname{sri}}\,}(\ensuremath{\operatorname{dom}} f^*-\ensuremath{\operatorname{dom}}\varphi^*),
\end{equation}
and let $x\in(\ensuremath{\operatorname{int}\operatorname{dom}f})\cap\ensuremath{\operatorname{int}}(\ensuremath{\operatorname{dom}} f+\ensuremath{\operatorname{dom}}\varphi)$.
Then the following hold.
\begin{enumerate}
\item
\label{t:puerto-princesa2011-03-06i}
$f(x)=(\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} f)(x)+(\varphi^*\ensuremath{\mbox{\Large$\,\diamond\,$}} f^*)
\big(\ensuremath{\nabla f}(x)\big)$.
\item
\label{t:puerto-princesa2011-03-06iia}
$x=\ensuremath{{\operatorname{aprox}}}^f_\varphi x+\nabla f^*
\big(\ensuremath{{\operatorname{bprox}}}^{f^*}_{\varphi^*}\big(\nabla f(x)\big)\big)$.
\item
\label{t:puerto-princesa2011-03-06iib}
$\bpair{\ensuremath{{\operatorname{aprox}}}^f_\varphi x}{\ensuremath{{\operatorname{bprox}}}_{\varphi^*}^{f^*}
\big(\ensuremath{\nabla f}(x)\big)}=\varphi\big(\ensuremath{{\operatorname{aprox}}}^f_\varphi x\big)+
\varphi^*\big(\ensuremath{{\operatorname{bprox}}}_{\varphi^*}^{f^*}\big(\ensuremath{\nabla f}(x)\big)\big)$.
\item
\label{t:puerto-princesa2011-03-06iic}
$\bpair{\ensuremath{{\operatorname{aprox}}}^f_\varphi x}{\ensuremath{\nabla f}\big(x-\ensuremath{{\operatorname{aprox}}}^f_\varphi x\big)}=
\varphi\big(\ensuremath{{\operatorname{aprox}}}^f_\varphi x\big)+
\varphi^*\big(\ensuremath{\nabla f}\big(x-\ensuremath{{\operatorname{aprox}}}^f_\varphi x\big)\big)$.
\end{enumerate}
\end{theorem}
\begin{proof}
Since $x\in\ensuremath{\operatorname{int}}(\ensuremath{\operatorname{dom}} f+\ensuremath{\operatorname{dom}}\varphi)$,
Lemma~\ref{l:5}\ref{l:5iii} yields
\begin{equation}
x\in\ensuremath{{\operatorname{sri}}\,}(\ensuremath{\operatorname{dom}} f+\ensuremath{\operatorname{dom}}\varphi)\quad\text{and}\quad
\ensuremath{\nabla f}^*\big(\ensuremath{\nabla f}(x)\big)\in\ensuremath{\operatorname{int}}\big(\ensuremath{\operatorname{dom}} f^{**}+\ensuremath{\operatorname{dom}}\varphi^{**}\big).
\end{equation}
Hence, it follows from Proposition~\ref{p:2011-04-14}
that $\ensuremath{{\operatorname{aprox}}}^f_\varphi x$ is well defined and, from
Lemma~\ref{l:5}\ref{l:5i} and
Proposition~\ref{p:2011-04-13}\ref{p:2011-04-13i}
(applied to $f^*$ and $\varphi^*$), that
$\ensuremath{\nabla f}^*(\ensuremath{{\operatorname{bprox}}}_{\varphi^*}^{f^*}(\ensuremath{\nabla f}(x)))$ is well defined.
In addition,
\begin{equation}
\label{e:2011-04-27}
(\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} f)(x)\in\ensuremath{\mathbb{R}}
\quad\text{and}\quad
(\varphi^*\ensuremath{\mbox{\Large$\,\diamond\,$}} f^*)\big(\ensuremath{\nabla f}(x)\big)\in\ensuremath{\mathbb{R}}.
\end{equation}
\ref{t:puerto-princesa2011-03-06i}:
It follows from \eqref{e:elnido2011-03-05},
Lemma~\ref{l:5}\ref{l:5iii}, and the Fenchel-Young identity
\cite[Theorem~2.4.2(iii)]{Zali02} that
\begin{align}
(\forall x^*\in\ensuremath{\mathcal X}^*)\quad
D_{f^*}\big(x^*,\ensuremath{\nabla f}(x)\big)
&=f^*( x^*)-f^*\big(\ensuremath{\nabla f}(x)\big)-\pair{x^*-\ensuremath{\nabla f}(x)}{x}_*\nonumber\\
&=f^*(x^*)+f(x)-\pair{x^*}{x}_*.
\end{align}
This, \eqref{e:inf-convo-breg}, \eqref{e:2011-04-15b}, and
Lemma~\ref{l:9}\ref{l:9i} imply that
\begin{align}
(\varphi^*\ensuremath{\mbox{\Large$\,\diamond\,$}} f^*)\big(\ensuremath{\nabla f} (x)\big)
&=\inf_{x^*\in\ensuremath{\mathcal X}^*}\big(\varphi^*(x^*)+f^*(x^*)+f(x)-
\pair{x^*}{x}_*\big)\nonumber\\
&=f(x)-\sup_{x^*\in\ensuremath{\mathcal X}^*}\big(\pair{x^*}{x}_*-\varphi^*(x^*)-
f^*(x^*)\big)\nonumber\\
&=f(x)-(\varphi^*+f^*)^*(x)\nonumber\\
&=f(x)-(\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} f)(x).
\end{align}
In view of \eqref{e:2011-04-27}, we obtain the announced identity.
\ref{t:puerto-princesa2011-03-06iia}:
Let $p\in\ensuremath{\mathcal X}$. Using Proposition~\ref{p:2011-04-14},
Lemma~\ref{l:5}\ref{l:5iii}, and
Proposition~\ref{p:2011-04-13}\ref{p:2011-04-13i}, we obtain
\begin{eqnarray}
p=\ensuremath{{\operatorname{aprox}}}^f_\varphi x
&\Leftrightarrow&\ensuremath{\nabla f}(x-p)\in\partial\varphi(p)
\label{e:puerto-princesa2011-03-06a}\\
&\Leftrightarrow&
p\in\partial\varphi^*\big(\ensuremath{\nabla f}(x-p)\big)\nonumber\\
&\Leftrightarrow&\ensuremath{\nabla f}^*\big(\ensuremath{\nabla f}(x)\big)-\ensuremath{\nabla f}^*\big(\ensuremath{\nabla f}(x-p)\big)\in
\partial\varphi^*\big(\ensuremath{\nabla f}(x-p)\big)\nonumber\\
&\Leftrightarrow&\ensuremath{\nabla f}(x-p)=\ensuremath{{\operatorname{bprox}}}_{\varphi^*}^{f^*}\big(\ensuremath{\nabla f}(x)\big)
\label{e:puerto-princesa2011-03-06b}\\
&\Leftrightarrow&x-p=\ensuremath{\nabla f}^*\big(\ensuremath{{\operatorname{bprox}}}_{\varphi^*}^{f^*}
\big(\ensuremath{\nabla f}(x)\big)\big).
\end{eqnarray}
\ref{t:puerto-princesa2011-03-06iib}:
Set $p=\ensuremath{{\operatorname{aprox}}}^f_\varphi x$. As seen in
\eqref{e:puerto-princesa2011-03-06b} and
\eqref{e:puerto-princesa2011-03-06a},
\begin{equation}
\label{e:puerto-princesa2011-03-06c}
\ensuremath{{\operatorname{bprox}}}_{\varphi^*}^{f^*}\big(\ensuremath{\nabla f}(x)\big)=\ensuremath{\nabla f}(x-p)\in
\partial\varphi(p).
\end{equation}
Hence, the Fenchel-Young identity yields
\begin{align}
\label{e:2010-11-04h}
\pair{p}{\ensuremath{{\operatorname{bprox}}}_{\varphi^*}^{f^*}\big(\ensuremath{\nabla f}(x)\big)}
&=\pair{p}{\ensuremath{\nabla f}(x-p)}\nonumber\\
&=\varphi(p)+\varphi^*\big(\ensuremath{\nabla f}(x-p)\big)\nonumber\\
&=\varphi(p)+\varphi^*\big(\ensuremath{{\operatorname{bprox}}}_{\varphi^*}^{f^*}
\big(\ensuremath{\nabla f}(x)\big)\big).
\end{align}
\ref{t:puerto-princesa2011-03-06iic}: This follows at once from
\ref{t:puerto-princesa2011-03-06iib} and
\eqref{e:puerto-princesa2011-03-06c}.
\end{proof}
Theorem~\ref{t:puerto-princesa2011-03-06} provides a range of new
decomposition schemes, even in the case when $\ensuremath{\mathcal X}$ is a Hilbert
space. Thus, in the following result, we obtain a new hilbertian
frame decomposition principle (for background on frames and their
applications, see \cite{Chri08}).
\begin{corollary}
\label{c:frames}
Suppose that $\ensuremath{\mathcal X}$ is a separable Hilbert space, let $I$ be a
countable set, and let $(e_i)_{i\in I}$ be a frame in $\ensuremath{\mathcal X}$, i.e.,
\begin{equation}
\label{e:frame-bounds}
(\ensuremath{\exists\,}\alpha\in\ensuremath{\left]0,+\infty\right[})(\ensuremath{\exists\,}\beta\in\ensuremath{\left]0,+\infty\right[})(\forall x\in\ensuremath{\mathcal X})\quad
\alpha\|x\|^2\leq\sum_{i\in I}|\pair{x}{e_i}|^2\leq\beta\|x\|^2.
\end{equation}
Let $S\colon\ensuremath{\mathcal X}\to\ensuremath{\mathcal X}\colon x\mapsto\sum_{i\in I}
\pair{x}{e_i}e_i$ be the associated frame operator and let
$(e^*_i)_{i\in I}=(S^{-1}e_i)_{i\in I}$ be the associated
canonical dual frame. Furthermore, let $\varphi\in\Gamma_0(\ensuremath{\mathcal X})$,
let $x\in\ensuremath{\mathcal X}$, and set
\begin{equation}
\label{e:frame2}
a(x)=\underset{y\in \ensuremath{\mathcal X}}{\operatorname{argmin}}
\left(\varphi(y)+\frac{1}{2}\sum_{i\in I}|\pair{x-y}{e_i}|^2
\right)
\end{equation}
and
\begin{equation}
\label{e:frame3}
b(x)=\underset{x^*\in\ensuremath{\mathcal X}}{\operatorname{argmin}}
\left(\varphi^*(x^*)-\pair{x^*}{x}+\frac{1}{2}\sum_{i\in I}
|\pair{x^*}{e^*_i}|^2\right).
\end{equation}
Then $x=a(x)+\sum_{i\in I}\pair{b(x)}{e^*_i}e^*_i$.
\end{corollary}
\begin{proof}
Set $f\colon\ensuremath{\mathcal X}\to\ensuremath{\mathbb{R}}\colon x\mapsto(1/2)
\sum_{i\in I}|\pair{x}{e_i}|^2$. It is easily seen that $f$ is
Fr\'echet differentiable on $\ensuremath{\mathcal X}$ with $\ensuremath{\nabla f}=S$. It therefore
follows from \cite[Theorem~5.6]{Ccm101} that $f$ is essentially
smooth. Now fix $x^*\in\ensuremath{\mathcal X}$. Since the frame operator of
$(e_i^*)_{i\in I}$ is $S^{-1}$ \cite[Lemma~5.1.6]{Chri08}, we have
\begin{equation}
\label{e:2011-04-18c}
\pair{S^{-1}x^*}{x^*}=
\bpair{\sum_{i\in I}\pair{x^*}{e_i^*}e_i^*}{x^*}
=\sum_{i\in I}|\pair{x^*}{e_i^*}|^2=2f(S^{-1}x^*).
\end{equation}
Now set $g\colon\ensuremath{\mathcal X}\to\ensuremath{\mathbb{R}}\colon x\mapsto f(x)-\pair{x}{x^*}$.
Then $g$ is a differentiable convex function and
$\nabla g\colon x\mapsto Sx-x^*$ vanishes at $x=S^{-1}x^*$.
Hence, using \eqref{e:2011-04-18c}, we obtain
\begin{equation}
\label{e:conju-frame}
f^*(x^*)=-\min_{x\in\ensuremath{\mathcal X}}g(x)=\pair{S^{-1}x^*}{x^*}-f(S^{-1}x^*)
=f(S^{-1}x^*)=\frac12\sum_{i\in I}|\pair{x^*}{e^*_i}|^2.
\end{equation}
Hence, as above, $f^*$ is Fr\'echet differentiable on $\ensuremath{\mathcal X}$ with
$\ensuremath{\nabla f}^*=S^{-1}$ and, in turn, essentially smooth, which makes $f$
essentially strictly convex \cite[Theorem~5.4]{Ccm101}. Altogether,
$f$ is a Legendre function with
\begin{equation}
\label{e:2011-04-18d}
\ensuremath{\operatorname{dom}} f=\ensuremath{\mathcal X},\quad\ensuremath{\operatorname{dom}} f^*=\ensuremath{\mathcal X},\quad\ensuremath{\nabla f}=S,
\quad\text{and}\quad\ensuremath{\nabla f}^*=S^{-1}.
\end{equation}
Moreover,
it follows from \eqref{e:avril2011a}, \eqref{e:avril2011b},
\eqref{e:2011-04-18d}, Lemma~\ref{l:5}\ref{l:5iii}, \eqref{e:frame2},
\eqref{e:frame3}, and \eqref{e:conju-frame} that
\begin{equation}
\ensuremath{{\operatorname{bprox}}}_{\varphi^*}^{f^*}(\ensuremath{\nabla f}(x))=b(x)
\quad\text{and}\quad
\ensuremath{{\operatorname{aprox}}}_{\varphi}^{f}(x)=a(x).
\end{equation}
The result is therefore an application of
Theorem~\ref{t:puerto-princesa2011-03-06}%
\ref{t:puerto-princesa2011-03-06iia}.
\end{proof}
\begin{remark}
\label{r:2011-04-18}
Corollary~\ref{c:frames} can be regarded as an extension of Moreau's
decomposition principle in separable Hilbert spaces. Indeed, in the
special case when $(e_i)_{i\in I}$ is an orthonormal basis in
Corollary~\ref{c:frames}, we recover Proposition~\ref{p:3}\ref{p:3i}.
\end{remark}
The next application is set in uniformly convex and uniformly smooth
Banach spaces.
\begin{corollary}
\label{c:puerto-princesa2011-03-06}
Suppose that $\ensuremath{\mathcal X}$ is uniformly convex and uniformly smooth,
let $J$ be its duality mapping, set $q=\|\cdot\|^2/2$, and
let $\varphi\in\Gamma_0(\ensuremath{\mathcal X})$.
Then $q^*=\|\cdot\|^2_*/2$ and the following hold for every
$x\in\ensuremath{\mathcal X}$.
\begin{enumerate}
\item
\label{c:puerto-princesa2011-03-06ii}
$q(x)=(\varphi\ensuremath{\mbox{\footnotesize$\,\square\,$}} q)(x)+(\varphi^*\ensuremath{\mbox{\Large$\,\diamond\,$}} q^*)(Jx)$.
\item
\label{c:puerto-princesa2011-03-06i}
$x=\ensuremath{{\operatorname{aprox}}}^q_\varphi x+
J^{-1}\big(\ensuremath{{\operatorname{bprox}}}^{q^*}_{\varphi^*}(Jx)\big)$.
\item
\label{c:puerto-princesa2011-03-06iii}
$\bpair{\ensuremath{{\operatorname{aprox}}}^q_\varphi x}{\ensuremath{{\operatorname{bprox}}}_{\varphi^*}^{q^*}(Jx)}=
\varphi\big(\ensuremath{{\operatorname{aprox}}}^q_\varphi x\big)+
\varphi^*\big(\ensuremath{{\operatorname{bprox}}}_{\varphi^*}^{q^*}(Jx)\big)$.
\item
\label{c:puerto-princesa2011-03-06iv}
$\bpair{\ensuremath{{\operatorname{aprox}}}^q_\varphi x}{J\big(x-\ensuremath{{\operatorname{aprox}}}^q_\varphi x\big)}=
\varphi\big(\ensuremath{{\operatorname{aprox}}}^q_\varphi x\big)+
\varphi^*\big(J\big(x-\ensuremath{{\operatorname{aprox}}}^q_\varphi x\big)\big)$.
\end{enumerate}
\end{corollary}
\begin{proof}
This is an application of Theorem~\ref{t:puerto-princesa2011-03-06}
with $f=q$. Indeed, $\ensuremath{\operatorname{dom}} f=\ensuremath{\mathcal X}$, $\ensuremath{\operatorname{dom}} f^*=\ensuremath{\mathcal X}^*$, and $\ensuremath{\nabla f}=J$.
\end{proof}
In particular, if $\ensuremath{\mathcal X}$ is a Hilbert space in
Corollary~\ref{c:puerto-princesa2011-03-06}, if follows from
Remark~\ref{r:2011-03-13a}\ref{r:2011-03-13aii} and
Remark~\ref{r:2011-03-13b} that we
recover Moreau's decomposition principle (Proposition~\ref{p:3}) and
a fortiori Propositions~\ref{p:1} and \ref{p:2}.
Another noteworthy instance of
Corollary~\ref{c:puerto-princesa2011-03-06}
is when $\varphi=\iota_K$, where $K$ is a nonempty closed convex
cone in $\ensuremath{\mathcal X}$. In this case, $\varphi^*=\iota_{K^\ominus}$,
$\ensuremath{{\operatorname{aprox}}}^q_\varphi=P_K$, and we derive
from \eqref{e:gproj} and \eqref{e:dualitymap} that
$\ensuremath{{\operatorname{bprox}}}^q_\varphi=\Pi_K$. Hence,
Corollary~\ref{c:puerto-princesa2011-03-06}%
\ref{c:puerto-princesa2011-03-06i}\&%
\ref{c:puerto-princesa2011-03-06iii} yields Proposition~\ref{p:4}.
\begin{remark}
\label{r:2011-03-16}
Consider the setting of
Theorem~\ref{t:puerto-princesa2011-03-06}
and set $A=\partial\varphi$. Then,
by Rockafellar's theorem, $A$ is a maximally monotone operator
\cite[Theorem~3.1.11]{Zali02}. Moreover, it follows from
\eqref{e:2011-03-11}, Lemma~\ref{l:5}\ref{l:5iii},
and \eqref{e:2011-04-13b} that we can rewrite
Theorem~\ref{t:puerto-princesa2011-03-06}%
\ref{t:puerto-princesa2011-03-06iia} as
\begin{equation}
\label{e:2011-03-16j}
x=(\ensuremath{\operatorname{Id}}\,+\nabla f^*\circ A)^{-1}x+
\nabla f^*\circ\big(\ensuremath{\nabla f}^*+A^{-1}\big)^{-1}x,
\end{equation}
where $\ensuremath{\operatorname{Id}}\,$ is the identity operator on $\ensuremath{\mathcal X}$.
The results of \cite[Section~3.3]{Sico03} suggest that this
decomposition holds for more general maximally monotone
operators $A\colon\ensuremath{\mathcal X}\to 2^{\ensuremath{\mathcal X}^*}$.
If $\ensuremath{\mathcal X}$ is a Hilbert space and
$f=\|\cdot\|^2/2$, \eqref{e:2011-03-16j} yields the well-known
resolvent identity $\ensuremath{\operatorname{Id}}\,=(\ensuremath{\operatorname{Id}}\,+A)^{-1}+(\ensuremath{\operatorname{Id}}\,+A^{-1})^{-1}$,
which is true for any maximally monotone operator $A$
\cite[Proposition~23.18]{Livre1}.
\end{remark}
\noindent
{\bfseries Acknowledgement.}
The work of the first author was supported by the Agence
Nationale de la Recherche under grant ANR-08-BLAN-0294-02.
The authors thank one of the referees for making some
valuable comments.
|
\section{Introduction}
The almost sure invariance principle is a powerful tool in both probability and statistics. It says that the partial sums of random variables can be approximated by those of independent Gaussian random variables, and that the approximation error between the trajectories of the two processes is negligible compared to their size. More precisely, when $(X_i)_{i \geq 1}$ is a sequence of i.i.d. centered real valued random variables with a finite second moment, a sequence $(Z_i)_{i \geq 1}$ of i.i.d. centered Gaussian variables may be constructed is such a way that
\beq \label{strassen}
\sup_{1 \leq k \leq n} | \sum_{i=1}^k (X_i - Z_i ) | = o(a_n) \text{ almost surely},
\eeq
where $(a_n)_{n \geq 1}$ is a nondecreasing sequence of positive reals tending to infinity. The first result of this type is due to Strassen (1964) who obtained (\ref{strassen}) with $a_n = (n\log \log n)^{1/2}$. To get smaller $(a_n)$ additional information on the moments of $X_1$ is necessary. If ${\mathbb E}|X_1|^p < \infty$ for $p $ in $]2, 4[$, by using the Skorohod embedding theorem, Breiman (1967) showed that (\ref{strassen}) holds with $a_n = n^{1/p} (\log n)^{1/2}$. He also proved that $a_n =n^{1/p}$ cannot be improved under the $p$-th moment assumption for any $ p >2$. The Breiman paper highlights the fact that there is a gap between the direct result and its converse when using the Skorohod embedding. This gap was later filled by Koml\'os, Major and Tusn\'ady (1976) for $p>3$ and by Major (1976) for $p$ in $]2,3]$: they obtained (\ref{strassen}) with $a_n = n^{1/p}$ as soon as ${\mathbb E}|X_1|^p < \infty$ for any $p>2$, using an explicit construction of the Gaussian random variables, based
on quantile transformations.
There has been a great deal of work to extend these results to dependent sequences: see for instance Philipp and Stout (1975), Berkes and Philipp (1979), Dabrowski (1982), Bradley (1983), Shao (1993), Eberlein (1986), Wu (2007), Zhao and Woodroofe (2008) among others, for extensions of (\ref{strassen}) under various dependence conditions.
In this paper, we are interested in the case of strictly stationary strongly mixing sequences. Recall that the strong mixing coefficient of Rosenblatt (1956) between two $\sigma$-algebras ${\mathcal F}$
and ${\mathcal G}$ is defined by
$$ \alpha({\mathcal F}, {\mathcal G})= \sup_{A \in {\mathcal F}, B \in {\mathcal G}}|{\mathbb P}(A \cap B)-{\mathbb P}(A){\mathbb P}(B)| \, .
$$
For a strictly stationary sequence $(X_i)_{i \in {\mathbb Z}}$ of real valued random variables, and the $\sigma$-algebra ${\mathcal F}_0=\sigma (X_i, i \leq 0)$ and ${\mathcal G}_n
= \sigma (X_i, i \geq n)$, define then
\begin{equation}\label{defalpharosen}
\alpha(0) = 1 \text{ and } \alpha(n)= 2\alpha({\mathcal F}_0,{\mathcal G}_n) \text{ for $n>0$} \, .
\end{equation}
Concerning the extension of (\ref{strassen}) in the strong mixing setting, Rio (1995-a) proved the following: assume that
\beq \label{quantrio}
\sum_{k=0}^\infty \int_0^{\alpha(k)} Q_{|X_0|}^2(u) du < \infty \, ,\eeq
where $Q_{|X_0|}$ is given in Definition \ref{defquant}. Then the series
$\bkE (X_0^2) + 2 \sum_{k \geq 1} \bkE (X_0 X_k)$ is convergent to some nonnegative real $\sigma^2$ and one can construct a sequence $(Z_i)_{i \geq 1}$ of zero mean i.i.d. Gaussian variables with variance $\sigma^2$ such that (\ref{strassen}) holds true with $a_n = ( n \log \log n)^{1/2}$. As shown in Theorem 3 of Rio (1995-a), the condition (\ref{quantrio}) cannot be improved. Recently Dedecker, Gou\"{e}zel and Merlev\`{e}de (2010) proved that this result still holds if we replace the Rosenblatt strong mixing coefficients $\alpha(n)$ by the weaker coefficients defined in (\ref{defalpha}), provided that the underlying sequence is ergodic.
Still in the strong mixing setting, the best extension, up to our knowledge, of the Koml\'os, Major and Tusn\'ady results is due to Shao and Lu (1987). Applying the Skorohod embedding, they obtained the following result (see also
Corollary 9.3.1 in Lin and Lu (1996)): Let $p \in ]2,4]$ and $r >p$. Assume that
\beq \label{Shaocond}
{\mathbb E}(|X_0|^{r})< \infty \quad \text{ and} \quad \sum_{n \geq 1} (\alpha(n))^{(r-p)/(rp)} < \infty \, .
\eeq
Then the series
$\bkE (X_0^2) + 2 \sum_{k \geq 1} \bkE (X_0 X_k)$ is convergent to some nonnegative real $\sigma^2$ and one can construct a sequence $(Z_i)_{i \geq 1}$ of zero mean i.i.d. Gaussian variables with variance $\sigma^2$ such that (\ref{strassen}) holds true with $a_n =n^{1/p}(\log n)^{1 + (1 + \lambda)/p}$, where $\lambda = (\log 2)/\log (r / (r-2))$.
\medskip
Comparing (\ref{Shaocond}) with (\ref{quantrio}) when $p$ is close to $2$, there appears to be a gap between the two above results. A reasonable conjecture is that
Shao and Lu's result still holds under the weaker condition
\beq \label{ros}
{\mathbb E}(|X_0|^{p})< \infty \quad \text{ and} \quad \sum_{k=1}^\infty k^{p-2}\int_0^{\alpha(k)} Q_{|X_0|}^p(u) du < \infty \, ,
\eeq
since the Rosenthal inequality of order $p$ is true under (\ref{ros}) (see Theorem 6.3 in Rio (2000)) and may fail to hold if this condition is not satisfied (see Rio (2000), chapter 9). To compare (\ref{ros}) with (\ref{Shaocond}), note that (\ref{ros}) is implied by: for $r>p$,
$$
\sup_{x >0} x^r {\mathbb P}(|X_0| > x) < \infty \quad \text{ and } \quad \sum_{n=1}^\infty
n^{p-2} (\alpha(n) )^{(r-p)/r} < \infty\, ,
$$
which is much weaker than (\ref{Shaocond}). For example, in the case of bounded random variables ($r = \infty$), (\ref{Shaocond}) needs $\alpha (n) = O (n^{-p})$,
while (\ref{ros}) holds as soon as $\alpha (n) = O (n^{1-p} (\log n)^{-1-\eps})$ for some positive $\eps$.
Let us now give an outline of our results and methods of proofs. Our main result is Theorem \ref{ThSM}, which ensures in particular that, for $p \in ]2,3[$, (\ref{strassen}) holds for $a_n = n^{1/p} (\log n)^{1/2-1/p}$ under (\ref{ros}). Furthermore the error in ${\mathbb L}^2$ is of the same order. The proof of our Theorem \ref{ThSM} is based on an explicit construction of the approximating sequence of i.i.d. Gaussian random variables with the help of conditional quantile transformations. From our construction, the ${\mathbb L}^2$ approximating error
between dyadic blocks of the initial sequence and the gaussian one can be handled with the help of a conditional version of a functional inequality due to Rio (1998), linking the Wasserstein distance $W_2$ with the Zolotarev distance $\zeta_2$ (see our Proposition \ref{condversion}).
This method allows us to get a smaller logarithmic factor than the extra factor $(\log n)^{1/2}$ induced by the Skorohod embedding.
Moreover, it is possible to adapt it (by conditioning up to the future rather than to the past) to deal with the partial sums of non necessarily bounded functions $f$ of iterates of expanding maps such as those considered in Section \ref{dynsys}. For such maps, Theorem \ref{ASmap} completes results obtained by Melbourne and Nicol (2005, 2009) when $f$ is H\"{o}lder continuous. The rest of the paper is organized as follows: Section \ref{proofs} is devoted to the proof of the main results whereas the technical tools are stated and proven in Appendix.
\section{Definitions and main result}\label{MR} Let $(\Omega,{\cal A}, {\mathbb P} )$ be a
probability space. Assume that there exists some strictly stationary sequence
$(Y_i)_{i \in {\mathbb Z}}$ of real valued random variables on this probability space, and that the probability space
$(\Omega,{\cal A}, {\mathbb P} )$ is large enough to contain a sequence
$(\delta_i)_{i \in {\mathbb Z}}$ of independent random variables with uniform distribution
over $[0,1]$, independent of $(Y_i)_{i \in {\mathbb Z}}$. Define the nondecreasing
filtration $({\cal F}_i)_{i \in {\mathbb Z}}$ by
${\cal F}_i = \sigma ( (Y_k, \delta_k) : k \leq i)$. Let ${\cal {F}}_{-\infty} = \bigcap_{i \in {\mathbb Z}} {\cal {F}}_{i}$ and ${\cal {F}}_{\infty} = \bigvee_{i \in
{\mathbb Z}} {\cal {F}}_{i}$. We shall denote sometimes by ${\mathbb
E}_i$ the conditional expectation with respect to ${\mathcal F}_i$.
\setcounter{equation}{0} In this section we give rates of
convergence in the almost sure and ${\mathbb L}^2$ invariance principle for functions of a stationary
sequence $(Y_i)_{i \in {\mathbb Z}}$ satisfying weak dependence conditions that we specify below.
\begin{definition} \label{defquant}
For any nonnegative random variable
$X$, define the ``upper tail'' quantile function $Q_X $ by $
Q_X (u) = \inf \left \{ t \geq 0 : \p \left(X >t \right) \leq
u\right \} $.
\end{definition}
This function is defined on $[0,1]$, non-increasing, right
continuous, and has the same distribution as $X$. This makes it
very convenient to express the tail properties of $X$ using
$Q_X$. For instance, for $0<\varepsilon<1$, if the distribution
of $X$ has no atom at $Q_X(\varepsilon)$, then
\begin{equation*}
\bkE ( X \I_{X > Q_X(\varepsilon)})=\sup_{\p(A)\leq \varepsilon} \bkE(X \I_A) = \int_0^\varepsilon Q_X(u) du\, .
\end{equation*}
\begin{definition}
\label{defclosedenv} Let $\mu$ be the probability distribution of a
random variable $X$. If $Q$ is an integrable quantile function,
let $\tMon( Q, \mu)$ be the set of functions $g$ which are
monotonic on some open interval of ${\mathbb R}$ and null
elsewhere and such that $Q_{|g(X)|} \leq Q$. Let $\widetilde{\mathcal{F}}( Q,
\mu)$ be the closure in ${\mathbb L}^1(\mu)$ of the set of
functions which can be written as $\sum_{\ell=1}^{L} a_\ell
f_\ell$, where $\sum_{\ell=1}^{L} |a_\ell| \leq 1$ and $f_\ell$
belongs to $\tMon( Q, \mu)$.
\end{definition}
\begin{definition}
For any integrable random variable $X$, let us write
$X^{(0)}=X- \bkE(X)$.
For any random variable $Y=(Y_1, \cdots, Y_k)$ with values in
${\mathbb R}^k$ and any $\sigma$-algebra $\F$, let
\[
\alpha(\F, Y)= \sup_{(x_1, \ldots , x_k) \in {\mathbb R}^k}
\left \| \bkE \Big(\prod_{j=1}^k (\I_{Y_j \leq x_j})^{(0)} \Big | \F \Big)^{(0)} \right\|_1.
\]
For the sequence ${\bf Y}=(Y_i)_{i \in {\mathbb Z}}$, let \begin{equation}
\label{defalpha} \alpha_{k, {\bf Y}}(0) =1 \text{ and }\alpha_{k, {\bf Y}}(n) = \max_{1 \leq l \leq
k} \ \sup_{ n\leq i_1\leq \ldots \leq i_l} \alpha(\F_0,
(Y_{i_1}, \ldots, Y_{i_l})) \text{ for $n>0$}.
\end{equation}
\end{definition}
For any positive $n$, $\alpha_{k, {\bf Y}}(n) \leq \alpha (n)$, where $\alpha (n)$ is defined by
(\ref{defalpharosen}). We now introduce some quantities involving the rate of
mixing and the quantile function $Q$. Define
\begin{equation}
\label{definitionR}
\alpha^{-1}_{2,{\bf Y}}(x) = \min\{q\in \BBN \,:\, \alpha_{2,{\bf Y}}(q)\leq x\}
\ \text{ and }\
R(x)=\alpha^{-1}_{2,{\bf Y}}(x)( Q(x) \vee 1)
\end{equation}
(note that $\alpha^{-1}_{2,{\bf Y}}(x) \geq 1$ for $x < 1$). Set, for $p\geq 1$,
\begin{equation}
\label{defmoments}
M_{p,\alpha} (Q) = \int_0^1 R^{p-1}(u) Q(u) du \ \text{ and }\
\Lambda_{p,\alpha} (Q) = \sup_{u\in]0,1]} u R^{p-1} (u) Q(u) .
\end{equation}
Note that, if $M_{p,\alpha} (Q) <\infty$ then $\Lambda_{p,\alpha} (Q) < \infty$,
Also, if $\Lambda_{p,\alpha} (Q) < \infty$, then $M_{r,\alpha} (Q) <\infty$ for
any $r<p$. Let us now state our main result.
\begin{Theorem} \label{ThSM} Let $X_i = f(Y_i) - \bkE ( f(Y_i))$ where $f$ belongs to $\widetilde{\mathcal{F}}( Q, P_{Y_0})$ (here $P_{Y_0}$ denotes the law of $Y_0$).
Assume that $M_{2,\alpha} (Q) < \infty$. Then the series
$\bkE (X_0^2) + 2 \sum_{k \geq 1} \bkE (X_0 X_k)$ is convergent to some nonnegative real $\sigma^2$.
Now let $p \in ]2,3]$ and suppose that $\Lambda_{p,\alpha} (Q) < \infty$ in the case $p<3$ or $M_{3,\alpha} (Q) < \infty$ in the case $p=3$.
\begin{enumerate}
\item
Assume that $\sigma^2 >0$. Then:
\begin{enumerate}
\item
there exists a sequence $(Z_i)_{i \geq 1}$ of iid random variables with law $N(0,\sigma^2)$
such that, setting $\Delta_k = \sum_{i=1}^k (X_i - Z_i)$,
$$
\sup_{k \leq n} |\Delta_k| = O ( n^{1/p} (\log n)^{1/2-1/p}) \text{ in }{\mathbb L}^2
\text{ and a.s. for $p< 3$ if } M_{p,\alpha} (Q) < \infty .
$$
\item For any $\varepsilon >0$, there exists a sequence
$(\widetilde Z_i)_{i\geq 1}$ of iid random variables with law $N(0,\sigma^2)$ such that,
setting $\widetilde \Delta_k = \sum_{i=1}^k (X_i -\widetilde Z_i)$,
$$
\sup_{k \leq n} |\widetilde \Delta_k| =
O ( n^{1/p} (\log n)^{1/2}(\log \log n)^{(1+ \varepsilon )/p}) \text{ a.s.}
$$
\end{enumerate}
\item Assume that $\sigma^2 =0$. Let $S_k = \sum_{i=1}^k X_i$.
Then
\begin{enumerate}
\item $\sup_{k \leq n} | S_k | = O ( n^{1/p} )
\text{ in }{\mathbb L}^2 \text{ and } \sup_{k \leq n} | S_k | = O ( (n\log n)^{1/p}
(\log\log n)^{(1+\varepsilon)/p} ) \text{ a.s.}$
\item If $p<3$ and $M_{p,\alpha} (Q) < \infty$, then
$\sup_{k \leq n} | S_k | = o ( n^{1/p} )$ a.s.
\end{enumerate}
\end{enumerate}
\end{Theorem}
\begin{Remark} \label{remarkequicond} The condition $M_{p,\alpha} (Q) < \infty$ can be rewritten in a complete equivalent way as
\beq \label{Condstrong}
\sum_{k\geq 0} (1\vee k)^{p-2} \int_0^{\alpha_{2, {\bf Y}}(k)} Q^p(u) du
< \infty \, .\eeq
(see Annexe C in Rio (2000)), which corresponds to (\ref{ros}) with $\alpha_{2,{\bf Y}} (k)$
instead of $\alpha (k)$.
\end{Remark}
\noindent
{\sl Applications to geometric or arithmetic rates of mixing. } Below we denote by $H$ the cadlag
inverse of the function $Q$. Assume first that, for some $a$ in $]0,1[$,
$\alpha_{2 , {\bf Y}} (n) = O ( a^n)$ as $n \rightarrow \infty$.
Then $\alpha_{2 , {\bf Y}}^{-1} (u) = O ( |\log u| )$ as $u$ decreases to $0$.
Consequently $M_{p,\alpha} ( Q ) < \infty$ as soon as
$$
\int_0^1 |\log u|^{p-1} Q^p (u) du < \infty .
$$
This condition holds if
$H(x) = O (\, (x\log x)^{-p} (\log\log x)^{-(1+\eps)} )$ as $x \rightarrow \infty$.
In a similar way $\Lambda_{p,\alpha} (Q) < \infty$ if one of the following equivalent weaker conditions holds:
\begin{eqnarray*}
Q (u) = O ( u^{-1/p} |\log u|^{-1+ (1/p)} ) \hbox{ as } u\downarrow 0 \, ,
\
H(x) = O (\, x^{-p} (\log x)^{1-p} ) \hbox{ as } x\uparrow \infty .
\end{eqnarray*}
\par
Suppose now that, for some real $q>2$, $\alpha_{2 , {\bf Y}} (n) = O ( n^{1-q} )$ as $n \rightarrow \infty$.
Then $\alpha_{2 , {\bf Y}}^{-1} (u) = O ( u^{-1/(q-1)} )$ as $u\rightarrow 0$. For $p$ in $[2,q[$,
we get that $M_{p,\alpha} ( Q ) < \infty$ as soon as
$$
\int_0^1 |u|^{-1/(q-1)} Q^p (u) du < \infty .
$$
This condition holds if $H(x) = O (\, (x^p \log (x) (\log\log x)^{1+\eps} )^{-(q-1)/(q-p)} )$ as $x\rightarrow \infty$.
In a similar way $\Lambda_{p,\alpha} (Q) < \infty$ if and only if
$H(x) = O (\, x^{-p(q-1)/(q-p)} )$ as $x \rightarrow \infty$.
Note also that $\Lambda_{q,\alpha} (Q) < \infty$ if and only if $Q$ is uniformly bounded over $]0,1]$.
\par
\section{Application to dynamical systems} \label{dynsys}
\setcounter{equation}{0}
In this section, we consider a class of piecewise expanding
maps $T$ of $[0,1]$ with a neutral fixed point, and their
associated Markov chain $Y_i$ whose transition kernel is the
Perron-Frobenius operator of $T$ with respect to the absolutely
continuous invariant probability measure. Applying Theorem \ref{ThSM}, we give a large class
of unbounded functions $f$ for which we can give rates of convergence close to optimal in the strong invariance principle of the partial sums of
both $f\circ T^i$ and $f(Y_i)$.
For $\gamma$ in $]0, 1[$, we consider the intermittent map
$T_\gamma$ from $[0, 1]$ to $[0, 1]$, which is a modification of the
Pomeau-Manneville map (1980):
$$
T_\gamma(x)=
\begin{cases}
x(1+ 2^\gamma x^\gamma) \quad \text{ if $x \in [0, 1/2[$}\\
2x-1 \quad \quad \quad \ \ \text{if $x \in [1/2, 1]$} \, .
\end{cases}
$$
We denote by $\nu_\gamma$ the unique $T_\gamma$-invariant
probability measure on $[0, 1]$ which is absolutely continuous with
respect to the Lebesgue measure. We denote by $K_\gamma$ the
Perron-Frobenius operator of $T_\gamma$ with respect to
$\nu_\gamma$. Recall that for any bounded measurable functions $f$ and $ g$,
$$
\nu_\gamma(f \cdot g\circ T_\gamma)=\nu_\gamma(K_\gamma(f) g) \, .
$$
Let $(Y_i)_{i \geq 0}$ be a stationary Markov chain with invariant
measure $\nu_\gamma$ and transition Kernel $K_\gamma$. It is well
known (see for instance Lemma XI.3 in Hennion and Herv\'e (2001))
that on the probability space $([0, 1], \nu_\gamma)$, the random
variable $(T_\gamma, T^2_\gamma, \ldots , T^n_\gamma)$ is
distributed as $(Y_n,Y_{n-1}, \ldots, Y_1)$.
To state our results for those intermittent maps, we need preliminary definitions.
\begin{Definition}
A function $H$ from ${\mathbb R}_+$ to $[0, 1]$ is a tail
function if it is non-increasing, right continuous, converges
to zero at infinity, and $x\rightarrow x H(x)$ is integrable.
\end{Definition}
\begin{Definition}
\label{defMonGPM}
If $\mu$ is a probability measure on $\mathbb R$ and $H$ is a
tail function, let $\Mon(H, \mu)$ denote the set of functions
$f:{\mathbb R}\to {\mathbb R}$ which are monotonic on some open interval and null
elsewhere and such that $\mu(|f|>t)\leq H(t)$. Let $\mathcal{F}(H,
\mu)$ be the closure in ${\mathbb L}^1(\mu)$ of the set of
functions which can be written as $\sum_{\ell=1}^L a_\ell
f_\ell$, where $\sum_{\ell=1}^L |a_\ell| \leq 1$ and $f_\ell\in
\Mon(H, \mu)$.
\end{Definition}
Note that a function belonging to $\mathcal{F}(H, \mu)$ is allowed to
blow up at an infinite number of points. Note also that any
function $f$ with bounded variation ($\BV$) such that $|f|\leq
M_1$ and $\|df\|\leq M_2$ belongs to the class $\mathcal{F}(H, \mu)$
for any $\mu$ and the tail function $H=\I_{[0, M_1+2M_2)}$
(here and henceforth, $\|df\|$ denotes the variation norm of
the signed measure $df$).
In the unbounded case, if a function $f$ is piecewise monotonic with $N$ branches, then
it belongs to $\mathcal{F}(H, \mu)$ for $H(t)=\mu(|f|>t/N)$. Finally, let us
emphasize that there is no requirement on the modulus of
continuity for functions in $\mathcal{F}(H, \mu)$.
Let $Q$ denote the cadlag inverse of $H$. Then, for the random variable $X$ defined by $X(\omega) =
\omega$, $\Mon(H, \mu) = \tMon( Q, \mu)$ and $\mathcal{F}(H, \mu) = \widetilde{\mathcal{F}} ( Q, \mu)$. Furthermore
Proposition 1.17 in Dedecker, Gou\"{e}zel and
Merlev\`{e}de (2010) states that there exists a positive constant $C$ such
that, for any $n>0$, $\alpha_{2,{\bf Y}}(n) \leq C n^{(\gamma -1)/\gamma}$. In addition,
the computations page 817 in the same paper show that, for $p\gamma<1$, the integrability conditions below are equivalent:
\begin{equation} \label{ratecond}
\int_0^1 R^{p-1}(u) Q(u) du < \infty \ \hbox{ and }\ \int_0^{\infty} x^{p-1} (H(x))^{\frac{1-p\gamma}{1-\gamma}} dx < \infty \, .
\end{equation}
Also, for $p$ in $]2, 1/\gamma[$,
\begin{equation} \label{ratecond2}
\Lambda_{p,\alpha} (Q) < \infty \ \hbox{if and only if}\ H(x) = O (x^{-p(1-\gamma)/(1-p\gamma)} )
\ \hbox{as}\ x\rightarrow\infty \,
\end{equation}
and, for $p= 1/\gamma$ and $H =\I_{[0, M)}$, $\Lambda_{p,\alpha} (Q) <\infty$ (see the previous section).
A modification of the proof of Theorem \ref{ThSM} leads to the result below for the
Markov chain or the dynamical system associated to the transformation $T_\gamma$.
\begin{Theorem} \label{ASmap} Let $\gamma <1/2$.
Let $f \in {\mathcal F}(H, \nu_\gamma)$ for some tail function $H$ satisfying (\ref{ratecond})
with $p=2$. Then the series
\begin{equation}\label{seriecov}
\sigma^2(f)= \nu_\gamma((f-\nu_\gamma(f))^2)+ 2
\sum_{k>0} \nu_\gamma ((f-\nu_\gamma(f))f\circ T_\gamma^k)
\end{equation}
converges absolutely to some nonnegative number $\sigma^2 (f)$. Let $p \in ]2,3]$ satisfying
$p\leq 1/\gamma$. Let $Q$ denote the cadlag inverse of $H$. Suppose that
$\Lambda_{p,\alpha} (Q) < \infty$ in the case $p<3$ or $M_{3,\alpha} (Q) < \infty$ in the case $p=3$.
\begin{enumerate}
\item Let $(Y_i)_{i \geq 1}$ be a stationary Markov chain with
transition kernel $K_{\gamma}$ and invariant measure $\nu_\gamma$, and let
$X_i=f(Y_i)-\nu_\gamma(f)$. The sequence $(X_i)_{i \geq 0}$ satisfies the conclusions of Items 1 and 2 of Theorem \ref{ThSM} with $\sigma^2 = \sigma^2(f)$.
\item If $\sigma^2(f) = 0$, the sequence $(f\circ T_\gamma^i- \nu_\gamma(f) )_{i \geq 1}$ satisfies the conclusions of Item 2 of Theorem \ref{ThSM}. If $\sigma^2(f) > 0$, enlarging the probability space $([0, 1], \nu_\gamma)$, there exist sequences $(Z^*_i)_{i \geq
1}$ and $(\tilde Z^*_i)_{i \geq
1}$ of iid random variables with law $N( 0, \sigma^2 (f) )$ such that the random variables
$\Delta_k = \sum_{i=1}^k (f\circ T_\gamma^i- \nu_\gamma(f) - Z_i^* )$
satisfy the conclusions of Item 1(a) of Theorem \ref{ThSM} and the random variables
$\tilde \Delta_k =\sum_{i=1}^k (f\circ T_\gamma^i- \nu_\gamma(f) - \tilde Z_i^* )$ satisfy the conclusion of Item 1(b).
\end{enumerate}
\end{Theorem}
Item 1 is direct by using Theorem \ref{ThSM} together with (\ref{ratecond}) and (\ref{ratecond2}). Item 2 requires a proof that is given in Section \ref{proofASmap}.
\begin{Remark}
Theorem \ref{ASmap} can be extended to generalized Pomeau-Manneville
map (or GPM map) of parameter $\gamma \in (0,1)$ as defined in Dedecker, Gou\"{e}zel and
Merlev\`{e}de (2010).
\end{Remark}
In the specific case of bounded variation functions, Theorem \ref{ASmap} provides the almost sure
invariance principle below for the dynamical system associated to $T_\gamma$. Below we give the results
in the case $\sigma^2 (f)>0$. The rates are slightly better in the case $\sigma^2 (f) = 0$.
\begin{Corollary} \label{ASmapB} Let $\gamma \in ]1/3 , 1/2[$ and $f$ be
a function of bounded variation. Then the series in (\ref{seriecov})
converges absolutely to some nonnegative number $\sigma^2 (f)$ and, for any $\varepsilon >0$,
there exists a sequence
$(\widetilde Z_i^*)_{i\geq 1}$ of iid random variables with law
$N(0,\sigma^2 (f) )$ such that
$$
\sup_{k \leq n} | \sum_{i=1}^k (f\circ T_\gamma^i- \nu_\gamma(f) - \tilde Z_i^* ) | =
O ( n^{\gamma} (\log n)^{1/2}(\log \log n)^{(1+ \varepsilon )\gamma}) \text{ a.s.}
$$
\end{Corollary}
For the maps under consideration and H\"older continuous functions $f$, by using an approximation argument introduced by Berkes and Philipp (1979),
Melbourne and Nicol (2009) obtained the following explicit error term in the almost sure invariance principle (see their Theorem 1.6 and their Remark 1.7): Let $p >2$ and $0<\gamma< 1/p$, then the error term in the almost sure invariance results is $O(n^{\beta + \varepsilon})$ where $\varepsilon >0$ is arbitrarily small and $\beta = \frac{\gamma}{2} + \frac{1}{4}$ if $\gamma$ belongs to $]1/4, 1/2[$
and $\beta = \frac{3}{8} $ if $\gamma \leq 1/4$. Consequently, for the modification of the
Pomeau-Manneville map and functions $f$ of bounded variation, Corollary \ref{ASmapB} improves
the error in the almost sure invariance principle obtained in Theorem 1.6 in Melbourne and Nicol (2009). Note also that, for $\gamma < 1/3$ and $f$
of bounded variation, condition (\ref{ratecond}) is satisfied with $p=3$, and Theorem \ref{ASmap}
gives the error term $O(n^{1/3} (\log n)^{1/2} (\log\log n)^{(1+ \varepsilon )/3})$
in the almost sure invariance principle.
\section{Proofs} \label{proofs}
\setcounter{equation}{0}
From now on, we denote by $C$ a numerical constant which may vary
from line to line. Throughout the proofs, to shorten the notations, we write
$\alpha (n) = \alpha_{2, {\bf Y}} (n)$ and $\alpha^{-1} (u) = \alpha_{2, {\bf Y}}^{-1} (u)$.
We also set, for $\lambda >0$,
\begin{equation}\label{3alphatronq}
M_{3,\alpha} (Q, \lambda) = \int_0^1 Q(u) R(u) (R(u) \wedge \lambda) du .
\end{equation}
We start by recalling some fact proved in Rio (1995-b), Lemma A.1.: for $p$ in $]2,3[$,
\begin{equation} \label{RioA1}
M_{3,\alpha} (Q, \lambda) = O ( \lambda^{3-p} ) \ \hbox{ as }\ \lambda \rightarrow +\infty
\ \hbox{ if }\ \Lambda_{p,\alpha} (Q) < \infty .
\end{equation}
\subsection{Proof of Theorem \ref{ThSM}.}
Assume first that $\sigma^2 >0$. For $L \in {\mathbb N}$, let $m(L) \in {\mathbb N}$ be such that $m(L)\leq L$. Let
\begin{equation*} \label{defUkL}
I_{k,L} = ]2^L + (k-1)2^{m(L)} , 2^L + k 2^{m(L)}] \cap {\mathbb N} \ \text{and}\ U_{k,L} = \sum_{i\in I_{k,L}} X_i \, ,
\, k \in \{1, \cdots, 2^{L-m(L)} \} \, .
\end{equation*}
For $k \in \{1, \cdots, 2^{L-m(L)} \} $, let $V_{k,L}$ be the
${\mathcal N}(0, \sigma^2 2^{m(L)} )$-distributed random variable
defined from $U_{k,L}$ via the conditional quantile transformation,
that is \beq \label{defVkN}
V_{k,L} = \sigma 2^{m(L)/2} \Phi^{-1}
(\widetilde F_{k,L}( U_{k,L}- 0 ) + \delta_{2^L + k 2^{m(L)}} (\widetilde
F_{k,L}( U_{k,L} ) - \widetilde F_{k,L} (U_{k,L} - 0) ) ) \, ,
\eeq where
$\widetilde F_{k,L}:= F_{U_{k,L} | {\mathcal F}_{2^{L}+ (k-1)2^{m(L)}}}$ is the d.f. of
$P_{U_{k,L} | {\mathcal F}_{2^L + (k-1)2^{m(L)}
}}$ (the conditional law of $U_{k,L}$ given $\mathcal{F}_{2^L +
(k-1)2^{m(L)} }$) and $\Phi^{-1}$ the inverse of the
standard Gaussian distribution function $\Phi$. Since $\delta_{2^L + k 2^{m(L)}}$ is
independent of ${\mathcal F}_{2^L + (k-1)2^{m(L)}}$, the random variable $V_{k,L}$
is independent of ${\mathcal F}_{2^L + (k-1)2^{m(L)}}$, and has the Gaussian distribution
$N (0, \sigma^2 2^{m(L)})$. By induction on $k$, the random variables $(V_{k,L})_k$
are mutually independent and independent of ${\mathcal F}_{2^L}$.
In addition
\begin{eqnarray*} \label{transquantcond}\bkE
(U_{k,L} - V_{k,L})^2 & = & \bkE \int_0^1 \big ( F^{-1}_{U_{k,L} |
{\mathcal F}_{2^L + (k-1)2^{m(L)} }}(u)- \sigma 2^{m(L)/2}\Phi^{-1}(u)
\big )^2 du \\ & : = & \bkE \big ( W_2^2 (P_{U_{k,L} | {\mathcal
F}_{2^L + (k-1)2^{m(L)} }} , G_{\sigma^2 2^{m(L)} }) \big ) \, ,
\end{eqnarray*}
where $G_{\sigma^2 2^{m(L)} }$ is the Gaussian distribution
$N (0, \sigma^2 2^{m(L)})$. Using Proposition
\ref{condversion} and stationarity, we then get that there exists a positive constant
$C$ such that
\beq\label{majw2cond} \bkE (U_{k,L} - V_{k,L})^2 \leq
C 2^{m(L)/2} M_{3,\alpha} (Q ,2^{m(L)/2}) \, .
\eeq
\par
Now we construct a
sequence $(Z'_i)_{i \geq 1}$ of i.i.d. Gaussian random variables with
zero mean and variance $\sigma^2$ as follows. Let $Z'_1 = \sigma \Phi^{-1} (\delta_1)$. For any
$L \in {\mathbb N}$ and any $k \in \{1, \cdots, 2^{L -m(L)} \}$
the random variables $(Z'_{2^L + (k-1)2^{m(L)} +1}, \ldots , Z'_{2^L + k 2^{m(L)}})$
are defined in the following way. If $m(L)=0$, then
$Z'_{2^L + k 2^{m(L)}} = V_{k,L}$.
If $m(L)>0$, then by the Skorohod lemma (1976), there exists some measurable function $g$ from ${\mathbb R} \times [0,1]$
in ${\mathbb R}^{2^{m(L)}}$ such that, for any pair
$(V,\delta)$ of independent random variables with respective laws $N(0,\sigma^2 2^{m(L)})$ and
the uniform distribution over $[0,1]$, $g(V,\delta) = (N_1, \ldots N_{2^{m(L)}})$ is a Gaussian random vector with i.i.d. components such that $V = N_1 + \cdots + N_{2^{m(L)}}$.
We then set
$$
(Z'_{2^L + (k-1)2^{m(L)} +1}, \ldots , Z'_{2^L + k 2^{m(L)}}) = g (V_{k,L} , \delta_{2^L + (k-1)2^{m(L)} +1} ).
$$
The so defined sequence $(Z'_i)$ has the prescribed distribution.
\par\medskip
Set $S_j = \sum_{i=1}^j X_i$ and $T_j = \sum_{i=1}^j Z'_i$. Let
\begin{eqnarray*} \label{0dec}
D_L := \sup_{ \ell \leq 2^{L}} | \sum_{i = 2^L +1}^{2^L + \ell} (X_i -Z'_i)| \, .
\end{eqnarray*}
Let $N\in {\BBN}^*$ and let $k \in ]1, 2^{N+1}]$.
We first notice that $D_L \geq | (S_{2^{L + 1} } - T_{2^{L + 1} }) - ( S_{2^L } - T_{2^L } )|$,
so that, if $K$ is the integer such that $2^K < k \leq 2^{K+1}$, $|S_k - T_k| \leq |X_1 - Z'_1| + D_0 + D_1 + \cdots + D_K$.
Consequently since $K \leq N$,
\begin{eqnarray} \label{1dec} \sup_{1 \leq k \leq 2^{N+1} }|S_k -T_k| \leq
|X_1 - Z'_1| + D_0 + D_1 + \cdots + D_N .
\end{eqnarray}
\par
We first notice that the following decomposition is valid:
\beq \label{decsup}
D_L \leq D_{L,1} + D_{L,2} \, ,
\eeq
where
$$
D_{L,1}:= \sup_{k \leq 2^{L-m(L)} } \Big|
\sum_{ \ell =1}^k (U_{\ell,L} - V_{\ell,L}) \Big|
\ \text{and}\
D_{L,2}:= \sup_{ k \leq 2^{L-m(L)}} \sup_{ \ell \in I_{k,L} }
\Big| \sum_{i = \inf I_{k,L} }^{\ell} (X_i -Z_i) \Big| .
$$
The main tools for proving Theorem \ref{ThSM} will be the two lemmas below.
The first lemma allows us to control the fluctuation term $D_{L,2}$.
\begin{Lemma} \label{LemDL2} There exists positive constants $c_1$, $c_2 \geq 2$, $c_3$ and $c_4$ such that,
for any positive $\lambda$,
\begin{equation}
\BBP ( D_{L,2} \geq 2\lambda ) \leq (c_1+2) 2^L \exp \Bigl( - \frac{\lambda^2}{c_2 \sigma^2
2^{m(L)}} \Bigr) + 2^L\lambda^{-3} \bigl( c_3
M_{3,\alpha} (Q,\lambda) + c_4 \sigma^3 \bigr) \, . \label{apppropineg}
\end{equation}
\end{Lemma}
The second lemma gives a bound in ${\mathbb L}^2$ on the Gaussian approximation term $D_{L,1}$.
\begin{Lemma} \label{LemDL1} Let $p\in ]2,3]$. Suppose that $\Lambda_{p,\alpha} (Q) < \infty$
in the case $p<3$ and $M_{3,\alpha} (Q) < \infty$ in the case $p=3$. Then
\begin{equation} \label{DL1borneL2}
\|D_{L,1}\|^2_2 \leq C 2^L \Bigl ( 2^{(2-p)m(L)} + 2^{-m(L)/2} M_{3,\alpha} (Q , 2^{m(L)/2}) \Bigr) \, .
\end{equation}
\end{Lemma}
\noindent
{\bf Proof of Lemma \ref{LemDL2}.} By the triangle inequality together with the stationarity of the sequences $(X_i)_i$ and $(Z_i)_i$, for any positive $\lambda$,
\beq \label{decdl2} {\mathbb P}
( D_{L,2} \geq 2 \lambda ) \leq 2^{L-m(L)} {\mathbb P} \Bigl( \sup_{
\ell \leq 2^{m(L)}} | S_{\ell} | \geq \lambda \Bigr) + 2^{L-m(L)} {\mathbb P}
\Bigl( \sup_{\ell \leq 2^{m(L)}} | T_{\ell} | \geq \lambda
\Bigr) \, .\eeq
By L\'evy's inequality (see for instance Proposition 2.3 in Ledoux and Talagrand (1991)),
\begin{equation} {\mathbb P} \Bigl( \sup_{
\ell\leq 2^{m(L)}} | T_{\ell} | \geq \lambda
\Bigr) \leq 2 \exp \Bigl( - \frac{\lambda^2}{2 \sigma^2 2^{ m(L)}
} \Bigr) \, . \label{LI}
\end{equation}
On the other hand, applying Proposition \ref{inegamax}, we get that
$$
\BBP \Bigl( \sup_{ \ell \leq 2^{m(L)}} | S_{\ell} | \geq \lambda \Bigr) \leq
c_1\exp \Bigl( - \frac{\lambda^2}{c_2 \sigma^2
2^{m(L)}} \Bigr) + 2^{m(L)}\lambda^{-3} \bigl( c_3
M_{3,\alpha} (Q,\lambda) + c_4 \sigma^3 \bigr) \, .
$$
Collecting the above inequalities, we then get Lemma \ref{LemDL2}.
\par\medskip\noindent
{\bf Proof of Lemma \ref{LemDL1}.} For any $\ell \in \{ 1, \cdots, 2^{L-m(L)} \} $, let
$\widetilde U_{\ell,L} =U_{\ell,L} - \bkE_{2^L + (\ell-1)2^{m(L)}} (U_{\ell,L})$.
Then $(\widetilde U_{\ell,L})_{\ell \geq 1}$ is a strictly
stationary sequence of martingale differences adapted to the
filtration $({\mathcal F}_{2^L + \ell 2^{m(L)}})_{\ell \geq 1}$. Notice first that
\beq \label{dec0item2}
\|D_{L,1}\|_2 \leq \Vert \sup_{k \leq 2^{L-m(L)}} | \sum_{ \ell =1}^k (\widetilde
U_{\ell,L} - V_{\ell,L}) | \Vert_2 +\Vert \sup_{k \leq 2^{L-m(L)}} | \sum_{ \ell =1}^k (\widetilde
U_{\ell,L} - U_{\ell,L}) | \Vert_2 \, .
\eeq
Let us deal with the first term on right hand. Since $V_{\ell,L}$ is independent of
${\mathcal F}_{2^L + (\ell - 1) 2^{m(L)} }$, the sequence $(\widetilde
U_{\ell,L} - V_{\ell,L})_\ell$ is a martingale difference sequence with respect to the
nondecreasing filtration $({\mathcal F}_{2^L + \ell 2^{m(L)}})_\ell$. Hence, by the
Doob-Kolmogorov maximal inequality, we get that
\begin{eqnarray*}
\Vert \sup_{k \leq 2^{L-m(L)}} \big | \sum_{ \ell =1}^k (\widetilde
U_{\ell,L} - V_{\ell,L}) \big | \Vert^2_2 & \leq & 4
\sum_{\ell=1}^{2^{L-m(L)}} \Vert \tilde U_{\ell, L} - V_{\ell, L}
\Vert^2_2 \\
& \leq & 8 \sum_{\ell=1}^{2^{L-m(L)}} \Vert \widetilde
U_{\ell,L} - U_{\ell,L} \Vert^2_2 +8 \sum_{\ell=1}^{2^{L-m(L)}}
\Vert U_{\ell, L} - V_{\ell, L}\Vert^2_2 \, .
\end{eqnarray*}
Since $V_{\ell,N}$ is independent of $\mathcal{F}_{2^L +
(\ell-1)2^{m(L)} }$, $\bkE_{2^L + (\ell-1)2^{m(L)}} (V_{\ell,L}) =0$.
Consequently,
$$ \Vert \tilde U_{\ell, L} - U_{\ell, L} \Vert^2_2 = \Vert \bkE_{2^L + (\ell-1)2^{m(L)}} (U_{\ell, L} - V_{\ell,L}) \Vert_2^2
\leq \Vert U_{\ell, L} - V_{\ell,L}\Vert_2^2 \, . $$ Using (\ref{majw2cond}), it follows that
\beq \label{maj5DR}
\Vert \sup_{k \leq 2^{L-m(L)}} | \sum_{ \ell =1}^k (\widetilde
U_{\ell,L} - V_{\ell,L}) | \Vert_2^2 \leq C 2^{L-m(L)/2} M_{3,\alpha} (Q, 2^{m(L)/2} )\, .
\eeq
We deal now with the second term in the right hand side of (\ref{dec0item2}). According to Dedecker and Rio's maximal inequality (2000, Proposition 1), we obtain that
\begin{eqnarray} \label{inegadedrio}
& & \Vert \sup_{k \leq 2^{L-m(L)}} | \sum_{ \ell =1}^k (\widetilde
U_{\ell,L} - U_{\ell,L}) | \Vert_2^2 \leq 4 \sum_{k=1}^{2^{L-m(L)}} \Vert\bkE_{2^L + (k-1)2^{m(L)}} (U_{k,L}) \Vert_2^2
\nonumber \\
& & \quad \quad + 8 \sum_{k=1}^{2^{L-m(L)} -1 }
\Vert\bkE_{2^L + (k-1)2^{m(L)}} (U_{k,L}) \big ( \sum_{i= k +1}^{2^{L-m(L)}}\bkE_{2^L + (k-1)2^{m(L)}} (U_{i,L}) \big ) \Vert_1 \, .
\end{eqnarray}
Stationarity leads to
\beq \label{maj1DR}
\Vert\bkE_{2^L + (k-1)2^{m(L)}} (U_{k,L}) \Vert_2^2= \Vert\bkE_{0} (S_{2^{m(L)}}) \Vert_2^2 \leq 2 \sum_{i=1}^{2^{m(L)}} \sum_{j=1}^i {\mathbb E} | X_j {\mathbb E}_0 (X_i)| \, .
\eeq
Using Lemma 4 (page 679) in Merlev\`ede and Peligrad (2006), we get that
$$
{\mathbb E} | X_j {\mathbb E}_0 (X_i)| \leq 3 \int_{0}^{\Vert{\mathbb E}_0 (X_i) \Vert_1} Q_{|X_0|} \circ G_{|X_0|} (u) du \, ,
$$
where $G_{|X_0|}$ is the inverse of $L_{|X_0|}(x) =
\int_0^{x} Q_{|X_0|} (u) du $. We will denote by $L$ and $G$
the same functions constructed from $Q$. Assume first that $X_i = f(Y_i)-\mathbb E(f(Y_i))$ with
$f=\sum_{\ell=1}^L a_{\ell} f_{\ell}$, where $f_\ell \in
\tMon( Q,P_{Y_0})$ and $\sum_{\ell=1}^L |a_\ell| \leq 1$. According
to Proposition \ref{covineg},
\begin{equation}
\label{majgamma2}\Vert{\mathbb E}_0 (X_i) \Vert_1 \leq 8 \int_0^{\alpha (i)} Q(u) du \, .
\end{equation}
Since $Q_{|X_0|}(u)\leq Q_{|f(Y_0)|} (u) + |\mathbb E(f(Y_0))|$, we see that
$\int_0^x Q_{|X_0|}(u) du \leq 2 \int_0^x Q_{|f(Y_0)|}(u) du$.
Since $f=\sum_{\ell=1}^L a_\ell f_\ell$, we get, according to Item (c) of
Lemma 2.1 in Rio (2000),
\[
\int_0^{x} Q_{|X_0|} (u) du \leq 2\sum_{\ell=1}^L \int_0^{x} Q_{|a_\ell
f_\ell(X_0)|} (u) du \leq 2\sum_{\ell=1}^L |a_\ell| \int_0^{x} Q (u) du \, .
\]
Since $\sum_{\ell=1}^L |a_\ell| \leq 1$, it follows that $
G(u/2) \leq G_{|X_0|} (u)$. In particular, $G_{|X_0|}(u)\geq
G(u/8)$. Using the fact that $Q_{|X_0|}$ is non-increasing and the change of variables $w = G(v)$,
\begin{align*}
\int_{0}^{\Vert{\mathbb E}_0 (X_i) \Vert_1} Q_{|X_0|} \circ G_{|X_0|} (u) du
&\leq \int_0^{\Vert{\mathbb E}_0 (X_i) \Vert_1} Q_{|X_0|} \circ G(u/8) du
= 8 \int_0^{\Vert{\mathbb E}_0 (X_i) \Vert_1/8} Q_{|X_0|}\circ G(v)dv
\\&
=8 \int_0^{ G(\Vert{\mathbb E}_0 (X_i) \Vert_1/8)} Q_{|X_0|}(w) Q(w)dw
\leq 8 \int_0^{\alpha (i)} Q_{|X_0|} (w) Q(w) dw \,
,
\end{align*}
where the last inequality follows from (\ref{majgamma2}). Consequently, by Item (c) of
Lemma 2.1 in Rio (2000),
\begin{eqnarray} \label{ineMP}
{\mathbb E} | X_j {\mathbb E}_0 (X_i)| \leq 48 \sum_{\ell=1}^L |a_\ell| \int_0^{\alpha (i)} Q_{|f_{\ell}(Y_0)|} (u) Q(u) du \leq 48 \int_0^{\alpha (i)} Q^2 (u)du\, ,
\end{eqnarray}
and the same inequality holds if
$f \in \widetilde{\mathcal{F}} (Q,P_{Y_0})$ by applying Fatou's lemma. Consequently starting from (\ref{maj1DR}), we derive that
\beq \label{maj2DR}
\sum_{k=1}^{2^{L-m(L)}} \Vert\bkE_{2^L + (k-1)2^{m(L)}} (U_{k,L}) \Vert_2^2 \leq 96.\, 2^{L-m(L)}\sum_{i=1}^{2^{m(L)}} i \int_0^{\alpha (i)} Q^2 (u)du \, .
\eeq
We now bound up the second term in the right hand side of (\ref{inegadedrio}). Stationarity yields that
$$
\Vert\bkE_{2^L + (k-1)2^{m(L)}} (U_{k,L}) \big ( \sum_{i= k +1}^{2^{L-m(L)}}\bkE_{2^L + (k-1)2^{m(L)}} (U_{i,L}) \big ) \Vert_1 \leq \sum_{j=1}^{2^{m(L)}}
\sum_{i=2^{m(L)}+ 1}^{2^L - (k-1) 2^{m(L)}} {\mathbb E} | X_j {\mathbb E}_0 (X_i)| \, .
$$
Using Inequality (\ref{ineMP}), we then derive that
\begin{equation} \label{maj3DR}
\sum_{k=1}^{2^{L-m(L)} -1 }
\Vert\bkE_{2^L + (k-1)2^{m(L)}} (U_{k,L}) \big ( \sum_{i= k +1}^{2^{L-m(L)}}\bkE_{2^L + (k-1)2^{m(L)}} (U_{i,L}) \big ) \Vert_1 \leq 48.\, 2^{L}
\sum_{i=2^{m(L)}+ 1}^{2^L } \int_0^{\alpha (i)} Q^2 (u)du \, .
\end{equation}
Starting from (\ref{inegadedrio}) and considering the bounds (\ref{maj2DR}) and (\ref{maj3DR}),
we get that
\begin{eqnarray} \label{maj4DR} \Vert \sup_{k \leq 2^{L-m(L)}} | \sum_{ \ell =1}^k (\widetilde
U_{\ell,L} - U_{\ell,L}) | \Vert_2^2 & \leq & C 2^{L-m(L)} \!\! \int_0^1 \!\!Q (u) R (u)
(\alpha^{-1} (u) \wedge 2^{m(L)} ) du \nonumber \\
& \leq & C 2^{L-m(L)} M_{3,\alpha} (Q , 2^{m(L)} ) \, ,
\end{eqnarray}
since $R(u) \geq \alpha^{-1} (u)$.
Starting from (\ref{dec0item2}) and considering the bounds (\ref{maj5DR}),
(\ref{maj4DR}) and (\ref{RioA1}) in the case $p<3$, we then get (\ref{DL1borneL2}),
which ends the proof of Lemma \ref{LemDL1}.
\par\medskip
\noindent{\bf Proof of Item 1(a). } We choose $Z_i=Z_i'$ with
\beq \label{defml}
m(L) = \Big [ \frac{2L}{p} -\frac{2}{p} \log_2 L \Big ] \, ,\ \text{so that}\
\frac 12 \Bigl(\frac{2^L}{L} \Bigr)^{2/p} \leq 2^{m(L)} \leq \Bigl(\frac{2^L}{L} \Bigr)^{2/p} \, ,
\eeq
square brackets designating as usual the integer part and $\log_2 (x)= (\log x)/(\log 2)$.
Starting from (\ref{apppropineg}), we now prove that
\begin{equation}
D_{L,2} = O ( 2^{L/p} L^{1/2-1/p}) \text{ in } {\mathbb L}^2
\text{ for $p \leq 3$ and a.s. for $p< 3$ if }\, M_{p,\alpha} (Q) < \infty . \label{DL2}
\end{equation}
To prove the almost sure part in (\ref{DL2}), take
\beq \label{deflambda}\lambda=\lambda_L= K 2^{m(L)/2} \sqrt{L}
\text{ with } K = \sqrt{2 c_2 \sigma^2 \log 2} \, . \eeq
Then, on one hand,
$$
\sum_{L > 0} 2^L \exp \Bigl( - \frac{\lambda_L^2}{ c_2 \sigma^2 2^{m(L)}} \Bigr)
= \sum_{L \geq 0} 2^{L-2L} < \infty
\ \text{and}\
\sum_{L > 0} 2^{L} \lambda_L^{-3} < \infty \, ,
$$
for $p<3$. On the other hand, since $M_{3,\alpha} (Q, a\lambda) \leq aM_{3,\alpha} (Q,\lambda)$
for any $a\geq 1$,
\begin{eqnarray*}
2^L\lambda_L^{-3} M_{3,\alpha} (Q,\lambda_L) \leq
2^{L - 3m(L)/2} L^{-1} M_{3,\alpha} (Q, K 2^{m(L)/2} ) .
\end{eqnarray*}
Consequently, from the choice of $m(L)$ made in (\ref{defml}),
\begin{equation*}
\sum_{L > 0} 2^L \lambda_L^{-3} M_{3,\alpha} (Q,\lambda_L)
\leq C \sum_{L > 0} (2^L / L)^{(p-3)/p}
M_{3,\alpha} (Q, (2^L / L)^{1/p} ).
\end{equation*}
Next, for $p \in ]2,3[$,
$$
\sum_{ L \, : \, \frac{2^L}{L} \geq R^p(x)} \Big (\frac{2^L}{L}
\Big )^{1 -3/p} \leq C R^{p-3} (x) \text{ and } \sum_{ L \, : \,
\frac{2^L}{L} \leq R^p(x)} \Big (\frac{2^L}{L} \Big )^{1 -2/p} \leq
C R^{p-2} (x) \, ,
$$
which ensures that
\beq \label{calcul} \sum_{L > 0} 2^L \lambda_L^{-3}
M_{3,\alpha} (Q,\lambda_L) \leq C M_{p,\alpha} (Q) .
\eeq
Consequently under (\ref{Condstrong}), we derive that $\sum_{L > 0} {\mathbb P} (
D_{L,2} \geq 2 \lambda_L ) < \infty $ implying the almost sure part of (\ref{DL2})
via the Borel-Cantelli lemma.
\par\ssk
We now prove the ${\mathbb L}^2$ part of (\ref{DL2}). Clearly
\beq \label{1equalityL2}
{\mathbb E} ( D_{L,2}^2 ) = 8 \int_0^\infty \lambda {\mathbb P} ( D_{L,2} \geq 2\lambda) d\lambda
\leq 4 \lambda_L^2 + 8 \int_{\lambda_L}^\infty \lambda {\mathbb P} ( D_{L,2} \geq 2\lambda) d\lambda .
\eeq
We now apply (\ref{apppropineg}). First, from (\ref{deflambda}),
$$
\int_{\lambda_L}^\infty \lambda \exp \Bigl( - \frac{\lambda^2}{c_2 \sigma^2
2^{m(L)}} \Bigr) d\lambda = c_2 \sigma^2 2^{m(L)-L} \ \text{and}\
2^L \int_{\lambda_L}^\infty \frac{c_4\sigma^3}{\lambda^2} d\lambda =
c_4\sigma^3 \frac{2^L}{\lambda_L}.
$$
In the case $p<3$ and $\Lambda_{p,\alpha} (Q) < \infty$, from (\ref{RioA1}),
there exists a positive constant $C$ depending on $p$ and $\Lambda_{p,\alpha} (Q)$ such that
\beq \label{2inequalityL2}
\int_{\lambda_L}^\infty \frac{c_3 2^L}{\lambda^2} M_{3,\alpha} ( Q , \lambda )
d\lambda \leq C \int_{\lambda_L}^\infty \lambda^{1-p} d\lambda \leq
\frac{ C 2^L}{(p-2) \lambda_L^{p-2} } \,\, .
\eeq
Now, by (\ref{deflambda}) again,
$(K/2) 2^{L/p} L^{1/2 - 1/p} \leq \lambda_L \leq K 2^{L/p} L^{1/2 - 1/p}$,
and consequently, collecting the above estimates, we get that
${\mathbb E} ( D_{L,2}^2 ) = O ( \lambda_L^2 )$, which implies the ${\mathbb L}^2$ part of
(\ref{DL2}).
\par\ssk
We now deal with $D_{L,1} $. We will prove that
\begin{equation}
D_{L,1} = O ( 2^{L/p} L^{1/2-1/p})
\text{ in } {\mathbb L}^2 \text{ for $p \leq 3$ and a.s. for $p< 3$ if }\,
M_{p,\alpha} (Q) < \infty . \label{DL1}
\end{equation}
We first derive from Lemma \ref{LemDL1} that
$\|D_{L,1}\|^2_2 \leq C 2^{L -m(L) (p-2)/2 }$ (applying (\ref{RioA1}) in the case $p<3$),
which implies the ${\mathbb L}^2$ part of (\ref{DL1}).
\par\ssk
Next, from (\ref{DL1borneL2}) together with the Markov inequality,
$$
\sum_{L>0} {\mathbb P} ( D_{L,1} \geq \lambda_L) \leq C \sum_{L>0} 2^{L+ (1-p) m(L)} +
C \sum_{L>0} \frac{ 2^L }{L 2^{3m(L)/2} } M_{3,\alpha} (Q , 2^{m(L)/2}) \, ,
$$
where $\lambda_L$ is defined by (\ref{deflambda}).
Repeating exactly the same arguments as in the proof of (\ref{calcul}), we get that
the second series on right hand in the above inequality is convergent for $p < 3$. Now
$2^{L + (1-p) m (L) } \leq 2^{p-1} 2^{L(2-p)/p} L^{2(p-1)/p}$,
which ensures the convergence of the first series on right hand. Hence, by the Borel-Cantelli
lemma $D_{L,1} = O (\lambda_L )$ almost surely, which completes the proof of (\ref{DL1}).
Finally Item 1(a) of Theorem \ref{ThSM} follows from both (\ref{DL1}), (\ref{DL2}) and (\ref{1dec})
and (\ref{decsup}).
\par\medskip\noindent
{\bf Proof of Item 1(b).} We choose $\widetilde Z_i= Z_i'$ with
$m(L) = [ (2L/p) + (2(1 + \varepsilon) / p ) \log_2 (1 \vee \log L) ]$.
Following the proof of Item 1(a) with this selection of $m(L)$, Item 1(b) follows.
\par\medskip
\noindent{\bf Proof of Item 2.} Starting from the decomposition (\ref{1dec}), we just have to bound both almost surely and in ${\mathbb L}^2$ the random variables
$D_L := \sup_{ \ell \leq 2^{L}} | S_{2^L + \ell} - S_{2^L}|$.
Applying Proposition \ref{inegamax} in case where $\sigma^2 =0$, we get that for any positive $\lambda$,
\beq \label{applpropmax0}
\BBP ( D_L \geq \lambda ) \leq c 2^{L }\lambda^{-3}
M_{3,\alpha} (Q , \lambda ) ,
\eeq
where $c$ is a positive constant. Using computations as in (\ref{1equalityL2}) and (\ref{2inequalityL2}), we then get that for any positive $\lambda_L$,
$\| D_L \|_2^2 \leq 4 \lambda_L^2 + C 2^L \lambda_L^{2-p}$. Choosing $\lambda_L = 2 ^{L/p}$ gives the
${\mathbb L}^2$ part of Item 2 (a). To prove the almost sure parts, we start from (\ref{applpropmax0}) and choose, for $\delta >0$ arbitrarily small,
$$
\lambda = 2 ^{L/p} L^{1/p} ( 1 \vee \log L)^{(1+\varepsilon)/p} \text{ and }
\lambda = \delta 2 ^{L/p} \text{ if $p\in ]2,3[$ and } M_{p,\alpha} (Q) < \infty .
$$
The Borel-Cantelli lemma then implies that almost surely
$$
D_L = O ( 2 ^{L/p} L^{1/p} ( 1 \vee \log L)^{(1+\varepsilon)/p})
\text{ a.s. and } D_L = o (2 ^{L/p})\text{ a.s. if $p\in ]2,3[$ and } M_{p,\alpha} (Q) < \infty \,.
$$
This ends the proof of the almost sure part of Item 2 and then of the theorem.
\subsection{Proof of Item 2 of Theorem \ref{ASmap}.} \label{proofASmap}
If $\sigma^2(f) >0$, similarly as for the proof of Theorem \ref{ThSM}, we start by constructing a sequence $(Z'^*_i)_{i \geq 1}$ of
i.i.d.~gaussian random variables with mean zero and
variance $\sigma^2(f)$ depending on the sequence $(m(L))_{L \geq 0}$ defined either as in (\ref{defml}) or as in the proof of Item 1(b).
Define for any $k \in \{1, \cdots, 2^{L-m(L)} \}$,
$$
I_{k,L} = ]2^L + (k-1) 2^{m(L)} , 2^L +k 2^{m(L)} ] \cap {\mathbb N} \ \text{and}\ U^*_{k,L} = \sum_{i\in I_{k,L}} (f\circ T_\gamma^i- \nu_\gamma(f)) \, .
$$
For $k \in \{1, \cdots, 2^{L-m(L)} \} $, let $V^*_{k,L}$ be the
${\mathcal N}(0, \sigma^2 2^{m(L)} )$-distributed random variable
defined from $U^*_{k,L}$ via the conditional quantile transformation,
that is \beq \label{defVkNT}
V^*_{k,L} = \sigma(f) 2^{m(L)/2} \Phi^{-1}
( F^*_{k,L}( U^*_{k,L}- 0 ) + \delta_{2^L +k2^{m(L)} }(
F^*_{k,L}( U^*_{k,L} ) - F^*_{k,L} (U^*_{k,L} - 0) ) ) \, ,
\eeq where
$ F^*_{k,L}:= F_{U^*_{k,L} | \tilde {\mathcal G}_{ 2^{L}+k2^{m(L)}+1}}$ is the d.f. of
the conditional law of $U^*_{k,L}$ given $\tilde {\mathcal G}_{2^{L}+k 2^{m(L)}+1}$,
where $\tilde {\mathcal G}_{m} = \sigma( T_{\gamma}^m, ( \delta_i)_{i \geq m}$ ) and $\Phi^{-1}$ the inverse of the
standard Gaussian distribution function $\Phi$. Since $\delta_{2^L +k2^{m(L)} }$ is
independent of $\tilde {\mathcal G}_{ 2^L +k2^{m(L)}+1}$, the random variable $V^*_{k,L}$
is independent of $\tilde {\mathcal G}_{2^{L}+k 2^{m(L)}+1}$, and has the Gaussian distribution
$N (0, \sigma^2(f) 2^{m(L)})$. By induction on $k$, the random variables $(V^*_{k,L})_k$
are mutually independent and independent of $\tilde {\mathcal G}_{ 2^{L+1} +1 }$. Let us construct now the
sequence $(Z'^*_i)_{i \geq 1}$ as follows. Let $Z'^*_1= \sigma(f) \Phi^{-1}(\delta_1)$. For any $L \in {\mathbb N}$ and any $k \in \{1, \cdots, 2^{L-m(L)} \}$,
the random variables $(Z'^*_{ 2^L +( k-1) 2^{m(L)}+1}, \ldots , Z'^*_{ 2^L +k2^{m(L)}})$
are defined in the following way. If $m(L)=0$, then $
Z'^*_{2^L + k 2^{m(L)}} = V^*_{k,L} $. If $m(L)>0$, then there exists some measurable function $g$ from ${\mathbb R} \times [0,1]$
in ${\mathbb R}^{2^{m(L)}}$ such that, for any pair
$(V,\delta)$ of independent random variables with respective laws $N(0,\sigma^2(f) 2^{m(L)})$ and
the uniform distribution over $[0,1]$, $g(V,\delta) = (N_1, \ldots N_{2^{m(L)}})$ is a Gaussian random vector with i.i.d. components such that $V = N_1 + \cdots + N_{2^{m(L)}}$.
We then set
$$
(Z'^*_{2^L+ ( k-1)2^{m(L)} +1}, \ldots , Z^*_{2^L +k 2^{m(L)}}) = g (V^*_{k,L} , \delta_{2^L
+ (k-1) 2^{m(L)}+1} ).
$$
The so defined sequence $(Z'^*_i)$ has the prescribed distribution.
\medskip
\par
Set now $ S^*_j= \sum_{i=1}^j (f\circ T_\gamma^i- \nu_\gamma(f))$, $T^*_j = \sum_{i=1}^j Z'^*_i$ if $\sigma^2(f) >0$ and $T^*_j =0$ otherwise, and let
$$
D^*_L := \sup_{ 0 \leq \ell \leq 2^{L}} |(S^*_{2^L + \ell}-T^*_{2^L + \ell}) - ( S^*_{ 2^{L+1} } -T^*_{ 2^{L+1} })| \, .
$$
Similarly as in the proof of (\ref{1dec}), we get that
\begin{eqnarray} \label{1decdyn} \sup_{1 \leq k \leq 2^{N+1} }|S^*_k -T^*_k| \leq
|S^*_1 -T^*_1| + 2 D^*_0 + 2 D^*_1 + \cdots + 2 D^*_N .
\end{eqnarray}
For any $L \in {\mathbb N}$, on the probability space $([0, 1], \nu_\gamma)$, the random
variable $(T^{2^{L}+1}_\gamma, T^{2^{L}+2}_\gamma, \ldots , T^{2^{L+1}}_\gamma)$ is
distributed as $(Y_{2^{L+1}},Y_{2^{L+1}-1}, \ldots, Y_{2^{L}+1})$, where $(Y_i)_{i \geq 1}$ is a stationary Markov chain with
transition kernel $K_{\gamma}$ and invariant measure $\nu_\gamma$. From our construction of the random variables $Z'^*_i$, for any $L \in {\mathbb N}$,
$$
(T^{2^{L}+1}_\gamma, \ldots , T^{2^{L+1}}_\gamma,Z'^*_{2^{L}+1}, \ldots , Z'^*_{2^{L+1}}) =^{{\mathcal D}}(Y_{2^{L+1}}, \ldots , Y_{2^{L}+1},Z'_{2^{L+1}}, \ldots , Z'_{2^{L}+1}) \, ,
$$
where the sequence $(Z'_i)_{2^L +1 \leq i \leq 2^{L+1}}$ is defined from $(Y_i,\delta_i)_{2^L < i \leq 2^{L+1}}$ as in the proof of Theorem \ref{ThSM}.
It follows that
\begin{equation*} \label{equalitylaw*} D^*_L =^{{\mathcal D}} D_L \text{ where }D_L := \sup_{ 0 \leq \ell \leq 2^{L}} |(S_{2^L + \ell}-T_{2^L + \ell}) - ( S_{ 2^{L} } -T_{ 2^{L} })|
\end{equation*}
and, for any $j \geq 1$, $T_j = \sum_{i=1}^j Z'_i$ if $\sigma^2(f) >0$ and $T_j =0$ otherwise.
Hence we have, for any positive $\lambda$,
$ {\mathbb P} (D^*_{L} \geq \lambda) = {\mathbb P} (D_{L} \geq \lambda)$.
Proceeding as in the proof of Theorem \ref{ThSM}, Item 2 follows.
\section{Appendix}
\setcounter{equation}{0}
Next lemma is a parametrized version of Theorem 1 of Rio (1998). We first need the following definition.
\begin{Definition}
$\Lambda_2$ is the class of
real functions $f$ which are continuously differentiable
and such that $|f'(x) - f'(y) |\leq | x - y |$ for any $(x,y) \in {\mathbb R} \times {\mathbb R}$.
\end{Definition}
\begin{Lemma} \label{riolma} Let $Z$ be a random variable with values in
a purely non atomic Lebesgue space $(E, {\mathcal L} (E) , m)$ and ${\mathcal F} = \sigma (Z)$.
For real random variables $U$ and $V$, let $P_{U | {\mathcal F}}$ be the law of $U$
given ${\mathcal F}$ and $P_V$ be the law of $V$. Assume that $V$ is
independent of $\mathcal F$. Let $\sigma^2 >0$ and $N$ be a ${\mathcal N} (0,\sigma^2)$-distributed random variable independent of $\sigma( Z,U,V)$. Then
$$
\bkE \big ( W_2^2 (P_{U | {\mathcal F}} , P_V) \big ) \leq 16 \sup_{f \in \Lambda_2(E)}\bkE \big ( f( U + N , Z) - f( V + N , Z) \big ) + 8 \sigma^2
\, ,
$$
where $\Lambda_2 (E)$ denotes the set of measurable functions $f:{\mathbb R}
\times E \rightarrow {\mathbb R}$ wrt the $\sigma$-fields
${\mathcal L} ( {\mathbb R} \times E) $ and ${\mathcal B} ({\mathbb R})$,
such that $f( \cdot, z) \in \Lambda_2$ and $f(0,z)=f'(0,z)=0$ for any
$z \in E$.
\end{Lemma}
\noindent{\it Proof of Lemma \ref{riolma}.} Notice first that
\beq \label{lissage}
\bkE \big ( W_2^2 (P_{U | {\mathcal F}} , P_V) \big ) \leq 2 \bkE \big ( W_2^2 (P_{U + N | {\mathcal F}} , P_{V + N}) \big ) + 8 \sigma^2 \, .
\eeq Let $G$ be the d.f. of $P_{V + N}$.
Since $E$ is a Lebesgue space, there exists a regular version of the conditional distribution function of $U + N$ conditionally to $Z$, that is, a function $(x, z) \rightarrow F_{z} (x)$ from ${\mathbb R} \times E$ in ${\mathbb R}$ such that, for any real $x$,
$F_Z (x) = {\mathbb E} ( \I_{ U + N\leq x} | Z )$ almost surely.
Notice in addition that, for any $z$ in $E$, $F_{z}$ is a $C^{\infty}$ increasing distribution function.
Let now $H_z (x) = F_z (x) - G(x)$, $ A_z= \{y \in {\mathbb R} \, : \, H_z (y)= 0 \}$, and for any $(x, z) \in {\mathbb R} \times E$, let
\begin{equation} \label{defflemma}
h(x, z) = d(x,A_z \cup \{ 0 \} ) \sign H_z (x) \text{ and } f(x ,z ) = \int_0^x h (y , z) dy \, ,
\end{equation}
where $d(x,A_z \cup \{ 0 \})$ is the distance of $x$ to the random set $A_z\cup \{ 0 \}$ and $\sign y=1$ for $y>0$, $0$ for $y=0$ and $-1$ for $y < 0$.
For $z$ fixed, $ f(0 ,z )= f'(0 ,z )=0$ and it is shown in Rio (1998, Inequality (7)) that $ f(
\cdot, z)$ belongs to $\Lambda_2$, and that for any $u \in ]0,1[$,
$$
f(F_z^{-1} (u) , z) - f(G^{-1} (u) , z) \geq \frac 18 \big (F_z^{-1} (u) - G^{-1} (u) \big )^2 \, ,$$
and therefore that for any $z \in E$,
\begin{eqnarray} \label{formulemajw2}
W_2^2 (P_{U + N| Z = z } , P_V) & = & \int_0^1 \big (F_z^{-1} (u) - G^{-1} (u) \big )^2 du \nonumber \\
& \leq & 8 \Big ( \int_{\mathbb R} f(
x, z) dP_{U+ N | Z=z } - \int_{\mathbb R} f(
x, z) dP_{V+N} \Big ) \, .
\end{eqnarray}
We prove now that the function $ f$ defined by (\ref{defflemma}) is ${\mathcal L} ( {\mathbb R} \times E) -{\mathcal B} ( {\mathbb R})$ measurable. Notice first that since for any fixed $z$, $x \mapsto h(x, z)$ is continuous we get that
$$
f(x ,z ) = \lim_{n \rightarrow \infty} \frac{x}{n} \sum_{i=1}^n h (itn^{-1} , z) \, .
$$
Therefore the mesurability of $ f$ will come from the mesurability of $h$. With this aim, it is
enough to prove the mesurability of the restriction $h_n$ of $h$ to $[-n,n] \times E$
for any positive integer $n$.
Let $\varphi\, : \, [-n,n] \rightarrow [0,1]$ be the one to one bicontinuous map defined by
$\varphi(x) = (n-x)/(2n)$. We then define
\begin{eqnarray} \label{defggrande}
g \, : \, [0,1]\times E& \rightarrow & {\mathbb R} \nonumber \\
(x, z) & \mapsto & h( \varphi^{-1} (x) ,z ) \, .
\end{eqnarray}
The mesurability of $h_n$ will then follow from the mesurability of $g$. Since $E$ is purely non atomic, $(E, {\mathcal L} (E), m)$ is isomorph to $([0,1], {\mathcal L} ([0,1]) , \lambda_{[0,1]})$ where ${\mathcal L} ([0,1])$ and $ \lambda_{[0,1]} $ are respectively the Lebesgue $\sigma$-algebra and the Lebesgue measure on $[0,1]$
(see for instance Theorem 4.3 in De La Rue (1993)). Consequently the following theorem due to Lipi\'nski (1972) which is recalled in Grande (1976) also holds in $[0,1]\times E$.
\begin{Theorem} (Lipi\'nski (1972)) Let $g$ be a bounded function from $[0,1] \times E$ into ${\mathbb R}$ such that
\begin{enumerate}
\item the cross sections $g_x (t) = g(x,t)$ and $g^z(t) = g(t,z)$ are respectively ${\mathcal L}(E)$ and ${\mathcal L}([0,1])$-measurable,
\item for all $t \in [0,1]$, $k_t(z) = \int_0^t g(x,z)dx$ is ${\mathcal L}(E)$-measurable,
\item for all $z \in E$, the cross section $g^z$ is a derivative.
\end{enumerate}
Then $g$ is measurable wrt the $\sigma$-fields ${\mathcal L} ( [0,1]\times E)$ and ${\mathcal B} ( {\mathbb R})$.
\end{Theorem}
Items 2 and 3 as well as the second part of Item 1 follows directly from the fact that if $z$ is fixed, then the function $x \rightarrow g(x,z)$ is continuous (recall that $h ( \cdot, z)$ and $\varphi^{-1}$ are continuous). It remains to show that for all $x \in [0,1]$ the cross section $g_x$ is Lebesgue-measurable. Let us then prove that for any $x \in [-n,n]$ and any $\delta >0$,
\begin{equation} \label{mesurabilite}
\{ z \in E \, : \, g(x, z) \geq \delta \} \in {\mathcal L}(E) \text{ and } \{ z \in E \, : \, g(x, z) \leq -\delta \} \in {\mathcal L}(E)
\end{equation}
which will end the proof of the mesurability of $g$ and then of the lemma. For any $x \in [-n,n]$ and any $\delta >0$, we notice that
$$
\{ z \in E \, : \, g(x, z) \geq \delta \} =
\begin{cases}
\{ z\in E \, : \, H_z (x) >0\} \cap \{ z \in E \, : \, d(x, A_z) \geq \delta \} \quad
\text{if $|x| \geq \delta$}\\
\ \emptyset \phantom{z \in E \, : \, H_z (x) >0\} \cap \{ z \in E \, : \, d(x, A_z) \geq \delta \}} \ \text{ if $|x| < \delta$}.
\end{cases}
$$
If $|x| \geq \delta$,
\begin{eqnarray*}
& & \{ z\in E \, : \, H_z(x) >0\} \cap \{ z \in E \, : \, d(x, A_z) \geq \delta \} \\
& & \quad \quad = \{ z \in E \, : \, H_z (x) >0\} \cap \{ z\in E \, : \, ]x - \delta, x + \delta [ \cap A_z = \emptyset \} \\
& & \quad \quad = \{ z \in E \, : \, H_z (y) >0 \, , \, \forall y \in ]x - \delta, x + \delta [\} \, .
\end{eqnarray*}
Using the fact that the function $H_z (\cdot)$ is continuous, we get that if $|x| \geq \delta$,
\begin{eqnarray*}
& & \{ z \in E \, : \, H_z (x) >0\} \cap \{ z \in E \, : \, d(x, A_z) \geq \delta \} \\
& & \quad \quad = \bigcup_{p \in {\mathbb N }^*} \big \{ z \in E \, : \, H_z(y) \geq \frac{1}{p}
\, , \, \forall y \in ]x - \delta, x + \delta [ \cap {\mathbb Q}\big \} \, ,
\end{eqnarray*}
which proves the first part of (\ref{mesurabilite}) since
$\{ z \in E \, : \, H_z (a) \geq p^{-1} \}$ belongs to ${\mathcal L} (E)$ for any $a \in {\mathbb Q}$ and any $p \in {\mathbb N }^*$. The second part of (\ref{mesurabilite}) follows from the same arguments by changing the sign. This ends the proof of the ${\mathcal L} ( {\mathbb R} \times E) -{\mathcal B} ( {\mathbb R})$ measurability of $ f$ defined by (\ref{defflemma}).
Next $P_{(U + N , Z)}$ and $P_{(V + N , Z)}$ are absolutely continuous wrt $\lambda \otimes P_Z$. Consequently, starting from (\ref{lissage}) and using (\ref{formulemajw2}), the lemma follows.
$\diamond$
\begin{Proposition} \label{condversion} Let $X_i = f(Y_i) - \bkE ( f(Y_i))$, where $f$ belongs to $\widetilde{\mathcal{F}}(Q, P_{Y_0})$. Assume that $M_{2,\alpha} (Q) < \infty$. Then the series
$\bkE (X_0^2) + 2 \sum_{k \geq 1} \bkE (X_0 X_k)$ is convergent to some nonnegative real $\sigma^2$. If $\sigma^2 >0$, then there exists a positive constant
$C$ depending on $\sigma^2$ such, that for any $n >0$,
\begin{equation}
\bkE \big ( W_2^2 (P_{S_n | {\mathcal F}_0} , G_{n \sigma^2}) \big ) \leq C n^{1/2}
M_{3,\alpha} ( Q , n^{1/2}) \, ,
\end{equation}
where $M_{3,\alpha} ( Q , n^{1/2})$ is defined in (\ref{3alphatronq}).
\end{Proposition}
\noindent{\it Proof of Proposition \ref{condversion}.}
Let $(N_i)_{i\in{\mathbb Z}}$ be a sequence of independent random variables
with normal distribution ${\mathcal N}(0, \sigma^2)$. Suppose furthermore that the sequence $(N_i)_{i\in{\mathbb Z}}$ is independent of ${\mathcal F}_{\infty}$.
Let $N$ be a ${\mathcal N}(0,\sigma^2)$-distributed random variable, independent
of ${\mathcal F}_{\infty}\vee \sigma(N_i , i \in
{\mathbb Z} )$. Set $T_n = N_1 + N_2 + \cdots + N_n$. Let $Z= ((Y_i, \delta_i) \, : \, i \leq 0)$ and $E = ({\mathbb R} \times [0,1] )^{{\mathbb Z}^-}$. Notice that
$(E, {\mathcal L}(E), P_Z)$ is a purely non atomic Lebesgue space. From Lemma \ref{riolma}, we have to bound
\begin{equation}\label{Delta}
\Delta (\varphi) = {\mathbb E} ( \varphi(S_n + N, Z) - \varphi (T_n + N,Z) ) \, ,
\end{equation}
for any function $\varphi$ in ${\Lambda_2 (E)}$. With this aim, we apply the Lindeberg method.
\begin{Notation}\label{not21}
Let $$\varphi_k (x,Z) = \int_{{\mathbb R}} \varphi (t, Z) \phi_{\sigma \sqrt{n-k+1}} (x-t) dt\, . $$ Let $S_0 = 0$, and, for $k>0$,
let $\Delta_k = \varphi_k (S_{k-1} + X_k,Z) - \varphi_k ( S_{k-1} + N_k,Z)$.
\end{Notation}
Since the sequence $(N_i)_{i\in{\mathbb Z}}$ is independent of the sequence $(X_i)_{i\in{\mathbb Z}}$,
\begin{equation}\label{sumdelta}
{\mathbb E} ( \varphi (S_n + N, Z) -\varphi (T_n + N,Z) ) = \sum_{k=1}^n {\mathbb E} ( \Delta_k ) .
\end{equation}
We first show that for any real $u \in [0,1]$,
\begin{equation} \label{1resrio95}
|\bkE (\Delta_k)| \leq C \big ( (n-k+1)^{-1/2} + D_k(u) \big ) \, ,
\end{equation}
where
\begin{eqnarray} \label{2resrio95}
& & D_k(u) = (n-k+1)^{1/2} \int_0^{\alpha (k)} Q(x)dx + \sum_{ i > [k/2]} \int_0^{\alpha (i)} Q^2(x) dx \nonumber \\
& & + \int_0^{u} Q(x) R(x) dx + (n-k+1)^{-1/2} \int_0^{1} Q (x) R(x) R(x\vee u) dx
\, .
\end{eqnarray}
We now prove (\ref{1resrio95}). For the sake of brevity, write $ \varphi_k (x, Z) = \varphi_k(x)$ and $ \varphi (x, Z) = \varphi(x)$ (the derivatives are taken wrt $x$). By the Taylor formula at order 3,
$$
\big | \bkE \big ( \varphi_k (S_{k-1} + N_k) - \varphi_k ( S_{k-1}) - \frac{\sigma^2}{2} \varphi''_k ( S_{k-1}) \big ) \big | \leq
\frac{\| \varphi^{(3)}_k \|_{\infty}}{6 } \bkE |N|^3 \, .
$$
Now Lemma 6.1 in Dedecker, Merlev\`ede and Rio (2009) gives that, almost surely,
\beq \label{cons1lmadmr}
\| \varphi^{(i)}_k\|_{\infty} \leq c_{i} \sigma^{2-i}(n-k+1)^{(2-i)/2} \mbox{ for any integer $i \geq 2$ } \,
\eeq
where the $c_i$'s are universal constants.
Therefore
$$
\big | \bkE \big ( \varphi_k (S_{k-1} + N_k) - \varphi_k ( S_{k-1}) - \frac{\sigma^2}{2} \varphi''_k ( S_{k-1}) \big ) \big | \leq
C (n-k+1)^{-1/2} \, .
$$
Consequently to prove (\ref{1resrio95}), it remains to show that
\begin{equation} \label{3resrio95}
\big | \bkE \big ( \varphi_k (S_{k-1} + X_k) - \varphi_k ( S_{k-1}) - \frac{\sigma^2}{2} \varphi''_k ( S_{k-1}) \big ) \big | \leq C D_k(u) \, ,
\end{equation}
where $D_k(u)$ is defined by (\ref{2resrio95}). To prove (\ref{3resrio95}), we follow the lines of the proof of Proposition 2(a) of Rio (1995-b) with $b_2 =\| \varphi^{(2)}_k\|_{\infty}$,
$b_3 = \| \varphi^{(3)}_k\|_{\infty}$ and the modifications below.
Since $f$ belongs to $\widetilde{\mathcal{F}}(Q, P_{Y_0})$, we can write
$$
X_i = \lim_{N \rightarrow \infty} {\mathbb L}^1 \sum_{\ell=1}^N a_{\ell,N}\big( f_{\ell,N}(Y_i)
- \bkE(f_{\ell,N}(Y_i)) \big )\, ,
$$
with $f_{\ell,N}$
belonging to $\tMon( Q, P_{Y_0})$ and $\sum_{\ell=1}^N |a_{\ell,N}|
\leq 1$. For $u \in [0,1]$, let the function $g_u$ be defined by $g_u(x) = (x \wedge Q(u)) \vee (-Q(u))$. Since there exists a subsequence $m(N)$ tending to infinity such that $\sum_{\ell=1}^{m(N)} a_{\ell, m(N)} g_u \circ f_{\ell, m(N)}(Y_0)$ is convergent in $ {\mathbb L}^1$, for any $i \geq 0$, we define
\[
\bar X_i=
\lim_{N \rightarrow \infty} {\mathbb L}^1 \sum_{\ell=1}^{m(N)} a_{\ell, m(N)}\big( g_u \circ f_{\ell, m(N)}(Y_i)
- \bkE(g_u \circ f_{\ell,m(N)}(Y_i)) \big ) \quad \text{and} \quad \tilde X_i=X_i - \bar X_i \, .
\]
Let also
$$
Q_u(x) := Q (x)\I_{x \leq u} \text{ and } \bar Q_u (x) := Q(x \vee u) \, .
$$
Since $Q_{|g_u \circ f_{\ell, m(N)} (Y_i)|} \leq \bar Q_u $, this means that $\bar X_i = r(Y_i) - \bkE (r(Y_i)) $ where $r$ belongs to $\widetilde{\mathcal{F}}(\bar Q_u, P_{Y_0})$.
By the Taylor integral formula,
\begin{eqnarray} \label{dt1}
\varphi_k( S_k ) - \varphi_k (S_{k-1} ) - \varphi_k^\prime (S_{k-1}) X_k & = &
X_k \int_0^1 (\varphi_k^\prime (S_{k-1} + v X_k ) - \varphi_k^\prime (S_{k-1} )) dv \nonumber \\
& = & X_k \int_0^1 (\varphi_k^\prime (S_{k-1} + v X_k ) -
\varphi_k^\prime (S_{k-1} + v \bar X_k )) dv \nonumber \\ & + &
X_k\bar X_k \int_0^1 \!\!\! \int_0^1 v \varphi_k'' ( S_{k-1} + vv'\bar X_k) dv dv' .
\end{eqnarray}
The first term on right hand is bounded up by
$b_2 |X_k (X_k - \bar X_k)|/2$. Moreover
$$
\Big|
\int_0^1 \!\!\! \int_0^1 v \varphi_k'' ( S_{k-1} + vv'\bar X_k) dv dv'
-{1\over 2} \varphi_k'' (S_{k-1}) \Big| \leq {b_3\over 6} |\bar X_k|.
$$
Setting $h_u(x) = x- g_u(x)$, we get that for any $f$ belonging to $\tMon( Q, P_{Y_0})$,
\begin{eqnarray*}
& & \bkE \big | (f(Y_k)
- \bkE(f(Y_k)) ) ( h_u \circ f_{\ell}(Y_k)
- \bkE(h_u \circ f(Y_k) )) \big | \\
& & \quad \quad \quad \quad \leq \bkE | f(Y_k)
h_u(f(Y_k)) | +3 \bkE | f(Y_k) | \bkE |
h_u(f(Y_k)) | \, .
\end{eqnarray*}
Since $Q_{| f
(Y_k) |} \leq Q $ and $Q_{|h_u ( f (Y_k))|} \leq (Q - Q(u))_+ \leq Q_u $, we derive that
\begin{eqnarray*}
& & \bkE \big | (f(Y_k)
- \bkE(f(Y_k)) ) ( h_u \circ f_{\ell}(Y_k)
- \bkE(h_u \circ f(Y_k) )) \big |\\
& & \quad \leq \int_0^u Q^2(x) dx +3 \big( \int_0^1 Q(x)dx \big) \big( \int_0^u Q(x)dx \big)
\leq 4 \int_0^u Q^2(x) dx \, ,
\end{eqnarray*}
by using Lemma 2.1(a) in Rio (2000). Now, by Fatou lemma,
\begin{eqnarray*}
& & \bkE | X_k ( X_k - \bar X_k) | \leq \liminf_{N \rightarrow \infty} \sum_{\ell =1}^{m(N)} \sum_{j =1}^{m(N)}
|a_{\ell, m(N)}| |a_{j,m(N)}|\\
& & \times \bkE \big | (f_{\ell, m(N)}(Y_k)
- \bkE(f_{\ell, m(N)}(Y_k)) ) ( h_u \circ f_{j, m(N)}(Y_k)
- \bkE(h_u \circ f_{j, m(N)}(Y_k) )) \big | \, ,
\end{eqnarray*}
whence
\beq \label{dt2}
\bkE | X_k ( X_k - \bar X_k) | \leq 4 \int_0^u Q^2(x) dx \, .
\eeq
Similarly using Lemma 2.1 in Rio (2000) and the fact that $Q_{| g_u\circ f
(Y_k) |} \leq \bar Q_u$ for any $f$ belonging to $\tMon( Q, P_{Y_0})$, we derive that
\beq \label{dt2bis}
\bkE | X_k ( \bar X_k)^2 | \leq 8 \int_0^{1} Q^2 (x) Q (x\vee u) dx \, .
\eeq
It follows that
\begin{eqnarray} \label{dt3}
& & \big | \bkE ( \varphi_k( S_k ) - \varphi_k (S_{k-1} ) - \varphi_k^\prime (S_{k-1}) X_k -
{1\over 2} \varphi_k'' (S_{k-1}) X_k \bar X_k ) \big | \nonumber \\
& & \leq
2b_2 \int_0^u Q^2(x) dx +
{4b_3\over 3} \int_0^{1} Q^2 (x) Q (x\vee u) dx .
\end{eqnarray}
Now we control the second order term.
Let \beq \label{defgamma} \Gamma_{k}(k,i) = \varphi_k'' (S_{k-i}) - \varphi_k'' (S_{k-i-1}) \, ,\eeq
and
\beq \label{defr}
r = \alpha^{-1}(u) \, .
\eeq
Clearly
$$
\varphi_k'' (S_{k-1}) X_k \bar X_k = \sum_{i=1}^{(r\wedge k)-1}
\Gamma_{k}(k,i) X_k \bar X_k + \varphi_k'' (S_{k-(r\wedge k)})X_k \bar X_k \, ,
$$
Since
$|\Gamma_{k}(k,i) | \leq b_3 |X_{k-i}|$, by stationarity we get that for any $i \leq (r\wedge k)-1$,
$$
\big |\mathop{\rm Cov}\limits(\Gamma_{k}(k,i) , X_k \bar X_k ) \big | \leq b_3 \Vert X_0 \big (\bkE_0 (X_i \bar X_i )- \bkE (X_k \bar X_k) \big ) \Vert_1 \, .
$$
Applying Proposition \ref{covineg} with $m=1$, $q=2$, $k_1=0$, $k_2=k_3=i$, $f_{j_1}=f_{j_2}=f$ and
$f_{j_3} \in \widetilde{\mathcal{F}}( \bar Q_u, P_{Y_0})$, we derive that
\begin{eqnarray*}
& & \big |\mathop{\rm Cov}\limits (\Gamma_{k}(k,i) , X_k \bar X_k ) \big | \leq 32 b_3 \int_0^{\alpha(i)} Q^2(x) Q (x \vee u) dx
\, .
\end{eqnarray*}
Since $| \varphi_k'' (S_{k-(r\wedge k)}) | \leq b_2 $ a.s., we also get
by stationarity that
$$
\big |\mathop{\rm Cov}\limits ( \varphi_k'' (S_{k-(r\wedge k)}), X_k\bar X_k)) \big | \leq b_2 \Vert \bkE_0 (X_{r\wedge k} \bar X_{r\wedge k})- \bkE (X_{r\wedge k} \bar X_{r\wedge k})\Vert_1 \, .
$$
Applying Proposition \ref{covineg} with $m=0$, $q=2$, $k_1=k_2=r$, $f_{j_1}=f$ and
$f_{j_2} \in \widetilde{\mathcal{F}}(\bar Q_u, P_{Y_0})$, and noting that $\alpha (r) \leq u$, we also get that
\begin{eqnarray*}
\big |\mathop{\rm Cov}\limits ( \varphi_k'' (S_{k-(r\wedge k)}), X_k\bar X_k)) \big | &\leq &16 b_2 \Big ( \int_0^{u} Q(x) Q (u) dx \I_{r \leq k} +
\int_0^{{\alpha(k)}} Q(x) Q (x \vee u) dx \I_{k < r} \Big ) \, .
\end{eqnarray*}
Hence
\begin{eqnarray*} \label{dt4}
{1\over 2} |\mathop{\rm Cov}\limits ( \varphi_k'' (S_{k-1}),
X_k\bar X_k) | & \leq & 8 b_2 \int_0^{u} Q(x) Q ( u) dx\I_{r \leq k} +8 b_2 \int_0^{{\alpha(k)}} Q(x) Q (x \vee u) dx \I_{k < r} \nonumber \\
& + & 16 b_3 \int_0^{1} Q^2(x) R (x \vee u) dx \, ,
\end{eqnarray*}
which together with (\ref{dt3}) and (\ref{dt2}) implies that
\begin{eqnarray} \label{dt5}
& & \big | \bkE ( \varphi_k( S_k ) - \varphi_k (S_{k-1} ) - \varphi_k^\prime (S_{k-1}) X_k) -
{1\over 2} \bkE (\varphi_k'' (S_{k-1})) \bkE (X_k^2) \big | \leq
12b_2 \int_0^u Q^2(x) dx +
\nonumber \\
& & 8 b_2 \int_0^{{\alpha(k)}} Q(x) Q (x \vee u) dx \I_{k < r}+ \frac{52}{3} b_3 \int_0^{1} Q^2(x) R (x \vee u) dx \, .
\end{eqnarray}
To give now an estimate of the expectation of $\varphi'_k ( S_{k-1})X_k$, we write
$$
\varphi'_k (S_{k-1} ) = \varphi'_k (0) + \sum_{i=1}^{k-1} ( \varphi'_{k} (S_{k-i}) - \varphi'_{k} ( S_{k-i-1}) ) .
$$
Hence
\begin{eqnarray} \label{dt6}
\bkE (\varphi'_k ( S_{k-1})X_k) & = & \sum_{i=1}^{k-1} {\rm Cov} \big (\varphi'_{k} (S_{k-i}) - \varphi'_{k} ( S_{k-i-1} ) , X_k \big ) +
\bkE ( \varphi'_k (0) X_k ) \, .
\end{eqnarray}
Now $\varphi'_k (0) $ is a ${\cal F}_0$-measurable random variable, and since $\varphi'(0)=0$ and $\varphi'$ is $1$-Lipschitz wrt $x$,
$$
|\varphi'_k (0)| = |\int (\varphi'(u) -\varphi'(0)) \phi_{\sigma \sqrt{ n-k+1 } } (-u) du | \leq \sigma \sqrt{ n-k+1 } \, \text{ a.s.}
$$
Applying Proposition \ref{covineg} with $m=0$, $q=1$, $k_1=k$ and $f_{j_1}=f$, it follows that
\beq \label{dt7}
{\mathbb E} ( \varphi'_k (0) X_k ) \leq \sigma\sqrt{ n-k+1} \Vert \bkE_0 (X_k) \Vert_1 \leq 8 \sigma \sqrt{n-k+1} \int_0^{\alpha (k)/2} Q(x)dx \, .
\eeq
We give now an estimate of $\sum_{i=1}^{k-1} {\rm Cov} \big (\varphi'_{k} (S_{k-i}) - \varphi'_{k} ( S_{k-i-1} ) , X_k \big )$. Using the
stationarity and noting that $|\varphi_k^\prime (S_{k-i} ) -\varphi_k^\prime (S_{k-i-1})| \leq b_2|X_{k-i}|$,
we have
$$
|\mathop{\rm Cov}\limits (\varphi_k^\prime (S_{k-i} ) -\varphi_k^\prime (S_{k-i-1} ), X_k)| \leq b_2 \Vert X_0 \bkE_0 ( X_i) \Vert_1 \, .
$$
Now, for any $ i \geq r$, $\alpha(i) \leq u$. So applying
Proposition \ref{covineg} with $m=1$, $q=1$, $k_1=0$, $k_2=i$,
$f_{j_1}=f_{j_2}=f$, we get, for any $k \geq i \geq r$, that
\begin{eqnarray} \label{dt8}
& & |\mathop{\rm Cov}\limits (\varphi_k^\prime (S_{k-i} ) -\varphi_k^\prime (S_{k-i-1} ), X_k)| \leq 16 b_2 \int_0^{u} Q^2(x) \I_{x < \alpha (i) } dx \, .
\end{eqnarray}
From now on, we assume that $i<r \wedge k$. Let us replace $X_k$ by $\bar X_k$. Since by stationarity,
$$
|\mathop{\rm Cov}\limits (\varphi_k^\prime (S_{k-i} ) -\varphi_k^\prime (S_{k-i-1} ), X_k- \bar X_k)| \leq b_2 \Vert X_0 \bkE_0 ( X_i- \bar X_i) \Vert_1 \, ,
$$
we can apply Proposition \ref{covineg} with $m=1$, $q=1$, $k_1=0$, $k_2=i$, $f_{j_1}=f$ and
$f_{j_2} \in \widetilde{\mathcal{F}}( \bar Q_u, P_{Y_0})$. Consequently,
\begin{eqnarray} \label{dt9}
& & |\mathop{\rm Cov}\limits (\varphi_k^\prime (S_{k-i} ) -\varphi_k^\prime (S_{k-i-1} ), X_k- \bar X_k)| \leq 16 b_2 \int_0^{u} Q^2(x) \I_{x < \alpha(i)} dx \, .
\end{eqnarray}
Now
$$
\varphi_k^\prime (S_{k-i} ) -\varphi_k^\prime (S_{k-i-1} ) - \varphi_k'' (S_{k-i-1}) X_{k-i}
= R_{k,i},
$$
where $R_{k,i}$ is ${\cal F}_{k-i}$-measurable and
$|R_{k,i}| \leq b_3 X_{k-i}^2 /2$. Consequently, by stationarity,
$$
|\mathop{\rm Cov}\limits (R_{k,i}, \bar X_k) | \leq b_3 \Vert X^2_0 \bkE_0 ( \bar X_i) \Vert_1 /2 \, .
$$
Applying Proposition \ref{covineg} with $m=2$, $q=1$, $k_1=k_2=0$, $k_3=i$, $f_{j_1}=f_{j_2}=f$ and
$f_{j_3} \in \widetilde{\mathcal{F}}(\bar Q_u, P_{Y_0})$, we get that
\begin{eqnarray} \label{dt10}
|\mathop{\rm Cov}\limits (R_{k,i}, \bar X_k) | & \leq & 32 b_3 \int_0^{ \alpha(i) } Q^2(x)Q(x \vee u) dx
\, .
\end{eqnarray}
In order to estimate the term $\mathop{\rm Cov}\limits (\varphi''_k (S_{k-i-1})X_{k-i}, \bar X_k)$,
we introduce the decomposition below:
$$
\varphi''_k (S_{k-i-1}) = \sum_{l=1}^{(i-1) \wedge (k-i-1)} ( \varphi''_k
(S_{k-i-l}) - \varphi''_k (S_{k-i-l-1}) ) + \varphi''_k (S_{(k-2i)\vee 0}) .
$$
For any $l\in \{ 1, \cdots, (i-1) \wedge (k-i-1) \}$, by using the notation (\ref{defgamma}) and stationarity, we get that
$$
|\mathop{\rm Cov}\limits (\Gamma_{k}(k,l+i) X_{k-i} , \bar X_k )| \leq b_3 \Vert X_{- l} X_0 \bkE_0 (\bar X_i ) \Vert_1 \, .
$$
Applying Proposition \ref{covineg} with $m=2$, $q=1$, $k_1=-\ell$, $k_2=0$, $k_3=i$, $f_{j_1}=f_{j_2}=f$ and
$f_{j_3} \in \widetilde{\mathcal{F}}(\bar Q_u, P_{Y_0})$, we then derive that
\begin{eqnarray} \label{dt11}
|\mathop{\rm Cov}\limits (\Gamma_{k}( k, l+i)X_{k-i} , \bar X_k )| & \leq & 64 b_3 \int_0^{ \alpha(i) } Q^2(x)Q(x \vee u) dx \, .
\end{eqnarray}
As a second step, we bound up
$|\mathop{\rm Cov}\limits ( \varphi_k'' (S_{(k-2i)\vee 0}) , X_{k-i} \bar X_k )|$. Assume first that $i \leq [k/2]$. Clearly, using the notation (\ref{defgamma}),
$$
\varphi_k'' (S_{k-2i}) = \sum_{l=i}^{(r-1) \wedge (k-i -1)} \Gamma_{k}(k, l+i) + \varphi'' (S_{(k-i-r) \vee 0}).
$$
Now for any $l\in \{ i, \cdots, (r-1) \wedge (k-i-1) \}$, by stationarity,
$$
|\mathop{\rm Cov}\limits (\Gamma_{k}(k, l+i) , X_{k-i}\bar X_k )| \leq b_3 \Vert X_{- l} \big (\bkE_{- l} (X_0\bar X_i )
- \bkE ( X_0 \bar X_i ) \big ) \Vert_1 \, .
$$
Hence applying Proposition \ref{covineg} with $m=1$, $q=2$, $k_1=-l$, $k_2=0$, $k_3=i$, $f_{j_1}=f_{j_2}=f$ and
$f_{j_3} \in \widetilde{\mathcal{F}}( \bar Q_u, P_{Y_0})$, we derive that
\begin{eqnarray} \label{dt12}
|\mathop{\rm Cov}\limits (\Gamma_{k}( k, l+i) , X_{k-i}\bar X_k )| & \leq & 32 b_3 \int_0^{ \alpha(l) } Q^2(x)Q(x \vee u) dx \, .
\end{eqnarray}
If $i \leq k-r$, then stationarity implies that
$$
|\mathop{\rm Cov}\limits (\varphi_k'' (S_{k-i-r } ), X_{k-i}\bar X_k )|\leq b_2 \Vert \bkE_{0} (X_r\bar X_{i+r} )
- \bkE ( X_{r} \bar X_{i+r} ) \big ) \Vert_1 \, .
$$
Noting that $\alpha(r)\leq u < \alpha(i)$ and applying Proposition \ref{covineg} with $m=0$, $q=2$, $k_0=0$, $k_1=r$, $k_2=i+r$, $f_{j_1}=f$ and
$f_{j_2} \in \widetilde{\mathcal{F}}(\bar Q_u, P_{Y_0})$, we also get that
\beq \label{dt13}
|\mathop{\rm Cov}\limits (\varphi_k'' (S_{k-i-r } ), X_{k-i}\bar X_k )|
\leq 16b_2 \int_0^u \I_{x < \alpha(i)} Q(x) Q(u) dx \, .
\eeq
Now if $i > k-r$, then we write that
$$
|\mathop{\rm Cov}\limits (\varphi_k'' (0 ), X_{k-i}\bar X_k )|\leq b_2 \Vert \bkE_{0} (X_{k-i}\bar X_k )
- \bkE (X_{k-i}\bar X_k ) \Vert_1 \, .
$$
Applying Proposition \ref{covineg} with $m=0$, $q=2$, $k_0=0$, $k_1=k-i$, $k_2=k$, $f_{j_1}=f$ and
$f_{j_2} \in \widetilde{\mathcal{F}}(\bar Q_u, P_{Y_0})$, and noting that for $i \leq [k/2]$, $ \alpha(k-i) \leq \alpha([k/2])$, we obtain that
\beq \label{dt13pri}
|\mathop{\rm Cov}\limits (\varphi_k'' (0 ), X_{k-i}\bar X_k )|
\leq 16b_2 \int_0^{\alpha([k/2])} Q(x) Q(x \vee u) dx \, .
\eeq
Assume now that $i \geq [k/2] + 1$. For any $i \leq k$, the stationarity entails that
$$
|\bkE ( \varphi_k''(0) X_{k-i} \bar X_k) |\leq b_2 \Vert X_0 \bkE_0
(\bar X_i) \Vert_1 \, .
$$
Hence applying Proposition \ref{covineg} with $m=1$, $q=1$, $k_0=0$, $k_1=i$, $f_{j_1}=f$ and
$f_{j_2} \in \widetilde{\mathcal{F}}(\bar Q_u, P_{Y_0})$, and noting that for $i \geq [k/2]+1$, $ \alpha(i) \leq \alpha([k/2])$, we obtain that
\begin{eqnarray} \label{dt13second}
|\bkE ( \varphi_k''(0) X_{k-i} \bar X_k) |&\leq & 16 b_2 \int_0^{\alpha([k/2])} Q(x) Q(x \vee u) dx \, .
\end{eqnarray}
Adding the inequalities (\ref{dt7}), (\ref{dt8}),
(\ref{dt9}), (\ref{dt10}), (\ref{dt11}), (\ref{dt12}) (\ref{dt13}), (\ref{dt13pri}) and (\ref{dt13second}), summing on $i$ and $l$, and using the fact that
$$\sum_{i=1}^{k-1}\I_{x < \alpha (i)} \leq \alpha^{-1} (x) \, , \, \sum_{i=1}^r \I_{x < \alpha (i)}
\leq \alpha^{-1} (x \vee u) \text{ and } \sum_{i=1}^r i \I_{x < \alpha(i)} \leq
(\alpha^{-1} (x \vee u))^2 \, , $$
we then get:
\begin{eqnarray} \label{dt14}
& & \quad |\bkE (\varphi^\prime ( S_{k-1}) X_k ) - \sum_{i=1}^{r-1}
\bkE (\varphi'' (S_{k-2i})) \bkE (X_{k-i} \bar X_k ) \I_{i \leq [k/2]}| \leq C (n-k+1)^{1/2} \int_0^{\alpha(k)}
Q(x)dx + \nonumber \\
& &
48 b_2 \int_0^{u} Q(x)R (x) dx + 24 k b_2 \int_0^{\alpha([k/2])} Q(x) Q(x \vee u) dx + 128 b_3 \int_0^{ 1 } Q^2(x)R(x \vee u) dx
\, .
\end{eqnarray}
It remains to bound up
$$
A_k:=\sum_{i=1}^{r-1} \bkE (\varphi_k'' (S_{k-2i})) \bkE ( X_{k-i}
\bar X_k) \I_{i \leq [k/2]}- \sum_{i=1}^{\infty} \bkE (\varphi_k''
(S_{k-1}) ) \bkE (X_{k-i}X_k) \, .
$$
We first note that by stationarity,
$$
\sum_{i\geq r}| \bkE (\varphi_k'' (S_{k-1}) ) \bkE (X_{k-i}X_k) |
\leq b_2 \sum_{i\geq r} | \bkE(f(Y_0) \bkE_0(X_i)) | \, .
$$
Applying Proposition \ref{covineg} and noting that $\alpha(i) \leq u$ for $i \geq r$, we get that
\begin{equation} \label{dt16}
\sum_{i\geq r}| \bkE (\varphi_k'' (S_{k-1}) ) \bkE (X_{k-i}X_k) | \leq
8b_2 \sum_{i \geq r} \int_0^{\alpha (i)} Q^2(x)dx \leq 8b_2 \int_0^u Q(x) R (x)dx \, .
\end{equation}
By stationarity we also have
$$
\sum_{i=1}^{r-1}| \bkE (\varphi_k'' (S_{k-1}) ) \bkE (X_{k-i}(X_k - \bar X_k)) |
\leq b_2 \sum_{i=1}^{r-1} | \bkE(f(Y_0) \bkE_0(X_i-\bar X_i)) | \, .
$$
Next, noting that $u < \alpha (i)$ for all $i <r$ and
applying Proposition \ref{covineg}, we get that
\begin{eqnarray} \label{dt17}
\sum_{i=1}^{r-1}| \bkE (\varphi_k'' (S_{k-1}) ) \bkE (X_{k-i}(X_k - \bar X_k)) | &\leq &
8b_2 \int_0^u Q^2(x) \sum_{i =1}^{r-1}\I_{x <\alpha (i)}dx \nonumber \\
& \leq & 8b_2 \int_0^u Q^2(x) \alpha^{-1} (x)dx \, .
\end{eqnarray}
In addition, another application of Proposition \ref{covineg} gives
\begin{eqnarray} \label{dt17bis}
\sum_{i=1+[k/2]}^{r-1}| \bkE (\varphi_k'' (S_{k-1}) ) \bkE (X_{k-i} \bar
X_k) | &\leq &
8b_2 \sum_{ i > [k/2]} \int_0^{\alpha(i)} Q^2(x) dx \, .
\end{eqnarray}
In order to bound up the last term, we still write
$$
\bkE (\varphi_k'' (S_{k-1}) - \varphi_k'' (S_{k-2i})) \bkE ( X_{k-i}
\bar X_k ) \I_{i \leq [k/2]} = \sum_{l=1}^{2i-1} \bkE (\Gamma_{k}(k,l)) \bkE (
f(Y_{0}) \bkE_0(\bar X_i) ) \I_{i \leq [k/2]}.
$$
Both this decomposition, Proposition \ref{covineg} and Lemma 2.1 in Rio (2000) then yield :
\begin{eqnarray} \label{dt18}
& & \sum_{i=1}^{r-1}| \bkE (\varphi_k'' (S_{k-1}) - \varphi_k'' (S_{k-2i})) \bkE ( X_{k-i} \bar X_k ) |\I_{i \leq [k/2]}
\leq 8 b_3 \sum_{i =1}^{r-1} i \int_0^{\alpha(i)} Q^2(x) Q(x \vee u)dx \nonumber \\
& & \quad \quad \quad \quad \leq 8 b_3 \int_0^1 Q(x)R(x) R(x \vee u) dx \, .
\end{eqnarray}
Hence (\ref{dt16}), (\ref{dt17}) and (\ref{dt18}) together entail that
\begin{equation} \label{dt19}
|A_k| \leq 16b_2 \int_0^u Q(x) R (x)dx + 8b_2 \sum_{ i > [k/2]} \int_0^{\alpha(i)} Q^2(x) dx + 8 b_3 \int_0^1 Q(x) R(x) R(x \vee u) dx \, .
\end{equation}
(\ref{dt19}), (\ref{dt14}), (\ref{dt5}) together with (\ref{cons1lmadmr}) then yield (\ref{1resrio95}).
Notice now that
$$
\sum_{k=1}^n\sqrt{n-k+1} \int_0^{\alpha(k)} Q(x)dx \leq n^{1/2} \int_0^1 (\alpha^{-1} (x)\wedge n ) Q(x) dx \, ,
$$
and that
\begin{eqnarray*}
& & \sum_{k=1}^n\sum_{ i > [k/2]} \int_0^{\alpha(i)} Q^2 (x) dx \leq 2 \sum_{ i \geq 1 } (i \wedge n)
\int_0^{\alpha(i)} Q^2(x)dx \\
& & \quad \leq 2
\int_0^1 Q(x) R(x) (\alpha^{-1}(x)\wedge n )dx
\leq 2n^{1/2} \int_0^1 Q(x)R(x) (R(x)\wedge n^{1/2} )dx \, .
\end{eqnarray*}
Moreover
\begin{eqnarray*}
n^{1/2} \int_0^1 (\alpha^{-1} (x)\wedge n ) Q(x) dx \leq n^{1/2}\int_0^1 Q(x)R(x) (R(x)\wedge n^{1/2} )dx \, .
\end{eqnarray*}
\par
Hence to prove Proposition \ref{condversion}, it remains to select $u=u_k$ in such a way that
\begin{equation} \label{b2}
\sum_{k=1}^n \int_0^{u_k} Q (x) R ( x) dx + \sum_{k=1}^n
\frac{1}{\sqrt{k}} \int_0^1 Q(x)R(x) R(x\vee u_k) dx
\leq C n^{1/2} M_{3,\alpha} (Q ,n^{1/2})\, .
\end{equation}
Let $R^{-1}(y) = \inf\{ v \in [0,1] \, : \, R(v) \leq y \}$ be the right continuous inverse of $R$.
Since $R$ is right continuous, $x < R^{-1}(y)$ if and only if $R(x) > y$. We now choose
$u_k = R^{-1}( k^{1/2} )$, so that
\begin{equation}\label{defuk}
R (u_k) \leq k^{1/2} \ \text{and}\ R(x) > k^{1/2} \ \text{for any}\ x<u_k.
\end{equation}
With this choice of $u_k$, on one hand,
\begin{eqnarray} \label{b3}
\sum_{k=1}^n \int_0^{u_k} Q(x) R(x) dx & = &
\int_0^{1} Q(x) R(x) \sum_{k=1}^n \I_{R(x) > \sqrt{k}} dx \leq \int_0^{1} Q(x)R(x) (R^2(x) \wedge n) dx \nonumber \\
& \leq & n^{1/2}\int_0^{1} Q (x) R(x) (R(x) \wedge n^{1/2}) dx \, .
\end{eqnarray}
On the other hand
\begin{equation} \label{b40}
\sum_{k=1}^n \frac{1}{\sqrt{k}} \int_0^{1} Q(x) R(x)R (x\vee u_k) dx
\leq
\sum_{k=1}^n \frac{1}{\sqrt{k}} \int_{u_k}^1 Q(x) R^2 (x) dx \\
+\sum_{k=1}^n \int_0^{u_k} Q(x) R(x) dx
\end{equation}
using (\ref{defuk}). Next
\begin{equation} \label{b42}
\sum_{k=1}^n \frac{1}{\sqrt{k}} \int_{u_k}^1 Q(x)R^2 (x) dx \leq
\sum_{k=1}^n \frac{1}{\sqrt{k}} \int_{u_n}^1 Q(x)R^2 (x) dx
\leq 2 n^{1/2} M_{3,\alpha} (Q ,n^{1/2})\, .
\end{equation}
Combining (\ref{b40}) with (\ref{b42}) and (\ref{b3}), we then get (\ref{b2}) ending the proof of the proposition.
$\diamond$
\begin{Proposition} \label{inegamax} For $f$ in $\widetilde{\mathcal{F}}( Q, P_{Y_0})$,
let $X_i = f(Y_i) - \bkE ( f(Y_i))$. Set $S_n^* = \max_{1 \leq k\leq n}| S_k|$. Assume that $M_{2,\alpha} (Q) < \infty$. Then the series
$\bkE (X_0^2) + 2 \sum_{k \geq 1} \bkE (X_0 X_k)$ is convergent to some nonnegative real $\sigma^2$ and for any positive real $\lambda$,
\begin{eqnarray*}
\BBP (S_n^* \geq 5 \lambda) & \leq & c_1\exp \Bigl( - \frac{\lambda^2}{c_2n
\sigma^2} \Bigr) + c_3 n \lambda^{-3} M_{3,\alpha} (Q , \lambda) + c_4 n \sigma^3 \lambda^{-3} \, ,
\end{eqnarray*}
where $M_{3,\alpha} ( Q , n^{1/2})$ is defined in (\ref{3alphatronq}) and $c_1$, $c_2$, $c_3$ and $c_4$ are positive constants not depending on $\sigma^2$, so that the first term vanishes if $\sigma^2 =0$.
\end{Proposition}
\noindent{\bf Proof of Proposition \ref{inegamax}.}
Assume first that $X_i = \sum_{\ell=1}^L a_{\ell} f_{\ell}(Y_i)
- \sum_{\ell=1}^L a_{\ell}\bkE(f_{\ell}(Y_i))$, with $f_\ell$
belonging to $\tMon(Q, P_{Y_0})$ and $\sum_{\ell=1}^L |a_\ell|
\leq 1$. Let $M>0$ and $g_M(x) = (x \wedge M) \vee (-M)$. For
any $i \geq 0$, we first define
\[
X_i'=
\sum_{\ell=1}^L a_{\ell}\big( g_M \circ f_{\ell}(Y_i)
- \bkE(g_M \circ f_{\ell}(Y_i)) \big ) \quad \text{and} \quad X_i''=X_i - X_i' \, .
\]
Let $q$ be a positive integer such that $q \leq n$. Let us first show that
\begin{equation} \label{dec1FN}
\max_{1 \leq k \leq n } |S_k | \leq
\max_{ 1 \leq k \leq n} | \bkE (S_n |\F_k)|
+ 2 q M + \max_{1 \leq k \leq n } \bkE_k \big (\sum_{i=1}^n |X_i''| \big ) + \max_{1 \leq k \leq n } \bkE_k \big (\sum_{i=1}^n |\bkE_{i-q} (X_i')| \big ) \, .
\end{equation}
Notice that
$$
S_k = \bkE (S_n |\F_k) - \sum_{i=k+1}^n \bkE (X''_i |\F_k) - \sum_{i=k+1}^n \bkE (X'_i |\F_k) \, .
$$
Now
$$
\sum_{i=k+1}^n \bkE (X'_i |\F_k) = \sum_{i=k+1}^n \bkE (X'_i - \bkE_{i-q} (X_i') |\F_k) - \sum_{i=k+1}^n \bkE ( \bkE_{i-q} (X_i') |\F_k) \, .
$$
The inequality (\ref{dec1FN}) follows by noticing that
$$
\sum_{i=k+1}^n \bkE (X'_i - \bkE_{i-q} (X_i') |\F_k) = \sum_{i=k+1}^{q+k} (\bkE_k (X'_i) - \bkE_{i-q} (X_i') ) \leq 2qM \, .
$$
Notice now that $(\bkE (S_n |\F_k))_{k \geq 1}$, $\Big (\bkE_k \big (\sum_{i=1}^n |X_i''| \big ) \Big )_{k \geq 1}$ and
$\Big (\bkE_k \big (\sum_{i=1}^n |\bkE_{i-q} (X_i')| \big ) \Big )_{k \geq 1}$ are martingales with respect to the filtration
$(\F_k)_{k \geq 1}$. Consequently from (\ref{dec1FN}) and the Doob maximal inequality, we infer that for any nondecreasing, non negative, convex and even function
$\varphi$ and if $qM \leq \lambda$,
\begin{eqnarray} \label{decproba}
\BBP (S_n^* \geq 5 \lambda) &\leq & \frac{\bkE(\varphi(S_n))}{\varphi(\lambda)} + \lambda^{-1}\sum_{i=1}^n \bkE |X_i''| +
\lambda^{-1}\sum_{i=1}^n \| \bkE_{i-q} (X_i') \|_1 \, .
\end{eqnarray}
Choose $u = R^{-1}(\lambda)$, $q = \alpha^{-1} (u) \wedge n $ and $M = Q (u)$. Since $R$ is right
continuous, we have $R(u)\leq \lambda$, hence $qM \leq R(u) \leq \lambda$. Note also that
\begin{equation} \label{dec13FN}
\sum_{k=1}^n \bkE ( | X_k''|)
\leq 2 n \int_0^u Q (x) dx
\leq 2n \int_0^1 Q(x) \I_{R(x)> \lambda} dx \, .
\end{equation}
In addition using Proposition \ref{covineg}, we get that
\begin{equation}
\label{majnorm1prime}
\| \bkE ( X'_i | {\mathcal F}_{i-q} )\|_1 \leq 8 \int_0^{\alpha (q)/2} Q(x) dx \, .
\end{equation}
Since $\alpha(q) /2 \leq u$,
$$
\sum_{i=1}^n \| \bkE_{i-q} (X_i') \|_1 \leq 8 n \int_0^1 Q(x) \I_{R(x) > \lambda} dx \, .
$$
It follows that
\begin{eqnarray} \label{decsecprime}
\lambda^{-1} \big ( \sum_{i=1}^n \bkE |X_i''| +
\sum_{i=1}^n \| \bkE_{i-q} (X_i') \|_1 \big )
& \leq & 10n \lambda^{-1} \int_0^1 Q(x) \I_{R(x) > \lambda} dx \nonumber \\ & \leq & 10n \lambda^{-2} \int_0^1 Q(x) R(x) \I_{R(x) > \lambda} dx\, .
\end{eqnarray}
To control now the first term in the inequality (\ref{decproba}), we choose the even convex function $\varphi$ such that
$$
\varphi(t) = \left\{
\begin{array}{ll}0 & \text{if $0 \leq t \leq \lambda/2$} \\
\frac{1}{6} ( t - \frac{\lambda}{2} )^3 & \text{if $\lambda/2 \leq t \leq \lambda $} \\
\frac{\lambda^3}{48} + \frac{\lambda}{4} ( t - \lambda)^2+ \frac{\lambda^2}{8} ( t - \lambda) & \text{if $t \geq \lambda $} \, .
\end{array}
\right.
$$
Clearly $ \|\varphi^{(2)}\|_{\infty} \leq \lambda/2$ and $\|\varphi^{(3)}\|_{\infty} \leq 1$. Let $(N_i)_{i\in{\mathbb Z}}$
be a sequence of independent random variables
with normal distribution ${\mathcal N}(0, \sigma^2)$. Suppose furthermore that the sequence $(N_i)_{i\in{\mathbb Z}}$ is independent of $(X_i)_{i\in{\mathbb N}}$.
Set $T_n = N_1 + N_2 + \cdots + N_n$ and
$\varphi_k (x) = {\mathbb E} ( \varphi (x + T_n - T_k) )$. With this notation
$$
{\mathbb E} ( \varphi (S_n) - \varphi (T_n) ) = \sum_{k=1}^n {\mathbb E} ( \varphi_k (S_{k-1} + X_k) - \varphi_k ( S_{k-1} + Y_k) ).
$$
To bound up ${\mathbb E} ( \varphi_k (S_{k-1} + X_k) - \varphi_k ( S_{k-1} + Y_k) ) $, we proceed as in the proof of Proposition \ref{condversion} with the following modifications. Firstly, $b_2 = \| \varphi^{(2)}_k\|_{\infty} \leq \lambda/2 $
and $b_3 = \| \varphi^{(3)}_k\|_{\infty} \leq 1 $. The following convention is also used: $S_0 = 0$ and for any positive integer $j$, $S_{-j}= - \sum_{i=1}^{j}X_{1-i}$. Notice that here the $\varphi_k$ are deterministic. Consequently $\bkE ( \varphi'_k (0) X_k) = 0$ and $\varphi_k'' (S_{\ell})$ is always
${\mathcal F}_{\ell}$-measurable for any $\ell \in {\mathbb Z}$. We then infer that the following bound is valid: for any $k=1, \dots, n$,
$$
{\mathbb E} ( \varphi_k (S_{k-1} + X_k) - \varphi_k ( S_{k-1} + Y_k) ) \leq
\sigma^3 + C \lambda \int_0^u Q(x) R(x) dx + C\int_0^1 Q (x) R(x) R(x\vee u) dx \, ,
$$
where $C$ is a positive constant not depending on $\sigma^2$.
Choosing $u= R^{-1}(\lambda)$, we get that
\begin{eqnarray*}
\int_0^{u} Q(x) R(x) dx = \int_0^{1} Q(x) R(x) \I_{R(x) > \lambda} dx \, ,
\end{eqnarray*}
and
\begin{eqnarray*}
\int_0^{1} Q(x)R(x) R(x\vee u) dx \leq \int_0^{1} Q (x) R(x) (R(x) \wedge \lambda) dx \, .
\end{eqnarray*}
It follows that
\begin{equation} \label{1resphi}
{\mathbb E} ( \varphi (S_n) - \varphi (T_n) ) \leq n \sigma^3 + 2 C n
M_{3,\alpha} (Q, \lambda) \, .
\end{equation}
It remains to compute ${\mathbb E} ( \varphi (T_n) ) $. We have that
$6{\mathbb E} ( \varphi (T_n) ) \leq {\mathbb E} \big ( T_n - \lambda/2 \big)_+^3$.
Hence, using the fact that $t^2 = \lambda^2/4 + (t-\lambda/2)^2 + \lambda (t-\lambda/2)$, we
obtain:
$$\bkE(\varphi(T_n)) \leq \frac{e^{-\lambda^2/(8n\sigma^2)}}{6}
\int_{0}^{\infty} e^{-\lambda x/(2n \sigma^2)} \frac{x^3}{\sigma \sqrt{2 n
\pi}} dx \, .
$$
Using the change of variables $y= \lambda x/(2n\sigma^2)$, we derive that
\begin{equation} \label{majgauss} \bkE(\varphi(T_n)) \leq \frac{\lambda^3}{\sqrt{2\pi} }
\Bigl( \frac{(2n\sigma^2)}{\lambda^2} \Bigr)^{7/2} e^{-\lambda^2/(8n\sigma^2)} \, .
\end{equation}
Starting from (\ref{decproba}) and collecting the bounds (\ref{decsecprime}), (\ref{1resphi}) and (\ref{majgauss}), the proposition is proved for any variable $X_i=f(Y_i) - \bkE(f(Y_i)) $ with
$f=\sum_{\ell=1}^L a_{\ell} f_{\ell}$ and $f_{\ell} \in \tMon (Q, P_{Y_0})$,
$\sum|a_\ell|\leq 1$. Since these functions are dense in $\widetilde{\mathcal{F}} (Q, P_{Y_0})$ by
definition, the result follows by applying Fatou's lemma.
\begin{Proposition} \label{covineg} Let $m$ and $q$ be two nonnegative integers.
For any $(m+q)$-tuple of integers $(j_{\ell})_{1\leq \ell \leq m+q}$, let $X^{(j_{\ell})}_i = f_{j_{\ell}}(Y_i) - \bkE ( f_{j_{\ell}}(Y_i))$, where $f_{j_{\ell}}$ belongs to $\widetilde{\mathcal{F}}( Q_{j_{\ell}}, P_{Y_0})$ for $1 \leq \ell \leq m+q$. Suppose that $Q_{j_{\ell}}^q$
is integrable for $\ell \geq m+1$. Define the coefficients
$\alpha_{k, {\bf Y}}(n) $ as in
(\ref{defalpha}). Then for any integers $(j_{\ell})_{1 \leq \ell \leq m+q}$ and any integers $(k_{\ell})_{1 \leq \ell \leq m+q}$ such that
$k_1\leq k_2 \leq \cdots \leq k_{m+q}$ and $k_{m+1}-k_m = \ell$, \begin{eqnarray*} \label{cov1}
\left \| \prod_{i=1}^{m}X^{(j_{i})}_{k_i}\Big ( \bkE_{k_m} \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i} \big ) - \bkE \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i} \big ) \Big )
\right \|_1 \leq 2^{m+q+2} \int_0^{2^{q-2} \alpha_{q, {\bf Y}}(\ell) } \prod_{i=1}^{m+q} Q_{j_i}(x) dx \, ,
\end{eqnarray*}
and
\begin{eqnarray*} \label{cov1}
\left \| \prod_{i=1}^{m}f_{j_{i}}(Y_{k_i})\Big ( \bkE_{k_m} \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i} \big ) - \bkE \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i} \big ) \Big )
\right \|_1 \leq 2^{q+2} \int_0^{2^{q-2} \alpha_{q, {\bf Y}}(\ell) }\prod_{i=1}^{m+q} Q_{j_i}(x) dx \, ,
\end{eqnarray*}
with the convention that $\prod_{i=1}^0 =\prod_{i=m+1}^m = 1$.
\end{Proposition}
{\bf Proof of proposition \ref{covineg}.} Assume first that $f_{j_{\ell}}= \sum_{r=1}^N a_{r} g_{j_{\ell},r}$ where $ \sum_{r=1}^N |a_{r}| \leq 1$ and $g_{j_{\ell},r}$ belongs to $\tMon( Q_{j_{\ell}}, P_{Y_0})$ for $1 \leq \ell \leq m+q$. To soothe the notation, let also
\beq \label{notasimpX}
X^{(j_{\ell})}_{i, r} = g_{j_{\ell},r}(Y_i)
- \bkE(g_{j_{\ell}, r }(Y_i)) \, .
\eeq
We then have that
\begin{eqnarray*}
& & \left \| \prod_{i=1}^{m}X^{(j_i)}_{k_i}\Big ( \bkE_{k_m} \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i} \big ) - \bkE \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i} \big ) \Big )
\right \|_1 \\
& & \leq \prod_{p=1}^{m+q} \big ( \sum_{r_p =1}^N
|a_{r_p}| \big )
\left \| \prod_{i=1}^{m}X^{(j_i)}_{k_i, r_i}\Big ( \bkE_{k_m} \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i} \big ) - \bkE \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i} \big ) \Big )
\right \|_1 \, .
\end{eqnarray*}
Now setting
$$A: =\Big | \prod_{i=1}^{m}X^{(j_i)}_{k_i, r_i} \Big | \sign \left \{ \bkE_{k_m} \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i} \big ) - \bkE \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i} \big )\right \} , $$ we get that
\begin{eqnarray*}
& & \left \|\prod_{i=1}^{m}X^{(j_i)}_{k_i, r_i}\Big ( \bkE_{k_m} \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i} \big ) - \bkE \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i} \big )\Big )
\right \|_1 \\
& & = \bkE \left ( A \Big (\bkE_{k_m} \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i} \big ) - \bkE \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i} \big ) \Big )\right )
= \bkE \left ( (A - \bkE(A)) \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i} \right ) \, .
\end{eqnarray*}
From Proposition A.1 and Lemma A.1 in Dedecker and Rio (2008), we have that
\begin{eqnarray*}
\bkE \left ( (A - \bkE(A)) \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i} \right ) \leq 2^{q+2} \int_0^{\bar \alpha /2} Q_{|A|}(x) \prod_{i=m+1}^{m+q} Q_{j_i}(x)dx\, ,
\end{eqnarray*}
where %
\[
\bar \alpha = \sup_{(t_1, \ldots , t_{q+1}) \in {\mathbf R}^{q+1}}\Big | \bkE \big ( (\I_{ A
\leq t_1} - \p (A \leq t_1)) \prod_{i=m+1}^{m+q}(\I_{ g_{j_i,r_i} (Y_{k_i}) \leq t_{i- m+1} } -
\p ( g_{j_i,r_i} (Y_{k_i}) \leq t_{i - m+1}) ) \big )\big | \, .\]
By monotonocity of the functions $g_{j_i,r_i}$, we then get that
\begin{eqnarray*}
\bar \alpha &\leq & 2^{q}\sup_{(t_1, \ldots ,t_{q+1}) \in {\mathbf R}^{q+1}} \big | \bkE \big ( (\I_{ A
\leq t_1} - \p (A \leq t_1)) \prod_{i=m+1}^{m+q}(\I_{Y_{k_i} \leq t_{i- m+1} } - \p (Y_{k_i} \leq t_{i - m+1})) \big ) \big | \\ & \leq & 2^{q-1} \alpha_{q, {\bf Y}}(\ell) \, .
\end{eqnarray*}
Consequently,
\begin{eqnarray*}
& & \bkE \left ( (A - \bkE(A)) \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i} \right ) \leq 2^{q+2} \int_0^{2^{q-2} \alpha_{q, {\bf Y}}(\ell) } Q_{|A|}(x) \prod_{i=m+1}^{m+q} Q_{j_i}(x)dx \\
& & \leq 2^{q+2} \int_0^{2^{q-2} \alpha_{q, {\bf Y}}(\ell) } \prod_{i=1}^m \big ( Q_{j_i}(x) + \int_0^1 Q_{j_i}(x) dx \big ) \prod_{i=m+1}^{m+q} Q_{j_i}(x)dx \, .
\end{eqnarray*}
Hence taking into account that $\prod_{i=1}^{m+q} \big ( \sum_{r_i =1}^N
|a_{r_i}| \big ) \leq 1$ and using Lemma 2.1 in Rio (2000), the inequality is proved for functions $f_{j_{\ell}}= \sum_{r=1}^N a_{r} g_{j_{\ell},r}$ where $ \sum_{r=1}^N |a_{r}| \leq 1$ and $g_{j_{\ell},r}$ belongs to $\tMon( Q_{j_{\ell}}, P_{Y_0})$ for $1 \leq \ell \leq m+q$.
It remains to prove that the inequality remains valid for $f_{j_{\ell}}$ belonging to
$\widetilde{\mathcal{F}}( Q_{j_{\ell}}, P_{Y_0})$ for $1 \leq \ell \leq m+q$. By definition,
$$
X^{(j_{\ell})}_{i} = \lim_{N \rightarrow \infty} {\mathbb L}^1 \sum_{r=1}^N a_{r,N}X^{(j_{\ell})}_{i, r, N} \, ,
$$
where $\sum_{r=1}^N |a_{r,N} |\leq 1$ and $
X^{(j_{\ell})}_{i, r, N} = g_{j_{\ell},r,N}(Y_i)
- \bkE(g_{j_{\ell}, r, N }(Y_i)) $ with the $g_{j_{\ell}, r, N }$ belonging to
$\tMon( Q_{j_{\ell}}, P_{Y_0})$ for $1 \leq \ell \leq m+q$.
Hence, by Fatou lemma the proposition will hold if we can prove that the
following inequality holds almost surely
\begin{eqnarray} \label{argfatou}
& & \bkE_{k_m} \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i} \big ) - \bkE \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i} \big )
\nonumber \\
& & \quad \quad =\lim_{N \rightarrow \infty} \prod_{i=m+1}^{m+q} \big (\sum_{r_i=1}^N a_{r_i,N} \big )
\Big (\bkE_{k_m} \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i, N} \big ) - \bkE \big ( \prod_{i=m+1}^{m+q}X^{(j_i)}_{k_i, r_i, N} \big ) \Big ) \, .
\end{eqnarray}
With this aim, notice that for any $m+1 \leq \ell \leq m+q$,
$$
X^{(j_{\ell})}_{i} = \sum_{r=1}^N a_{r,N}X^{(j_{\ell})}_{i, r, N} + \epsilon^{(j_{\ell})}_{i,N} \, ,
$$
with $\lim_{N \rightarrow \infty} \Vert \epsilon^{(j_{\ell})}_{i,N} \Vert_1 =0$. In addition, since for $m+1 \leq \ell \leq m+q$, $Q_{j_\ell}^q$ is integrable and $g_{j_{\ell}, r, N }$ belongs to $\tMon( Q_{j_{\ell}}, P_{Y_0})$,
it follows that $\Vert X^{(j_{\ell})}_{i, r, N} \Vert_q \leq 2 \Vert Q_{j_\ell}\Vert_q$ and next
$\Vert X^{(j_{\ell})}_{i}\Vert_q \leq 2 \Vert Q_{j_\ell}\Vert_q$ by an application of Fatou lemma. Consequently the $\epsilon^{(j_{\ell})}_{i,N}$'s are in ${\mathbb L}^q$ and satisfy
$\Vert \epsilon^{(j_{\ell})}_{i,N} \Vert_q \leq 4 \Vert Q_{j_\ell}\Vert_q$. Now
\begin{eqnarray*}
\Vert \epsilon^{(j_{m+1})}_{k_{m+1},N} \prod_{i=m+2}^{m+q}X^{(j_i)}_{k_i} \Vert_1 & \leq & 2^{q-1}
\int_0^1 Q_{|\epsilon^{(j_{m+1})}_{k_{m+1},N}|} (x) \prod_{i=m+2}^{m+q}Q_{j_i}(x) dx \\
& \leq & 2^{q-1}
\int_0^1 Q_{|\epsilon^{(j_{m+1})}_{k_{m+1},N}|} (x) Q^{q-1}_*(x) dx \,
,
\end{eqnarray*}
where $Q_*= \max_{m+2 \leq i \leq m+q}Q_{j_i}$. Now for any positive $M$, $Q^{q-1}_* \leq M^{q-1} + Q^{q-1}_* \I_{Q_* >M}$. Hence,
\begin{eqnarray*}
2^{1-q} \Vert \epsilon^{(j_{m+1})}_{k_{m+1},N} \prod_{i=m+2}^{m+q}X^{(j_i)}_{k_i} \Vert_1 & \leq & M^{q-1} \Vert \epsilon^{(j_{m+1})}_{k_{m+1},N}\Vert_1 + \Vert \epsilon^{(j_{m+1})}_{k_{m+1},N} \Vert_q \Vert Q_* \I_{Q_* >M} \Vert_q^{q-1} \\
& \leq & M^{q-1} \Vert \epsilon^{(j_{m+1})}_{k_{m+1},N}\Vert_1 + 4 \Vert Q_{j_{m+1}}\Vert_q \Vert Q_* \I_{Q_* >M} \Vert_q^{q-1} \, ,
\end{eqnarray*}
which tends to zero by letting first $N$ tends to infinity and after $M$. Similarly, we can show that for any $\ell \in \{ 1, \dots, q-1 \}$,
$$
\lim_{N \rightarrow \infty}\Big \| \Big ( \prod_{i=1}^{\ell} X^{(j_{m+i})}_{k_{m+i}, r, N} \Big ) \epsilon^{(j_{m+\ell +1})}_{k_{m+\ell + 1},N} \prod_{i=m+\ell +2}^{m+q}X^{(j_i)}_{k_i} \Big \|_1 =0 \, .
$$
This ends the proof of (\ref{argfatou}) and then of the proposition.
|
\section{Introduction}
Dynamical properties of many particle system can generally be studied by
employing an external perturbation, which disturbs the system only slightly
from its equilibrium state, and thus measuring the spontaneous
response/fluctuations of the system to this external perturbation.
In general, the fluctuations are related to the correlation function through
the symmetry of the system, which provides important inputs for quantitative
calculations of complicated many-body system. Also, many of the properties
of the deconfined strongly interacting matter are reflected in the structure
of the correlation and the spectral functions~\cite{hasimoto} of the
vector current.
The static thermal dilepton rate describing the production of lepton pairs is
related to the spectral function in the vector current~\cite{munshi1,karsch}.
Within the Hard thermal loop perturbation theory (HTLpt) the vector spectral
function has been obtained~\cite{munshi1,beradau,moore}, which is found
to diverge due to its spatial part at the low
energy regime. This is due to the fact that the HTL quark-photon vertex
is inversely proportional to the photon energy and it sharply rises at
zero photon energy. On the other hand, the fluctuations of conserved
quantities, such as baryon number and electric charge, are considered
to be a signal~\cite{expt,stephanov} for quark-gluon plasma (QGP)
formation in heavy-ion experiments. These conserved density fluctuations
are closely related to the
temporal correlation function in the vector channel through derivatives
of a thermodynamic quantity associated with the symmetry, known as
the thermodynamic sum rule~\cite{calen}. It is expected that the
temporal part of the spectral function associated with the symmetry
should be a finite quantity and would not encounter any such infrared
divergence unlike the spatial part at low energy. A very recent lattice
calculation~\cite{karsch2} in quenched approximation has obtained the
temporal part of the Euclidian
correlation function associated with the response of the conserved
density fluctuations, which is found to be a finite quantity. In view
of this we would like to compute the
temporal correlation function in the vector channel from the quark number
susceptibility associated with the quark number density fluctuations
within the HTLpt~\cite{munshi2} and to compare it with recent lattice
data~\cite{karsch2} in quenched approximation.
The temporal correlator follows from the QCD polarization diagram. To lowest
order perturbation theory it is given by the one-loop diagram containing bare
quark propagators. The HTL resummation technique provides a consistent
perturbative expansion for gauge theories at finite temperature by using
HTL resummed propagators and vertices \cite{braaten}. As in usual perturbation
theory it is strictly applicable only in the weak coupling limit but takes
into account all dynamical effects. Going beyond the lowest order perturbation
theory in the case of the temporal correlator, we use HTL resummed quark
propagators and quark-gluon vertices in the polarization diagram
(see \cite{munshi2,munshi}). The HTL resummed quark propagators correspond to
static external quarks (valence quarks). Within this approximation no internal
quark-loops appear. In this sense our approximation can be compared to results
from quenched lattice QCD. The inclusion of dynamical quark-loops requires to
consider higher-order diagrams within the HTL resummed perturbation theory in
which HTL resummed gluon propagators will show up. However, we do not expect
a significant change in the result as higher contributions give less than 5\%
corrections in the case of thermodynamic quantities such as pressure
\cite{andersen1}. Of course, close to the transition temperature higher order
effects may become important~\cite{andersen2}.
The paper is organised as follows. In sec.II we briefly discuss some
generalities on correlation functions, fluctuation and its response
(susceptibility) associated with conserved charges.
In sec.III we obtain the relation between the response of the density
fluctuation of the conserved charge and the corresponding temporal part of
the Euclidian correlation function in the vector current. Next we compute them in
HTLpt~\cite{munshi2} and compare with lattice data. Finally, we conclude
in sec.VI.
\section{Generalities}
In this section we summarize some of the basic relations and also
describe in details their important features relevant as well as
required for our purpose.
\subsection{Correlation Functions}
The two-point correlation function~\cite{hasimoto,munshi1,karsch} of
the vector current,
$J_\mu={\bar {\psi}}(\tau,{\vec {\mathbf x}})\Gamma_\mu \psi(\tau,{\vec
{\mathbf x}})$ with three point function $\Gamma_\mu$,
is defined at fixed momentum ${\vec {\mathbf p}}$ as
\begin{eqnarray}
G_{\mu\nu}(\tau,{\vec{\mathbf p}})=\int d^3x \ \langle J_\mu(\tau,
{\vec {\mathbf x}}) J_\nu^\dagger(0, {\vec {\mathbf 0}})\rangle \
e^{i{\vec{\mathbf p}}\cdot {\vec{\mathbf x}}} \ , \label{cor_p}
\end{eqnarray}
where the Euclidian time $\tau$ is restricted to the interval
$[0,\beta=1/T]$. The thermal two-point vector correlation function
in coordinate space, $ G_{\mu\nu}(\tau, {\vec {\mathbf x}})$, can be written as
\begin{equation}
G_{\mu\nu}(\tau, {\vec {\mathbf x}})=\langle J_\mu(\tau, {\vec {\mathbf x}})
J_\nu^\dagger(0, {\vec {\mathbf 0}})\rangle =T\sum_{n=-\infty}^{\infty} \int
\frac{d^3p}{(2\pi)^3} \ e^{-i(w_n\tau+{\vec{\mathbf p}}
\cdot {\vec{\mathbf x}})}\ G_{\mu\nu}(w_n,{\vec{\mathbf p}}) \ ,
\label{vec_cor}
\end{equation}
where the Fourier transformed correlation function
$G_{\mu\nu}(w_n,{\vec{\mathbf p}})$ is given at the discrete
Matsubara modes, $w_n=2\pi n T$. The imaginary part of the momentum space
correlator gives the spectral function $\sigma(\omega,{\vec{\mathbf p}})$ as
\begin{eqnarray}
G_H(w_n,{\vec{\mathbf p}}) &=& - \int_{-\infty}^{\infty} d\omega
\frac{\sigma_H(\omega,{\vec{\mathbf p}})}{iw_n-\omega+i\epsilon} \nonumber \\
\ \Rightarrow \
\sigma_H(\omega,{\vec{\mathbf p}})&=& \frac{1}{\pi} {\mbox{Im}}
\ G_H(\omega,{\vec{\mathbf p}}) \ , \label{spec_mom}
\end{eqnarray}
where $H=(00,ii,V)$ denotes (temporal, spatial, vector). We have
also introduced the vector spectral function
as $\sigma_V=\sigma_{00}+ \sigma_{ii}$, where $\sigma_{ii}$ is the sum
over the three space-space components and $\sigma_{00}$ is the time-time
component of $\sigma_{\mu\nu}$.
Using (\ref{vec_cor}) and (\ref{spec_mom}) in (\ref{cor_p})
the spectral representation of
the thermal correlation functions at fixed momentum
can be obtained~\cite{munshi1} as
\begin{equation}
G_H(\tau, {\vec {\mathbf p}})=\int_0^\infty\ d\omega \
\sigma_H (\omega, {\vec {\mathbf p}}) \
\frac{\cosh[\omega(\tau-\beta/2)]}{\sinh[\omega\beta/2]} \ . \label{corr_mom}
\end{equation}
We note that the Euclidian correlation function is
usually restricted to vanishing three momentum, $\vec{\mathbf p}=0$,
in the analysis of lattice gauge theory and one can write
$G_H(\tau T)=G_H(\tau,{\vec{\mathbf 0}})$.
A finite temperature
lattice gauge theory calculation is performed on
lattices with finite temporal extent $N_\tau$, which provides information
on the Euclidian correlation function, $G_H(\tau T)$, only for a discrete
and finite set of Euclidian times
$\tau =k/(N_\tau T), \ \ k=1,\cdots \ N_\tau$. The vector correlation
function, $G_V(\tau T)$, had been computed~\cite{karsch} within the
quenched approximation of QCD\footnote{In comparison to thermodynamic
quantities some quantities like mesonic correlation and spectral functions
due to their structures are still too exhaustive and expensive to
calculate in full QCD with improved lattice action. Hence the quenched
approximation is still very useful to understand the various features of
correlation and spectral functions.}
using
non-perturbative improved clover fermions~\cite{clov} through a probabilistic
application based on the maximum entropy method (MEM)~\cite{mem} for
temporal extent $N_\tau= 16$ and spatial extent $N_\sigma = 64$. Then by
inverting the integral in (\ref{corr_mom}), the spectral function\footnote{We
note that lattice technique cannot directly be used to obtain the spectral
function due to its difficulty to perform an analytic continuation of
(\ref{spec_mom}) from imaginary time to real time.} was
reconstructed~\cite{karsch,aarts} in lattice QCD. The vector spectral functions
above the deconfinement temperature ({\em viz.}, $T=1.5T_c \ {\mbox{and}}
\ 3T_c$) show an oscillatory behaviour compared to the free one. In the high energy
regime, $\omega/T\ge 4$ the vector spectral function, $\sigma_V(\omega,
{\vec{\mathbf 0}})$ agreed with
that of the HTLpt~\cite{munshi1,carsten}. On the other hand, the
lattice spectral
functions and dilepton rates~\cite{karsch} were found to fall off very
fast and became vanishingly small for $\omega/T\le 4$ due to the sharp
cut-off used in the reconstruction.
In a very recent lattice analysis~\cite{karsch2} the low energy
behaviour ({\it viz.}, downward slope) of the spatial and vector spectral
functions has been improved substantially on lattices up to size
$128^3\times 48$ by changing the slope upward through a fit to the free
plus Breit-Wigner (BW) spectral functions at lower energy limit. This upward
slope in the vector spectral functions in lattice gauge theory~\cite{karsch2}
resembles up to some extent that of the HTLpt spectral
function~\cite{munshi1,beradau}
and dilepton rate~\cite{moore,carsten} at low energy regime despite the
infrared problem of HTLpt at vanishing energy. Nonetheless, the existence
of the van Hove peaks in the vector spectral function
in HTLpt~\cite{munshi1,yaun,markus, wong} has not
been realized yet in the improved lattice analysis~\cite{karsch2}, probably
due to the ansatz that the lattice spectral function is proportional to
that of the free plus BW one in the low energy regime. The existence of
van Hove peaks cannot yet be ruled out, which are general features of
massless fermions~\cite{yaun,markus,wong} in a relativistic plasma in
the low energy regime. Also, the high energy behaviour of
the spectral function agrees well with those of free and HTLpt results,
respectively, and are in conformity with its earlier
analysis on the lattice~\cite{karsch}. On the other hand, the temporal component of
spectral and correlation functions associated with the symmetry of the system
are finite and do not encounter any
infrared problem unlike their spatial part. Now, it would be interesting
to analyse the temporal correlation and spectral functions, associated
with the symmetry, within HTLpt and compare them with those of lattice
gauge theory~\cite{karsch2}.
\subsection{ Density Fluctuation and its Response}
Let \( {\cal O}_{\alpha } \) be a Heisenberg operator where $\alpha$
may be associated with a degree of freedom in the system.
In a static and uniform
external field \( {\cal F}_{\alpha } \), the (induced)
expectation value of the operator \( {\cal O}_\alpha \left( 0,
\overrightarrow{\mathbf x}
\right) \) is written~\cite{calen,kunihiro} as
\begin{equation}
\phi _{\alpha }\equiv \left\langle {\cal O} _{\alpha }\left
( 0,\overrightarrow{\mathbf x}\right) \right\rangle _{\cal F} = \frac{{\rm Tr}\left
[ {\cal O} _{\alpha }\left( 0,\overrightarrow{\mathbf x}\right) e^{-\beta \left
( {\cal H}+{\cal H}_{ex}\right) }\right] }{{\rm Tr}\left[ e^{-\beta
\left( {\cal H}+{\cal H}_{ex}\right) }
\right] }=\frac{1}{V}\int d^{3}x\, \left\langle {\cal O} _{\alpha }
\left( 0,\overrightarrow{\mathbf x}\right) \right\rangle \: , \label{eq1}
\end{equation}
where translational invariance is assumed, $V$ is the volume of the
system and
\({\cal H}_{ex} \) is given by
\begin{equation}
{\cal H}_{ex}=-\sum _{\alpha }\int d^{3}x\, {\cal O} _{\alpha }\left( 0,
\overrightarrow{\mathbf x}\right) {\cal F}_{\alpha }\: .\label{eq2}
\end{equation}
The (static) susceptibility \( \chi _{\alpha \sigma } \) is defined as
the rate with which the expectation value changes
in response to that external field,
\begin{eqnarray}
\chi _{\alpha \sigma }(T) & = & \left. \frac{\partial \phi _{\alpha }}
{\partial {\cal F}_{\sigma }}\right| _{{\cal F}=0}
= \beta \int d^{3}x\, \left\langle {\cal O} _{\alpha }\left
( 0,\overrightarrow{\mathbf x}\right) {\cal O} _{\sigma }
( 0,\overrightarrow{\mathbf 0})
\right\rangle \: , \label{eq3}
\end{eqnarray}
where
$\langle {\cal O}_\alpha (0,{\vec {\mathbf x}})
{\cal O}_\sigma(0,{\vec {\mathbf 0}})\rangle $
is the two point correlation function with operators evaluated
at equal times. There is no broken symmetry as
\begin{equation}
\left.\left\langle {\cal O} _{\alpha }
\left ( 0,\overrightarrow{\mathbf x}\right ) \right\rangle
\right|_{{\mathcal F}\rightarrow 0}
=\left. \left\langle {\cal O} _{\sigma }
( 0,\overrightarrow{\mathbf 0}) \right\rangle
\right |_{{\mathcal F}\rightarrow 0}=0 \ . \label{eq3i}
\end{equation}
\section{{Quark Number Susceptibility (QNS) and Temporal Euclidian
Correlation Function:}}
The QNS is a measure of the response of the quark number density with
infinitesimal change in the quark chemical
potential, $\mu+\delta\mu $. Under such a situation the external field,
${\cal F}_\alpha$, in ({\ref{eq2}}) can be identified as the
quark chemical potential and the operator ${\cal O}_\alpha$ as
the temporal component ($J_0$) of the vector current,
$J_\sigma(t,{\vec {\mathbf x}})= \overline{\psi} \Gamma_{ \sigma}\psi$,
where $\Gamma^\sigma$ is in general a three point function.
Then the QNS for a given quark flavour follows from
(\ref{eq3}) as
\begin{eqnarray}
\chi_q(T) &=& \left.\frac{\partial \rho}{\partial \mu}\right |_{\mu=0}
= \left.\frac{\partial^2 {\cal P}}{\partial \mu^2}\right |_{\mu=0}
= \int \ d^4x \ \left \langle J_0(0,{\vec {\mathbf x}})J_0(0,{\vec {\mathbf 0}})
\right \rangle \ =- {\lim_{\vec{\mathbf p}\rightarrow 0}}
{\mbox {Re}}\ G_{00}^R(0,{\vec{\mathbf p}}),
\label{eq4}
\end{eqnarray}
where $G_{00}^R$ is the retarded correlation function. To obtain
(\ref{eq4}) in concise form, we have used the fluctuation-dissipation
theorem given as
\begin{equation}
G_{00}(\omega,{\vec{\mathbf p}})=-\frac{2}{1-e^{-\omega/T}} {\mbox{Im}}
G_{00}^R(\omega,{\vec{\mathbf p}}), \label{eq4a}
\end{equation}
and the Kramers-Kronig dispersion relation
\begin{equation}
{\mbox{Re}}G_{00}^R(\omega,{\vec{\mathbf p}})=\int_{-\infty}^{\infty}
\frac{d\omega^\prime}{2\pi}
\frac{{\mbox{Im}}G_{00}^R(\omega,{\vec{\mathbf p}})}{\omega^\prime-\omega},
\label{eq4b}
\end{equation}
where $\lim_{\vec{\mathbf p}\rightarrow 0}{\mbox{Im}} G_{00}^R
(\omega,{\vec{\mathbf p}})$ is proportional to
$\delta(\omega)$ due to the quark number conservation~\cite{calen,kunihiro}.
Also the number density for a given quark flavour can be written as
\begin{equation}
\rho=\frac{1}{V}
\frac{{\rm{Tr}}\left [ {\cal N} e^{-\beta \left({\cal H}-\mu {\cal N}
\right )}\right ]}
{{\rm{Tr}}\left [e^{-\beta \left({\cal H}-\mu {\cal N}\right )}\right ]} \frac{\langle {\cal N}\rangle}{V} = \frac{\partial {\cal P}}
{\partial \mu} \ , \label{eq5}
\end{equation}
with the quark number operator, ${\cal N}=\int J_0(t, {\vec x}) \ d^3x
=\int {\bar \psi}(x)\Gamma_0\psi(x)d^3x$, and
${\cal P}=\frac{T}{V} \ln {\mathcal Z}$ is the pressure and
${\mathcal Z}$ is the partition
function of a quark-antiquark gas. The quark number density vanishes
if $\mu\rightarrow 0$, i.e., there is no broken CP symmetry.
Now, (\ref{eq3}) or (\ref{eq4}) indicates that the thermodynamic derivatives
with respect to the external source
are related to the temporal component of the static correlation function
associated with the number conservation of the system. This relation
in (\ref{eq4}) is known as {\em the thermodynamic sum rule}~\cite{calen}.
Owing to the quark number conservation the temporal spectral function
$\sigma_{00}(\omega,{\vec{\mathbf 0}})$ in (\ref{spec_mom}) becomes
\begin{equation}
\sigma_{00}(\omega,{\vec{\mathbf 0}})=\frac{1}{\pi} {\mbox{Im}} G_{00}^R
(\omega,{\vec{\mathbf 0}})= - \omega \delta(\omega) \chi_q(T) \, . \label{eq6}
\end{equation}
Using (\ref{eq6}) in (\ref{corr_mom}), it is straight forward to obtain
the temporal correlation function as
\begin{equation}
G_{00}(\tau T)=-T\chi_q(T), \label{eq7}
\end{equation}
which is proportional to the QNS $\chi_q$ and $T$, but independent of $\tau$.
The QNS has been calculated within the framework of lattice
gauge theory
\cite{lat1,lat3,peter,bazavov,milc,lat2,qmass,peter1}, perturbative
QCD~\cite{vuorinen}, Nambu-Jona-Lasinio (NJL) Model~\cite{kunihiro,fuji},
Polyakov-Nambu-Jona-Lasinio (PNJL) Model~\cite{ghosh}, Ads/CFT
correspondence and Holographic QCD~\cite{kim}, Renormalisation Group
approach~\cite{schaefer},
two loop approximately self-consistent
$\Phi$-derivable HTL resummation~\cite{blaizot}
and HTLpt~\cite{munshi2,munshi,jiang}. We note that
in a resummed perturbation theory~\cite{pisarski,braaten}, the higher
order loops contribute to the lower order due to the fact that the
loop expansion and the coupling expansion are not symmetric. So, unlike
conventional perturbation theory~\cite{kapusta} one needs to take a proper
measure in order to calculate a quantity in a given order
of $\alpha_s$ correctly using HTLpt. We further note that various HTL
approaches~\cite{blaizot,munshi,jiang,andersen,blaizot1} have been
used for calculating LO thermodynamic quantities and
QNS in the literatures, which led to different results within the same
approximation. Recently, we have developed a resummed HTLpt~\cite{munshi2} by
employing a variation of an external probe that disturbs the system only
slightly from its equilibrium positions. In this way the effect of higher
order variations of the external source is taken into account, which are
essential to get the LO quantities correct. Once this is done the LO
quantities in HTLpt~\cite{munshi2} agree with the HTL
approaches of Ref.~\cite{blaizot,blaizot1}.
Here we would like to exploit the LO QNS in
HTLpt obtained in Ref.~\cite{munshi2} for our purpose.
Because of the structure of the HTL propagator~\cite{pisarski,braaten} the
LO QNS in HTLpt contains quasiparticle (QP) contributions due to the
poles of the HTL quark propagator and Landau damping (LD) contributions
due to the space like part of the HTL quark propagator. The QNS can be
decomposed as
\begin{equation}
\chi_q^{HTL}(T)= \chi_{q}^{QP}+\chi_{q}^{LD}. \label{s5}
\end{equation}
The LO QNS in HTLpt due to QP is obtained~\cite{munshi2}
as
\begin{eqnarray}
\chi_q^{QP}(T)
&=& 4N_cN_f\beta\int \frac{d^3k} {(2\pi)^3}\left [
n(\omega_+)\left(1-n(\omega_+)\right)
\ + \ n(\omega_-)\left(1-n(\omega_-)\right) \right. \nonumber \\
&&\left. \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
-\ n(k)\left(1-n(k)\right)
\right ] \ , \label{H8}
\end{eqnarray}
and the LD part is obtained~\cite{munshi2} as
\begin{eqnarray}
\chi_q^{LD}(T)
=2N_cN_f\beta\int\frac{d^3k}{(2\pi)^3}
\int\limits_{-k}^k d\omega\left(\frac{2m_q^2}{\omega^2-k^2}\right)\
\beta_+(\omega,k)\ n(\omega)\left(1-n(\omega)\right) \ , \label{L4}
\end{eqnarray}
where $n(y)$ is the Fermi-Dirac distribution, $\omega_+$ corresponds to
the energy of a quasiparticle having chirality to helicity ratio $+1$,
$\omega_-$ is the energy of a mode called plasmino having chirality to
helicity ratio $-1$, and $\beta_\pm$ are the cut spectral functions
of the HTL quark propagator. The QP part results in (\ref{H8})
is identical to that of the $2$-loop approximately self-consistent
$\Phi$-derivable HTL resummation approach of Blaizot {\em et al}
~\cite{blaizot}. The LD part (\ref{L4}) cannot be compared directly to
the LD part of Ref.~\cite{blaizot} as no closed expression is given
there. However, numerical results of the both QNS agree very well.
We also note that Jiang {\em et al}~\cite{jiang} used
HTLpt but did not take into account properly the effect of the variation
of the external field to the density fluctuation, which resulted in
an overcounting in the LO QNS. Moreover, in their approach an ad hoc
separation scale is required to distinguish between soft and hard momenta and
the thermodynamic sum rule is violated.
In the HTLpt approach in Ref.~\cite{munshi}
the HTL N-point functions were used uniformly for all momenta scale,
{\em i.e.}, both soft and hard momenta, which resulted in an overcounting
within the LO contribution~\cite{blaizot}. The reason is that the HTL
action is accurate only for
soft momenta and for hard ones only in the vicinity of light cone.
\begin{figure}[!tbh]
\begin{center}
\includegraphics[height=0.55\textwidth, width=0.65\textwidth]{chi.ps}
\caption{(Color online) The 2-flavour scaled QNS with that of free
one as a function of $T/T_C$.
The solid lines are for LO in HTLpt whereas the dashed lines are
for LO (proportional to $g^2$) in pQCD~\cite{blaizot,toimela,kapusta}.
The different choices of the renormalisation
scale are $Q =2\pi T \ {\mbox{(red)}}, \ \
{\mbox{and}} \ \ 4\pi T \ {\mbox{(blue)}}$. The symbols represent the
various lattice data~\cite{lat3,peter,bazavov, milc,lat2}.
The violet triangles (with $T_c=204\pm2$ MeV), brown crosses (with $T_c196\pm 3$ MeV), green triangles (with $T_c=191\pm 2$ MeV) and
purple squares (with $T_c=185 \pm 5$ MeV) represent
$p4$ lattice QCD data~\cite{peter,bazavov}.
The squares (cyan) and stars (saffron) are from asqtad lattice QCD
data~\cite{milc}.
The solid circles (purple) represent quenched QCD
data~\cite{lat2} for $T_c=0.49\Lambda_{\overline{MS}}$.
The quark mass ranges between (0.1 to 0.2)$m_s$,
where $m_s$ is the strange quark mass near its physical value.
Note that further lowering the quark
mass to its physical value seems to have a small effect~\cite{qmass} for
$T> 200$ MeV. The details of these lattice results are also summarised
in Ref.~\cite{peter1}. }
\label{qns_I}
\end{center}
\end{figure}
Now we display in Fig.~\ref{qns_I} the 2-flavour\footnote{We note
that the QNS has a very weak flavour dependence that enters through the
temperature dependence of the strong coupling as
$\alpha_s(T)=\frac{12\pi}{(33-2N_f)\ln(Q^2/\Lambda_0^2)}$
where $Q$ is the momentum scale and $T_C=0.49\Lambda_0$. }
scaled QNS in LO with that of free gas as a function of temperature that
shows significant improvement over pQCD results of order $g^2$
~\cite{blaizot,toimela,kapusta}. Moreover,
it also shows the same trend as the available lattice
results~\cite{lat2,lat3,peter,bazavov,milc,qmass,peter1}, though there
is a large variation among the various lattice results within the improved
lattice (asqtad and p4) actions~\cite{peter,milc} due to the higher order
discretisation of the relevant operator associated with the thermodynamic
derivatives. A detailed analysis
on uncertainties of the ingredients in the lattice QCD calculations is
presented in Refs.~\cite{bazavov,qmass}. This calls for further investigation
both on the analytic side by improving the HTL resummation schemes and on
the lattice side by refining the various lattice ingredients.
\begin{figure}[!tbh]
{\includegraphics[height=0.48\textwidth, width=0.55\textwidth,
angle=270]{htl_corr_00.ps}}
{\includegraphics[height=0.48\textwidth, width=0.55\textwidth,
angle=270]{htl_corr_00_3nf.ps}}
\caption{(Color online) The scaled temporal
correlation function with $T^3$ for $N_f=1$ (left panel) and
$N_f=3$ (right panel) at $T=1.45T_C$ for $Q=2\pi T$ (red) and
$4\pi T$ (blue) as a function of scaled Euclidian time, $\tau T$.
The symbols represent the recent lattice data~\cite{karsch2} on
lattices of size $128^3\times 48$ for quark mass $0.1T$ in quenched
QCD.}
\label{corr1}
\end{figure}
Recently, an improved lattice calculation~\cite{karsch2} has been
performed within the quenched approximation of QCD where the temporal
correlation
function is determined to better than $1\%$ accuracy.
Using the LO HTLpt QNS in (\ref{s5}) we now obtain the temporal
correlation function in (\ref{eq7}) and compare with the recent lattice
data~\cite{karsch2}. In Fig.~\ref{corr1} the scaled temporal correlation
function with $T^3$ is shown for $N_f=1$ (left panel) and $N_f=3$ (right panel)
at $T=1.45T_c$. We first note that the correlation functions both in
HTLpt and pQCD have weak flavour dependence due to the temperature dependent
coupling, $\alpha_s$ as discussed before. The LO HTLpt result indicates
an improvement over that of the pQCD one~\cite{blaizot,toimela,kapusta} for
different choices
of the renormalisation scale as shown in Fig.~\ref{corr1}. Also, the HTLpt
result shows a good agreement
to that of recent lattice gauge theory calculation~\cite{karsch2} performed
on lattices up to size $128^3\times 48$ in quenched approximation for a quark
mass $\sim 0.1T$. We also note that unlike the
dynamical spatial part of the correlation function in the vector channel
the temporal part does not encounter any infrared problem in the
low energy part
as it is related to the static quantity through the thermodynamic sum rule
associated with the corresponding symmetry, {\it viz.}, the number conservation of
the system. We also presented two extreme cases of HTLpt temporal
correlation function at $T=1.2T_C$ in Fig.~\ref{corr2} and at $T\sim 3T_C$
in Fig.~\ref{corr3}, respectively, for two different flavours and
compare with the corresponding preliminary lattice data~\cite{karsch3}, which
are also found to be in good agreement. Finally, we also note that
even if one compares improved lattice action (asqtad) data~\cite{milc}
and recent quenched data~\cite{karsch2} for QNS, the quantitative difference
is within $5\%$ in the temperature domain $T_C\leq T\leq 3T_C$.
\begin{figure}[!tbh]
{\includegraphics[height=0.48\textwidth, width=0.55\textwidth,
angle=270]{htl_corr_00_12tc.ps}}
{\includegraphics[height=0.48\textwidth, width=0.55\textwidth,
angle=270]{htl_corr_00_3nf_12tc.ps}}
\caption{(Color online) Same as Fig.~\ref{corr1} but at
$T=1.2T_C$ and the corresponding lattice data are preliminary~\cite{karsch3}
with
lattice size $148^3\times 40$.}
\label{corr2}
\end{figure}
\begin{figure}[!tbh]
{\includegraphics[height=0.48\textwidth, width=0.55\textwidth,
angle=270]{htl_corr_00_3tc.ps}}
{\includegraphics[height=0.48\textwidth, width=0.55\textwidth,
angle=270]{htl_corr_00_3nf_3tc.ps}}
\caption{(Color online) Same as Fig.~\ref{corr1} but at
$T=2.98T_C$ and the corresponding lattice data are preliminary~\cite{karsch3}
with lattice size $148^3\times 16$.}
\label{corr3}
\end{figure}
\section{Conclusion}
The LO QNS as a response of the conserved density fluctuation in HTLpt when
compared with the available lattice data with improved lattice
actions~\cite{lat3,peter,bazavov,milc,lat2}
in the literature within their wide variation shows the same trend
but deviates from those in certain extent. The same HTL QNS is used to
compute the temporal part of the Euclidian correlation in the vector current
which agrees quite well with that of improved lattice gauge theory
calculations~\cite{karsch2,karsch3} recently performed within the quenched
approximation on lattices up to size $128^3\times48$ for a quark mass
$\sim 0.1T$. It is also interesting to note that the quantitative difference
between the recent quenched approximation data~\cite{karsch2,karsch3} and
the full QCD data with improved (asqtad) lattice action~\cite{milc} for QNS
is within $5\%$ in the temperature range $T_c\le T\le 3T_c$. Leaving aside
the difference
in ingredients in various lattice calculations, one can expect that the HTLpt
and lattice calculations are in close proximity for quantities associated
with the conserved density fluctuation.
\begin{acknowledgments}
We are thankful to P. Petreczky for providing us the lattice data on QNS in
Fig.~\ref{qns_I} and
also for clarifying the details of those data. We are also thankful to
F. Karsch for supplying us the preliminary lattice data in Figs.~\ref{corr2}
and \ref{corr3}, and for very useful discussions. MGM is thankful to
A. De for useful discussions.
\end{acknowledgments}
|
\section{Introduction}
Solar flares are huge bursts of energy in the atmosphere of the Sun.
Satellite observations have emphasized their observations in the extreme ultraviolet (EUV) and Soft X-rays (SXR) domain, which were not possible before the space age. At these wavelengths, the emission increases drastically and can stand for a few hours (depending on the flare magnitude). Flares however were first observed from the ground in the visible domain \citep{1859MNRAS..20...13C,1966SSRv....5..388S,Neidig:1989ft}
and the advent of spectroscopic observations revealed the increase of solar flux in visible spectral lines, sometimes changing from absorption to a emission profile \citep{Hale:1931lr,1998SoPh..182..447S}. \\
Flare emission in the visible domain is thus well known to occur in visible chromospheric lines like the Balmer and Ca II K lines \citep{1990ApJ...363..318C,1992A&A...256..255F,1994SoPh..152..393H}. There is, however, another -and less understood- contribution to the visible emission that is due to the enhancement of the continuum. The term 'white-light' (WL) is used to refer to visible continuum enhancement. The study of WL flares have been rendered very difficult by their short duration and very low contrast which makes their observations from Earth rare and of poor quality. \cite{Neidig:1989ft} emphasized this aspect together with the suggestion that WL emission can be quite common and can constitute up to 90$\%$ of the total energy radiated by the flare. This has, however, never been confirmed and several questions remain about WL emission, like: What is the contribution of WL emission to the total flare energy ? Does it appear only in very large flares ? How and where is it produced ? \\
There are only a few observations of WLFs in space as the available measurements do not in general have the required high spatial and temporal resolution -as well as duty cycle-. Several WLFs have however been observed with Yohkoh \citep[e.g.][]{Matthews:2003fk} and TRACE \citep[e.g.][]{Hudson:2006aa}; these studies have confirmed the importance of WLFs and their association with the hard X-rays in the impulsive phase of the flare. Furthermore, many flares were observed with no associated white-light emission and with no means to determine if this was due to instrumental limitations or to the actual absence of WL emission. Recent studies \citep{2008ApJ...688L.119J,1674-4527-9-2-001} identified WL emission in relatively small flares, which support the view that WL emission could be a peculiar feature of flare and not only associated to the largest flares. \\
Knowing how much energy is radiated as a whole and at each wavelength is important both to understand the physical processes in the solar atmosphere and the impact of flares on Earth. Ideally, it would however be necessary to observe the Sun at all wavelengths with sufficient contrast (implying good spatial and temporal resolution) and duty cycle (to avoid missing the flare). One must often deal with partial observations only, especially at ultraviolet and visible light where contrast is weak and space instrumentation is rare. This explains why we still have a partial view of how the flare energy is spectrally distributed. \cite{2004GeoRL..3110802W,Woods:2006aa} provided the first direct observation of the flare signal in the total solar irradiance (TSI), allowing thus a precise estimate of the the total energy radiated by this extremely large (X17) flare of October 28, 2003. Using additional solar irradiance observations in the XUV (shorter than 27nm) and the FISM flare model up to 190nm, they concluded that about half of the energy is radiated below 190nm, most of it coming from wavelengths shorter than 14nm and the other half coming thus from near UV, visible and infrared light. These important results for very large flares suffer, however, from the absence of direct observations at wavelengths longer than 27nm and assume that the TSI time profile follows the flare time profile in XUV. \\
In this paper, we show that white-light continuum is commonly produced in basically all flares and that it represents most of the flare energy. In section \ref{sec_data}, we present the data and the analysis method. We follow \cite{2010NatPh...6..690K} and perform a superposed epoch analysis of solar \emph{irradiance} (i.e. integrated solar flux) in the visible domain during flares. The exceptional duty cycle and long-term monitoring of the Sun Photometer (SPM) instrument counterbalance the absence of spatial resolution and the presence of background temporal fluctuations due to p modes, which allows to reveal the flare signal at visible wavelengths and to identify it as WL continuum (section \ref{sec_res}). In section \ref{sec_discuss}, we estimate the spectral distribution of the radiated energy for average flares of different amplitude as well as for the single X17 flare of October 28, 2003, for which we present new observations. We conclude in section \ref{sec_conclu}.
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=0.8\textwidth,height=0.45\textheight]{Fig1.pdf}
\end{center}
\caption{{\bf Average flare light curves for TSI (black), and the three visible SPM channels (blue, green, and red)}. \textit{Upper left panel}: average over 43 flares from X17.2 to X1.3. \textit{Upper right panel}: average over 68 flares from X1.3 to M6.4. \textit{Lower left panel}: average over 140 flares from M6.4 to M2.8. \textit{Lower right panel}: average over 1850 flares from M2.8 to C4.\label{fig_res}. Horizontal lines show the $\pm 2\sigma$ limits computed from -90 to 150 minutes excluding 60min centered around the peak time.}
\end{figure*}
\section{Data and Analysis \label{sec_data}}
The data we used in this study are full Sun fluxes (i.e. solar irradiance) observed by the SOHO and GOES spacecrafts: The \textit{Total Solar Irradiance (TSI)} (solar flux integrated over all wavelengths), measured by the VIRGO/PMOv6 instrument onboard SOHO; three \textit{Visible Solar Irradiance} from the VIRGO/SPM instrument (still SOHO) consisting of passbands of 5nm respectively centered on 402nm (blue), 500nm (green), and 862nm (red).
These irradiance time series cover the period from 1996 to mid 2008 with a 1-minute time step with very few data gaps. They have been corrected for degradations and can be downloaded from the VIRGO ftp website\footnote{ftp.pmodwrc.ch$/$pub$/$data$/$irradiance$/$virgo$/$1$-$minute\_Data$/$}. Additionally we use the Extreme Ultraviolet (EUV) irradiance in the ranges 0.1-50nm and 26-34nm measured by SOHO/SEM \citep{1998SoPh..177..161J} and the Soft X-ray (SXR) irradiance measured by the GOES satellites.\\
The superposed epoch analysis is relatively simple and consists in superposing several time series in which a similar event (here a flare) occurs at the same time. It is useful when the signal of the event is faint with respect to the random (incoherent) background fluctuations. The resulting superposed (or average) curve exhibits thus smaller background fluctuation (typically attenuated by a factor $1/\sqrt{n}$, $n$ being the number of samples) but a stronger signal at the time of the event, which results from the superposition of all the coherent faint signals in each individual time series. This technique has been used to detect flares in total solar irradiance by \cite{2010NatPh...6..690K} and we extend here this work by looking at visible, EUV, and SXR irradiance time series. \\
The events we considered here are flares observed in Soft X-ray by the GOES satellites from 1996 to 2007, and occurring at heliocentric angle $\theta \le 60\deg$ because at higher $\theta$ the visible emission is likely to be strongly absorbed.
Before superposing the light curves for the TSI and visible irradiance, we subtracted a linear fit from the time series in order to eliminate trends. This in principle could hide meaningful trends (for example linked to the evolution of the active region pre- or post- flare) but it allows us to reach a better signal-to-noise ratio and the analysis without trend removal did not reveal clear results. In section \ref{sec_discuss} we analyse the spectral distribution of the flare energy; for this purpose, we convert the TSI and visible light curves from ppm to irradiance units (W.m$^{-2}$) by simply multiplying the average time series by their average value (1365 W.m$^{-2}$, for the TSI, 1.67W.m$^{-2}$, 1.83W.m$^{-2}$, 0.97W.m$^{-2}$ respectively for the blue, green, and red channel); small departures from theses values (for example by taking 1361W.m$^{-2}$ instead of 1365W.m$^{-2}$ for the TSI) do not affect the result because of the difference of many orders of magnitude between the quiet irradiance and the flare flux. The averaged light curves in physical units for the EUV and soft X-rays channels are obtained by subtracting the background of each time series before averaging.
\section{White-light Flare observations in Sun-as-a-star measurements \label{sec_res}}
Fig.\ref{fig_res} shows the average flare light curves for 4 sets of flares of decreasing magnitude in the visible and TSI channels. The GOES SXR peak time has been chosen as the key time for each flare. The first immediate result is the appearance in all channels of a peak around the key time that is caused by the flares. Importantly, this is true down to C-class flare, although it can be noted that the signal-to-noise ratio is smaller. This shows in particular that visible emission is ubiquitous in flares. We can note that other peaks exceed the 2$\sigma$ limits (about 5 points over 100, as expected for random noise); they are, however, less pronounced and appear at different times for the different flare sets. The curves also display oscillations with roughly a 10 to 20-minute period, whose amplitude, however, is below the 2$\sigma$ level (see also fig.\ref{fig_spec0_2100}, before peak time); it is not clear whether these oscillations can be related to actual solar processes (pre- and post- flare accoustic waves) or if they are artifacts of the instrumental noise. The fact that they are not in phase for different flare sets rather points towards an instrumental effect.\\
The curves exhibit several additional interesting features. \\
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.4\textwidth,height=0.22\textheight]{fig2b.pdf}
\end{center}
\caption{{\bf Influence of the key time on the analysis}. Dashed lines are average flare time series using the GOES 0.1-0.8 nm flux peak time, thick lines are average flare time series using the peak time of the derivative of the GOES 0.1-0.8 nm flux. Black is for TSI, red for 862nm, blue for 402nm, and green for 500nm. The average is made over the 150 largest flares in our sample, from X28 to M5. \label{fig_peaktime}}
\end{figure}
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.4\textwidth,height=0.22\textheight]{Fig2c.pdf}
\end{center}
\caption{{\bf Influence of the impulsive phase duration on the analysis}. Thick lines show average time series for flares with short impulsive phase (equivalent X-ray class M2), dashed lines for long impulsive phase (equivalent X-ray class M5). SPM green is not shown for clarity but behaves similarly to SPM blue. \label{fig_duration}}
\end{figure}
First, most of the visible emission belongs to the impulsive phase of the flare. This can be better seen on fig.\ref{fig_peaktime}:
when the key time is the GOES SXR peak time, the maximum of the curves precedes it, while when the key time is the peak of the derivative of the SXR flux, the maximum occurs at t=0. The amplitude of the increase, however, is larger when using the SXR flux peak; we attribute this to the fact that the Neupert effect does not hold rigorously in all flares. Fig. \ref{fig_duration} shows in a similar way the effect of the duration of the impulsive phase, computed as the duration of the rising phase of the SXR flare, i.e. $t_{SXR}^{peak} -t_{SXR}^{start}$. The flare signal in the visible channels and in the TSI nearly disappears after the SXR peak time for flares with short impulsive phase ($\le 6$min), while it persists for flares with longer impulsive phase ($> 6$min). This suggests that visible emission is present in both compact and gradual flare, but that it also persists for some time during the gradual phase of gradual flares.\\
The second point to note on fig.\ref{fig_res} is the longer decay of the red channel after the peak, which can be attributed to the chromospheric Ca II line emission that is included in this channel. This constitutes a clear signature of the gradual phase, that is not seen in the blue and green channel, and is much less obvious in the TSI measurements. \\
Thirdly, the flare emission in the blue tends to be larger than in the two other visible channels. We come back to this point below, but let's note now that it agrees with the general statement that stellar flares are "blue", i.e. display more emission towards the UV \citep[e.g.][]{2003ApJ...597..535H}. \\
Last, but not least, the relative increase in the visible channels is nearly of the same amplitude as for the TSI. This strongly suggests that the visible emission has a dominant contribution in the total energy radiated by flares. We also come back to this point later. \\
Fig.\ref{fig_res} shows that the SPM visible emission is observed on average in all flares down to C-class ones. Does this emission come from lines or continuum ? The SPM channels have been chosen to be in the continuum of the solar spectrum and their response functions computed for the quiet Sun \citep{2005ApJ...623.1215J} shows that the the blue and green channel corresponds to the deep photosphere (near $\tau _{500}=1$), while the red channel has a larger contribution of the upper photosphere (the Ca II line).
Since, however, a multitude of lines are present all over the optical solar spectrum, we have looked at the modeled spectrum from \cite{Chance:2010lr}, the observed spectrum by \cite{2004AdSpR..34..256T} and the observed identified line lists by \cite{Allende-Prieto:1998qy}. The blue and green channels include only a few weak lines from neutrals (mainly Fe I), which strongly suggests that the increase of emission, observed during the impulsive phase, is caused by a general increase of the continuum level. The next section shows that this continuum is furthermore consistent with a blackbody spectrum at T$\sim$9000K, which agrees with solar and stellar observations of WLFs. \\
Since we access only light curves averaged over several flares, we can wonder how often this WL continuum emission does actually occur. To test this, we have substituted in our sample some of the time series that include flares, by time series that do not include them (randomly selected at time where no flares occur). The flare signal for the less energetic flare set (bottom right panel in fig.\ref{fig_res}) becomes visible when only one out of five time series is replaced. This indicates roughly that at least 80\% of the flares have WL continuum emission and can thus be considered as WLFs. \\
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=0.7\textwidth]{Fig4.pdf}
\end{center}
\caption{{\bf Flare light curves averaged over 2100 flares} (from X17.2 to C4) in various spectral ranges.\label{fig_spec0_2100}}
\end{figure*}
\section{Spectral distribution of flare energy \label{sec_discuss}}
\begin{table*}[!t]
\caption{\label{tab_nrj1} {\bf Spectral distribution of flare energy.} From first row to the last one the average corresponds to the five sets of flares shown in Figs.\ref{fig_res} and \ref{fig_spec0_2100}. The two ratio in column 3 to 5 correspond respectively to integrations over all the flare duration and limited to the TSI flare signal period. The three last columns show mean blackbody and flare parameters that can explain the visible emission at the flare peak time.}
\centering
\begin{tabular}{llcccccc}
\hline\hline
Mean & Total Energy &Ratio&Ratio&Ratio & T$_{bb}$ & S$_{f}$ & Ratio\\
X-ray class & TSI (Ergs)&26-34nm/TSI&0-50nm/TSI&0.1-0.8nm/TSI & ($^{\circ}$K) & (arcsec$^{2}$) & Continuum/TSI\\
\hline
X3.2 & 5.9 10$^{31}$ & 0.9\% - 0.8\% & 12\% - 9\% & 1.2\% - 1\% & 9345 & 16.7 & 67\% \\
M9.1 & 1.6 10$^{31}$ & 1.7\% - 0.4 \% & 23\% - 5\% & 1.0\% - 0.4\% & 8993 & 13.2 & 85\% \\
M4.2 & 1.3 10$^{31}$ & 2.2\% - 0.5\% & 18\% - 6\% & 0.6\% - 0.3\% & 9244 & 7.3 & 74\% \\
C8.7 & 3.6 10$^{30}$ & 1.5\% - 0.5\% & 16\% - 5\% & 0.4\% - 0.2\% & 8655 & 2.4 & 72\%\\
\hline
M2.0 & 5.1 10$^{30}$ & 1.7\% - 0.6\% & 18\% - 6\% & 0.7\% - 0.4\% & 8941K & 2.8 & 69\%\\
\hline
\hline
\end{tabular}
\end{table*}
Fig.\ref{fig_spec0_2100} shows flare light curves in several spectral bands from Soft X-ray to visible and averaged over 2100 flares ranging from X-class to C-class. The EUV and SXR bands show a long gradual phase that can also be seen in the visible red channel that contains a chromospheric contribution. The TSI light curve should also exhibit the gradual phase that occurs at all chromospheric and coronal wavelengths but it does not. The most probable explanation is that the gradual phase is below the noise level, which emphasizes the predominance of the impulsive phase. Because of this fact, we cannot exclude that the blue and green channels also have a gradual phase although smaller than the red channel. We use fig.\ref{fig_spec0_2100} to estimate the spectral distribution of the flare energy and we repeat this for the four sets of flares shown on fig.\ref{fig_res}. Because the gradual phase is hardly observed in the TSI average light curves, we computed the energy radiated in each passband both over the total duration of the flares, i.e. including the long gradual phase when it is present, and over the restricted time for which the flare signal is seen in the TSI. For the visible irradiance and the TSI (EUV and SXR channels respectively) we limit the start of the flare integration at 10min (20min resp.) before the key time, in order to avoid signals potentially due to background fluctuations in the visible and TSI light curves.\\
Emission excess observed at Earth can be converted in energy release on the Sun with the factor $d_{AU}^2 \times f$ where $f$ takes into account the angular distribution of the emission. We use the following factors: $f=2\pi$ for optically thin emission (0-50nm and 0.1-0.8nm), and $f=1.4\pi$ for the TSI and 26-34nm channel, as suggested by \cite{Woods:2006aa}; assuming the same angular distribution for all channels decreases the optically thin contribution by a factor 1.43. The resulted flare energy distribution is shown in tab.\ref{tab_nrj1} for the two integrations (next section explains the three last columns); it reveals how the short wavelength passbands are a minor contributor to the total energy, even when the gradual phase is not taken into account only for the TSI. The GOES SXR energy is less than 1\% of the total energy, while all wavelengths below 50nm represent between 10\% and 20\%. The visible and near UV part of the flare spectrum must thus constitute the bulk of the flare energy.
\subsection{WL energy}
It is difficult to reproduce the exact shape of the continuum spectrum from the observations of only three passbands. However we can make simple considerations and try to estimate the total energy that goes in the continuum.
\begin{figure}[t]
\centering
\includegraphics[width=0.35\textwidth,height=0.19\textheight]{Fig3.pdf}
\caption{{\bf Ratio between the flare maximum emission in the blue and green channel versus the mean SXR flux peak.} Dotted line is the average ratio of 1.34 and the dashed line is the best fit obtained with the following expression $I^{f}_{Blue}/I^{f}_{Green}=0.04\times\log(E_{SXR})+1.51$. The mean ratio can be reproduced by assuming the flare emission follow a blackbody curve at $\sim9100$K. }
\label{Fig_ratioBlueGreen}
\end{figure}
WLFs are usually divided in two types according to whether the emission exhibits or not a jump at the Balmer and Paschen edges; in our case, the blue and green passbands are just above the Balmer edge and the red channel is just above the Paschen edge; the Hydrogen free-bound emission can thus difficultly explain the observed emission (which does not mean that this emission is not present). Fig.\ref{Fig_ratioBlueGreen} shows the ratio of the flare emission in the blue and green channel $I_{blue}/I_{green}$ at peak time; this ratio is about 1.34 and changes slowly for different flare amplitudes. We can easily convert the ratio values in blackbody temperature; a linear extrapolation of the observed increasing ratio with flare amplitude leads to temperatures ranging from T$\sim$ 8430K for C1 flares to T$\sim$ 9560K for X10 flares. The slow change in temperature with the SXR flare magnitude indicates that the spectrum shape is similar for different flare magnitudes, the amplitude of the WL continuum depending mainly on the flaring area. Table \ref{tab_nrj1} gives the exact temperature found for each set of flares.
These blackbody temperatures are in good agreement with what has been found previously. \cite{2003ApJ...597..535H} explained the continuum emission of a flare observed on the M dwarf AD Leo with a blackbody temperature near 10000K; \cite{2007ApJ...656.1187F} found a very upper limit of $2.5\times 10^{4}$K to reproduce the spectral ratio between UV and visible emission in several solar WL flares, while recently \cite{Kowalski:2010fj} have provided evidence that the flare emission observed on a dMe4.5e star is composed of a Balmer component at wavelengths shorter than $\sim 380$nm and of a $T \sim 10^{4}$K blackbody component above 400nm. WL continuum emission is thought to occur in the minimum temperature region or below, where heating is provided by radiative backwarming from the chromosphere as has been found in simulations \citep{2005ApJ...630..573A,2006ApJ...644..484A,Cheng:2010qy}. As a consequence of the very disturbed and unknown state of the flaring atmosphere, the exact processes at play remain unidentified and it is possible that an unknown mechanism produces a flare spectrum that looks similar to a blackbody spectrum near 9000K.
As supported by our results and the other findings cited above, and in order to compute the flare continuum energy, we assume in the following that the WL continuum follows a blackbody spectrum, keeping thus in mind that its origin is unclear, but that it is supported by various observations. Hydrogen free-bound continua should also be present but are unlike to contribute significantly in the SPM passbands.\\
Assuming that the flare signal in the blue and green channels come from blackbody radiation, we can make simple estimates; we first concentrate on the "average" M2 flare shown on fig.\ref{fig_spec0_2100}. The ratio $I_{blue}/I_{green}$ leads to a blackbody temperature of T$\sim$ 8941K (see table \ref{tab_nrj1}). Multiplying the corresponding blackbody spectrum by the SPM blue channel response and matching the observed value at peak time,
we obtain a flaring area on the solar surface of 2.8 $arcsec^{2}$. Integrating the blackbody spectrum with this flaring area leads to a total continuum emission at peak time of 6.2 10$^{-3}$W.m$^{-2}$, i.e. about $\sim$70\% of the total radiated energy at peak time (9 10$^{-3}$W.m$^{-2}$ for the TSI). We stress here that the estimation of the total energy contained in the WL continuum is independent of the one deduced from the TSI observations.
Doing the same exercise for the total energy radiated over time where the TSI flare profile is available, we find a mean flaring area of 1.4 $arcsec^{2}$ corresponding to a total continuum energy of 3.4 10$^{30}$erg to be compared with the observed total 5.2 10$^{30}$erg, i.e. 60\%. 6\% comes from wavelengths below 50nm; the remaining could come from EUV wavelengths above 50nm and from the visible and near UV emission of chromospheric origin as well as from H free-bound continuum. We did this exercise for all 5 sets of flares and the deduced parameters of the blackbody radiation are presented in table \ref{tab_nrj1}. The contribution of the WL continuum is similar in all cases, i.e. around 70\%. The estimated flaring areas are consistent with WLF observations \citep[e.g.][]{Hudson:2006aa} and increase with the flare magnitude. Finally let's also note that other emission mechanisms with similar spectral distribution should lead to similar estimates of the flare energy contained in the white-light and near UV continuum.\\
\subsection{The case of the X17 flare on 28 October 2003\label{sec_28oct}}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=0.5\textwidth]{Fig5.pdf}
\end{center}
\caption{{\bf Flare light curves for the October 28, 2003 flare} Single flare light curves for the X17 flare of October 28, 2003. The grey curve in the 3$^{rd}$ panel is the flux coming from a region of $\sim$900$arcsec^{2}$ that includes the flare site and observed by the VIRGO/LOI instrument \citep{1997SoPh..170...27A} in the same passband than SPM/Green. SEM and GOES 0.01-0.5nm are not shown as they are polluted by saturation and particle effects. \label{fig_28oct}}
\end{figure}
The X17 flare thAT occurred on October 28, 2003 is a very large flare and the only one that has been unambiguously detected in TSI \citep{2004GeoRL..3110802W}. Fig.\ref{fig_28oct} shows irradiance light curves for this flare and confirms this finding by showing the flare signature in the SOHO/VIRGO instrument, while Woods et al. used the TIM instrument onboard SORCE. The increase in VIRGO is 264ppm while it is of 268ppm in the TIM measurements, i.e. in very good agreement. It also shows for the first time A clear white-light signature in Sun-as-a-star observations of a single flare using the VIRGO/SPM channels (relative increase of 267ppm, 191ppm, and 176ppm respectively for the blue, green, and red channel). Note that the visible light and TSI peak about 5 minutes before SXR, confirming the importance of the impulsive phase.\\
Following the analysis of the previous section, we find that the ratio $I_{blue}/I_{green}$ is 1.28 which corresponds to a blackbody temperature of 8545K, relatively lower than in the average cases of the previous sections, although still consistent. The signal-to-noise ratio is lower in this single event than in the average cases but this somehow gives an idea of the uncertainty on the determination of the blackbody temperature and we should speak of temperature {\it roughly around} 9000K, that does not change very much with the amplitude of the flare. Matching the observed blue signal gives a flaring area of $\sim$130 $arcsec^{2}$; this is much larger than areas shown in tab.\ref{tab_nrj1} but it should not be so surprising for such a flare; it also agrees with white-light observations from the Global Oscillation Network Group (GONG) -see fig.4 and fig.7 of \cite{2009SoPh..258...31M}). Integrating then a blackbody spectrum at 8545K over this flaring area, we find that the blackbody spectrum accounts for 64\% of the total energy, in good agreement with what we found for the average cases in table \ref{tab_nrj1}.
\section{Conclusions\label{sec_conclu}}
In this study, we identify and analyze for the first time the visible light produced by flares in Sun-as-a-star observations. We use a superposed epoch analysis to show that visible emission is present on average for all flares from X-ray class X to C, that it occurs mainly during the impulsive phase and must be considered as continuum emission, i.e. white-light flares. Analyzing the intensity ratio of the SPM blue and green channels, we found this emission to be consistent with a blackbody spectrum near 9000K. We next match the increase observed in the blue channel and deduce flaring areas that are consistent with previous observations. Using these results, we compute the total energy contained in the continuum and find it to represent about 70\% of the total energy radiated by flares. This study shows that that the white light continuum is ubiquitous in flares and that it represents about 2 thirds of the energy radiated by the flares. Furthermore, we reveal the existence of white light signature in solar flux for the single X17 flare that occurred on October 28, 2003 and find similar energy distribution. \\
These results show the very large predominance of the lower atmosphere with respect to the corona in freeing the flare energy that is initially stored in the solar magnetic field, and put constraints on models. Additionally, since each flare releases most of its energy in white light, and since it exists a continuum of flares, we can wonder if the visible emission released by flare can contribute to the variations of the TSI. We plan to address this issue in a forthcoming study.
\section{acknowledgements}
This work has received funding from the European Community's Seventh Framework Programme (FP7/2007-2013) under the grant agreement n° 218816 (SOTERIA project, www.soteria-space.eu). The author thanks T. Appourchaux for the VIRGO/LOI data, C. Wehrli for providing the SPM response functions, T. Dudok de Wit for useful discussions, and one anonymous referee for useful comments and suggestions.
\bibliographystyle{plainnat}
|
\section{Introduction}
In this note we consider the Cauchy proble
\begin{eqnarray}
\partial _{t}u(t,x) &=&Lu(t,x)+f(t,x),(t,x)\in H=[0,T]\times \mathbf{R}^{d},
\label{intr1} \\
u(0,x) &=&0 \notag
\end{eqnarray
in H\"{o}lder spaces for a class of integrodifferential operators $L=A+B$ of
the order $\alpha \in (0,2)$ whose principal part $A$ is of the for
\begin{eqnarray}
Au(t,x) &=&A_{t}u(t,x) \label{0} \\
&=&\int \left[ u(x+y)-u(x)-\chi _{\alpha }(y)(\nabla u(x),y)\right] m(t,x,y
\frac{dy}{|y|^{d+\alpha }} \notag
\end{eqnarray
with $\chi _{\alpha }(y)=1_{\alpha >1}+1_{\alpha =1}1_{\left\{ |y|\leq
1\right\} }$. We notice that the operator $A$ is the generator of an $\alpha
$-stable process. If $m=1,$then $A=c\left( -\Delta \right) ^{\alpha /2}$
(fractional Laplacian) is the generator of a spherically symmetric $\alpha
-stable process. The part $B$ is a perturbing, subordinated operator.
In \cite{MiP922}, the problem was considered assuming that $m$ is Holder
continuous in $x$, homogeneous of order zero and smooth in $y$ and for some
\eta >0$ {\
\begin{equation}
\int_{S^{d-1}}|(w,\xi )|^{\alpha }m(t,x,w)\mu _{d-1}(dw)\geq \eta ,\quad
(t,x)\in H,|\xi |=1, \label{1}
\end{equation
where }$\mu _{d-1}$ is the Lebesgue measure on the unit sphere $S^{d-1}$ in
\mathbf{R}^{d}$. In \cite{AbK09}, the existence and uniqueness of a solution
to (\ref{intr1}) in H\"{o}lder spaces was proved analytically for $m$ H\"{o
lder continuous in $x$, smooth in $y$ and such that
\begin{equation}
C\geq m\geq \delta >0 \label{2}
\end{equation
without assumption of homogeneity in $y$. \ The elliptic problem $Lu=f$ in
\mathbf{R}^{d}$ was considered in \cite{bas}, \cite{cafsilv}, \cite{KimDong1
$.$ In \cite{cafsilv}, the interior H\"{o}lder estimates (in a non-linear
case as well) were studied assuming (\ref{2}) and $m(x,y)=m(y)=m(-y)$. In
\cite{bas}, the apriori estimates were derived in Holder classes assuming
\ref{2}) and Holder continuity of $m$ in $x$, except the case $\alpha =1$.
Similar results, including the case $\alpha =1$ were proved in \cit
{KimDong1}. The equation (\ref{intr1}) with $\alpha =1$ can be regarded as a
linearization of the quasigeostrophic equation (see \cite{cav1}).
In this note, we consider he problem (\ref{intr1}), assuming that $m$ is
measurable, Holder continuous in $x$ and
\begin{equation}
C\geq m\geq m_{0}, \label{3}
\end{equation}
where the function $m_{0}=m_{0}(t,x,y)$ is smooth and homogeneous in $y$ and
satisfies (\ref{1}). So, the density $m$ can degenerate on a substantial set.
A certain aspect of the problem is that the symbol of the operator $A,
\begin{equation*}
\psi (t,x,\xi )=\int \left[ e^{i(\xi ,y)}-1-\chi _{\alpha }(y)i(\xi ,y
\right] m(t,x,y)\frac{dy}{|y|^{d+\alpha }}
\end{equation*
is not smooth in $\xi $ and the standard Fourier multiplier results (for
example, used in \cite{MiP922}) do not apply in this case. Instead we use
direct analytic and probabilistic arguments. We start with equation (\re
{intr1}) assuming that $B=0$, the input function $f$ is smooth and the
function $m=m(t,Y)$ is smooth and homogeneous in $y,$ satisfies (\ref{1})
and does not depend on $x$. This case of equation (\ref{intr1}) was
considered in \cite{MiP922} and \cite{MiP09} and the estimates of its
solution in H\"{o}lder spaces were derived. Then we use the Ito-Wentzell
formula to pass to $m(t,y)$ which is only measurable and satisfies (\ref{3})
and obtain a solution of (\ref{intr1}) with all the estimates retained. The
case of variable coefficients is considered by using partition of unity and
deriving apriori Schauder estimates in H\"{o}lder-Zygmund spaces. Finally,
we apply the continuation by parameter method to extend solvability of an
equation with constant coefficients to that of (\ref{intr1}).
As an application, we consider the martingale problem associated to $L$.
Since the coefficients are H\"{o}lder the existence of a martingale solution
is trivial. Applying the Ito formula to the solution of (\ref{intr1}), we
prove the weak uniqueness of the solution to the martingale problem,
generalizing so the uniqueness results in \cite{AbK09} and \cite{MiP923}$.$
The note is organized as follows. In Section 2, the main theorem is stated.
In Section 3, the essential technical results are presented. The case of the
equation with constant coefficients not depending on the spacial variable is
considered in Section 4. The main theorem is proved in Section 5. In Section
6 the uniqueness of the associated martingale problem is considered.
\section{Notation and main results}
Denote $H=[0,T]\times \mathbf{R}^{d}$, $\mathbf{N}=\{0,1,2,\ldots \}$,
\mathbf{R}_{0}^{d}=\mathbf{R}^{d}\backslash \{0\}$. If $x,y\in \mathbf{R
^{d} $, we write
\begin{equation*}
(x,y)=\sum_{i=1}^{d}x_{i}y_{i},|x|=(x,x)^{1/2}.
\end{equation*}
For a function $u=u(t,x)$ on $H$, we denote its partial derivatives by
\partial _{t}u=\partial u/\partial t,\partial _{i}u=\partial u/\partial
x_{i},\partial _{ij}^{2}u=\partial ^{2}u/\partial x_{i}\partial x_{j}$ and
D^{\gamma }u=\partial ^{|\gamma |}u/\partial x_{i}^{\gamma _{1}}\ldots
\partial x_{d}^{\gamma _{d}},$ where multiindex $\gamma =(\gamma _{1},\ldots
,\gamma _{d})\in \mathbf{N}^{d},\nabla u=(\partial _{1}u,\ldots ,\partial
_{d}u)$ denotes the gradient of $u$ with respect to $x$.
For a function $u$ on $H$ and $\beta \in (0,1]$, we writ
\begin{eqnarray*}
|u|_{0} &=&\sup_{t,x}|u(t,x)|, \\
\lbrack u]_{\beta } &=&\sup_{t,x,h\neq 0}\frac{|u(t,x+h)-u(t,x)|}{h^{\beta }
\text{ if }\beta \in (0,1), \\
\lbrack u]_{\beta } &=&\sup_{t,x,h\neq 0}\frac{|u(t,x+h)+u(t,x-h)-2u(t,x)|}
|h|}\text{ if }\beta =1.
\end{eqnarray*
For $\beta =[\beta ]^{-}+\left\{ \beta \right\} ^{+}>0$, where $[\beta
]^{-}\in \mathbf{N}$ and $\left\{ \beta \right\} ^{+}\in (0,1]$, we denote
C^{\beta }(H)$ denote the space of measurable functions $u$ on $H$ such that
the norm
\begin{equation*}
|u|_{\beta }=\sum_{|\gamma |\leq \lbrack \beta ]^{-}}|D^{\gamma
}u|_{0}+\sup_{|\gamma |=[\beta ]^{-}}[D^{\gamma }u]_{\left\{ \beta \right\}
^{+}}.
\end{equation*
Accordingly, $C^{\beta }(\mathbf{R}^{d})$ denotes the corresponding space of
functions on $\mathbf{R}^{d}$.~The classes $C^{\beta }$ coincide with H\"{o
lder spaces if $\beta \notin \mathbf{N}$ (see 1.2.2 of \cite{Tri92}).
For $\alpha \in (0,2)$ and $u\in C^{\alpha +\beta }(H),$ we define the
fractional Laplacia
\begin{equation}
\partial ^{\alpha }u(t,x)=\int [u(t,x+y)-u(t,x)-\left( \nabla
u(t,x),y\right) \chi _{\alpha }(y)]\frac{dy}{|y|^{d+\alpha }}, \label{fo4}
\end{equation
where $\chi ^{(\alpha )}(y)=\mathbf{1}_{\{|y|\leq 1\}}\mathbf{1}_{\{\alpha
=1\}}+\mathbf{1}_{\{\alpha \in (1,2)\}}$.
We denote $C_{b}^{\infty }(H)$ the space of bounded infinitely
differentiable in $x$ functions whose derivatives are bounded.
$C=C(\cdot ,\ldots ,\cdot )$ denotes constants depending only on quantities
appearing in parentheses. In a given context the same letter is (generally)
used to denote different constants depending on the same set of arguments.
Let $(U,\mathcal{U)}$ be a measurable space with a non-negative measure $\pi
(d\upsilon )$ on it$.$
Let $\alpha \in (0,2)$ and $\beta \in (0,1]$ be fixed. Let $m:H\times
\mathbf{R}_{0}^{d}\rightarrow \lbrack 0,\infty ),b:H\rightarrow \mathbf{R
^{d},c:H\times U\rightarrow \mathbf{R}^{d}$ and $\rho :H\times U\rightarrow
\mathbf{R}$ be measurable functions. We also introduce an auxiliary function
$m_{0}:[0,T]\times \mathbf{R}_{0}^{d}\rightarrow \lbrack 0,\infty )$ and fix
positive constants $K$ and $\mu $. Throughout the paper we assume that the
function $m_{0}$ satisfies the following conditions. \medskip
\noindent \textbf{Assumption} $\mathbf{A}_{0}.$ (i) The function
m_{0}=m_{0}(t,y)\geq 0$ is measurable, homogeneous in $y$ with index zero,
differentiable in $y$ up to the order $d_{0}=[\frac{d}{2}]+1$ an
\begin{equation*}
|D_{y}^{\gamma }m_{0}^{(\alpha )}(t,y)|\leq K
\end{equation*
for all $t\in \lbrack 0,T]$, $y\in \mathbf{R}_{0}^{d}$ and multiindices
\gamma \in \mathbf{N}_{0}^{d}$ such that $|\gamma |\leq d_{0}$;
(ii) If $\alpha =1$, then for all $t\in \lbrack 0,T]$
\begin{equation*}
\int_{S^{d-1}}wm_{0}(t,w)\mu _{d-1}(dw)=0,
\end{equation*
where $S^{d-1}$ is the unit sphere in $\mathbf{R}^{d}$ and $\mu _{d-1}$ is
the Lebesgue measure on it;
(iii) For all $t\in \lbrack 0,T]$
\begin{equation*}
\inf_{|\xi |=1}\int_{S^{d-1}}|(w,\xi )|^{\alpha }m_{0}(t,w)\mu
_{d-1}(dw)\geq \eta >0.
\end{equation*}
\begin{remark}
\label{r10}\emph{The nondegenerateness assumption $A_{0}$ (iii) holds with
certain $\delta >0$ if, e.g
\begin{equation*}
\inf_{t\in \lbrack 0,T],w\in \Gamma }m_{0}^{(\alpha )}(t,w)>0
\end{equation*
for a measurable subset $\Gamma \subset S^{d-1}$ of positive Lebesgue
measure.}
\end{remark}
Further we will use the following assumptions.
\textbf{Assumption }$\mathbf{A.}$ (i) For all $(t,x)\in H,y\in \mathbf{R
_{0}^{d},
\begin{equation*}
|m(\cdot ,y)|_{\beta }\leq K
\end{equation*
an
\begin{equation*}
m(t,x,y)\geq m_{0}(t,y),
\end{equation*
where the function $m_{0}$ satisfies Assumption $\mathbf{A}_{0}$;
(ii) If $\alpha =1$, then for all $(t,x)\in H$ and $r\in (0,1),
\begin{equation*}
\int_{r<|y|\leq 1}ym(t,x,y)\frac{dy}{|y|^{d+\alpha }}=0.
\end{equation*}
We will assume that there is a decreasing sequence of subsets $U_{n}\in
\mathcal{U}$ such that $U=\cup _{n}U_{n}^{c}$ and the following assumptions
hold.
\textbf{Assumption B1. }(i) for all $(t,x)\in H,
\begin{equation*}
\int_{U_{1}}|c(t,x,\upsilon )|^{\alpha }\pi (d\upsilon
)+\int_{U_{1}^{c}}|c(t,x,\upsilon )|^{\alpha \wedge 1}\wedge 1\pi (d\upsilon
)\leq K
\end{equation*
(ii) for $\alpha \in (0.2)$\textbf{\
\begin{equation*}
\lim_{\varepsilon \rightarrow 0}\sup_{t,x}\int 1_{|c(t,x,\upsilon )|\leq
\varepsilon }|c(t,x,\upsilon )|^{\alpha }\pi (d\upsilon )=0
\end{equation*}
\textbf{Assumption B2}. (i)
\begin{equation*}
|b|_{\beta }+|l|_{\beta }\leq K;
\end{equation*}
(ii) If $\alpha \in \lbrack 1,2)$, then there is a constant $C$ such that
for all $(t,x)\in H,h\in \mathbf{R}^{d},
\begin{equation*}
\int_{U_{1}}|c(t,x,\upsilon )-c(t,x+h,\upsilon )|^{\alpha }\pi (d\upsilon
)\leq C|h|^{\alpha \beta },
\end{equation*
an
\begin{equation*}
\int_{U_{1}^{c}}[|c(t,x,\upsilon )-c(t,x+h,\upsilon )|\wedge 1]\pi
(d\upsilon )\leq C|h|^{\beta };
\end{equation*}
(iii) If $\alpha <1,$ then there is $\beta ^{\prime }\,\ $such that $\alpha
+\beta >\alpha +\beta ^{\prime }\geq \beta $ and there is a constant $C$
such that for all $(t,x),\in H,h\in \mathbf{R}^{d},$
\begin{eqnarray*}
\int_{U_{1}}|c(t,x,\upsilon )-c(t,x+h,\upsilon )|^{(\alpha +\beta ^{\prime
})\wedge 1}\pi (d\upsilon ) &\leq &C|h|^{\beta }, \\
\int_{U_{1}^{c}}|c(t,x,\upsilon )-c(t,x+h,\upsilon )|^{(\alpha +\beta
^{\prime })\wedge 1}\wedge 1\pi (d\upsilon ) &\leq &C|h|^{\beta };
\end{eqnarray*}
(iv) For all $\upsilon \in U,
\begin{equation*}
|\rho \left( \cdot ,\upsilon \right) |_{\beta }\leq K.
\end{equation*}
For $(t,z)\in H,u\in C^{\alpha +\beta }(\mathbf{R}^{d})$ we introduce the
operator
\begin{equation*}
A_{t,z}u(x)=A_{t,z}^{m}u(x)=\int_{\mathbf{R}^{d}}[u(x+y)-u(x)-(\nabla
u(x),y)\chi _{\alpha }(y)]m(t,z,y)\frac{dy}{|y|^{d+\alpha }},
\end{equation*}
\begin{eqnarray*}
B_{t,z,\bar{z}}u(x) &=&(b(t,z)\nabla u(x))1_{1\leq \alpha
<2}+\int_{U}[u(x+c(t,z,\upsilon ))-u(x) \\
&&-(\nabla u(x),c(t,z,\upsilon ))1_{U_{1}}(\upsilon )1_{1<\alpha <2}]\rho (t
\bar{z},\upsilon )\pi (d\upsilon ) \\
&&+l(t,z)u(x),
\end{eqnarray*
and
\begin{equation}
L_{t,z}u(x)=A_{t,z}u(x)+B_{t,z,z}u(x). \label{for0}
\end{equation
For brevity of notation, we writ
\begin{eqnarray}
Au(t,x) &=&A_{t}u(x)=A_{t,x}u(x),Bu(t,x)=B_{t}u(x)=B_{t,x,x}u(x),
\label{for2} \\
Lu(t,x) &=&L_{t}u(x)=L_{t,x}u(x),L=A+B. \notag
\end{eqnarray
According to Assumptions \textbf{A, B1, B2}, the operator $A$ represents the
principal part of $L$ and the operator $B$ is a lower order operator.
\begin{remark}
A simple example of $U,U_{n},\pi (d\upsilon )$ is $U=\mathbf{R
_{0}^{d},U_{n}=\{\upsilon :|\upsilon |<1/n\},c(t,x,\upsilon )=\upsilon ,\pi
(d\upsilon )=d\upsilon /|\upsilon |^{d+\alpha ^{\prime }},\alpha ^{\prime
}<\alpha ,$ an
\begin{equation*}
Bu(t,x)=\int_{\mathbf{R}_{0}^{d}}[u(x=y)-u(x)-(\nabla u(x),y)\mathbf{1
_{\left\{ |y|\leq 1\right\} }\mathbf{1}_{1\leq \alpha ^{\prime }<2}]\rho
(t,x,y)\frac{dy}{|y|^{d+\alpha ^{\prime }}}.
\end{equation*}
\end{remark}
For a fixed $\alpha \in (0,2),\beta \in (0,1)$ we consider the following
Cauchy proble
\begin{eqnarray}
\partial _{t}u(t,x) &=&(L-\lambda )u(t,x)+f(t,x),(t,x)\in H, \label{eq1} \\
u(0,x) &=&0,x\in \mathbf{R}^{d}, \notag
\end{eqnarray
in Holder classes $C^{\alpha +\beta }(H)$, where $\lambda \geq 0$ and $f\in
C^{\beta }(H)$.
\begin{definition}
Let $f$ be a bounded measurable function on $H.$ We say that $u\in C^{\alpha
+\beta }(H)$ is a solution of (\ref{eq1}), if for each $(t,x)\in H,
\begin{equation}
u(t,x)=\int_{0}^{t}[Lu(s,x)-\lambda u(s,x)+f(s,x)]ds. \label{defs}
\end{equation}
\end{definition}
If Assumptions \textbf{A} and \textbf{B1} are satisfied, then $Lu$ is
bounded (see Proposition \ref{prop2} and Lemma \ref{l7} below). So, (\re
{defs}) is well defined.
The main result of the paper is the following theorem.
\begin{theorem}
\label{main}Let $\alpha \in (0,2),\beta \in (0,1]$ and Assumptions \textbf
A, B1, }and \textbf{B2 }be satisfied.
Then for any $f\in C^{\beta }(H)$ there exists a unique solution $u\in
C^{\alpha +\beta }(H)$ to (\ref{eq1}). Moreover, there is a constant
C=C(\alpha ,\beta ,d,K,\mu )$ such tha
\begin{equation*}
|u|_{\alpha +\beta }\leqslant C|f|_{\beta },
\end{equation*
and for all $s\leq t\leq T,$
\begin{equation*}
|u(t,\cdot )-u(s,\cdot )|_{\frac{\alpha }{2}+\beta }\leq
C(t-s)^{1/2}|f|_{\beta }.
\end{equation*}
\end{theorem}
\section{Auxiliary results}
We will use the following equality for the H\"{o}lder norm estimates.
\begin{lemma}
\label{r1}$($Lemma 2.1 in \textup{\cite{Kom84}}$)$ For $\delta \in (0,1)$
and $u\in C_{0}^{\infty }(\mathbf{R}^{d})$,
\begin{equation}
u\left( x+y\right) -u(x)=C\int k^{(\delta )}(y,z)\partial ^{\delta }u(x-z)dz,
\label{22}
\end{equation
where the constant $C=C(\delta ,d)$ and
\begin{equation*}
k^{(\delta )}(y,z)=|z+y|^{-d+\delta }-|z|^{-d+\delta }.
\end{equation*
Moreover, there is a constant $C=C(\delta ,d)$ such that for each $y\in
\mathbf{R}^{d}$
\begin{equation*}
\int |k^{(\delta )}(y,z)|dz\leq C|y|^{\delta }.
\end{equation*}
\end{lemma}
The following Lemmas \ref{le4}, \ref{l8} are deterministic counterparts of
the statements proved in \cite{MiP09}.
\begin{lemma}
\label{le4}(see Corollary 15 in \cite{MiP09})Let $\beta \in (0,1],f\in
C^{\beta }(H)$. Then there is a sequence $f_{n}\in C_{b}^{\infty }(H)$ such
tha
\begin{equation*}
|f_{n}|_{\beta }\leq 2|f|_{\beta },|f|_{\beta }\leq \lim
\inf_{n}|f_{n}|_{\beta },
\end{equation*
and for any $0<\beta ^{\prime }<\beta
\begin{equation*}
|f_{n}-f|_{\beta ^{\prime }}\rightarrow 0\text{ as }n\rightarrow \infty .
\end{equation*}
\end{lemma}
\begin{proof}
The proof in \cite{MiP09} (Corollaries 13 and 15) for $\beta \in (0,1)$
covers without any changes the case $\beta =1$ as well.
\end{proof}
\begin{lemma}
\label{rem1}$($see Theorem 6.3.2 in\cite{bergl}$)$ For $\alpha \in (0,2)$,
\beta >0$, the norms $|u|_{\alpha ,\beta }=|u|_{0}+|\partial ^{\alpha
}u|_{\beta }$, and $|u|_{\alpha +\beta }$ are equivalent in $C^{\alpha
+\beta }$.
\end{lemma}
Let us introduce an operator $A^{0}$ defined as operator $A$ with $m$
replaced by $m_{0}$. In terms of Fourier transforms, for $u\in C_{b}^{\infty
}(H),
\begin{equation*}
\mathcal{F}\left( A^{0}u\right) (t,\xi )=\psi _{0}(t,\xi )\mathcal{F}u(t,\xi
),
\end{equation*
wher
\begin{eqnarray*}
\psi _{0}(t,\xi ) &=&-C\int_{S^{d-1}}|(w,\xi )|^{\alpha }[1-i(\tan \frac
\alpha \pi }{2}\text{sgn}(w,\xi )1_{\alpha \neq 1} \\
&&-\frac{2}{\pi }\text{sgn}(w,\xi )\ln |(w,\xi )|1_{\alpha =1}]m_{0}(t,w)\mu
_{d-1}(dw)
\end{eqnarray*
and the constant $C=C(\alpha )>0.$ Denot
\begin{eqnarray*}
K_{s,t}(\xi ) &=&\exp \left\{ \int_{s}^{t}\psi _{0}(r,\xi )dr\right\} ,s\leq
t, \\
G_{s,t}(x) &=&\mathcal{F}^{-1}K_{s,t},G_{s,t}^{\lambda }(x)=e^{-\lambda
(t-s)}G_{s,t}(x).
\end{eqnarray*
According to Assumption $\mathbf{A}_{0}$, $\int |K_{s,t}(\xi )|d\xi <\infty
,s<t$. Therefore $G_{s,t}$ is the density function of a random variable
whose characteristic function is $K_{s,t}$. Hence
\begin{equation}
G_{s,t}\geq 0,\int G_{s,t}(y)dy=1,s<t. \label{4}
\end{equation
Let $f\in C_{b}^{\infty }(H)$ an
\begin{equation}
R_{\lambda }f(t,x)=\int_{0}^{t}\left[ G_{s,t}^{\lambda }\ast f(s,\cdot
\right] (x)ds, \label{5}
\end{equation
where $\ast $ denotes the convolution with respect to $x$.
\begin{lemma}
\label{l8}$($see Lemmas 7 and 17 in \cite{MiP09})Let $\alpha \in (0,2)$,
\beta \in (0,1]$, $f\in C_{b}^{\infty }(H)$ and Assumption $\mathbf{A}_{0}$
be satisfied. Then the Cauchy proble
\begin{eqnarray}
\partial _{t}u(t,x) &=&A^{0}u(t,x)-\lambda u(t,x)+f(t,x),(t,x)\in H,
\label{for1} \\
u(0,x) &=&0,x\in \mathbf{R}^{d}, \notag
\end{eqnarray
has a unique solution $u=R_{\lambda }f\in C_{b}^{\infty }(H)$. Moreover,
there are constants $C_{1}=C(\alpha ,\ \beta ,T,\ d$,$\mu ,K)$ and
C_{2}=C_{2}(\alpha ,d)$ such that
\begin{equation}
|u|_{\alpha +\beta }\leqslant C_{1}|f|_{\beta }, \label{cc1}
\end{equation
\begin{equation}
|u|_{\beta }\leq C_{2}(\lambda ^{-1}\wedge T)|f|_{\beta } \label{cc2}
\end{equation
and for all $0\leq s\leq t\leq T$
\begin{equation}
|u(t,\cdot )-u(s,\cdot )|_{\alpha /2+\beta }\leq C(t-s)^{1/2}|f|_{\beta }.
\label{cc3}
\end{equation}
\end{lemma}
\begin{proof}
The statement is proved in \cite{MiP09} for $\beta \in (0,1)$ (Lemmas 7 and
17). According to Lemma 7 in \cite{MiP09}, for each $f\in C_{b}^{\infty }(H)$
there is a unique solution $u=R_{\lambda }f\in C_{b}^{\infty }(H)$.
Obviously $\partial ^{1/2}u$ solve the equation (\ref{for2}) in
C_{b}^{\infty }(H)$ with $\partial ^{1/2}f$ as input function. Applying the
statement with $\beta =1/2$ we hav
\begin{eqnarray*}
|\partial ^{1/2}u|_{\alpha +1/2} &\leq &C|\partial ^{1/2}f|_{1/2}, \\
|\partial ^{1/2}u|_{1/2} &\leq &C_{2}(\lambda ^{-1}\wedge T)|\partial
^{1/2}f|_{1/2}, \\
|\partial ^{1/2}u(t,\cdot )-\partial ^{1/2}u(s,\cdot )|_{\alpha /2+1/2}
&\leq &C(t-s)^{1/2}|\partial ^{1/2}f|_{1/2},s\leq t\leq T.
\end{eqnarray*
By Lemma \ref{rem1} (using equivalence of norms), we see that there are
constants $C_{1}=C(\alpha ,\ \beta ,T,\ d$,$\mu ,K)$ and $C_{2}=C_{2}(\alpha
,d)$ such tha
\begin{eqnarray*}
|u|_{\alpha +1} &\leq &C_{1}|f|_{1}, \\
|u|_{1} &\leq &C_{2}(\lambda ^{-1}\wedge T)|f|_{1}, \\
|u(t,\cdot )-u(s,\cdot )|_{\alpha /2+1} &\leq &C_{1}(t-s)^{1/2}|f|_{1},s\leq
t\leq T.
\end{eqnarray*
The statement follows immediately for $\beta =1$ by repeating the proof of
Theorem 6 in \cite{MiP09} and using Lemma \ref{le4} with $\beta =1.$
\end{proof}
Let $c_{i}:U\rightarrow \mathbf{R}^{d},i=1,2,$ be measurable functions and
\nu (d\upsilon )$ be a $\sigma $-finite signed measure on $(U,\mathcal{U}).$
Consider the operators ($i=1,2$
\begin{equation*}
L^{i}=L^{c_{i}}u(x)=\int_{U}[u(x+c_{i}(\upsilon ))-u(x)-\mathbf{1}_{\alpha
\in (1,2)}\mathbf{1}_{U_{1}}(\upsilon )(\nabla u(x),c_{i}(\upsilon ))]\nu
(d\upsilon ),
\end{equation*
where $U_{1}\in \mathcal{U},|\nu |(U_{1}^{c})<\infty $ ($|\nu |$ is the
total variation of $\nu $).
\begin{lemma}
\label{le9}Let $\beta \in (0,1]$. Assum
\begin{eqnarray*}
&&\mathbf{1}_{\alpha \in (1,2)}\{\int_{U_{1}}|c_{i}(\upsilon )|^{\alpha
}|\nu |(d\upsilon )+\int_{U_{1}^{c}}|c_{i}(\upsilon )|\wedge 1|\nu
|(d\upsilon )\} \\
&&+\mathbf{1}_{\alpha \in (0,1]}\int_{U}|c_{i}(\upsilon )|^{\alpha }\wedge
1|\nu |(d\upsilon ) \\
&\leq &K_{1},i=1,2.
\end{eqnarray*}
Then there is $\beta ^{\prime }\in (0,\beta )$ and a constant $C$ such that
for each $\kappa \in (0,1)
\begin{equation*}
\sup_{x}|L^{1}u(x)|\leq C|u|_{\alpha +\beta ^{\prime }}K_{1},
\end{equation*
an
\begin{eqnarray*}
\left[ L^{1}u\right] _{\beta } &\leq &C|u|_{a+\beta }[1_{\alpha \in
(1,2)}\int_{U_{1},|c_{1}|\leq \kappa }|c_{1}|^{\alpha }d|\nu |+\mathbf{1
_{\alpha \in (0,1]}\int_{|c_{1}|\leq \kappa }|c_{1}|^{\alpha }d|\nu |] \\
&&+|u|_{\alpha +\beta ^{\prime }}\kappa ^{-\alpha }K_{1}.
\end{eqnarray*}
Also, there is $\beta ^{\prime }\in (0,\beta )$ such that $\alpha +\beta
^{\prime }\geq \beta $ an
\begin{eqnarray*}
&&\sup_{x}|L^{1}u(x)-L^{2}u(x)| \\
&\leq &C|u|_{\alpha +\beta ^{\prime }}\{\mathbf{1}_{\alpha \in (0,1]}\int
|c_{1}-c_{2}|^{(\alpha +\beta ^{\prime })\wedge 1}\wedge 1d|\nu | \\
&&+\mathbf{1}_{\alpha \in (1,2)}[\int_{U_{1}^{c}}|c_{1}-c_{2}|\wedge 1d|\nu
|+(\int_{U_{1}}|c_{1}-c_{2}|^{\alpha }d|\nu |)^{1/\alpha }]\}.
\end{eqnarray*}
\end{lemma}
\begin{proof}
If $\alpha \in (0,1]$, then for any $\beta ^{\prime }\in (0,\beta )$,
\begin{equation*}
\sup_{x}|L^{1}u(x)|\leq C|u|_{\alpha +\beta ^{\prime }}\left( \int
|c_{1}|^{\alpha }\wedge 1d|\nu |\right) .
\end{equation*}
If $\alpha \in (1,2)$, then
\begin{equation*}
L^{1}u=\int_{U_{1}}...+\int_{U_{1}^{c}}...=L_{1}^{1}u+L_{2}^{2},
\end{equation*
an
\begin{eqnarray*}
\sup_{x}|L_{1}^{1}u(x)| &\leq &C|u|_{\alpha }\int_{U_{1}}|c_{1}|^{\alpha
}d|\nu |, \\
\sup_{x}|L_{2}^{2}u(x)| &\leq &C|u|_{\alpha }\int_{U_{1}^{c}}|c_{1}|\wedge
1d|\nu |.
\end{eqnarray*
If $\alpha \in (0,1],$ then for each $\kappa \in (0,1)
\begin{equation*}
L^{1}u=\int_{|c_{1}|\leq \kappa }...+\int_{|c_{1}|>\kappa
}=L_{1}^{1}u+L_{2}^{1}u
\end{equation*
and
\begin{equation*}
|L_{1}^{1}u|_{\beta }\leq C|u|_{\alpha +\beta }\int_{|c_{1}|\leq \kappa
}|c_{1}|^{\alpha }d|\nu |
\end{equation*
\begin{eqnarray*}
|L_{2}^{2}u|_{\beta } &\leq &C|u|_{\beta }|\nu |\left( |c_{1}|>\kappa
\right) \\
&\leq &C\kappa ^{-\alpha }|u|_{\beta }\int |c_{1}|^{\alpha }\wedge 1d|\nu |.
\end{eqnarray*
If $\alpha \in (1,2),$ the
\begin{equation*}
L^{1}u=\int_{U_{1}}...+\int_{U_{1}^{c}}...=L_{1}^{1}u+L_{2}^{1}u
\end{equation*
an
\begin{eqnarray*}
L_{1}^{1}u(x) &=&\int_{U_{1},|c|\leq \kappa }\int_{0}^{1}(\partial ^{\alpha
-1}\nabla u(x-z)k^{(\alpha -1)}(z,sc_{1}),c_{1})dzd\nu \\
+\int_{U_{1},|c_{1}|>\kappa }... &=&L_{11}^{1}u(x)+L_{12}^{1}u(x).
\end{eqnarray*
We hav
\begin{eqnarray*}
\lbrack L_{11}^{1}u]_{\beta } &\leq &C|\partial ^{\alpha -1}\nabla u|_{\beta
}\int_{U_{1},|c_{1}|\leq \kappa }|c_{1}|^{\alpha }d|\nu |, \\
\lbrack L_{12}^{1}u]_{\beta } &\leq &C\kappa ^{-\alpha }|u|_{1+\beta
}\int_{U_{1}}|c_{1}|^{\alpha }d|\nu |.
\end{eqnarray*
Also,
\begin{eqnarray*}
L_{2}^{1}u &=&\int_{U_{1}^{c},|c_{1}|>1}...+\int_{U_{1}^{c},|c_{1}|\leq 1}...
\\
&=&\int_{U_{1}^{c},|c_{1}|\leq 1}\int_{0}^{1}(\nabla u(x+sc_{1}),c_{1})d\nu
+\int_{U_{1}^{c},|c_{1}|>1}... \\
&=&L_{21}^{1}u+L_{22}^{1}u,
\end{eqnarray*
an
\begin{eqnarray*}
\lbrack L_{21}^{1}u]_{\beta } &\leq &C|\nabla u|_{\beta }\int |c_{1}|\wedge
1d|\nu |, \\
\lbrack L_{22}^{1}u]_{\beta } &\leq &C|u|_{\beta }\int |c_{1}|\wedge 1d|\nu
|.
\end{eqnarray*
Finally,
\begin{eqnarray*}
|L^{1}u(x)-L^{2}u(x)| &\leq &\mathbf{1}_{\alpha \in (0,1]}\int
|u(x+c_{1})-u(x+c_{2})|d|\nu | \\
&&+\mathbf{1}_{\alpha \in
(1,2)}[\int_{U_{1}^{c}}|u(x+c_{1})-u(x+c_{2})|d|\nu | \\
&&+\int_{U_{1}}\int_{0}^{1}|(\nabla u(x+sc_{1})-\nabla u(x),c_{1}) \\
&&-(\nabla u(x+sc_{2})-\nabla u(x),c_{2})|d|\nu |].
\end{eqnarray*
So, there is $\beta ^{\prime }\in (0,\beta )$ such that $\alpha +\beta
^{\prime }\geq \beta $ an
\begin{eqnarray*}
&&|L^{1}u(x)-L^{2}u(x)| \\
&\leq &C|u|_{\alpha +\beta ^{\prime }}[\mathbf{1}_{\alpha \in (0,1]}\int
|c_{1}-c_{2}|^{(\alpha +\beta ^{\prime })\wedge 1}\wedge 1d|\nu |+\mathbf{1
_{\alpha \in (1,2)}\int_{U_{1}^{c}}|c_{1}-c_{2}|\wedge 1d|\nu |] \\
&&+\mathbf{1}_{\alpha \in (1,2)}|\nabla u|_{\alpha
-1}[\int_{U_{1}}|c_{1}-c_{2}|^{\alpha -1}|c_{1}|d|\nu
|+\int_{U_{1}}|c_{2}|^{\alpha -1}|c_{1}-c_{2}|d|\nu |]
\end{eqnarray*
and the last term can be estimated by H\"{o}lder inequality.
\end{proof}
\begin{lemma}
\label{l7}(cf. Lemma 23 in \cite{MiP09})Let $\beta \in (0,1]$ and
assumptions B1-B2 be satisfied. Then for each $\varepsilon >0$ there is a
constant $C_{\varepsilon }$ such that for any $u\in C^{\alpha +\beta }
\mathbf{R}^{d}),t\in \lbrack 0,T],
\begin{equation*}
|Bu|_{\beta }\leq \varepsilon |u|_{\alpha +\beta }+C_{\varepsilon }|u|_{0}.
\end{equation*}
\end{lemma}
\begin{proof}
Let $\beta \in (0,1)$. Since for $(t,x)\in H,z,h\in \mathbf{R}^{d},
\begin{eqnarray*}
B_{t,x+h,z}u(x+h)-B_{t,x,z}u(x) &=&B_{t,x+h,z}u(x+h)-B_{t,x+h,z}u(x) \\
&&+B_{t,x+h,z}u(x)-B_{t,x,z}u(x),
\end{eqnarray*
it follows by Lemma \ref{le9} that for each $\varepsilon >0$ there is a
constant $C_{\varepsilon }$ such that for all $z\in \mathbf{R}^{d},
\begin{equation}
|B_{\cdot ,z}u|_{\beta }\leq \varepsilon |u|_{\alpha +\beta }+C_{\varepsilon
}|u|_{0}. \label{nulis}
\end{equation
Let $\beta =1$. Since for $(t,x)\in H,z,h\in \mathbf{R}^{d},
\begin{eqnarray*}
&&B_{t,x+h,z}u(x+h)-2B_{t,x,z}u(x)+B_{t,x-h,z}u(x-h) \\
&=&[B_{t,x+h,z}u(x+h)-B_{t,x,z}u(x+h)]+[B_{t,x-h,z}u(x+h)-B_{t,x,z}u(x+h)] \\
&&+[B_{t,x-h,z}u(x-h)-B_{t,x,z}u(x-h)]-[B_{t,x-h,z}u(x+h)-B_{t,x,z}u(x+h)] \\
&&+B_{t,x,z}u(x+h)-2B_{t,x,z}u(x)+B_{t,x,z}u(x-h)
\end{eqnarray*
it follows again by Lemma \ref{le9} that for each $\varepsilon >0$ there is
a constant $C_{\varepsilon }$ such that for all $z\in \mathbf{R}^{d},
\begin{equation}
|B_{\cdot ,z}u|_{1}\leq \varepsilon |u|_{\alpha +1}+C_{\varepsilon }|u|_{0}.
\label{vienas}
\end{equation
Finally, if $\beta \in (0,1)$, then for all $(t,x)\in H,z\in \mathbf{R}^{d},
\begin{eqnarray*}
&&B_{t,x+h,x+h}u(x+h)-B_{t,x,x}u(x) \\
&=&B_{t,x+h,x+h}u(x+h)-B_{t,x,x+h}u(x) \\
&&+B_{t,x,x+h}u(x)-B_{t,x,x}u(x),
\end{eqnarray*
an
\begin{eqnarray*}
&&|B_{t,x+h,x+h}u(x+h)-B_{t,x,x+h}u(x)| \\
&\leq &|h|^{\beta }\sup_{z}[B_{\cdot ,z}u]_{\beta }.
\end{eqnarray*
So, the statement follows by (\ref{nulis}) and Lemma \ref{le9}.
If $\beta =1$, then $(t,x)\in H,z,h\in \mathbf{R}^{d},
\begin{eqnarray*}
&&B_{t,x+h,x+h}u(x+h)-2B_{t,x,x}u(x)+B_{t,x-h,x-h}u(x-h) \\
&=&\{B_{t,x+h,x+h}u(x+h)-2B_{t,x+h,x}u(x+h)+B_{t,x+h,x-h}u(x+h)\} \\
&&
\{[B_{t,x-h,x-h}u(x-h)-B_{t,x-h,x}u(x-h)]-[B_{t,x+h,x-h}u(x+h)-B_{t,x+h,x}u(x+h)]\}
\\
&&+\{B_{t,x+h,x}u(x+h)-2B_{t,x,x}u(x)+B_{t,x-h,x}u(x-h)\}
\end{eqnarray*
and the statement follows by (\ref{vienas}) and Lemma \ref{le9}.
\end{proof}
Let $n:\mathbf{R}_{0}^{d}\rightarrow \mathbf{R}$ be a measurable function
satisfying the following conditions:
(i) there is a constant $k_{1}$ such that for all $(t,x)\in H,y\in \mathbf{R
_{0}^{d},
\begin{equation}
|n(y)|\leq k_{1}; \label{6}
\end{equation}
(ii) if $\alpha =1$, then for all $r\in (0,1),
\begin{equation}
\int_{r<|y|\leq 1}yn(y)\frac{dy}{|y|^{d+1}}=0. \label{63}
\end{equation
For $u\in C^{\alpha +\beta }(\mathbf{R}^{d})$, we introduce the operator
\begin{equation*}
\mathcal{A}u(x)=\int_{\mathbf{R}^{d}}\nabla _{y}^{\alpha }u(x)n(y)\frac{dy}
|y|^{d+\alpha }}.
\end{equation*}
\begin{proposition}
\label{prop2}Let $\alpha \in (0,2)$, $\beta \in (0,1],\beta ^{\prime }\in
(0,\beta )$ and (\ref{6}), (\ref{63}) be satisfied.
Then there are constants $C_{1}=C_{1}(\alpha ,\beta ^{\prime
},d),C_{2}=C_{2}(\alpha ,\beta ,d)$ such that for all $u\in C^{\alpha +\beta
}(H)
\begin{eqnarray*}
\sup_{x}|\mathcal{A}u(x)| &\leq &C_{1}k_{1}|u|_{\alpha +\beta ^{\prime }}, \\
\lbrack \mathcal{A}u]_{\beta } &\leq &C_{2}k_{1}|u|_{\alpha +\beta }.
\end{eqnarray*}
\end{proposition}
\begin{proof}
For $\alpha \in (0,1)$, let $\beta ^{\prime }\in (0,\beta )$ be such that
\alpha +\beta ^{\prime }<1$. Then for $u\in C^{\alpha +\beta }(\mathbf{R
^{d}),$
\begin{equation*}
\sup_{x}|\mathcal{A}u(x)|\leq Ck_{1}|u|_{a+\beta ^{\prime }}\int
(|y|^{\alpha +\beta ^{\prime }}\wedge 1)\frac{dy}{|y|^{d+\alpha }}\leq
CK|u|_{\alpha +\beta ^{\prime }}.
\end{equation*
For $\alpha \in \lbrack 1,2),
\begin{eqnarray*}
\mathcal{A}u(x) &=&\int_{|y|\leq 1}\left( \int_{0}^{1}(\nabla u(x+sy)-\nabla
u(x),y\right) ds]m(y)\frac{dy}{|y|^{d+\alpha }} \\
&&+\int_{|y|>1}[u(x+y)-u(x)-\mathbf{1}_{\alpha \in (1,2)}\left( \nabla
u(x),y\right) ]m(y)\frac{dy}{|y|^{d+\alpha }} \\
&=&L_{1}u(x)+L_{2}u(x).
\end{eqnarray*}
Obviously, for any $\mu \in (0,1)$ such that $1+\mu \in (\alpha ,\alpha
+\beta )
\begin{equation*}
\sup_{x}|L_{1}u(x)|\leq CK|\nabla u|_{\mu }\int_{|y|\leq 1}|y|^{1+\mu
-d-\alpha }dy\leq CK|u|_{1+\mu },
\end{equation*
an
\begin{equation*}
\sup_{x}|L_{2}u(x)|\leq CK\sup_{x}(|u(x)|+|\nabla u(x)|).
\end{equation*}
In order to estimate the differences, first we note that for $u\in C^{1+\mu
}(\mathbf{R}^{d}),\mu \in (0,1),
\begin{eqnarray*}
&&u(x+h)-u(x)+u(x-h)-u(x) \\
&=&\int_{0}^{1}\left( \nabla u(x+sh)-\nabla u(x-sh),t\right) ds
\end{eqnarray*
and
\begin{equation}
|u(x+h)-u(x)+u(x-h)-u(x)|\leq C|\nabla u|_{\mu }|h|^{1+\mu },x,h\in \mathbf{
}^{d}. \label{fo3}
\end{equation
Also, for $u\in C^{\mu }(\mathbf{R}^{d}),\mu \in (0,1),
\begin{equation}
|u(x+h)-u(x)|\leq C|u|_{\mu }|h|^{\mu },x,h\in \mathbf{R}^{d}. \label{fo5}
\end{equation}
Fix $h\in \mathbf{R}^{d}$ with $a=|h|\in (0,1).$ Then
\begin{equation*}
\mathcal{A}u(x)=\int_{|y|\leq a}...+\int_{|y|>a}...=I_{1}(x)+I_{2}(x),x\in
\mathbf{R}^{d},
\end{equation*
wher
\begin{eqnarray*}
I_{1}(x) &=&\int_{|y|\leq a}[u(x+y)-u(x)-\mathbf{1}_{\alpha \in \lbrack
1,2)}\left( \nabla u(x),y\right) ]m(y)\frac{dy}{|y|^{d+\alpha }}, \\
I_{2}(x) &=&\int_{|y|>a}[u(x+y)-u(x)-\mathbf{1}_{\alpha \in (1,2)}\left(
\nabla u(x),y\right) ]m(y)\frac{dy}{|y|^{d+\alpha }}.
\end{eqnarray*
For $\alpha \in (0,1),\beta \in (0,1]$, let $\beta ^{\prime }\in (0,\beta )$
and $\alpha +\beta ^{\prime }<1$. Then for $u\in C^{\alpha +\beta }(\mathbf{
}^{d})$ by Lemma \ref{r1}
\begin{eqnarray*}
|I_{1}(x+h)-I_{1}(x)| &\leq &k_{1}\int_{|y|\leq a}|\partial ^{\alpha +\beta
^{\prime }}u(x+h-z)-\partial ^{\alpha +\beta ^{\prime }}u(x-z)|~|k^{(\alpha
+\beta ^{\prime })}(z,y)|\frac{dy}{|y|^{d+\alpha }} \\
&\leq &Ck_{1}|\partial ^{\alpha +\beta ^{\prime }}|_{\beta -\beta ^{\prime
}}a^{\beta -\beta ^{\prime }}\int_{|y|\leq a}|y|^{\alpha +\beta ^{\prime }
\frac{dy}{|y|^{d+\alpha }}\leq Ck_{1}|u|_{\alpha +\beta }a^{\beta }.
\end{eqnarray*
For $\alpha \in \lbrack 1,2),\beta \in (0,1]$, let $\beta ^{\prime }\in
(0,1) $ and $\alpha <1+\beta ^{\prime }<\alpha +\beta $. Then for $u\in
C^{\alpha +\beta }(\mathbf{R}^{d})$ by Lemma \ref{r1}
\begin{eqnarray*}
&&|I_{1}(x+h)-I_{1}(x)| \\
&\leq &Ck_{1}\int_{|y|\leq a}\int_{0}^{1}\int |\left( \partial ^{\beta
^{\prime }}\nabla u(x+h-z)-\partial ^{\beta ^{\prime }}\nabla
u(x-z),y\right) |~|k^{(\beta ^{\prime })}(z,sy)|dzds\frac{dy}{|y|^{d+\alpha
} \\
&\leq &Ck_{1}|\partial ^{\beta ^{\prime }}\nabla u|_{\alpha +\beta -\beta
^{\prime }-1}a^{\alpha +\beta -1-\beta ^{\prime }}\int_{|y|\leq
a}|y|^{1+\beta ^{\prime }}\frac{dy}{|y|^{d+\alpha }}\leq Ck_{1}|u|_{\alpha
+\beta }a^{\beta }.
\end{eqnarray*
Let $\alpha \in (0,1],\beta \in (0,1),$ and $\alpha ^{\prime }<\alpha $ be
such that $\beta +\alpha -\alpha ^{\prime }<1$. By Lemma \ref{r1} and (\re
{fo5})
\begin{eqnarray*}
|I_{2}(x+h)-I_{2}(x)| &\leq &k_{1}\int_{|y|>a}|\partial ^{\alpha ^{\prime
}}u(x+h-z)-\partial ^{\alpha ^{\prime }}u(x-z)|~|k^{(\alpha ^{\prime
})}(y,z)|\frac{dy}{|y|^{d+\alpha }} \\
&\leq &Ck_{1}|\partial ^{\alpha ^{\prime }}u|_{\alpha +\beta -\alpha
^{\prime }}a^{\alpha +\beta -\alpha ^{\prime }}\int_{|y|>a}|y|^{\alpha
^{\prime }}\frac{dy}{|y|^{d+\alpha }}\leq Ck_{1}|u|_{\alpha +\beta }a^{\beta
}.
\end{eqnarray*
For $\alpha \in (0,1],$ $\beta =1,$ let $\alpha ^{\prime }\in (0,\alpha )$.
Then $1<\beta +\alpha -\alpha ^{\prime }<2$,
\begin{equation*}
I_{2}(x)=\int_{|y|>a}\int \partial ^{\alpha ^{\prime }}u(x-z)k^{(\alpha
^{\prime })}(z,y)m(y)\frac{dy}{|y|^{d+\alpha }}.
\end{equation*
and by (\ref{fo3})
\begin{eqnarray*}
|I_{2}(x+h)-2I_{2}(x)+I_{2}(x-h)| &\leq &Ck_{1}|\partial ^{\alpha ^{\prime
}}u|_{\alpha +\beta -\alpha ^{\prime }}a^{\alpha +\beta -\alpha ^{\prime
}}\int_{|y|>a}|y|^{\alpha ^{\prime }}\frac{dy}{|y|^{d+\alpha }} \\
&\leq &Ck_{1}|u|_{\alpha +\beta }a^{\beta }.
\end{eqnarray*
Let $x,\bar{x}\in \mathbf{R}^{d},a=|x-\bar{x}|$, and $\beta ^{\prime }<1$ be
such that $\alpha +\beta ^{\prime }<2$ and $0\leq \beta -\beta ^{\prime }<1
. By Lemma \ref{r1},
\begin{eqnarray*}
&&|L_{1}u(x)-L_{1}(\bar{x})| \\
&\leq &K\int_{|y|\leq a}\int_{0}^{1}\int |\partial ^{\alpha +\beta ^{\prime
}-1}\nabla u(x-z)-\partial ^{\alpha +\beta ^{\prime }-1}\nabla u(x-z)|\times
\\
&&\times ~|k^{(\alpha +\beta ^{\prime }-1)}(sy,z)||y|\frac{dsdzdy}
|y|^{d+\alpha }} \\
&\leq &CK|u|_{\alpha +\beta }a^{\beta -\beta ^{\prime }}\int_{|y|\leq
a}|y|^{\alpha +\beta ^{\prime }}\frac{dy}{|y|^{d+\alpha }}=CK|u|_{\alpha
+\beta }a^{\beta }.
\end{eqnarray*
For $\alpha \in (1,2),\beta \in (0,1)$, let $1<\alpha ^{\prime }<\alpha $ be
such that $\alpha -\alpha ^{\prime }+\beta <1$. By Lemma \ref{r1},
\begin{eqnarray*}
&&|I_{2}(x+h)-I_{2}(x)| \\
&=&|\int_{|y|>a}\int_{0}^{1}\int \left( \partial ^{\alpha ^{\prime
}-1}\nabla u(x+h-z)-\partial ^{\alpha ^{\prime }-1}\nabla u(x-z),y\right)
k^{(\alpha ^{\prime }-1)}(z,sy)m(y)\frac{dzdsdy}{|y|^{d+\alpha }}| \\
&\leq &Ck_{1}|\partial ^{\alpha ^{\prime }-1}\nabla u|_{\alpha +\beta
-\alpha ^{\prime }}a^{\alpha +\beta -\alpha ^{\prime
}}\int_{|y|>a}|y|^{\alpha ^{\prime }}\frac{dzdsdy}{|y|^{d+\alpha }}\leq
Ck_{1}|u|_{\alpha +\beta }a^{\beta }.
\end{eqnarray*
For $\alpha \in (1,2),\beta =1,$ we hav
\begin{equation*}
I_{2}(x)=\int_{|y|>a}\int_{0}^{1}\left( \nabla u(x+sy)-\nabla u(x),y\right)
m(y)ds\frac{dy}{|y|^{d+\alpha }},
\end{equation*
and, by (\ref{fo3})
\begin{eqnarray*}
|I_{2}(x+h)-2I_{2}(x)+I_{2}(x-h)| &\leq &Ck_{1}|\nabla u|_{\alpha +\beta
-1}a^{\alpha }\int_{|y|>a}|y|\frac{dy}{|y|^{d+\alpha }} \\
&=&Ck_{1}|u|_{a+\beta }a.
\end{eqnarray*}
The statement follows.
\end{proof}
We will need a generalization of this statement. Let $n:\mathbf{R}^{d
\mathbf{\times R}_{0}^{d}\rightarrow \mathbf{R}$ be a measurable function
satisfying the following conditions:
(i) there is a constant $k_{1}$ such that for all $x\in \mathbf{R}^{d},y\in
\mathbf{R}_{0}^{d}
\begin{equation}
|n(x,y)|\leq k_{1}; \label{91}
\end{equation}
(ii) for $\beta \in (0,1]$ there is a constant $k_{2}$ such that for all
y\in \mathbf{R}_{0}^{d}
\begin{equation}
\lbrack n(\cdot ,y)]_{\beta }\leq k_{2}; \label{9}
\end{equation}
(iii) if $\alpha =1$, then for all $x\in \mathbf{R}^{d},r\in (0,1),
\begin{equation}
\int_{r<|y|\leq 1}yn(x,y)\frac{dy}{|y|^{d+1}}=0. \label{92}
\end{equation
For $u\in C^{\alpha +\beta }(\mathbf{R}^{d})$, we introduce an operato
\begin{equation*}
\mathcal{A}u(x)=\int_{\mathbf{R}^{d}}\nabla _{y}^{\alpha }u(x)n(x,y)\frac{d
}{|y|^{d+\alpha }}.
\end{equation*}
\begin{corollary}
\label{cor1}Let $\alpha \in (0,2)$, $\beta \in (0,1],\beta ^{\prime }\in
(0,\beta )$ and (\ref{91})-(\ref{92}) be satisfied. Then there is a constant
$C=C(\alpha ,\beta ,\beta ^{\prime },d)$ such that for all $u\in C^{\alpha
+\beta }(H)
\begin{equation*}
|\mathcal{A}u|_{\beta }\leq C[k_{1}|u|_{\alpha +\beta }+k_{2}|u|_{\alpha
+\beta ^{\prime }}].
\end{equation*}
\end{corollary}
\begin{proof}
For $u\in C^{\alpha +\beta }(\mathbf{R}^{d})$ conside
\begin{equation*}
\mathcal{A}_{z}u(x)=\int_{\mathbf{R}^{d}}\nabla _{y}^{\alpha }u(x)n(z,y
\frac{dy}{|y|^{d+\alpha }},x,z\in \mathbf{R}^{d}.
\end{equation*
Let $\beta \in (0,1)$. Since for $h,x\in \mathbf{R}^{d},
\begin{eqnarray*}
\mathcal{A}_{x+h}u(x+h)-\mathcal{A}_{x}u(x) &=&\left( \mathcal{A
_{x+h}u(x+h)-\mathcal{A}_{x+h}u(x)\right) \\
&&+\left( \mathcal{A}_{x+h}u(x)-\mathcal{A}_{x}u(x)\right) ,
\end{eqnarray*
the statement follows by Proposition \ref{prop2}.
Let $\beta =1$. Since, similarly, for $h,x\in \mathbf{R}^{d},
\begin{eqnarray*}
&&\mathcal{A}u(x+h)-2\mathcal{A}u(x)+\mathcal{A}u(x-h) \\
&=&\{\mathcal{A}_{x-h}u(x+h)-2\mathcal{A}_{x}u(x+h)+\mathcal{A}_{x+h}u(x+h)\}
\\
&&+\{[\mathcal{A}_{x-h}u(x+t)-\mathcal{A}_{x}u(x+t)]-[\mathcal{A
_{x-h}u(x-t)-\mathcal{A}_{x}u(x-t)]\} \\
&&+\{\mathcal{A}_{x}u(x+h)-2\mathcal{A}_{x}u(x)+\mathcal{A}_{x}u(x-h)\},
\end{eqnarray*
the statement follows by Proposition \ref{prop2}.
\end{proof}
\section{Equation with coefficients independent of spatial variable}
In this section, we consider the Cauchy problem
\begin{equation}
\left\{
\begin{array}{ll}
\partial _{t}u(t,x)=A_{t}u(t,x)-\lambda u(t,x)+f(t,x), & (t,x)\in H \\
u(0,x)=0, & x\in \mathbf{R}^{d}
\end{array
\right. \label{one}
\end{equation
assuming that the function $m(t,x,y)$ does not depend on $x$.
\begin{theorem}
\label{t1}Let $\alpha \in (0,2),\beta \in (0,1],m(t,x,y)=m(t,y)$ and
Assumption A be satisfied.
Then for each $f\in C^{\beta }(H)$ there is a unique solution $u\in
C^{\alpha +\beta }(H)$ to (\ref{one}). Moreover, the solution satisfies (\re
{cc1})-(\ref{cc3}).
\end{theorem}
\begin{proof}
\emph{Uniqueness. }Let $u^{1},u^{2}\in C^{\alpha +\beta }$ be two solutions
to (\ref{one}). Then the function $u=u^{1}-u^{2}$ satisfies (\ref{one}) with
$f=0$.
Let a nonnegative $\zeta \in C_{0}^{\infty }(\mathbf{R}^{d})$ be such that
\int \zeta dx=1$. Denote
\begin{equation*}
\zeta _{\varepsilon }(x)=\varepsilon ^{-d}\zeta (x/\varepsilon ),x\in
\mathbf{R}^{d},\varepsilon \in (0,1),
\end{equation*
and
\begin{equation*}
u_{\varepsilon }(t,x)=u(t,\cdot )\ast \zeta _{\varepsilon }(x),(t,x)\in
\text{.}
\end{equation*
Then $u_{\varepsilon }$ solves (\ref{one}) with $f=0$.
Let $\left( \Omega ,\mathcal{F},\mathbf{P}\right) $ be a complete
probability space with a filtration of $\sigma $-algebras $\mathbb{F}=\left(
\mathcal{F}_{t}\right) _{t\geq 0}$ satisfying the usual conditions. We fix
t_{0}\in (0,T)$ and introduce an $\mathbb{F}$-adapted Poisson point measure
p(dt,dy)$ on $[0,t_{0}]\times \mathbf{R}_{0}^{d}$ with a compensator
m(t_{0}-t,y)dtdy/|y|^{d+\alpha }$. Le
\begin{equation*}
q(dt,dy)=p(dt,dy)-m(t_{0}-t,y)\frac{dtdy}{|y|^{d+\alpha }}
\end{equation*
be the corresponding martingale measure and
\begin{equation*}
X_{t}=\int_{s_{0}}^{t}\int \chi _{\alpha }(y)yq(ds,sy)+\int_{0}^{t}\int
(1-\chi _{\alpha }(y))yp(ds,dy)
\end{equation*
for $0\leq t\leq t_{0}.$ By Ito's formul
\begin{eqnarray*}
u_{\varepsilon }(t_{0},x) &=&u_{\varepsilon }(t_{0},x)-\mathbf{E
u_{\varepsilon }(0,x+X_{t_{0}})e^{-\lambda t_{0}} \\
&=&\mathbf{E}\int_{0}^{t_{0}}e^{-\lambda t}\left[ \frac{\partial
u_{\varepsilon }}{\partial t}-Au_{\varepsilon }+\lambda u_{\varepsilon
\right] (t-t_{0},x+X_{t})dt=0.
\end{eqnarray*
Since $\varepsilon ,t_{0}$ and $x$ are arbitrary, we have $u=0.$
\emph{Existence. }First we prove the existence of a solution to (\ref{one})
for a smooth input function $f$.
We introduce an $\mathbb{F}$-adapted Poisson measure $\bar{p}(dt,dz)$ on
[0,\infty )\times \mathbf{R}_{0}$ with a compensator $dtdz/z^{2}$. Le
\begin{equation*}
\bar{q}(dt,dz)=\bar{p}(dt,dz)-\frac{dtdz}{z^{2}}
\end{equation*
be the corresponding martingale measure. According to Lemma 14.50 in \cit
{Jac79}, there is a measurable function $\bar{c}:[0,T]\times \mathbf{R
_{0}\rightarrow \mathbf{R}^{d}$ such that for every Borel $\Gamma \subseteq
\mathbf{R}_{0}^{d}
\begin{equation*}
\int_{\Gamma }\left( m(t,y)-m_{0}(t,y)\right) \frac{dy}{|y|^{d+\alpha }
=\int 1_{\Gamma }(\bar{c}(t,z))\frac{dz}{z^{2}}.
\end{equation*
Let
\begin{equation*}
Y_{t}=\int_{0}^{t}\int (1-\chi _{\alpha }(\bar{c}(t,z)))\bar{c}(t,z)\bar{p
(dz,dz)+\int_{0}^{t}\int \chi _{\alpha }(\bar{c}(t,z))c(t,z)\bar{q}(ds,dz).
\end{equation*
For $f\in C_{b}^{\infty }(H),$ we consider the equatio
\begin{eqnarray}
\partial _{t}u(t,x) &=&A_{\alpha }^{0}u(t,x)+f(t,x-Y_{t}),(t,x)\in H,
\label{d1} \\
u(0,x) &=&0,x\in \mathbf{R}^{d}. \notag
\end{eqnarray}
By Lemma \ref{l8}, there is a unique solution $u\in C_{b}^{\infty }(H)$ to
\ref{d1}). Moreover, the solution satisfies (\ref{cc1})-(\ref{cc3}) $\mathbf
P}$-a.s. In addition, the solution is $\mathbb{F}$-adapted becaus
\begin{equation*}
u(t,x)=\int_{0}^{t}\int G_{s,t}^{\lambda }(y)f(s,x-y-Y_{s})dyds.
\end{equation*}
Using (\ref{4}), we have for any multiindex $\gamma ,
\begin{equation*}
D^{\gamma }u(t,x)=\int_{0}^{t}\int G_{s,t}^{\lambda }(y)D_{x}^{\gamma
}f(s,x-y-Y_{s})dyds
\end{equation*
an
\begin{equation}
\text{essup}_{\omega \in \Omega }\sup_{t,x}|D_{x}^{\gamma }u(t,x)|<\infty .
\label{a1}
\end{equation}
Let $\bar{A}$ be the operator defined as the operator $A$ with $m$ replaced
by $m-m_{0}$. According to (\ref{d1}) and the Ito-Wentzell formula (see \cit
{mik})
\begin{eqnarray}
u(t,x+Y_{t})-u(0,x) &=&\int_{0}^{t}[\partial _{s}u(s,x+Y_{s})+\bar{A
u(s,x+Y_{s})+M_{t} \label{a2} \\
&=&\int_{0}^{t}[Au(s,x+Y_{s})-\lambda u(s,x+Y_{s})+f(s,x)]ds+M_{t}, \notag
\end{eqnarray
wher
\begin{equation*}
M_{t}=\int_{0}^{t}\int [u(s,x+Y_{s-}+\bar{c}(t,z))-u(s,x+Y_{s-})]\bar{q
(ds,dz).
\end{equation*}
Taking expectation on both sides of (\ref{a2}) and using (\ref{a1}), we
conclude that the function $v(t,x)=\mathbf{E}u(t,x+Y_{t})$ belongs to
C_{b}^{\infty }(H)$ and solves (\ref{one}). Moreover, $v$ satisfies (\re
{cc1})-(\ref{cc3}) because $\ u$ satisfies (\ref{cc1})-(\ref{cc3}) $\mathbf{
}$-a.s.
Next we prove the existence of a solution to (\ref{one}) for $f\in C^{\beta
}(H)$. By Lemma \ref{le4}, there is a sequence $f_{n}\in C_{b}^{\infty }(H)$
such tha
\begin{equation}
|f_{n}|_{\beta }\leq 2|f|_{\beta },|f|_{\beta }\leq \lim
\inf_{n}|f_{n}|_{\beta }, \label{a3}
\end{equation
and for every $\kappa \in (0,\beta )$
\begin{equation}
\lim_{n\rightarrow \infty }|f_{n}-f|_{\kappa }=0. \label{a4}
\end{equation}
According to the first part of the proof and (\ref{a3}), for each $n$ there
is a unique solution $u_{n}\in C^{\alpha +\beta }(H)$ to (\ref{one}) with $f$
replaced by $f_{n}.$ Moreover, there are constants $C_{1}=C_{1}(\alpha
,\beta ,\mu ,K,T)$ and $C_{2}=C_{2}(\alpha ,d)$ such tha
\begin{equation}
|u_{n}|_{\alpha +\beta }\leqslant C_{1}|f|_{\beta }, \label{ine0}
\end{equation
\begin{equation}
|u_{n}|_{\beta }\leq C_{2}(\alpha ,d)(\lambda ^{-1}\wedge T)|f|_{0,\beta ;p},
\label{a5}
\end{equation
and for all $s\leq t\leq T,$
\begin{equation}
|u_{n}(t,\cdot )-u_{n}(s,\cdot )|_{\alpha /2+\beta }\leq
C(t-s)^{1/2}|f|_{\beta }. \label{ine2}
\end{equation
Fix an arbitrary $\kappa \in (0,\beta ).$ Again, by the first part of he
proof, there is a constant $C$ not depending on $n$ such tha
\begin{equation*}
|u_{n}|_{\alpha +\kappa }\leqslant C|f_{n}|_{\kappa }.
\end{equation*
Moreover, by Lemma \ref{le4} and (\ref{a4})
\begin{equation*}
|u_{n}-u_{k}|_{\alpha +\kappa }\leqslant C|f_{n}-f_{k}|_{\kappa }\rightarrow
0
\end{equation*
as $n,k\rightarrow \infty $. Hence, there is $u\in C^{\alpha +\kappa }(H)$
such tha
\begin{equation}
\lim_{n\rightarrow \infty }|u_{n}-u|_{\alpha +\kappa }=0. \label{a6}
\end{equation
According to (\ref{ine0}) and (\ref{a6}), we have $[\partial ^{\alpha
}u_{n}]_{\beta }\leq C_{1}|f|_{\beta }$ and \TEXTsymbol{\vert}$\partial
^{\alpha }u_{n}-\partial ^{\alpha }u|_{0}\rightarrow 0$ as $n\rightarrow
\infty $. Therefore $\left[ \partial ^{\alpha }u\right] _{\beta }\leq
C_{1}|f|_{\beta }$ and $u\in C^{\alpha +\beta }(H)$. Passing to the limit in
(\ref{ine0})-(\ref{ine2}) as $n\rightarrow \infty $ we conclude that $u$
satisfies (\ref{cc1})-(\ref{cc3}). Finally, passing to the limit in the
equatio
\begin{equation*}
u_{n}(t,x)=\int_{0}^{t}\left[ Au_{n}-\lambda u_{n}+f_{n}\right] (s,x)ds
\end{equation*
and using Corollary \ref{cor1}, we conclude that $u$ solves (\ref{one}).
The theorem is proved.
\end{proof}
\section{Proof of Theorem \protect\ref{main}}
We follow the proof of Theorem 5 in \cite{MiP09} with obvious changes.
It is well known that for an arbitrary but fixed $\delta >0$ there is a
family of cubes $D_{k}\subseteq \tilde{D}_{k}\subseteq \mathbf{R}^{d}$ and a
family of deterministic functions $\eta _{k}\in C_{0}^{\infty }(\mathbf{R
^{d})$ with the following properties:
1. For all $k\geq 1,D_{k}$ and $\tilde{D}_{k}$ have a common center $x_{k},$
diam $D_{k}\leq \delta ,$dist$(D_{k},\mathbf{R}^{d}\backslash \tilde{D
_{k})\leq C\delta $ for a constant $C=C(d)>0,\cup _{k}D_{k}=\mathbf{R}^{d},$
and $1\leq \sum_{k}1_{\tilde{D}_{k}}\leq 2^{d}.$
2. For all $k$, $0\leq \eta _{k}\leq 1,\eta _{k}=1$ in $D_{k},\eta _{k}=0$
outside of $\tilde{D}_{k}$ and for all multiindices $\gamma ,|\gamma |\leq 3,
\begin{equation*}
|\partial ^{\gamma }\eta _{k}|\leq C(d)\delta ^{-|\gamma |}.
\end{equation*
For $\alpha \in (0,2),k\geq 1,$ denot
\begin{eqnarray*}
A_{k}u(t,x) &=&A_{t,x_{k}}u(t,x), \\
E_{k}u(t,x) &=&\int [u(t,x+y)-u(t,x)][\eta _{k}(x+y)-\eta
_{k}(x)]m(t,x_{k},y)\frac{dy}{|y|^{d+\alpha }}, \\
E_{k,1}u(t,x) &=&\int [u(t,x+y)-u(t,x)][\eta _{k}(x+y)-\eta _{k}(x)]\frac{d
}{|y|^{d+\alpha }}, \\
F_{k}u(t,x) &=&u(t,x)A_{k}\eta _{k}(x),F_{k,1}u(t,x)=u(t,x)\partial ^{\alpha
}\eta _{k}(x).
\end{eqnarray*
We notice that
\begin{equation}
A_{k}(u\eta _{k})=\eta _{k}A_{k}u+E_{k}u+F_{k}u \label{f0}
\end{equation
an
\begin{equation}
\partial ^{\alpha }(u\eta _{k})=\eta _{k}\partial ^{\alpha
}u+E_{k,1}u+F_{k,1}u. \label{f00}
\end{equation}
It is readily checked that there is $\beta ^{\prime }\in (0,\beta )$ and a
constant $C=$ $C=C(\alpha ,\beta ,d,K,\delta )$ such tha
\begin{equation*}
\sup_{k}\left( |E_{k}^{(\alpha )}u(t,\cdot )|_{\beta }+|E_{k,1}^{(\alpha
)}u(t,\cdot )|_{\beta }\right) \leq C|u|_{\alpha +\beta ^{\prime }}
\end{equation*
and, by Corollary \ref{cor1}
\begin{equation}
\sup_{k}\left( |F_{k}^{(\alpha )}u(t,\cdot )|_{\beta }+|F_{k,1}^{(\alpha
)}u(t,\cdot )|_{\beta }\right) \leq C|u|_{\beta }. \label{f2}
\end{equation
Hence, for each $\varepsilon >0$ there exists a constant $C=C(\alpha ,\beta
,d,K,\delta ,\varepsilon )$ such tha
\begin{equation}
\sup_{k}\left( |E_{k}^{(\alpha )}u(t,\cdot )|_{\beta }+|E_{k,1}^{(\alpha
)}u(t,\cdot )|_{\beta }\right) \leq \varepsilon |u|_{\alpha +\beta
}+C|u|_{0}. \label{f20}
\end{equation}
Elementary calculation shows that for every $u\in C^{\alpha +\beta }(H),
\begin{eqnarray}
|u|_{0} &\leq &\sup_{k}\sup_{x}|\eta _{k}(x)u(x)|, \notag \\
|u|_{\beta } &\leq &\sup_{k}|\eta _{k}u|_{\beta }+C|u|_{0}, \label{f3} \\
\sup_{k}|\eta _{k}u|_{\beta } &\leq &|u|_{\beta }+C|u|_{0}, \notag
\end{eqnarray}
the constant $C=C(\beta ,d,\delta ).$ By (\ref{f00}) and (\ref{f3}), we have
\begin{eqnarray*}
|u|_{\alpha ,\beta } &=&|u|_{0}+|\partial ^{\alpha }u|_{\beta }\leq
\sup_{k}|\eta _{k}\partial ^{\alpha }u|_{\beta }+C|u|_{0} \\
&=&\sup_{k}|\partial ^{\alpha }(\eta _{k}u)-E_{k,1}u-F_{k,1}u|_{\beta
}+C|u|_{0}.
\end{eqnarray*}
By (\ref{f20}) and (\ref{f2}), for each $\varepsilon >0$ there is a constant
$C=C(\varepsilon ,\alpha ,\beta ,d,\delta )$ such that for every $u\in
C^{\alpha +\beta }(H)
\begin{equation*}
|u|_{\alpha ,\beta }\leq \sup_{k}|\partial ^{\alpha }\left( u\eta
_{k}\right) |_{\beta }+\varepsilon |\partial ^{\alpha }u|_{\beta }+C|u|_{0}.
\end{equation*
Therefore
\begin{equation*}
|u|_{\alpha ,\beta }\leq 2\sup_{k}|u\eta _{k}|_{\alpha ,\beta }+C|u|_{0},
\end{equation*
where the constant $C=C(\alpha ,\beta ,d,\delta )$. This estimate, together
with Lemma \ref{rem1}, implie
\begin{equation}
|u|_{\alpha +\beta }\leq C_{1}\sup_{k}|u\eta _{k}|_{\alpha +\beta
}+C_{2}|u|_{0}, \label{f4}
\end{equation
where the constants $C_{1}=C_{1}(\alpha ,\beta ,d),C_{2}=C_{2}(\alpha ,\beta
,d,\delta ).$
Let $u\in C^{\alpha +\beta }(H)$ be a solution of (\ref{eq1}). Then $\eta
_{k}u$ satisfies the equatio
\begin{eqnarray}
\partial _{t}(\eta _{k}u) &=&A_{k}(\eta _{k}u)-\lambda (\eta _{k}u)+\eta
_{k}(Au-A_{k}u)+\eta _{k}f+\eta _{k}Bu \label{eqk} \\
&&-F_{k}u-E_{k}u. \notag
\end{eqnarray
By Theorem \ref{t1}
\begin{equation*}
|\eta _{k}u|_{\alpha +\beta }\leq C\left( |\eta _{k}(Au-A_{k}u)|_{\beta
}+|\eta _{k}Bu|_{\beta }+|\eta _{k}f|_{\beta }+|F_{k}u|_{\beta
}+|E_{k}u|_{\beta }\right) ,
\end{equation*
where the constant $C=C(\alpha ,\beta ,d,\mu ,K,T)$. By (\ref{f4})
\begin{equation}
|u|_{\alpha +\beta }\leq C_{1}\left( \sup_{k}|\eta _{k}f|_{\beta }+I\right)
+C_{2}|u|_{0}, \label{form00}
\end{equation
where the constants $C_{1}=C_{1}(\alpha ,\beta ,d,\mu
,K,T),C_{2}=C_{2}(\alpha ,\beta ,d,K,\delta ,T)$ an
\begin{equation*}
I=\sup_{k}\left( |\eta _{k}(Au-A_{k}u)|_{\beta }+|\eta _{k}Bu|_{\beta
}+|F_{k}u|_{\beta }+|E_{k}u|_{\beta }\right) .
\end{equation*
By Corollary \ref{cor1}, there is $\beta ^{\prime }\in (0,\beta )$ such that
\begin{equation*}
|\eta _{k}(Au-A_{k}u)|_{\beta }\leq C_{1}\left( \delta ^{\beta }|u|_{\alpha
+\beta }+C_{2}|u|_{\alpha +\beta ^{\prime }}\right) ,
\end{equation*
where the constants $C_{1}=C_{1}(\alpha ,\beta ,d,K),C_{2}=C_{2}(\alpha
,\beta ,d,K,\delta )$. Therefore, for each $\varepsilon >0$ we can choose
\delta >0$ so tha
\begin{equation*}
|\eta _{k}(Au-A_{k}u)|_{\beta }\leq \varepsilon |u|_{\alpha +\beta
}+C|u|_{0},
\end{equation*
where the constant $C=C(\alpha ,\beta ,d,K,\varepsilon )$. Hence, by (\re
{f20}), (\ref{f2}) and Lemma \ref{l7}, for each $\varepsilon >0\,$\ we can
choose $\delta >0$ such that
\begin{equation}
I\leq \varepsilon |u|_{\alpha +\beta }+C|u|_{0}, \label{f9}
\end{equation
where the constant $C=C(\alpha ,\beta ,d,K,\varepsilon )$. This estimate,
together with (\ref{form00}) and (\ref{f0}), implies
\begin{equation}
|u|_{\alpha +\beta }\leq C[|f|_{\beta }+|u|_{0}], \label{form01}
\end{equation
where the constant $C=C(\alpha ,\beta ,d,K,\mu ,T)$.
\bigskip On the other hand, according to (\ref{eqk}) and Theorem \ref{t1}
\begin{eqnarray*}
|u|_{0} &\leq &\sup_{k}|\eta _{k}u|_{\beta }\leq \mu (\lambda
)\sup_{k}[|f|_{\beta }+|\eta _{k}(Au-A_{k})|_{\beta }+|\eta _{k}Bu|_{\beta }
\\
&&+|F_{k}u|_{\beta }+|E_{k}u|_{\beta }],
\end{eqnarray*
where $\mu (\lambda )\rightarrow 0$ as $\lambda \rightarrow \infty $. So,
\begin{equation}
|u|_{0}\leq C\mu (\lambda )(|f|_{\beta }+|u|_{\alpha +\beta }).
\label{form02}
\end{equation
The inequalities (\ref{form01}) and (\ref{form02}) imply that there is
\lambda _{0}>0$ and a constant $C$ not depending on $u$ such tha
\begin{equation}
|u|_{\alpha +\beta }\leq C|f|_{\beta } \label{form03}
\end{equation
if $\lambda \geq \lambda _{0}.$ If $u\in C^{\alpha +\beta }(H)$ solves (\re
{eq1}) with $\lambda \leq \lambda _{0}$, then $\tilde{u}(t,x)=e^{-\lambda
(\lambda _{0}-\lambda )t}u(t,x)$ solves the same equation with $\lambda
=\lambda _{0}$, and by (\ref{form03}
\begin{equation*}
|u|_{\alpha +\beta }\leq e^{(\lambda _{0}-\lambda )T}|\tilde{u}|_{\alpha
+\beta }\leq Ce^{(\lambda _{0}-\lambda )T}|f|_{\beta }.
\end{equation*
So, (\ref{form03}) holds for all $\lambda \geq 0$.
By Theorem \ref{t1} and (\ref{f4}), there is a constant $C$ such that for
all $s\leq t\leq T,$
\begin{eqnarray*}
|u(t,\cdot )-u(s,\cdot )|_{\alpha /2+\beta } &\leq &C\sup_{k}|\eta
_{k}u(t,\cdot )-\eta _{k}u(s,\cdot )|_{\alpha /2+\beta } \\
&\leq &C(t-s)^{1/2}\left( |f|_{\beta }+|u|_{\alpha +\beta }\right) .
\end{eqnarray*
Therefore there is a constant $C$ such that for all $s\leq t\leq T,
\begin{equation*}
|u(t,\cdot )-u(s,\cdot )|_{\alpha /2+\beta }\leq C(t-s)^{1/2}|f|_{\beta }.
\end{equation*}
We finish the proof applying the continuation by parameter argument. Let
\begin{equation*}
\,L_{\tau }u=\tau Lu+\left( 1-\tau \right) \partial ^{\alpha }u,\tau \in
\left[ 0,1\right] .
\end{equation*}
We introduce the space $\hat{C}^{\alpha +\beta }\left( H\right) $ of
functions $u\in C^{\alpha +\beta }(H)$ such that for each $\left( t,x\right)
$, $u\left( t,x\right) =\int_{0}^{t}F\left( s,x\right) \,ds,$where $F\in
C^{\beta }\left( H\right) .$ It is a Banach space with respect to the norm
\begin{equation*}
|\left\vert u\right\vert |_{\alpha ,\beta }=\left\vert u\right\vert _{\alpha
+\beta }+\left\vert F\right\vert _{\beta }.
\end{equation*
Consider the mappings $T_{\tau }:\hat{C}^{\alpha +\beta }\left( H\right)
\rightarrow C^{\beta }(H)$ defined b
\begin{equation*}
u\left( t,x\right) =\int_{0}^{t}F\left( s,x\right) \,ds\longmapsto F-L_{\tau
}u.
\end{equation*
Obviously, for some constant $C$ not depending on $\tau ,$
\begin{equation*}
\left\vert T_{\tau }u\right\vert _{\beta }\leq C|\left\vert u\right\vert
|_{\alpha ,\beta }.
\end{equation*
On the other hand, there is a constant $C$ not depending on $\tau $ such
that for all $u\in \hat{C}^{\alpha +\beta }\left( H\right)
\begin{equation}
|\left\vert u\right\vert |_{\alpha ,\beta }\leq C\left\vert T_{\tau
}u\right\vert _{\beta }. \label{cp1}
\end{equation
Indeed,
\begin{equation*}
u\left( t,x\right) =\int_{0}^{t}F\left( s,x\right) \,ds=\int_{0}^{t}\left(
L_{\tau }u+(F-L_{\tau }u)\right) (s,x)\,ds,
\end{equation*
and, according to the estimate (\ref{form03}), there is a constant $C$ not
depending on $\tau $ such that
\begin{equation}
\left\vert u\right\vert _{\alpha +\beta }\leq C\left\vert T_{\tau
}u\right\vert _{\beta }=C\left\vert F-L_{\tau }u\right\vert _{\beta }.
\label{cp2}
\end{equation
Thus
\begin{eqnarray*}
|\left\vert u|\right\vert _{\alpha ,\beta } &=&\left\vert u\right\vert
_{\alpha +\beta }+\left\vert F\right\vert _{\beta }\leq \left\vert
u\right\vert _{\alpha +\beta }+\left\vert F-L_{\tau }u\right\vert _{\beta
}+\left\vert L_{\tau }u\right\vert _{\beta } \\
&\leq &C\left( \left\vert u\right\vert _{\alpha +\beta }+\left\vert
F-L_{\tau }u\right\vert _{\beta }\right) \leq C\left\vert F-L_{\tau
}u\right\vert _{\beta }=C\left\vert T_{\tau }u\right\vert _{\beta },
\end{eqnarray*
and (\ref{cp1}) follows. Since $T_{0}$ is an onto map, by Theorem 5.2 in
\cite{GiT83} all the $T_{\tau }$ are onto maps and the statement follows.
\section{Martingale problem}
In this section, we consider the martingale problem associated with the
operato
\begin{equation*}
L^{0}=A+B^{0}\text{,}
\end{equation*
where $B^{0}\,$\ is the operator $B$ defined by (\ref{for2}) with $\rho \geq
0$ and $l=0.$
Let $D=D([0,T],\mathbf{R}^{d})$ be the Skorokhod space of cadlag $\mathbf{R
^{d}$-valued trajectories and let $X_{t}=X_{t}(w)=w_{t},w\in D,$ be the
canonical process on it.
Le
\begin{equation*}
\mathcal{D}_{t}=\sigma (X_{s},s\leq t),\mathcal{D}=\vee _{t}\mathcal{D}_{t}
\mathbb{D=}\left( \mathcal{D}_{t+}\right) ,t\in \lbrack 0,T].
\end{equation*
We say that a probability measure $\mathbf{P}$ on $\left( D,\mathcal{D
\right) $ is a solution to the $(s,x,L)$-martingale problem (see \cite{st},
\cite{MiP923}) if $\mathbf{P}(X_{r}=x,0\leq r\leq s)=1$ and for all $u\in
C_{0}^{\infty }(H)$ the proces
\begin{equation}
u(t,X_{t})-\int_{s}^{t}[\partial _{t}u(r,X_{r})+L^{0}u(r,X_{r})]dr
\label{mart1}
\end{equation}
is a $(\mathbb{D},\mathbf{P})$-martingale. We denote $\mathcal{S}(s,x,L^{0})$
the set of all solutions to the problem $(s,x,L^{0})$-martingale problem.
\begin{lemma}
\label{ml1}let $\alpha \in (0,2),\beta \in (0,1]$ and Assumptions A, B1 and
B2 with $\rho \geq 0$ and $l=0$ be satisfied. Let $\mathbf{P\in }\mathcal{S
(s,x,L^{0}),f\in C^{\beta }(H),$ and let $u\in C^{\alpha +\beta }(H)$ be a
solution to the Cauchy proble
\begin{eqnarray}
\partial _{t}u(t,x)+L_{t}^{0}u(t,x) &=&f(t,x),(t,x)\in H, \label{mart0} \\
u(T,x) &=&0. \notag
\end{eqnarray
Then the process (\ref{mart1}) is a ($\mathbb{D},\mathbf{P})$-martingale an
\begin{equation}
u(s,x)=-\mathbf{P}_{s,x}\int_{s}^{T}f(r,X_{r})dr,(s,x)\in H. \label{mart2}
\end{equation}
\end{lemma}
\begin{proof}
Let $\zeta _{\varepsilon }$ be the function introduced in the proof of
Theorem \ref{t1} and
\begin{equation*}
u_{\varepsilon }(t,x)=u(t,\cdot )\ast \zeta _{\varepsilon }(x),(t,x)\in
H,\varepsilon \in (0,1).
\end{equation*}
Let $r<t$ and $h$ be a bounded $\mathcal{F}_{r}$-measurable random variable$.
$The
\begin{equation*}
\mathbf{P}_{s,x}\{h[u_{\varepsilon }(t,X_{t})-u_{\varepsilon
}(r,X_{r})-\int_{r}^{t}(\partial _{t}u_{\varepsilon
}(s,X_{s})+L^{0}u_{\varepsilon }(s,X_{s}))ds]\}=0.
\end{equation*
Passing to the limit in this equality as $\varepsilon \rightarrow 0$ and
using (\ref{mart0}), we ge
\begin{equation*}
\mathbf{P}_{s,x}\{h[u(t,X_{t})-u(r,X_{r})-\int_{r}^{t}f(s,X_{s}))ds]\}=0.
\end{equation*
In particular, for $t=T,r=0,h=1,$ (\ref{mart2}) follows.
\end{proof}
\begin{proposition}
\label{prop3}Let Assumptions A, B1 and B2 with $\rho \geq 0$ and $l=0$ be
satisfied. Then for each $(s,x)\in H$ there is a unique solution $\mathbf{P
_{s,x}$ to the martingale problem $(s,x,L^{0}),$ and the process $\left(
X_{t},\mathbb{D},(\mathbf{P}_{s,x})\right) $ is strong Markov.
If, in addition
\begin{equation*}
\lim_{R\rightarrow \infty }\int_{0}^{T}\sup_{x}\int_{|c(t,x,\upsilon
)|>R}\rho (t,x,\upsilon )\pi (d\upsilon )dt=0,
\end{equation*
then the function $\mathbf{P}_{s,x}$ is weakly continuous in $(s,x)$.
\end{proposition}
\begin{proof}
Since the coefficients of $L^{0}$ are H\"{o}lder continuous, it follows by
Theorem IX.2.31 in \cite{jaks} that the set $\mathcal{S}(s,x,L^{0})\neq
\emptyset .$ For $f\in C^{\beta }(H)$, let $u\,\in C_{\alpha +\beta }(H)$ be
the solution to (\ref{eq1}). By Lemma \ref{ml1}
\begin{equation*}
u(s,x)=\mathbf{P}_{s,x}\int_{s}^{T}f(r,X_{r})dr,\mathbf{P}_{s,x}\in \mathcal
S}(s,x,L).
\end{equation*
Therefore, by Lemma 2.4 \cite{MiP923}, the measure $\mathbf{P}_{s,x}\in
\mathcal{S}(s,x,L^{0})$ is unique. By Lemma 2.2 in \cite{MiP923}, the
process $\left( X_{t},\mathbb{D},(\mathbf{P}_{s,x})\right) $ is strong
Markov. The continuity of the function $(s,x)\rightarrow \mathbf{P}_{s,x}$
follows from Theorems IX.2.22 and IX.3.9 in \cite{jaks}.
\end{proof}
|
\section{Introduction}
The seesaw mechanism \cite{Yanagida:1979as} offers an exceptionally elegant explanation for the origin of neutrino masses. In its minimal incarnation---the so-called type I seesaw---the standard model is supplemented by three generations of sterile neutrinos, $\nu^c_i$, which couple via
\begin{eqnarray}
\mathcal{L} &=& \lambda_{ij}\ell_i \nu^c_j h + \frac{1}{2} M_{i} \nu^c_i \nu^c_i.
\end{eqnarray}
At energies well below the sterile neutrino masses, $M_i$, and the electroweak symmetry breaking scale, $v$, the active neutrinos acquire masses, $m_i$, which are the eigenvalues of the matrix
\begin{eqnarray}
m_{ij} &=& v^2 (\lambda M^{-1} \lambda^T)_{ij}.
\end{eqnarray}
The active neutrino masses are stringently bounded by cosmological measurements \cite{Seljak:2006bg}
\begin{eqnarray}
\sum_{i} m_i < 0.17\, \mbox{eV},
\label{eq:neutrinomassbound}
\end{eqnarray}
while experimental measurements of solar, atmospheric, accelerator and reactor neutrinos constrain the neutrino mass differences and mixing angles \cite{Nakamura:2010zzi},
\begin{eqnarray}
|\Delta m_{32}^2| &=& (2.43 \pm 0.13)\cdot 10^{-3} \textrm{ eV}^2 \nonumber \\
|\Delta m_{21}^2| &=& (7.59 \pm 0.20)\cdot 10^{-5} \textrm{ eV}^2 \nonumber \\
\sin^2 2\theta_{12} &=& 0.87 \pm 0.03 \nonumber \\
\sin^2 2 \theta_{23} &>& 0.92 \nonumber \\% for 0.44 \\
\sin^2 2\theta_{13} &<& 0.15,
\label{eq:neutrinoprops}
\end{eqnarray}
where $\Delta m^2_{ij} \equiv m_i^2 - m_j^2$.
In the high-scale seesaw, $M_i\simeq 10^{15}$ GeV, allowing for order unity Yukawa couplings,
$\lambda_{ij} \simeq \mathcal{O}(1)$. On the other hand, the sterile neutrinos can also be much lighter, $M_i\lesssim 100$ GeV, if one is willing to accommodate markedly smaller Yukawa couplings, $\lambda_{ij} \lesssim \mathcal{O}(10^{-6})$. Considering that the electron Yukawa is $\mathcal{O}(10^{-5})$, the low-scale seesaw is not terribly unreasonable. Small Yukawas notwithstanding, there is still a substantial ``seesaw effect'' at work in this regime, since the active neutrino masses are hierarchically smaller than the sterile neutrino masses. Furthermore, although thermal leptogenesis typically requires $M_i \gtrsim 10^9$ GeV, it is nevertheless possible to generate the observed baryon asymmetry in other models with $M_i$ near the weak scale \cite{Murayama:1993em, Pilaftsis:2003gt, Grossman:2004dz}.
A direct probe of the seesaw mechanism at colliders is considered impossible because the sterile neutrinos are either incredibly heavy or incredibly weakly coupled to the standard model. As a consequence, the seesaw is typically constrained by very indirect probes, for example testing the Majorana nature via measurements of neutrinoless double $\beta$ decay or lepton flavor violation, such as $\mu \rightarrow e \gamma$, in models with weak-scale supersymmetry.
In this paper, we show that the seesaw may be directly probed and in some cases {\it reconstructed} at colliders if there is weak-scale supersymmetry and the LSP is a sterile sneutrino. The superpotential is given by
\begin{eqnarray}
W &=& \lambda_{ij} L_i N^c_j H_u + \frac{1}{2} M_{i} N^c_i N^c_i ,
\end{eqnarray}
where $M_i$ is at or below the weak-scale in order to accommodate a sterile sneutrino LSP. Throughout, we assume that supersymmetry breaking in the $N^c_i$ sector is highly suppressed, as occurs in gauge mediation\footnote{In this case the sneutrino decays to the true LSP, the gravitino, but as we will see our analysis will be unaffected by this point.}, so the sterile neutrinos and sneutrinos are essentially degenerate.
Consider, for the sake of example, the case of chargino NLSP\footnote{While chargino NLSP is difficult to achieve in the MSSM \cite{Kribs:2008hq}, we consider it here for concreteness.}, which decays to a lepton and a sterile sneutrino LSP via $\tilde\chi^\pm \rightarrow \ell^\pm_i \tilde \nu^c_j$.
The chargino lifetime, along with the nine branching ratios
\begin{eqnarray} \textrm{BR}(\tilde\chi^\pm \rightarrow \ell^\pm_i \tilde \nu^c_j) \propto |\lambda_{ij}|^2
\end{eqnarray}
are in direct correspondence with the magnitudes of the nine entries of the Yukawa coupling matrix. Furthermore,
the sterile sneutrino masses are directly related to the chargino mass and the energy of the outgoing lepton in the chargino rest frame through
\begin{eqnarray}
M_{j}^2 = m_{\tilde \chi^\pm}^2 + m_{\ell^\pm_i}^2 - 2 m_{\tilde \chi^\pm} E_{\ell^\pm_i}.
\label{eq:massrelation}
\end{eqnarray}
Because $|\lambda_{ij}|$ and $M_j$ are related to physical quantities that can, in principle, be measured directly, we are left with the remarkable prospect of actually reconstructing the origin of neutrino masses at colliders. At the order of magnitude level, the seesaw may be verified by measuring
\begin{eqnarray}
\frac{\lambda^2 v_u^2}{M} \sim 0.1\, \mbox{eV},
\end{eqnarray}
in which $\lambda$ and $M$ are representative values of the Yukawa coupling matrix and the sterile sneutrino masses. As we will see, similar statements can be made for squark, slepton, and gluino NLSPs.
In what follows, we evaluate to what extent this idealized proposal might be realistically achieved at the LHC. Given existing measurements of the active neutrino masses, the decay of the NLSP to a sterile sneutrino LSP is typically displaced ($\gamma c \tau \gtrsim 1 \textrm{ mm}$) or even long-lived ($\gamma c \tau \gtrsim 1 \textrm{ m}$) at colliders. In the latter case, metastable charged or colored NLSPs may actually be produced in supersymmetric cascades and stopped in the ATLAS or CMS detector apparatus \cite{Khachatryan:2010uf, Asai:2009ka, Arvanitaki:2005nq}. Precision measurements can then be employed to determine $|\lambda_{ij}|$ from the lifetime and branching ratios of the NLSP, while $M_i$ may be ascertained from energy spectrum of the decay products. We dub this procedure seesaw spectroscopy.
Similar techniques have in the past been proposed as a sensitive probe of split supersymmetry \cite{Arvanitaki:2005nq} as well as very weakly coupled states like gravitinos \cite{Buchmuller:2004rq}, axinos \cite{Brandenburg:2005he}, goldstini \cite{Cheung:2010mc,Cheung:2010qf}, and dark matter \cite{Feng:2004mt,Cheung:2010gk}. Our analysis will address how stopped NLSPs might allow for the seesaw mechanism to be verified at collider experiments.
In Section 2 we discuss the conditions under which the low-scale seesaw might be verified or excluded given measurements of the neutrino Yukawa couplings and the sterile sneutrino masses. We go on to discuss how these parameters might be realistically obtained from collider experiments in Section 3. Finally, in Section 4 we conclude with a discussion of a remarkable version of weak scale leptogenesis---asymmetric freeze-in---in which the neutrino sector parameters relevant for the lepton asymmetry are directly accessible via seesaw spectroscopy.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.4]{decay.pdf}
\end{center}
\caption{The lifetime and branching ratios associated with the chargino decay $\tilde \chi^\pm \rightarrow \ell^\pm_i \tilde \nu^c_j$ can be used to determine $|\lambda_{ij}|$, while the lepton spectrum can be used to reconstruct $M_j$. }
\label{fig:decay}
\end{figure}
\section{Seesaw Signals}
Due to R-parity the NLSP is required to have some non-zero decay width to sterile sneutrinos via the seesaw Yukawa couplings. In this section we show how existing neutrino experiments can provide precise correlations which relate the NLSP lifetime and branching ratios to the sneutrino masses. For the moment, let us assume that these quantities can be successfully measured at colliders. We will show that this assumption is reasonable in Section 3.
The decay width of a pure charged higgsino or left-handed slepton NLSP into sterile sneutrinos is given, up to phase space factors, by
\begin{eqnarray}
\label{eq:gammas}
\Gamma(\tilde \chi^\pm \rightarrow \ell^\pm\tilde \nu^c) &\simeq&\frac{ m_{\tilde \chi^\pm}}{32\pi} \sum_{ij} |\lambda_{ij}|^2 \\
\Gamma(\tilde \ell^\pm_i \rightarrow W^\pm \tilde\nu^c) &\simeq& \frac{1}{{16\pi m_{\tilde \ell^\pm}}}\sum_{j} |\lambda_{ij} (M_j \sin \beta - \mu^* \cos \beta)|^2 \nonumber \\
\Gamma(\tilde \ell^\pm_i \rightarrow h^\pm \tilde\nu^c) &\simeq& \frac{1}{{16\pi m_{\tilde \ell^\pm}}}\sum_{j} |\lambda_{ij} (M_j \cos \beta+ \mu^* \sin \beta)|^2, \nonumber
\end{eqnarray}
where the slepton decay occurs through the supersymmetric trilinear Yukawa couplings, and $j$ implicitly sums over kinematically accessible sneutrinos. In any realistic scenario, the NLSP will of course include some admixture of states, in which case the decay width of the corresponding chargino or slepton will be dressed by the relevant mixing angles. Likewise, if the NLSP is a squark or gluino, then its decay will be mediated by an off-shell charged higgsino, and so its decay width will be suppressed by additional phase space factors and ratios of particle masses.
\subsection{Discovering the Seesaw}
The above formulae immediately allow for the following simple-minded, order of magnitude check of the neutrino seesaw at colliders. The lifetime, $\tau = \Gamma^{-1}$, and branching ratios of the NLSP yield an indirect measurement of the entries of $|\lambda_{ij}|$. Moreover, $M_i$ can in principle be measured from the energy spectrum of the NLSP decay products. For the case of pure charged higgsino NLSP,
\begin{eqnarray}
c \tau \sim 1 \textrm{ mm} \times \sin^2 \beta \left(\frac{150 \text{ GeV}}{m_\chi}\right)\left(\frac{0.1 \text{ eV}}{m}\right)\left(\frac{10 \text{ GeV}}{M}\right)
\label{eq:longtau}
\end{eqnarray}
so the decay length can be quite long. Note that for this order of magnitude confirmation we have inserted degenerate values for $\lambda_{ij} = \lambda \delta_{ij}$, $M_i = M$, and $m_i = m$ such that $\lambda^2 v_u^2 / M = m \sim 0.1 \textrm{ eV}$, which is the natural scale expected for the active neutrino masses given \Eq{eq:neutrinomassbound} and \Eq{eq:neutrinoprops}. This number is roughly the same for a pure left-handed slepton NLSP, but can be substantially greater if one includes additional mixing angles or if the NLSP is a squark or gluino. For example, gauge mediation typically results in right-handed slepton NLSPs, enhancing the slepton NLSP lifetime by a factor of $m_{\tilde{\ell}^{\pm}}^2/m_{\ell^{\pm}}^2$ once left-right mixing is included. These long lifetimes allow for stopping charged or colored NLSPs in the detector, improving the reconstructability of the sterile sneutrino masses from kinematic measurements.
Observing this order of magnitude relationship would provide substantial evidence for the seesaw as the origin of neutrino masses. That said, there are a number of important subtleties which one must address even in this simplest litmus test.
For example, the NLSP may have additional decay channels beyond the sneutrino which dominate the full width. This is true in very low-scale gauge mediated supersymmetry breaking, $\sqrt{F} \lesssim 100$ TeV, where the NLSP has a dominant decay width to gravitinos. In this scenario the overall scale of $\lambda_{ij}$ is naively inaccessible because $\tau$ is fixed by the supersymmetry breaking scale, $\sqrt{F}$, which is unrelated to the seesaw. Nonetheless, one can still measure the branching ratio of sneutrinos versus gravitinos, in which case the overall scale of $\lambda_{ij}$ can still be inferred and the above analysis applies. Alternatively, decays to gravitinos can easily be sub-dominant, requiring only $\sqrt{F} \gtrsim 100$ TeV.
The story is also more complicated if the NLSP is a slepton. Typically, in this case only a single row of $\lambda_{ij}$ is accessible from slepton NLSP decays, since the generational index of the NLSP is fixed. However, some theories possess sufficiently mass degenerate slepton co-NLSPs \cite{Giudice:1998bp}, in which case more information may be available.
In addition to the correlation between the active neutrino mass scale and the NLSP lifetime, further evidence for the seesaw may be discovered by measuring the branching ratios of NLSP decays to each of the lepton flavors. The existence of large neutrino mixing angles suggests that one should measure roughly equal parts $e$, $\mu$, and $\tau$ arising from each sneutrino in the NLSP decays.
\subsection{Probing the Seesaw}
Checking the order of magnitude validity of \Eq{eq:longtau} would be an important first step towards performing precision measurements of the seesaw mechanism at colliders. Of course, this coarse analysis disregards all of the flavor information which can be extracted from a measurement of the full $|\lambda_{ij}|$ matrix and the sneutrino masses $M_{i}$. Indeed, all seesaw parameters are accessible to experiment except for 1) the phases of the Yukawa couplings, $\arg \lambda_{ij}$, and 2) any Yukawa couplings or masses related to kinematically inaccessible sneutrinos. In the following section, we discuss how additional flavor information can provide essential data about the neutrino seesaw.
The measured values of $|\lambda_{ij}|$ and $M_i$ can be combined to form an ``effective'' neutrino mass matrix
\begin{eqnarray}
\hat m_{ij} &=& v_u^2 \sum_k |\lambda_{ik}| \frac{1}{M_k} |\lambda^T_{kj}| ,
\label{eq:mhat}
\end{eqnarray}
where $k$ sums over all sneutrinos lighter than the NLSP and $\tan\beta$ is assumed to be measured. Clearly, $\hat m_{ij}$ is missing all information in the theory involving phases or kinematically inaccessible sneutrinos. In general, $\hat m_{ij} \neq m_{ij}$, the actual neutrino mass matrix, and consequently the neutrino masses and mixing angles derived from $\hat m_{ij}$ will not match the observations of existing neutrino experiments shown in \Eq{eq:neutrinomassbound} and \Eq{eq:neutrinoprops}. However, in many cases the eigenvalues and mixing angles associated with $\hat m_{ij}$ may be found to match $m_{ij}$. If this occurs, then the information provided in $\hat m_{ij}$ can provide new and substantive information about the neutrino sector.
For example, if all three sneutrinos are kinematically accessible from the NLSP decay and one discovers that $\hat m_{ij} = m_{ij}$ identically, then this would indicate that the CP phases in the neutrino sector are negligible. This would be strong evidence that the neutrino Yukawa couplings respect CP symmetry. If this were true, then colliders could provide unambiguous information about as of yet unknown quantities like $\theta_{13}$ or the nature of the active neutrino mass hierarchy!
Alternatively, suppose that the heaviest active neutrino mass is generated by integrating out the lightest sterile neutrino, $\nu^c_1$. If the lightest sneutrino is the only sterile neutrino accessible in NLSP decay, so that only $k=1$ contributes in \Eq{eq:mhat}, then the NLSP lifetime will be precisely correlated with $m_{\mbox{atm}} = \sqrt{|\Delta m_{32}^2|}$. Furthermore, the ratio of $\tau:\mu:e$ events will be $0.5:0.5:\theta_{13}^2$. In this case, the phases in the neutrino Yukawa couplings may be rotated into the elements corresponding to the heavier sneutrinos, so that $\hat m_{ij} \approx m_{ij}$, up to corrections of $\mathcal{O}(m_{\mbox{sol}}/m_{\mbox{atm}})$ where $m_{\mbox{sol}} = \sqrt{|\Delta m_{21}^2|}$. Such measurements would convincingly demonstrate the weak-scale seesaw origin of neutrino masses and provide a measurement of the presently unknown angle $\theta_{13}$.
\subsection{Excluding the Seesaw}
Collider measurements can be used to exclude the seesaw as well. For instance, consider a scenario in which the NLSP can decay to all three sterile sneutrinos. For concreteness, we assume a pure charged higgsino NLSP, although our discussion is easily generalized to arbitrary NLSP. The decay width may be rewritten as
\begin{eqnarray}
\Gamma(\tilde\chi^\pm \rightarrow \ell^\pm \tilde \nu^c)
&=& \frac{ m_{\tilde \chi^\pm}}{32\pi} \sum_{i}\lambda_{i}^2,
\label{eq:partwidth}
\end{eqnarray}
where $\lambda_i$ are the real and positive semi-definite diagonal entries of the diagonalized Yukawa interaction matrix, $(L \lambda R^\dagger)_{ij}$.
The active neutrino masses are given by
\begin{eqnarray}
\sum_i m_i^2 &=& v_u^4 \sum_{ijkl} \lambda_i^2 R^\dagger_{ij} \frac{1}{M_j} R^*_{jk} \lambda^2_k R^T_{kl} \frac{1}{M_l} R_{li} \\
&\leq& v_u^4 \sum_{ijkl} \lambda^2_i \frac{1}{M_j} \lambda^2_k \frac{1}{M_l} \\
&=& v_u^4 (\sum_i \lambda_i^2)^2 (\sum_j M^{-1}_j)^2,
\end{eqnarray}
where the inequality in the second line arises from the fact that every entry of a unitary matrix is bounded by unity, so $|R_{ij}|<1$. Using this expression in \Eq{eq:partwidth} yields
\begin{eqnarray}
\frac{32 \pi \Gamma}{m_{\tilde \chi^\pm}} &\geq& \frac{ \sqrt{\sum_i m_i^2} (\sum_j M_j^{-1})^{-1}}{v_u^2}.
\end{eqnarray}
Since the sum of active neutrino masses is bounded from below by measurements of $|\Delta m_{32}^2|$ and the partial decay width of the chargino NLSP to sterile sneutrinos in \Eq{eq:partwidth} is constrained from above by the inverse lifetime, $\tau^{-1}$, we obtain our final bound,
\begin{eqnarray}
(\sum_i M_i^{-1})^{-1} &\leq& 81 \textrm{ GeV} \times\sin^2 \beta \times
\left(\frac{200 \textrm{ GeV}}{m_{\tilde \chi^\pm}}\right) \left(\frac{1 \textrm{ mm}}{c\tau}\right). \nonumber
\label{eq:taubound}
\end{eqnarray}
If the sneutrino masses are successfully measured and this inequality fails, then the physics being probed is not the origin of neutrino masses. Analogous bounds to \Eq{eq:taubound} can of course be derived for alternative NLSP candidates.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=1]{2body.pdf}
\end{center}
\caption{Histogram of the lepton energy from $\tilde \chi^\pm \rightarrow \ell^\pm_i \tilde \nu_j^c$ for $m_{\tilde \chi^\pm} =$ 150 GeV. The \{purple, red, yellow\} peaks correspond to the sterile sneutrinos at masses $M_j =$ \{50, 80, 100\} GeV and the arrows correspond to the values predicted by \Eq{eq:massrelation}. We have taken $\delta E_{\ell^\pm_i} = 0.004 E_{\ell^\pm_i}$.
}
\label{fig:2body}
\end{figure}
\section{Seesaw Spectroscopy }
In this section we evaluate to what extent the NLSP lifetimes, branching ratios, and sneutrino masses might be realistically measured at the LHC.
Assuming that the dominant decay channel of the NLSP is to sterile sneutrinos, then \Eq{eq:longtau} implies that the NLSP can easily be long-lived when one includes various mixing angles and kinematic factors. If the NLSP is charged or colored, then some fraction of them will be stopped inside the ATLAS or CMS detectors.
It has been shown \cite{Arvanitaki:2005nq} that upwards of $10^6$ long-lived, colored particles may be stopped each year at the LHC. Indeed, searches for such particles have already been performed \cite{Khachatryan:2010uf}, using dedicated triggers which detect decays occurring out of time with the beam. The aforementioned analysis considered the dominant interactions between the NLSP and the detector to be electromagnetic in nature, with the NLSP energy loss described by the Bethe-Bloch equation, and so it is equally applicable to both slepton NLSPs and squarks or gluinos which hadronize into charged particles. Since these squarks or gluinos must decay through an intermediate chargino, they will decay to a three-body final state, extending their lifetime by a three-body phase space factor and an off-shell propagator. As discussed above, slepton and chargino NLSPs will typically have their lifetimes extended by left-right mixing and higgsino-wino mixing, respectively. Thus, we expect NLSP lifetimes to be sufficiently long in many cases to produce substantial numbers of stopped particles.
The exact number of stopped NLSPs is strongly dependent on the lifetime of the NLSP, which is in turn fixed by the nature of the NLSP and the mixing angles in the supersymmetric spectrum. In order to sidestep this model dependence, we will divide the remainder of our discussion according to whether a sizeable fraction of NLSPs is stopped or not.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=.8]{3body.pdf}
\end{center}
\caption{Scatter plot of the quark and lepton energies from $\tilde q \rightarrow q \ell^\pm_i \tilde \nu_j^c$ for $m_{\tilde q} =$ 700 GeV. The \{purple, red, yellow\} points correspond to the sterile sneutrinos at masses $M_j =$ \{250, 450, 550\} GeV and the dashed lines correspond to endpoints analytically determined by the kinematics. We have taken $\delta E_q = 0.033 E_q$ and $\delta E_{\ell^\pm_i} = 0.004 E_{\ell^\pm_i}$.
}
\label{fig:3body}
\end{figure}
\subsection{Stopped NLSPs}
A large number of NLSPs will be stopped at LHC if the NLSP lifetime is sufficiently long. In this case the ensuing decays occur in the rest frame of the detector, and so the initial four-vector of the NLSP is known, assuming the NLSP mass has been measured by other means. Additionally, since decays from stopped particles occur out of time with respect to the main collision events, we expect signals from such decays to be exceptionally clean.
If the NLSP decay is two-body then the visible decay product will be monochromatic with an energy uniquely fixed by the NLSP mass and the sneutrino mass, as shown for the case of chargino NLSP in \Eq{eq:massrelation}. In this case each sneutrino decay product will produce a separate peak in the energy spectrum of the outgoing lepton.
In many cases it should be easy to distinguish the peaks corresponding to each sneutrino, even assuming uncertainties in the energy measurement of the outgoing charged lepton, gauge boson, or Higgs. \Fig{fig:2body} shows a histogram of the energy spectrum of the outgoing lepton\footnote{Representative values for the energy uncertainties in \Fig{fig:2body} and \Fig{fig:3body} are taken from the ATLAS TDR.} arising from the decay of a chargino NLSP, $\tilde \chi^\pm \rightarrow \ell^\pm_i \tilde \nu^c_j$. By measuring the central value of each peak, one can reconstruct the mass of each corresponding sneutrino. Likewise, by counting the total number of events within each peak and tagging the flavor of $\ell_i^\pm$, one obtains the branching ratio $\textrm{BR}(\tilde \chi^\pm \rightarrow \ell^\pm_i \tilde \nu^c_j)$, and hence the value of each entry of $|\lambda_{ij}|$.
Due to the small uncertainty in the lepton energy resolution, the dominant error in measuring the sneutrino mass by this method comes from uncertainty in the NLSP mass, $\delta M_j \approx \delta m_{\tilde{\chi}^{\pm}}(m_{\tilde{\chi}^{\pm}} - E_{\ell^{\pm}_i})/M_j$. Additionally, cosmic rays provide a significant background to these events. As a result, current searches veto on reconstructed muons in the final state; however, for the purposes of reconstructing the seesaw, it will be necessary to distinguish cosmic ray muons from muonic NLSP decays.
If the NLSP decays to a three-body final state, then the analysis is slightly more involved. For example, consider the case of a three-body decay of a squark NLSP, $\tilde q \rightarrow q \ell_i^\pm \tilde \nu^c_j$, via an off-shell chargino. While the quark and lepton energies are not monochromatic, they are constrained by kinematic endpoints dictated by the mass of the squark and sneutrino. For example, \Fig{fig:3body} depicts a scatter plot of quark and lepton energies from the squark NLSP decay. As expected, one obtains three distinct bands corresponding to the decay of the squark to each sneutrino. As in the two-body case, it should be possible to distinguish between these bands for a broad range of theories. If this is the case, then the shape and location of each band can be used to determine the mass of the corresponding sneutrino, and the number of events in each band along with the flavor of the outgoing lepton can provide the associated branching ratio, $\textrm{BR}(\tilde q \rightarrow q \ell^\pm_i \tilde \nu^c_j)$. Note that even in the case where the bands overlap, it may still be possible to distinguish them since there is additional information in the number of events in each band. For example, for a constant matrix element, the bands will appear as three distinct plateaus, each sitting atop the next.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.75]{edge.pdf}
\end{center}
\caption{Histogram of the invariant mass of the outgoing quark and lepton from $\tilde q \rightarrow q \tilde\chi^\pm \rightarrow q \ell^\pm_i \tilde \nu_j^c$ for $m_{\tilde q} =$ 600 GeV and $m_{\tilde \chi^\pm}=$ 200 GeV. The edges correspond to sterile sneutrinos at masses $M_j =$ \{20, 100, 160\} GeV.}
\label{fig:edge}
\end{figure}
\subsection{Non-Stopped NLSPs}
If the NLSP is too short-lived to be efficiently stopped, then it will typically be moving when it decays. In this scenario the initial four-vector of the NLSP is unknown, so event reconstruction is a greater challenge. Nonetheless, if the NLSP arises from a cascade decay, then kinematical information from the entire cascade can be used to reconstruct the sneutrino masses.
For example, for a squark to chargino to sneutrino cascade decay, $\tilde q \rightarrow q \tilde\chi^\pm \rightarrow q \ell^\pm_i \tilde \nu_j^c$, there is a distinct edge in the invariant mass constructed from the quark and lepton which occurs at
\begin{eqnarray}
m_{q\ell^\pm_i}^2 &=& \frac{(m_{\tilde q}^2-m_{\tilde \chi^\pm}^2)(m_{\tilde \chi^\pm}^2-M_j^2)}{m_{\tilde \chi^\pm}^2}.
\end{eqnarray}
As shown in \Fig{fig:edge}, each sneutrino edge may be visible in the $m_{q\ell^\pm_i}$ spectrum and it may be possible to reconstruct the branching ratios of the NLSP as well as the masses of each sneutrino.
Nevertheless, there is some subtlety required in measuring the $m_{q\ell^\pm_i}$ spectrum. Since each chargino decay is displaced, one can efficiently identify the associated leptons by selecting for hard leptons displaced from the primary vertex. On the other hand, there are intrinsic combinatoric jet backgrounds which remain even after cutting on soft jets in the event. If the NLSP decays are truly two-body, it may be possible to mitigate problems associated with jet combinatorics by considering only the lepton spectrum and the missing transverse momentum. Techniques based on the $M_{T2}$ kinematic variable \cite{Konar:2009wn} could provide a measurement of the sneutrino mass independent of both the jet kinematics and the NLSP mass in cases where the NLSP is pair produced. Note that because these measurements do not require stopped NLSPs, they apply to the left-handed sneutrino NLSP as well as charged and colored NLSPs.
\section{Conclusion}
In this paper we have discussed the possibility of seesaw spectroscopy at colliders: a procedure whereby detailed information concerning the masses and mixing angles of the neutrino sector might be directly extracted from the decays of long-lived NLSPs. Given the robustness of this probe, this proposal naturally begs the question: might seesaw spectroscopy be used as a tool to study the origin of the cosmological baryon asymmetry via leptogenesis?
Such an ambitious endeavor is not possible in many theories, including standard high-scale thermal leptogenesis, since $M_i \gtrsim 10^9$ GeV \cite{Davidson:2002qv}. Soft leptogenesis via CP violating mixing of sterile sneutrinos allows significantly lower values of $M_i$, but still very much larger than the weak scale \cite{Grossman:2003jv,D'Ambrosio:2003wy}. Soft leptogenesis with $M_i$ of order the weak scale is possible via CP violating decays of sterile sneutrinos \cite{Grossman:2004dz}, and weak scale leptogenesis can also result from the decay of sterile sneutrino condensate formed during inflation \cite{Murayama:1993em}. However, in both these cases the LSP is a standard model superpartner and not a sterile sneutrino, so there can be no seesaw spectroscopy signals of long-lived NLSP decays to sterile sneutrinos. Resonant leptogenesis can occur at the weak scale \cite{Pilaftsis:2003gt} but the size of the asymmetry is linked to the very high degeneracy of the $M_i$ which again cannot be probed via seesaw spectroscopy, even if the theory is supersymmetric.
On the other hand, a reconstructable theory of leptogenesis is possible at weak-scale temperatures if the NLSP decays to LSP sterile sneutrinos. This asymmetric ``freeze-in" mechanism is more or less identical to the one investigated in \cite{Hall:2010jx} using the superpotential interaction $L_i X H_u$, with $X$ re-interpreted as $N^c_j$. The lepton asymmetry produced in decays of superpartners to final states involving $\ell_i \tilde{\nu}^c_j$ is proportional to $|\lambda_{ij}|^2 \sin \phi$, where the CP phase $\phi$ arises from soft supersymmetry breaking interactions which might be probed indirectly through electric dipole moments. As a consequence of \Eq{eq:gammas}, the parameter $|\lambda_{ij}|$ can actually be experimentally accessed via seesaw spectroscopy.
We leave a more thorough investigation of this remarkable scheme of low-scale leptogenesis for future work \cite{CHP}.
|
\section{Introduction}
Complexity is a tremendous challenge in the field of marine ecology. It impacts the development of theory, the conduct of field studies, and the practical application of ecological knowledge. It is encountered at all scales (Loehle, 2004). There is growing concern over the excessive and unsustainable exploitation of marine resources due to overfishing, climate change, petroleum development, long-range pollution, radioactive contamination, aquaculture etc. Ecologists, marine biologists and now economists are searching for possible solutions to the problem (Kar and Matsuda, 2007). A qualitative and quantitative study of the interaction of different species is important for the management of fisheries. From the study of several fisheries models, it was observed that the type of functional response greatly affect the model predictions (Steele and Henderson, 1992; Gao et al., 2000; Fu et al., 2001; Upadhyay et al., 2009). An inappropriate selection of functional response alters quantitative predications of the model results in wrong conclusions. Mathematical models of plankton dynamics are sensitive to the choice of the type of predator's functional response, i.e., how the rate of intake of food varies with the food density. Based on real data obtained from expeditions in the Barents Sea in 2003-2005, show that the rate of average intake of algae by Calanus glacialis exhibits a Holling type III functional response, instead of responses of Holling types I and II found previously in the laboratory experiments. Thus the type of feeding response for a zooplankton obtained from laboratory analysis can be too simplistic and misleading (Morozov et al., 2008). Holling type III functional response is used when one wishes to stabilize the system at low algal density (Truscott and Brindley, 1994; Scheffer and De Boer, 1996; Bazykin, 1998; Hammer and Pitchford, 2005).
Modeling of phytoplankton-zooplankton interaction takes into account zooplankton grazing with saturating functional response to phytoplankton abundance called Michaelis-Menten models of enzyme kinetics (Michaelis and Menten, 1913). These models can explain the phytoplankton and zooplankton oscillations and monotonous relaxation to one of the possible multiple equilibria (Steele and Henderson, 1981, 1992; Scheffer, 1998; Malchow, 1993; Pascual, 1993; Truscott and Brindley, 1994). The problems of spatial and spatiotemporal pattern formation in plankton include regular and irregular oscillations, propagating fronts, spiral waves, pulses and stationary spiral patterns. Some significant contributions are dynamical stabilization of unstable equilibria (Petrovskii and Malchow, 2000; Malchow and Petrovskii, 2002) and chaotic oscillations behind propagating diffusive fronts in prey-predator models with finite or slightly inhomogeneous distributions (Sherratt et al., 1995, 1997; Petrovskii and Malchow, 2001).
Conceptual prey-predator models have often and successfully been used to model phytoplankton-zooplankton interactions and to elucidate mechanisms of spatiotemporal pattern formation like patchiness and blooming (Segel and Jackson, 1972; Steele and Hunderson, 1981; Pascual, 1993; Malchow, 1993). The density of plankton population changes not only in time but also in space. The highly inhomogeneous spatial distribution of plankton in the natural aquatic system called ``Plankton patchiness" has been observed in many field observations (Fasham, 1978; Steele, 1978; Mackas and Boyd, 1979; Greene et al., 1992; Abbott, 1993). Patchiness is affected by many factors like temperature, nutrients and turbulence, which depend on the spatial scale. Generally the growth, competition, grazing and propagation of plankton population can be modeled by partial differential equations of reaction-diffusion type. The fundamental importance of spatial and spatiotemporal pattern formation in biology is self-evident. A wide variety of spatial and temporal patterns inhibiting dispersal are present in many real ecological systems. The spatiotemporal self-organization in prey-predator communities modelled by reaction-diffusion equations have always been an area of interest for researchers. Turing spatial patterns have been observed in computer simulations of interaction-diffusion system by many authors (Malchow, 1996, 2000; Xiao et al., 2006; Brentnall et al., 2003; Grieco et al., 2005; Medvinsky et al., 2001; 2005; Chen and Wang, 2008; Liu et al., 2008). Upadhyay et al. (2009, 2010) investigated the wave phenomena and nonlinear non-equilibrium pattern formation in a spatial plankton-fish system with Holling type II and IV functional responses.
In this paper, we have tried to find out the analytical solution to plankton pattern formation which is in fact necessity for understanding the complex dynamics of a spatial plankton system with Holling type III functional response. And the analytical finding is supported by extensive numerical simulation.
\vspace*{0.5cm}
\setcounter{equation}{0}
\section{ The model system}
We consider a reaction-diffusion model for phytoplankton-zooplankton-fish system where at any point $(x, y)$ and time $t$, the phytoplankton $P(x, y, t)$ and zooplankton $H(x, y, t)$ populations. The phytoplankton population $P(x, y, t)$ is predated by the zooplankton population $H(x, y, t)$ which is predated by fish. The per capita predation rate is described by Holling type III functional response. The system incorporating effects of fish predation satisfy the following:
\begin{equation}\label{21}
\begin{array}{l}
\frac{\partial P}{\partial t}=rP-B_1P^2-\frac{B_2P^2H}{P^2+D^2}+d_1\nabla^2P,\\[6pt]
\frac{\partial H}{\partial t}=C_1H-C_2\frac{H^2}{P}-\frac{FH^2}{H^2+D_1^2}+d_2\nabla^2H
\end{array}
\end{equation}
with the non-zero initial conditions
\begin{equation}\label{22}
\begin{array}{l}
P(x,y,0)>0,\,\, H(x,y,0)>0,\qquad (x,y)\in\Omega=[0, R]\times[0, R],
\end{array}
\end{equation}
and the zero-flux boundary conditions
\begin{equation}\label{23}
\frac{\partial P}{\partial n}=\frac{\partial H}{\partial n}=0,\,
(x,y)\in\partial\Omega\quad \text{for all}\quad t.
\end{equation}
where, $n$ is the outward unit normal vector of the
boundary $\partial \Omega$ which we will assume is smooth.
And $\nabla^2$ is the Laplacian operator, in two-dimensional space, $\nabla^2=\frac{\partial^2}{\partial x^2}+\frac{\partial^2}{\partial y^2}$, while in one-dimensional case, $\nabla^2=\frac{\partial^2}{\partial x^2}$.
The parameters $r, B_1, B_2, D, C_1, C_2, D_1, d_1, d_2$ in model~\eqref{21} are positive constants. We explain the meaning of each variable and constant. $r$ is the prey's intrinsic growth rate in the absence of predation, $B_1$ is the intensity of competition among individuals of phytoplankton, $B_2$ is the rate at which phytoplankton is grazed by zooplankton and it follows Holling type--III functional response, $C_1$ is the predator's intrinsic rate of population growth, $C_2$ indicates the number of prey necessary to support and replace each individual predator. The rate equation for the zooplankton population is the logistic growth with carrying capacity proportional to phytoplankton density, $P/C_2$. $D, D_1$ is the half-saturation constants for phytoplankton and zooplankton density respectively, $F$ is the maximum value of the total loss of zooplankton due to fish predation, which also follows Holling type-III functional response. $d_1, d_2$ diffusion coefficients of phytoplankton and zooplankton respectively. The units of the parameters are as follows. Time $t$ and length $x, y\in[0, R]$ are measured in days $[d]$ and meters $[m]$ respectively. $r, P, H, D$ and $D_1$ are usually measured in $mg$ of dry weight per litre [$mg\cdot dw/l$]; the dimension of $B_2$, and $C_2$ is [$d^{-1}$], are measured in [$(mg\cdot dw/l)^{-1}d^{-1}$] and [$d^{-1}$], respectively. The diffusion coefficients $d_1$ and $d_2$ are measured in [$m^{2}d^{-1}$]. $F$ is measured in [$(mg\cdot dw/l)d^{-1}$]. The significance of the terms on the right hand side of Eq.(1) is explained as follows: The first term represents the density-dependent growth of prey in the absence predators. The third term on the right of the prey equation, $B_2P^2/(P^2+D^2)$, represents the action of the zooplankton. We assume that the predation rate follows a sigmoid (Type III) functional response. This assumption is suitable for plankton community where spatial mixing occur due to turbulence (Okobo, 1980). The dimensions and other values of the parameters are chosen from literatures (Malchow, 1996; Medvinsky et al., 2001, 2002 Murray, 1989), which are well established for a long time to explain the phytoplankton-zooplankton dynamics. Notably, the system (1) is a modified (Holling III instead of II) and extended (fish predation) version of a model proposed by May (1973).
Holling type III functional response have often been used to demonstrate cyclic
collapses which form is an obvious choice for representing the behavior of predator
hunting (Real, 1977; Ludwig et al., 1978). This response function is sigmoid, rising
slowly when prey are rare, accelerating when they become more abundant, and finally
reaching a saturated upper limit. Type III functional response also levels off at some prey
density. Keeping the above mentioned properties in mind, we have considered the
zooplankton grazing rate on phytoplankton and the zooplankton predation by fish follows
a sigmoidal functional response of Holling type III.
\vspace*{0.5cm}
\setcounter{equation}{0}
\section{Stability analysis of the non-spatial model system}
In this section, we restrict ourselves to the stability analysis of the model system in the
absence of diffusion in which only the interaction part of the model system are taken into account. We find the non-negative equilibrium states of the model system and discuss
their stability properties with respect to variation of several parameters.
\subsection{Local stability analysis}
We analyze model system~\eqref{21} without diffusion. In such case, the model system reduces to
\begin{equation}\label{31}
\begin{array}{l}
\frac{\partial P}{\partial t}=rP-B_1P^2-\frac{B_2P^2H}{P^2+D^2},\\[6pt]
\frac{\partial H}{\partial t}=C_1H-C_2\frac{H^2}{P}-\frac{FH^2}{H^2+D_1^2}
\end{array}
\end{equation}
with $P(x, 0)>0, H(x, 0)>0$.
The stationary dynamics of system~\eqref{31} can be analyzed from $dP/dt = 0, dH/dt=0$.
Then, we have
\begin{equation}\label{32}
\begin{array}{l}
rP-B_1P^2-\frac{B_2P^2H}{P^2+D^2}=0,\\[6pt]
C_1H-C_2\frac{H^2}{P}-\frac{FH^2}{H^2+D_1^2}=0.
\end{array}
\end{equation}
An application of linear stability analysis determines the stability of the two stationary
points, namely $E_1(\frac{r}{B_1}, 0)$ and $E^*(P^*, H^*)$.
Notably, $E^*(P^*, H^*)$ presents the kind of nontrivial coexistence equilibria, not only
mean one equilibrium. In fact, it may be noted that $P^*$ and $H^*$ satisfies the following
equations:
$$H^*-\frac{(r-B_2P^*)(P^{*2}+D^2)}{B_2P^*}=0,\, \frac{FH^*}{H^{*2}+D_1^2}+C_2\frac{H^*}{P^*}-C_1=0.$$
After solving $P^*$ from the second equation above and substituting it into the first
equation, we can obtain a polynomial equation about $H^*$ with degree nine; therefore the
equation at least has one real root. That is to say, system~\eqref{31} may be having one or more positive equilibria, namely $E^*(P^*, H^*)$.
From Eqs.~\eqref{31}, it is easy to show that the equilibrium point $E_1$ is a saddle point with stable manifold in $P-$direction and unstable manifold in $H-$direction. The stationary point
$E^*(P^*, H^*)$ is stable or unstable depend on the coexistence of phytoplankton and
zooplankton. In the following theorem we are able to find necessary and sufficient
conditions for the positive equilibrium $E^*(P^*, H^*)$ to be locally asymptotically stable.
The Jacobian matrix of model~\eqref{31} at $E^*(P^*, H^*)$ is
$$J=\left(\begin{array}{cc}P^*\left(-B_1+B_2H^*f_1(P^*, D)\right) &
-\frac{B_2P^{*2}}{P^{*2}+D^2} \\
\frac{C_2H^{*2}}{P^{*2}}&H^*\left(-\frac{C_2}{P^*}+Ff_2(H^*, D_1)\right)\end{array}
\right),$$
where,
$$f_1(P^*, D)=\frac{P^{*2}-D^2}{(P^{*2}+D^2)^2},\quad f_2(H^*, D)=\frac{H^{*2}-D_1^2}{(H^{*2}+D_1^2)^2}.$$
Following the Routh-Hurwitz criteria, we get,
\begin{equation}\label{33}
\begin{array}{l}
A=P^*(B_1-B_2H^*f_1(P^*, D))+H^*(\frac{C_2}{P^*}-Ff_2(H^*, D_1)),\\[6pt]
B=P^*H^*(B_1-B_2H^*f_1(P^*, D))\left(\frac{C_2}{P^*}-Ff_2(H^*, D_1)\right)+\frac{B_2C_2H^{*2}}{(P^{*2}+D^2)^2}.
\end{array}
\end{equation}
Summarizing the above discussions, we can obtain the following theorem.
\vspace*{0.25cm}
\begin{theorem}
The positive equilibrium $E^*(P^*, H^*)$ is locally asymptotically stable in the
$PH-$plane if and only if the following inequalities hold:
$$A > 0 \quad \text{and}\quad B > 0.$$
\end{theorem}
\vspace*{0.25cm}
\begin{remark}
Let the following inequalities hold
\begin{equation}\label{34}
\begin{array}{l}
B_1>B_2H^*f_1(P^*, D),\quad
C_2>FP^*f_2(H^*, D_1).
\end{array}
\end{equation}
Then $A>0$ and $B>0$. It shows that if inequalities in Eq.~\eqref{34} hold, then $E^*(P^*, H^*)$ is locally asymptotically stable in the $PH-$plane.
\end{remark}
Fig.1 shows the dynamics of system~\eqref{31} that for increasing fish predation rates at $F=0$
(limit cycle---no fish), $F=0.02$, 0.03, 0.04 (limit cycle) and $F=0.044$ (phytoplankton
dominance). We observe the limit cycle behavior for the fish predation rate $F$ in the
range $0\leq F\leq 0.043$, and for higher values phytoplankton dominance is reported for a
non-spatial system~\eqref{31}.
\begin{figure}[htbp]
\centerline{\includegraphics[scale=0.8]{fig1.eps}}
\caption{(a) predator-prey zero-isoclines (b) prey zero-isoclines for different value of $F$ with parameters value $r=1$, $B_1=0.2$, $B_2=0.91$, $C_1=0.22, C_2=0.2$, $D^2=0.3$, $D_1=0.1$.}
\label{figure1}
\end{figure}
\subsection{Global stability analysis}
In order to study the global behavior of the positive equilibrium $E^*(P^*, H^*)$, we need
the following lemma which establishes a region for attraction for model system~\eqref{31}.
\vspace*{0.25cm}
\noindent{\bf Lemma 1.}
The set $\mathcal{R}=\{(P, H): 0\leq P\leq \frac{r}{B_1}, 0\leq H\leq\frac{rC_1}{C_2B_1}\}$ is a region of attraction for all solutions initiating in the interior of the positive quadrant $\mathbb{R}$.
\vspace*{0.25cm}
\begin{theorem}
If
\begin{equation}\label{35}
P^*(r-B_1P^*)^2<4B_1rD^2,\quad rC_1H^*<D_1^2C_2B_1
\end{equation}
hold, then $E^*$ is globally asymptotically stable with respect to all solutions in the interior of the positive quadrant $\mathbb{R}$.
\end{theorem}
\noindent {\bf Proof.} Let us choose a Lyapunov function
\begin{equation}\label{Lyap}V(P, H)=\int_{P^*}^P\frac{\xi-P^*}{\xi\varphi(\xi)}d\xi+w\int_{H^*}^H\frac{\eta-H^*}{\eta}d\eta,
\end{equation}
where $\varphi(P)=\frac{B_2P^2}{P^2+D^2}$.
Differentiating both sides with respect to $t$, we get
$$\frac{dV}{dt}(P, H)=\frac{P-P^*}{P\varphi(P)}\frac{dP}{dt}+w\frac{H-H^*}{H}\frac{dH}{dt},$$
Substituting the expressions of
$\frac{dP}{dt}$ and $\frac{dH}{dt}$
from Eq.~\eqref{31} and putting
$w=\frac{P^*}{C_2H^*}$, we obtained:
$$\begin{array}{l}
\frac{dV}{dt}=\frac{P-P^*}{P}\Bigl(\frac{rP-B_1P^2}{\varphi(P)}-H^*\Bigr)-(H-H^*)^2\frac{P^*}{H^*P}
-w(H-H^*)^2\frac{F(D_1^2-H^*H)}{(H^2+D_1^2)(H^{*2}+D_1^2)}\\[8pt]
\quad\,\, =\frac{(P-P^*)^2}{BP^2P^*}\Bigl(rP^*P-rD^2-B_1PP^*(P^*+P)\Bigr)-(H-H^*)^2\frac{P^*}{H*P}\\[8pt]
\qquad\,\,\,\, -w(H-H^*)^2\frac{F(D_1^2-H^*H)}{(H^2+D_1^2)(H^{*2}+D_1^2)}
\end{array}$$
then $\frac{dV}{dt}<0$
for all $P, H\in\mathcal{R}$ if the condition~\eqref{35} hold.
\vspace*{0.5cm}
\setcounter{equation}{0}
\section{Stability analysis of the spatial model system}
In this section, we study the effect of diffusion on the model system about the interior
equilibrium point. Complex marine ecosystems exhibit patterns that are bound to each
other yet observed over different spatial and time scales (Grimm et al., 2005). Turing instability can occur for the model system because the equation for predator is nonlinear
with respect to predator population, $H$ and unequal values of diffusive constants.
To study the effect of diffusion on the system~\eqref{31}, we derive conditions for stability
analysis in one and two dimensional cases.
\subsection{One dimensional case}
The model system~\eqref{21} in the presence of one-dimensional diffusion has the following
form:
\begin{equation}\label{41}
\begin{array}{l}
\frac{\partial P}{\partial t}=rP-B_1P^2-\frac{B_2P^2H}{P^2+D^2}+d_1\frac{\partial^2P}{\partial x^2},\\[6pt]
\frac{\partial H}{\partial t}=C_1H-C_2\frac{H^2}{P}-\frac{FH^2}{H^2+D_1^2}+d_2\frac{\partial^2H}{\partial x^2}.
\end{array}
\end{equation}
To study the effect of diffusion on the model system, we have considered the
linearized form of system about $E^*(P^*, H^*)$ as follows:
\begin{equation}\label{42}
\begin{array}{l}
\frac{\partial U}{\partial t}=b_{11}U+b_{12}V+d_1\frac{\partial^2U}{\partial x^2},\\[8pt]
\frac{\partial V}{\partial t}=b_{21}U+b_{22}V+d_2\frac{\partial^2V}{\partial x^2},
\end{array}
\end{equation}
where $P=P^*+U, H=H^*+V$ and
$$\begin{array}{l}
b_{11}=-P^*(B_1-B_2H^*f_1(P^*, D)),\quad b_{12}=-B_2P^*/(P^{*2}+D^2),\\[6pt]
b_{21}=C_2H^{*2}/P^{*2},\qquad\qquad\qquad\quad\,\,\,\,\, b_{22}=H^*(Ff_2(H^*, D_1)-C_2/P^*).
\end{array}$$
It may be noted that $(U, V)$ are small perturbations of $(P, H)$ about the equilibrium point
$E^*(P^*, H^*)$.
In this case, we look for eigenfunctions of the form
$$\sum_{n=0}^{\infty}{a_n\choose b_n}\exp{(\lambda t+ikx)},$$
and thus solutions of system~\eqref{42} of the form
$${U\choose V}=\sum_{n=0}^{\infty}{a_n\choose b_n}\exp{(\lambda t+ikx)},$$
where $\lambda$ and $k$ are the frequency and wave-number respectively. The characteristic
equation of the linearized system~\eqref{42} is given by
\begin{equation}\label{43}
\lambda^2+\rho_1\lambda+\rho_2=0,
\end{equation}
where
\begin{equation}\label{44}
\rho_1=A+(d_1+d_2)k^2,
\end{equation}
\begin{equation}\label{45}
\rho_2=B+d_1d_2k^4+\Bigl[d_2P^*(B_1-B_2H^*f_1(P^*, D))+d_1H^*(C_2/P^*-Ff_2(H^*, D_1))\Bigr]k_2,
\end{equation}
where $A$ and $B$ are defined in Eq.~\eqref{33}. From Eqs.~\eqref{43}---\eqref{45}, and using Routh-Hurwitz criteria, we can know that the positive equilibrium
$E^*$ is locally asymptotically stable in the presence of diffusion if and only if
\begin{equation}\label{46}
\rho_1>0\quad \text{and}\quad \rho_2>0.
\end{equation}
Summarizing the above discussions, we can get the following theorem immediately.
\vspace*{0.25cm}
\begin{theorem}
(i) If the inequalities in Eq.~\eqref{34} are satisfied, then the positive equilibrium point $E^*$ is locally asymptotically stable in the presence as well as absence of diffusion.
(ii) Suppose that inequalities in Eq.~\eqref{34} are not satisfied, i.e., either $A$ or $B$ is negative or both $A$ and $B$ are negative. Then for strictly positive wave-number $k > 0$, i.e., spatially inhomogeneous perturbations, by increasing $d_1$ and
$d_2$ to sufficiently large values, $\rho_1$
and $\rho_2$ can be made positive and hence $E^*$ can be made locally asymptotically stable.
\end{theorem}
Diffusive instability sets in when at least one of the conditions in Eq.~\eqref{46} is
violated subject to the conditions in Theorem 1. But it is evident that the first condition
$\rho_1 > 0$ is not violated when the condition $A > 0$ is met. Hence only the violation of
condition $\rho_2>0$ gives rise to diffusion instability. Hence the condition for diffusive
instability is given by
\begin{equation}\label{47}
H(k^2)=d_1d_2k^4+\Bigl[d_2P^*(B_1-B_2H^*f_1(P^*, D))+d_1H^*(C_2/P^*-Ff_2(H^*, D_1))\Bigr]k_2+B<0.
\end{equation}
$H$ is quadratic in $k^2$ and the graph of $H(k^2)=0$ is a parabola. The minimum of $H(k^2)$
occurs at $k^2=k_m^2$ where
\begin{equation}\label{48}
k_m^2=\frac{1}{2d_1d_2}\Bigl[d_2P^*(B_2H^*f_1(P^*, D)-B_1)+d_1H^*(Ff_2(H^*, D_1)-C_2/P^*)\Bigr]>0.
\end{equation}
Consequently the condition for diffusive instability is $H(k_m^2)$. Therefore
\begin{equation}\label{49}
\frac{1}{4d_1d_2}\Bigl[d_2P^*(B_2H^*f_1(P^*, D)-B_1)+d_1H^*(Ff_2(H^*, D_1)-C_2/P^*)\Bigr]>B.
\end{equation}
\vspace*{0.25cm}
\begin{theorem}
The criterion for diffusive instability for the model system is obtained by combining the
result of Theorem 1, \eqref{48} and \eqref{49} and leading to the following condition:
\begin{equation}\label{410}
d_2P^*(B_2H^*f_1(P^*, D)-B_1)+d_1H^*(Ff_2(H^*, D_1)-C_2/P^*>2(Bd_1d_2)^{\frac{1}{2}}>0.
\end{equation}
\end{theorem}
In the following theorem, we are able to show the global stability behaviour of the
positive equilibrium in the presence of diffusion.
\vspace*{0.25cm}
\begin{theorem}
(i) If the equilibrium $E^*$ of system~\eqref{31} is globally asymptotically stable, the
corresponding uniform steady state of system~\eqref{42} is also globally asymptotically stable.
(ii) If the equilibrium $E^*$ of system~\eqref{31} is unstable even then the corresponding uniform steady state of system~\eqref{42} can be made globally asymptotically stable by increasing the diffusion coefficient $d_1$ and $d_2$ to a sufficiently large value for strictly positive wave-number $k > 0$.
\end{theorem}
\noindent {\bf Proof.} For the sake of simplicity, let $P(x, t)=P, H(x, t)=H$. For $x\in[0, R]$ and $t\in[0, \infty]$, let us define a functional:
$$V_1=\int_0^RV(P, H)dx$$
where $V(P, H)$ is defined as Eq.~\eqref{Lyap}.
Differentiating $V_1$ with respect to time $t$ along the solutions of system~\eqref{42}, we get
$$\frac{dV_1}{dt}=\int_0^R\Bigl(\frac{\partial V}{\partial P}\frac{\partial P}{\partial t}+\frac{\partial V}{\partial H}\frac{\partial H}{\partial t}\Bigr)dx
=\int_0^R\frac{dV}{dt}dx+\int_0^R
\Bigl(d_1\frac{\partial V}{\partial P}\frac{\partial^2 P}{\partial x^2}+d_2\frac{\partial V}{\partial H}\frac{\partial^2 H}{\partial x^2}\Bigr)dx.$$
Using the boundary condition~\eqref{23}, we obtain
\begin{equation}\label{411}
\frac{dV_1}{dt}=\int_0^R\frac{dV}{dt}dx-d_1\int_0^R
\frac{P^*(P^2+3D^2)-2D^2P}{B_2P^4}\Bigl(\frac{\partial P}{\partial x}\Bigr)^2dx-d_2\int_0^R\frac{wH^*}{H^2}\Bigl(\frac{\partial H}{\partial x}\Bigr)^2dx.
\end{equation}
From Eq.~\eqref{411}, we note that if $\frac{dV}{dt}<0$ then $\frac{dV_1}{dt}$ in the interior of the positive quadrant of the $PH-$plane, and hence the first part of the theorem follows. We also note that if $\frac{dV}{dt}<0$ then $\frac{dV_1}{dt}$
can be made negative by increasing $d_1$ and $d_2$ a sufficiently
large value, and hence the second part of the theorem follows.
\subsection{Two dimensional case}
In two-dimensional case, the model system~\eqref{21} can be written as
\begin{equation}\label{412}
\begin{array}{l}
\frac{\partial P}{\partial t}=rP-B_1P^2-\frac{B_2P^2H}{P^2+D^2}+d_1\Bigl(\frac{\partial^2P}{\partial x^2}+\frac{\partial^2P}{\partial y^2}\Bigr),\\[6pt]
\frac{\partial H}{\partial t}=C_1H-C_2\frac{H^2}{P}-\frac{FH^2}{H^2+D_1^2}+d_2\Bigl(\frac{\partial^2H}{\partial x^2}+\frac{\partial^2H}{\partial y^2}\Bigr).
\end{array}
\end{equation}
In this section, we show that the result of theorem 5 remain valid for two dimensional
case. To prove this result we consider the following functional (Dubey and Hussain,
2000):
$$V_2=\iint\limits_{\Omega}V(P, H)dA$$
where $V(P, H)$ is defined as Eq.~\eqref{Lyap}.
Differentiating $V_2(t)$ with respect to time $t$ along the solutions of system~\eqref{412}, we obtain
\begin{equation}\label{413}
\frac{dV_2}{dt}=I_1+I_2,
\end{equation}
where
$$I_1=\iint\limits_{\Omega}\frac{dV}{dt}dA,\qquad I_2=\iint\limits_{\Omega}\Bigl(d_1\frac{\partial V}{\partial P}\nabla^2P+d_2\frac{\partial V}{\partial H}\nabla^2H\Big)dA.$$
Using Green's first identity in the plane
$$\iint\limits_{\Omega}F\nabla^2GdA=\int\limits_{\partial\Omega}F\frac{\partial G}{\partial
n}ds-\iint\limits_{\Omega}(\nabla F\cdot \nabla G)dA.$$
and under an analysis similar to Dubey and Hussain(2000), one can show that
$$d_1\iint\limits_{\Omega}\left(\frac{\partial V}{\partial P}\nabla^2 P\right)dA
=-d_1\iint\limits_{\Omega}\frac{\partial^2V}{\partial P^2}\left[(\frac{\partial P}{\partial x})^2+(\frac{\partial P}{\partial y})^2\right]dA\leq 0,$$
$$d_2\iint\limits_{\Omega}\left(\frac{\partial V}{\partial H}\nabla^2 H\right)dA
=-d_2\iint\limits_{\Omega}\frac{\partial^2V}{\partial H^2}\left[(\frac{\partial H}{\partial x})^2+(\frac{\partial H}{\partial y})^2\right]dA\leq 0.$$
This shows that $I_2\leq 0$. From the above analysis we note that if $I_1<0$, then
$\frac{dV_2}{dt}$.
This implies that if in the absence of diffusion $E^*$ is globally asymptotically stable, then
in the presence of diffusion $E^*$ will remain globally asymptotically stable.
We also note that if $\frac{dV}{dt}>0$, then $I_1>0$. In such a case diffusion, $E^*$ will be unstable in the absence of diffusion. Even in this case by increasing $d_1$ and $d_2$ to a sufficiently large value $\frac{dV_2}{dt}$ can be made negative. This shows that if in the absence of diffusion $E^*$ is unstable, then in the presence of diffusion $E^*$ can be made stable by increasing diffusion coefficients to sufficiently large value.
\vspace*{0.5cm}
\setcounter{equation}{0}
\section{Numerical simulations}
In this section, we perform numerical simulations to illustrate the results obtained in
previous sections. The dynamics of the model system~\eqref{21} is studied with the help of
numerical simulation, both in one and two dimensions, to investigate the spatiotemporal
dynamics of the model system~\eqref{21}. For the one-dimensional case, the plots (space vs.
population densities) are obtained to study the spatial dynamics of the model system. The
temporal dynamics is studied by observing the effect of time on space vs. density plot of
prey populations. For the two-dimensional case, the spatial snapshots of prey densities
are obtained at different time levels for different values of $F$ and we have tried to study
the spatiotemporal dynamics of the spatial model system. The rate of fish predation, $F$
acting as forcing term or distributed control parameter in the model system~\eqref{21} which is
strictly nonnegative. Mathematically, we assume that $F$ can be manipulated at every point
in space and time. However, from practical point of view, we only have direct control of
the net density of fish population at any instant. The parameters determining the shape of
the type-III functional response of fish to zooplankton density are set more or less
arbitrarily. They will in fact very much depend on the fish species, and also change
significantly within a species with the size of the individuals.
\subsection{Spatiotemporal dynamics in One-dimensional system}
Now we start our insight into the spatiotemporal dynamics of the model system~\eqref{21} in one-dimension with the following set of parameter values. The spatiotemporal dynamics of
the system depends to a large extent on the choice of initial conditions. In a real aquatic
ecosystem, the details of the initial spatial distribution of the species can be caused by
spatially homogenous initial conditions. However, in this case, the distribution of species
would stay homogenous for any time, and no spatial pattern can emerge. To get a
nontrivial spatiotemporal dynamics, we have perturbed the homogenous distribution. We
start with a hypothetical ``constant-gradient" distribution (Malchow et al., 2008):
\begin{equation}\label{51}
P(x,0)=P^*,\quad H(x,0)=H^*+\varepsilon x+\delta,
\end{equation}
where $\varepsilon$ and $\delta$ are parameters. $(P^*, H^*)=(1.8789, 1.3984)$ is the non-trivial equilibrium point of system~\eqref{31}, with fixed parameter set
$$r=1, B_1=0.2, B_2=0.91, C_1=0.22, C_2=0.2, D^2=0.3, D_1=0.1, F=0.1.$$
It appears that the type of the system dynamics depends on $\varepsilon$ and $\delta$. In Figs. 2 and 3, we show the population distribution over space at time $t = 2000$ (Fig.2) and $t=1000, 2000$(Fig.3), respectively. In Fig. 4, we illustrate the pattern formation about the above two cases, the system size is 3000, iteration numbers from 17800 to 19800, i.e., time $t$ from 1780 to 1980. In Fig.2a and Fig.4a, $\varepsilon=-1.5\times 10^{-6}$ and $\delta=10^{-6}$, the spatial distributions gradually vary in time, and the local temporal behavior of the dynamical variables $P$ and $H$ is strictly periodic following limit cycle of the non-spatial system, which is perhaps what is intuitively expected from system~\eqref{41}, there emerge a periodic wave pattern (Fig.4a, for $P$). However, in Fig.2b and Fig.4b, for a slightly different set of $\varepsilon$ and $\delta$, i.e.,
$\varepsilon=-1.5\times 10^{-3}$ and $\delta=10^{-6}$, the dynamics of the system undergoes principal changes, and there emerge an oscillations wave pattern with a ``flow" pattern around $x=1500$ (c.f., Fig.4b). In Fig.3 and Fig.4c, $\varepsilon=-1.5\times 10^{-2}$ and $\delta=10^{-3}$, more different with the two cases above,
the initial conditions~\eqref{51} lead to the formation of strongly irregular ``jagged" dynamics pattern inside a sub-domain of the system. There is a growing region of periodic
travelling waves ($x = 2000$ to $x =2500$) followed by oscillations with a large spatial
wavelength. This is a typical scenario which precedes either the formation of
spatiotemporal irregular oscillations or spatially homogeneous oscillations (c.f., Fig.4c).
\begin{figure}[htbp]
\centerline{\includegraphics[scale=0.6]{fig2.eps}}
\caption{Population distributions over space at $t = 2000$ obtained for the parameter values $r=1, B_1=0.2, B_2=0.91, C_1=0.22, C_2=0.2, D^2=0.3, D_1=0.1, F=0.1$ and the initial conditions ~\eqref{51} with (a) $\varepsilon=-1.5\times 10^{-6}$ and $\delta=10^{-6}$; (b) $\varepsilon=-1.5\times 10^{-3}$ and $\delta=10^{-6}$.}
\label{figure2}
\end{figure}
\begin{figure}[htbp]
\centerline{\includegraphics[scale=0.6]{fig3.eps}}
\caption{Population distributions over space at (a) $t=1000$ (b) $t=2000$ obtained for the parameter values and the initial conditions~\eqref{51} with $\varepsilon=-1.5\times 10^{-2}$ and $\delta=10^{-3}$. The values of the other parameters are the same as given in Fig. 2.}
\label{figure3}
\end{figure}
\begin{figure}[htbp]
\centerline{\includegraphics[scale=0.8]{fig4.eps}}
\caption{Spatiotemporal dynamical of system~\eqref{41} with the initial conditions~\eqref{51}. Parameters: (a) $\varepsilon=-1.5\times 10^{-6}$, $\delta=10^{-6}$; (b) $\varepsilon=-1.5\times 10^{-3}$, $\delta=10^{-6}$; (c) $\varepsilon=-1.5\times 10^{-2}$, $\delta=10^{-3}$.}
\label{figure4}
\end{figure}
\subsection{Spatiotemporal dynamics in Two-dimensional system}
In this subsection, we show the spatiotemporal dynamics in two-dimensional case of
system~\eqref{412}.
\subsubsection{Turing pattern}
First, we show Turing pattern of system~\eqref{412} obtained by performing numerical
simulations with initial and boundary conditions given in Eqs.~\eqref{22}--\eqref{23} with the
parameter values given below in Eq. (27) with system size $200\times 200$ for different values
of $F$ and time $t$. The values of the other parameters are fixed as
\begin{equation}\label{52}
\begin{array}{l}
r=1, B_1=0.2, B_2=0.91, C_1=0.22, C_2=0.2, D^2=0.3, D_1=0.1, \\[6pt]
d_1=0.005, d_2=1,
\Delta x=\Delta y=1/3, \Delta t=1/40.
\end{array}
\end{equation}
Initially, the entire system is placed in the stationary state
$(P^*, H^*)$, and the propagation velocity of the initial
perturbation is thus on the order of $5\times 10^{-4}$ space units
per time unit. And the system is then integrated for $10^5$ or $2\times 10^5$ time steps and some images saved. After the initial period
during which the perturbation spreads, either the system goes into a
time dependent state, or to an essentially steady state (time
independent). Here, we show the
distribution of phytoplankton $P$, for instance.
Fig.5 shows five typical Turing pattern of phytoplankton $P$ in system~\eqref{412} with
parameters set~\eqref{52}. From Fig. 5, one can see that values for the concentration P are
represented in a color scale varying from blue (minimum) to red (maximum), and on
increasing the control parameter $F$, the sequences spots $\rightarrow$ spot-stripe mixtures$\rightarrow$ stripes$\rightarrow$ hole-stripe mixtures holes pattern is observed. When $F > 0.077$, other parameters are
the same as above case, wave pattern emerges and are presented in Fig. 6.
\begin{figure}[htbp]
\centerline{\includegraphics[scale=0.8]{fig5.eps}}
\caption{Typical Turing patterns of $P$ in the system~\eqref{412} with fixed parameters set ~\eqref{52}. (a) spots pattern, $F=0.0001$; (b) spot-stripe mixtures pattern, $F=0.01$; (c) stripes pattern, $F=0.035$; (d) hole-stripe mixtures, $F=0.065$; (e) holes pattern, $F=0.0765$. Iterations: Pattern-(a) and (e): $2\times 10^5$, others: $10^5$.}
\label{figure5}
\end{figure}
\begin{figure}[htbp]
\centerline{\includegraphics[scale=0.52]{fig6.eps}}
\caption{Wave pattern obtained with system~\eqref{412} with $F=0.085$. Iterations: (a) 0, (b) 10000, (c) 25000, (d) 40000.}
\label{figure6}
\end{figure}
\subsubsection{Spiral wave pattern}
Thanks to the insightful work of Medvinsky et al. (2002), we have studied the spiral wave pattern for two different set of initial condition discussed in Eqs.~\eqref{53} and~\eqref{54} respectively. In these two cases, we employ $\Delta x=\Delta y=1, \Delta t=1/3$ and the system size is $900\times 300$. Other parameters set is fixed as~\eqref{52}.
The first case is:
\begin{equation}\label{53}
\begin{array}{l}
P(x,y,0)=P^*-\varepsilon_1(x-0.2y-225)(x-0.2y-675),\\[6pt] H(x,y,0)=H^*-\varepsilon_2(x-450)-\varepsilon_3(y-150).
\end{array}
\end{equation}
where $\varepsilon_1=2\times 10^{-7}, \varepsilon_2=3\times 10^{-5}, \varepsilon_3=1.2\times 10^{-4}$.
The initial conditions are deliberately chosen to be unsymmetrical in order to make any influence of the corners of the domain more visible. Snapshots of the spatial distribution arising from~\eqref{53} are shown in Fig. 7 for $t = 0, 2500, 5000, 10000$. Fig.7a shows that for the system~\eqref{412} with initial conditions~\eqref{53}, the formation of the irregular patchy structure can be preceded by the evolution of a regular spiral spatial pattern. Note that the appearance of the spirals is not induced by the initial conditions. The center of each spiral is situated in a critical point $(x_{cr}, y_{cr})$, where $U(x_{cr}, y_{cr}) = P^*$, $V(x_{cr}, y_{cr}) = H^*$. The distribution~\eqref{53} contains two such points; for other initial conditions, the number of spirals may be different. After the spirals form (Fig. 7b), they grow slightly for a certain time, their spatial structure becoming more distinct (Figs. 7c and 7d).
\begin{figure}[htbp]
\centerline{\includegraphics[scale=0.8]{fig7.eps}}
\caption{Spiral pattern of P in system~\eqref{412} with special initial condition~\eqref{53} and $F=0.1$. Time: (a) 0, (b) 2500, (c) 5000, (d) 10000.}
\label{figure7}
\end{figure}
In the second case, the initial conditions describe a phytoplankton (prey) patch placed into a domain with a constant-gradient zooplankton (predator) distribution:
\begin{equation}\label{54}
\begin{array}{l}
P(x,y,0)=P^*-\varepsilon_1(x-180)(x-720)-\varepsilon_2(y-90)(y-210),\\[6pt] H(x,y,0)=H^*-\varepsilon_3(x-450)-\varepsilon_4(y-135).
\end{array}
\end{equation}
where $\varepsilon_1=2\times 10^{-7}, \varepsilon_2=6\times 10^{-7}, \varepsilon_3=3\times 10^{-5}, \varepsilon_4=6\times 10^{-5}$.
Fig. 8 shows the snapshots of the spatial distribution arising from~\eqref{53} for $t = 0, 1000, 4000, 9000$. Although the dynamics of the system preceding the formation of patchy spatial structure is somewhat less regular than in the previous case, it follows a similar scenario. Again the spirals appear, with their centers located in the vicinity of critical points (Figs. 8c and 8d).
\begin{figure}[htbp]
\centerline{\includegraphics[scale=0.8]{fig8.eps}}
\caption{Spiral pattern of P in system~\eqref{412} with special initial condition~\eqref{54} and $F=0.1$. Time: (a) 0, (b) 1000, (c) 4000, (d) 9000.}
\label{figure7}
\end{figure}
\section{Discussions and Conclusions}
In this paper, we have made an attempt to provide a unified framework to understand the complex dynamical patterns observed in a spatial plankton model for phytoplankton-zooplankton-fish interaction with Holling type III functional response. Since most marine mammals and boreal fish species are considered to be generalist predators (including feeders, grazers, etc.), a sigmoidal (type III) functional response might be more appropriate (Magnusson and Palsson, 1991). This has been proposed as being likely for minke whales, harp seals, and cod in the Barents Sea where switching between prey species has been hypothesized (Nilssen et al., 2000; Schweder et al., 2000). Sigmoidal functional response arising due to heterogeneity of space is less addressed in the literature. It is well known that real vertical distribution of plankton is highly heterogeneous, especially during plankton blooms and this can seriously affect the behavior of plankton models (Poggiale et al. 2008). Morozov (2010) shows that emergence of Holling type III in plankton system is due to mechanisms different from those well known in the literature (e.g. food search learning by predator, existence of alternative food, refuse for prey). It has certainly been useful in both theoretical investigations and practical applications (Turchin, 2003). Movement of phytoplankton and zooplankton population with different velocities can give rise to spatial patterns (Malchow, 2000). We have studied the reaction-diffusion model in both one and two dimensions and investigated their stability conditions. By analyzing the model system~\eqref{21}, we observed that two points of equilibrium exist for the model system. The non-trivial equilibrium state of prey-predator coexistence is locally as well as globally asymptotically stable under a fixed region of attraction when certain conditions are satisfied. It has been shown that the instability observed in the system is diffusion-driven under the condition~\eqref{410}. The expression for critical wave-number has been obtained. It has been observed that the solution of the model system converges faster to its equilibrium in case of diffusion in two-dimensional space as compared to the diffusion in one dimensional space.
In the numerical simulation, we adopt the rate of fish predation F in the system~\eqref{21} as the control parameter which is strictly nonnegative. In the two-dimensional case, on increasing the value of $F$, the sequences spots$\rightarrow$spot-stripe mixtures$\rightarrow$stripes$\rightarrow$ hole-stripe mixtures$\rightarrow$holes$\rightarrow$wave pattern is observed. For the sake of learning the wave pattern further, we show the time evolution process of the pattern formation with two special initial conditions and find spiral wave pattern emerge. That is to say, our two-dimensional spatial patterns may indicate the vital role of phase transience regimes in the spatiotemporal organization of the phytoplankton and zooplankton in the aquatic systems. It is also important to distinguish between ``intrinsic" patterns, i.e., patterns arising due to trophic interaction such as those considered above, and ``forced" patterns induced by the inhomogeneity of the environment. The physical nature of the environmental heterogeneity, and thus the value of the dispersion of varying quantities and typical times and lengths, can be essentially different in different cases.
\vspace*{0.5cm}
\section*{Acknowledgements}
This work is supported by University Grants Commission, Govt. of India under grant no. F.33-116/2007(SR) to the corresponding author (RKU) who has visited to Budapest, Hungary under Indo-Hungarian Educational exchange programme. Authors are also grateful to Dr. Sergei V. Petrovskii and the reviewer for their critical review and helpful suggestions.
|
\section{Introduction}\label{sec:introduction}
Adaptive coding techniques are frequently employed, especially in wireless communications, in order to dynamically
adjust the coding rate to changing channel conditions. An example of adaptive coding technique consists in puncturing a
mother code. When the channel conditions are good more bits are punctured and the coding rate is increased. In poor
channel conditions all redundant bits are transmitted and the coding rate drops. However, in harsh conditions, the
receiver might not be able to successfully decode the received signal, even if all the redundant bits have been
transmitted. In such a case, the coded block can be retransmitted until the sent information is successfully decoded.
This is equivalent to additional repetition coding, which further lowers the coding rate below the mother coding rate.
However, the use of retransmission techniques might not be suitable nor possible in some situations, such as
multicast/broadcast transmissions, or whenever the return link is strictly limited or not available (such situations
are generally encountered in satellite communications). The main alternative in this case is the use of {\em erasure
codes} that operate at the transport or the application layer of the communication system: source data packets are
extended with redundant (also referred to as {\em repair}) packets that are used to recover the lost data at the
receiver. Physical (PHY) and upper layer (UL) codes are not mutually exclusive, but they are complementary to each
other. Adaptive coding schemes are also required at the upper layer, in order to dynamically adjust to variable loss
rates. Besides, codes with very small rates or even {\em rateless} \cite{luby2002lc,shokrollahi2006rc} are sometimes
used at the application layer for fountain-like content distribution applications.
In this paper we propose a coding technique that allows to produce extra redundant bits, such as to decrease the coding
rate below the mother coding rate. Extra redundant bits can be produced in an incremental way, yielding very small
coding rates, or can be optimized for a given target rate below the mother coding rate. As for puncturing, the proposed
technique allows for using the same decoder, regardless of how many extra redundant bits have been produced, which
considerably increases the flexibility of the system, without increasing its complexity.
The proposed coding scheme is based on non-binary low density parity check (NB-LDPC) codes \cite{gall-monograph} or,
more precisely, on their {\em extended binary image} \cite{Valentin10}. If $q=2^p$ denotes the size of the non-binary
alphabet, each non-binary symbol corresponds to a $p$-tuple of bits, referred to as its binary image. Extra redundant
bits, called {\em extended bits}, are generated as the XOR of some bits from the binary image of the same non-binary
coded symbol. If a certain number of extended bits are transmitted over the channel, we obtain an {\em extended code},
the coding rate of which is referred to as {\em extended (coding) rate}. In the extreme case when all the extended bits
are transmitted, the mother code is turned into a {\em very small rate} code, and can be used for fountain-like content
distribution applications \cite{Valentin10}. A similar approach to fountain codes, by using multiplicatively repeated
NB-LDPC codes, has been proposed in \cite{kasai10}. If some extended rate is targeted, we show that the extended code
can be optimized by using density evolution methods.
The paper is organized as follows. Section \ref{sec:nbldpc_codes} gives the basic definitions and the notation related
to NB-LDPC codes. In Section \ref{sec:extended_nbldpc_codes}, we introduce the extended NB-LDPC codes and discuss their
erasure decoding. The analysis and optimization of extended NB-LDPC codes are addressed in Section
\ref{sec:analysis_optimization}. Section \ref{sec:code_design_performance} focuses on the code design and presents
simulation results, and Section \ref{sec:conclusions} concludes the paper.
\section{Non-binary LDPC Codes}\label{sec:nbldpc_codes}
We consider NB-LDPC codes defined over $\f_q$ \cite{Davey-MacKey}, the finite field with $q$ elements, where $q = 2^p$
is a power of $2$ (this condition is only assumed for practical reasons). We fix once for all an isomorphism of vector
spaces:
\begin{equation}
\label{eq:identify}
\f_q \stackrel{\sim}{\longrightarrow} \f_2^p \nonumber
\end{equation}
Elements of $\f_q$ will also be referred to as {\em symbols}, and we say that $\ux = (x_0,\dots,x_{p-1})\in\f_2^p$ is
the {\em binary image} of the symbol $X\in{\f_q}$, if they correspond to each other by the above isomorphism. A
non-binary LDPC code over $\f_q$ is defined as the kernel of a sparse parity-check matrix $H\in
\mathbf{M}_{M,N}(\f_q)$. Alternatively, it can be represented by a bipartite (Tanner) graph \cite{Tann} containing
symbol-nodes and constraint-nodes associated respectively with the $N$ columns and $M$ rows of $H$. A symbol-node and a
constraint-node are connected by an edge if and only if the corresponding entry of $H$ is non-zero; in this case, the
edge is assumed to be {\em labeled} by the non-zero entry. As usually \cite{Rich-Urba}, we denote by $\lambda$ and
$\rho$ the left (symbol) and right (constraint) edge-perspective degree distribution polynomials. Hence, $\lambda(x) =
\sum_d \lambda_d x^{d-1}$ and $\rho(x) = \sum_d \rho_d x^{d-1}$, where $\lambda_d$ and $\rho_d$ represent the fraction
of edges connected respectively to symbol and constraint nodes of degree-$d$. The design coding rate is defined as $r =
1 - \frac{\int_{0}^{1}\rho(x)\text{d}x}{\int_{0}^{1}\lambda(x)\text{d}x}$, and it is equal to the coding rate if and
only if the parity-check matrix is full-rank.
\section{Extended Non-binary LDPC Codes}\label{sec:extended_nbldpc_codes}
\subsection{Extended code description}
For any integer $1 \leq k \leq q-1$, let $[k] = (k_0, \dots, k_{p-1})^{\text{T}}$ denote the column vector
corresponding to the binary decomposition of $k$; {\em i.e.} $k = \sum_{i = 0}^{p-1}k_i2^i$, with $k_i\in\{0,1\}$. Let
$X\in{\f_q}$ be a non-binary symbol, and $\ux = (x_0,\dots,x_{p-1})$ be its binary image. The $k$-th {\em extended bit}
of $X$ is by definition:
$$\alpha_k = \ux \times [k] = \sum_{i = 0}^{p-1} k_i x_i$$
The vector $\ualpha = (\alpha_1,\dots,\alpha_{q-1})$ is called {\em extended binary image} of $X$. Note that
$\alpha_{2^i} = x_i$, for any $0\leq i \leq p-1$. An extended bit $\alpha_k$ is said to be {\em nontrivial} if $k$ is
not a power of $2$ (hence, $\alpha_k$ is a linear combination of at least two bits from the binary image $\ux$ of $X$).
Now, consider a NB-LDPC code defined over $\f_q$, with coding rate $r = \frac{K}{N}$. Let $(X_1, \dots, X_N) \in
\f_q^N$ be a non-binary codeword, $(\ux_1, \dots, \ux_N) \in \f_2^{Np}$ be its binary image, and $(\ualpha_1, \dots,
\ualpha_N) \in \f_2^{N(q-1)}$ be its extended binary image. By transmitting the extended binary image over the channel,
we obtain a code with rate $\frac{Kp}{N(q-1)} = r\frac{p}{q-1}$, which can be advantageously used for application
requiring very small coding rates \cite{Valentin10}.
We define an {\em extension of the NB-LDPC code}, as a family of matrices $\{A_1,\dots,A_N\}$, where each
$A_n\in\mathbf{M}_{p,t_n}(\f_2)$ is a binary matrix with $p$ rows and $t_n$ columns, with $t_n\geq 0$. Let $\ua_n =
\ux_n \times A_n \in \f_2^{t_n}$; hence, $\ua_n$ is constituted of $t_n$ extended bits of $X_n$ (possibly with
repetitions, if $A_n$ contains two or more identical columns). The binary vector $(\ua_1,\dots,\ua_N)$ is called {\em
extended codeword}, and the {\em extended coding rate} is given by $r_e = \frac{Kp}{T}$, where $T = \sum_{n=1}^N t_n$.
Note that the above definition is very broad, and it can yield extended rates below as well as extended rates above the
mother coding rate. In particular, it includes punctured codes: if $A_n = \text{col}([2^1], \dots, [2^{p-1}])$, then
$\ua_n = (x_{n,1},\dots,x_{n,p-1})$, which is the same as puncturing the first bit, $x_{n,0}$, from the binary image of
$X_n$. Moreover, taking some $t_n = 0$ is equivalent to puncturing the whole symbol $X_n$. The optimization of
puncturing distributions for NB-LDPC codes has been addressed in \cite{ISITA2010}. In this paper, we restrict ourselves
to the case when matrices $A_n$ are of the form $A_n = [I_p \mid B_n ]$, where $I_p$ is the $p\times p$ identity
matrix, meaning that each $\ua_n$ contains the binary image $\ux_n$ (the use of ``extension'' complies with its literal
meaning). We will further assume that any two columns of a matrix $A_n$ are different. It follows that $p\leq t_n \leq
q-1$, and each $\ua_n$ is constituted of the binary image $\ux_n$ and $t_n-p$ pairwise different (nontrivial) extended
bits. In this case, we shall say that {\em the symbol $X_n$ is extended by $k=t_n-p$ bits}.
For instance, let $p=3$ and consider that the following matrix $A$ is used to extend by $2$ bits some coded symbol $X$:
$$
\ua = \ux \times A, \text{ where } A = \left[
\begin{array}{lllll}
1 & 0 & 0 & 1 & 0 \\
0 & 1 & 0 & 0 & 1 \\
0 & 0 & 1 & 1 & 1
\end{array}
\right]
$$
Then $\ua = (\alpha_1, \alpha_2, \alpha_4, \alpha_5, \alpha_6) = (x_0, x_1, x_2, x_0 \wedge x_2, x_1 \wedge x_2)$,
where $\wedge$ denotes the bit-xor operator.
In order to determine the coding rate of the extended code, we denote by $\fdk$ the {\em fraction of degree-$d$ symbols
with $k$ nontrivial extended bits}, $0 \leq k < q - p$; thus
$\sum_{k=0}^{q-p-1}\!\fdk \!= \!1$.\break
The average number of nontrivial extended bits per coded symbol is given by
${f} = \sum_{d=1}^{d_s} \Lambda_d\sum_{k=0}^{q-p-1}k\fdk$,
where $d_s$ is the maximum symbol node degree, and $\Lambda_d = \frac{\lambda_d}{d \int_0^1{\lambda(x)dx}}$ is the
fraction of degree-$d$ symbol nodes. It follows that the extended coding rate is given by $r_e = r\frac{p}{p+f}$. Thus,
we can achieve an arbitrary extended rate within the interval $\left[r \frac{p}{q - 1} , r\right]$ by varying the
parameter $f$.
Figure \ref{fig:extended_codes} illustrates an extended code defined over $\f_8$, with $f_{2,0} = f_{2,1} = f_{2,2} =
f_{2,4} = 1/4$, and $f_{3,2} = f_{3,3} = 1/2$, which correspond to $f=2$. The mother coding rate is $r=0.5$ and the
extended coding rate is $r_e = 0.3$.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{extended_codes
\caption{Extension of a NB-LDPC code. Blue circles represent bits of the binary image ($p=3$), while red circles
represent (nontrivial) extended bits.}
\label{fig:extended_codes
\vspace{-3mm}
\end{figure}
\subsection{Iterative erasure decoding}
We consider that the extended codeword $(\ua_1,\dots,\ua_N)$ is transmitted over a binary erasure channel (BEC). At the
receiver part, the received bits (both from the binary image and extended bits) are used to reconstruct the
corresponding non-binary symbols. Precisely, for each received bit we know its position within the extended binary
image of the corresponding symbol. Hence, for each symbol node we can determine a set of {\em eligible symbols} that is
constituted of symbols whose extended binary images match the received bits. These sets are then iteratively updated,
according to the linear constraints between symbol-nodes \cite{Valentin08}. Alternatively (and equivalently), the
extended code can be decoded by using the linear-time erasure decoding proposed in \cite{Valentin10}.
The {\em asymptotic threshold} of an ensemble of codes is defined as the maximum erasure probability $\pth$ that allows
transmission with an arbitrary small error probability when the code length tends to infinity \cite{Rich-Urba}. Given
an ensemble of codes, its threshold value can be efficiently computed by tracking the fraction of erased messages
passed during the belief propagation decoding; this method is referred to as {\em density evolution}. In this paper,
the density evolution is approximated based on the Monte-Carlo simulation of an infinite code, similar to the method
presented in \cite{ISITA2010}. This method has two main advantages: it can easily incorporate the extending
distribution $\{\fdk\}_{d,k}$, and it can be extrapolated to more general channel models.
\section{Analysis and optimization}\label{sec:analysis_optimization}
\begin{figure}[!b]
\centering\vspace{-3mm}
\includegraphics[width=\columnwidth]{1_ext_bit_selection
\caption{1-bit extension for regular NB-LDPC codes over $\f_{16}$
\label{fig:ext_bit_selec_expl
\end{figure}
The goal of this section is to answer the following questions.
First of all, assume that we have given a symbol node that has to be extended by $k$ bits. How should they be chosen
among the $q-p-1$ (nontrivial) extended bits?
Secondly, given an extended coding rate $r_e$, how should be extended bits distributed over the symbol-nodes? Put
differently, which is the optimal extending distribution $\left\{\fdk\right\}$?
\subsection{Extended bits selection strategy} \label{subsubsec:ext_bit_selec}
We assume that we have given a symbol-node that has to be extended by $k$ bits. A choice of the $k$ bits among the
$q-p-1$ extended bits corresponds to an extending matrix $A = [I_p \mid B]$ of size $p\times (p+k)$, with pairwise
distinct columns. For each such a matrix, assume that the extended symbol $\ua = \ux \times A$ is transmitted over the
BEC, and let $E(A)$ be the expected number of eligible symbols at the receiver. Recall that an eligible symbol is a
symbol whose extended binary image match the received bits. If all transmitted bits have been erased, any symbol is
eligible. Conversely, if the received bits completely determine the non-binary symbol, then there is only one eligible
symbol. More generally, let $\ua_{\text{rec}}$ denote the sequence of received bits, and $A_{\text{rec}}$ denote the
submatrix of $A$ determined by the columns that correspond to the received positions of $\ua$. Then the eligible
symbols are the solutions of the linear system $\ux \times A_{\text{rec}} = \ua_{\text{rec}}$, and their number is
equal to $2^{p-\text{rank}(A_{\text{rec}})}$. Now, if $\epsilon$ denotes the erasure probability of the BEC, it can be
easily verified that:
$$E(A) = \sum_{i=0}^{p+k}(1-\epsilon)^i\epsilon^{p+k-i}\left(\sum_{A_i\subseteq A} 2^{p-\text{rank}(A_i)}\right),$$
where the second sum takes over all the submatrices $A_i$ cons\-ti\-tuted of $i$ among the $p+k$ columns of $A$. Hence,
in order to minimize the expected number of eligible symbols $E(A)$, we choose $A$ such that $\dmin(A)$ is maximal,
where $\dmin(A)$ is the smallest number of linearly dependent columns of $A$.
\subsubsection{One extended bit per symbol node} \label{expl:selec_one_ext_bit} Consider the ensemble of regular
$\left(\lambda(x) = x, \rho(x) = x^3\right)$ LDPC codes defined over the $\f_{16}$. Assume that each symbol-node is
extended by $k=1$ bit, such as to achieve an extended rate $r_e = 0.4$. According to the choice of the extended bit
(among the $11$ nontrivial extended bits), $\dmin$ may be equal to $3,4$, or $5$. The asymptotic threshold
corresponding to each choice of the extended bit is shown in Figure \ref{fig:ext_bit_selec_expl}. Note that extended
bits are ordered on the abscissa according to the corresponding $\dmin$. For comparison purposes, we show also the
asymptotic threshold corresponding to the repetition of some bit from the binary image (trivial extended bit
$\alpha_{2^i} = x_i$), in which case $\dmin=2$. Also, the blue line correspond to the threshold obtained if each symbol
node was extended by choosing a random nontrivial extended bit. We observe that the best threshold is obtained when
each symbol node is extended by $\alpha_{15}$, which is\break the XOR of the four bits $x_{0},x_{1},x_{2},x_{3}$ of the
binary image.
\begin{figure}[!b]
\centering \vspace{-3mm}
\includegraphics[width=\columnwidth]{ext_bit_selection
\caption{$k$-bits extension for regular and semi-regular NB-LDPC codes over $\f_{16}$
\label{fig:ext_bit_selec_expl_overall
\end{figure}
\subsubsection{Several extended bits per symbol node} \label{expl:selec_ext_bit}
We consider two ensembles of regular codes $\mathcal{C}_1\left(\lambda(x)=x, \rho(x)=x^3\right)$,
$\mathcal{C}_2\left(\lambda(x)=x^2, \rho(x)=x^5\right)$ and one ensemble of semi-regular codes
$\mathcal{C}_3\left(\lambda(x)=0.5x+0.5x^4, \rho(x)=0.25x^4+0.75x^5\right)$, of coding rate $r=1/2$, defined over
$\f_{16}$. For each ensemble of codes, we consider five different cases, in which all symbol nodes are extended by the
same number $k$ of bits, with $k = 1$, $2$, $3$, $4$, and $5$. Accordingly, the extended coding rate $r_e = 0.4$,
$0.33$, $0.29$, $0.25$, and $0.22$.
The {\em normalized gap to capacity}, defined as:
$$\delta = \frac{\text{capacity} - \text{threshold}}{\text{capacity}} = \frac{1-r-\pth}{1-r},$$
is shown in Figure \ref{fig:ext_bit_selec_expl_overall}. Solid curves correspond to a $\dmin$-optimized choice of the
extended bits, while dashed curves correspond to a random choice of the extended bits. For $k = 5$, there is only a
small difference between these two strategies. However, when $k \leq 4$, the gain of the $\dmin$-optimized choice is
significant for both regular and semi-regular codes.
\subsection{Extending distribution analysis} \label{sec:extending_distr}
First of all, we discuss the case of regular codes. In Figure~\ref{fig:spread_cluster_regular_expl}, we consider three
ensembles of regular codes over $\f_{16}$, with coding rate $r = 0.5$. For each ensemble of codes, we consider five
cases, corresponding to values of $k$ between $1$ and $5$. In each case, a fraction $f_k$ of symbol-nodes are extended
by $k$ bits, while the remaining symbol-nodes have no extended bit. The fraction $f_k$ is chosen such that the extended
coding rate $r_e = 0.4$. Hence, $f_k = 1, 0.5, 0.33, 0.25,0.20,0.09$, for $k = 1,2,3,4,5$, respectively. The right most
point on the abscissa corresponds to a sixth case, in which the extended rate $r_e = 0.4$ is achieved by extending
$9\%$ of symbol-nodes by $k=11$ bits (hence, $k = q-p-1$, which is the maximum number of extended bits). For any of the
three ensembles, we can observe that the smallest gap to capacity is obtained for $k=1$, which means that extended bits
are {\em spread} over as many symbol nodes as possible (in this case, $100$\%), instead of being {\em clustered} over
the smallest possible number of symbol-nodes.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{spreading_clustering_regular
\caption{Comparison of spreading vs. clustering extending distributions for regular NB-LDPC codes over $\f_{16}$
\label{fig:spread_cluster_regular_expl
\vspace{-4mm}
\end{figure}
In case of irregular NB-LDPC codes, let $\phi = \left\{f_{d,k}\right\}_{d,k}$ be an extending distribution. Thus,
$\fdk$ is the fraction of degree-$d$ symbols with $k$ nontrivial extended bits, $0 \leq k < (q - p)$. Let ${f}_d$
denote the average number of extended bits per symbol-node of degree-$d$; that is:
$$ {f}_d = \sum_{k = 0}^{q-p-1}k\fdk \in [0, q-p-1]$$
\noindent We say that {\em the extending distribution $\phi$ is of spreading-type} if for any degree $d$, $f_{d,k} \neq
0$ only if $k=\lfloor{f}_d\rfloor$ or $k=\lceil{f}_d\rceil$. In different words, for any degree $d$, the extended bits
are uniformly spread over all the symbol-nodes of degree $d$. Clearly, a spreading-type distribution is completely
determined by the parameters $\{{f}_d\}$, as we have $f_{d,\lfloor{f}_d\rfloor} = \lceil{f}_d\rceil - {f}_d$,
$f_{d,\lceil{f}_d\rceil} = \lfloor{f}_d\rfloor - {f}_d$, and $f_{d,k} = 0$ for $k \not\in \{\lfloor{f}_d\rfloor,
\lceil{f}_d\rceil\}$.
\smallskip
\noindent We say that {\em the extending distribution $\phi$ is of clustering-type} if for any degree $d$, $f_{d,k}
\neq 0$ only if $k=q-p+1$. In different words, for any degree $d$, the extended bits are clustered over the smallest
possible fraction of symbol-nodes of degree $d$. Clearly, a clustering-type distribution is completely determined by
the parameters $\{{f}_d\}$, as we have $f_{d,q-p+1} = \frac{{f}_d}{q-p+1}$ and $f_{d,k} = 0$ for $k \neq q-p+1$.
Now, let us consider the ensemble of semi-regular LDPC codes over $\f_{16}$ with edge-perspective degree distribution
polynomials $\lambda(x) =0.5x+0.5x^4$ and $\rho(x) =0.25x^4+0.75x^5$. The mother coding rate is $r=0.5$, and we intend
to extend symbol-nodes such as to achieve extended coding rates $r_e \in \{0.45, 0.4, 0.35, 0.3\}$. Several extending
distributions are compared in Figure \ref{fig:ext_distr_irreg_expl}. There are three spreading-type distributions,
which spread the extended bits over all the symbol-nodes, or only over the symbol-nodes of degree either $2$ or $5$,
and two clustering-type distributions, which cluster the extended bits over the symbol-nodes of degree either $2$ or
$5$. In all cases, extended bits (or, equivalently, extending matrices $A_n$) are chosen such as to maximize the
corresponding $\dmin$ values. We observe that the smallest gap to capacity is obtained for extending distributions that
spread extended bits either over the degree-$5$ symbol nodes only ($r_e = 0.45,0.4$), or over all the symbol-nodes
($r_e = 0.35,0.3$).
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{spreading_clustering_irregular
\caption{Comparison of several extending-distributions, for semi-regular NB-LDPC codes over $\f_{16}$
\label{fig:ext_distr_irreg_expl
\vspace{-4mm}
\end{figure}
\subsection{Extending distribution optimization} \label{subsec:strategy}
Based on the above analysis, we only consider spreading-type extending distributions. Such an extending distribution is
completely determined by the parameters $\{{f}_d\}$, and the extended coding rate can be computed by $r_e = \frac{r}{1+
\frac{1}{p}\sum_{d} \Lambda_d {f}_d}$, where $\Lambda_d$ is the fraction of symbol-nodes of degree $d$.
For given degree distribution polynomials $\lambda$ and $\rho$, and a given extending rate $r_e$, we use the
differential evolution algorithm \cite{DiffEvol} to search for parameters $\{{f}_d\}$ that minimize the asymptotic gap
to capacity. We assume that, for each symbol-node, the extended bits are chosen such as to maximize the corresponding
$\dmin$. The optimized extended codes are presented in the next section.
\section{Code Design and Performance}\label{sec:code_design_performance}
In this section we present optimized extending distributions for an irregular mother code over $\f_{16}$. The mother
code has coding rate $r = 1/2$, and it has been optimized by density evolution. The asymptotic threshold is $\pth =
0.4945$, and the edge-perspective degree distribution polynomials are:
$$ \begin{array}{l}
\lambda(x) = 0.596 x + 0.186 x^4 + 0.071 x^7 + 0.147x^{17} \\
\rho(x) = 0.2836 x^4 + 0.7164 x^5
\end{array}$$
We optimized extending distributions for extended rates $r_e\in\{0.45, 0.40, 0.35, 0.30, 0.25, 0.20\}$. Optimized
distributions $\{{f}_{d}\}$ are shown in Table \ref{tab:opt_ext_distr}, together with the corresponding asymptotic
threshold $\pth$ and normalized gap to capacity $\delta$. For comparison purposes, we have also indicated the
normalized gap to capacity $\delta_{\text{rand}}$ corresponding to a random choice of extended bits. The last column
corresponds to extended rate $r_e = 2/15$, obtained by extending each symbol-node by the maximum number of extended
bits, {\em i.e.} $q-p-1 = 11$ bits. It can be observed that the optimized distributions allow to maintain an almost
constant value of $\delta\approx 0.01$, for all extended rates $0.45 \geq r_e \geq 2/15$.
\begin{table}[!t]
\centering \caption{Optimized extending distributions for a mother NB-LDPC code with $r=0.5$ over $\f_{16}$, for $r_e =
\left\{0.45\right.$, 0.4, 0.35, 0.3, 0.25, $\left.0.2\right\}$} \label{tab:opt_ext_distr
\begin{tabular}{|@{\ }l@{\ }|*{8}{@{\ }c@{\ }|}}
\hline
\textbf{$r_e$} & \textbf{0.5} & \textbf{0.45} & \textbf{0.4} & \textbf{0.35} & \textbf{0.3} & \textbf{0.25} & \textbf{0.2} & \textbf{2/15} \\
\hline
\hline
$f_2$ & 0 & 0.4610 & 1.0164 & 1.7851 & 2.7442 & 4.1290 & 6.1737 & 11 \\
\hline
$f_5$ & 0 & 0.3731 & 1.2113 & 1.2981 & 2.5055 & 3.5864 & 5.3409 & 11 \\
\hline
$f_8$ & 0 & 0.2487 & 0.0359 & 1.8748 & 1.6831 & 2.3393 & 4.7494 & 11 \\
\hline
$f_{18}$ & 0 & 0.1309 & 0.4871 & 0.8511 & 1.6415 & 2.9800 & 4.0234 & 11 \\
\hline
\hline
$\pth$ & 0.4945 & 0.544 & 0.5939 & 0.6406 & 0.69 & 0.74 & 0.7872 & 0.8543 \\
\hline
$\delta$ & 0.011 & 0.0109 & 0.0102 & 0.0145 & 0.0143 & 0.0133 & 0.016 & 0.0143 \\
\hline
\hline
$\delta_{\text{rand}}$ & 0.011 & 0.0234 & 0.0284 & 0.0266 & 0.0251 & 0.0213 & 0.0172 & 0.0143 \\
\hline
\end{tabular}
\vspace{-2mm}
\end{table}
Finally, Figure \ref{fig:flp_opt_ext_distr_irreg} presents the Bit Erasure Rate (BER) performance of optimized
extending distributions for finite code lengths. All the codes have binary dimension (number of source bits) equal to
5000 bits (1250 $\f_{16}$-symbols). The mother code with rate $1/2$ has been constructed by using the Progressive Edge
Growth (PEG) algorithm \cite{PEG}, and symbol nodes have been extended according to the optimized distributions
(extension matrices $A_n$ being chosen such as to maximize $\dmin(A_n)$).
\section{Conclusions}\label{sec:conclusions}
Based on the extended binary image of NB-LDPC codes, we presented a coding technique that allows to produce extra
redundant bits, such as to decreases the coding rate of a mother code. The proposed method allows for using the same
decoder as for the mother code: extra redundant bits transmitted over the channel are only used to ``improve the
quality of the decoder input''.
Extending distributions for regular and irregular codes have been analyzed by using simulated density evolution
thresholds of extended codes over the BEC. We have also presented optimized extending distributions, which exhibit a
normalized gap to capacity $\delta\approx 0.01$, for extended rates from $0.45$ to $2/15$
Finally, although this paper dealt only with NB-LDPC codes over the BEC, the results presented here can be easily
generalized to different channel models.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{finite_length_deopt
\vspace{-2mm}
\caption{Finite length performance of optimized extended NB-LDPC codes
\label{fig:flp_opt_ext_distr_irreg
\vspace{-3mm}
\end{figure}
\bibliographystyle{IEEEbib}
|
\section{Introduction}\label{sec:intro}
There is overwhelming evidence that the Universe is filled with nonbaryonic dark matter (DM). However, despite the wealth of cosmological evidence for its existence, its particle physics properties remain a complete mystery. Among the many candidates proposed, weakly interacting massive particles (WIMPs) are among the best motivated, due to the possibility of understanding their abundance as thermal relics and potential connection to electroweak symmetry-breaking. The hunt is on to discover dark matter through direct observation of its scattering with nuclei, indirectly through its annihilation into high energy particles, and through its production at colliders.
This is a particularly exciting time for direct detection experiments. Detectors are achieving unprecedented sensitivity, and will improve their discovery reach by orders of magnitude. A positive signal at such an experiment would be our first hint that dark matter has interactions beyond gravitational, and would confirm the WIMP picture of dark matter and open the door to improving our quantitative understanding of the early Universe through its relic density. In fact, the experimental situation is somewhat confusing. DAMA/LIBRA \cite{Bernabei:2010mq} and CoGeNT
\cite{Aalseth:2010vx} both report signals which have defied explanation in terms of conventional backgrounds to WIMP searches, but XENON \cite{Aprile:2010um}, and CDMS \cite{Ahmed:2009zw} seem to exclude some or all of the parameter space necessary to explain them.
One of the key properties which distinguish different visions for the nature of dark matter are its spin and the nature of its interactions with Standard Model (SM) particles. The most common assumption is that heavy particles (including perhaps the SM $Z$ and Higgs bosons) mediate the interactions, resulting in interactions which are essentially contact interactions at the length scales probed by the direct searches. However, it remains possible that dark matter interacts with ordinary matter by exchanging photons. To remain dark, it must be electrically neutral, so the coupling to photons must be described by a higher multipole interaction. In particular, a Dirac fermion WIMP could couple to photons through an electric or magnetic dipole moment, and this could provide the dominant mechanism for scattering off heavy nuclei either elastically
\cite{Raby:1987ga,Bagnasco:1993st,Sigurdson:2004zp,Masso:2009mu,Barger:2010gv,Fitzpatrick:2010br,Banks:2010eh} or inelastically \cite{Chang:2010en}.
In Ref.~\cite{Banks:2010eh} it was found that dipole-moment interacting dark matter (DMDM) with mass around $10$ GeV and magnetic gyromagnetic ratio around $0.02$ (i.e. with magnetic dipole moment around $10^{-17}$ $e$-cm) is (marginally) consistent with results from direct detection experiments including CDMS-II \cite{Ahmed:2009zw}, XENON100 \cite{Aprile:2010um}, DAMA \cite{Bernabei:2010mq}, and CoGeNT \cite{Aalseth:2010vx}.
As part of this work, we update these bounds to include the recent CDMS low threshold analysis \cite{Ahmed:2010wy} and find that it excludes the CoGeNT and DAMA regions for elastic scattering at the $90\%$ CL. Nonetheless, the electromagentic dipole moment is an interesting portal for a WIMP to interact with the Standard Model, and it is an important part of the over-all picture of dark matter interactions to understand the bounds from existing experiments as well as future prospects for discovery.
A final important question is whether or not DMDM can realize the WIMP(less) miracle \cite{Feng:2008ya}. Early proposals for DMDM had large dipole moments by virtue of the WIMP itself being composite \cite{Banks:2005hc}. The dark matter relic abundance was generated non-thermally through strong dynamics, and ultimately determined by a primordial asymmetry between WIMPs and anti-WIMPS. This is an attractive picture, but the magnetic Land\'{e} factor consistent with the direct detection experiments is small enough to suggest that the dark sector is weakly coupled. This opens the possibility that the dark matter may be a more standard thermal relic, and it would be interesting to see how the parameter space consistent with a thermal relic is related to that relevant for direct detection experiments.
In this article, we build upon Refs \cite{Bagnasco:1993st,Sigurdson:2004zp,Masso:2009mu,Barger:2010gv,Banks:2010eh} to explore DMDM from the points of view of the thermal relic density and in light of constraints from both LEP and hadron colliders \cite{Goodman:2010yf,Goodman:2010ku,Bai:2010hh,Cao:2009uw}. It is organized as follows: Section~\ref{sec:DMDM} briefly reviews the theory of DMDM. Section~\ref{sec:Abundance} computes the thermal relic density of DMDM and identifies the parameter space consistent with a thermal relic. Section \ref{sec:Collider} discusses collider bounds on DMDM from LEP and the Tevatron as well as the reach of the LHC. Finally Section \ref{sec:conc} summarizes the DMDM bounds from direct detection experiments \cite{Banks:2010eh}, the DMDM thermal relic density constraint for weakly-coupled dark sectors and the DMDM bounds from colliders.
\section{Dark Matter with Dipole Moments}\label{sec:DMDM}
We work in an effective theory framework, assuming that the only relevant degree of freedom of the dark sector at the energies of interest is the WIMP itself. The WIMP is a Dirac fermion DM $\psi$ of mass $m_{\rm DM}$ which interacts with the SM through electromagnetic dipole moments \cite{Bagnasco:1993st,Sigurdson:2004zp,Masso:2009mu,Barger:2010gv,Banks:2010eh},
\begin{equation}\label{eqn:effectiveLagrangian}
\mathcal{L}_{\rm DM}=
\bar{\psi}(i\gamma^\mu\partial_\mu-m_{\rm DM})\psi
+\frac{g_Me}{8m_{\rm DM}}\bar{\psi}\sigma^{\mu\nu}\psi F_{\mu\nu}
+\frac{g_Ee}{8m_{\rm DM}}\bar{\psi}\sigma^{\mu\nu}\psi\widetilde{F}_{\mu\nu}
\end{equation}
where $g_M$ and $g_E$ are the DM magnetic and electric gyromagnetic ratios, and $F_{\mu \nu}$ and $\widetilde{F}_{\mu\nu}$ are the electromagnetic (dual) field strengths. Along with the Higgs portal operator $|H|^2 \bar{\psi} \psi$ (considered elsewhere \cite{Kanemura:2011nm}), the magnetic and electric dipole operators are the gauge-invariant dimension $5$ operators coupling the DM to the SM and are thus among the most relevant operators mediating SM-DM interaction.
The electric dipole moment operator violates the discrete symmetries P and T. In models where these symmetries are spontaneously broken, the electric dipole operator is effectively of dimension $5+n$ operator where $n\geq1$. The magnetic and electric Land\'{e} factors are related to the DM scale $\Lambda_{\rm DM}$ and CP-violating scale $\Lambda_{\rm CP}>\Lambda_{\rm DM}$ through $g_M\sim4m_{\rm DM}/\Lambda_{\rm DM}$ and $g_E\sim4m_{\rm DM}/\Lambda_{\rm DM}(\Lambda_{\rm DM}/\Lambda_{\rm CP})^n$ respectively.
In terms of the parameters of $g_M$ and $g_E$, the DMDM has a magnetic dipole moment $\mu_{\rm DM}=g_Me/4m_{\rm DM}$ and an electric dipole moment
$d_{\rm DM}=g_Ee/4m_{\rm DM}$. Naive dimensional analysis suggests that fermionic DM Land\'{e} factors $g_M$ and $(\Lambda_{\rm CP}/\Lambda_{\rm DM})^ng_E$ are order one in theories with a strongly-coupled dark sector, and loop-suppressed when the dark sector is weakly-coupled.
\section{Relic Abundance}\label{sec:Abundance}
In this section we examine the thermal relic density of DMDM in theories with a weakly-coupled dark sectors. We perform the thermal relic density computation under the assumption that there is no primordial asymmetry between $\psi$ and $\bar{\psi}$. At freeze-out, the WIMP is non-relativistic for the masses of interest, and its relic density is determined from its annihilation cross section into Standard Model particles as usual \cite{Kolb:1988aj}.
The annihilation cross sections for DMDM to SM particles at lowest order in the relative velocity $v_{\rm rel}$ and neglecting the fermion masses are given by\footnote{Note that the work of \cite{Masso:2009mu} does not take into account annihilation into photons.}
\begin{equation}\label{eqn:DM-SM}
\begin{array}{rclcrcl}
\sigma_{\bar{\psi}\psi\rightarrow\bar{f}f}^{\rm MDM}v_{\rm rel} &=&
\frac{\pi(g_MQ_f\alpha)^2}{4m_{\rm DM}^2}
&\hspace{2cm}& \sigma_{\bar{\psi}\psi\rightarrow\bar{f}f}^{\rm EDM}v_{\rm rel} &=&
\frac{\pi(g_EQ_f\alpha)^2}{48m_{\rm DM}^2}v_{\rm rel}^2\vspace{0.2cm}\\
\sigma_{\bar{\psi}\psi\rightarrow\gamma\gamma}^{\rm MDM}v_{\rm rel} &=&
\frac{\pi(g_M^2\alpha)^2}{64m_{\rm DM}^2}
&\hspace{2cm}& \sigma_{\bar{\psi}\psi\rightarrow\gamma\gamma}^{\rm EDM}v_{\rm rel} &=&
\frac{\pi(g_E^2\alpha)^2}{64m_{\rm DM}^2}
\end{array}
\end{equation}
where $Q_f$ is the fermion charge. In the limit of zero SM fermion mass, magnetic-dipole interacting dark matter has predominantly $s$-wave annihilation into SM fermions, while the leading electric-dipole interaction is $p$-wave. Annihilation into photons is $s$-wave for both. We neglect annihilation into a pair of $W$ bosons and/or top quarks, which will correct these expressions for $m_{\rm DM} \gtrsim 80$~GeV, but as we will see, the possibility of a thermal relic for this range of masses is already excluded by direct detection experiments.
The thermally averaged annihilation cross section $\sigma_A$ can be written,
\begin{equation}\label{eqn:sigmaA}
\langle\sigma_Av_{\rm rel}\rangle=
\sum_{m_f<m_{\rm DM}}\langle\sigma_{\bar{\psi}\psi\rightarrow\bar{f}f}v_{\rm rel}\rangle
+\langle\sigma_{\bar{\psi}\psi\rightarrow\gamma\gamma}v_{\rm rel}\rangle
\equiv \sigma_0
\left[1+b\left(\frac{T}{m_{\rm DM}}\right
\right]
\end{equation}
where $\sigma_0$ and $b$ are extracted in terms of $g_M$ or $g_E$ and $m_{\rm DM}$ from Eq.~(\ref{eqn:DM-SM}) (and $b=0$ for magnetic interactions).
To good approximation, the freeze-out temperature $T_f$ is given by \cite{Kolb:1988aj}
\begin{eqnarray}\label{eqn:Tf}
x_f=\frac{m_{\rm DM}}{T_f} &=& \ln[0.038~(g/g_*^{1/2})m_{\rm Pl}m_{\rm DM}\sigma_0]\nonumber\\
&& -\frac{1}{2} \ln \{ \ln[0.038~(g/g_*^{1/2})m_{\rm Pl}m_{\rm DM}\sigma_0]\}\nonumber\\[0.2cm]
&& +\ln(1+b\{\ln[0.038(g/g_*^{1/2})m_{\rm Pl}m_{\rm DM}\sigma_0]\}^{-1}),
\end{eqnarray}
where $m_{\rm Pl}$ is the Planck mass, $g=2$ is the number of degrees of freedom of the DM and $g_*$ is the effective number of relativistic degrees of freedom at the time of decoupling. The DMDM relic density is then \cite{Kolb:1988aj,Servant:2002aq}
\begin{equation}\label{eqn:Omegahsq}
\Omega_{\rm DM}h^2=
\frac{1.07\times 10^9 ~{\rm GeV}^{-1}}{m_{\rm Pl}} ~
\frac{x_f}{\sigma_0 \left[1+ 3 b x_f^{-1} \right](g_{*S}/g_*^{1/2})}
\end{equation}
where $g_{*S}$ is the effective number of degrees of freedom relevant for the entropy at the time of decoupling. At the time of freeze-out, all particle species have a common temperature and $g_{*S}=g_*$. To saturate the observed DM relic density, DMDM which is a thermal relic should satisfy $\Omega_{\rm DM}h^2=0.1120\pm0.0056$ \cite{Komatsu:2010fb}. The values of $g_M$ ($g_E$) for which the correct thermal relic density is obtained for a give WIMP mass are plotted in Figures~\ref{fig:90pclm} and~\ref{fig:90pcle}, respectively.
\section{Collider Constraints}\label{sec:Collider}
DMDM can also be produced at colliders, which may provide complimentary information to the direct detection experiments \cite{Banks:2010eh}. The analysis is somewhat similar to studies which probe contact interactions between WIMPs and quarks and gluons \cite{Goodman:2010yf,Goodman:2010ku,Bai:2010hh} or leptons \cite{Fox:2011fx},\footnote{For example, \cite{Fox:2011fx} investigates WIMP interactions through heavy mediators with leptons, whereas we consider WIMPs interacting through photons.} but with the added feature that the mediator particle is now the massless photon. We examine the implications of collider searches for photon plus missing energy and jet plus missing energy final states on DMDM.
Any theory with non-renormalizable interactions must be understood as an effective theory, breaking down at some finite energy scale. Firm predictions can only be made at energies well below this scale. In the case of DMDM, theories with a weakly coupled dark sector such that $g_M, g_E \ll 1$ have robust predictions for WIMP production at the LHC. For a strongly coupled dark sector such as a composite WIMP, the momentum transfer will be of order the compositeness scale, which could invalidate bounds (and also could potentially lead to a variety of other interesting signals).
\subsection{LEP II Constraints}\label{subsec:Lepton}
In lepton colliders, production of $\psi \bar{\psi}$ (missing transverse energy in the detector) plus a photon can lead to an observable signal,
\begin{equation*}
e^-+e^+\rightarrow\bar{\psi}+\psi+\gamma.
\end{equation*}
The photon can be produced either as initial state radiation from the incoming leptons, or directly from the DMDM final state through is dipole moment. Initial state radiation enjoys a large collinear singularity which causes the cross section to grow as $\ln E_\gamma / m_e$ when the photon is approximately collinear with the incoming $e^\pm$ beams. Thus, it completely dominates the final state radiation (which we neglect). The center-of-mass production cross section for this process is given in the massless electron limit by
\begin{eqnarray*}
\frac{d\sigma_\gamma}{dx_{E_T^\gamma}d\eta_\gamma} &=& \frac{g^2\alpha^3}{12s}\sqrt{1-\frac{4x_{m_{\rm DM}}^2}{1-2x_{E_T^\gamma}\cosh(\eta_\gamma)}}\\
&& \times\frac{[1-2x_{E_T^\gamma}\cosh(\eta_\gamma)+2(1+3\sigma)x_{m_{\rm DM}}^2][1-2x_{E_T^\gamma}\cosh(\eta_\gamma)+x_{E_T^\gamma}^2\cosh(2\eta_\gamma)]}{x_{E_T^\gamma}x_{m_{\rm DM}}^2[1-2x_{E_T^\gamma}\cosh(\eta_\gamma)]}
\end{eqnarray*}
where $\sqrt{s}$ is the center-of-mass energy, $E_T^\gamma=x_{E_T^\gamma}\sqrt{s}$ is the photon transverse energy, $\cos(\theta_\gamma)=\tanh(\eta_\gamma)$ is the photon polar angle, and $m_{\rm DM}=x_{m_{\rm DM}}\sqrt{s}$. For DMDM with magnetic dipole moment $g=g_M$ and $\sigma=1$ while for DMDM with electric dipole moment $g=g_E$ and $\sigma=-1$. Note that the full cross section with massive electrons is used in generating the plots.
A search for this signal was performed using the L3 detector based on several different LEP II collider energies \cite{Achard:2003tx}. The most significant SM background is production of a $Z$-boson (decaying to neutrinos, and this producing missing transverse energy in the final state)
together with initial state radiation. Events with missed particles or particles with mismeasured energies also contribute a ``fake" background. Of these, Bhabha events where both charged leptons are lost turn out to be the most important \cite{Achard:2003tx}.
L3 considered two event topologies containing one final state photon \cite{Achard:2003tx}:
\begin{itemize}
\item High energy single photon:
\begin{itemize}
\item[-]$E_T^\gamma>0.02\sqrt{s}$;
\item[-]$14^\circ<\theta_\gamma<166^\circ$;
\item[-]No other photon with $E_\gamma>1$ GeV.
\end{itemize}
\item Low energy single photon:
\begin{itemize}
\item[-]$0.008\sqrt{s}<E_T^\gamma<0.02\sqrt{s}$;
\item[-]$43^\circ<\theta_\gamma<137^\circ$;
\item[-]No other photon with $E_\gamma>1$ GeV.
\end{itemize}
\end{itemize}
The data set is divided in $8$ subsets with specific luminosities and center-of-mass energies $\sqrt{s}$ between $189$ and $209$ GeV. The signal efficiencies vary between $69.8\%$ and $73.7\%$ across these data sets. We choose a flat efficiency of $71\%$ close to the median value for all energies. We simulate events and use the LEP bound to derive limits on the cross section. To be conservative, we derive $90\%$ CL bounds from the single event topology and data subset that leads to the best constraints, and do not attempt to combine the various datasets.
Our results are summarized on Figures~\ref{fig:90pclm} and \ref{fig:90pcle}, from which we see that bounds from LEP II are effective up to about the kinematic limit of $m_{\rm DM} \lesssim 100$~GeV. For both types of interaction, the LEP limits do better than those from direct detection for masses less than about 5 GeV, but in neither case is LEP able to probe couplings small enough to explain the thermal relic density.
\subsection{Hadron Collider Constraints}\label{subsec:Hadron}
In hadron colliders, fermionic DMDM can be searched for in photon plus missing transverse energy and jet plus missing transverse energy final states. We focus here on the jet plus missing transverse energy final state
\begin{equation*}
p^++p^\pm\rightarrow\bar{\psi}+\psi+{\rm jet}.
\end{equation*}
Photon plus missing transverse energy final states are interesting but the rate is suppressed by the electromagnetic coupling $\alpha$ compared to the jet plus missing transverse energy process which is proportional to the strong coupling $\alpha_S$. In estimating our signal rates, we use MSTW parton distribution functions (PDFs) \cite{Martin:2009iq}. The center-of-mass production cross section for this process, in which the final jet can originate from a quark(anti-quark) or a gluon at the parton level, is given in the massless parton limit by
\begin{eqnarray*}
\frac{d\sigma_{q-{\rm jet}}}{dx_{E_T^{\rm jet}}d\eta_{\rm jet}} &=& C_qQ^2\frac{g^2\alpha^2\alpha_S}{24s}\sqrt{1-\frac{4x_{m_{\rm DM}}^2}{1-2x_{E_T^{\rm jet}}\cosh(\eta_{\rm jet})}}\\
&& \times\frac{[1-2x_{E_T^{\rm jet}}\cosh(\eta_{\rm jet})+2(1+3\sigma)x_{m_{\rm DM}}^2]}{x_{m_{\rm DM}}^2[1-2x_{E_T^{\rm jet}}\cosh(\eta_{\rm jet})]\cosh(\eta_{\rm jet})[1+\tanh(\eta_{\rm jet})]}\\
&& \times\{1-2x_{E_T^{\rm jet}}\cosh(\eta_{\rm jet})[1-\tanh(\eta_{\rm jet})]\\
&& \hspace{1cm}+x_{E_T^{\rm jet}}^2\cosh^2(\eta_{\rm jet})[5-2\tanh(\eta_{\rm jet})+\tanh^2(\eta_{\rm jet})]\}\\
\frac{d\sigma_{g-{\rm jet}}}{dx_{E_T^{\rm jet}}d\eta_{\rm jet}} &=& C_gQ^2\frac{\alpha_S}{\alpha}\frac{d\sigma_\gamma}{dx_{E_T^{\rm jet}}d\eta_{\rm jet}}
\end{eqnarray*}
where again $\sqrt{s}$ is the center-of-mass energy, $E_T^{\rm jet}=x_{E_T^{\rm jet}}\sqrt{s}$ is the jet transverse energy, $\cos(\theta_{\rm jet})=\tanh(\eta_{\rm jet})$ is the jet polar angle, $m_{\rm DM}=x_{m_{\rm DM}}\sqrt{s}$, $Q$ is the initial parton electric charge and $C_q,C_g$ are the appropriate color factors. As usual $g=g_M$ and $\sigma=1$ for DMDM with magnetic dipole moment while $g=g_E$ and $\sigma=-1$ for DMDM with electric dipole moment.
The most significant SM background comes from $Z$-bosons decaying to neutrinos with associated jet production. Background also includes events with missed particles or particles with mismeasured energies. Of these, events leading to $W$-boson plus jets where the charged lepton from the $W$-boson decay is missed are the most important \cite{Aaltonen:2008hh,CDF}.
\subsubsection*{Tevatron $90\%$ Confidence Limits}
Both D0 and CDF searched for the jet plus missing transverse energy final state, inspired by models with large extra dimensions. We compare with the CDF \cite{Aaltonen:2008hh,CDF} search, which is based on a significantly larger integrated luminosity and thus provides the strongest limits.
The CDF event selection included \cite{CDF}:
\begin{itemize}
\item[-]$E_T^{\rm jet}>80$ GeV;
\item[-]$|\eta_{\rm jet}|<1.1$;
\item[-]$\not{\!\!E_T}>80$ GeV;
\item[-]At most one extra jet with $E_T<30$ GeV;
\item[-]Any additional jet with $E_T>20$ GeV vetoed.
\end{itemize}
The SM predicted $8663$ events in the $1$ fb$^{-1}$ of data at $\sqrt{s}=1.96$ TeV studied while only $8449$ events were observed by CDF. The signal event efficiency was found in \cite{Goodman:2010ku} to be roughly $40\%$, rather independent of the WIMP mass. From Figures~\ref{fig:90pclm} and \ref{fig:90pcle}, we see that bounds from the Tevatron are weaker than those from LEP II for masses smaller than about 100 GeV, where direct detection experiments are providing the best limits, several orders of magnitude better than those from colliders.
\subsubsection*{LHC $5\sigma$ Reach}
Predictions for the jet plus missing transverse energy final state was studied at the LHC in \cite{Vacavant:2001sd} (also in the context of large extra dimensions). The study \cite{Vacavant:2001sd} selected events with:
\begin{itemize}
\item[-]$E_T^{\rm jet}>100$ GeV;
\item[-]$|\eta_{\rm jet}|<3.2$;
\item[-]$\not{\!\!E_T}>500$ GeV.
\end{itemize}
At center-of-mass energy $\sqrt{s}=14$ TeV, $100$ fb$^{-1}$ of data lead to an expected number of background events of approximatively $B=2\cdot10^4$. To compare with these bounds we follow the results of \cite{Goodman:2010ku} which found an efficiency of $80\%$. We define a $5\sigma$ discovery region in which DMDM can be discovered in end-stage LHC running.
\section{Discussion and Conclusion}\label{sec:conc}
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.60]{Figure-90pclMDMwithLowThCDMS2to5.eps}
\end{center}
\caption{$90\%$ confidence level (CDMS-II, XENON100, DAMA, CoGeNT, LEP, Tevatron) and $5\sigma$ reach (LHC) plots for direct detection and collider experiments for DMDM with magnetic dipole moment. The dash line corresponds to the $90\%$ confidence level plot for the low threshold CDMS analysis.}
\label{fig:90pclm}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.60]{Figure-90pclEDMwithLowThCDMS2to5.eps}
\end{center}
\caption{$90\%$ confidence level (CDMS-II, XENON100, DAMA, CoGeNT, LEP, Tevatron) and $5\sigma$ reach (LHC) plots for direct detection and collider experiments for DMDM with electric dipole moment. The dash line corresponds to the $90\%$ confidence level plot for the low threshold CDMS analysis.}
\label{fig:90pcle}
\end{figure}
We have considered a dark matter particle which interacts with the Standard Model through an electric or magnetic dipole moment. This is an interesting portal connecting the dark sector to ordinary matter, mediated by the massless photon. As such, it is also one of the most challenging cases for colliders, because its rate drops rapidly with the mass of the WIMP.
We have considered updated direct detection bounds from CDMS, the thermal relic density, and collider constraints from LEP II and the Tevatron. We have also considered the long-term LHC prospects for a discovery in the channel of $\psi \bar{\psi} +$~jet. Our results are summarized in Figures~\ref{fig:90pclm} and \ref{fig:90pcle}.
For masses between a few to 100 GeV, direct detection constraints are already quite strong \cite{Banks:2010eh}, somewhat stronger on the electric dipole moment than on the magnetic dipole moment. This is a consequence of the low energy enhancement of the electric dipole moment-mediated scattering (see, for example, \cite{Banks:2010eh}), and results in direct detection experiments discriminating between the two. Conversely, at colliders the dark matter is produced at high energy, and the rates are comparable for both interaction types. A thermal relic is consistent with all constraints only for masses of order a few GeV, and in the case of a magnetic dipole interaction, would be consistent with the regions favored by the DAMA and CoGeNT signals, had that same region not been (marginally) excluded by the low threshold CDMS analysis. It is interesting that light dark matter with a dipole moment induced at the loop level seems to pass all of the usual criteria to be a good WIMP\footnote{We note in passing that it is difficult to explain the PAMELA \cite{:2008zzr} positron excess. To explain PAMELA the total annihilation cross section must satisfy $\langle\sigma_Av_{\rm rel}\rangle\gtrsim10^{-23}$ cm$^3$/s. For $m_{\rm DM}\lesssim50$ GeV this translates into $g_M\gtrsim0.05m_{\rm DM}/$GeV and $g_E\gtrsim0.76\sqrt{m_{\rm DM}/{\rm GeV}}$. Since LEP requires $g_{M,E}\lesssim0.05m_{\rm DM}/$GeV for $m_{\rm DM}\lesssim50$ GeV, the former is tightly constrained by LEP while the latter is excluded by LEP.}. At larger masses, a non-thermal production mechanism such as could occur in theories with a strongly-coupled dark sector \cite{Banks:2005hc} must be invoked.
Collider bounds are typically somewhat weaker than the direct detection bounds, which are also stronger than typical indirect detection bounds from annihilation into $\gamma \gamma$ \cite{Sigurdson:2004zp,Goodman:2010qn}. Although, due to a cleaner environment, the LEP bounds are stronger by an order of magnitude compared to the Tevatron bounds, both bounds are barely consistent with a weakly-coupled theory. The LHC discovery prospects in the jets + missing energy channel are not very promising, being already excluded by LEP II at low masses, and direct detection experiments at large masses. Of course, with its large center-of-mass energy, one could hope that the LHC could also directly produce even heavier states in the dark sector. Their signatures are model-dependent, but there is much potential for discovery of states with higher production rates and/or more striking experimental signals with lower backgrounds.
Dark matter coupled through the photon portal is an interesting and motivated vision which leads to a somewhat different perspective from the standard WIMP. Understanding its allowed parameter space is an important step in its search.
\subsection*{Acknowledgments}
We would like to thank Kunal Kumar, Patipan Uttayarat, and especially Wouter Waalewijn for valuable discussions. The research of JFF was supported in part by DOE grant DOE-FG03-97ER40546. The research of TT was supported in part by NSF grant PHY-0970171 and in addition he gratefully acknowledges the hospitality of the SLAC theory group, where part of this work was completed.
|
\section{Introduction}
Constructions of fundamental domains of generalized modular groups usually
rely on geometric considerations. By considering the different
possible symmetry transformations acting on some generalized upper-half plane,
the precise shape of the fundamental domain is narrowed down step-by-step until
one arrives at its final shape. Especially for higher rank groups (such
as $SL_n(\mathbb{Z})$) this poses a considerable computational and combinatorial
problem since one has to consider a large number of possible successive symmetry transformations (already the determination of the fundamental domain of the standard
modular group $PSL_2(\mathbb{Z})$ along these lines takes more than two pages of computations,
see e.g. \cite{Apo90}). Although one can show that the precise shape of the fundamental
domain can be determined within a finite number of steps, in the actual computation of a
domain it is not always clear how many steps are actually necessary.
In this paper we show that, at least for modular groups arising as (even) Weyl groups
of certain hyperbolic Kac--Moody algebras, such cumbersome constructions can be
altogether avoided. More specifically, we present an easy method for obtaining the
complete geometric information about the associated fundamental domains. All we
require as information for determining the explicit shape and volume is the Cartan
matrix of the corresponding Kac-Moody algebra and its Coxeter labels. As we will
demonstrate this construction works for {\em all} hyperbolic Kac--Moody algebras
\footnote{An indefinite Kac--Moody algebra is called {\em hyperbolic} if the removal of
any one node from its Dynkin diagram leaves an algebra which is either affine
or finite \cite{Kac90}.}
$\mak{g}^{++}$ of over-extended type, which are generally obtained by extending a
given finite dimensional simple Lie algebra $\mak{g}$ via its affine extension
$\mak{g}^+$ by adding two nodes to the Dynkin diagram in a specified way. Likewise
it applies to the {\em twisted} algebras obtained by inverting the arrows in the Dynkin
diagram, because their Weyl groups are the same (but note that these twisted algebras,
while being indefinite Kac--Moody algebras, in general are not of over-extended type).
In particular, our construction also applies to those hyperbolic Kac--Moody algebras
whose even Weyl groups can be identified with generalized modular groups
defined over rings of integers in division algebras \cite{FKN09}. The first
example of such an identification was given in \cite{FF} where it was shown that the
rank-3 hyperbolic Kac--Moody algebra $A_1^{++}$ (also denoted $AE_3$ or $\cF$ in
the literature) has the usual modular group $PSL_2(\mathbb{Z})$ as its even Weyl group,
the full Weyl group being $W(A_1^{++}) = PGL_2(\mathbb{Z})$. In \cite{FKN09} more
complicated examples were given, involving for instance the quaternionic integers
(Hurwitz numbers), and admitting a M\"obius-like realization \cite{KNP10}.
The most interesting (and most complicated) example is the even Weyl group
$W^+(E_{10})$ which can be identified with the arithmetic group $PSL_2(\mathtt{O})$
(where $\mathtt{O}$ are octonionic integers, also called {\it octavians}). For this
example we will explicitly display the coordinates of the vertices of the
fundamental domain of the Weyl group.
Knowledge of the shape of the fundamental domain allows one to compute its volume.
In the non-linear realization of the hyperbolic Weyl group on some generalized upper
half plane \cite{KNP10} (a hyperbolic space of constant negative curvature) the fundamental domains are realized as higher dimensional simplices. We present a very simple general
formula for the volume of the domain in terms of integrals involving a quadratic form
which contains all the information about the Lie algebra $\mak g^{++}$ (see (\ref{main})
below). We note that our considerations would also apply to cases where analogs of
the so-called congruence subgroups of $PSL_2(\mathbb{Z})$ can de defined: the volume
is then simply a multiple of the original volume, with the factor equal to the index of the congruence subgroup in the given generalized modular group. Such congruence subgroups
presumably do exist for the generalized arithmetic groups studied in \cite{FKN09},
but we are not aware of any concrete results along these lines.
As an historic aside, we mention that
the first computation of hyperbolic volumes in terms of the dihedral angles of
the simplex under consideration is due to one of the inventors of hyperbolic
geometry, N.I. Lobachevsky \cite{Lob04}. His results were extended by Schl\"afli
and Coxeter \cite{Cox35}, see also Vinberg \cite{Vin93b}. Further work on this problem
can be found in \cite{JKRT99} which gives a list of numerical values for the volumes of
hyperbolic Coxeter simplices, as well as analytical expressions for some special cases.
Using (\ref{main}) these values can be easily reproduced. We also note that in the physical context, these Coxeter simplices appear in the cosmological billiards setting, see \cite{KKN09,Koe11} for the implications of the quantum treatment of the cosmological billiards for an initial spacelike singularity.
\section{Hyperbolic roots and weights}
Let $\mak{g}$ be a finite-dimensional Lie algebra. We denote the simple roots
of $\mak g$ by ${\mathbf{a}}_i\in\mathbb{R}^n$ and their associated fundamental weights by ${\boldsymbol{\lambda}}_i$,
where $i=1,\ldots ,n$ with $n=\mathrm{Rank}(\mak g)$ (see e.g. \cite{Hum72} for
details). With the Cartan matrix of $\mak g$
\begin{eqnarray}
A_{ij} = \langle {\mathbf{a}}_i | {\mathbf{a}}_j\rangle \equiv \frac{2 {\mathbf{a}}_i \cdot {\mathbf{a}}_j}{{\mathbf{a}}_j\cdot {\mathbf{a}}_j}
\end{eqnarray}
we define the {\em symmetrized Cartan matrix} $B_{ij}$ as
\begin{eqnarray}
B_{ij} \equiv (AD)_{ij} =2 {\mathbf{a}}_i \cdot {\mathbf{a}}_j = A_{ij} \, {\mathbf{a}}_j^2 \ , \label{symmcart}
\end{eqnarray}
where ${\mathbf{a}}_j^2\equiv {\mathbf{a}}_j\cdot {\mathbf{a}}_j$ and there is no summation
over double indices. Unlike $A_{ij}$, the matrix $B_{ij}$ and the symmetrizing
matrix $D_{ij} = \delta_{ij} {\mathbf{a}}_j^2$ depend on the normalization of ${\mathbf{a}}_j$. Following \cite{FKN09} we choose this normalization such that always ${\boldsymbol{\theta}}^2=1$ for
the highest root
\begin{eqnarray}
{\boldsymbol{\theta}} = \sum_{j=1}^n m_j {\mathbf{a}}_j
\end{eqnarray}
with the Coxeter labels $m_j$. When ${\boldsymbol{\theta}}$ is a {\em long} root
we therefore have ${\mathbf{a}}_j^2=1$ for the long roots.
The associated fundamental weights ${\boldsymbol{\lambda}}_j$ constitute a basis dual to the
simple roots \cite{Hum72}
\begin{eqnarray}\label{lambda}
\braket{{\boldsymbol{\lambda}}_i}{{\mathbf{a}}_j}\equiv \frac{2{\boldsymbol{\lambda}}_i \cdot {\mathbf{a}}_j}{{\mathbf{a}}_j\cdot {\mathbf{a}}_j}=\delta_{ij}
\end{eqnarray}
implying
\begin{eqnarray}
{\boldsymbol{\lambda}}_i\cdot {\mathbf{a}}_j = \frac{1}{2} \, \delta_{ij} \, {\mathbf{a}}_j^2 \ .
\end{eqnarray}
With the inverse Cartan matrix $A^{-1}$ we thus have
\begin{eqnarray}
{\boldsymbol{\lambda}}_i = \sum_k (A^{-1})_{ik} {\mathbf{a}}_k \ ,
\end{eqnarray}
from which we deduce
\begin{eqnarray}
{\boldsymbol{\lambda}}_i \cdot {\boldsymbol{\lambda}}_j = \frac{1}{2} (A^{-1})_{ij} \, {\mathbf{a}}_j^2 \ .
\end{eqnarray}
or
\begin{eqnarray}\label{lambda1}
{\boldsymbol{\lambda}}_i \cdot {\boldsymbol{\lambda}}_j = \frac{1}{2} {\mathbf{a}}_i^2 \,(B^{-1})_{ij} \, {\mathbf{a}}_j^2 \ .
\end{eqnarray}
Next we consider the {\em hyperbolic extension} $\mak g^{++}$ of the finite-dimensional
algebra $\mak g$ obtained by adjoining to the Dynkin diagram of $\mak g$ the affine node (labeled `0') and the over-extended node (labeled `$-1$'). This entails extending
the Euclidean root space $\mathbb{R}^n$ to the {\em Lorentzian} space $\mathbb{R}^{1,n+1}=\mathbb{R}^{1,1} \oplus \mathbb{R}^n$.
We denote the roots of $\mak g^{++}$ by $\alpha_I$, $I=-1,0,1,\ldots , n$, and define
them according to
\begin{eqnarray}
\alpha_{-1}\equiv -\de - \bd \ , \quad \alpha_0\equiv \de- {\boldsymbol{\theta}} \ , \quad \alpha_i\equiv {\mathbf{a}}_i
\end{eqnarray}
with the affine null vector $\de\in\mathbb{R}^{1,1}$ and the conjugate null vector $\bd\in\mathbb{R}^{1,1}$
obeying $\de \cdot \bd=\frac{1}{2}$. In this way we obtain the Cartan matrix of
${\mak g}^{++}$ as
\begin{eqnarray}
A_{IJ}= \langle \alpha_I | \alpha_J\rangle \equiv
\frac{2 \alpha_I \cdot \alpha_J}{\alpha_J \cdot \alpha_J}
\end{eqnarray}
with the {\em Lorentzian} inner product
\begin{eqnarray}
\alpha_I\cdot \alpha_J \equiv \eta_{\mu\nu} \alpha_I^\mu \alpha_J^\nu
\end{eqnarray}
where the signature of $\eta_{\mu\nu}$ is $(-+\cdots +)$.
Notice that the affine and over-extended simple roots are also
normalized as $\alpha_{-1}^2 = \alpha_0^2 =1$. The normalization
${\boldsymbol{\theta}}^2=1$ is necessary to obtain a single line between the affine
and the hyperbolic node (connecting $\alpha_0$ and $\alpha_{-1}$).
The fundamental weights $\Lambda_I$ for the hyperbolic extension
$\mak g^{++}$ are defined in analogy with (\ref{lambda})
\begin{eqnarray}\label{scalarhyp}
\braket{\Lambda_I}{\alpha_J}\equiv \frac{2\Lambda_I\cdot \alpha_J}{\alpha_J \cdot \alpha_J} = \delta_{IJ}\ .
\end{eqnarray}
By a standard construction (see e.g. \cite{DHJN01}), the fundamental weights of
$\mak g^{++}$ can be expressed in terms of the null vectors $\de$ and $\bd$
and the finite weights ${\boldsymbol{\lambda}}_j$ as
\begin{eqnarray}\label{Lambda}
\Lambda_{-1}=-\de \ , \quad \Lambda_0=\bd - \de \ , \quad \Lambda_j=n_j \Lambda_0+{\boldsymbol{\lambda}}_i \ .
\end{eqnarray}
The coefficients $n_j$ are fixed by requiring $\alpha_0\cdot \Lambda_j=0$
(cf. \eqref{scalarhyp}), which gives
\begin{eqnarray}
n_j = m_j {\mathbf{a}}_j^2 \ ,
\end{eqnarray}
The fundamental Weyl chamber $\cC_0 \subset \mathbb{R}^{1,n+1}$ is
\begin{eqnarray}\label{F1}
\cC_0 := \big\{ X\in\mathbb{R}^{1,n+1} \, | \, X \cdot \alpha_I \geq 0 \;\mbox{ for $I=-1,0,1,...,n$} \big\}
\nonumber
\end{eqnarray}
With the fundamental weights $\Lambda_I$ one obtains a more convenient
representation of $\cC_0$
\begin{eqnarray}\label{F2}
\cC_0 = \big\{ X \in \mathbb{R}^{1,n+1}\, |\, X = \sum_I s_I \Lambda_I \;\; \mbox{with $s_I\geq 0$ for
all $I$} \big\}\nonumber\\
\end{eqnarray}
The null vector $\de$ lies on the forward
light-cone in root space. The fundamental Weyl chamber $\cC_0$ is the convex hull of the hyperplanes
orthogonal to the simple roots of the algebra. The fundamental weights are vectors
pointing along the edges of $\cC_0$. In other words, $\cC_0$ is a `wedge' in
$\mathbb{R}^{1,n+1}$. For the hyperbolic algebras ${\mak g}^{++}$ of over-extended type
considered here this wedge lies inside the forward lightcone, always touching it with the
lightlike weight vector $\Lambda_{-1}$, while all other fundamental weights obey
$\Lambda^2_j\leq 0$. By contrast, for general indefinite (Lorentzian) ${\mak g}^{++}$
the fundamental Weyl chamber may stretch beyond the lightcone and also contain space-like vectors.
A schematic picture of the fundamental Weyl chamber $\cC_0$ for hyperbolic $\mak{g}^{++}$ is
shown in Fig.\,\ref{fig:kegel}. We have included the forward light-cone and the intersecting unit hyperboloid.
\begin{figure}
\begin{center}
\begin{tikzpicture}[domain=-2.84:2.84]
\filldraw[fill=gray!40!white, draw=black] (0,0) -- (-0.5,3.15) -- (2,3.23) -- (0,0);
\filldraw[fill=gray!20!white, draw=black] (0,0) -- (0.0,3.0) -- (-0.5,3.15) -- (0,0);
\filldraw[fill=gray!20!white, draw=black] (0,0) -- (0.0,3.0) -- (2,3.23) -- (0,0);
\draw (0,0) -- (-0.5,3.15) -- (2,3.23) -- (0,0);
\draw (0,0) -- (0.0,3.0) -- (2,3.23) -- (0,0);
\draw (-3.0,-3.0) -- (3.0,3.0);
\draw (-3.0,3.0) -- (3.0,-3.0);
\draw (0.0,3.0) ellipse (3 and 0.3);
\draw[color=black] (0.0,3.0) ellipse (2.84 and 0.25);
\draw[color=black] (1.35,2.15) arc (38:-259:1.71 and 0.2);
\draw[dotted,color=black] (-0.25,2.22) arc (101:43:1.71 and 0.2);
\draw[dotted] (3.0,-3.0) arc (0:180:3 and 0.3);
\draw (-3.0,-3.0) arc (180:360:3 and 0.3);
\draw[color=black] plot ( \x ,{sqrt(\x*\x+1.0)});
\draw[->] (0,-3.5) -- (0,4);
\draw[->] (-3.5,0) -- (4,0);
\draw[->] (-3,-2) -- (2.25,1.5);
\end{tikzpicture}
\caption{Sketch of the fundamental Weyl chamber $\cC_0$ as a wedge inside the forward light cone that is
intersected by the unit hyperboloid.}\label{fig:kegel}
\end{center}
\end{figure}
As it turns out the assumptions made suffice to cover all cases of interest.
This concerns in particular the {\em twisted} algebras: as these are obtained
by inverting the arrows in the relevant Dynkin diagrams, the associated
Coxeter Weyl groups, not being sensitive to the direction of the arrows,
coincide with those of the untwisted diagrams. We therefore note the following
isomorphisms of Weyl groups using Kac' notation \cite{Kac90}:
\begin{eqnarray}
W(G_2^{(1)+}) &\cong& W(D_4^{(3)+} ) \nonumber\\
W(B_n^{(1)+}) &\cong& W(A_{2n-1}^{(2)+} ) \nonumber\\
W(C_n^{(1)+}) &\cong& W(D_{n+1}^{(2)+}) \nonumber\\
W(F_4^{(1)+}) &\cong& W(E_{6}^{(2)+})
\end{eqnarray}
where the superscript $^+$ on the r.h.s. indicates the extension
of the affine algebra by another node. But note that the twisted
algebras, though perfectly well-defined as indefinite Kac--Moody
algebras, are not necessarily of over-extended type.
In the notation of Fuchs
and Schweigert \cite{FS}, the later three isomorphisms are
\begin{eqnarray}
W(B_n^{(1)+}) &\cong& W(C_{n}^{(2)+}) \nonumber\\
W(C_n^{(1)+}) &\cong& W(B_{n}^{(2)+}) \nonumber\\
W(F_4^{(1)+}) &\cong& W(F_{4}^{(2)+})
\end{eqnarray}
The corresponding volumes of the fundamental domains
therefore also coincide.
\section{Volume Formula}
The linear action of the Weyl group in $\mathbb{R}^{1,n+1}$ preserves the (Lorentzian) length,
and therefore induces a non-linear {\em modular action} on the forward unit hyperboloid
\begin{eqnarray}
X\cdot X \equiv -x^+x^-+ {\mathbf{x}}^2= - 1 \;\; , \quad x^{\pm}>0
\end{eqnarray}
with light-cone coordinates $x^\pm\equiv (x^0\pm x^{n+1})/\sqrt{2}$ in $\mathbb{R}^{1,1}$
and ${\mathbf{x}}\in\mathbb{R}^n$. For the cases $n=1,2,4$ and 8 studied in \cite{FKN09}, where the
dual of the Cartan subalgebra of $\mak g$ can be endowed with the structure of a division
algebra, the induced non-linear action takes the form of a generalized M\"obius
transformation over a (possibly non-commutative and non-associative)
ring of integers.
The intersection of the fundamental Weyl chamber $\cC_0$ with the
unit hyperboloid defines a corresponding fundamental domain $\cF_0$
on the unit hyperboloid. The corresponding domain on the (compactified)
unit hyperboloid ({\it alias} the Poincar\'e disk) is depicted in
Fig.\,\ref{fig:poincare}. In the remainder, however, we will study this domain
as a subset of the {\em generalized (Poincar\'e) upper half plane} $\cH$
rather than the unit hyperboloid \footnote{Note that the fundamental domain
$\cF_0$ is half of the fundamental domain $\cF$ of the ordinary
modular group. The latter corresponds to the {\em even} subgroup
of the Weyl group.}. This upper half plane is defined as
\begin{eqnarray}
\cH \equiv \cH_{n+1} := \big\{ ({\mathbf{u}}, v)\, | \, {\mathbf{u}}\in \mathbb{R}^n \,,\, v >0 \big\}
\end{eqnarray}
\begin{figure}
\begin{center}
\begin{tikzpicture}
\draw (0,0) circle (3);
\filldraw[fill=gray!20!white, draw=black,pattern=north east lines] (-3,0) -- (3,0) arc (90:120:5.2) -- (0,0);
\draw (-1.5,-2.598) -- (1.5,2.598);
\draw (-1.5,2.598) -- (1.5,-2.598);
\draw (3,0) arc (90:150:5.2) -- (-1.5,-2.598) arc (-30:30:5.2) -- (-1.5,2.598) arc (-150:-90:5.2) -- (3,0);
\draw (0,0) -- (3,0);
\end{tikzpicture}
\caption{Example of a fundamental domain on the Poincar\'e disk obtained by intersecting
the Weyl chamber with the (compactified) unit hyperboloid, here for the
algebra $A_1^{++}$.}\label{fig:poincare}
\end{center}
\end{figure}
and is thus of dimension $n+1$. $\cH_{n+1}$ is isometric to the forward unit hyperboloid
in $\mathbb{R}^{1,n+1}$ by means of the standard coordinate transformation
\begin{eqnarray}\label{UHP1}
x^- = \frac1{v} \;, \quad x^+ = v + \frac{{\mathbf{u}}^2}v \;, \quad {\mathbf{x}}= \frac{{\mathbf{u}}}{v}
\end{eqnarray}
The Minkowskian line element is transformed to
\begin{eqnarray}
\d s^2 = \frac{\d{\mathbf{u}}^2 + \d v^2}{v^2}
\end{eqnarray}
where, of course, $\d{\mathbf{u}}^2 \equiv \d u_1^2 + \cdots + \d u_n^2$.
The fundamental domain $\cF_0\subset\cH$ is now rather easy to determine
from the representation (\ref{F2}) by identifying the points where the rays
along $\Lambda_I$ `pierce' the unit hyperboloid, and then mapping these points
to $\cH$ by means of (\ref{UHP1}). We first notice that the over-extended fundamental
weight $\Lambda_{-1}$ ({\it alias} the affine null vector $\de$) corresponds
to the `cusp' at infinity in $\cH$ with coordinates $v=\infty\,,\, {\mathbf{u}}=0$ \cite{KNP10},
while $\Lambda_0$ corresponds to the point $v=1\,,\, {\mathbf{u}}=0$ in $\cH$. From (\ref{Lambda}) we see that the remaining fundamental weights are mapped to the points
\begin{eqnarray}\label{UHPMap}
v_j= \sqrt{ 1 - \frac{{\boldsymbol{\lambda}}_j^2}{n_j^2}} \;\; , \quad {\mathbf{u}}_j = \frac{{\boldsymbol{\lambda}}_j}{n_j}
\end{eqnarray}
on the unit hemisphere $v^2 + {\mathbf{u}}^2 =1\, ,\,v>0$ in $\cH_{n+1}$. If $|{\mathbf{u}}_j|=1$ for some
$j$ we have another cusp in addition to the cusp at infinity, but now lying
on the boundary $v=0$ of $\cH$. Therefore, the fundamental region always
has the shape of a `skyscraper' that extends to infinite height
over the simplex $\Sigma\subset\mathbb{R}^n$ defined by the points $0$ and ${\mathbf{u}}_j$,
and whose `bottom' is cut off by the unit hemisphere. See Fig.\,\ref{fig:lqs}
for an artist's view; the `bottom' of the skyscraper is the excised shaded region
on the unit sphere.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\draw[dashed] (0,0) -- (0,1.2);
\draw[->] (0,2.8) -- (0,6) node[right]{$v$};
\draw (3,0) arc (0:180:3);
\draw (-3,0) arc (-180:0:3 and 1);
\draw[dotted] (-3,0) arc (180:0:3 and 1);
\filldraw[fill=gray!20!white, draw=black, pattern=north east lines] (0,2.8) arc (110:166:2.05 and 3) -- (-1.3,0.7) arc (121.5:58.5:2.5 and 3) -- (1.3,0.7) arc (14:70:2.05 and 3);
\draw (-1.3,0.7) -- (-1.3,5.5);
\draw[dashed] (1.3,5.5) -- (1.3,6.5);
\draw[dashed] (-1.3,5.5) -- (-1.3,6.5);
\draw[dashed] (0,6) -- (0,7);
\draw (1.3,0.7) -- (1.3,5.5);\draw[dashed] (-1.3,0.7) -- (-1.3,-0.5);
\draw[dashed] (1.3,0.7) -- (1.3,-0.5);
\filldraw[fill=gray!20!white, draw=black, pattern=north east lines] (-1.3,-0.5) -- (0,0) -- (1.3,-0.5) -- (-1.3,-0.5);
\end{tikzpicture}
\caption{Schematic depiction of a Weyl chamber on the UHP, corresponding
in this case to $A_2^{++}$.}\label{fig:lqs}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{tikzpicture}[>=stealth]
\filldraw[fill=gray!20!white, draw=black, pattern=north east lines] (0,0) -- (1.5,0.866) -- (0,1.732) -- (0,0);
\draw[->,thick] (0,0) -- (3,0) node[right]{${\mathbf{a}}_1$};
\draw[->,thick] (0,0) -- (-1.5,2.598) node[above]{${\mathbf{a}}_2$};
\draw[->,thick] (0,0) -- (0,1.732) node[above]{${\boldsymbol{\lambda}}_1$};
\draw[->,thick] (0,0) -- (1.5,0.866) node[right]{${\boldsymbol{\lambda}}_2$};
\end{tikzpicture}
\caption{Schematic example for the projection of the fundamental domain
for the Weyl chamber onto the hypersurface $v=0$ for $A_2^{++}$. In
accordance with (\ref{lambda1}) the roots and weights here are normalized as
${\mathbf{a}}_1=(1,0)$, ${\mathbf{a}}_2=(-\frac{1}{2},\frac{\sqrt{3}}{2})$, and ${\boldsymbol{\lambda}}_1=(0,\frac{1}{\sqrt{3}})$,
${\boldsymbol{\lambda}}_2=(\frac{1}{2},\frac{1}{2\sqrt{3}})$.}\label{fig:projv}
\end{center}
\end{figure}
Using the above formulas we obtain
\begin{eqnarray}\label{S}
{\mathbf{u}}_i\cdot {\mathbf{u}}_j \equiv S_{ij} = \frac{1}{2m_i m_j} (B^{-1})_{ij} \ .
\end{eqnarray}
The matrix $S_{ij}$ encodes all the Lie algebraic
information about the over-extended algebra $\mak g^{++}$ via the
inverse symmetrized Cartan matrix $B^{-1}$ and the Coxeter labels $m_j$.
By a general result valid for all finite $\mak g$ \cite{Hum72} the matrices
$B^{-1}$ are positive definite; furthermore their individual
entries $B^{-1}_{ij}$ are also positive. It thus follows that
\begin{eqnarray}\label{Spos}
S > 0 \;\; \mbox{(as a matrix) and}\;\; S_{ij} > 0
\;\; \mbox{for all $i,j$}
\end{eqnarray}
Note that this formula holds for simply-laced as well as non-simply-laced
(untwisted) algebras. In particular, in the non-simply laced case one has to
distinguish between the Coxeter/dual Coxeter labels of the untwisted and
the Coxeter/dual Coxeter labels of the twisted version of the over-extension of the algebra.
As we just explained the fundamental domain $\cF_0\subset\cH$ rises over the
simplex $\Sigma \subset\mathbb{R}^n$ defined by
\begin{eqnarray}
\Sigma :=
\big\{ {\mathbf{x}}\in\mathbb{R}^n\, | \, {\mathbf{x}} =\sum\limits_{i=1}^{n} t_i {\mathbf{u}}_i\,; \ t_i\geq 0,\
\sum\limits_{i=1}^{n} t_i\leq1 \big\} \ .
\end{eqnarray}
With the above definitions we get
\begin{eqnarray}
{\mathbf{x}}(t)^2 =\sum\limits_{i,j=1}^n S_{ij}t_i t_j \ .
\end{eqnarray}
From the positivity properties (\ref{Spos}) we deduce the following
chain of inequalities valid for all points ${\mathbf{x}}(t)\in\Sigma$
\begin{eqnarray}\label{S>0}
0\leq \sum_{i,j} S_{ij} t_i t_j \,\leq\, \max_{i,j} S_{ij} \left( \sum_k t_k\right)^2
\leq \max_{i,j} S_{ij}
\end{eqnarray}
Therefore ${\mathbf{x}}(t)^2 < 1$ as long as all matrix entries satisfy $S_{ij} < 1$.
From \eqref{Lambda} it is straighforward to see that
\begin{eqnarray}
\Lambda_j^2 = n_j^2 \,\big( {\mathbf{u}}_j^2 -1 \big)
\end{eqnarray}
and it therefore follows that $S_{ii}=1$ when the corresponding
hyperbolic weight $\Lambda_i$ becomes null; for spacelike weights
($\Lambda_j^2 >0$) we have $|{\mathbf{u}}_j| > 1$, and the corresponding
point $(v_j,{\mathbf{u}}_j)$ no longer lies in the generalized upper half-plane.
This happens when $\mak g^{++}$ is Lorentzian, but no longer
hyperbolic, as is for instance the case for all $A_n^{++}$ with $n\geq 8$
and $B_n^{++}$ and $D_n^{++}$ for $n\geq 9$.
With the hyperbolic volume element
\begin{eqnarray}
\d{\rm vol} ({\mathbf{u}},v) = \frac{\d^nu\, \d
v}{v^{n+1}}
\end{eqnarray}
we thus obtain
\begin{eqnarray}
\text{vol}(\mathcal{F}_0)=\sqrt{\det S}\int\limits_{\Delta_n}
\d t_1 \cdots \d t_n\int\limits_{\sqrt{1- {\mathbf{x}}(t)^2}}^{\infty}\frac{\d v}{v^{n+1}}
\end{eqnarray}
where $\Delta_n$ is the {\em standard simplex} in $\mathbb{R}^n$
\begin{eqnarray}
\Delta_n := \big\{ (t_1,\dots,t_n) \, | \, t_i \geq 0 \, , \, \sum t_i \leq 1 \big\}
\end{eqnarray}
Performing the integral over $v$ we arrive at
\begin{eqnarray}\label{main}{}\nonumber\\
\boxed{\text{vol} (\cF_0 ) =
\frac1{n}\int_{\Delta_n} \d t_1 \cdots \d t_n
\frac{\sqrt{\det S}}{(1-\sum t_i S_{ij}t_j)^{\frac{n}{2}}} \ .}\nonumber\\
\end{eqnarray}
This simple formula is our main result: it expresses the hyperbolic volume as an integral
over a standard simplex $\Delta_n$ in $\mathbb{R}^n$ with the single matrix $S_{ij}$
encoding all the Lie algebraic information about the hyperbolic Weyl group.
The integral is manifestly convergent if all $S_{ij} <1$. When $S_{ii}=1$
the corresponding point has $|{\mathbf{u}}_i|=1$ and $v_i=0$ and thus lies on the
boundary of $\cH$, but the integral is still convergent (see below for
examples when this happens). For non-hyperbolic Lorentzian
algebras the integral diverges and therefore $\text{vol}(\cF_0) = \infty$.
When evaluating this formula it may be convenient to diagonalize
the quadratic form in terms of new integration variables $\xi_i$
such that
\begin{eqnarray}
\sum_{i,j} S_{ij} t_i t_j = \xi_1^2 + \cdots + \xi_n^2
\end{eqnarray}
and the determinant factor $(\det S)^{1/2}$ is cancelled by the Jacobian.
The variables $\xi_i$ always exist by the positivity properties of the
matrix $S$. However, the (still simplicial) domain of integration is then
more complicated to parametrize.
\section{Analytic Results}
We now show how our formula (\ref{main}) immediately yields the volumes
for various hyperbolic reflection groups corresponding to over-extended
hyperbolic algebras $\mak g^{++}$ of low rank. It is straightforward to check that
for $A_1$ (corresponding to the rank-3 Feingold-Frenkel algebra $A_1^{++}$)
we have $S=\frac12$, and one easily recovers the well known
result $\text{vol} (\cF_0 [A_1^{++}])= \frac{\pi}6$. For this reason we proceed
right away to the case of rank 4.
The rank-4 algebras of over-extended type are $A_2^{++}, C_2^{++}$ and
$G_2^{++}$. For $\mak g^{++}= A_2^{++}$ we have $m_1 = m_2 =1$ and
thus the matrix $S_{ij}$ is 1/2 the inverse of the $A_2$ Cartan matrix
\begin{eqnarray}
B^{-1} = \left( \begin{array}{cc} \frac23 & \frac13 \\[2mm]
\frac13 & \frac23 \end{array}\right)
\;\; \Rightarrow \quad
S = \left( \begin{array}{cc} \frac13 & \frac16 \\[2mm]
\frac16 & \frac13 \end{array}\right)
\end{eqnarray}
Transforming to new coordinates $\xi_1=\frac12(t_1 + t_2)$ and
$\xi_2= (1/2\sqrt{3}) (t_1 -t_2)$ such that the Jacobi determinant
cancels the factor $(\det S)^{1/2}= 1/2\sqrt{3}$ and
\begin{eqnarray}\label{b2}
\frac13 \big( t_1^2 + t_1 t_2 + t_2^2 \big) = \xi_1^2 + \xi_2^2
\end{eqnarray}
we obtain
\begin{eqnarray}
\text{vol} (\cF_0[A_2^{++}]) &=& \frac12
\int_{0}^\frac{1}{2} \d \xi_1 \int_{- \frac{\xi_1}{\sqrt{3}}}^{\frac{\xi_1}{\sqrt{3}}}
\frac{\d\xi_2}{1- \xi_1^2 - \xi_2^2} \nonumber\\
&&\hspace{-2cm}
= \frac{1}{2}\int_0^\frac{1}{2}\frac{\d \xi}{\sqrt{1-\xi^2}}\ln \left(
\frac{\sqrt{1-\xi^2}+\frac{1}{\sqrt{3}}\xi}{\sqrt{1-\xi^2} -\frac{1}{\sqrt{3}}\xi} \right)
\end{eqnarray}
The substitution $ \xi =\sin\theta$ leads to
\begin{align}
\text{vol}(\cF_0[A_2^{++}]) & =
\frac{1}{2} \int_{0}^{\frac{\pi}{6}}\d\theta \ln \left(\frac{\cos\theta+\frac{1}{\sqrt{3}}\sin\theta}{\cos\theta-\frac{1}{\sqrt{3}}\sin\theta}\right) \nonumber\\
& \hspace{-1cm}= \frac{1}{2} \int_{0}^{\frac{\pi}{6}}
\d\theta \ln \left(\frac{2\sin(\theta+\frac{\pi}{3})}{2\sin(\frac{\pi}{3} - \theta)}\right)
\end{align}
After a suitable shift of integration variables and using
the definition and properties of the Lobachevsky function,
this reduces to
\selectlanguage{russian}
\begin{eqnarray}
& \phantom{=} & \hskip-3em \text{vol}(\cF_0[A_2^{++}])\\
&=&\frac12 \left[ \CYRL\left(\frac{\pi}3\right) - \CYRL\left(\frac{\pi}6\right)
- \CYRL\left(\frac{\pi}2\right) + \CYRL\left(\frac{\pi}3\right)\right] \nonumber\\
&=& \frac14 \, \CYRL\left(\frac{\pi}3\right)
\end{eqnarray}
For $\mathfrak{g}^{++} = G_2^{++}$ we have the Coxeter labels $m_1 =2 \, , \, m_2 = 3$
and the relevant matrices are
\begin{eqnarray}
B^{-1} = \left( \begin{array}{cc} 2 & 3 \\[2mm]
3 & 6 \end{array}\right)
\;\; \Rightarrow \quad
S = \left( \begin{array}{cc} \frac14 & \frac14 \\[2mm]
\frac14 & \frac13 \end{array}\right)
\end{eqnarray}
Now the substitution to diagonalize the quadratic form is
$\xi_1=\frac12(t_1 + t_2)\, , \, \xi_2= (1/2\sqrt{3}) t_2$, and we get
\begin{eqnarray}
\text{vol} (\cF_0[G_2^{++}]) &=& \frac12
\int_{0}^\frac{1}{2} \d \xi_1 \int_0^{\frac{\xi_1}{\sqrt{3}}}
\frac{\d \xi_2}{1- \xi_1^2 - \xi_2^2} \nonumber\\
&& \hspace{-1.5cm} = \frac12 \, \text{vol} \big( \cF_0 [A_2^{++}] \big)
= \frac18 \, \CYRL\left(\frac{\pi}3\right)
\end{eqnarray}
Finally, for $C_2^{++}$ we have $m_1=m_2 = 1$ and
\begin{eqnarray}
B^{-1} = \left( \begin{array}{cc} \frac12 & \frac12 \\[2mm]
\frac12 & 1 \end{array}\right)
\;\; \Rightarrow \quad
S = \left( \begin{array}{cc} \frac14& \frac14 \\[2mm]
\frac14 & \frac12 \end{array}\right)
\end{eqnarray}
Note that the corresponding matrix for $B_2^{++}$ is
\mbox{$S = \frac14 \left(\begin{smallmatrix} 2& 1 \\ 1 & 1 \end{smallmatrix}\right)$}
and yields the same volume.
\\
Now we substitute $\xi_1 = \frac12(t_1 + t_2)$ and $\xi_2 = \frac12 t_2$ to get
\begin{eqnarray}
\text{vol} \big (\cF_0[C_2^{++}] \big) &=& \frac12
\int_{0}^\frac{1}{2} \d \xi_1 \int_0^{\xi_1} \frac{\d\xi_2}{1- \xi_1^2- \xi_2^2} \nonumber\\
&& \hspace{-1.5cm}= \frac{1}{2} \int_{0}^{\frac{\pi}{6}}
\d\theta \ln \left(\frac{2\sin(\theta+\frac{\pi}{4})}{2\sin(\frac{\pi}{4} - \theta)}\right)
\end{eqnarray}
Similar manipulations as before lead to the result
\begin{eqnarray}
& \phantom{=} & \hskip-3em \text{vol}\left( \cF_0[C_2^{++}] \right) \nonumber\\
& = & \frac14 \left[
\CYRL\left(\frac{\pi}4\right) - \CYRL\left(\frac{\pi}{12}\right) - \CYRL\left(\frac{5\pi}{12}\right) + \CYRL\left(\frac{\pi}4\right) \right] \nonumber\\
&=&\frac16 \, \CYRL\left(\frac{\pi}{4}\right)
\end{eqnarray}
\section{Higher rank algebras}
What about higher rank algebras? Although the integrals (\ref{main}) look
elementary it turns out that calculations become rapidly more complicated with
increasing dimension, and we have not been able to derive `simple' closed form expressions
for them when $n>2$. The complications are mainly due to the integration boundaries which must
be analyzed case by case. Although (\ref{main}) is suggestive of higher order Lobachevsky functions (see appendix), this expectation (as expressed, for instance, in
\cite{Vin93b}) is not borne out by the concrete calculations, nor have such expressions
been explicitly exhibited in the literature, see e.g. \cite{Vin93b}. One possibility, to
be explored in future work, would be to expand the integrand in (\ref{main}) whereby
the integral is expressed as an infinite sum of terms each one of which involves
an integral of a monomial over the standard simplex $\Delta_n$. Such integrals have been studied in the literature \cite{Bri88,BBDLKV11} but the resulting expressions are still quite
involved. Numerically these series converge rapidly, as all terms are of the same sign.
Using \eqref{main} one can compute the volume of different fundamental domains
numerically. The only input information that is needed is the matrix $S$, which is
calculated from the matrix $B^{-1}$ and the Coxeter labels via \eqref{S}. As already
mentioned above, $B^{-1}$ is the inverse symmetrized Cartan matrix and its form for the different algebras can be found in the standard Lie algebra literature, see e.g. \cite{Hum72}).
\selectlanguage{english}
\begin{longtable*}[b]{C{1cm}|C{7cm}|C{7cm}}
$G$ & Untwisted & Twisted \\ \hline\hline
\begin{minipage}[b]{1cm}
\centering $B_n$\\ \phantom{ }
\end{minipage}
&
\begin{tikzpicture}
[place/.style={circle,draw=black,fill=black,
inner sep=0pt,minimum size=6}]
\draw (0,1) -- (0,0) -- (1,0);
\draw[dashed] (1,0) -- (2,0);
\draw (2,0.1) -- (3,0.1);
\draw (2,-.1) -- (3,-.1);
\draw (-2,0) -- (0,0);
\draw (2.6,0) -- (2.4,.2);
\draw (2.6,0) -- (2.4,-.2);
\node at (-2,0) [place,label=below:$-1$] {};
\node at (0,1) [place,label=right:$1$] {};
\node at (-1,0) [place,label=below:$0$] {};
\node at (0,0) [place,label=below:$2$] {};
\node at (1,0) [place,label=below:$3$] {};
\node at (2,0) [place,label=below:$n-1$] {};
\node at (3,0) [place,label=below:$n\!\!\!\phantom{1}$] {};
\end{tikzpicture}
&
\begin{tikzpicture}
[place/.style={circle,draw=black,fill=black,
inner sep=0pt,minimum size=6}]
\draw (-1,0) -- (0,0);
\draw (0,.1) -- (1,.1);
\draw (0,-.1) -- (1,-.1);
\draw (1,0) -- (2,0);
\draw[dashed] (2,0) -- (3,0);
\draw (3,.1) -- (4,.1);
\draw (3,-.1) -- (4,-.1);
\draw (0.4,0) -- (0.6,0.2);
\draw (0.4,0) -- (0.6,-0.2);
\draw (3.6,0) -- (3.4,0.2);
\draw (3.6,0) -- (3.4,-0.2);
\node at (-1,0) [place,label=below:$-1$] {};
\node at (0,0) [place,label=below:$0$] {};
\node at (1,0) [place,label=below:$1$] {};
\node at (2,0) [place,label=below:$2$] {};
\node at (3,0) [place,label=below:$n-1$] {};
\node at (4,0) [place,label=below:$n\!\!\!\phantom{1}$] {};
\end{tikzpicture}
\\ \hline
\begin{minipage}[b]{1cm}
\centering $C_n$\\ \phantom{ }
\end{minipage} & \begin{tikzpicture}
[place/.style={circle,draw=black,fill=black,
inner sep=0pt,minimum size=6}]
\draw (-1,0) -- (0,0);
\draw (0,.1) -- (1,.1);
\draw (0,-.1) -- (1,-.1);
\draw (1,0) -- (2,0);
\draw[dashed] (2,0) -- (3,0);
\draw (3,.1) -- (4,.1);
\draw (3,-.1) -- (4,-.1);
\draw (0.6,0) -- (0.4,0.2);
\draw (0.6,0) -- (0.4,-0.2);
\draw (3.4,0) -- (3.6,0.2);
\draw (3.4,0) -- (3.6,-0.2);
\node at (-1,0) [place,label=below:$-1$] {};
\node at (0,0) [place,label=below:$0$] {};
\node at (1,0) [place,label=below:$1$] {};
\node at (2,0) [place,label=below:$2$] {};
\node at (3,0) [place,label=below:$n-1$] {};
\node at (4,0) [place,label=below:$n\!\!\!\phantom{1}$] {};
\end{tikzpicture}
&
\begin{tikzpicture}
[place/.style={circle,draw=black,fill=black,
inner sep=0pt,minimum size=6}]
\draw (0,1) -- (0,0) -- (1,0);
\draw[dashed] (1,0) -- (2,0);
\draw (2,0.1) -- (3,0.1);
\draw (2,-.1) -- (3,-.1);
\draw (-2,0) -- (0,0);
\draw (2.4,0) -- (2.6,.2);
\draw (2.4,0) -- (2.6,-.2);
\node at (-2,0) [place,label=below:$-1$] {};
\node at (0,1) [place,label=right:$1$] {};
\node at (-1,0) [place,label=below:$0$] {};
\node at (0,0) [place,label=below:$2$] {};
\node at (1,0) [place,label=below:$3$] {};
\node at (2,0) [place,label=below:$n-1$] {};
\node at (3,0) [place,label=below:$n\!\!\!\phantom{1}$] {};
\end{tikzpicture}
\\ \hline
\rule{0pt}{1.2cm}
\begin{minipage}[b]{1cm}
\centering $F_4$\\ \phantom{ }
\end{minipage} & \begin{tikzpicture}
[place/.style={circle,draw=black,fill=black,
inner sep=0pt,minimum size=6}]
\draw (-1,0) -- (2,0);
\draw (2,.1) -- (3,.1);
\draw (2,-.1) -- (3,-.1);
\draw (3,0) -- (4,0);
\draw (2.6,0) -- (2.4,0.2);
\draw (2.6,0) -- (2.4,-0.2);
\node at (-1,0) [place,label=below:$-1$] {};
\node at (0,0) [place,label=below:$0$] {};
\node at (1,0) [place,label=below:$1$] {};
\node at (2,0) [place,label=below:$2$] {};
\node at (3,0) [place,label=below:$3$] {};
\node at (4,0) [place,label=below:$4$] {};
\end{tikzpicture}
&
\begin{tikzpicture}
[place/.style={circle,draw=black,fill=black,
inner sep=0pt,minimum size=6}]
\draw (-1,0) -- (2,0);
\draw (2,.1) -- (3,.1);
\draw (2,-.1) -- (3,-.1);
\draw (3,0) -- (4,0);
\draw (2.4,0) -- (2.6,0.2);
\draw (2.4,0) -- (2.6,-0.2);
\node at (-1,0) [place,label=below:$-1$] {};
\node at (0,0) [place,label=below:$0$] {};
\node at (1,0) [place,label=below:$1$] {};
\node at (2,0) [place,label=below:$2$] {};
\node at (3,0) [place,label=below:$3$] {};
\node at (4,0) [place,label=below:$4$] {};
\end{tikzpicture}
\\ \hline
\rule{0pt}{1.2cm}
\begin{minipage}[b]{1cm}
\centering $G_2$\\ \phantom{ }
\end{minipage} & \begin{tikzpicture}
[place/.style={circle,draw=black,fill=black,
inner sep=0pt,minimum size=6}]
\draw (-1,0) -- (2,0);
\draw (1,.1) -- (2,.1);
\draw (1,-.1) -- (2,-.1);
\draw (1.6,0) -- (1.4,0.2);
\draw (1.6,0) -- (1.4,-0.2);
\node at (-1,0) [place,label=below:$-1$] {};
\node at (0,0) [place,label=below:$0$] {};
\node at (1,0) [place,label=below:$1$] {};
\node at (2,0) [place,label=below:$2$] {};
\end{tikzpicture}
&
\begin{tikzpicture}
[place/.style={circle,draw=black,fill=black,
inner sep=0pt,minimum size=6}]
\draw (-1,0) -- (2,0);
\draw (1,.1) -- (2,.1);
\draw (1,-.1) -- (2,-.1);
\draw (1.4,0) -- (1.6,0.2);
\draw (1.4,0) -- (1.6,-0.2);
\node at (-1,0) [place,label=below:$-1$] {};
\node at (0,0) [place,label=below:$0$] {};
\node at (1,0) [place,label=below:$1$] {};
\node at (2,0) [place,label=below:$2$] {};
\end{tikzpicture}\\ \hline\hline
\multicolumn{3}{c}{} \\[1mm]
\caption{Dynkin diagrams of the over-extended twisted and untwisted non-simply laced finite-dimensional algebras with Dynkin labeling of nodes}\label{fig:nsla}
\end{longtable*}
\selectlanguage{russian}
Here we list the matrices $S$ for the Lie Algebras of
$A_n, B_n, C_n, D_n, F_4, E_6, E_7$
and $E_8$ in the Cartan classification (the matrices for the rank 2 algebras were already
given in the previous section). In addition we list the set of Coxeter (or dual Coxeter)
labels $m_i$ used in the computation of $S$. Note that in the case of the non-simply laced algebras it is necessary to distinguish between the labels of the \textit{twisted} and \textit{untwisted} algebra. For these algebras we label the matrix $S$ with a superscript $^{(1)}$
or $^{(2)}$, respectively, indicating whether it corresponds to the \textit{untwisted} or
\textit{twisted} over-extension of the algebra. Considering Table \ref{fig:nsla} containing Dynkin diagrams of over-extensions of the non-simply laced algebras, we note that the twisted Dynkin diagram of $B_n$ is the same as the untwisted Dynkin diagram of $C_n$, simply with the direction of the arrows reversed. This tells us that the volumes of the corresponding fundamental domains have to be the same, since the matrix $S$ is the \textit{symmetrized} version the Cartan matrix and therefore contains no information about the direction of the arrows. A similar correspondence holds for the untwisted diagram of $B_n$ and the twisted diagram of $C_n$, as well as for $G_2$ and $F_4$.\\
\\
The condition for the over-extension of each algebra to be of hyperbolic type is that \textit{all} of the diagonal entries $S_{ii}$ of the underlying finite-dimensional algebra must satisfy $S_{ii}\leq1$. Geometrically each $S_{ii}=1$ corresponds to an additional fundamental weight (edge of the Weyl chamber) lying \textit{on} the forward light-cone. For each $S_{ii}>1$ an additional fundamental weight lies \textit{outside} the light cone and the over-extension is not of hyperbolic type. For each algebra we state the range of $n$ for which the over-extension is hyperbolic.\\
\\
\noindent{$\mak{g}^{++} \! = \!A_n^{++}$:}
The Coxeter labels are $m_i=\left(1,...,1\right)$, and thus
the matrix $S$ is
\begin{align}
& S[A_n]=\frac{1}{(n+1)}\times \nonumber\\
&\times\begin{pmatrix}
\frac{n}{2} & \frac{n-1}{2} & \frac{n-2}{2} & \frac{n-3}{2} & \cdots & \frac12 \\[2mm]
\frac{n-1}{2} & n-1 & n-2 & n-3 & \cdots & 1 \\[2mm]
\frac{n-2}{2} & n-2 & \frac{3(n-2)}{2} & \frac{3(n-3)}{2} & \cdots & \frac32 \\[2mm]
\frac{n-3}{2} & n-3 & \frac{3(n-3)}{2} & 2(n-3) & \cdots & 2 \\[2mm]
\vdots & \vdots & \vdots & \vdots & \ddots & \vdots\\[2mm]
\frac12 & 1 & \frac32 & 2 & \cdots & \frac{n}{2}
\end{pmatrix}
\end{align}
From the explicit form of the matrix it is obvious that
\begin{eqnarray}
S_{ij}\leq 1 \quad \Leftrightarrow\quad
\frac{j(n+1-j)}{2(n+1)} \leq 1
\end{eqnarray}
for all $j=1,\dots, n$. Hence $A_n^{++}$ is hyperbolic for $n\leq 7$.
\vspace{1mm}
\noindent $\mak{g}^{++} \!=\! B_n^{++}$: the Coxeter labels
are $m_i=\left(1,2,...,2\right)$ (as untwisted Coxeter labels for $B_n$,
and as twisted dual Coxeter labels for $C_n$), and the matrix $S$ is
\begin{align}
S^{(1)} [B_n] & =\left( \begin{array}{c|cccccc}
\frac12 & \frac14 & \frac14 & \frac14 & \cdots & \frac14 & \frac14 \\[2mm] \mhlines
\frac14 & \frac14 & \frac14 & \frac14 & \cdots & \frac14 & \frac14 \\[2mm]
\frac14 & \frac14 & \frac38 & \frac38 & \cdots & \frac38 & \frac38 \\[2mm]
\frac14 & \frac14 & \frac38 & \frac12 & \cdots & \frac12 & \frac12 \\[2mm]
\vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\[2mm]
\frac14 & \frac14 & \frac38 & \frac12 & \cdots & \frac{n-1}{8} & \frac{n-1}{8} \\[2mm]
\frac14 & \frac14 & \frac38 & \frac12 & \cdots & \frac{n-1}{8} & \frac{n}{8}
\end{array}\right)
\end{align}
All matrix entries are $\leq 1$ for $n\leq 8$, whence
$B_n^{++}$ is hyperbolic for $n\leq8$. Inverting the arrow
in the Dynkin diagram we infer that
\begin{eqnarray}
S^{(1)}[B_n] = S^{(2)}[C_n].
\end{eqnarray}
\vspace{1mm}
\noindent $\mak{g}^{++} \!=\! C_n^{++}$: the Coxeter labels are
$m_i=\left(1,...,1\right)$ (as twisted Coxeter labels for $B_n$ and
untwisted dual Coxeter labels for $C_n$), so
\begin{align}
S^{(1)} [C_n] & =\begin{pmatrix}
\frac14 & \frac14 & \frac14 & \frac14 & \cdots & \frac14 & \frac14 \\[2mm]
\frac14 & \frac12 & \frac12 & \frac12 & \cdots & \frac12 & \frac12 \\[2mm]
\frac14 & \frac12 & \frac34 & \frac34 & \cdots & \frac34 & \frac34 \\[2mm]
\frac14 & \frac12 & \frac34 & 1 & \cdots & 1 & 1 \\[2mm]
\vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\[2mm]
\frac14 & \frac12 & \frac34 & 1 & \cdots & \frac{n-1}{4} & \frac{n-1}{4} \\[2mm]
\frac14 & \frac12 & \frac34 & 1 & \cdots & \frac{n-1}{4} & \frac{n}{4}
\end{pmatrix}
\end{align}
Clearly, $C_n^{++}$ is hyperbolic for $n\leq4$. As before we get
\begin{eqnarray}
S^{(1)}[C_n] = S^{(2)}[B_n]
\end{eqnarray}
\vspace{1mm}
\noindent $\mak{g}^{++} \!=\! D_n^{++}$: the Coxeter labels are
$m_i=\left(1,2,...,2,1,1\right)$, and therefore
\begin{eqnarray}
S[D_n] = \left(\begin{array}{c|ccccc|cc}
\frac12 & \frac14 & \frac14 & \frac14 & \cdots & \frac14 & \frac14 & \frac14 \\[2mm] \mhlinee
\frac14 & \frac14 & \frac14 & \frac14 & \cdots & \frac14 & \frac14 & \frac14 \\[2mm]
\frac14 & \frac14 & \frac38 & \frac38 & \cdots & \frac38 & \frac38 & \frac38 \\[2mm]
\frac14 & \frac14 & \frac38 & \frac12 & \cdots & \frac12 & \frac12 & \frac12 \\[2mm]
\vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots & \vdots \\[2mm]
\frac14 & \frac14 & \frac38 & \frac12 & \cdots & \frac{n-2}{8} & \frac{n-2}{8} & \frac{n-2}{8} \\[2mm]\mhlinee
\frac14 & \frac14 & \frac38 & \frac12 & \cdots & \frac{n-2}{8} & \frac{n}{8} & \frac{n-2}{8} \\[2mm]
\frac14 & \frac14 & \frac38 & \frac12 & \cdots & \frac{n-2}{8} & \frac{n-2}{8} & \frac{n}{8}
\end{array}\right)
\end{eqnarray}
We see that $D_n^{++}$ is hyperbolic for $n\leq 8$.
\vspace{2mm}
\noindent $\mak{g}^{++} \!=\! F_4^{++}$: the Coxeter labels are
$\left(2,3,2,1\right)$ for the untwisted dual Coxeter labels as
well as for the twisted Coxeter labels:
\begin{eqnarray}
S^{(1)}[F_4] = S^{(2)}[F_4] = \left(\begin{array}{cccc} \frac14 & \frac14 & \frac14 & \frac14 \\[2mm]
\frac14 & \frac13 & \frac13 & \frac13 \\[2mm]
\frac14 & \frac13 & \frac38 & \frac38 \\[2mm]
\frac14 & \frac13 & \frac38 & \frac12 \end{array} \right)
\end{eqnarray}
\vspace{2mm}
\noindent $\mak{g}^{++} \!=\! E_6^{++}$: the Coxeter labels are
$\left(1,2,3,2,1,2\right)$, and we have
\begin{eqnarray}
S[E_6] = \left(\begin{array}{cccccc} \frac23 & \frac{5}{12} & \frac13 & \frac13 & \frac13 & \frac14 \\[2mm]
\frac{5}{12} & \frac{5}{12} & \frac13 & \frac13 & \frac13 & \frac14 \\[2mm]
\frac13 & \frac13 & \frac13 & \frac13 & \frac13 & \frac14 \\[2mm]
\frac13 & \frac13 & \frac13 & \frac{5}{12} & \frac{5}{12} & \frac14 \\[2mm]
\frac13 & \frac13 & \frac13 & \frac{5}{12} & \frac23 & \frac14 \\[2mm]
\frac14 & \frac14 & \frac14 & \frac14 & \frac14 & \frac14 \end{array} \right)
\end{eqnarray}
\vspace{2mm}
\noindent $\mak{g}^{++} \!=\! E_7^{++}$: the Coxeter labels are
$\left(2,3,4,3,2,1\right)$, and we have
\begin{eqnarray}
S[E_7] = \left(\begin{array}{ccccccc} \frac14 & \frac14 & \frac14 & \frac14 & \frac14 & \frac14 & \frac14 \\[2mm]
\frac14 & \frac13 & \frac13 & \frac13 & \frac13 & \frac13 & \frac13 \\[2mm]
\frac14 & \frac13 & \frac38 & \frac38 & \frac38 & \frac38 & \frac38 \\[2mm]
\frac14 & \frac13 & \frac38 & \frac{5}{12} & \frac{5}{12} & \frac{5}{12} & \frac38 \\[2mm]
\frac14 & \frac13 & \frac38 & \frac{5}{12} & \frac12 & \frac12 & \frac38 \\[2mm]
\frac14 & \frac13 & \frac38 & \frac{5}{12} & \frac12 & \frac34 & \frac38 \\[2mm]
\frac14 & \frac13 & \frac38 & \frac38 & \frac38 & \frac38 & \frac{7}{16} \end{array} \right)
\end{eqnarray}
\\
\vspace{2mm}
\noindent $\mak{g}^{++} \!=\! E_8^{++}$: with the Coxeter labels
$\left(2,3,4,5,6,4,2,3\right)$ we have
\begin{eqnarray}
S[E_8] =\left(\begin{array}{cccccccc} \frac14 & \frac14 & \frac14 & \frac14 & \frac14 & \frac14 & \frac14 & \frac14 \\[2mm]
\frac14 & \frac13 & \frac13 & \frac13 & \frac13 & \frac13 & \frac13 & \frac13 \\[2mm]
\frac14 & \frac13 & \frac38 & \frac38 & \frac38 & \frac38 & \frac38 & \frac38 \\[2mm]
\frac14 & \frac13 & \frac38 & \frac25 & \frac25 & \frac25 & \frac25 & \frac25 \\[2mm]
\frac14 & \frac13 & \frac38 & \frac25 & \frac{5}{12} & \frac{5}{12} & \frac{5}{12} & \frac{5}{12} \\[2mm]
\frac14 & \frac13 & \frac38 & \frac25 & \frac{5}{12} & \frac{7}{16} & \frac{7}{16} & \frac{5}{12} \\[2mm]
\frac14 & \frac13 & \frac38 & \frac25 & \frac{5}{12} & \frac{7}{16} & \frac12 & \frac{5}{12} \\[2mm]
\frac14 & \frac13 & \frac38 & \frac25 & \frac{5}{12} & \frac{5}{12} & \frac{5}{12} & \frac29 \end{array} \right)
\end{eqnarray}
\vspace{2mm}
Using equations \eqref{UHPMap} one can determine the coordinates of
the vertices of the fundamental domain. For example, the vertices of the domain
corresponding to the Weyl group of $\mak{e}_{10}$ are given by
\begin{eqnarray}
\left(v_1,{\mathbf{u}}_1\right)& = &\left(\frac{\sqrt{3}}{2},\frac12 {\mathbf{e}}_0 \right) \nonumber \\
\left(v_2,{\mathbf{u}}_2\right)& = &\left(\sqrt{\frac23},\frac12 {\mathbf{e}}_0+\frac16\left({\mathbf{e}}_1+{\mathbf{e}}_5+{\mathbf{e}}_6\right)\right) \nonumber \\
\left(v_3,{\mathbf{u}}_3\right)& = &\left(\sqrt{\frac58},\frac12 {\mathbf{e}}_0+\frac14\left({\mathbf{e}}_5+{\mathbf{e}}_6\right)\right) \nonumber \\
\left(v_4,{\mathbf{u}}_4\right)& = &\left(\sqrt{\frac35},\frac12 {\mathbf{e}}_0+\frac{1}{10}\left({\mathbf{e}}_2+3{\mathbf{e}}_5+2{\mathbf{e}}_6-{\mathbf{e}}_7\right) \right)\nonumber \\
\left(v_5,{\mathbf{u}}_5\right)& = &\left(\frac{1}{\sqrt{6}},\frac12 {\mathbf{e}}_0+\frac16\left(2{\mathbf{e}}_5+{\mathbf{e}}_6-{\mathbf{e}}_7\right) \right) \nonumber \\
\left(v_6,{\mathbf{u}}_6\right)& = &\left(\frac34,\frac12 {\mathbf{e}}_0+\frac18\left({\mathbf{e}}_3+3{\mathbf{e}}_5+{\mathbf{e}}_6-{\mathbf{e}}_7\right) \right) \nonumber \\
\left(v_7,{\mathbf{u}}_7\right)& = &\left(\frac{1}{\sqrt{2}},\frac12 {\mathbf{e}}_0+\frac12{\mathbf{e}}_5 \right) \nonumber \\
\left(v_8,{\mathbf{u}}_8\right)& = &\left(\frac{\sqrt{5}}{3},\frac12 {\mathbf{e}}_0+\frac16\left({\mathbf{e}}_4+2{\mathbf{e}}_5+{\mathbf{e}}_6-{\mathbf{e}}_7\right) \right)
\end{eqnarray}
The special feature of this example is that the vectors ${\mathbf{u}}_j$ now belong to
the octonions $\mathbb{O}$, the non-commutative and non-associative maximal division
algebra. Accordingly, the unit vectors ${\mathbf{e}}_j$ (for $j=1,\dots,7$)
are just the octonionic imaginary units. The vertex coordinates of the fundamental
domains of other Weyl groups are obtained similarly.
By evaluating the integrals in \eqref{main} numerically we obtain the volumes of all the fundamental domains of the hyperbolic Weyl groups of the algebras listed above. Employing a deterministic adaptive integration scheme with a sufficient number of evaluation points of the integrand, the
values we find agree to high accuracy with those found in \cite{JKRT99} where the
volumes of all hyperbolic Coxeter simplices were obtained by a different method.
\acknowledgments{We are very grateful to Axel Kleinschmidt for discussions
and helpful comments on an earlier version of this paper. We would also like
to thank Jakob Palmkvist for discussions and correspondence. The work of
P. Fleig is supported by an IRAP Erasmus Mundus Joint Doctorate Fellowship
and the University of Nice -- Sophia Antipolis.}
\vspace{1cm}
|
\section{Introduction.}\label{Intro}
For $k$ an even integer, the space of holomorphic forms of weight $k$ for the full modular group $\Gamma = PSL(2,\mathbb{Z})$ is of dimension $\frac{k}{12} + O(1)$ (see \cite{Ser} for exact definitions as well as other basic facts). Such a form $F$ has $\frac{k}{12} + O(1)$ zeros in $\mathfrak{X} = \Gamma\backslash\mathcal{H}$. More precisely,
\begin{equation}
\nu_{\infty}(F) + \frac{\nu_{i}(F)}{2} + \frac{\nu_{\rho}(F)}{3} + \sum_{p \in \mathfrak{X}} \nu_{p}(F) = \frac{k}{2}
\end{equation}
where $\nu_{p}(F)$ is the order of vanishing of $F$ at the point $p$. Here $\infty$, $i$ and $\rho$ are points in $\mathfrak{X}$ depicted in the familiar Figure 1. Other than vanishing at $z=i$ and $z=\rho$ when forced by (1) for $k$ in various progressions modulo 12, the distribution of the zeros of such an $F$ is not restricted. However, for the arithmetically interesting case of $F$ being a Hecke eigenform, which we will assume henceforth, there are constraints on the distribution of the zeros. In particular, if $F$ is an Eisenstein series $E_{k}(z)$, then it has been shown by Rankin and Swinnerton-Dyer \cite{RanSD} that all of its zeros are on the geodesic segment $\delta_{3}$ in Figure 1 (there have been many generalizations of this result to functions constructed from Eisenstein series (see \cite{DukeJenkins}) and to other Fuchsian groups (see \cite{Hahn})). For the rest of the Hecke eigenforms, namely the cusp forms, which we denote by $f$, the distribution of the zero set $\mathcal{Z}(f)$ (counted with multiplicities) is very different. By definition such a $f$ vanishes at the cusp $z = i\infty$ and from the well known properties of the Hecke operators it follows that $\nu_{\infty}(f)=1$. The expansion of $f$ at $i\infty$ takes the form
\begin{equation}
f(z) = \sum_{n=1}^{\infty} a_{f}(n) e(nz),
\end{equation}
where we normalize $f$ with $a_{f}(1)=1$ and write $e(x) = exp(2\pi ix)$.
\centerline{\includegraphics[width=3.5in]{delta.png}}
\vspace{10pt}
One of the striking consequences of the recent proof of the holomorphic QUE conjecture by Holowinsky and Soundararajan \cite{HS} is that $\mathcal{Z}(f)$ is equidistributed in $\mathfrak{X}$ as $k \rightarrow \infty$ (see also \cite{Rudnick} and the report \cite{Sa} for a discussion). That is for any nice set $\Omega \subset \mathfrak{X}$
\begin{equation}
\frac{|\mathcal{Z}(f)\cap \Omega|}{|\mathcal{Z}(f)|} \rightarrow \frac{Area(\Omega)}{Area(\mathfrak{X})},
\end{equation}
as $k \rightarrow \infty$. Here $Area$ is the hyperbolic area with $dA =\frac{dx dy}{y^2}$.
This paper is concerned with the zeros of $f$ lying on the geodesic segments $\delta_{1}$, $\delta_{2}$ and $\delta_{3}$ in Figure 1, which we call the `real' zeros of $f$.\footnote{As Zeev Rudnick notes, these segments are the points at which the $j$-invariant is real.} The reason for this name is that the $a_{f}(n)$'s in (2) are all real and hence $f$ is a real-valued function on the segments $\delta_{1}$ and $\delta_{2}$, while $z^{\frac{k}{2}}f(z)$ is real-valued on $\delta_{3}$. These follow from the relations
\[
f(S_{1}z) = f(S_{2}z) = \overline{f(z)},
\]
and
\begin{equation}
f(S_{3}z) = \overline{z}^{k}\overline{f(z)}
\end{equation}
where $S_{1}$, $S_{2}$ and $S_{3}$ are the reflections
\[
S_{1}(z)= -\overline{z},\hspace{20pt} S_{2}(z)= 1 -\overline{z},\hspace{20pt} S_3(z)=\frac{1}{\overline z}.
\]
It follows that $\mathcal{Z}(f)$ is invariant under these involutions whose fixed points are the segments $\delta_{1}$, $\delta_{2}$ and $\delta_{3}$ respectively. Let $\delta^{*} = \delta_{1}\cup \delta_{2} \cup \delta_{3}$. One might expect that the number of real zeros, $N_{real}(f) := |\mathcal{Z}(f)\cap \delta^{*}|$ to grow with $k$, much like the number of real zeros of a random polynomial with real coefficients. According to (3), $N_{real}(f) = o(|\mathcal{Z}(f)|) = o(k)$ as $k \rightarrow \infty$.
\vspace{20pt}
\centerline{\includegraphics[width=3.5in]{stromberg.png}}
\vspace{20pt}
Our first result concerning real zeros of $f$ is that their number does in fact grow with $k$.
\begin{thm}\label{ThmOne} For any $\epsilon > 0$, as $k \rightarrow \infty$ and any Hecke cusp form of even weight $k$,
\[
N_{real}(f) \gg_{\epsilon} k^{(\frac{1}{4} - \frac{1}{80} - \epsilon)}.
\]
\end{thm}
As discussed below and in Section 6, the true order of magnitude of $N_{real}(f)$ is probably $\sqrt{k}\log k$.
\vspace{20pt}
Theorem \ref{ThmOne} follows from a more detailed investigation of the zeros of $f$ in regions which move into the cusp with $k$. To quantify this statement, we define the Siegel sets
\begin{equation}
\mathcal{F}_{Y} = \{ z \in \mathfrak{X}: \Im (z) \geq Y\},
\end{equation}
which have hyperbolic area $\frac{1}{Y}$ and we restrict our considerations to $Y$ satisfying
\begin{equation}
\sqrt{k\log k} \ll Y < \frac{1}{100}k.
\end{equation}
We show that the equidistribution (3) holds (approximately) for the zeros of $f$ in these shrinking (relative to the area) regions $\mathcal{F}_{Y}$ and moreover that many of these are real zeros.
\begin{thm}\label{ThmTwo} Let $\epsilon > 0$ and $k \rightarrow \infty$, $f$ any Hecke cusp form of even weight $k$ and $Y$ satisfying (6). Then
\begin{itemize}
\item[(i)] $$ \frac{k}{Y} \ll |\mathcal{Z}(f) \cap \mathcal{F}_{Y}| \ll \frac{k}{Y}, $$
\item[(ii)] $$|\mathcal{Z}(f) \cap \mathcal{F}_{Y} \cap (\delta_{1}\cup \delta_{2}| \gg _\epsilon \big{(}\frac{k}{Y}\big{)}^{\frac{1}{2} - \frac{1}{40} -\epsilon}.$$
\end{itemize}
The implied constants in (i) above are absolute while that in (ii) depend only on $\epsilon$.
\end{thm}
\vspace{20pt}
Our proof of Theorem \ref{ThmTwo} leads to questions of the nonvanishing of many of the $a_{f}(n)$'s with $n \leq \sqrt{k}$ and it suggests, somewhat surprisingly, that
\begin{equation}
|\mathcal{Z}(f) \cap \mathcal{F}_{Y} \cap (\delta_{1}\cup \delta_{2}| \sim |\mathcal{Z}(f) \cap \mathcal{F}_{Y}|
\end{equation}
as $k \rightarrow \infty$ (see Corollary \ref{Corl 5}). In other words, almost all the zeros in $\mathcal{F}_{Y}$, for $Y$ restricted in the range (6), are real zeros and futhermore half of these are on $\delta_{1}$ and the other half on $\delta_{2}$.
The proof of (ii) in Theorem \ref{ThmTwo} (and so Theorem \ref{ThmOne}) does not specify on which of $\delta_{1}$ or $\delta_{2}$ the zeros that are being produced lie. On $\delta_{2}$, we are able to take advantage of a natural oscillation induced on the fourier coefficients to produce some zeros of $f(z)$. The situation on $\delta_{1}$ is substantially harder. Our analysis reduces the problem of finding zeros to producing $n$'s with $1 \leq n \ll k^{\frac{1}{2} - \eta}$ and $a_{f}(n) < -\epsilon_{0}$ for some $\eta > 0$ and $\epsilon_{0} >0$ (both fixed independent of $k$). There has been recent progress on the problem of estimating from above the least $n$ for which $a_{f}(n) < 0$, see (\cite{IKS}, \cite{KLSW} and most recently \cite{Mato}). Remarkably the optimization in \cite{Mato} of the smooth number argument from \cite{KLSW}, together with the sharp subconvex bounds of \cite{Peng} and \cite{J-M} for the critical values of the $L$-function $L(s,f)$, allow us by the closest of margins to produce the requisite $n$'s. In either case (see Section 4), we then have
\begin{thm}\label{ThmThree} The number of zeros of $f(z)$ on $\delta_{1}$ and separately $\delta_{2}$ goes to infinity as k goes to infinity. More quantitatively, for $j =1$ and $2$
\[
|\mathcal{Z}(f) \cap \delta_{j}| \gg \log k .
\]
\end{thm}
\vspace{20pt}
It is natural to ask if $N_{real}(f)$ has an asymptotic law. To try to answer this, we determined the number of real zeros such an $f$ would have if the coefficients $a_{f}(n)$ were to behave in some random manner (satisfying the requisite bounds). This model is not completely accurate since the $a_{f}(n)$'s have a multiplicative structure which we ignore (for the sake of simplicity); however we still believe it yields the correct order of magnitude. Whether the constants obtained in the asymptotics are reliable is best checked by numerical experimentation. In any case, such a random model predicts that as $k \rightarrow \infty$
\[
|\mathcal{Z}(f) \cap \delta_{1}| \sim |\mathcal{Z}(f) \cap \delta_{2}| \sim \frac{\sqrt{k}}{4\pi} \log k,
\]
while
\begin{equation}
|\mathcal{Z}(f) \cap \delta_{3}| \sim \frac{\sqrt{k}}{4\pi} \log 3
\end{equation}
and in particular that
\[
N_{real}(f) \sim \frac{\sqrt{k}}{2\pi} \log k.
\]
The random model can also be examined for zeros of $f(z)$ on $\delta_{1}$ and $\delta_{2}$ with $y \gg \sqrt{k}$ (see (26) in Section 6). It predicts that almost all of the zeros of $f(z)$ for $y \gg \sqrt{k}$ are real, which is consistent with the statements made in (7) that were obtained by purely arithmetic considerations. In fact, it is even possible that all of these zeros are real (see remark 5.3). This lends support to the believe that the random model is appropriate even for $y \ll \sqrt{k}$.
\vspace{20pt}
To end this introduction, we outline our proofs of the Theorems. The analysis is based on a suitable approximation to $f(z)$ in the regions $\mathcal{F}_{Y}$ when $k$ and $Y$ are large. This is derived in Sections 2 and 3. Recall that for a Hecke cuspform, one can write
\[
a_{f}(n) = a_{f}(1)\lambda_{f}(n)n^\frac{k-1}{2},
\]
with $ a_{f}(1)$ nonzero (normalised to $1$) and with $\lambda_{f}(n)$ multiplicative and real. Among the various inputs into this asymptotic analysis are Deligne's \cite{Deligne} bounds $|\lambda_{f}(n)| \leq d(n) \ll n^{\epsilon}$ for any $\epsilon > 0$. The upshot of the analysis is that for integers $1 \ll l \ll \sqrt{\frac{k}{\log k}}$, with $y_{l} = \frac{k-1}{4\pi l}$ (or for $y$ close enough to $y_{l}$) and $0 \leq \alpha \leq \frac{1}{2}$, both $f(\alpha + iy_{l})$ and $\frac{f'}{f}(\alpha + iy_{l})$ can be approximated by simple functions as long as the $\lambda_{f}(n)$'s are not too small (see Cor 3.3). This condition on $\lambda_{f}(n)$ allows us to conclude that $f$ has exactly $l$ zeros in $\mathcal{F}_{\frac{k-1}{4\pi l}}$. Thus, part (i) of Theorem \ref{ThmTwo} is reduced to finding some $l$'s in suitable ranges with $\lambda_{f}(l)$ not small. This is a nontrivial problem since $f$ is varying and $l \ll \sqrt{k}$ , which is small in terms of the conductors of the associated $L$-functions $L(s,f)$ and $L(s,sym^{2}f)$. We proceed by using a much exploited and robust feature, that since $\lambda_{f}(n)$ is multiplicative, either $|\lambda_{f}(p)|$ or $|\lambda_{f}(p^2)|$ is at least $\frac{1}{2}$ for any prime $p$. In section 4 we use this together with a combinatorial analysis to construct sufficiently many such $l$'s. For part (ii) of Theorem \ref{ThmTwo} we proceed by looking for many sign-changes of $f$ on $\delta_{1} \cup \delta_{2}$, and for this we have to elaborate the analysis above by constructing many pairs of $l$'s with opposite parity with $\lambda_{f}(l)$'s not small. We reduce the problem to seeking a full density set of integers $m$ in $[M,2M]$, for which the short intervals $[m,m+\Delta]$ have at least one prime number. Assuming the Riemann Hypothesis for the Riemann zeta function, Selberg \cite{Sel} showed that the above is true it $\Delta = (\log M)^{2 + \epsilon}$. It appears that the smallest $\Delta$ for which the above is known unconditionally is $\Delta = M^{\frac{1}{20}}$ \cite{Jia}, and this is what we use and it is responsible for the various exponents in our theorems. As noted earlier, the proof of Theorem \ref{ThmThree} for $\delta_{1}$ relies on strong subconvex bounds for $L(s,f)$, as well as optimised smooth number arguments.
An alternate but equivalent approach is to find $l < l'$ of opposite parity which are close to each other and for which both $\lambda_{f}(l)$ and $\lambda_{f}(l')$ are not too small. This ensures that in the region $\frac{k-1}{4\pi l'} \leq y \leq \frac{k-1}{4\pi l}$, the number of zeros of $f$ is odd and hence by the symmetry associated with $S_{1}$ and $S_{2}$, there must be at least one zero of $f$ on $\delta_{1} \cup \delta_{2}$, in this region.
The last section is devoted to modelling $\lambda_{f}(n)$ by random numbers, namely identical, independently distributed Gaussians of mean zero and variance one. For a random such $f$, we determine the expected density of zeros on each segment of $\delta^{*}$.
\vskip 0.2in
{\small {\bf Acknowledgements.}
We thank Fredrik Str\"omberg whose revealing computations and pictures such as those shown in Figure 2, led us to investigate the real zeros of the forms $f$. Thanks also to Andre Reznikov with whom we are preparing a followup to this paper which investigates the `real nodal domains' of Maass forms on $\mathfrak{X}$, to K. Soundararajan for pointing us to the recent preprint \cite{Mato} and to K. Matomaki for the reference \cite{Jia}.
The first author also thanks the Institute for Advanced Study for providing the possibility
of an extended visit during which most of this work took place. He also gratefully acknowledges the support from the Ellentuck Fund of the Institute for Advanced Study.}
\vskip 0.2in
\section{Basic proposition.}
We begin with a detailed steepest descent analysis of the behavior of $f(z)$ when $k$ and $y$ are large. In connection with $L^{\infty}$-norms, related approximations are derived in \cite{Sarnak2}(pages 26-29) for Maass forms and by \cite{Xia} for holomorphic forms.
Let
\[
I_{s}(y) = y^{\frac{s-1}{2}}e^{-y}
\]
for $y>0$ and $s \in \mathbb{C}$, and define
\[
\Phi_{f}(s;\alpha,y) = \sum_{1}^{\infty} \lambda_{f}(n)e(n\alpha)I_{s}(2\pi ny)
\]
for any real $\alpha$. We then have
\begin{equation}
f(\alpha +iy) = a_{f}(1)(2\pi y)^{-k'}\Phi_{f}(k;\alpha,y),
\end{equation}
where we will use the notation $k' = \frac{k-1}{2}$. More generally, for any $m \geq 1$, we have
\begin{equation}
\big{(}\frac{1}{2\pi i}\big{)}^{m}f^{(m)}(\alpha +iy) = a_{f}(1)(2\pi y)^{-k' -m}\Phi_{f}(k + 2m;\alpha,y).
\end{equation}
This implies that if $f(\alpha +iy) \neq 0$, then
\begin{equation}
\frac{1}{2\pi i} \frac{f'}{f}(\alpha +iy) = \frac{1}{2\pi y}\frac{\Phi_{f}(k+2;\alpha,y)}{\Phi_{f}(k;\alpha,y)}.
\end{equation}
We will first prove our basic
\begin{prop}\label{Prop 1} Let $\delta > 0$. Then there is a $N(\delta)$ sufficiently large such that for all real $s > N(\delta)$, for all $y$ satisfying $ \sqrt{s} \ll y < \frac{1}{100}s$, and with $B = \sqrt{\delta s\log{s}}$, we have
\begin{equation}
\frac{\Phi_{f}(s;\alpha,y)}{I_{s}(s')} = \sum_{\substack{n \\ |2\pi ny - s'|\leq B}} \lambda_{f}(n)e(n\alpha)e^{-\frac{|2\pi ny -s'|^{2}}{2s'}} + O(s^{-\delta}),
\end{equation}
where $s' = \frac{s-1}{2}$.
\end{prop}
To prove the theorem, we will need some elementary lemmas regarding the behaviour of $I_{s}(y)$.
\begin{lem}\label{Lemma 1.} For a fixed $s$, $I_{s}(y)$ is strictly increasing for $0<y<s'$, and strictly decreasing if $y>s'$.
\end{lem}
\begin{lem}\label{Lemma 2.} Suppose $|h| \ll s^{\frac{2}{3} -\delta}$ for some positive $\delta$ sufficiently small. Then
\[
I_{s}(s' +h) = I_{s}(s')e^{-\frac{h^2}{2s'}}(1 + O(s^{-3\delta})).
\]
\end{lem}
{\noindent\bf Proof. } We write
\[
I_{s}(s' +h) = e^{-s' -h}(s')^{s'}(1+\frac{h}{s'})^{s'}.
\]
To a first approximation,
\[
\log(1+\frac{h}{s'})^{s'} = h - \frac{h^2}{2s'} + O(\frac{h^3}{s'^2})
\]
from which the lemma follows.
\vskip 0.1in
{\noindent \bf Proof of Prop. \ref{Prop 1} .} We write, for a paramater $B$ to be chosen later
\[
\Phi_{f}(s;\alpha,y) = \sum_{i=1}^{3} \Phi_{f}^{(i)}(s;\alpha,y)
\]
where
\[
\Phi_{f}^{(1)}(s;\alpha,y) = \sum_{\substack{n \geq 1\\ 2\pi ny < s' - B}} \lambda_{f}(n)e(n\alpha)I_{s}(2\pi ny),
\]
\[
\Phi_{f}^{(2)}(s;\alpha,y) = \sum_{\substack{n \\ |2\pi ny - s'|\leq B}} \lambda_{f}(n)e(n\alpha)I_{s}(2\pi ny),
\]
and
\[
\Phi_{f}^{(3)}(s;\alpha,y) = \sum_{ 2\pi ny > s' + B} \lambda_{f}(n)e(n\alpha)I_{s}(2\pi ny).
\]
We choose $h= 2\pi ny -s'$ (in Lemma \ref{Lemma 2.}), so that $I_{s}(2\pi ny)=I_{s}(s' + h)$ and consequently assume that $1 \leq B \ll s^{\frac{2}{3}-\delta}$.
We will first estimate $\Phi_{f}^{(3)}(s;\alpha,y)$. Using the fact that $I_{s_{1}}(t) = t^{\frac{s_{1} -s}{2}}I_{s}(t)$, and the upper-bound for $\lambda_{f}(n)$, we have for any $\epsilon >0$ and $s$ sufficiently large,
\[
\Phi_{f}^{(3)}(s;\alpha,y) \ll y^{-\epsilon} \sum_{n > \frac{s' + B}{2\pi y}} I_{s + 2\epsilon}(2\pi ny).
\]
The maximum for $I_{s + 2\epsilon}(t)$ is attained at $s' + \epsilon$ and since $B \geq 1$, we see that
$I_{s + 2\epsilon}(2\pi ny)$ is decreasing in this sum. We may the approximate the sum by the appropriate integral
to get
\[
\Phi_{f}^{(3)}(s;\alpha,y) \ll y^{-\epsilon} \Big{(} \int_{\frac{s' + B}{2\pi y}}^{\infty} (2\pi yt)^{s' + \epsilon}e^{-2\pi yt} \, dt + I_{s + 2\epsilon}(s' + B)\Big{)}
\]
\begin{equation}
\hspace{40pt} \ll y^{-\epsilon} \Big{(} \frac{1}{y}\Gamma(s'+\epsilon +1,s' +B) + I_{s + 2\epsilon}(s' + B)\Big{)}
\end{equation}
where
\[
\Gamma(s,x) = \int_{x}^{\infty} t^{s-1} e^{-t} \, dt
\]
is the incomplete gamma function.
First, we have
\[
\frac{I_{s + 2\epsilon}(s' + B)}{I_{s}(s')} \ll e^{-B}(1+\frac{B}{s'})^{s'}s^{\epsilon},
\]
so that on using
\[
\log(e^{-B}(1+\frac{B}{s'})^{s'}) = -\frac{B^2}{2s'} +O(\frac{B^3}{s'^2}),
\]
we conclude that
\begin{equation}
I_{s + 2\epsilon}(s' + B) \ll I_{s}(s')s^{\epsilon}e^{-\frac{B^2}{2s'}}.
\end{equation}
To estimate the incomplete gamma function, we use the following inequality due to Natalini-Palumbo \cite{NP}
\begin{lem}\label{Lemma 3} If $a>1$, $\sigma > 1$ and $x > \frac{\sigma}{\sigma -1}(a-1)$, one has
\[
x^{a-1}e^{-x} < |\Gamma(a,x)| < \sigma x^{a-1}e^{-x}.
\]
\end{lem}
We shall use this lemma with $a=s' +\epsilon +1$ and $\sigma = 1+\epsilon + \frac{s'}{B}$. Then
\[
\Gamma(s'+\epsilon +1,s' +B) \ll \frac{s'}{B}(s' +B)^{s' +\epsilon}e^{-s'-B}
\]
\begin{equation}
\ll \frac{s'}{B}(1+\frac{B}{s'})^{s'}e^{-B}I_{s}(s')s^{\epsilon}
\ll \frac{s'}{B}e^{-\frac{B^2}{2s'}}I_{s}(s')s^{\epsilon}.
\end{equation}
Collecting the estimates from (7),(6) and (5) gives us
\[
\Phi_{f}^{(3)}(s;\alpha,y) \ll y^{-\epsilon}(\frac{s'}{By}+1)e^{-\frac{B^2}{2s'}}I_{s}(s')s^{\epsilon}.
\]
We choose
\[
\sqrt{\delta s\log s} \leq B \ll s^{\frac{2}{3} - \delta}
\]
and $y \gg \sqrt{s}$ to conclude that
\[
\Phi_{f}^{(3)}(s;\alpha,y) \ll I_{s}(s')s^{-\frac{\delta}{2}}.
\]
We next estimate $\Phi_{f}^{(1)}(s;\alpha,y)$.
Since $n$ is bounded by $s$, we replace $\lambda_{f}(n)$ with $s^{\epsilon}$. Moreover, $I_{s}(2\pi ny)$ is strictly increasing in our interval so that we may approximate the modified sum with the appropriate integral to get
\begin{equation}
\Phi_{f}^{(1)}(s;\alpha,y) \ll s^{\epsilon}\big{(}\int_{1}^{\frac{s'-B}{2\pi y}}(2\pi yt)^{s'}e^{-2\pi yt} \, dt + I_{s}(2\pi y) + I_{s}(s' -B)\big{)}.
\end{equation}
Now, we have
\[
\frac{I_{s}(2\pi y)}{I_{s}(s')} \ll (\frac{2\pi ye}{s'})^{s'}e^{-2\pi y},
\]
which decays exponentially in $s$ if we choose $y < \frac{s}{100}$. Using the analysis for (6), we have
\[
\frac{I_{s}(s'-B)}{I_{s}(s')} \ll e^{-\frac{B^2}{2s'}}.
\]
To estimate the integral in (8), we break it up into two pieces: let $B_{1} > B$ so that we may write the integral as
\[
\int_{1}^{\frac{s'-B}{2\pi y}} + \int_{\frac{s'-B_{1}}{2\pi y}}^{\frac{s'-B}{2\pi y}} (2\pi yt)^{s'}e^{-2\pi yt} \, dt.
\]
Replacing $t$ with $\frac{s't}{2\pi y}$ and simplifying gives us
\begin{equation}
\frac{s'}{y}I_{s}(s')\big{(}\int_{\frac{2\pi y}{s'}}^{1-\frac{B_{1}}{s'}} + \int_{1-\frac{B_{1}}{s'}}^{1-\frac{B}{s'}} t^{s'}e^{s'(1-t)} \, dt \big{)}.
\end{equation}
The integrand is strictly increasing so that the second integral is bounded by
\[
\frac{B_{1} -B}{s'}(1-\frac{B}{s'})^{s'}e^{B} \ll \frac{B_{1} -B}{s'}e^{-\frac{B^2}{2s'}}.
\]
The first integral is trivially bounded by $e^{-\frac{B_{1}^2}{2s'}}$ provided $B_{1} \ll s^{\frac{2}{3} -\epsilon}$. Hence, from (9), the integral in (8) is
\begin{equation}
\ll \frac{s'}{y}I_{s}(s')\big{(} e^{-\frac{B_{1}^2}{2s'}} + \frac{B_{1} -B}{s'}e^{-\frac{B^2}{2s'}} \big{)}.
\end{equation}
We choose
\[
B = \sqrt{\delta s\log s},\hspace{20pt} B_{1} = \sqrt{\frac{1}{\delta} s\log s}
\]
with $\delta$ sufficiently small to get that (10) is
\[
\ll \frac{s'}{y}I_{s}(s')(s^{-\frac{1}{\delta}} + \frac{\sqrt{s\log s}}{s'}s^{-\delta}) \ll \frac{\sqrt{s\log s}}{y}I_{s}(s')s^{-\delta} \ll I_{s}(s')s^{-\frac{1}{2}\delta},
\]
since $y>\sqrt{s}$. Collecting all thses estimates together gives us
\[
\Phi_{f}^{(1)}(s;\alpha,y) \ll I_{s}(s')s^{-\frac{1}{3}\delta}.
\]
Finally, for $\Phi_{f}^{(2)}(s;\alpha,y)$ we use Lemma \ref{Lemma 2.}. The error-term contributes
\[
\ll s^{\epsilon}(\sum_{|2\pi ny -s'|\leq B}1)s^{-3\delta} \ll (1+\frac{B}{y})s^{-2\delta} \ll s^{-\delta}.
\] The main-term in Lemma \ref{Lemma 2.} gives the sum stated in (4), with $h=2\pi ny -s'$. This completes our proof.
\section{Main approximation theorem}
Let $l \in \mathbb{N}$ and put $y_{l}(s)= \frac{s-1}{4\pi l}$. Then,
\[
|2\pi ny_{l}(s) -s'|\leq B \iff |n-l|\leq \frac{Bl}{s'}
\]
so that we must have $n=l$ if $l<\frac{s'}{B}$. Putting in the restrictions on $B$ and $y$ in Prop. \ref{Prop 1} gives us
\begin{thm}\label{Theorem 2} There are positive constants $\beta_{1}$ and $\beta_{2}$ such that for all integers $l$ satisfying $\beta_{1} < l < \beta_{2}\sqrt{\frac{s}{\log s}}$, for all $s$ sufficiently large, the numbers $y_{l}(s)= \frac{s-1}{4\pi l}$ satisfy the equation
\[
\frac{\Phi_{f}(s;\alpha,y_{l}(s))}{I_{s}(\frac{s-1}{2})} = \lambda_{f}(l)e(\alpha l) + O(s^{-\delta})
\]
for some $\delta > 0$, uniformly for any real number $\alpha$.
\end{thm}
We now extend this theorem as follows. Let $m \in \mathbb{N}$. We analyse the behaviour of $\Phi_{f}(s +2m;\alpha,y_{l}(s))$. We may use Prop. \ref{Prop 1} without modification provided $m = o(s)$ so that in the sum in (4), we have
\[
|2\pi ny_{l}(s) -\frac{s+2m-1}{2}| \leq B \iff |n-l+ \frac{ml}{s'}| \leq \frac{Bl}{s'}.
\]
If we choose $l \ll \epsilon \sqrt{\frac{s}{\log s}}$ and $m \ll \sqrt{s\log s}$, we must have $n=l$ in our sum, so that under these conditions, we have
\[
\frac{\Phi_{f}(s +2m;\alpha,y_{l}(s))}{I_{s+2m}(s' +m)} = \lambda_{f}(l)e(l\alpha)e^{-\frac{m^2}{s+2m-1}} + O(s^{-\delta}).
\]
The exponential term above is $ 1 + O(s^{-\delta})$, provided we choose $m \ll s^{\frac{1}{2} - \delta}$. Next, we see that
\[
\frac{I_{s+2m}(s'+m)}{I_{s}(s')} = (s' +m)^{m} (1 + O(s^{-\delta})) = s'^{m} +O(s^{m-\delta}).
\]
Combining these estimates gives us
\begin{thm}\label{Theorem3} Let $\delta > 0$ be sufficiently small. For all $s$ sufficiently large (depending on $\delta$), let $m$ be a real number such that $0 \leq m \ll s^{\frac{1}{2} - \delta}$, and let $l \in \mathbb{N}$ satisfy $ 1 \ll l < \delta \sqrt{\frac{s}{\log s}}$. Then,
\[
\frac{\Phi_{f}(s+2m;\alpha,y_{l}(s))}{I_{s}(\frac{s-1}{2})} = (\frac{s-1}{2})^{m}\lambda_{f}(l)e(\alpha l) + O(s^{m-\delta})
\]
\end{thm}
We apply this theorem with $m=1$ and prove the following
\begin{cor}\label{Theorem 4} Let $\delta > 0$ be sufficiently small. For all $k$ sufficiently large (depending on $\delta$), let $l \in \mathbb{N}$ satisfy $ 1 \ll l < \delta \sqrt{\frac{k}{\log k}}$ such that $|\lambda_{f}(l)| \gg k^{-\frac{\delta}{2}}$. Then, we have for $y_{l}= \frac{k-1}{4\pi l}$
\[
\frac{1}{2\pi i}\frac{f'}{f}(\alpha + iy_{l}) = l + O(lk^{-\delta}),
\]
uniformly for all $\alpha \in \mathbb{R}$.
\end{cor}
{\noindent\bf Proof. } This is a direct consequence of (3) and the theorem above.
\begin{rem} It is easy to see that the theorems above hold with $y_{l}(s)$ replaced by $y$ with $|y-y_{l}(s)| \ll \frac{s^{\frac{1}{2}- 2\delta}}{l}$.
\end{rem}
\section{Sign-changes of \mbox{\boldmath$\Phi_{f}(k;\alpha,y)$}: lower bounds.}
We first note that there are no zeros of $f(z)$ if $y > Ck$ for some absolute constant $C$ (one may take $C = \frac{\log 4}{4\pi}$ for example) except for the isolated zero at infinity. This may be deduced directly from the fourier expansion of $f(z)$ by isolating the first fourier coefficient.
To obtain a lower bound for the number of zeros of $f(z)$ on $\delta^{*}$, it sufficies to detect sign-changes of $\Phi_{f}(k;\alpha,y)$ which we recall is real valued when $\alpha = 0$ or $\frac{1}{2}$ (corresponding to
$z=\alpha + iy$ lying on $\delta_{1}$ or $\delta_{2}$ respectively). Our theorems in the previous section are valid only for
\begin{equation}
\sqrt{k\log k} \ll y < \frac{1}{100}k
\end{equation}
and consequently, we restrict our attention to this region. We let $Y$ be a parameter satisfying (11) and we define the Siegel set
\[
\mathcal{F}_{Y} = \{ z = \alpha + iy : -\frac{1}{2}<\alpha \leq \frac{1}{2}, y \geq Y\},
\]
which is a part of the standard fundamental domain containing the cusp. Then, for $z \in \mathcal{F}_{Y} \cap \delta^{*}$, we will determine a lower bound for the number of sign-changes of $\Phi_{f}(k;\alpha,y)$ by detecting sign-changes of $\lambda_{f}(l)e(\alpha l)$ and utilising Theorem \ref{Theorem 2}. To this end, we have to ensure that $\lambda_{f}(l)$ is also not too small.
For the latter, we use the Hecke relations
\begin{equation}
\lambda_{f}(p)^{2} = \lambda_{f}(p^2) +1
\end{equation}
valid for all prime numbers $p \geq 2$. Put $\beta = \frac{\sqrt{5}-1}{2}$. It follows that either $|\lambda_{f}(p)| \geq \beta$ or if not, then $|\lambda_{f}(p^2)| \geq \beta$. We define
\[
\omega = \left\{
\begin{array}{ll}
2 & \quad \mbox{if $|\lambda_{f}(2)| \geq \beta$};\\
4 & \quad \mbox{if $|\lambda_{f}(2)| < \beta$},\\ \end{array} \right.
\]
so that $|\lambda_{f}(\omega)| \geq \beta$.
To obtain a result that is as strong as possible,it would be preferable if we could detect sign-changes of $\lambda_{f}(l)$ for $l$ in suitably short intervals, but as discussed in the introduction, the current methods do not give such sharp results. To circumvent this problem, we look for integers $u$ and $v$ of opposite parity, both in the same short interval such that both $\lambda_{f}(u)$ and $\lambda_{f}(v)$ are not small. The parity condition will then ensure that $\Phi_{f}(k;\alpha,y)$ changes sign for either $\alpha = 0$ or $\alpha = \frac{1}{2}$, for $y$ in a short interval. We will first prove a much weaker result which however has the benefit of localising the detection of sign-changes to each of the lines $\delta_{1}$ and $\delta_{2}$.
\vspace{10pt}
\subsection{Sign-changes on \mbox{\boldmath$\delta_{2}$}.}\
\vspace{10pt}
In this section, we show that there are infinitely many zeros of $f(z)$ on $\delta_{2}$. We will need
\begin{lem}\label{Lemma A} Let $p \geq 2$ be a prime number. For a fixed number $J \geq 1$, there is a constant $B$ depending at most on $J$ and a number $b = b(f,p)$, with $1 \leq b \leq B$ such that if $a = p^b$ we have
\[
\lambda_{f}(a^j) \geq \frac{1}{10}
\]
for all $1 \leq j \leq J$.
\end{lem}
{\noindent\bf Proof. } Recall that there is a number $\theta_{p} = \theta(f,p)$ with $0 \leq \theta_{p} \leq \pi$ such that for all non-negative integers $n$,
\[
\lambda_{f}(p^{n}) = \frac{\sin{\big{(}(n+1)\theta_{p} \big{)}}}{\sin \theta_{p}}.
\]
If $\theta_{p} = 0$ or $\pi$, then we may take $ b = 2$ since $\lambda_{f}(p^{2j}) \geq 3$ for all $j \geq 1$. By continuity, we see that $b=2$ still suffices for $\theta_{p}$ near $0$ or $\pi$. In other words, there is number $\theta_{0} > 0$ depending at most on $J$, such that the conclusion of the lemma holds unless $\theta_{0} < \theta_{p} < \pi - \theta_{0}$, which we now assume.
By Dirichlet's approximation theorem, for any integer $B \geq 1$, there is an integer $1 \leq b \leq B$ such that $\|b\frac{\theta_{p}}{2\pi}\| \leq \frac{1}{B+1}$. Hence, we can find integers $b$ and $b'$ so that $b\theta_{p} = 2\pi b' + \eta$ with $|\eta| \leq \frac{2\pi}{B+1}$ so that
\[
\lambda_{f}(p^{bj}) = \frac{\sin{\big{(}j\eta + \theta_{p} \big{)}}}{\sin \theta_{p}}.
\]
We shall choose $B$ sufficiently large so that $0 < j\eta + \theta_{p} < \pi$ for all $1 \leq j \leq J$. Using estimates for trignometric functions, we conclude that
\[
\lambda_{f}(p^{bj}) \geq 1 - \frac{1}{2}(j\eta)^{2} - j\eta\cot(\theta_{p}) \geq \frac{1}{10}
\]
by choosing $B$ sufficiently large.
\begin{thm}\label{Theorem A} There is a constant $C>0$ such that $f(z)$ has at least $C\log{k}$ zeros on the line $\delta_{2}$ with $z=\frac{1}{2} + iy $ and $ y \geq \sqrt{k\log k}$ for $k$ sufficiently large.
\end{thm}
{\noindent\bf Proof. } Let $X \rightarrow \infty$ with $k$ such that $X \ll (\frac{k}{\log k})^{\frac{1}{4}}$, where the implied constant is chosen suitably so as to satisfy the conditions of Theorem \ref{Theorem 2}. We decompose the interval $[1,X]$ into dyadic subintervals $[(2a)^{i},(2a)^{i+1}]$ with $0 \leq i \leq R$ with $R \gg \log k$ and $a \geq 1$ some integer. Each such subinterval we denote by $\mathcal{I}=[m,2am]$ and by $\mathcal{I}^{2}$ the corresponding subinterval $[m^{2},(2am)^{2}]$. Every interval $[m,2m]$ contains a prime number $q \geq 3$, so that both $q$ and $aq$ lie in $\mathcal{I}$. We call a prime $q$ ``good'' if $|\lambda_{f}(q)|\geq \beta$ (see (12)) and ``bad'' otherwise, in which case $|\lambda_{f}(q^{2})|\geq \beta$.
Suppose $q$ is ``good''. In Lemma \ref{Lemma A}, we take $p=2$ and $J=2$ and choose $a = 2^b$ in the dyadic subdivision above. Then $|\lambda_{f}(q)|\geq \beta$, $|\lambda_{f}(aq)|\geq \frac{\beta}{10}$ and
\[
(-1)^{q}\lambda_{f}(q)(-1)^{aq}\lambda_{f}(aq) < 0.
\]
This shows by Theorem \ref{Theorem 2} that $\Phi_{f}(k,\frac{1}{2},y)$ has a sign-change between $\frac{k-1}{4\pi q}$ and $\frac{k-1}{4\pi aq}$.
Now suppose $q$ is ``bad''. In this case, both $q^{2}$ and $a^{2}q^{2}$ lie in $\mathcal{I}^{2}$, $|\lambda_{f}(q^2)|\geq \beta$, $|\lambda_{f}(a^{2}q^{2})|\geq \frac{\beta}{10}$ and
\[
(-1)^{q^{2}}\lambda_{f}(q^{2})(-1)^{a^{2}q^{2}}\lambda_{f}(a^{2}q^{2}) < 0.
\]
This time we have a sign-change between $\frac{k-1}{4\pi q^{2}}$ and $\frac{k-1}{4\pi a^{2}q^{2}}$.
By considering only subintervals with $\frac{R}{2} \leq i \leq R$, we can ensure that all our subintervals of the type $\mathcal{I}$ and $\mathcal{I}^{2}$ are disjoint, so that there are at least $\frac{R}{2}$ zeros of $f(z)$ with $z=\frac{1}{2} + iy$ and $y \gg \frac{k}{X^{2}}$, from which the theorem follows.
\vspace{10pt}
\subsection{Sign-changes on \mbox{\boldmath$\delta_{1}$}.}\
\vspace{10pt}
We show in this section
\begin{thm}\label{Theorem20} There is a constant $C>0$ such that $f(z)$ has at least $C\log{k}$ zeros on the line $\delta_{1}$ with $z=iy $ and $ y \geq \sqrt{k\log k}$ for $k$ sufficiently large.
\end{thm}
As mentioned in the introduction, the proof is more involved and we require the following proposition that relies on strong subconvexity estimates for $L$-functions. Our notation here will follow that of \cite{KLSW} and \cite{Mato}.
\begin{prop}\label{Prop 3} There is $\epsilon_{0} > 0$ (independent of $k$) such that if $k$ is large enough and $f$ is of even weight $k$, then there is an $n < k^{0.4963}$ ($n$ a power of a prime) such that
\[
\lambda_{f}(n) \leq -\epsilon_{0} .
\]
\end{prop}
{\noindent\bf Proof. } This follows by a modification of the recent developments connected with the first sign-change in $\lambda_{f}(n)$'s for such $f$'s. Fortuitously, the optimization in \cite{Mato} of the smooth number argument in \cite{KLSW} coupled with the subconvex bounds in \cite{Peng} and \cite{J-M} just allows us to secure an exponent less than $\frac{1}{2}$ in the Proposition.
In more detail, if for $\epsilon > 0$, $\lambda_{f}(p^{e})\geq -\epsilon$ for all prime powers $p^{e} \leq y$, that is $\frac{\sin{\big{(}(m+1)\theta_{p} \big{)}}}{\sin \theta_{p}} \geq -\epsilon$, for $m \leq K$ ($K$ will be fixed, say at 100), then
\[
\lambda_{f}(p) \geq 2\cos\big{(}\frac{\pi}{m+1}\big{)} - \eta(\epsilon)
\]
for $p \leq y^{\frac{1}{m}}$. Here, we may choose $\eta(\epsilon) = C_{K}\epsilon$ for a constant $C_{K}$ depending at most on $K$ and with $\epsilon >0$ chosen suitable small in what follows.
Define the multiplicative function $h_{y}$ on squarefree numbers by
\begin{equation}
h_{y}(p) =
\begin{cases}
-2 \hspace{10pt} $if$ \hspace{4pt} p>y, \\
2\cos\big{(}\frac{\pi}{m+1}\big{)} - \eta(\epsilon) \hspace{4pt} $if$ \hspace{4pt} y^{\frac{1}{m+1}} \leq p \leq y^{\frac{1}{m}}, 1 \leq m \leq K\\
2\cos\big{(}\frac{\pi}{K+1}\big{)} - \eta(\epsilon), \hspace{4pt} p \leq y^{\frac{1}{K+1}}.
\end{cases}
\end{equation}
Following the analysis in \cite{Mato} and in particular the continuity of the solution $\sigma(u)$ to the corresponding difference-differential equation, we conclude that uniformly for $\frac{1}{2} \leq u \leq 3$ and if $\epsilon = \epsilon_{0}$ is small enough but fixed
\begin{equation}
\sum_{n\leq y^{u}} h_{y}(n) \geq \big{(} \sigma_{\epsilon_{0}}(u) + o(1)\big{)}y^{u}
\end{equation}
where $\sigma_{\epsilon_{0}}(u) > 0$ for $\frac{1}{4} \leq u \leq 1.3434 := \kappa$.
As in \cite{KLSW}, define the multiplicative function $g_{y}$ on squarefree numbers by
\begin{equation}
\lambda_{f} = g_{y}*h_{y},
\end{equation}
so that
\begin{equation}
g_{y}(p) = \lambda_{f}(p) - h_{y}(p).
\end{equation}
Then, by construction $g_{y}(p) \geq 0$ for all p and hence $g_{y}(n) \geq 0$. Now
\begin{equation}
\sideset{}{^\flat }\sum_{n\leq y^{\kappa}} \lambda_{f}(n) = \sideset{}{^\flat }\sum_{d\leq y^{\kappa}}g_{y}(d) \Big{(}\sideset{}{^\flat }\sum_{l\leq \frac{y^{\kappa}}{d}}h_{y}(l)\Big{)},
\end{equation}
where the sums are over squarefree numbers. Since $h_{y}(l) \geq 0$ for $l \leq y^{\frac{1}{3}}$, it follows that $\sideset{}{^\flat }\sum_{l\leq \xi}h_{y}(l) \geq 0$ if $\xi \leq y^{\frac{1}{3}}$ while the same is true for $\xi > y^{\frac{1}{3}}$ from (22). Hence the coefficients of the sum over $d$ in (25) are all non-negative so that
\begin{equation}
\sideset{}{^\flat }\sum_{n\leq y^{\kappa}} \lambda_{f}(n) \geq \sideset{}{^\flat }\sum_{l\leq y^{\kappa}}h_{y}(l) \geq \frac{1}{2} \sigma_{\epsilon_{0}}(\kappa)y^{\kappa}.
\end{equation}
On the other hand, it follows directly from the subconvex bounds for $L(s,f)$ of \cite{Peng} and \cite{J-M} that for any $\delta > 0$
\begin{equation}
\sum_{n\leq y^{\kappa}} \lambda_{f}(n) \ll _{\delta} k^{\frac{1}{3} + \delta}y^{\frac{\kappa}{2}}.
\end{equation}
Combining (26) and (27) leads to a contradiction if $y > k^{\frac{2}{3\kappa} + \delta'}$, which is the case if we assumed the Proposition to be false.
\begin{lem}\label{Lemma14} Given $\xi \geq 1000$ and any cusp form $f$, there are six integers $m$ in the interval $(\xi,50\xi)$ which are relatively prime in pairs and for which
\[
|\lambda_{f}(m)| \geq \frac{1}{10}.
\]
\end{lem}
{\noindent\bf Proof. } The interval $(\sqrt{\xi},\sqrt{50\xi})$ contains at least 18 primes $p$ and for each either $|\lambda_{f}(p)| \geq \beta$ or $|\lambda_{f}(p^{2})| \geq \beta$ or both (see (20)). If six of these have $|\lambda_{f}(p^{2}| \geq \beta$, then we choose our $m$'s to be these $p^{2}$'s. Otherwise, we can find twelve distinct primes $p_{j}$ with $|\lambda_{f}(p_{j})| \geq \beta$. We now take for our $m$'s the six products $p_{1}p_{2}$, $p_{3}p_{4}$, $...$ , $p_{11}p_{12}$.
\begin{lem}\label{lemma15} Given $\xi \geq 1000$ and $f$, there are relatively prime integers $m_{1}$ and $m_{2}$ in the interval $(\xi,2500\xi)$ such that
\[
\lambda_{f}(m_{j}) \geq \frac{1}{100},\hspace{10pt} j=1,2.
\]
\end{lem}
{\noindent\bf Proof. } Consider the interval $(\sqrt{\xi},50\sqrt{\xi})$. By the previous lemma, there are integers $n_{1}, ..., n_{6}$ in the interval that are relatively prime in pairs such that $|\lambda_{f}(n_{j})| \geq \frac{1}{10}$. Of the three numbers $n_{1}, n_{2}$ and $n_{3}$. at least two have the same sign (we assume the first two) so that $\lambda_{f}(n_{1}n_{2}) \geq \frac{1}{100}$, giving us $m_{1}$ and similarly for $m_{2}$ using the remaining three integers.
\vspace{10pt}
{\noindent \bf Proof of Theorem \ref{Theorem20}.} According to Prop. \ref{Prop 3}, there is an integer $\hat{n}=\hat{n}_{f}$ equal to a prime power $p^{e}$ such that $\hat{n} < k^{0.4963}$ and $\lambda_{f}(\hat{n}) \leq -\epsilon_{0}$ for some fixed $\epsilon_{0} > 0$. Let $\mathcal{I} = (\eta,2500\eta)$ be a subinterval of $(k^{0.4963},\sqrt{\frac{k}{\log k}})$. By Lemma \ref{lemma15}, there is a $m_{1} \in \mathcal{I}$ such that $\lambda_{f}(m_{1}) \geq \frac{1}{100}$. Also applying Lemma \ref{lemma15} but now to the interval $(\frac{\eta}{\hat{n}},2500\frac{\eta}{\hat{n}})$, we find two relatively prime integers $v_{1}$ and $v_{2}$ such that $\lambda_{f}(v_{j}) \geq \frac{1}{100}$ for $j = 1$ and $2$. At least one of the $v_{j}$'s is coprime to $\hat{n} = p^{e}$, say $v_{1}$. We set $m_{2} = \hat{n}v_{1}$ so that $m_{2} \in \mathcal{I}$ and $\lambda_{f}(m_{2}) \leq -\frac{\epsilon_{0}}{100}$. Then, using Theorem \ref{Theorem 2} we find that $f(iy)$ has a sign-change for a $y$ between $\frac{k-1}{4\pi m_{1}}$ and $\frac{k-1}{4\pi m_{2}}$. Since there are $C_{1}\log k$ such disjoint subintervals $\mathcal{I}$ for some positive constant $C_{1}$, we complete our proof.
\vspace{10pt}
\subsection{Sign-changes on \mbox{\boldmath$\delta^{*}$}.}\
\vspace{10pt}
Let $X \ll (\frac{k}{\log k})^{\frac{1}{4}}$ be as in the proof of Theorem \ref{Theorem A}. We let $\mathcal{I}$ denote the interval $(X,\sqrt{\omega}X)$ and for any interval $I=(a,b)$, we denote the interval $(\frac{1}{\sqrt{\omega}}a,\frac{1}{\sqrt{\omega}}b)$ by $\frac{1}{\sqrt{\omega}}I$. The intervals $\frac{1}{\sqrt{\omega}}\mathcal{I}$ and $\mathcal{I}$ will form a disjoint pair in our considerations. For any positive $H$ such that $\frac{X}{H} \rightarrow \infty$, let $\mathcal{I}_{j}$ denote the interval $\big{(}X+jH,X+(j+1)H\big{)}$ with $j=0,1,...,R$ with $R$ chosen so that $\mathcal{I}_{j} \subset \mathcal{I}$ for all $j$. We will use the following
\begin{lem}\label{Lemma 4} There is a $H = H(X) >0$ (as above) such that of the corresponding pairs of subintervals $\{\frac{1}{\sqrt{\omega}}\mathcal{I}_{j},\mathcal{I}_{j}\}$, one can find a positive proportion such that each subinterval of the pair contains at least two (odd) prime numbers.
\end{lem}
We first indicate how we use this lemma to prove
\begin{thm}\label{Theorem31} Let $N_{f}^{Y}(\delta^{*})$ denote the number of zeros of $f(z)$ lying on $\mathcal{F}_{Y} \cap \delta^{*}$ and let $X$ have order of magnitude $\big{(}\frac{k}{Y}\big{)}^{\frac{1}{2}}$. Then, for all $k$ sufficiently large, $N_{f}^{Y}(\delta^{*}) \gg \frac{X}{H}$.
\end{thm}
{\noindent\bf Proof. } Let $\frac{1}{\sqrt{\omega}}\mathcal{I}_{j_{1}}$ and $\mathcal{I}_{j_{1}}$ be a generic such pair satisfying the lemma so that $\frac{1}{\sqrt{\omega}}\mathcal{I}_{j_{1}}$ contains (at least) two primes denoted by $q$ and $q'$ and $\mathcal{I}_{j_{1}}$ contains $p$ and $p'$. We will construct pairs of integers $\{u,v\}$ with $u$ odd and $v$ even such that neither $|\lambda_{f}(u)|$ nor $|\lambda_{f}(v)|$ are too small. We have 4 cases to consider:
\begin{itemize}
\item[(i)] Suppose $|\lambda_{f}(q)|\leq \beta$ or $|\lambda_{f}(q')|\leq \beta$ (we assume q). If both $|\lambda_{f}(p)|$ and $|\lambda_{f}(p')|$ exceed $\beta$, we put
\[
u=pp', \hspace{20pt} v=\omega q^{2}.
\]
Then, by the multiplicativity of $\lambda_{f}(n)$ and the Hecke relations (11)
\[
|\lambda_{f}(u)|\geq \beta^{2},\hspace{20pt} |\lambda_{f}(v)|= |\lambda_{f}(\omega)||\lambda_{f}(q^{2})|\geq \beta^{2}.
\]
Moreover, both $u$ and $v$ lie in the subinterval $\mathcal{J}_{j_{1}}$ where we denote the interval $\big{(}(X+jH)^{2},(X+(j+1)H)^{2}\big{)}$ by $\mathcal{J}_{j}$.
\item[(ii)] If $|\lambda_{f}(q)|\leq \beta$ or $|\lambda_{f}(q')|\leq \beta$ (we assume q) and if either $|\lambda_{f}(p)|$ or $|\lambda_{f}(p')|$ does not exceed $\beta$ (we assume the former), then put
\[
u=p^{2}, \hspace{20pt} v=\omega q^{2},
\]
with the same conclusions as above.
\item[(iii)] Now suppose $|\lambda_{f}(q)|\geq \beta$ and $|\lambda_{f}(q')|\geq \beta$ but with $|\lambda_{f}(p)|\leq \beta$ (or $|\lambda_{f}(p')|\leq \beta$). We choose
\[
u=p^{2}, \hspace{20pt} v=\omega qq',
\]
where this time $|\lambda_{f}(v)| \geq \beta^{3}$ with the condition on $u$ as before.
\item[(iv)] Lastly if all $|\lambda_{f}(.)|$ exceed $\beta$ for the four primes, then we take
\[
u=pp', \hspace{20pt} v=\omega qq',
\]
with estimates involving $u$ and $v$ as before.
\end{itemize}
By Lemma \ref{Lemma 4}, we conclude that there are $\gg R$ disjoint subintervals $\mathcal{J}_{j}$ that contain a pair of integers $\{u,v\}$ such that $|\lambda_{f}(u)|,|\lambda_{f}(v)| \geq \beta^{3}$, with $u$ odd and $v$ even. These disjoint subintervals are contained in the interval $\mathcal{J}=(X^{2},4X^{2})$ so that the conditions of Theorem \ref{Theorem 2} are satisfied with $u$ and $v$ taking the value $l$. We put $y_{u}=\frac{k-1}{2\pi u}$ and $y_{v}=\frac{k-1}{2\pi v}$ and observe that our conditions on $X$ and $Y$ ensure that they exceed $Y$. Using Theorem \ref{Theorem 2} with the two values of $\alpha = 0$ and $\frac{1}{2}$, we see that the size and sign of $\Phi_{f}(k;0,y_{u})$ and $\Phi_{f}(k;0,y_{v})$ are determined by the pair $\lambda_{f}(u)$ and $\lambda_{f}(v)$ respectively. On the other hand, the size and sign of $\Phi_{f}(k;\frac{1}{2},y_{u})$ and $\Phi_{f}(k;\frac{1}{2},y_{v})$ are determined by the pair $-\lambda_{f}(u)$ and $\lambda_{f}(v)$ (due to the parity difference). Consequently, without any further input, we can conclude that at least one of these pairs must be of opposite sign, implying that either $\Phi_{f}(k;0,y)$ or $\Phi_{f}(k;\frac{1}{2},y)$ has a zero for some $y$ between $y_{u}$ and $y_{v}$. Thus there are at least $\frac{1}{2}R$ zeros of $f(z)$ on either $\delta_{1}$ or $\delta_{2}$ and since $R \asymp \frac{X}{H}$, the theorem follows.
\vspace{10pt}
{\noindent \bf Proof of Lemma \ref{Lemma 4}.} To verify the lemma, we appeal to the well-known result that for the interval $(A,2A)$, for $A$ large enough, there is a number $J$ depending on $A$ such that almost all subintervals of length $J$ in $(A,2A)$ have at least one prime number. It is easy to check (by combining consecutive subintervals and then discarding subintervals amongst the pairs that do not satisfy the conditions of the lemma) that the lemma is satisfied with $H=\frac{1}{4}J$, for example.
If one assumes the Riemann Hypothesis (for the Riemann zeta-function), then Selberg \cite{Sel} showed in 1943 that one may take $J=g(A)(\log A)^{2}$ for any function $g(A)$ tending to infinity with $A$. This was improved by Heath-Brown \cite{HB} to $J=g(A)\log A$ subject to additional assumptions on the vertical distribution of zeros of the zeta-function. The best uncondtional result to date is due to Jia \cite{Jia}, where $J=A^{\frac{1}{20} + \epsilon}$ (this follows a sequence of similar results by Harman, Watt and Li). As a consequence, we have our
\begin{cor}\label{Corl 1}
\mbox{}
\begin{itemize}
\item[(i)] Let $N_{f}^{Y}(\delta^{*})$ denote the number of zeros of $f(z)$ lying on $\mathcal{F}_{Y} \cap \delta^{*}$. Then, for any $\epsilon >0$ and all $k$ sufficiently large, $N_{f}^{Y}(\delta^{*}) \gg \big{(}\frac{k}{Y}\big{)}^{\frac{1}{2} - \frac{1}{40} -\epsilon}$. If one assumes the Riemann Hypothesis for the Riemann zeta-function, then $N_{f}^{Y}(\delta^{*}) \gg \big{(}\frac{k}{Y}\big{)}^{\frac{1}{2}-\epsilon}$ .
\item[(ii)] If $N_{f}(\delta^{*})$ denotes the number of zeros of $f(z)$ lying on $\delta^{*}$, then for any $\epsilon >0$ and all $k$ sufficiently large, $N_{f}(\delta^{*}) \gg k^{\frac{1}{4} - \frac{1}{80} -\epsilon}$. Moreover, on the Riemann Hypothesis for the Riemann zeta-function, one has $N_{f}(\delta^{*}) \gg k^{\frac{1}{4} -\epsilon}$.
\end{itemize}
\end{cor}
\begin{rem}
One can give an alternative construction of the numbers $u$ and $v$ used in the proof of Theorem \ref{Theorem31}
above using Lemma \ref{Lemma A} which we do as follows:
\end{rem}
Let $X$ be as before and we consider the intervals $\mathcal{I}_{j}$ and $\mathcal{J}_{j}$ as in the proof of Theorem \ref{Theorem31} (we drop the subscripts in what follows). We choose a number $a$ satisfying Lemma \ref{Lemma A} with $p=2$ and $J=4$ and consider the six subintervals $\mathcal{I}$, $a \mathcal{I}$, $a^{2}\mathcal{I}$, $\mathcal{J}$, $a^{2} \mathcal{J}$ and $a^{4}\mathcal{J}$. We also choose $H$ so that there are prime numbers in each of the subintervals $\mathcal{I}$, $a \mathcal{I}$ and $a^{2}\mathcal{I}$, denoted by $p_{1}$, $p_{2}$ and $p_{3}$ respectively. We consider ``good'' and ``bad'' primes as in the proof of Theorem \ref{Theorem A}. Clearly at least two of the primes $p_{1}$, $p_{2}$, $p_{3}$ are ``good'' (which we call case I) or at least two are bad (case II).
In case I, we have that either $a \mathcal{I}$ or $a^{2}\mathcal{I}$ contains two elements $u = a^{j}p$ with $j = 1$ or $2$, and $v=p'$ with both $p$ and $p'$ ``good" odd primes. These numbers are of odd parity and satisfy a lower-bounds of the type described in the proof of Theorem \ref{Theorem31} and so can be used to detect sign-changes. Similarly, in case II, $a^{2} \mathcal{J}$ or $a^{4}\mathcal{J}$ contains two elements $u = a^{2j}p^{2}$ with $j = 1$ or $2$, and $v=(p')^{2}$ with both $p$ and $p'$ ``bad'' odd primes and in this case we get a similar conclusion. The rest of the argument follows that given in Theorem \ref{Theorem31}.
\section{Bounds for the zeros of $\bf{f(z)}$ in \mbox{\boldmath$\mathcal{F}_{Y}$}. }
We now prove a conditional result that gives the precise number of zeros of $f(z)$ in some special Siegel sets, from which we obtain unconditional upper and lower bounds for the general case.
\begin{thm}\label{Theorem 41} Let $\delta>0$ and k be sufficiently large. Suppose there exists an integer $l$ satisfying $ 1 \ll l < \delta \sqrt{\frac{k}{\log k}}$ such that $|\lambda_{f}(l)| \gg k^{-\frac{\delta}{2}}$. Then there are exactly $l$ zeros of $f(z)$ in the region $\{z=\alpha +iy: -\frac{1}{2}<\alpha \leq \frac{1}{2}, y \geq \frac{k-1}{4\pi l}\}$.
\end{thm}
{\noindent\bf Proof. } We integrate $\frac{1}{2\pi i}\frac{f'}{f}(z)$ along the boundary of the indicated region to count the number of zeros (one makes the standard indentations to avoid zeros on the vertical paths). We observe that the vertical integrals cancel due to periodicity and opposite orientations, and that there are no zeros on the horizontal path. Consequently, the number of zeros in our region is
\[
\frac{1}{2\pi i}\int_{-\frac{1}{2}}^{\frac{1}{2}} \frac{f'}{f}(x + iy_{l}) \, dx \hspace{10pt} = \frac{1}{2\pi i}\big{(} \log f(\frac{1}{2} + iy_{l}) - \log f(-\frac{1}{2} + iy_{l})\big{)}
\]
\[
= l + O(k^{-\delta}),
\]
by Theorem \ref{Theorem 2}, from which our result follows.
\begin{cor}\label{Corl 5} For sufficiently large $k$, suppose there exists an integer $1 \ll L_{k} \ll \sqrt{\frac{k}{\log k}}$ such that for all integers $1 \leq l \leq L_{k}$, $|\lambda_{f}(l)| \geq \frac{1}{\Delta_{k}}$, where $\Delta_{k} > 0$ also satisfies the condition $\frac{\log \Delta_{k}}{\log k} \rightarrow 0$. Then, all the zeros of $f(z)$ in the region $\mathcal{F}_{Y}$ are real zeros with $Y = A \frac{k}{ L_{k}}$ for a sufficiently large number $A$.
Moreover, if $L_{k}^{+}$ denotes the number of $l$'s above satisfying $\lambda_{f}(l)\lambda_{f}(l+1) > 0$, then $L_{k}^{+}$ of the real zeros above lie on $\delta_{2}$, $L_{k} -1 - L_{k}^{+}$ lie on $\delta_{1}$ and they are all necessarily simple zeros.
\end{cor}
\begin{rem}
From this we expect that asymptotically half of the zeros lie on each segment. As we noted in the introduction, this together with a standard expectation that the $\lambda_{f}(l)$'s are not zero (see for example \cite{F-J}) suggests that perhaps all the zeros of $f(z)$ in $y \gg \sqrt{k}$ are real. There is some numerical evidence provided by Fredrik Str\"omberg that bears this out.
\end{rem}
{\noindent\bf Proof. } This follows directly from the above Theorem by looking at the intersection of the Siegel domains for consecutive values of $l$. There is only one zero in such an intersection, and since the zeros are symmetric with respect with the line $\Re(z) = 0$, it must lie on the boundary. Moreover, if $\lambda_{f}(l)\lambda_{f}(l+1) > 0$, then this zero is located on $\delta_{2}$ by Theorem \ref{Theorem 2} and the conclusions follow.
\begin{cor}\label{Corl 2} Suppose $\sqrt{k\log k} \ll Y < \frac{1}{100}k$. Let $N_{f}^{Y}$ denote the number of zeros of $f(z)$ in the region $\mathcal{F}_{Y}$. Then there are absolute positive constants $C_{1}$ and $C_{2}$ such that $C_{1}\frac{k}{Y} \leq N_{f}^{Y} \leq C_{2}\frac{k}{Y}$.
\end{cor}
{\noindent\bf Proof. } Put $X= \sqrt{\frac{k-1}{4\pi Y}}$. By the proof of Theorem \ref{Theorem31}, we see that there are integers $u$ and $u'$ with $u \in (\frac{1}{2}X^{2},X^{2})$ and $u' \in (X^{2},4X^{2})$ satisfying $|\lambda_{f}(u)|, |\lambda_{f}(u')|\geq \beta^{3}$. Then, $ y_{u'} < Y < y_{u}$, where we recall that $y_{l} = \frac{k-1}{4\pi l}$. Applying Theorem \ref{Theorem 41}, we conclude $N_{f}^{Y} \leq N_{f}^{y_{u'}} = u' \ll \frac{k}{Y}$, and $N_{f}^{Y} \geq N_{f}^{y_{u}} = u \gg \frac{k}{Y}$, from which our corollary follows.
\section{A probabilistic model.}
We determine a probabilistic model to predict the expected number of zeros of $f(z)$ on our curves. To consider the distribution of zeros on $\delta_{1}$ and $\delta_{2}$ we may generalise our problem to any vertical line segment $\mathcal{L}_{\alpha}$ consisting of points $z=\alpha + iy$ with $y > c$ for any fixed number $c>0$ and $0 \leq |\alpha| \leq \frac{1}{2}$. We ask for the number of zeros of $f(z)$ as $k$ becomes unbounded. The analysis in the previous sections focused on the part of the lines with $y > \sqrt{k\log k}$, and in fact Corollary \ref{Corl 2} gives us an upper-bound of at most $\sqrt{k}$. In what follows, we will first focus on the range $1 < y \ll \sqrt{k}$.
Writing $\frac{f(\alpha _ iy)}{a_{f}(1)} = \sum_{1}^{\infty} \lambda_{f}(n)e(n\alpha)n^{\frac{k-1}{2}}e^{-2\pi ny}$, we denote the real part by $R_{f}(\alpha,y)$ and the imaginary part by $I_{f}(\alpha,y)$. Since $\sum_{n=1}^{N} \lambda_{f}(n) = o(N)$ and $\sum_{n=1}^{N} |\lambda_{f}(n)|^{2} \sim c(f)N$ for some constant $c(f) > 0$, we replace the $\lambda_{f}(n)$ with independent standard normal random coefficients with mean zero and variance 1 (the constant c(f) plays no role in the subsequent analysis and so may be absorbed in $f$). Following the general principles as shown in Edelman-Kostlan \cite{EK}, we consider the vectors
\[
\mathbf{v} = \mathbf{v}(\alpha,y) = \sum_{n=1}^{\infty} \cos({2\pi n\alpha}) n^{\frac{k-1}{2}}e^{-2\pi ny} \mathbf{e}_{n}
\]
and
\[
\mathbf{w} = \mathbf{w}(\alpha,y) = \sum_{n=1}^{\infty} \sin({2\pi n\alpha}) n^{\frac{k-1}{2}}e^{-2\pi ny} \mathbf{e}_{n},
\]
where $\mathbf{e}_{n} = (...,0,0,1,0,0,...)$ denotes the vector with the $1$ in the nth-coordinate. If $\mathbf{u}$ denotes such a vector function, not identically zero, then the probablity density function for the real zeros of the the associated random wave function is given by
\begin{equation}
\mathcal{P}(\mathbf{u}) = \mathcal{P}(\mathbf{u},\alpha,y) = \frac{1}{\pi} \sqrt{\frac{<\mathbf{u},\mathbf{u}><\mathbf{u'},\mathbf{u'}> - <\mathbf{u},\mathbf{u'}>^{2}}{<\mathbf{u},\mathbf{u}>^{2}}}
\end{equation}
where $\mathbf{u'}$ is the derivative with respect to $y$ and with the standard inner-product. This is the expected number of real zeros of the associated wave function per unit length at the point $y$. We need to compute both $\mathcal{P}(\mathbf{v},\alpha,y)$ and $\mathcal{P}(\mathbf{w},\alpha,y)$. Let
\begin{equation}
S(k,\alpha,y) = \sum_{n=1}^{\infty} n^{k-1}\cos({4\pi n\alpha})e^{-4\pi ny}.
\end{equation}
Then
\[
<\mathbf{v},\mathbf{v}>= \frac{1}{2}(S(k,0,y) + S(k,\alpha,y)),
\]
\[
<\mathbf{v},\mathbf{v}'> = -\pi (S(k+1,0,y) + S(k+1,\alpha,y)),
\]
and
\[
<\mathbf{v}',\mathbf{v}'> = 2\pi ^{2}(S(k+2,0,y) + S(k+2,\alpha,y)).
\]
The analogous formulae for $\mathbf{w}$ are the same except the sum of the $S$ functions are replaced with their difference. To determine the asymptotics of $S(k,\alpha,y)$, we apply the Poisson summation formula to the series $S(k,0,y - i\alpha)$ and take real parts so that we get
\[
S(k,\alpha,y) = \Re \frac{\Gamma (k)}{(2\pi i)^{k}} \sum_{h \in \mathbb{Z}} \frac{1}{(2\alpha + h + 2iy)^{k}}.
\]
We truncate this sum with $|h| < 2y$ and see easily that the tailend is bounded by $\frac{y}{(2y)^{k}k}$. If $2\alpha $ is not zero modulo one (that is if $\alpha \neq 0, \pm \frac{1}{2})$) for a fixed $\alpha $, the contribution from $|h| < 2y$ is bounded by $\frac{y}{(2y)^{k}\sqrt{k}}$ so that we conclude that the sum $S(k,\alpha,y)$ is negligible for $1<y<\epsilon \sqrt{k}$. Thus, the inner-products are all determined asymptotically by $S(*,0,y)$ for all $\alpha$. It is easy to evaluate $S(k,0,y)$ using the elementary fact that if a positive function $g(x)$ is increasing in the interval $0 \leq x < x_{0}$ and decreasing for $x > x_{0}$, then
\[
|\sum_{n=0}^{\infty} g(n) - \int_{0}^{\infty} g(x) \, dx| \ll g(x_{0}).
\]
We apply this with $x_{0} = \frac{k-1}{4\pi y}$, to get
\[
S(k,0,y) = \frac{\Gamma(k)}{(4\pi y)^{k}} \big{(}1+ O(\frac{y}{\sqrt{k}})\big{)},
\]
where here we have used Stirling's formula to estimate the error-term. To apply this result to our distribution function, we have to assume that $\frac{y}{\sqrt{k}}$ tends to zero so that $c<y<\epsilon \sqrt{k}$, with $\epsilon$ positive and sufficiently small. Then we have
\begin{equation}
\mathcal{P}(\mathbf{u},\alpha,y) \sim \frac{1}{2\pi} \frac{\sqrt{k}}{y}.
\end{equation}
for all fixed $\alpha \neq 0, \pm \frac{1}{2}$ with $\mathbf{u} = \mathbf{v}$ and $ \mathbf{w}$. If $\alpha = 0, \pm \frac{1}{2}$, then $\mathbf{u} = \mathbf{v}$.
\vspace{10pt}
We next model the distribution of zeros on the segment $\delta_{3}$. A direct approach is unnecessary since one can map the segment $\delta_{3}$ onto the line $\Re(z)=\frac{1}{2}$ by the M\"{o}bius transformation $\sigma(z) = \frac{z}{z+1}$. This sends $\delta_{3}$ to the segment $z= \frac{1}{2} + iy$ with $\frac{1}{2} \leq y \leq \frac{\sqrt{3}}{2}$ and the number of zeros of $f(z)$ is the same in both segments. Thus we may apply the analysis above to obtain the same density function (30), which on integrating gives us (8).
\vspace{10pt}
It remains for us to consider the analog of the above when $y \gg \sqrt{k}$. The situation is quite different than in the case $y \ll \epsilon \sqrt{k}$ since only a few terms dominate in the sums $S(k,\alpha,y)$, as can be seen by Theorem \ref{Theorem 2}. We will consider the case with $\alpha = 0$ (the case $\alpha=\frac{1}{2}$ is the same) and we will assume $y \gg k^{\frac{1}{2} + \delta}$ with $\delta$ small. We first rewrite $$<\mathbf{v},\mathbf{v}><\mathbf{v'},\mathbf{v'}> - <\mathbf{v},\mathbf{v'}>^{2}$$ in (28) as
\begin{equation}
4\pi^{2}\sum_{m=2}^{\infty}\sum_{n=1}^{m-1} (mn)^{k-1}(m-n)^{2}e^{-4\pi (m+n)y}.
\end{equation}
The summand, considered as a function of real variables $m$ and $n$ has a maximum at the point $m_{0} = \frac{k + \sqrt{k}}{4\pi y}$ and $n_{0} = \frac{k - \sqrt{k}}{4\pi y}$, and then has exponential decay beyond a small neighbourhood of this point, as $k$ becomes unbounded. The denominator in (28) also localises in a similar manner so that in (29), a maximum occurs at the real value $n= \frac{k-1}{4\pi y}$. Let us put $t = \frac{k-1}{4\pi y}$ and let $l$ denote the integer closest to $t$ in what follows. We will assume that $y \gg k^{\frac{1}{2}+\delta}$, so that necessarily $1\leq l \ll k^{\frac{1}{2}-\delta}$. We observe then that the closest integer to $m_{0}$ and $n_{0}$ is precisely $l$ and see that only the following integer pairs $(l+1,l), (l+1,l-1)$ and $(l,l-1)$ contribute to the sums in (31) while the other terms are exponentially small. Similarly, the integers $l-1, l$ and $l+1$ contribute to $S(k,0,y)$ in (29). We thus conclude that
\[
\mathcal{P}(\mathbf{v},0,y)^{2} \sim 4 \frac{(1-\frac{1}{l})^{k-1}e^{4\pi y} + 4(1-\frac{1}{l^{2}})^{k-1} + (1+\frac{1}{l})^{k-1}e^{-4\pi y}}{\Big{(}(1-\frac{1}{l})^{k-1}e^{4\pi y} + 1 + (1+\frac{1}{l})^{k-1}e^{-4\pi y}\Big{)}^{2}}.
\]
Our goal is to compute the integral of $\mathcal{P}(\mathbf{v},0,y)$ between say, $Y_{1}<y<Y_{2}$ to give us the expected number of zeros for $y$ on that range on $\delta_{0}$. Thus, we write
\begin{equation}
N := \int_{Y_1}^{Y_2} \mathcal{P}(\mathbf{v},0,y) \, dy = \sum_{l = L_2}^{L_1} \int_{l-\frac{1}{2}}^{l+\frac{1}{2}} \mathcal{P}(\mathbf{v},0,\frac{k-1}{4\pi t}) \Big{(}\frac{k-1}{4\pi t^2}\Big{)} \, dt,
\end{equation}
where $L_{i} \sim \frac{k-1}{4\pi Y_{i}}$.
To evaluate (32) asymptotically, it simplifies our analysis to assume that $k^{\frac{1}{3} + \delta} \ll l \ll k^{\frac{1}{2} - \delta}$, so that we assume $ k^{\frac{1}{2} + \delta} \ll y \ll k^{\frac{2}{3} - \delta}$. This allows us to use approximations of the type
\[
(1-\frac{1}{l})^{k-1} \sim e^{-\frac{k-1}{l} - \frac{k-1}{2l^2}}
\]
so that we obtain from (32) after some simplifications
\[
N \sim 4 \sum_{l = L_2}^{L_1}(\frac{k-1}{4\pi l^{2}}) \int_{0}^{\frac{1}{2}} \sqrt{\frac{e^{\sigma(\frac{1}{2} - u)} + 4 + e^{\sigma(\frac{1}{2} + u)}}{(e^{-\sigma u} + e^{\frac{1}{2}\sigma} + e^{\sigma u})^2}} \, du,
\]
where we made the change of variable $ t = l + u$ and have written $\sigma = \frac{k-1}{l^{2}}$. Noting that $\sigma \rightarrow \infty $, the integral above is well aproximated by
\[
\int_{0}^{\frac{1}{2}} \frac{e^{\frac{1}{2}\sigma(\frac{1}{2} + u)}}{e^{-\sigma u} + e^{\frac{1}{2}\sigma} + e^{\sigma u}} \, du \sim \frac{\pi}{2\sigma},
\]
so that the expected number of zeros for $f(iy)$ for $y \geq Y$ with $ k^{\frac{1}{2} + \delta} \ll Y \ll k^{\frac{2}{3} - \delta}$ is $\frac{1}{2}\frac{k}{4\pi Y}$ (see Corollary \ref{Corl 5}). We expect that a more refined analysis will verify this result for all $Y \gg k^{\frac{1}{2} + \delta}$. Thus the random coefficient model predicts that for $Y \gg k^{\frac{1}{2} + \delta}$
\begin {equation}
|\mathcal{Z}(f) \cap \mathcal{F}_{Y} \cap \delta_{1}| \sim |\mathcal{Z}(f) \cap \mathcal{F}_{Y} \cap \delta_{2}| \sim \frac{1}{2}|\mathcal{Z}(f) \cap \mathcal{F}_{Y}|.
\end{equation}
\vskip 0.5in
|
\section{Introduction}
Models of social networks are structures made up of individuals that are
tied based on their interdependency. Such models explain the confidence or
influence flow in populations without relying on detailed social
psychological findings. The process of opinion dynamics evolves along the
networks of social confidence or influence and affects the structure of the
network itself.
A common feature among
many models of opinion dynamics is bounded confidence or influence, which
means that an individual only interacts with those whose opinions are close
enough to its own. This idea reflects the psychological concept called
\emph{selective exposure} \cite{WM-FC-RS:08}. Broadly defined, ``selective
exposure refers to behaviors that bring the communication content within
reach of one's sensory apparatus'' \cite{DZ-JB:85}.
In the field of social networks, opinion dynamics is of high interest in
many areas including: politics, as in voting prediction \cite{EBN:05};
physics, as in spinning particles \cite{EN-PK-FV-SR:03}; sociology, as in
the diffusion of innovation \cite{TV:96}, the electronic exchange of personal
information \cite{EM-NL-KF-RP-HL-MM-SE-XZ:08}, and language change
\cite{JT-MT-AN:07,FC-SS:07}; and finally economics, as in price
change \cite{KS-RW:02}.
\subsection{Literature Review}
The study of opinion dynamics and social networks goes back to the early
work by J.R.P.\ French \cite{JF:56} on ``A Formal Theory of Social Power.''
This work explores the patterns of interpersonal relations and agreements
that can explain the influence process in groups of agents. Subsequently,
F.\ Harary provides a necessary and sufficient condition to reach a
consensus in French's model of power networks \cite{FH:59}. The modeling of
``continuous opinion dynamics'', in which opinions are represented by real
positive numbers, is initially studied in \cite{MS:61,SC-ES:77,CJ-EJ:86}.
In contrast to the classical case of ``binary opinion dynamics''
\cite{GW-GB:02,SG:97}, the continuous case deals with the problem of what
happens to the worthiness of a choice or the probability of choosing one
decision over another. Recently, \emph{bounded confidence} (BC) models of
opinion dynamics, a label coined by Krause in 1998 \cite{UK:00}, have
received significant attention. BC models are models of continuous opinion
dynamics in which agents have bounded confidence in others opinions. The
first version of BC models was formulated by Hegselmann and Krause
\cite{RH-UK:02}, called HK model, where agents synchronously update their
opinions by averaging all opinions in their confidence bound. The other
popular version of BC models was developed and investigated by Deffuant and
Weisbuch \cite{GW-GD-FA-JN:02}, called DW model. The HK and DW models are
very similar, they differ in their update rule: in DW model a
pairwise-sequential updating procedure is employed instead of the
synchronized one.
In the HK model, the set of neighbors of the $i$th agent is defined as
those agents whose opinions differ from the $i$th opinion by less than the
$i$th confidence bound. Hence, this model is dealing with
\emph{endogenously} changing topologies, that is, state dependent or
changing from inside, in contrast to the \emph{exogenously} changing
topologies. For instance,
\cite{AJ-JL-ASM:02,LM:05,FC-SS:07,MC-ASM-BDOA:06} study a synchronized
linear averaging model with time-dependent exogenously changing
topologies. The HK models are classified based on various factors: a model
is called agent- or density-based if its number of agents is finite or
infinite, respectively; and a model is called homogeneous or heterogeneous
if its confidence bounds are uniform or agent-dependent, respectively. The
convergence of both agent- and density-based homogeneous HK models are
discussed in \cite{VDB-JMH-JNT:09}. The agent-based homogeneous HK system
is proved to reach a fixed state in finite time \cite{JD:01}, the time
complexity of this convergence is discussed in \cite{SM-FB-JC-EF:05m}, its
stabilization theorem is given in \cite{JL:05a}, and its rate of
convergence to a global consensus is studied in \cite{AO-JNT:07}. The
heterogeneous HK model is studied by Lorenz who reformulated the HK
dynamics as an interactive Markov chain \cite{JL:06b} and analyzed the
effects of heterogeneous confidence bounds \cite{JL:09}. %
The convergence of the agent-based heterogeneous HK systems is
experimentally observed, but its proof is still an open problem.
\subsection{Contributions}
In this paper, to distinguish between the HK and DW models, we call a
discrete-time agent-based heterogeneous HK model a \emph{synchronized
bounded confidence} (\emph{SBC}) model. Additionally, we introduce a
model similar to the SBC model and call it the \emph{synchronized bounded
influence} (\emph{SBI}) model. The difference is that in an SBI model the
set of neighbors of the $i$th agent is defined as those agents $j$ whose
influence range contain the $i$th agent's opinion. We analyze SBC and SBI
models with heterogeneous bounds of confidence or influence, respectively.
Indeed, if the SBC and SBI models have agents with homogeneous bounds, then
both models are equivalent to the homogeneous HK model. These
heterogeneous models of opinion dynamics, in spite of the considerable
complexity of their dynamics, describe features of a real society that
cannot be explained by homogeneous models. Specifically, the behavior of a
heterogeneous opinion dynamics model is richer than that of a homogeneous
model in the following ways: (1) an agent may trust an individual but not
be trusted back by that same individual; (2) in steady state, an agent can
keep its own opinion constant while listening to dissimilar opinions,
(whereas, in the steady state of a homogeneous society, any two agents are
either disconnected or in consensus); (3) one can observe
\emph{pseudo-stable} configurations, i.e., ``configurations that have a
subset of agents that is stationary, and the rest are the reason for
further dynamics'' \cite{DU:03}; (4) it is possible for two disconnected
agents to reconnect, and an agent with a large bound of confidence or
influence can pull clusters of agents towards or away from each other; (5)
the order of opinions is not preserved along the system evolution; (6)
"given an average level of confidence, diversity of bounds of confidence
enhances the chances for consensus" \cite{JL:09}; and (7) convergence in
infinite time is possible.
Based on numerical evidence, we formulate our main conjecture that along
the evolution of an SBC or SBI system there exists a finite time, after
which the topology of the interconnection network remains unchanged, and as
a result, the trajectory converges to a limiting opinion vector. We also
observe that each trajectory either reaches a fixed state in finite time or
exhibits a pseudo-stable behavior. This observation is verified assuming
that the main conjecture is true. Furthermore, the following results put
together partly prove our main conjecture: (1) We design an appropriate
classification of agents in both SBC and SBI systems. This classification
is a function of state-dependent interconnection topology of the system,
and can explain the observed pseudo-stable behavior. (2) We introduce the
new notion of \emph{final value at constant topology}, and based on our
classification, we formulate the map under which this value is an image of
the current opinion vector. The set of final values at constant topology is
a superset of the equilibria of the system. We derive necessary and
sufficient conditions for the final value at constant topology to be an
equilibrium vector. (3) For each equilibrium opinion vector, we define its
\emph{equi-topology neighborhood} and \emph{invariant equi-topology
neighborhood}. We show that if a trajectory enters the invariant
equi-topology neighborhood of an equilibrium vector, then it remains
confined to its equi-topology neighborhood, and sustains an interconnection
topology equal to that of the equilibrium vector. This fact establishes a
novel and simple sufficient condition under which: the initial opinion
vector converges to a steady state; the topology of the interconnection
network remains unchanged; and the limiting opinion vector is equal to the
final value at constant topology of the initial opinion vector. (4) We
explore some interesting behavior of classes of agents when they update
their opinions under fixed interconnection topology for infinite time. For
instance, we compute agents rates and directions of convergence, and show
the existence of a leader group for each group of agents that determines
the follower's rate and direction of convergence.
In our extensive simulation results, we observe that for uniformly randomly
generated initial opinion vector and bounds vector, the SBC and SBI
trajectories eventually satisfy our novel sufficient condition for
convergence with probability one. We give some intuitive explanation for
this observation. Finally, we conjecture that the SBI trajectories reach a
fixed state in finite time more often than the SBC trajectories. To
substantiate this conjecture, we present a sufficient condition for SBC and
SBI systems separately that guarantees reaching an \emph{agreement opinion
vector}, which occurs in finite time and is a general case of reaching a
global consensus. Based on these sufficient conditions, we explain our last
conjecture.
\subsection{Organization}
This paper is organized as follows. In Section~\ref{model}, the
mathematical models, conjectures, agents classification, and spectral
properties of the adjacency matrices are presented. In
Section~\ref{final_value}, the final value at constant topology is
introduced and characterized. Section~\ref{sufficientcond} contains the
novel sufficient condition for constant topology and convergence, and for
convergence in finite time. In Section~\ref{expsec}, the simulation
results and intuitive explanations are presented. In
Section~\ref{fixedtopbehavior}, the behavior of the system assuming that
its interconnection topology remains unchanged in a long run is
analyzed. Finally, Section~\ref{conclusion} contains the conclusion and
open questions.
\section{Mathematical Models}\label{model}
Consider $n$ interacting agents and assume that each agent's opinion is
expressed by a real number, say $y_i$ for agent $i\in\until{n}$. In
\emph{bounded confidence interaction}, the opinion $y_i$ is affected by the
opinion $y_j$ if $|y_i-y_j| \le r_i$, where the positive number $r_i$ is the
\emph{confidence bound} of agent $i$. In \emph{bounded influence
interaction}, the opinion $y_i$ is affected by the opinion $y_j$ if
$|y_i-y_j| \le r_j$, where the positive number $r_j$ is the \emph{influence
bound} of agent $j$. The \emph{opinion vector} $y\in \mathbb{R}^n$ and the
\emph{bounds vector} $r\in\mathbb{R}^n_{>0}$ are obtained by stacking all
$y_i$'s and $r_i$'s, respectively. We associate to each opinion vector $y$
two digraphs, both with nodes $\until{n}$ and edge set defined as follows:
denoting the set of out-neighbors of node $i$ by $\mathcal{N}_i (y)$
\begin{itemize}
\item in a \emph{synchronized bounded confidence} (SBC) digraph,
$\mathcal{N}_i (y) = \{ j\in\until{n} : |y_i - y_j| \le r_i \}$; and
\item in a \emph{synchronized bounded influence} (SBI) digraph,
$\mathcal{N}_i (y) = \{ j\in\until{n} : |y_i - y_j| \le r_j \}$.
\end{itemize}
We let $G_r(y)$ denote one of the two \emph{proximity digraphs}, its
precise meaning being clear from the context.
We associate to the SBC and SBI digraphs two dynamical systems, called the
\emph{SBC} and \emph{SBI systems} respectively. Both dynamical systems
update a trajectory $x: \mathbb{N} \rightarrow \mathbb{R}^n$ according to the
discrete-time and continuous-state rule
\begin{equation} \label{system}
x(t + 1) = A(x(t)) x(t),
\end{equation}
where the $i,j$ entry of the \emph{adjacency matrix} $A(y) \in
\mathbb{R}^{n\times n}$ for any $y \in \mathbb{R}^n$ is defined by
\begin{align*}
a_{ij} (y) & =
\begin{cases}
\frac{1}{\displaystyle |\mathcal{N}_i (y)|}, & \mbox{if } j \in \mathcal{N}_i(y),\\
0, & \mbox{if } j \notin \mathcal{N}_i(y),
\end{cases}
\end{align*}
and $|\mathcal{N}_i (y)|$ is the cardinality of $\mathcal{N}_i (y)$. Note
that $i \in \mathcal{N}_i (y)$, in other words, every agent has some
self-confidence or self-influence. This assumption is a key factor in the
convergence of infinite products of adjacency matrices \cite{JL:06}. In the
following, we present our conjectures on SBC and SBI
systems, and the trajectories of Figure~\ref{evolutionfig} support these
conjectures.
\begin{conjecture}[Existence of a limiting opinion vector]\label{xinfconj}
Every trajectory of an SBC or SBI system converges to a limiting opinion
vector.
\end{conjecture}
\begin{conjecture}[Constant-topology in finite time]\label{conjtau}
For any trajectory $x(t)$ of an SBC or SBI system, there exists a finite
time $\tau$ after which the state-dependent interconnection topology, or
equivalently $G_r(x(t))$, remains constant.
\end{conjecture}
Before proceeding, let us define a term borrowed from \cite{DU:03}. A trajectory $x(t)\in\mathbb{R}^n$ that is converging to limiting opinion vector $x_\infty\in\mathbb{R}^n$ is said to have a \emph{pseudo-stable} behavior after $\tau$, if the node set $\mathcal{V} = \until{n}$ is composed of two non-empty subsets $\mathcal{V}_{\text{fixed}}$ and $\mathcal{V}_{\text{converging}}$ such that, for all $t\ge \tau$,
\begin{equation}\label{pseudoeq}
\begin{cases}
x_i(t) = x_{\infty,i}, & \mbox{if } \; i \in \mathcal{V}_{\text{fixed}},\\
x_i(t) < x_i(t+1) < x_{\infty,i} \;\; or \;\; x_i(t) > x_i(t+1) >
x_{\infty,i} , & \mbox{if } \; i \in \mathcal{V}_{\text{converging}}.
\end{cases}
\end{equation}
\begin{conjecture}[Pseudo-stable behavior]\label{pseudoconj}
For any SBC or SBI trajectory, there exists a finite time after which the
trajectory either reaches a fixed state or exhibits a pseudo-stable
behavior.
\end{conjecture}
\begin{conjecture}[Convergence of SBI systems versus SBC systems]\label{inffiniteconj}
For any initial opinion vector and
bounds vector that are generated uniformly randomly, the SBI system is more likely to converge in finite time than the SBC system.
\end{conjecture}
\begin{figure}
\centering
\includegraphics[width=.47\textwidth,keepaspectratio]{compareSBC.pdf}
\includegraphics[width=.47\textwidth,keepaspectratio]{compareSBI.pdf}
\caption{The trajectory of an SBC system (left) and an SBI system (right) are illustrated.
Both systems have the same initial opinion vector and bounds vectors that are randomly generated. However, the SBI
trajectory reaches a fixed state in six time steps, while the SBC trajectory converges in infinite time.
The interconnection topology of the agents in the SBC system
remains constant after $t=64$, and hence its trajectory exhibits a pseudo-stable behavior.}\label{evolutionfig}
\end{figure}
This paper aims to study these conjectures.
\subsection{Agents Classification}\label{class}
In this section, we introduce a classification of agents for both SBC and SBI systems based on their state-dependent interaction topology at each time step. This classification is used later to find the limiting opinion vector and explain the
pseudo-stable behavior.
First, let us quote some relevant definitions from graph theory,
e.g. see \cite{FB-JC-SM:09}. A node of a digraph is \emph{globally
reachable} if it can be reached from any other node by traversing a
directed path. A digraph is \emph{strongly connected} if every node is
globally reachable. A digraph is \emph{weakly connected} if replacing all
of its directed edges with undirected edges produces a connected undirected
graph. A maximal subgraph which is strongly or weakly connected forms a
\emph{strongly connected component} (SCC) or a \emph{weakly connected
component} (WCC), respectively. Every digraph $G$ can be decomposed into
either its SCC's or WCC's. Accordingly, the \emph{condensation digraph} of
$G$, denoted $C (G)$, is defined as follows: the nodes of $C (G)$ are the
SCC's of $G$, and there exists a directed edge in $C (G)$ from node $H_1$
to node $H_2$ if and only if there exists a directed edge in $G$ from a
node of $H_1$ to a node of $H_2$. A node with out-degree zero is named a
\emph{sink}. Knowing that the condensation digraphs are acyclic, each WCC
of $C(G)$ is also acyclic and thus has at least one sink. In a digraph, $i$ is a
\emph{predecessor} of $j$ and $j$ is a \emph{successor} of $i$ if
there exists a directed path from node $i$ to node $j$.
For opinion vector $y \in \mathbb{R}^n$, let $G_r(y)$ denote either of its SBC or SBI digraphs. We classify the SCC's of $G_r(y)$ into three classes.
An SCC of $G_r(y)$ is called a \emph{closed-minded component} if it is a complete subgraph of $G_r(y)$ and corresponds to a sink of $C(G_r(y))$.
An SCC of $G_r(y)$ is called a \emph{moderate-minded component} if it is a non-complete subgraph of $G_r(y)$ and corresponds to a sink of $C(G_r(y))$.
The rest of SCC's of $G_r(y)$ are called \emph{open-minded SCC's}.
Now, the \emph{open-minded subgraph} of $G_r(y)$ is the remaining subgraph after removing $G_r(y)$'s closed- and moderate-minded components and their edges. A WCC of the open-minded subgraph of $G_r(y)$ will be called an \emph{open-minded WCC}, see Figure~\ref{proximityfig}.
\begin{remark}
Previously, (Lorenz, 2006) classified the agents of an SBC system into two classes of \emph{essential} and \emph{inessential}. An agent is
essential if any of its successors is also a predecessor, and an agent is inessential if it has a successor who is not a predecessor \cite{JL:06}.
This classification is similar to the one used for Markov chains \cite[Chapter 1.2]{ES:81}. It is easy to see that the closed- and moderate-minded components are in essential class, and the open-minded components are inessential.
\end{remark}
\begin{figure}
\centering
\includegraphics[width=3.4in,keepaspectratio]{opinion_sample.pdf}
\caption{Consider the opinion vector $x=$ [0.1 0.24 0.27 0.3 0.34 0.37
0.39 0.4 0.5 0.6 0.67 0.68 0.75 0.85 0.86 0.87 1]$^T$ and bounds vector
$r=$[0.5 0.04 0.04 0.04 0.031 0.021 0.011 0.061 0.25 0.01 0.04 0.03 0.3
0.07 0.07 0.07 0.135]$^T$: (a) shows the SBC digraph of $x$, $G_r(x)$,
with its closed- (red), moderate- (green), and open-minded (blue)
components, and each thick gray edge represents multiple edges to all
agents in one component; (b) shows the condensation digraph of
$G_r(x)$; and (c) shows the open-minded subgraph of $G_r(x)$ that is
composed of two open-minded WCC's.}\label{proximityfig}
\end{figure}
\subsection{Spectral Properties of Adjacency Matrix}\label{blocks}
For any opinion vector $y \in \mathbb{R}^n$ in an SBC or SBI system \eqref{system}, the adjacency matrix $A(y)$ is a non-negative row-stochastic matrix, and its nonzero diagonal establishes its aperiodicity.
Since $C(G_r(y))$ is an acyclic digraph, its adjacency matrix is lower-triangular in an appropriate ordering \cite{FB-JC-SM:09}. In such ordering, the adjacency matrix of $G_r(y)$ is lower block triangular.
Based on the classification of the SCC's in $G_r(y)$, we put $A(y)$ into the \emph{canonical form} $\overline{A}(y)$, by an appropriate
\emph{canonical permutation matrix} $P(y)$,
\begin{equation}\label{canonical}
\overline{A}(y) = P(y) A(y) P^T(y) = \left[ \begin{array}{cccccccccccccccccccccccccc} C(y) & 0 & 0\\0 & M(y) & 0\\ \Theta_C(y) & \Theta_M(y)& \Theta(y) \end{array} \right].
\end{equation}
The submatrices $C(y)$, $M(y)$, and $\Theta(y)$ are block diagonal. Each diagonal block $C_i(y)$, with size $n_i(y)$, is the adjacency matrix of the $i$th closed-minded component, and is equal to $C_i(y)= {\bf 1}_{n_i(y)}{\bf 1}^T_{n_i(y)} / n_i(y).$ Let us call a matrix with such structure a \emph{complete consensus matrix}, whose spectrum is found to be $\{ 1, 0, \dots,0 \}.$
Similarly, each diagonal block $M_i(y)$ is the adjacency matrix of the $i$th moderate-minded component.
Each entry in $\Theta_C(y) $ or $ \Theta_M(y)$ represents an edge from an open-minded node to a closed- or moderate-minded node, respectively.
Finally, in the submatrix $\Theta(y)$, each diagonal block $\Theta_i(y)$ corresponds to one open-minded WCC, and is block lower triangular and strictly row-substochastic. By strictly row-substochastic we mean a square matrix with nonnegative entries so that every row adds up to at most one, and there exists at least one row whose sum is strictly less than one.
Note that the adjacency matrix of each SCC in $G_r(y)$ is a diagonal block of $\overline{A}(y)$ and is row-stochastic, nonnegative and primitive. On account of the properties of the open-minded class, the following lemma is proved.
\begin{lemma}\label{rowsum}
For any row $k$ of the submatrix $\Theta(y)$, there exists $p_k \in \mathbb{N}$ such that the $k$th row sum of $\Theta(y)^{p_k}$ is strictly less than 1.
\end{lemma}
\begin{proof}
Every WCC of $C(G_r(y))$ contains at least one sink.
Hence, from any open-minded agent $k$, there exists a directed path of length $p_k$ to an agent $s$ in either a closed- or moderate-minded component. Now, consider the canonical adjacency matrix to the power~$p_k$,
$$\overline{A}(y)^{p_k} = \left[ \begin{array}{cccccccccccccccccccccccccc} C(y)^{p_k} & 0 & 0\\0 & M(y)^{p_k} & 0\\ \Theta_C^{(p_k)}(y) & \Theta_M^{(p_k)}(y) & \Theta(y)^{p_k}\end{array} \right].$$
Existence of such directed path, by \cite[Lemma 1.32]{FB-JC-SM:09}, implies that the $(k,s)$ entry of $\overline{A}(y)^{p_k}$, which belongs to either of the submatrices $\Theta_C^{(p_k)}(y)$ or $\Theta_M^{(p_k)}(y)$, is nonzero. Consequently, the $k$th row sum of $\Theta(y)^{p_k}$ is strictly less than 1.
\end{proof}
It follows from Lemma~\ref{rowsum} that $\lim_{t \rightarrow \infty} \Theta(y)^t = {\bf 0}, $
for a proof of which refer to \cite[Theorem 4.3]{ES:81}.
Therefore, the spectral radius of $\Theta(y)$ is strictly less than one.
\begin{example}\label{exampleblocks}
Consider the SBC system of Figure~\ref{proximityfig} with the permuted opinion vector $x = [ x_2 \;\; x_3 \;\; x_4 \;\; x_5 \;\; x_6 \;\; x_7 \;\; x_8\;\; x_{10} \;\; x_{11} \;\; x_{12} \;\; x_{14} \;\; x_{15} \;\; x_{16} \;\; x_{17} \;\; x_9 \;\; x_1 ]^T$.
Then, the canonical form of the adjacency matrix $A(x)$ contains the following submatrices:
$$C(x) = \left[ \begin{array}{cccccccccccccccccccccccccc}
1 &&&&& \\
& \frac{1}{2} & \frac{1}{2} &&&\\
& \frac{1}{2} & \frac{1}{2} &&&\\
&&& \frac{1}{3} & \frac{1}{3} & \frac{1}{3} \\
&&& \frac{1}{3} & \frac{1}{3} & \frac{1}{3} \\
&&& \frac{1}{3} & \frac{1}{3} & \frac{1}{3} \end{array} \right], \;\;
M(x) = \left[ \begin{array}{cccccccccccccccccccccccccc}
\frac{1}{2} & \frac{1}{2} &&&&&\\
\frac{1}{3} & \frac{1}{3} & \frac{1}{3} &&&&\\
& \frac{1}{2} & \frac{1}{2} && &&\\
&&& \frac{1}{2} & \frac{1}{2} &&\\
&&&& \frac{1}{2} & \frac{1}{2} &\\
&&&&& \frac{1}{2} & \frac{1}{2} \\
&&& \frac{1}{4} & \frac{1}{4} & \frac{1}{4} & \frac{1}{4} \end{array} \right],$$
$$\Theta(x) = \left[ \begin{array}{cccccccccccccccccccccccccc}
\frac{1}{2} && &\\
& \frac{1}{8} & \frac{1}{8}&\\
& \frac{1}{11} & \frac{1}{11} &\\
&& \frac{1}{10} & \frac{1}{10} \end{array} \right], \;\;
\Theta_C(x) = \left[ \begin{array}{cccccccccccccccccccccccccc}
&&&&& \frac{1}{2} \\
\frac{1}{8} & \frac{1}{8} & \frac{1}{8} & \frac{1}{8} & \frac{1}{8} & \frac{1}{8} \\
\frac{1}{11} & \frac{1}{11} & \frac{1}{11} &\frac{1}{11} &&\\
\frac{1}{10} &&&&& \end{array} \right], $$
$$\Theta_M(x) = \left[ \begin{array}{cccccccccccccccccccccccccc}
\tiny 0 &&& \dots &&& 0\\
0 &&& \dots &&& 0 \\
& & & \frac{1}{11} & \frac{1}{11} & \frac{1}{11} & \frac{1}{11} \\
\frac{1}{10} & \frac{1}{10} & \frac{1}{10} & \frac{1}{10} & \frac{1}{10} & \frac{1}{10} & \frac{1}{10} \end{array} \right].$$
\end{example}
\section{Equilibria and Final Value at Constant Topology}\label{final_value}
An opinion vector $y_0$ is an \emph{equilibrium opinion vector} of the dynamical system~\eqref{system}
if and only if $y_0$ is an eigenvector of the adjacency matrix $A(y_0)$ for the eigenvalue one or, equivalently, $y_0 = A(y_0) y_0$. Next, based on Conjecture~\ref{conjtau}, we introduce the following definition.
\begin{definition}[Final value at constant topology]
For any opinion vector $y\in\mathbb{R}^n$ we define its \emph{final value at constant topology} $\operatorname{fvct} : \mathbb{R}^n \rightarrow \mathbb{R}^n$ to be the limiting opinion vector of an SBC or SBI system whose initial opinion vector is $y$ and the interconnection topology of its agents remains unchanged for all $t \ge 0$. That is,
\begin{equation*}
\operatorname{fvct}(y) = \lim_{t\to\infty}A(y)^ty \in\mathbb{R}^n.
\end{equation*}
\end{definition}
The final value at constant topology of any equilibrium opinion vector is equal to itself, that is, $\operatorname{fvct}(y_0)= \lim_{t\to\infty}A(y_0)^t y_0 =y_0 $. Therefore, the set of final values at constant topology is a superset of the equilibria of the system, and the set of equilibria is a superset of the limiting opinion vectors. The condition under which a final value at constant topology is an equilibrium is discussed as follows.
\begin{proposition}[Properties of the final value at constant topology]
\label{prop:properties-final-value}
For any opinion vector $y\in\mathbb{R}^n$ in an SBC or SBI system, whose adjacency matrix can be found from equation~\eqref{canonical}:
\begin{enumerate}
\item $\operatorname{fvct}(y)$ is well defined, and is equal to
\begin{equation*
\operatorname{fvct}(y) = P^T(y)
\left[ \begin{array}{cccccccccccccccccccccccccc} C & 0 & 0\\
0 & M^*& 0 \\
(I - \Theta)^{-1}\Theta_CC & (I - \Theta)^{-1}\Theta_MM^* & 0\end{array} \right](y) P(y) y,
\end{equation*}
where
the submatrix $M^*(y)$ is set equal to $\lim_{t\rightarrow \infty} M(y)^t$ and is well defined.
\item If the two networks of agents with opinion vectors $y$ and $\operatorname{fvct}(y)$ have the same interconnection topology or, equivalently, $G_r(y) =G_r(\operatorname{fvct}(y))$, then \label{nomoderate}
\begin{enumerate}
\item\label{iiparta} $\operatorname{fvct}(y)$ is an equilibrium opinion vector,
\item\label{iipartb} $G_r(y)$ contains no moderate-minded component, and
\item\label{midclosed} in any WCC of $G_r(\operatorname{fvct}(y))$, the maximum and minimum opinions belong to its closed-minded components.
\end{enumerate}
\end{enumerate}
\end{proposition}
\begin{proof}
Let us first drop the $y$ argument for matrices for readability. Regarding part~(i),
\begin{gather*}
\operatorname{fvct}(y) = \lim_{t\to\infty} A^t y = P^T \lim_{t\to\infty}\overline{A}^t P y = P^T \lim_{t\to\infty} \left[ \begin{array}{cccccccccccccccccccccccccc} C^t & 0 & 0\\ 0 & M^t & 0 \\ \Theta_C^{(t)} & \Theta_M^{(t)} & \Theta^t \end{array} \right] P y.
\end{gather*}
From Section~\ref{blocks}, $C^t = C$ for any $t \ge 1$, $\lim_{t\to\infty}\Theta^t = {\bf 0}$, and each diagonal block $M_i$, with size $n_i$, is a row-stochastic primitive nonnegative matrix. For such matrices the Perron-Frobenius Theorem tells us that the spectral radius is equal to one, and the essential spectral radius is strictly less than one. Thus, if we let $\nu_i \in \mathbb{R}^{n_i}$ be a left eigenvector of $M_i$ for the eigenvalue one, then from \cite[Remark 1.69]{FB-JC-SM:09},
\begin{equation}\label{m_infinity}
M_i^* = \lim_{t\to\infty} M_i^t = (\nu_i{\bf 1}_{n_i})^{-1}{\bf 1}_{n_i} \nu_i.
\end{equation}
Using the solution to the infinite products of transition matrices of a Markov chain, given in \cite[Chapter 5]{FG:08}, it can be shown that
$\lim_{t\to\infty} \Theta_C^{(t)} = (I - \Theta)^{-1}\Theta_CC$, and $\lim_{t\to\infty} \Theta_M^{(t)}= (I - \Theta)^{-1}\Theta_MM^*.$
Regarding part~\ref{iiparta}, $G_r(y) =G_r(\operatorname{fvct}(y))$ results in $A(y) =A(\operatorname{fvct}(y))$, and hence
\begin{equation*}
A(\operatorname{fvct}(y)) \operatorname{fvct}(y) = A(y) \lim_{t\to\infty}A(y)^t y =\lim_{t\to\infty}A(y)^t y = \operatorname{fvct}(y).
\end{equation*}
Regarding part~\ref{iipartb}, by contradiction assume that $G_r(y)$ contains at least one moderate-minded component with the opinion vector $y_{M_1}$ and the adjacency matrix $M_1$. The trajectory of each sink in $C(G_r(y))$ is independent of other nodes, hence $\operatorname{fvct}_{M_1}(y) = \lim_{t\to\infty}M_1^t y_{M_1},$ and by equation~\eqref{m_infinity}, $\operatorname{fvct}_{M_1}(y) = (\nu_1{\bf 1}_{n_1})^{-1}{\bf 1}_{n_1} \nu_1 y_{M_1}$. Since $(\nu_1{\bf 1}_{n_1})^{-1}\nu_1 y_{M_1} $ is a scalar, all agents in one moderate-minded component are in consensus in final value at constant topology, and their adjacency matrix is no longer $M_1$, but rather a complete consensus matrix, which contradicts the assumption of $G_r(y) = G_r(\operatorname{fvct}(y))$.
Regarding part~\ref{midclosed}, let $i$ denote the agent with the minimum final value at constant topology in one WCC of $G_r(\operatorname{fvct}(y))$. By contradiction, assume that $i$ is open-minded. Granted that the confidence or influence bounds are strictly greater than zero, in the set of out-neighbors of each open-minded agent there exists at least one agent with distinct opinion. Since $i$ has the smallest opinion among its neighbors, $\operatorname{fvct}_i(y)$ increases after taking an average of $i$'s out-neighbors opinions. In other words, the $i$th entry in the vector $A(\operatorname{fvct}(y)) \operatorname{fvct}(y) $ is strictly larger than $\operatorname{fvct}_i(y)$, which contradicts the fact that $\operatorname{fvct}(y)$ is an equilibrium opinion vector and invariant under matrix $A(\operatorname{fvct}(y))$. Same can be proved for the agent with the maximum opinion.
\end{proof}
\section{Convergence Analysis}\label{sufficientcond}
In this section, motivated by Conjectures~\ref{xinfconj} and \ref{conjtau}, we drive sufficient condition which guarantees that an SBC or SBI trajectory converges to a limiting opinion vector.
Next, to explain Conjecture~\ref{inffiniteconj}, we study sufficient conditions for SBC and SBI systems separately that guarantee reaching a fixed state in finite time.
\subsection{Sufficient Condition for Constant Topology and Convergence}
Our sufficient condition is based on specific neighborhoods of each opinion vector, which is introduced in the following.
\begin{definition}[Equi-topology distances and neighborhoods]
Consider an SBC or SBI system with opinion vector $z\in \mathbb{R}^n$.
\begin{enumerate}
\item The \emph{equi-topology distance} of $z$ is a vector of
non-negative entries $\epsilon(z)\in\mathbb{R}^n_{\geq0}$ defined by, for
$i\in\until{n}$,
\begin{equation}\label{epsiloneq}
\epsilon_i(z) = 0.5 \min \{||z_i - z_j| - R| : j\in\until{n}\setminus\{i\}, R \in \{r_i,r_j\} \},
\end{equation}
and the \emph{equi-topology neighborhood} of $z$ is the set
$\subscr{\mathcal{B}}{et}(z)$ of opinion vectors $y\in\mathbb{R}^n$ such that
\begin{align*}
|y_i - z_i| < \epsilon_i(z), & \text{ for all } i\in\until{n} \text{ with }
\epsilon_i(z) > 0, \text{ and}\\
|y_i - z_i| = \epsilon_i(z), & \text{ for all } i\in\until{n} \text{ with }
\epsilon_i(z) = 0.
\end{align*}
\item The \emph{invariant equi-topology distance} of $z$ is the vector of
non-negative entries $\delta(z)\in\mathbb{R}^n_{\geq0}$ defined by, for
$i\in\until{n}$,
\begin{equation}\label{deltaeq}
\delta_i(z) =
\min \{\epsilon_j(z) : j \mbox{ is a predecessor of } i \text{
in the graph } G_r(z) \},
\end{equation}
and the \emph{invariant equi-topology neighborhood} of $z$ is the set
$\subscr{\mathcal{B}}{iet}(z)$ of opinion vectors $y\in\mathbb{R}^n$ such
that
\begin{align*}
|y_i - z_i| < \delta_i(z), & \text{ for all } i\in\until{n} \text{ with }
\delta_i(z) > 0, \text{ and}\\
|y_i - z_i| = \delta_i(z), & \text{ for all } i\in\until{n} \text{ with }
\delta_i(z) = 0.
\end{align*}
\end{enumerate}
\end{definition}
Note that in any SBC or SBI digraph, each node has a self-loop, and hence each agent is a predecessor of
itself. Therefore, for any opinion vector $z\in\mathbb{R}^n$ and for all $i\in\until{n}$, we have $\delta_i(z) \le \epsilon_i(z)$, which results in $\subscr{\mathcal{B}}{iet}(z) \subset \subscr{\mathcal{B}}{et}(z)$.
\begin{lemma}[Sufficient condition for equal topologies]\label{epsilonlemma}
Consider an SBC or SBI system with opinion vectors $y,z\in \mathbb{R}^n$. If
$y$ belongs to the equi-topology neighborhood of $z$, then the two
networks of agents with opinion vectors $y$ and $z$ have the same
interconnection topology, or equivalently $G_r(y) = G_r(z)$.
\end{lemma}
\begin{remark}
For any $y\in\mathbb{R}^n$, if $y \in \subscr{\mathcal{B}}{et}(\operatorname{fvct}(y))$, then by Lemma~\ref{epsilonlemma} we have
$G_r(y) = G_r(\operatorname{fvct}(y))$.
Hence, by Proposition~\ref{prop:properties-final-value}, $\operatorname{fvct}(y)$
is an equilibrium opinion vector, and $G_r(y)$ contains no moderate-minded
component.
\end{remark}
\begin{proof}
For any $i,j\in \until{n}$ two cases exists:\\
1. $j$ is an out-neighbor of $i$ in $G_r(z)$, hence $|z_i - z_j| \le r$,
where in an SBC system $r=r_i$ and in an SBI system $r=r_j$. In either
system, since $y \in \subscr{\mathcal{B}}{et}(z)$,
\begin{equation*}
|y_i-y_j| \le |z_i - z_j| +\epsilon_i(z)+\epsilon_j(z) \le |z_i - z_j|+ ||z_i - z_j|| -r| = r.
\end{equation*}
2. $j$ is not an out-neighbor of $i$ in $G_r(z)$, hence
$|z_i - z_j|> r$, with $r$ defined above. If both $\epsilon_i(z)$ and $\epsilon_j(z)$ are zero, then $y \in \subscr{\mathcal{B}}{et}(z)$ gives us
$$|y_i-y_j| = |z_i - z_j| > r,$$
and if at least one is nonzero, then
\begin{equation*}
|y_i-y_j| > |z_i - z_j| -\epsilon_i(z)-\epsilon_j(z) \ge |z_i - z_j| - ||z_i - z_j| -r| = r.
\end{equation*}
Therefore, the neighboring relation of agents in $G_r(z)$ is preserved in $G_r(y)$.
One can also prove that any neighboring relation in $G_r(y)$ is preserved in $G_r(z)$.
\end{proof}
\begin{theorem}[Sufficient condition for constant topology and convergence]\label{epsilonthm}
Consider a trajectory $x(t)$ of an SBC or SBI system. Assume that there exists an equilibrium opinion vector $z\in\mathbb{R}^n$ for the system such that $x(0) \in \mathbb{R}^n$ belongs to the invariant equi-topology neighborhood of $z$. Then, for all $t \ge 0$:
\begin{enumerate}
\item\label{epsst} $x(t)$ takes value in the equi-topology neighborhood of
$z$, and hence $G_r(z) = G_r(x(t))$;
\item\label{nomoderatest} $G_r(x(t))$ contains no moderate-minded component; and
\item\label{convergencest} $x(t)$ converges to $\operatorname{fvct}(x(0))$ as time goes to infinity.
\end{enumerate}
\end{theorem}
\begin{remark}[Interpretation of Theorem~\ref{epsilonthm}]
This theorem tells us that if the trajectory of an SBC or SBI system enters a specific ball around any equilibrium opinion vector of that system, then it remains in some larger ball around that vector for all future iterations. Moreover, the proximity digraph of the trajectory and the equilibrium opinion vector remain equal.
\end{remark}
\begin{remark}\label{remarknotgeneral}
Under the condition of Theorem~\ref{epsilonthm}, a trajectory $x(t)$ converges to its final value at constant topology $\operatorname{fvct}(x(t))=\operatorname{fvct}(x(0))$. However, $\operatorname{fvct}(x(t))$ is not necessarily equal to the equilibrium opinion vector $z$, and the proximity digraphs $G_r(\operatorname{fvct}(x(t)))$ and $G_r(x(t))$ can be different, see Figure~\ref{znotxstarfig}.
\end{remark}
\begin{remark}\label{specialcase}
One special case of Theorem~\ref{epsilonthm} is when $x(0) \in \subscr{\mathcal{B}}{iet}(\operatorname{fvct}(x(0))) $, which implies that $\operatorname{fvct}(x(0))$ is an equilibrium opinion vector, again see Figure~\ref{znotxstarfig}.
\end{remark}
\begin{figure}[h!]
\begin{minipage}[b]{.33\linewidth}
\includegraphics[width=.9\textwidth,keepaspectratio]{zxstar.pdf}
\end{minipage}\hfill\begin{minipage}[b]{.67\linewidth}
\caption{The SBC trajectory $x(t)$ with $x(0)=[0\;\; 0.6 \;\; 1]^T$ and confidence bounds $r=[0.25\;\;
1 \;\; 0.25]^T$ is illustrated. It can be computed that $\operatorname{fvct}(x(0)) =
[0\;\; 0.5 \;\; 1]$ and $\delta(\operatorname{fvct}(x(0))) = [0.25 \;\; 0.25 \;\;
0.25]$. Clearly, $x(0) \in \subscr{\mathcal{B}}{iet}(\operatorname{fvct}(x(0)))$,
i.e., the initial vector satisfies the special case of
Theorem~\ref{epsilonthm} stated in Remark~\ref{specialcase}. Hence,
$x(t)$ converges to $\operatorname{fvct}(x(0))$, and their proximity digraphs are
equal. However, if the confidence bounds are equal to $r=[0.5\;\; 1
\;\; 0.25]^T$, then $\delta(\operatorname{fvct}(x(0))) = [0\;\;0\;\;0]$, and $x(t)
\notin \subscr{\mathcal{B}}{iet}(\operatorname{fvct}(x(0)))$ for all $t\ge 0$.
Therefore, $x(t)$ converges to $\operatorname{fvct}(x(0))$, while their proximity
digraphs are different. Both trajectories with the two
confidence bounds vectors are the
same. \label{znotxstarfig}}
\end{minipage}
\end{figure}
\begin{proof}[Proof of Theorem~\ref{epsilonthm}]
Regarding statement~\ref{epsst}, by induction we prove that $x(t) \in \subscr{\mathcal{B}}{et} (z)$ for all $t \ge 0$, which by Lemma~\ref{epsilonlemma} results in $G_r(x(t)) = G_r(z)$.
The first induction step is $x(0) \in \subscr{\mathcal{B}}{et} (z)$, which is true knowing that $x(0) \in \subscr{\mathcal{B}}{iet} (z)$ and $\subscr{\mathcal{B}}{iet} (z) \subset \subscr{\mathcal{B}}{et} (z)$.
To complete the induction argument, assume that the statement~\ref{epsst} holds at times $t =0 ,\dots, \tau$, which implies that $A(z) = A(x(t))$.
The equilibrium opinion vector $z$ satisfies $z=A(z)z$, thus we have $x(t+1) - z = A(x(t))x(t) - A(z)z = A(z)(x(t) - z)$ or, equivalently
\begin{equation}\label{deltadistancekids2}
x_i(t+1) - z_i = \frac{1}{|\mathcal{N}_i(z)|} \sum_{j\in \mathcal{N}_i(z)} (x_j(t) - z_j).
\end{equation}
One can see that
\begin{multline}\label{ineqkids2}
|x_i(\tau + 1) - z_i| \le \max_{j \in \mathcal{N}_i(z)} |x_j(\tau) - z_j| \le \max_{\ell \in \mathcal{N}_j(z), j \in \mathcal{N}_i(z)} |x_\ell(\tau-1) - z_\ell|\\
\le \dots \le \max_{k\in \mathcal{M}} |x_k(0) - z_k|,
\end{multline}
where $\mathcal{M}$ is a subset of successors of $i$ in $G_r(z)$, and thus for any $k\in\mathcal{M}$, equation~\eqref{deltaeq} tells us that $\delta_k(z) \le \epsilon_i(z)$.
Here again two cases exists: First, if for all $k\in\mathcal{M}$, $\delta_k(z) = 0$, then the condition $x(0) \in \subscr{\mathcal{B}}{iet} (z)$ implies that $x_k(0) - z_k = 0$, and it follows from inequality~\eqref{ineqkids2} that $x_i(\tau+1) - z_i =0$. Second, if there exists $\ell \in \mathcal{M}$ such that $\delta_\ell(z) > 0$, then $\epsilon_i(z) > 0$ and
$$|x_i(\tau+1) - z_i| \le \max_{k\in \mathcal{M}} |x_k(0) - z_k| < \max_{k\in \mathcal{M}} \delta_k(z) \le \epsilon_i(z).$$
Therefore, $x(\tau+1) \in \subscr{\mathcal{B}}{et} (z)$.
Regarding statement~\ref{nomoderatest}, according to Section~\ref{final_value}, an equilibrium opinion vector is equal to its own final value at constant topology. Hence $G_r(z) = G_r(z^*(z))$, and by Proposition~\ref{prop:properties-final-value}, $G_r(z)$ and thus $G_r(x(t))$ contain no moderate-minded component.
Regarding statement~\ref{convergencest}, according to the definition of the final value at constant topology, if the topology remains constant for all $t \ge 0$, then $x(t)$ converges to $\operatorname{fvct}(x(0))$.
\end{proof}
Motivated by Conjecture~\ref{xinfconj}, the existence of a limiting opinion vector is required in the following lemma.
\begin{lemma}[Sufficient condition for a limiting opinion vector to be an equilibrium]\label{xinflemma}
Pick a trajectory $x(t)$ of an SBC or SBI system that is convergent. Denote the limiting opinion vector of $x(t)$ by $x_\infty$. If $\min_{i \in \until{n}} \epsilon_i(x_\infty) >0$, where $\epsilon(x_\infty)$ is the equi-topology distance of $x_\infty$ , then there exists time $T$ such that for all $t\ge T$:
\begin{enumerate}
\item $G_r(x_\infty) = G_r(x(t))$, and
\item $x_\infty = \operatorname{fvct}(x(t))$, and is an equilibrium opinion vector.
\end{enumerate}
\end{lemma}
\begin{proof}
According to the definition of convergence, for any $\delta \in \mathbb{R}_{>0}$, there exists $T$ such that for all $t \ge T$, $\|x(t) - x_\infty \|_\infty < \delta$.
Now, if we let $\delta$ be equal to $\min_{i \in \until{n}} \epsilon_i(x_\infty)$, then $\|x(t) - x_\infty \|_\infty < \min_{i \in \until{n}} \epsilon_i(x_\infty)$ for all $t \ge T$, and it follows from Lemma~\ref{epsilonlemma} that $ G_r(x_\infty) = G_r(x(t))$. Under fixed topology, $x(t)$ converges to its final value at constant topology, thus $x_\infty = \operatorname{fvct}(x(t))$. Moreover, the equality $G_r(x(t)) = G_r(\operatorname{fvct}(x(t))) $ tells us that $\operatorname{fvct}(x(t))$, and hence $x_\infty$, is an equilibrium opinion vector.
\end{proof}
\begin{corollary}
An equilibrium opinion vector $z$ is a Lyapunov stable equilibrium vector for the system if $\min_{i \in \until{n}} \epsilon_i(z) > 0$.
\end{corollary}
\subsection{Sufficient Condition for Convergence in Finite Time and Consensus}\label{consensus}
In this subsection, we discuss the sufficient conditions for SBC and SBI
systems to converge to an \emph{agreement opinion vector}. In an agreement
opinion vector, any two agents are either disconnected or in consensus. Reaching global consensus, in which all agents hold the same opinion, is an special case of convergence to an agreement opinion vector. Note that it is possible for an SBC or SBI system to reach an opinion vector that
contains two neighbor agents with separate opinions in finite time. For instance, the trajectory of an SBC system with initial opinion vector $ [0 \;\; 2 \;\; 3 \;\; 4.5\;\; 7 ]^T$ and bounds vector $ [0.01 \;\; 3 \;\; 0.01 \;\; 3 \;\; 0.01 ]^T$ exhibits convergence to a fixed profile in finite time. While, the SBC digraph of the limiting opinion vector contains open-minded agents.
\begin{proposition}[Properties of agreement opinion vectors]
For any agreement opinion vector $\tilde{y} \in \mathbb{R}^n$ in an SBC or SBI system:
\begin{enumerate}
\item\label{epsilongezero} $\min_{i \in \until{n}} \epsilon_i(\tilde{y}) >0$, where $\epsilon(\tilde{y})$ is the equi-topology distance of $\tilde{y}$; and
\item\label{agreementfinite} if $\tilde{y}$ is the limiting opinion vector of a trajectory, then the trajectory reaches $\tilde{y}$ in finite time.
\end{enumerate}
\end{proposition}
\begin{proof}
Regarding statement~\ref{epsilongezero}, by contradiction assume that $\min_{i
\in \until{n}} \epsilon_i(\tilde{y}) =0$. Then based on
equation~\eqref{epsiloneq}, there exist agents $i$ and $j$ such that
$|\tilde{y}_i-\tilde{y} _j| = r_i$. The latter equation tells us that $j
\in \mathcal{N}_i(\tilde{y})$ in an SBC digraph or $i \in
\mathcal{N}_j(\tilde{y})$ in an SBI digraph, while their opinions are
different from each other by $r_i$, which contradicts the definition of
agreement opinion vectors. Regarding statement~\ref{agreementfinite},
consider trajectory $x(t)$ that converges to $\tilde{y}$. Then,
previous statement shows that the limiting opinion vector of $x(t)$ satisfies the condition of
Lemma~\ref{xinflemma}. Therefore, there exists time step $\tau$ such that
the proximity digraphs $G_r(\tilde{y})$ and $G_r(x(t))$ are equal for all
$t \ge \tau$. On the other hand, the proximity digraph of an agreement
opinion vector contains only closed-minded components. Hence, the agents
in each WCC of $G_r(x(\tau))$ reach consensus at the next iteration.
\end{proof}
One sufficient condition that guarantees asymptotic consensus in
``agreement algorithms'', which includes SBC and SBI systems, is given in \cite[Theorem 2.4]{AO-JNT:07} and is as follows. Take a trajectory $x(t)$ of an SBC or SBI
system with proximity digraph $G_r(x(t)) = \big(V, E(x(t))\big)$, where $V$ and
$E(x(t))$ are the sets of nodes and edges of the digraph,
respectively. If there exists $\tau$ such that the graph $\big( V,
E(x(k\tau))\cup E(x(k\tau+1))\cup \dots \cup E(x((k+1)\tau-1))\big)$ is
strongly connected for all $k \in \mathbb{Z}_{\ge 0}$, then all entries of $x(t)$ converge to one real number.
However, this sufficient condition requires knowledge of the system for
infinite time, and is the same for both SBC and SBI systems. Hence, we
derive sufficient conditions that are required to hold in one time step, and
also make it possible to compare SBC and SBI systems in support of
Conjecture~\ref{inffiniteconj}. Let us first define the \emph{opinion
interval} of any subgraph of an SBC or SBI digraph be a closed
interval in $\mathbb{R}$ between that subgraph's minimum and maximum opinions.
\begin{proposition}[Sufficient conditions for convergence to an agreement opinion vector]\label{agreementprop}
Consider the opinion vector $y\in\mathbb{R}^n$ in an SBC or SBI system with the following properties:
\begin{enumerate}
\item\label{separateintervals} the opinion intervals of any two WCC's of the proximity digraph are separated from each other by a distance strictly larger than the maximum confidence or influence bounds of the agents in those WCC's; and
\item\label{sinkforever} it is true that:
\begin{itemize}
\item for any WCC of $y$'s SBC digraph, with $m$ agents, at least $m-1$ agents have confidence bounds larger than that WCC's opinion interval; and
\item for any WCC of $y$'s SBI digraph, at least one agent has influence bound larger than that WCC's opinion interval.
\end{itemize}
\end{enumerate}
Then, the trajectories of both SBC and SBI systems with the initial opinion vector $y$ converge to agreement opinion vectors in finite time. Moreover, in every WCC of either of the SBC or SBI digraphs, at least one node is an out-neighbor of all nodes in that WCC for all $t \ge 0$.
\end{proposition}
\begin{remark}
Any trajectory of an SBC or SBI system that converges to an agreement opinion vector will eventually satisfy the conditions of Proposition~\ref{agreementprop}.
\end{remark}
\begin{proof}[Proof of Proposition~\ref{agreementprop}]
Let us denote either of the SBC or SBI digraphs of $y$ by $G_r(y)$.
In an SBC or SBI system, the smallest and largest opinions in a separate WCC of the proximity digraph are, respectively, non-decreasing and non-increasing in one iteration \cite{VDB-JMH-JNT:09}. This fact tells us that for all $t\ge0$:
first, under the condition~\ref{separateintervals}, the two sets of nodes of two separate WCC's in $G_r(x(0))$ remain separate in $G_r(x(t))$;
second, if the condition~\ref{sinkforever} holds for $G_r(x(0))$, then it also holds for $G_r(x(t))$.
Now, under condition~\ref{sinkforever} for both SBC and SBI systems, any WCC of $G_r(x(t))$ contains at least one agent that is an out-neighbor of all agents in that WCC for all $t \ge 0$.
Denote one such agent in a WCC by $s$, then that WCC's agents with maximum and minimum opinions update their opinions by taking an average of their out-neighbors, including $s$. Hence, at the next iteration, their opinions will converge to $s$'s opinion, which results in an strict decrease in the opinion interval of the WCC. Since the confidence or influence bounds are strictly greater than zero, there exists a time step after which the opinion interval of the WCC is larger than the minimum confidence or influence bound. Consequently, all agents become each others out-neighbors, and the WCC becomes one closed-minded component.
\end{proof}
\section{Numerical Analysis}\label{expsec}
In this section, we provide extensive simulation results that demonstrate the results of Section~\ref{sufficientcond} and are consistent with our conjectures.
We performed 2000 simulations: 100 simulations of both SBC and SBI systems for
ten different agent numbers. In each
simulation, the initial opinion vector and
bounds vector are generated randomly and uniformly distributed on
[0, 1] and [0, 0.3], respectively.
The time steps $\tau$ at which trajectories satisfied the condition of Theorem~\ref{epsilonthm} are plotted
in Figures~\ref{tfConf} and \ref{tfInfl}.
\begin{figure}
\centering
\includegraphics[width=.49\textwidth,keepaspectratio]{finiteConf.pdf} \hfill
\includegraphics[width=.49\textwidth,keepaspectratio]{infiniteConf.pdf}
\caption{In one thousand simulations of SBC systems, the time step $\tau$ at which the trajectory of each system satisfied the sufficient condition of Theorem~\ref{epsilonthm} is plotted versus the number of agents in that system. The left and right plots, respectively, illustrate time $\tau$ for trajectories that converged in finite time and infinite time.
Each initial opinion vector and bounds vector are generated randomly and uniformly distributed on [0, 1] and [0, 0.3], respectively.
For each agent number hundred simulations are performed. All trajectories satisfied the special case of sufficient condition of Theorem~\ref{epsilonthm}, stated in Remark~\ref{specialcase} in finite time.}\label{tfConf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=.49\textwidth,keepaspectratio]{finiteInfl.pdf} \hfill
\includegraphics[width=.49\textwidth,keepaspectratio]{infiniteInfl.pdf}
\caption{In one thousand simulations of SBI systems, the time step $\tau$ at which the trajectory of each system satisfied the sufficient condition of Theorem~\ref{epsilonthm} is plotted versus the number of agents in that system. The left and right plots, respectively, illustrate time $\tau$ for trajectories that converged in finite time and infinite time. As shown, only four SBI trajectories converged in infinite time.
Each initial opinion vector and bounds vector are generated randomly and uniformly distributed on [0, 1] and [0, 0.3], respectively.
For each agent number hundred simulations are performed. All trajectories satisfied the special case of sufficient condition of Theorem~\ref{epsilonthm}, stated in Remark~\ref{specialcase} in finite time.}\label{tfInfl}
\end{figure}
All the 2000 SBC and SBI trajectories eventually satisfied the special case of the
sufficient condition of Theorem~\ref{epsilonthm}, stated in
Remark~\ref{specialcase}. In other words, for each trajectory $x(t)$,
there exists time $\tau$ such that $x(\tau)$ belongs to the invariant
equi-topology neighborhood of its own final value at constant topology
$\operatorname{fvct}(x(\tau))$. Thus, $\operatorname{fvct}(x(\tau))$ is an equilibrium opinion vector, and is
equal to the limiting opinion vector of $x(t)$. The frequency of occurrence
of this special case is intuitively explained by the following statements: First, by
Conjecture~\ref{xinfconj}, for each trajectory a limiting opinion vector
$x_\infty$ exists. Second, for any randomly generated opinion vector $y$
and bounds vector $r$, the probability of having
$\min_{i\in \until{n}} \epsilon_i(y) = 0$, where $\epsilon(y)$ is the
equi-topology distance of $y$, is equal to zero, and one can assume that
the same holds for any limiting opinion vector. Third,
Lemma~\ref{xinflemma} tells us that if the limiting opinion vector
satisfies $\min_{i\in \until{n}} \epsilon_i(x_\infty) > 0$, then the trajectory eventually satisfies the mentioned special case of condition of Theorem~\ref{epsilonthm}.
\begin{figure}
\centering
\includegraphics[width=.57\textwidth,keepaspectratio]{percentboth.pdf}
\caption{For each agent number 100 SBC systems and 100 SBI systems are simulated, as explained in Figures~\ref{tfConf} and \ref{tfInfl}.
The percentage of SBC (blue) and SBI (green) trajectories that reached their limiting opinion vector, which is an agreement opinion vector, in finite time is plotted versus agent number.}\label{percentfig}
\end{figure}
In above mentioned simulations, for each agent number, the percentage of SBC and SBI trajectories that reached a fixed profile in finite time are plotted in Figure~\ref{percentfig}. Clearly, Figure~\ref{percentfig} supports Conjecture~\ref{inffiniteconj}. To explain this frequency of convergence of SBI trajectories in finite time as compared with SBC trajectories, we use the results of Subsection~\ref{consensus}.
For uniformly randomly generated opinion vector and bounds vector, an SBI digraph is more likely to satisfy condition~\ref{separateintervals} of Proposition~\ref{agreementprop} than an SBC digraph.
One can assume that the same holds for a trajectory with uniformly randomly generated initial opinion vector and bounds vector.
In the next section, based on our ``constant topology in finite time'' conjecture, we assume that the interconnection topology in an SBC or SBI system remains constant for infinite time and address the following questions: How the three classes of agents behave? How groups of agents affect each other? And can one explain the observed pseudo-stable behavior of trajectories, as stated in Conjecture~\ref{conjtau}?
\section{The Rate and Direction of Convergence under Fixed Topology}\label{fixedtopbehavior}
In this section, we analyze the rates and directions of convergence of separate classes of agents in the SBC and SBI systems under fixed interconnection topology as time goes to infinity. This analysis proves that the system shows a pseudo-stable behavior under fixed topology.
\begin{definition}[Agent's per-step convergence factor]\label{k}
In an SBC or SBI system with trajectory $x(t) \in\mathbb{R}^n$, we define the \emph{per-step convergence factor} of an agent $i$ whose $x_i(t) - \operatorname{fvct}_i(x(t))$ is nonzero to be
\begin{equation*}
k_i(x(t)) = \frac{x_i(t+1) - \operatorname{fvct}_i(x(t)) }{ x_i(t) - \operatorname{fvct}_i(x(t))}.
\end{equation*}
\end{definition}
The per-step convergence factor of a network of agents was previously introduced in~\cite{LX-SB:04} to measure the overall speed of convergence toward consensus
\begin{remark}\label{pseudoperstep}
Consider a converging trajectory $x(t)$ whose limiting opinion vector is equal to $\operatorname{fvct}(x(t))$ for all $t \ge 0$. Then, $x(t)$ exhibits a pseudo-stable behavior, see equation~\eqref{pseudoeq}, if and only if for all $i \in \until{n}$
\begin{equation*}
\begin{cases}
0 < k_i(x(t)) < 1, \hspace{.6in} & \mbox{if} \hspace{.1in} k_i(x(t)) \mbox{ exists}, \\
x_i(t) = x_i(t+1) = \operatorname{fvct}_i(x(t)), \hspace{.6in} & \mbox{otherwise}.
\end{cases}
\end{equation*}
\end{remark}
\begin{definition}[Leader SCC]\label{leaderSCC}
For opinion vector $y\in\mathbb{R}^n$, let $G_r(y)$ denote either its SBC or SBI digraph.
For any open-minded SCC of $G_r(y)$, $S_k(y)$, denote the set of its open-minded successor SCC's by $\mathcal{M}(S_k(y))$, which includes $S_k(y)$.
We define $S_k(y)$'s \emph{leader SCC} to be the SCC whose adjacency matrix has the largest spectral radius among all SCC's of $\mathcal{M}(S_k(y))$.
\end{definition}
Note that in SBC and SBI digraphs, the adjacency matrix of a large SCC has a large spectral radius, hence that SCC tends to become a leader SCC for its predecessors.
\begin{theorem}[Evolution under constant topology]\label{convlemma}
Consider an SBC or SBI system, denote its trajectory by $x(t)$ and proximity digraph by $G_r(x(t))$.
Assume that there
exists a time $\tau$ after which $G_r(x(t))$ remains
unchanged, that is, $G_r(x(t)) = G_r(x(\tau))$. Then, the following statements hold:
\begin{enumerate}
\item\label{convfirst} $\operatorname{fvct}(x(t)) = \operatorname{fvct}(x(\tau))$ for all $t \ge \tau$.
\item\label{convsecond} $G_r(x(\tau))$ contains no moderate-minded component.
\item\label{convthird} Consider any open-minded SCC $S_k(x(t))$ of $G_r(x(t))$ and its leader SCC $S_m(x(t))$, with adjacency matrices denoted by $\Theta_{k}$ and $\Theta_{m}$, respectively. Then,
\begin{enumerate}
\item\label{convthirda} for any $i \in S_k(x(t))$, either $x_i(t) - \operatorname{fvct}_i(x(t)) = 0$ or its per-step convergence factor converges to the spectral radius of $\Theta_{m}$ as time goes to infinity, and
\item\label{convthirdb} if the spectral radius of $\Theta_{k}$ is strictly less than that of $\Theta_{m}$, then there exists $t_1 \ge \tau$ such that for all $i \in S_k(x(t))$, $j \in S_m(x(t))$, and $t \ge t_1$,
\begin{gather*}
x_j(t_1) < \operatorname{fvct}_j(x(t_1)) \quad\implies\quad x_i(t) \leq \operatorname{fvct}_i(x(t)),
\\
x_j(t_1) > \operatorname{fvct}_j(x(t_1)) \quad\implies\quad x_i(t) \geq \operatorname{fvct}_i(x(t)).
\end{gather*}
\end{enumerate}
\item\label{convforth} There exists time $t_2 \ge \tau$ such that for all $t \ge t_2$, $x(t)$ exhibits a pseudo-stable behavior, see equation~\eqref{pseudoeq}.
\end{enumerate}
\end{theorem}
\begin{remark}[Interpretation of statement~\ref{convthird} in Theorem~\ref{convlemma}]
Parts (a) and (b) tell us, respectively, that the rates and directions of convergence of opinions in an open-minded SCC toward the final value at constant topology are governed by the rate and direction of convergence of its leader SCC.
It is easy to see that the per-step convergence factor has an inverse relation with the rate of convergence to the final value at constant topology. Therefore, Theorem~\ref{convlemma} implies that under fixed interconnection topology, individuals converge to a final decision as slow as the slowest group of agents whom they listen to.
\end{remark}
\begin{proof}
Statement \ref{convfirst} is a direct consequence of $A(x(t)) = A(x(\tau))$ for all $t \ge \tau$.
Statement \ref{convsecond} can be proved similar to part~\ref{iipartb} of Proposition~\ref{prop:properties-final-value}. It was shown that under fixed interconnection topology, all agents in one moderate-minded SCC of an SBC or SBI digraph reach consensus as time goes to infinity. Since, the bounds vector is strictly greater than zero, there exists a time step after which the adjacency matrix of one moderate-minded SCC transforms into a complete consensus matrix, which contradicts the assumption of having fixed topology for infinite time.
Before proving statement~\ref{convthird},
since the canonical permutation matrix remains unchanged, let us assume that the opinions in $x(\tau)$ are ordered such that $A(x(\tau))=\overline{A}(x(\tau))$. Furthermore, owing to the fixed interconnection topology, we drop the $x(t)$ argument for simplicity.
Therefore, by equation~\eqref{canonical}, $$A(x(t)) = \left[ \begin{array}{cccccccccccccccccccccccccc} C & 0 \\ \Theta_C & \Theta \end{array} \right].$$
Now, for all $t > \tau$ we have
\begin{equation*}
x(t) - \operatorname{fvct} (x(\tau))= \left[ \begin{array}{cccccccccccccccccccccccccc} C x_C(\tau) \\ x_\Theta(t) \end{array} \right] - \left[ \begin{array}{cccccccccccccccccccccccccc} C x_C(\tau) \\ \operatorname{fvct}_\Theta(x(\tau)) \end{array} \right]
= \left[ \begin{array}{cccccccccccccccccccccccccc} 0 \\ x_\Theta(t) - \operatorname{fvct}_\Theta(x(\tau)) \end{array} \right],
\end{equation*}
where $x_C(t)$ and $x_\Theta(t)$ are the opinion vectors of agents in closed- and open-minded classes respectively. Using $\operatorname{fvct}(x(\tau)) = A(x(\tau)) \operatorname{fvct}(x(\tau))$, the following recurrence relation holds
\begin{equation}\label{WCCdelta}
x_\Theta(t+1) - \operatorname{fvct}_\Theta(x(\tau)) = \Theta ( x_\Theta(t) - \operatorname{fvct}_\Theta(x(\tau))) \hspace{.2in}\forall~t\ge\tau.
\end{equation}
Consider an open-minded WCC of $G_r(x(t))$, denoted by $W_1$. Let $\Theta_1$ denote $W_1$'s adjacency matrix, and
$x_1(t)$ denote the trajectory of nodes of $W_1$.
Under fixed interconnection topology, the trajectory of each WCC is independent of others, thus for all $t \ge 0$,
$x_1(t+\tau) - \operatorname{fvct}_1(x(\tau)) = \Theta_1^t ( x_1(\tau) - \operatorname{fvct}_1(x(\tau))).$
According to the block lower triangular from of $\Theta_1$,
\begin{equation*}
\Theta_1^t= \left[ \begin{array}{cccccccccccccccccccccccccc} \Theta_{11}^t & & 0\\ \Theta_{21}^{(t)} & \Theta_{22}^t & \\ \vdots & & \ddots \end{array} \right],
\end{equation*}
where each $\Theta_{ii}$ is the adjacency matrix of an SCC, denoted by $S_{ii}$, of $W_1$. Let $x_{ii}(t)$ be the opinion trajectory of nodes of $S_{ii}$. Clearly, $S_{11}$ is one of the sink SCC's in $W_1$, and $\Theta_{ii}$'s are ordered in $\Theta_1$ according to the distance of $S_{ii}$'s to the sinks. For simplicity, we prove statement~\ref{convthird} for $S_{11}$ and an SCC that is the direct predecessor of $S_{11}$.
Without loss of generality, let $S_{22}$ be one such SCC. The proof for the rest of open-minded SCC's is similar.
Each block $\Theta_{ii}$ is nonnegative and primitive. By Perron-Frobenius Theorem: the spectral radius of $\Theta_{ii}$, denoted by $\lambda_i$, is positive and a simple eigenvalue of $\Theta_{ii}$; and there exists a positive eigenvector $\nu_i$ for $\Theta_{ii}$ associated to $\lambda_i$. Any $\Theta_{ii}$ can be written in Jordan normal form by some similarity transformation
\begin{equation*}
\Theta_{ii} = QJQ^{-1} = \left[ \begin{array}{cccccccccccccccccccccccccc} \nu_i \hspace{.1in}\vline & Q_e \end{array} \right] \left[ \begin{array}{cccccccccccccccccccccccccc} \lambda_i & {\bf 0} \\ {\bf 0} & J_e \end{array} \right] \left[ \begin{array}{cccccccccccccccccccccccccc} w_i \\\hline Q_e^{(-1)} \end{array} \right],
\end{equation*}
where $w_i $ is the first row of $Q^{-1}$. Consequently,
\begin{equation}\label{SCCconv}
\lim_{t\rightarrow \infty}\Theta_{ii}^t = \lim_{t\rightarrow \infty} (\lambda_i^t \nu_i w_i+ Q_e J_e^tQ_e^{(-1)}) = \lim_{t\rightarrow \infty} \lambda_i^t \nu_i w_i.
\end{equation}
In any open-minded WCC, the sink SCC is not effected by other SCC's, hence a sink SCC is its own leader. Therefore, for all $t \ge 0$, $$x_{11}(t+\tau) - \operatorname{fvct}_{11}(x(\tau)) = \Theta_{11}^t ( x_{11}(\tau) - \operatorname{fvct}_{11}(x(\tau))).$$
In the interest of simplicity, let us denote the vector $x_{ii}(t)- \operatorname{fvct}_{ii}(x(\tau))$ by $\Delta_{ii}(t)$, then we have
\begin{equation}\label{theta11eq}
\lim_{t \rightarrow \infty} \Delta_{11}(t) = \nu_1 w_1\Delta_{11}(\tau)\lim_{t\rightarrow \infty} \lambda_1^t.
\end{equation}
Regarding part~\ref{convthirda} for $S_{11}$, for the per-step convergence factor of any $i \in S_{11}$ we have
\begin{equation*}
\lim_{t\rightarrow \infty} k_i(x(t)) = \lim_{t\rightarrow \infty}\frac{\lambda_1^{t+1} w_1\Delta_{11}(\tau)\nu_{1i}}{\lambda_1^t w_1\Delta_{11}(\tau)\nu_{1i}} = \lambda_1.
\end{equation*}
Regarding part~\ref{convthirdb} for $S_{11}$, since $\lambda_1^t w_1\Delta_{11}(\tau)$ is a scalar, $\lambda_1$ is positive, and $\nu_1$ is a positive vector, all entries of vector $\nu_1 w_1\Delta_{11}(\tau) \lambda_1^t$ have the same sign. Therefore, there exists time $T \ge \tau$ after which all entries of $\Delta_{11}(t)$ have the same sign for all $t\ge T$.
Here, we prove the two statements for $S_{22}$. It can be computed that for all $t \ge 0$
\begin{gather*}
\Delta_{22}(t+\tau)= \sum_{i=0}^{t-1} \Theta_{22}^i \Theta_{21} \Theta_{11}^{t-i-1} \Delta_{11}(\tau) + \Theta_{22}^t \Delta_{22}(\tau).
\end{gather*}
Now, to find $\lim_{t \rightarrow \infty}\Delta_{22}(t)$, we consider three cases:\\
1) If $\lambda_1 > \lambda_2$, then $S_{11}$ is $S_{22}$'s leader. According to the transient analysis of the reducible Markov chains from \cite[Section 5.6]{FG:08}, and granted that $\lambda_1$ and $\lambda_2$ are strictly less than one, see Section~\ref{blocks},
\begin{multline*}
\lim_{t\rightarrow \infty}\sum_{i=0}^{t-1}\Theta_{22}^i\Theta_{21}\Theta_{11}^{t-i-1}=\big(\lim_{t\rightarrow \infty}\sum_{i=0}^{t-1} \Theta_{22}^i \big)\Theta_{21}\lim_{t\rightarrow \infty} \Theta_{11}^t \\ = (I -\Theta_{22} )^{-1} \Theta_{21}\lim_{t\rightarrow \infty}\lambda_1^t \nu_1 w_1.
\end{multline*}
Therefore,
\begin{equation}\label{theta22eq}
\lim_{t \rightarrow \infty}\Delta_{22}(t) = (I -\Theta_{22} )^{-1} \Theta_{21} \nu_1 w_1 \Delta_{11}(\tau)\lim_{t\rightarrow \infty} \lambda_1^t.
\end{equation}
Regarding part~\ref{convthirda} for $S_{22}$, for any $i\in S_{22}$ we have
\begin{gather*}
\lim_{t \rightarrow \infty} k_i(x(t)) = \lim_{t \rightarrow \infty} \frac{\lambda_1^{t+1} w_1 \Delta_{11}(\tau) [(I -\Theta_{22} )^{-1} \Theta_{21} \nu_1]_i }{\lambda_1^t w_1 \Delta_{11}(\tau)[(I -\Theta_{22} )^{-1} \Theta_{21} \nu_1]_i} = \lambda_1.
\end{gather*}
Regarding part~\ref{convthirdb} for $S_{22}$, since $(I-\Theta_{22} )^{-1} \Theta_{21} \nu_1 $ is a nonnegative vector, all entries of the vector on the right hand side of equations \eqref{theta11eq} and \eqref{theta22eq} have the same sign as the scalar $w_1 \Delta_{11}(\tau)$.\\
2) If $\lambda_1 < \lambda_2$, then $S_{22}$ is its own leader. Similarly,
\begin{equation}\label{theta22selfleadeq}
\lim_{t \rightarrow \infty}\Delta_{22}(t) = w_2\big( \Theta_{21} (I -\Theta_{11} )^{-1} \Delta_{11}(\tau)+\Delta_{22}(\tau)\big)\nu_2\lim_{t \rightarrow \infty}\lambda_2^t ,
\end{equation}
where $ w_2\big( \Theta_{21} (I -\Theta_{11} )^{-1} \Delta_{11}(\tau)+\Delta_{22}(\tau)\big)$ is a scalar, $\lambda_2$ is positive, and $\nu_2$ is a positive vector. Regarding part~\ref{convthirda} for $S_{22}$, for any $i \in S_{22}$ we have
\begin{gather*}
\lim_{t \rightarrow \infty} k_i(x(t)) =\lim_{t \rightarrow \infty} \frac{\lambda_2^{t+1} w_2(\Theta_{21} (I -\Theta_{11} )^{-1} \Delta_{11}(\tau)+ \Delta_{22}(\tau)) \nu_{2i}}{\lambda_2^t w_2 (\Theta_{21} (I -\Theta_{11} )^{-1} \Delta_{11}(\tau)+\Delta_{22}(\tau) ) \nu_{2i}} = \lambda_2.
\end{gather*}
Regarding part~\ref{convthirdb} for $S_{22}$, all entries of the vector on the right hand side of equation~\eqref{theta22selfleadeq} have the same sign.\\
3) if $ \lambda_1 = \lambda_2 = \lambda$, then we have
\begin{gather*}
\lim_{t \rightarrow \infty} \Delta_{22}(t)= \big(\alpha\nu_2 +\beta(I -\Theta_{22} )^{-1} \Theta_{21}\nu_1 \big)\lim_{t\rightarrow \infty} \lambda^t,
\end{gather*}
where $\beta=w_1 \Delta_{11}(\tau)$ and $\alpha = w_2 \Theta_{21} (I -\Theta_{11} )^{-1}\Delta_{11}(\tau) +w_2 \Delta_{22}(\tau)$.
Regarding part~\ref{convthirda} for $S_{22}$, for any $i \in S_{22}$ we have
\begin{equation*}
\lim_{t \rightarrow \infty} k_i(x(t)) =\lim_{t \rightarrow \infty} \frac{\lambda^{t+1} (\alpha \nu_{2i} + \beta [(I -\Theta_{22} )^{-1} \Theta_{21} \nu_1]_i )}{\lambda^t (\alpha \nu_{2i} + \beta [(I -\Theta_{22} )^{-1} \Theta_{21} \nu_1]_i )} = \lambda.
\end{equation*}
Regarding part~\ref{convthirdb} for $S_{22}$, notice that the theorem does not discuss the case with equal spectral radii.
Finally, statement~\ref{convforth} is proved utilizing previous statements. For any $i \in G_r(x(t))$ two cases exists. First, if $i$ belongs to a closed-minded SCC, then $x_i(t) = x_i(\tau +1)$ for all $t > \tau$, and hence $k_i(x(t))$ does not exist. Second, if $i$ belongs to an open-minded SCC $S_k(x(t))$, then according to part~\ref{convthirda}, either $x_i(t) = \operatorname{fvct}_i(t)$ or $k_i(x(t))$ converges to the spectral radius of the adjacency matrix of $S_k(x(t))$'s leader SCC. This spectral radius is proved in Section~\ref{blocks} to be strictly larger than zero and strictly smaller than one. In other words, there exists time $t_2$ such that for all $t \ge t_2$, $ 0 < k_i(x(t)) < 1$. Therefore, according to Remark~\ref{pseudoperstep}, $x(t)$ exhibits pseudo-stable behavior.
\end{proof}
In Figures~\ref{itsownleaderfig} and \ref{differentdirfig}, we provide
numerical examples to facilitate the understanding of the conditions and
results of Theorem~\ref{convlemma}.
\begin{figure}[h!]
\centering
\includegraphics[width=.47\textwidth,keepaspectratio]{evolutionitsownleader.pdf}
\includegraphics[width=.47\textwidth,keepaspectratio]{perstepitsownleader.pdf}
\caption{The trajectory of an SBC system (left) and the per-step
convergence factor of its open-minded agents (right) are
illustrated. The initial opinion vector and confidence bounds vector
are generated randomly. This system satisfies the condition of
Theorem~\ref{epsilonthm} at $t=50$. Moreover, since $x(50) \in
\subscr{\mathcal{B}}{iet}(\operatorname{fvct}(x(50)))$, the SBC digraph $G_r(x(t))$ is
equal to $G_r(\operatorname{fvct}(x(50)))$ for all $t\ge 50$. The digraph
$G_r(\operatorname{fvct}(x(50)))$ contains two open-minded SCC's, denoted by $S_1$ and
$S_2$, while $S_2$ is a predecessor of $S_1$. The spectral radii of the
adjacency matrices of $S_1$ and $S_2$ are equal to 0.6667 and 0.8381,
respectively. Therefore, both $S_1$ and $S_2$ are their own leader
SCC's, and by Theorem~\ref{convlemma} the per-step convergence factors
of their agents converge to 0.6667 and 0.8381, respectively. The right
plot verifies that the per-step convergence factors of all open-minded
agents converge to those two values.}
\label{itsownleaderfig}
\end{figure}
\begin{figure}
\centering
$\begin{array}{ccc}
\includegraphics[width=2in,keepaspectratio]{ev_differentdir.pdf}&
\includegraphics[width=2in,keepaspectratio]{perstep_differentdir.pdf} \\
\includegraphics[width=2in,keepaspectratio]{delta_differentdir.pdf}&
\includegraphics[width=.8in,keepaspectratio]{ex1.pdf}
\end{array}$
\caption{An SBC trajectory $x(t)$ is
plotted on the top left, the open-minded agents per-step convergence
factors on the top right, the open-minded agents distances to their final
values at constant topology $x_i(t) - x_i^*(x(t))$ on the bottom left,
and the open-minded subgraph of $G_r(x(t))$ is illustrated on the bottom
right. This system is simulated with the initial vector $x(0) = [0
\; 1.5 \; 3.5 \; 5 \; 1 \; 1 \; 4 \; 2.1]^T$ and confidence bounds $r =
[0.01 \; 0.01 \; 0.01 \; 0.01 \; 1 \; 1 \; 1 \; 3]^T$.
For all $t \ge 0$, the SBC digraph $G_r(x(t))$ remains unchanged
and contains three open-minded SCC's: $\{ x_5, x_6 \}$, $\{ x_7 \}$, and
$\{ x_8\}$. The spectral radii of the adjacency matrices of these SCC's
are $0.5$, $0.333$, and $0.125$, respectively. The two SCC's $\{ x_5, x_6
\}$ and $\{ x_7 \}$ are successors of $\{ x_8\}$, and based on their
spectral radii, $\{ x_5, x_6\}$ is $\{ x_8\}$'s leader SCC. We can see
that the per-step convergence factor of $x_8$ converges to
$0.5$. Furthermore, the sing of its direction of convergence toward the
final value, i.e, the sign of $x_8(t) - x_8^*$, is the same as the
leader's after $t=1$. These facts support Theorem~\ref{convlemma}. }
\label{differentdirfig}
\end{figure}
\section{Conclusion and Future Work}\label{conclusion}
This paper introduced a synchronized bounded influence (SBI) model of
opinion dynamics, which is similar to the heterogeneous bounded confidence
model introduced by Hegselmann and Krause, which we called synchronized
bounded confidence (SBC) model. First, we conjectured that in both SBC and
SBI systems, for each trajectory there exists a finite time, after which
the topology of the interconnection network remains unchanged, hence, the
trajectory converges to a limiting opinion vector. Second, we conjectured
that if a trajectory does not reach a fixed profile in finite time, then it
eventually shows a pseudo-stable behavior. We partly proved our first
conjecture, and the second conjecture is proved assuming that the first one
is true. We designed a classification of agents that is employed in
computing the equilibria of the system. We introduced the equi-topology
neighborhood and the invariant equi-topology neighborhood of the equilibria
of the system. Based on these neighborhoods, we derived sufficient
condition for both SBC and SBI systems to guarantee that the
interconnection topology remains unchanged for infinite time in a
trajectory, and therefore, the trajectory converges to a steady state. In
our simulation results, it is observed that for uniformly randomly
generated initial opinion vector and bounds vector, the trajectories of
both systems eventually satisfy the mentioned sufficient condition with
probability one. However, the eventual convergence of every trajectory of
the SBC and SBI systems to a steady state is still an open problem. Third,
we conjectured that, for uniformly randomly generated initial opinion
vector and bounds vector, the simulations of SBI systems converge in fewer
time steps and more often in finite time than SBC systems. We derived
sufficient conditions for convergence in finite time for SBC and SBI
systems separately that intuitively explains our third conjecture.
Finally, we studied the trajectory of both SBC and SBI systems when they
update their opinions under fixed interconnection topology for infinite
time. We showed the existence of a leader group for each group of agents
that determines the follower's rate and direction of
convergence.
The main future challenge is to prove that all SBC and SBI systems converge
to steady states. One approach is to prove that in each system, any
trajectory is eventually confined to the invariant equi-topology
neighborhood of an equilibrium opinion vector of the
system.
Moreover, the fact that the SBI systems converge in finite time more often
than the SBC systems might be explained by a probability analysis on the
topology of proximity digraphs.
\bibliographystyle{plain}
|
\section{Introduction}
\label{sec:Introduction}
Electron-positron annihilation at fixed center-of-mass (c.m.)~energies
has long been a mainstay of research in elementary particle physics.
The idea of utilizing initial-state radiation (ISR) to explore \ensuremath{e^+e^-}\xspace
reactions below the nominal c.m.~energies was outlined in
Ref.~\cite{baier},
and discussed in the context of high-luminosity $\rm \phi$ and
$B$ factories in Refs.~\cite{arbus, kuehn, ivanch}.
At high c.m.\ energies, \ensuremath{e^+e^-}\xspace annihilation is dominated by quark-level
processes producing two or more hadronic jets.
Low-multiplicity processes dominate
below or around 2~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, and the region near the charm threshold, 3.0--4.5~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
features a number of resonances~\cite{PDG}.
Thus, studies with ISR events allow us
to probe a wealth of physics topics,
including cross sections, spectroscopy and form factors.
Charmonium and other states with $J^{PC}=1^{--}$ can be observed,
and intermediate states may contribute to the final state hadronic system.
Measurements of their decay modes and branching fractions are important
to an understanding of the nature of such states.
Of particular current interest (see Ref.~\cite{y2175theory}) is
the $Y(2175)$ state
observed to decay to $\phi(1020)f_0(980)$
in our previous study~\cite{isr2k2pi}
and confirmed by the BES~\cite{y2175bes} and Belle~\cite{belle_phif0} Collaborations.
With twice the integrated luminosity
(compared to Ref.~\cite{isr2k2pi})
in the present analysis,
we perform a more detailed study of this structure.
The study of $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\,$hadrons reactions in data is also critical to
hadronic-loop corrections to the muon magnetic anomaly, $a_\mu = (g_\mu-2)/2$.
The theoretical predictions of this anomaly rely on these measurements~\cite{dehz}.
Improving this prediction requires not only more precise measurements,
but also measurements from threshold to the highest c.m.~energy possible.
In addition,
all the important sub-processes should be studied
in order to properly incorporate possible
acceptance effects.
Events produced via ISR at $B$ factories provide independent and
contiguous measurements of
hadronic cross sections from the production threshold to a
c.m.~energy of $\sim$5~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace. With more data we also are able
to reduce systematic uncertainties in the
cross section measurements.
The cross section for the radiation of a photon of energy $E_{\gamma}$
in the c.m. frame,
followed by the production of a particular hadronic final state $f$,
is related to the corresponding direct $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace f$ cross section
$\sigma_f(s)$ by
\begin{equation}
\frac{d\sigma_{\gamma f}(s_0,x)}{dx} = W(s_0,x)\cdot \sigma_f(s_0(1-x))\ ,
\label{eq1}
\end{equation}
where $\sqrt{s_0}$ is the nominal \ensuremath{e^+e^-}\xspace c.m.\@ energy,
$x\! =\! 2E_{\gamma}/\sqrt{s_0}$ is the fraction of the beam energy
carried by the ISR photon,
and $\ensuremath {\rm E_{c.m.}}\xspace \!\equiv\! \sqrt{s_0(1-{\it x})}\!\equiv\! \sqrt{\it s} $ is
the effective c.m.\@ energy at
which the final state $f$ is produced.
The probability density function $W(s_0,x)$ for ISR photon emission has
been calculated with better than 1\% precision (see, e.g.\, Ref.~\cite{ivanch}).
It falls rapidly as $E_{\gamma}$ increases from zero, but has a long
tail, which in combination with
the increasing $\sigma_f(s_0(1-x))$ produces
a sizable event rate at very low \ensuremath {\rm E_{c.m.}}\xspace.
The angular distribution of the ISR photon peaks along the beam
directions. For a typical $e^+ e^-$
detector, around 10-15\% of the ISR photons fall within the
experimental acceptance~\cite{ivanch} .
Experimentally, the measured invariant mass of the hadronic final
state defines \ensuremath {\rm E_{c.m.}}\xspace.
An important feature of ISR data is that a wide range of energies is
scanned continuously in a single experiment,
so that no structure is missed, and
the relative normalization uncertainties in
data from different experiments are avoided.
Furthermore, for large values of $x$ the hadronic system is collimated,
reducing acceptance issues and allowing measurements down to
production threshold.
The mass resolution is not as good as the typical beam energy spread used in
direct measurements,
but resolution and absolute energy scale can be monitored
by means of the measured values of the width and mass of
well-known resonances, such as the $J/\psi$
produced in the reaction $\ensuremath{e^+e^-}\xspace \ensuremath{\rightarrow}\xspace J/\psi\gamma$.
Backgrounds from $\ensuremath{e^+e^-}\xspace \!\ensuremath{\rightarrow}\xspace\,$hadrons events at the nominal $\sqrt{s_0}$
and from other ISR processes can be
suppressed by a combination of particle identification and
kinematic fitting techniques.
Studies of $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\ensuremath{\mu^+\mu^-}\xspace\gamma$ and several multi-hadron ISR processes using
\mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}\ data have been performed~
\cite{isr2k2pi,Druzhinin1,isr3pi,isr4pi,isr6pi,isr5pi,isrkkpi,isr2pi},
demonstrating the viability of such measurements.
These analyses have led to improvements in
background reduction procedures for more rare ISR processes.
The $K^+K^- \ensuremath{\pi^+\pi^-}\xspace$ final state has been measured directly by the
DM1 Collaboration~\cite{2k2pidm1} for $\sqrt{s} <\! 2.2~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$,
and we have previously published ISR measurements of the
$K^+K^- \ensuremath{\pi^+\pi^-}\xspace$ and $K^+K^-K^+K^-$ final states~\cite{isr4pi} for
$\ensuremath {\rm E_{c.m.}}\xspace \!<\! 4.5~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$.
Later we reported an updated measurement of the
$K^+K^- \ensuremath{\pi^+\pi^-}\xspace$ final state with a larger data sample, together with
the first measurement of the $K^+K^- \ensuremath{\pi^0\pi^0}\xspace$ final state, in which we
observed a structure near threshold in the $\phi f_0$ intermediate state~\cite{isr2k2pi}.
In this paper we present a more detailed study of these two final states
along with an updated measurement of the $K^+K^-K^+K^-$ final state.
In all cases we require the detection of the ISR photon and perform a set
of kinematic fits.
We are able to suppress backgrounds sufficiently to study these
final states from their respective production thresholds up to \ensuremath {\rm E_{c.m.}}\xspace=5~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
In addition to measuring the overall cross sections,
we study the internal structure of the final states
and measure cross sections for a number of
intermediate states that contribute to them.
We also study the charmonium region,
measure several $J/\psi$ and $\psi(2S)$ products of branching fraction
and electron width,
and set limits on other states.
\section{\boldmath The \mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}\ detector and dataset}
\label{sec:babar}
The data used in this analysis were collected with the \mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}\ detector at
the PEP-II\ asymmetric-energy \ensuremath{e^+e^-}\xspace\ storage rings at the
SLAC National Accelerator Laboratory.
The total integrated luminosity used is 454.2~\ensuremath{\mbox{\,fb}^{-1}}\xspace,
which includes 413.1~\ensuremath{\mbox{\,fb}^{-1}}\xspace collected at the $\Upsilon(4S)$ peak,
$\sqrt{s_0}=10.58~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$,
and 41.1~\ensuremath{\mbox{\,fb}^{-1}}\xspace collected at about $\sqrt{s_0}=10.54~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$.
The \mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}\ detector is described elsewhere~\cite{babar}.
In the present work, we use charged-particle tracks reconstructed in the tracking system,
which is comprised of a five double-sided-layer silicon vertex tracker (SVT)
and a 40-layer drift chamber (DCH) in a 1.5 T axial magnetic field.
Separation of charged pions, kaons, and protons is achieved using a
combination of Cherenkov angles measured in the detector of internally-reflected
Cherenkov light (DIRC) and specific-ionization measurements in
the SVT and DCH.
For the present study we use a kaon identification algorithm that provides 90--95\%
efficiency, depending on momentum, and pion and proton rejection
factors in the 20--100 range.
Photon and electron energies are measured in a CsI(Tl)
electromagnetic calorimeter (EMC).
We use muon identification provided by an instrumented flux return (IFR)
to select the $\ensuremath{\mu^+\mu^-}\xspace\gamma$ final state used for photon
efficiency studies.
To study the detector acceptance and efficiency,
we use a simulation package developed for radiative processes.
The simulation of hadronic final states, including
$K^+K^- \ensuremath{\pi^+\pi^-}\xspace \gamma$, $K^+K^- \ensuremath{\pi^0\pi^0}\xspace\gamma$ and $K^+K^- K^+K^-\gamma$,
is based on the approach suggested by Czy\.z and K\"uhn \cite{kuehn2}.
Multiple soft-photon emission from the initial-state charged particles is
implemented with a structure-function technique~\cite{kuraev, strfun},
and photon radiation from the final-state particles is simulated by
the PHOTOS package~\cite{PHOTOS}.
The precision of the radiative corrections is about 1\%~\cite{kuraev, strfun}.
We simulate the two $K^+K^-\pi\pi$ ($\ensuremath{\pi^+\pi^-}\xspace, \ensuremath{\pi^0\pi^0}\xspace$) final states
uniformly in phase
space, and also according to models that include the $\phi(1020)\ensuremath{\rightarrow}\xspace K^+K^-$ and/or
$f_{0}(980)\ensuremath{\rightarrow}\xspace \pi\pi$ channels.
The $K^+K^- K^+K^-$ final state is simulated according to phase space, and
also including the $\phi \ensuremath{\rightarrow}\xspace K^+K^-$ channel.
The generated events are subjected to
a detailed detector simulation~\cite{GEANT4},
and we reconstruct them with the same software chain used for the experimental data.
Variations in detector and background conditions
over the course of the experiment are taken into account.
We also generate a large number of potential background processes,
including the ISR reactions
$\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\ensuremath{\pi^+\pi^-}\xspace\pipi\gamma$, $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\ensuremath{\pi^+\pi^-}\xspace\ensuremath{\pi^0\pi^0}\xspace\gamma$, and $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace K_S K\pi\gamma$,
which can contribute due to particle misidentification.
We also simulate
$\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\phi\eta\gamma$, $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\phi\pi^0\gamma$, and $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\ensuremath{\pi^+\pi^-}\xspace\pi^0\gamma$,
which have larger cross sections and can contribute background via missing or
spurious tracks or photons.
In addition, we study non-ISR backgrounds resulting from
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! q \ensuremath{\overline q}\xspace$ $(q = u, d, s, c)$ generated using
JETSET~\cite{jetset} and from
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \tau^+\tau^-$ generated using KORALB~\cite{koralb}.
The cross sections for these processes are known to about 10\% accuracy
or better, which is sufficiently precise for the purposes of the
measurements in this paper.
The contribution from
\Y4S decays is found to be negligible.
\section{\boldmath Event Selection and Kinematic Fit}
\label{sec:Fits}
In the selection of candidate events, we consider
photon candidates in the EMC with energy above 0.03~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
and charged-particle tracks reconstructed in either or both of the DCH
and SVT, that
extrapolate within 0.25 cm of the collision
axis in the transverse plane and within 3 cm of the nominal collision point
along this axis.
We require a photon with c.m.~energy
$E_\gamma > 3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ in each event,
and either four charged-particle tracks with zero net charge
and total momentum roughly (within 0.3 radians) opposite to the photon direction,
or two oppositely-charged tracks that combine with
other photons to roughly balance the high-energy photon momentum.
We assume that the photon with the largest value of $E_\gamma$
is the ISR photon.
We fit the set of charged-particle tracks to a common vertex
and use this as the point
of origin in calculating the photon direction(s).
If additional well-reconstructed tracks exist, the nearest
four (two) to the interaction region are chosen for the four-track (two-track) analysis.
Most events contain additional soft photons due to
machine background or interactions in the detector material.
We subject each candidate event to a set of constrained
kinematic fits and use the fit results,
along with charged-particle identification,
both to select the final states of interest and to measure backgrounds
from other processes.
The kinematic fits use the ISR photon direction and energy along with the
four-momenta and covariance matrices of the initial \ensuremath{e^+e^-}\xspace and the
set of selected tracks and photons.
The ISR photon energy and position are additionally aligned
and calibrated using the $\ensuremath{\mu^+\mu^-}\xspace\gamma$ ISR process, since the two
well-identified muons predict precisely
the position and energy of the photon. This
process is also used to identify and measure
data - Monte Carlo (MC) simulation differences in the photon detection efficiency and
resolution.
The fitted three-momentum for each charged-particle track and the photon are used
in further kinematical calculations.
For the four-track event candidates the fits have four constraints (4C).
We first fit to the $\ensuremath{\pi^+\pi^-}\xspace\pipi$ hypothesis, obtaining
the chi-squared value \ensuremath {\chi^2_{4\pi}}\xspace.
If the four tracks include one identified $K^+$ and one identified $K^-$,
we fit to the $K^+K^-\ensuremath{\pi^+\pi^-}\xspace$ hypothesis and retain the event as a
\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace candidate.
For events with one identified kaon, we perform fits with each of
the two oppositely charged tracks given the kaon hypothesis, and the
combination with the lower \ensuremath {\chi^2_{2K2\pi}}\xspace is retained if its value is less
than $\ensuremath {\chi^2_{4\pi}}\xspace$.
If the event contains three or four identified $K^\pm$,
we fit to the $K^+K^-K^+K^-$ hypothesis and retain the event as a
\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace candidate with chi-squared value \ensuremath {\chi^2_{4K}}\xspace.
For the events with two charged-particle tracks and five or more photon candidates,
we require that both tracks be identified as kaons
to suppress background from ISR $\ensuremath{\pi^+\pi^-}\xspace\ensuremath{\pi^0\pi^0}\xspace$ and $K^\pm\KS\pi^\mp$ events.
We then pair all non-ISR photon candidates and consider combinations with
invariant mass within $\pm$30~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace of the \ensuremath{\pi^0}\xspace mass~\cite{PDG} as \ensuremath{\pi^0}\xspace candidates.
We perform a six-constraint (6C) fit to each set of two non-overlapping
\ensuremath{\pi^0}\xspace candidates, the ISR photon, the two charged-particle tracks, and
the beam particles.
Both \ensuremath{\pi^0}\xspace candidates are constrained to the \ensuremath{\pi^0}\xspace mass,
and we retain the combination with the lowest chi-squared value,
\ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace.
\section{The {\boldmath $K^+ K^- \ensuremath{\pi^+\pi^-}\xspace$} final state}
\subsection{Final Selection and Backgrounds}
\label{sec:selection1}
\begin{figure}[t]
\includegraphics[width=0.9\linewidth]{fig1.eps
\vspace{-0.4cm}
\caption{
Distribution of \ensuremath{\chi^2}\xspace from the four-constraint fit for \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace candidates
in the data (points).
The open histogram is the distribution for simulated signal events,
normalized as described in the text.
The shaded, cross-hatched, and hatched regions
represent, respectively, the background from non-ISR events,
from the ISR $K_S K\pi$ process, and backgrounds with
dominant contribution from mis-identified ISR $4\pi$ events.
Signal and control regions are indicated.
}
\label{2k2pi_chi2_all}
\end{figure}
The \ensuremath {\chi^2_{2K2\pi}}\xspace distribution in data for the
\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace candidates is shown in Fig.~\ref{2k2pi_chi2_all} (points);
the open histogram is the distribution for the simulated \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace events.
The distributions are broader than those for a typical 4C \ensuremath{\chi^2}\xspace
distribution due to higher order ISR, and
the experimental distribution has contributions from background processes.
The simulated distribution is normalized to the data in the region
$\ensuremath {\chi^2_{2K2\pi}}\xspace\!\! <\! 10$ where the contributions of the
backgrounds and radiative corrections do not exceed 10\%.
The shaded histogram in Fig.~\ref{2k2pi_chi2_all} represents
the background from non-ISR $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{q\overline q}\xspace$ events
obtained from the JETSET simulation.
It is dominated by events with a hard $\ensuremath{\pi^0}\xspace$ that results in a fake ISR
photon. These events otherwise have kinematics similar to the signal,
resulting in the peaking structure at low values of \ensuremath {\chi^2_{2K2\pi}}\xspace.
We evaluate this background in a number of \ensuremath {\rm E_{c.m.}}\xspace ranges by
combining the ISR photon candidate with another photon candidate in
both data and simulated events,
and comparing the \ensuremath{\pi^0}\xspace signals in the resulting $\gamma\gamma$
invariant mass distributions.
The simulation gives an \ensuremath {\rm E_{c.m.}}\xspace-dependence consistent with the data,
so we normalize it using an overall factor.
The cross-hatched region in Fig.~\ref{2k2pi_chi2_all} represents
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K_S K\pi\gamma$ events with $K_S\ensuremath{\rightarrow}\xspace\ensuremath{\pi^+\pi^-}\xspace$ decays
close to the interaction region, and one pion mis-identified as a kaon. The process
has similar kinematics to the signal process, and a contribution of about
1\% is estimated using the cross
section measured in our previous study~\cite{isrkkpi}.
The hatched region represents the contribution from
ISR $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{\pi^+\pi^-}\xspace\pipi$ events with one or two misidentified
pions; this process contributes mainly at low \ensuremath{\chi^2}\xspace values.
We estimate the contribution as a function of \ensuremath {\rm E_{c.m.}}\xspace from a
simulation using the cross section value and shape
from our previous study~\cite{isr4pi}.
All remaining background sources are either negligible or give a
\ensuremath {\chi^2_{2K2\pi}}\xspace distribution that is nearly uniform over the range
shown in Fig.~\ref{2k2pi_chi2_all}.
We define the signal region by requiring $\ensuremath {\chi^2_{2K2\pi}}\xspace\!\! <\! 30$,
and estimate the sum of the remaining backgrounds from the difference
between the number of data and simulated entries in the control region,
$30\! <\!\ensuremath {\chi^2_{2K2\pi}}\xspace\!\! <\!60$, as shown in
Fig.~\ref{2k2pi_chi2_all}.
The background contribution to any distribution other than \ensuremath{\chi^2}\xspace is
estimated as the difference between the distributions in the relevant
quantity for data and MC events from the control region of Fig.~\ref{2k2pi_chi2_all},
normalized to the difference between the number of data and MC events
in the signal region. The non-ISR background is subtracted separately.
The signal region contains 85598 data and 63784 simulated events;
the control region contains 9684 data and 4315 simulated events.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig2.eps
\vspace{-0.4cm}
\caption{
The invariant mass distribution for \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace candidates in the data (points):
the shaded, cross-hatched and hatched regions show, respectively,
the non-ISR background from JETSET simulation, the $K_S K\pi$
background with a small
contribution from the control
region of Fig.~\ref{2k2pi_chi2_all}, and the
dominant
contribution resulting from ISR mis-identified $\ensuremath{\pi^+\pi^-}\xspace\pipi$ events.
}
\label{2k2pi_babar}
\end{figure}
Figure~\ref{2k2pi_babar} shows the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace invariant mass distribution
from threshold up to 5.0~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace for events in the signal region.
Narrow peaks are apparent at the $J/\psi$ and $\psi(2S)$ masses.
The shaded histogram represents the \ensuremath{q\overline q}\xspace background,
which is negligible at low mass but dominates at higher masses.
The cross-hatched region represents the background from
the $K_S K\pi$ channel (which exhibits a $\phi(1680)$ peak~\cite{isrkkpi}) and
from the \ensuremath{\chi^2}\xspace control region.
The hatched region represents the contribution from
mis-identified ISR $\ensuremath{\pi^+\pi^-}\xspace\pipi$, and is dominant for masses below 3.0~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
The total background is 6--8\% at low mass, but accounts for 20-25\% of the
observed distribution near 4 \ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace,
and increases further for higher masses.
We subtract the sum of backgrounds in each mass interval to obtain the number
of signal events.
Considering uncertainties in the cross sections for the background processes,
the normalization of events in the control region, and
the simulation statistics,
we estimate a systematic uncertainty on the signal yield that is
2\% or less in the 1.6--3.3~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace mass region, but
increases linearly to 10\% in the 3.3-5.0~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace region, and is about
20\% for the masses below 1.6~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
\subsection{Selection Efficiency}
\label{sec:eff1}
The selection procedure applied to the data is also applied to the
simulated signal samples.
The resulting \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace invariant-mass distributions in the signal and
control regions are shown in Fig.~\ref{mc_acc1}(a) for the uniform phase space
simulation. This model reproduces the observed distributions of kaon and pion
momenta and polar angles.
A broad, smooth mass distribution is chosen to facilitate the estimation
of the efficiency as a function of mass.
We divide the number of reconstructed simulated events in each
mass interval by the number generated in that interval to obtain the
efficiency shown by the points in Fig.~\ref{mc_acc1}(b).
The result of fitting a third-order polynomial to the points
is used for further calculations.
We simulate events with the ISR photon confined to the
angular range 20--160$^\circ$ with respect to the electron beam in the
\ensuremath{e^+e^-}\xspace c.m.\ frame;
this angular range is wider than the actual EMC acceptance.
The calculated efficiency is for this fiducial region,
and includes the acceptance for the final-state hadrons,
the inefficiencies of the detector subsystems,
and event loss due to additional soft-photon emission.
\begin{figure}[t]
\includegraphics[width=0.9\linewidth]{fig3.eps
\vspace{-0.4cm}
\caption{
(a)
The invariant mass distributions for \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace MC events that are
simulated uniformly in phase space, reconstructed
in the signal (open) and control
(hatched) regions of Fig.~\ref{2k2pi_chi2_all};
(b) net reconstruction and selection efficiency as a function of
mass obtained from this simulation
(the curve represents a third-order polynomial fit). The dashed curve is obtained
for the $\phi(1020)\ensuremath{\pi^+\pi^-}\xspace$ final state.
}
\label{mc_acc1}
\end{figure}
The simulations including the $\phi(1020)\ensuremath{\pi^+\pi^-}\xspace$ and/or $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace f_0(980)$
channels give very different mass and angular distributions in the
\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace rest frame.
However, the angular acceptance is quite uniform for ISR events (see
Ref.~\cite{isr4pi}), and the efficiencies are within $1\%$ of those
from the uniform phase space simulation, as shown
by the dashed curve in Fig.~\ref{mc_acc1}(b) for the
$\phi(1020)\ensuremath{\pi^+\pi^-}\xspace$ final state.
To study possible mis-modeling of the acceptance, we repeat the analysis
with tighter requirements.
All charged tracks are required to lie within
the DIRC acceptance, $0.45 \!<\! \theta_{\rm ch} \!<\! 2.4$ radians,
and the ISR photon must not appear near the edges of the EMC,
$0.35 \!<\! \theta_{\rm ISR} \!<\! 2.4$ radians.
The fraction of selected data events satisfying the tighter
requirements differs from the simulated ratio by 1.5\%.
We take the sum in quadrature of this variation and the
1\% model variation (2\% total)
as the systematic uncertainty due to acceptance and model dependence.
\input{2k2pi_table1_22}
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig4.eps
\vspace{-0.4cm}
\caption{
The $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace$ cross section as a function of
\ensuremath{e^+e^-}\xspace c.m.\@ energy measured with ISR data at \mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}\ (dots).
The direct measurements from DM1~\cite{2k2pidm1} are shown as the
open circles.
Only statistical errors are shown.
}
\label{2k2pi_ee_babar}
\end{figure}
Our data sample contains about 3000 events in the $J/\psi$ peak.
Comparing this number
with and without selection on \ensuremath {\chi^2_{2K2\pi}}\xspace we find less than a 1\%
difference between
data and MC simulation due to mis-modeling of the shape of
the \ensuremath {\chi^2_{2K2\pi}}\xspace distribution. This value is taken as an estimate of the
systematic uncertainty associated with
the \ensuremath {\chi^2_{2K2\pi}}\xspace selection criterion.
To measure tracking efficiency, we consider data and simulated events
that contain a high-energy photon and exactly three charged-particle
tracks, which satisfy a set of kinematical criteria,
including a good $\chi^2$ from a kinematic fit to the $\ensuremath{\pi^+\pi^-}\xspace\pipi$ hypothesis,
assuming one missing pion track in the event.
We find that the simulated track-finding efficiency is overestimated by
$(0.75\pm0.25)\%$ per track, so we apply a correction of $+(3\pm1)\%$
to the signal yield.
The kaon identification efficiency is studied in \mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}~ using many different test
processes (e.g.\, $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi(1020)\gamma \!\ensuremath{\rightarrow}\xspace\! \ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\gamma$), and we
conservatively estimate a systematic uncertainty of $\pm 1.0$\% per kaon due to
data-MC differences in our kaon momentum range.
The data-MC simulation correction due to ISR photon detection efficiency
was studied with a sample of $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\ensuremath{\mu^+\mu^-}\xspace\gamma$ events and
was found to be $+(1.0\pm0.5)\%$.
\begin{table}[b]
\caption{
Summary of corrections and systematic uncertainties
for the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace$ cross section measurements.
The total correction is the linear sum of the contributions, and the
total uncertainty is obtained by summing the individual uncertainties in quadrature.
}
\label{error_tab}
\begin{ruledtabular}
\begin{tabular}{l c r@{}l}
Source & Correction & \multicolumn{2}{c}{Uncertainty}\\
\hline
& & & \\[-0.2cm]
Rad. Corrections & -- & 1\% & \\
Backgrounds & -- & 2\% &, $\ensuremath {\rm E_{c.m.}}\xspace <3.3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
& & 2-10 \%&, $\ensuremath {\rm E_{c.m.}}\xspace >3.3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
Model Acceptance & -- & 2\% & \\
\ensuremath {\chi^2_{2K2\pi}}\xspace Distribution & -- & 1\% & \\
Tracking Efficiency & +3\% & 1\% & \\
Kaon ID Efficiency & -- & 2\% & \\
Photon Efficiency & +1.0\% & 0.5\% & \\
ISR Luminosity & -- & 1\% & \\[0.1cm]
\hline
& & & \\[-0.2cm]
Total & +4.0\% & 4\% &, $\ensuremath {\rm E_{c.m.}}\xspace <3.3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
& & 4-11\% &, $\ensuremath {\rm E_{c.m.}}\xspace >3.3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
\end{tabular}
\end{ruledtabular}
\end{table}
\subsection{\boldmath Cross Section for $\ensuremath{e^+e^-}\xspace \!\ensuremath{\rightarrow}\xspace \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace$}
\label{sec:xs2k2pi}
We calculate the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace$ cross section as a
function of the effective c.m.\ energy from
\begin{equation}
\sigma_{\ensuremath {2K2\pi}\xspace}(\ensuremath {\rm E_{c.m.}}\xspace)
= \frac{{\it dN}_{\ensuremath {2K2\pi}\xspace\gamma}(\ensuremath {\rm E_{c.m.}}\xspace)}
{{\it d}{\cal L}(\ensuremath {\rm E_{c.m.}}\xspace) \cdot \epsilon_{\ensuremath {2K2\pi}\xspace}(\ensuremath {\rm E_{c.m.}}\xspace)
\cdot{\it R}(\ensuremath {\rm E_{c.m.}}\xspace)}\ ,
\label{xseqn}
\end{equation}
where
$\ensuremath {\rm E_{c.m.}}\xspace \equiv {\it m}_{\ensuremath {2K2\pi}\xspace}{\it c}^2$ with
$m_{\ensuremath {2K2\pi}\xspace}$ the measured invariant mass of the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace system,
$dN_{\ensuremath {2K2\pi}\xspace\gamma}$ the number of selected events after background
subtraction in the interval $d\ensuremath {\rm E_{c.m.}}\xspace$,
$\epsilon_{\ensuremath {2K2\pi}\xspace}(\ensuremath {\rm E_{c.m.}}\xspace)$ the corrected detection efficiency,
and $R$ a radiative correction.
We calculate the differential luminosity $d{\cal L}(\ensuremath {\rm E_{c.m.}}\xspace)$
in each interval $d\ensuremath {\rm E_{c.m.}}\xspace$, with the photon in the same fiducial range
as that used for the simulation, using the simple leading order (LO) formula
described in Ref.~\cite{isr3pi}. From the mass spectra, obtained from the
MC simulation with and without extra-soft-photon (ISR and FSR) radiation, we
extract $R(\ensuremath {\rm E_{c.m.}}\xspace)$, which gives a correction less than 1\%.
Our data, calculated according to Eq.~\ref{xseqn}, include vacuum
polarization
(VP) and exclude any radiative effects, as is conventional for the reporting
of \ensuremath{e^+e^-}\xspace cross sections. Note that VP should be excluded and FSR included
for calculations of $a_\mu$.
From data-simulation comparisons for the $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\ensuremath{\mu^+\mu^-}\xspace\gamma$ events
we estimate a systematic uncertainty on $d{\cal L}$ of 1\%~\cite{isr2pi}.
We show the cross section as a function of $\ensuremath {\rm E_{c.m.}}\xspace$ in
Fig.~\ref{2k2pi_ee_babar} with statistical errors only
in comparison with the direct measurements from DM1~\cite{2k2pidm1},
and list our results in Table~\ref{2k2pi_tab}.
The results are consistent
with our previous measurements for this
reaction~\cite{isr4pi,isr2k2pi}, but have increased statistical precision.
Our data lie systematically below the DM1 data
for \ensuremath {\rm E_{c.m.}}\xspace above 1.9~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
The systematic uncertainties, summarized in Table~\ref{error_tab},
affect the normalization, but have little effect on the energy dependence.
The cross section rises from threshold to a peak value of about 4.6~nb near
1.86~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, then generally decreases with increasing energy.
In addition to narrow peaks at the $J/\psi$ and $\psi(2S)$ mass values,
there are several possible wider structures in the 1.8--2.8~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace
region.
Such structures might be due to thresholds for
intermediate resonant states, such as $\phi f_0(980)$ near 2~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
Gaussian fits to the distributions of the mass difference between generated
and reconstructed MC data yield \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace mass resolution values
that vary from 4.2~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace in the
1.5--2.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace region to 5.5~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace in the 2.5--3.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace region.
The resolution functions are not purely Gaussian due to soft-photon
radiation,
but less than 10\% of the signal is outside the 0.025~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace mass
interval used in Fig.~\ref{2k2pi_ee_babar}.
Since the cross section has no sharp structure other than the $J/\psi$
and $\psi(2S)$ peaks discussed in Sec.~\ref{sec:charmonium} below,
we apply no correction for mass resolution.
\subsection{\boldmath Substructures in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace Final State}
\label{sec:kaons}
Our previous study~\cite{isr4pi,isr2k2pi} showed evidence for
many intermediate resonances in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace final state.
With the larger data sample used here, these can be seen more clearly
and, in some cases, studied in detail.
Figure~\ref{kkstar}(a) shows a plot of the invariant mass of
the $\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace$ pair versus that of the $\ensuremath{K^+}\xspace\ensuremath{\pi^-}\xspace$ pair.
Signal for the $K^{*}(892)^{0}$ is clearly visible.
Figure~\ref{kkstar}(b) shows the $K^\pm\pi^\mp$ mass distribution
(two entries per event) for all selected \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace events.
As we show in our previous study~\cite{isr2k2pi}, the signal at about
1400~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace has parameters consistent with $K_2^{*}(1430)^{0}$.
Therefore,
we perform a fit to this distribution using P- and D-wave Breit-Wigner (BW)
functions for the $K^{*0}$ and $K_2^{*0}$ signals, respectively,
and a third-order polynomial
function for the remainder of the distribution, taking into account the
$K\pi$ threshold.
The fit result is shown by the curves in Fig.~\ref{kkstar}(b).
The fit yields a $K^{*0}$ signal of $53997\pm526$ events with
$m(K^{*0}) = 0.8932\pm0.0002$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace and $\Gamma(K^{*0}) = 0.0521\pm0.0007$~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
and a $K_2^{*0}$ signal of $4361\pm235$ events with
$m(K_2^{*0}) = 1.4274\pm0.0019$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace and $\Gamma(K_2^{*0}) = 0.0902\pm0.0056$~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
These values are consistent with current world averages for
$K^{*}(892)^0$ and $K_2^{*}(1430)^{0}$~\cite{PDG} ,
and the fit describes the data well, indicating that contributions
from other resonances decaying into $K^\pm\pi^\mp$, like
$K^{*}(1410)^{0}$ and/or $K_0^{*}(1430)^{0}$, are small.
\begin{figure}[tbh]
\includegraphics[width=1.0 \linewidth]{fig5.eps
\vspace{-0.2cm}
\caption{
(a) Invariant mass of the $\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace$ pair versus that of the $\ensuremath{K^+}\xspace\ensuremath{\pi^-}\xspace$ pair;
(b) The $K^\pm\pi^\mp$ mass distribution (two entries per event) for all
selected \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace events:
the solid line represents a fit including two resonances and a
polynomial background function, which is
shown as the hatched region.
}
\label{kkstar}
\end{figure}
We combine $K^{*0}/\kern 0.2em\overline{\kern -0.2em K}{}\xspace^{*0}$ candidates within the lines in
Fig.~\ref{kkstar}(a) with the remaining pion and kaon to obtain the
$K^{*}(892)^{0}\pi^{\pm}$ invariant mass distribution shown in Fig.~\ref{kstark}(b),
and the $K^{*}(892)^{0}\pi^{\pm}$ versus\ $K^{*}(892)^{0} K^{\mp}$ mass plot in
Fig.~\ref{kstark}(a).
The bulk of Fig.~\ref{kstark}(a) shows a strong positive correlation,
characteristic of $K^{*0}K\pi$ final states with no higher resonances.
The horizontal bands in Fig.~\ref{kstark}(a) correspond to the peak
regions of the projection plot of Fig.~\ref{kstark}(b) and are
consistent with the contribution
from the $K_1(1270)$ and $K_1(1400)$ resonances.
There is also an indication of a vertical band in Fig.~\ref{kstark}(a),
perhaps corresponding to a $K^{*}(892)^{0} K$ structure at
$\sim$1.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace. The projection plot of Fig.~\ref{kstark}(c) for
events with $m(K^{*}(892)^{0}\pi^{\pm})>1.5$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace shows the enhancement
not consistent with phase space behavior.
\begin{figure}[tbh]
\includegraphics[width=1.0\linewidth]{fig6.eps
\vspace{-0.2cm}
\caption{
(a)
Invariant mass of the $K^{*}(892)^{0}\pi^{\pm}$ system versus that of
the $K^{*}(892)^{0} K^{\mp}$ system;
(b) the $K^{*}(892)^{0}\pi^{\pm}$ projection plot of (a);
(c) the $K^{*}(892)^{0} K^{\mp}$ projection plot of (a) for $m(K^{*}(892)^{0}\pi^{\pm})>1.5$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
}
\label{kstark}
\end{figure}
We next suppress the $K^{*}(892)^{0}K\pi$ contribution by considering only events
outside the lines in Fig.~\ref{kkstar}(a).
In Fig.~\ref{kpipi}(a) the $K^\pm\ensuremath{\pi^+\pi^-}\xspace$ invariant mass (two
entries per event) shows evidence of the $K_1(1270)$ and $K_1(1400)$ resonances,
both of which decay into $K\rho(770)$,
although the latter decay is very weak~\cite{PDG}.
In Fig.~\ref{kpipi}(b) we plot
the $\ensuremath{\pi^+\pi^-}\xspace$ invariant mass for events with $m(K^\pm\ensuremath{\pi^+\pi^-}\xspace)>1.3$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
There is a strong $\rho(770) \!\ensuremath{\rightarrow}\xspace\! \ensuremath{\pi^+\pi^-}\xspace$ signal, and
there are indications of additional structures in the
$f_0(980)$ and $f_2(1270)$ regions.
\begin{figure}[tbh]
\includegraphics[width=1.0\linewidth]{fig7.eps
\vspace{-0.2cm}
\caption{
(a) The invariant mass of the $K^\pm\ensuremath{\pi^+\pi^-}\xspace$ combinations with
$K^{*}(892)^{0}K\pi$ events excluded;
(b) the $\ensuremath{\pi^+\pi^-}\xspace$ invariant mass for events from (a) with the $K_1(1270)$ region
suppressed by requiring $m(K^{\pm}\ensuremath{\pi^+\pi^-}\xspace)>1.3$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace as shown by vertical line in (a).
}
\label{kpipi}
\end{figure}
The separation of all these, and any other, intermediate states
involving relatively broad resonances requires a partial wave analysis.
This is beyond the scope of this paper.
Instead we present the cross sections for the sum of all states that include
$K^{*}(892)^{0}$, $K_2^{*}(1430)^{0}$ or $\rho(770)$ signals,
and study intermediate states that include a narrow $\phi$ or $f_0$ resonance.
\subsection{\boldmath The $\ensuremath{e^+e^-}\xspace \ensuremath{\rightarrow}\xspace K^{*}(892)^{0} K \pi$,
$K_2^{*}(1430)^{0} K \pi$ and $K^+ K^- \rho(770)$ Cross Sections}
\label{kstarxs}
Signals for the $K^{*}(892)^{0}$ and $K_2^{*}(1430)^{0}$ are clearly visible
in the $K^\pm\pi^\mp$ mass distributions in Fig.~\ref{kkstar}(a,b).
To extract the number of events with correlated production of
$K^{*}(892)^{0}\kern 0.2em\overline{\kern -0.2em K}{}\xspace^{*}(892)^{0}$
and $K^{*}(892)^{0}\kern 0.2em\overline{\kern -0.2em K}{}\xspace_2^{*}(1430)^{0}+c.c.$, we perform the same fit as that
shown in Fig.~\ref{kkstar}(b), but to the $K^+ \pi^-$ invariant mass
distribution in each 0.04~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace interval of $K^- \pi^+$
invariant mass. From each fit we obtain the number of
$K^{*}(892)^{0}$ and $K_2^{*}(1430)^{0}$ events and plot these values
as a function of $K^- \pi^+$ mass in Fig.~\ref{kstar_sel}(a)
and Fig.~\ref{kstar_sel}(b), respectively.
The fit to the data of Fig.~\ref{kstar_sel}(a)
indicates that only $548\pm 263$ events are associated with correlated
$\kern 0.2em\overline{\kern -0.2em K}{}\xspace^{*}(892)^{0} K^{*}(892)^{0}$ production (about 1\% of the total number of
$K^{*}(892)^{0}$ events), and that $1680\pm343$ events correspond to
$\kern 0.2em\overline{\kern -0.2em K}{}\xspace^{*}(892)^{0} K_2^{*}(1430)^{0}$ pairs, compared to $4361\pm235$, the
total number of events with a $K_2^{*}(1430)^{0}$ in the final state.
The distribution of the events from the $K_2^{*}(1430)^{0}$ peak shows a
strong signal at the
$\kern 0.2em\overline{\kern -0.2em K}{}\xspace^{*}(892)^{0}$ mass in Fig.~\ref{kstar_sel}(b), which contains
$1648\pm32$ events, in
agreement with the number of
$K^{*}(892)^{0}\kern 0.2em\overline{\kern -0.2em K}{}\xspace_2^{*}(1430)^{0}$ pairs obtained above.
\begin{figure}[tbh]
\includegraphics[width=1.0\linewidth]{fig8.eps
\vspace{-0.2cm}
\caption{
The $K^- \pi^+$ invariant mass distribution corresponding to the number of
$K^{*}(892)^{0}$ (a) and $K_{2}^{*}(1430)^{0}$ (b) events obtained from the fits to the
$K^+ \pi^-$ invariant mass distribution for each interval of $K^-\pi^+$ mass.
The curves result from the fits described in the text.
}
\label{kstar_sel}
\end{figure}
\begin{figure}[tbh]
\includegraphics[width=1.0\linewidth]{fig9a.eps
\vspace{-0.2cm}
\caption{
The $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace K^{*}(892)^{0}K^- \pi^+ $ cross section,
obtained from the $K^{*}(892)^{0}$ signal of
Fig.~\ref{kkstar}(b).
}
\label{kstar_xs}
\end{figure}
\begin{figure}[tbh]
\includegraphics[width=1.0\linewidth]{fig9b.eps
\vspace{-0.2cm}
\caption{
The $K_2^{*}(1430)^{0}K^- \pi^+ $ cross section,
obtained from the $K_2^{*}(1430)^{0}$ signal of
Fig.~\ref{kkstar}(b).
}
\label{kstar1_xs}
\end{figure}
\begin{figure}[tbh]
\vspace{-0.2cm}
\includegraphics[width=1.0\linewidth]{fig10.eps
\vspace{-0.2cm}
\caption{
(a) The $\pi^+ \pi^-$ mass distribution for all
selected \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace events with the $\phi$ and $K^{*0}$ regions excluded:
the solid curve represents a fit as described in the text, and the background
contribution is shown separately as the hatched region;
(b) the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K^+ K^- \rho(770)$ cross section
obtained from the $\rho$ signal from the fit in
each 0.025~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace c.m.~energy interval.
}
\label{kkrho_sel}
\end{figure}
\input{kstarkpi_table_22}
We perform a fit similar to that shown in Fig.~\ref{kkstar}(b)
to the data in intervals of \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace invariant
mass, with the resonance masses and widths fixed to the values
obtained from the overall fit.
Since correlated $K^{*}$ production is small,
we convert the
resulting $K^{*}$ yield in each interval into a
cross section value for
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K^{*}(892)^{0} K^- \pi^+$ or $K_2^{*}(1430)^{0} K^- \pi^+$,
\footnote{The use of charge conjugate reactions
is implied throughout the paper}
following the procedure described in Sec.~\ref{sec:xs2k2pi}.
These cross section values take into account only the $K\pi$ decay of
the $K^{*}(892)^{0}$ and the $K_2^{*}(1430)^{0}$.
Note that the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K^{*}(892)^{0} K \pi$ ($K_2^{*}(1430)^{0} K \pi$)
cross section includes a small contribution from the
$K_2^{*}(1430)^{0} K \pi$ ($K^{*}(892)^{0} K \pi$) channel, because
the $K_2^{*}(1430)^{0} K^{*}(892)^{0}$ final state
has not been taken into account.
These cross sections are shown in Fig.~\ref{kstar_xs} and
Fig.~\ref{kstar1_xs}, and
the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K^{*}(892)^{0} K^{-} \pi^{+} $ channel is listed in
Table~\ref{kstar_tab} for \ensuremath {\rm E_{c.m.}}\xspace energies from threshold up to 4.0~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
At higher energies the signals are small and contain an unknown, but
possibly large, contribution from $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{q\overline q}\xspace$ events.
There is a rapid rise from threshold to a peak value of about 4~nb at
1.84~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace for the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K^{*}(892)^{0} K^- \pi^+$ cross section,
followed by a very rapid decrease with increasing energy.
There are suggestions of narrow structures in the peak region,
but the only statistically significant structure is the $J/\psi$ peak,
which is discussed below.
There are some structures in the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K_2^{*}(1430)^{0} K^-
\pi^+$ cross section, but the signal size is too small to make any
definite statement.
The $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K^{*}(892)^{0} K^- \pi^+$ contribution is a large fraction of
the total \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace cross
section at all energies above its threshold,
and dominates in the 1.8--2.0~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace region.
The $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\rho^0(770)$ intermediate state makes up the majority of the
remainder of the cross section. We exclude a small $\phi$ contribution by
requiring
$|m(K^+ K^- )-m(\phi)| >0.01$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace, and suppress the large
$K^{*}(892)^{0}$ contribution by means of the anti-selection
$|m([K^\pm \pi^\mp )-0.892| >0.035$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
Figure~\ref{kkrho_sel}(a) shows the $\ensuremath{\pi^+\pi^-}\xspace$ mass distribution for
the remaining events.
The combinatorial background is relatively large, and includes
a small contribution from $f_0(980)\ensuremath{\rightarrow}\xspace\ensuremath{\pi^+\pi^-}\xspace$ decays.
We fit the $\rho(770)$ signal with a single BW (mass and
width are fixed to 0.77~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace and 0.15~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, respectively) and a
polynomial background (contribution shown by the hatched area) in each
0.025~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace c.m.~energy interval. The cross section obtained
is shown in Fig.~\ref{kkrho_sel}(b), and has no significant
structures except the $J/\psi$ signal.
The uncertainty in the $\rho(770)$ shape, and also in the background shape,
provides the largest contribution to the systematic error, estimated
to be 20-30\%. A small contribution to the background from $f_0(980)\ensuremath{\rightarrow}\xspace\ensuremath{\pi^+\pi^-}\xspace$
is ignored in the fit, which does not result in a significant
uncertainty.
\subsection{\boldmath The $\phi(1020)\ensuremath{\pi^+\pi^-}\xspace$ Intermediate State}
\label{sec:phipipi}
\begin{figure}[tbh]
\includegraphics[width=1.0\linewidth]{fig11.eps
\vspace{-0.4cm}
\caption{
(a) $m(\ensuremath{\pi^+\pi^-}\xspace)$ versus $m(\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace)$
for all selected \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace events;
(b) the $\ensuremath{\pi^+\pi^-}\xspace$ invariant mass projections for events in the $\phi$ peak
(open histogram), sidebands (hatched), and background control region
(cross-hatched);
(c) the $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ mass projections for all events (open) and control
region (cross-hatched);
(d) the difference between the open histogram and the sum of the
other contributions to (b).
}
\label{phif0_sel}
\end{figure}
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig12.eps
\vspace{-0.4cm}
\caption{
The $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^+\pi^-}\xspace$ cross section
as a function of \ensuremath{e^+e^-}\xspace c.m.\ energy obtained by \mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}~
(dots) and Belle (squares) ~\cite{belle_phif0}.
}
\label{phipipixs}
\end{figure}
Intermediate states containing narrow resonances can be
studied more easily. For the \ensuremath {\rm E_{c.m.}}\xspace energy range below 3.0~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
Fig.~\ref{phif0_sel}(a) shows a plot of the
invariant mass of the $\ensuremath{\pi^+\pi^-}\xspace$ pair versus that of the $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ pair.
Horizontal and vertical bands corresponding to the $\rho^0(770)$ and
$\phi$, respectively, are visible,
and there is a concentration of entries in the $\phi$ band corresponding
to the correlated production of $\phi$ and $f_{0}(980)$,
as demonstrated by the open histogram of Fig.~\ref{phif0_sel}(b).
The $\phi$ signal is clearly visible in the $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ mass projection of
Fig.~\ref{phif0_sel}(c).
The large contribution from the $\rho(770)$
is nearly uniform in $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ mass,
and the cross-hatched histogram shows the non-\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace background
estimated from the control region in \ensuremath {\chi^2_{2K2\pi}}\xspace.
The cross-hatched histogram
also shows a $\phi$ peak, but this is a small fraction of the events.
When we subtract this background and fit the remaining data with a double-Gaussian
function for the $\phi$ signal, and a first-order polynomial function for the non-$\phi$
background (with a cut-off at the KK threshold), we obtain
3951$\pm$91 events corresponding to the $\phi\ensuremath{\pi^+\pi^-}\xspace$ intermediate state.
To study the $\phi\ensuremath{\pi^+\pi^-}\xspace$ channel, we select candidate events with a
$\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ invariant mass
within 10~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace of the $\phi$ mass,
indicated by the inner vertical lines in Figs.~\ref{phif0_sel}(a,c),
and estimate the non-$\phi$ contribution from the mass sidebands between
the inner and outer vertical lines.
In Fig.~\ref{phif0_sel}(b) we show the \ensuremath{\pi^+\pi^-}\xspace invariant mass distributions
for $\phi$ candidate events, sideband events, and \ensuremath{\chi^2}\xspace control region events
as the open, hatched and cross-hatched histograms, respectively,
and in Fig.~\ref{phif0_sel}(d) we show the $\ensuremath{\pi^+\pi^-}\xspace$ distribution after
subtracting the non-$\phi$ background contributions. We observe a
clear, narrow peak in the $f_0(980)$ mass region, together with a broad
enhancement that reaches a maximum at about 0.6~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace, which could
indicate $f_0(600)$ production. We defer a
detailed analysis of this distribution to Secs.~\ref{sec:phif01},
~\ref{phipipistudy}, and ~\ref{phif0bump}.
\input{phi2pi_table_22}
We obtain the number of $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^+\pi^-}\xspace$ events in
0.025~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace intervals of the
$\phi\ensuremath{\pi^+\pi^-}\xspace$ invariant mass by fitting the $K^+ K^-$ invariant mass
projection in that interval after subtracting non-\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace background.
Each projection is a subset of Fig.~\ref{phif0_sel}(c),
where the curve represents the fit to the full sample.
In each mass interval, all parameters other than number of events in the
$\phi$ peak and the normalization of the background distribution are fixed
to the values obtained from the overall fit. As a check, we also describe
the background as a linear function, with all parameters free in each mass
interval; the alternative fit yields consistent results with the nominal
fit to within 5\%, which is taken as a systematic uncertainty.
The reconstruction efficiency may depend on the details of the production mechanism.
Using the two-pion mass distribution in Fig.~\ref{phif0_sel}(d)
as input,
we simulate the \ensuremath{\pi^+\pi^-}\xspace system as an S-wave composition of two
structures both described by the BW amplitudes,
with parameters set to the values obtained in Sec.~\ref{phipipistudy}.
The BW amplitudes represent the $f_0(980)$ and the bump at 0.6~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace,
which we call $f_0(600)$ (see Sec.~\ref{phipipistudy}).
We describe the $\phi\ensuremath{\pi^+\pi^-}\xspace$ mass distribution using a simple model
with one resonance of mass 1.68~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace and
width 0.3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, which decays to $\phi\ensuremath{\pi^+\pi^-}\xspace$ or $\phi f_0(980)$ when
phase space allows.
The reconstructed spectrum that results then has
a sharp increase at about 2~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace due to the
$\phi f_0(980)$ threshold.
We obtain the efficiency as a function of $\phi\ensuremath{\pi^+\pi^-}\xspace$ mass by
dividing the number of reconstructed events in each interval by the number
generated; the result is
shown in Fig.~\ref{mc_acc1} by the dashed curve.
Comparison with the solid curve
in the same figure shows that the model dependence is weak,
giving confidence in the efficiency calculation.
We calculate the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^+\pi^-}\xspace$ cross section as described
in Sec.~\ref{sec:xs2k2pi}, and divide by the $\phi \!\ensuremath{\rightarrow}\xspace\! K^+ K^-$
branching fraction (0.489~\cite{PDG}).
We show our results as a function of c.m.~energy in
Fig.~\ref{phipipixs}, and list them
in Table~\ref{phi2pi_tab}.
The cross section has a peak value of about 0.6~nb at about 1.7~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
then decreases with increasing energy until the $\phi(1020) f_0(980)$
threshold, around 2.0~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
From this point it rises, falls sharply at about 2.2~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
and then decreases slowly.
Except in the charmonium region, the results
at energies above 3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace are not meaningful due to
small signals and potentially large backgrounds, and are omitted from
Table~\ref{phi2pi_tab}.
Figure~\ref{phipipixs}
displays the cross section up to 4.0~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace in order to show
the $J/\psi$ and $\psi(2S)$ signals, which
are discussed in Sec.~\ref{sec:charmonium}.
The cross section obtained is in agreement with our previous
measurement~\cite{isr2k2pi}.
The cross section measured by
the Belle Collaboration~\cite{belle_phif0}, also shown in Fig.~\ref{phipipixs}, presents
very similar features, and a general consistency with
our data, although a small systematic
difference at higher c.m. energies is visible.
\input{phif0_table_22}
\begin{figure*}[tbh]
\includegraphics[width=0.32\linewidth]{fig13.eps
\includegraphics[width=0.32\linewidth]{fig14.eps
\includegraphics[width=0.32\linewidth]{fig15.eps
\vspace{-0.2cm}
\caption{
Distributions of the cosine of
(a) the $\phi$ production angle,
(b) the pion helicity angle, and
(c) the kaon helicity angle (see text) for $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^+\pi^-}\xspace$
events:
the curves (normalized to the data) represent the distributions
expected if the \ensuremath{\pi^+\pi^-}\xspace system
recoiling against the vector $\phi$ meson is an S-wave system
produced in an S-wave orbital angular momentum state.
}
\label{phi_angle}
\end{figure*}
\begin{figure}[tbh]
\includegraphics[width=1.0\linewidth]{fig16.eps
\vspace{-0.4cm}
\caption{
The $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^+\pi^-}\xspace$ cross section derived from the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace final
state as a function of c.m. energy, for (a) the $0.85<m(\ensuremath{\pi^+\pi^-}\xspace)<1.1$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace
region, dominated by the $\phi(1020) f_0(980)$, and (b) $m(\ensuremath{\pi^+\pi^-}\xspace)<0.85$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
}
\label{phif0xs}
\end{figure}
We perform a study of the angular distributions in the $\phi(1020)\ensuremath{\pi^+\pi^-}\xspace$
final state by considering all \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace candidate events with mass
below 3~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace in intervals
of the cosine of each angle defined below,
and fitting the background-subtracted $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ mass projection in each
interval.
The efficiency is nearly uniform in the cosine of each angle, and
so we study the number of events in each interval.
We define the $\phi$ production angle, $\Theta_\phi$, as the angle
between the $\phi$ direction and the ISR photon direction in the rest
frame of the $\phi\ensuremath{\pi^+\pi^-}\xspace$ system (i.e., the effective $\ensuremath{e^+e^-}\xspace$ collision axis).
The distribution of $\cos\Theta_\phi$, shown in Fig.~\ref{phi_angle}(a),
is consistent with the uniform distribution expected if
the quasi-two-body final state $\phi X$, $X \!\ensuremath{\rightarrow}\xspace\! \ensuremath{\pi^+\pi^-}\xspace$, is produced
in an S-wave angular-momentum state.
We define the pion helicity angle, $\Theta_{\ensuremath{\pi^+}\xspace}$, as that between
the \ensuremath{\pi^+}\xspace and the recoil $\phi$ direction in the \ensuremath{\pi^+\pi^-}\xspace rest frame.
The kaon helicity angle, $\Theta_{\ensuremath{K^+}\xspace}$ is defined as that between the \ensuremath{K^+}\xspace
direction and the ISR photon direction in the $\phi$ rest frame.
The distributions of $\cos\Theta_{\ensuremath{\pi^+}\xspace}$ and $\cos\Theta_{\ensuremath{K^+}\xspace}$,
shown in Figs.~\ref{phi_angle}(b) and~\ref{phi_angle}(c), respectively,
are consistent with those expected from scalar (uniform) and vector
($\cos^2 \Theta_{\ensuremath{K^+}\xspace}$) meson decays,
where for the latter the $\phi$ retains the helicity of the virtual photon
to which the $\phi X$ system couples.
\input{phis_table_22}
\subsection{\boldmath The $\phi(1020) f_{0}(980)$ and
$\phi(1020) f_{0}(600)$ Intermediate States}
\label{sec:phif01}
The narrow $f_0(980)$ peak seen in Fig.~\ref{phif0_sel}(d) allows the
selection of a fairly clean sample of $\phi f_0(980)$ events.
We repeat the analysis just described with the additional requirement
that the \ensuremath{\pi^+\pi^-}\xspace invariant mass be in the range 0.85--1.10~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
A fit to the $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ mass spectrum
for this sample, analogous to that
shown in Fig.~\ref{phif0_sel}(c),
yields about 1350 events;
all of these contain a true $\phi$, with a small fraction of events
with the pion pair not produced through the $f_0(980)$,
but the latter contribution is relatively small (see discussion
in Sec.~\ref{phipipistudy}).
By selecting events with the \ensuremath{\pi^+\pi^-}\xspace invariant mass below
0.85~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace, we similarly obtain a sample composed mostly of $\phi f_0(600)$ events.
We convert the above two samples of $f_0(980)$ and $f_0(600)$ events
in each mass interval
into measurements of the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi(1020) f_{0}(980)$
and $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi(1020) f_{0}(600)$
cross sections as described above, dividing by the
$f_0 \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{\pi^+\pi^-}\xspace$ branching fraction of 2/3
to account for $f_0\ensuremath{\rightarrow}\xspace\ensuremath{\pi^0\pi^0}\xspace$ decays.
The cross sections are shown in Fig.~\ref{phif0xs} as functions of
c.m.~energy
and are listed in Table~\ref{phif0_tab} and Table~\ref{phis_tab}.
The $\phi(1020) f_{0}(980)$ cross section
behavior near threshold does not appear to be smooth,
but is more consistent with a steep rise to a value of about 0.3~nb
at 2.0~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace followed by a slow decrease that is interrupted
by a structure around 2.175~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
In contrast, the $\phi(1020) f_{0}(600)$ cross section has a smooth
threshold increase to about 0.8 nb, followed by a smooth decrease
thereafter, and can be interpreted as the $\phi(1680)$ resonance.
It is
important to note that all structures above 2.0~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace seen in
Fig.~\ref{phipipixs} relate only to the $f_0(980)$ resonance.
Possible interpretations of these structures are discussed
in Sec.~\ref{phif0bump}.
Again, the cross section values are not meaningful for c.m.~energy
above about 3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
except for the $J/\psi$ and $\psi(2S)$ signals,
discussed in Sec.~\ref{sec:charmonium}.
\section{The {\boldmath $K^+ K^-\ensuremath{\pi^0\pi^0}\xspace$} Final State}
\subsection{Final Selection and Backgrounds}
The \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace sample contains background from the ISR processes
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^0}\xspace\gamma$ and $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\eta\gamma$,
in which two soft photon candidates from machine- or detector-related
backgrounds combine with the relatively energetic photons from
the \ensuremath{\pi^0}\xspace or $\eta$ to form two fake \ensuremath{\pi^0}\xspace candidates.
We reduce this background using the angle between each
reconstructed \ensuremath{\pi^0}\xspace direction and the direction of its higher-energy photon
daughter calculated in the \ensuremath{\pi^0}\xspace rest frame.
If the cosines of both angles are larger than 0.85,
we remove the event.
Figure~\ref{2k2pi0_chi2_all} shows the distribution of \ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace for
the remaining candidates together with the
simulated \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace events.
Again, the distributions are broader than those for a typical 6C \ensuremath{\chi^2}\xspace
distribution due to higher order ISR,
and we normalize the histogram to the data in the region
$\ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace\! <\! 15$.
The cross-hatched histogram in Fig.~\ref{2k2pi0_chi2_all} represents
background from $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{q\overline q}\xspace$ events,
evaluated in the same way as for the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace final state.
The hatched region represents the
ISR backgrounds from final states with similar kinematics.
The first of these is
$\ensuremath{\pi^+\pi^-}\xspace\ensuremath{\pi^0\pi^0}\xspace$, which yields events with
both charged pions misidentified
as kaons, and the second is the $K_S K\pi$,
which yields $K_S\ensuremath{\rightarrow}\xspace\ensuremath{\pi^0\pi^0}\xspace$ and a misidentified pion.
Each contribution is small.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig17.eps
\vspace{-0.4cm}
\caption{
Distribution of \ensuremath{\chi^2}\xspace from the six-constraint fits to \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace candidates
in the data (points).
The open histogram is the distribution for simulated signal events,
normalized as described in the text.
The cross-hatched, hatched, and dashed regions represent, respectively,
the backgrounds from non-ISR \ensuremath{q\overline q}\xspace events, ISR-produced $\ensuremath{\pi^+\pi^-}\xspace\ensuremath{\pi^0\pi^0}\xspace$ and
$K_S K\pi$ events,
and ISR-produced $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^0}\xspace$, $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\eta$ and $\ensuremath{\Kp \kern -0.16em \Km}\xspace\ensuremath{\pi^0\pi^0}\xspace\ensuremath{\pi^0}\xspace$ events.
}
\label{2k2pi0_chi2_all}
\end{figure}
The dominant background in this case is from residual ISR $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^0}\xspace$
and $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\eta$ events,
as well as ISR-produced $\ensuremath{\Kp \kern -0.16em \Km}\xspace\ensuremath{\pi^0\pi^0}\xspace\ensuremath{\pi^0}\xspace$ events.
Their net simulated contribution,
indicated by the dashed contour in Fig.~\ref{2k2pi0_chi2_all},
is consistent with the data in the high \ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace region.
All other backgrounds are either negligible or distributed uniformly
in \ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace.
We define the signal region by $\ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace\! <\! 50$, which
contains 7967 data and 7402 simulated events,
and a control region by $50\! <\! \ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace\! <\! 100$, which
contains 2007 data and 704 simulated signal events.
Figure~\ref{2k2pi0_babar} shows the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace invariant mass
distribution from threshold up to 5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace for events in the signal region.
The \ensuremath{q\overline q}\xspace background (cross-hatched histogram) is negligible at low
masses but yields a significant fraction of the selected events above about 4~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
The ISR $\ensuremath{\pi^+\pi^-}\xspace\ensuremath{\pi^0\pi^0}\xspace$ contribution (hatched region) is negligible except
in the 1.5--2.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace region.
The sum of all other backgrounds, estimated from the control region,
is the dominant contribution below 2.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace and is non-negligible everywhere.
The total background varies
from ~100\% below 1.6~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace to ~25\% at higher masses.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig18.eps
\vspace{-0.4cm}
\caption{
Invariant mass distribution for \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace candidates in the signal region for
data (points).
The cross-hatched, hatched, and open regions represent, respectively,
the non-ISR \ensuremath{q\overline q}\xspace background, the contribution from ISR-produced
$\ensuremath{\pi^+\pi^-}\xspace\ensuremath{\pi^0\pi^0}\xspace$ and $K_SK\pi$ events,
and the contribution from the other ISR processes described in the text.
}
\label{2k2pi0_babar}
\end{figure}
We subtract the sum of the estimated background contributions
from the number of selected events
in each mass interval to obtain the number of signal events.
Considering uncertainties in the cross sections for the background processes,
the normalization of events in the control region and the simulation
statistics, we estimate a systematic uncertainty on the signal yield
after background subtraction of
about 5\% in the 1.6--3.0~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace region;
this increases linearly from 5\% to 15\% in the region above 3~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
\subsection{Selection Efficiency}
The detection efficiency is determined in the same manner as in
Sec.~\ref{sec:eff1}.
Figure~\ref{mc_acc3}(a) shows the simulated \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace invariant mass
distributions in the signal and control regions obtained from the phase space
model.
We divide the number of reconstructed events in each
0.04~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace mass interval by the number generated in that interval to
obtain the efficiency estimate shown by the points in Fig.~\ref{mc_acc3}(b);
a third-order-polynomial fit to the efficiency
is used in calculating the cross section.
Again, the simulation of the ISR photon
covers a limited angular range, which is
about 30\% wider than the EMC acceptance.
Simulations assuming dominance of the $\phi \!\ensuremath{\rightarrow}\xspace\! \ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ and/or
the $f_0 \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{\pi^0\pi^0}\xspace$ channels give results consistent with those of
Fig.~\ref{mc_acc3}(b), and we apply
a 3\% systematic uncertainty for possible model dependence,
as in Sec.~\ref{sec:eff1}.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig19.eps
\vspace{-0.4cm}
\caption{
(a) Invariant mass distribution for simulated \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace events in the
signal (open) and control (hatched) regions (see Fig.~\ref{2k2pi0_chi2_all});
(b) net reconstruction and selection efficiency as a function of mass
obtained from this simulation
(the curve represents the result of a third-order-polynomial fit).
}
\label{mc_acc3}
\end{figure}
We correct for mis-modeling of the track-finding and kaon identification
efficiencies as in Sec.~\ref{sec:eff1} (corrections of $(+1.9\pm0.6)$\% and
$(0\pm2.0)$\%, respectively).
We do not observe any large discrepancy in the shape of the \ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace
distribution, and so apply no correction for the $\ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace<50$ selection,
but introduce 3\% as an associated systematic uncertainty.
We correct the \ensuremath{\pi^0}\xspace-finding efficiency using the procedure described in
detail in Ref.~\cite{isr6pi}.
From ISR $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \omega\ensuremath{\pi^0}\xspace\gamma \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{\pi^+\pi^-}\xspace\ensuremath{\pi^0\pi^0}\xspace\gamma$ events
selected with and without the \ensuremath{\pi^0}\xspace from the $\omega$ decay,
we find that the simulated efficiency for one \ensuremath{\pi^0}\xspace is too large by (3.0$\pm$1.0)\%,
and we apply a correction of $(+6.0\pm 2.0)$\%
because of the two $\ensuremath{\pi^0}\xspace$s in each event.
\input{2k2pi0_table_22}
\begin{table}[tbh]
\caption{
Summary of corrections and systematic uncertainties
for the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace$ cross section measurements.
The total correction is the linear sum of the contributions, and the
total uncertainty is obtained by summing the individual contributions
in quadrature.
}
\label{error2_tab}
\begin{ruledtabular}
\begin{tabular}{l c r@{}l}
Source & Correction & \multicolumn{2}{c}{Uncertainty} \\
\hline
& & & \\[-0.2cm]
Rad. Corrections & -- & 1\% & \\
Backgrounds & -- & 5\% &, $\ensuremath {\rm E_{c.m.}}\xspace <3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
& & 5-15\%&, $\ensuremath {\rm E_{c.m.}}\xspace >3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
Model Dependence & -- & 3\% & \\
\ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace Distribution & -- & 3\% & \\
Tracking Efficiency & +1.9\% & 0.6\%& \\
Kaon ID Efficiency & -- & 2\% & \\
\ensuremath{\pi^0}\xspace Efficiency & +6\% & 2\% & \\
ISR-photon Efficiency & +1.0\% & 0.5\%& \\
ISR Luminosity & -- & 1\% & \\[0.1cm]
\hline
& & & \\[-0.2cm]
Total & +8.9\% & 7\% &, $\ensuremath {\rm E_{c.m.}}\xspace <3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
& & 7-16\%&, $\ensuremath {\rm E_{c.m.}}\xspace >3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
\end{tabular}
\end{ruledtabular}
\end{table}
\subsection{\boldmath Cross Section for $\ensuremath{e^+e^-}\xspace \ensuremath{\rightarrow}\xspace \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace$}
\label{sec:2k2pi0xs}
We calculate the cross section for $\ensuremath{e^+e^-}\xspace \ensuremath{\rightarrow}\xspace \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace$ in 0.04~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace \ensuremath {\rm E_{c.m.}}\xspace
intervals from the analog of Eq.(\ref{xseqn}),
using the invariant mass of the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace system to determine the
c.m.\ energy.
We show the results
in Fig.~\ref{2k2pi0_ee_babar} and list the values and statistical errors
in Table~\ref{2k2pi0_tab}.
The cross section rises to a peak value near 0.8~nb at 2~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace then
shows a rapid decrease, which is
interrupted by a large $J/\psi$ signal; the charmonium region is
discussed in Sec.~\ref{sec:charmonium} below.
The drop at 2.2~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace is similar to that seen for the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace final state.
Again, the differential luminosity includes corrections for vacuum
polarization that should be omitted for calculations of $a_\mu$.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig20.eps
\vspace{-0.5cm}
\caption{
The $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace$ cross section as a function of
\ensuremath{e^+e^-}\xspace c.m.\ energy measured with ISR data at \mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}.
The errors are statistical only.
}
\label{2k2pi0_ee_babar}
\end{figure}
The simulated \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace invariant mass resolution is 8.8~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace in the
1.5--2.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace mass range,
and increases with mass to 11.2~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace in the 2.5--3.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace range.
Since less than 20\% of the events in a 0.04~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace interval are
reconstructed outside that interval,
and the cross section has no sharp structure other than the $J/\psi$ peak,
we again make no correction for resolution.
The point-to-point systematic uncertainties are much smaller than the statistical
uncertainties, and
the errors on the normalization are summarized in Table~\ref{error2_tab},
along with the corrections that were applied to the measurements.
The total correction is $+8.9$\%, and the total systematic uncertainty
is 7\% at low mass, increasing linearly from 7\% to 16\% above 3~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
\subsection{\boldmath Substructure in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace Final State}
\label{sec:kaons2}
\begin{figure}[tbh]
\includegraphics[width=1.0\linewidth]{fig21.eps
\vspace{-0.4cm}
\caption{
(a) Invariant mass of the $\ensuremath{K^-}\xspace\ensuremath{\pi^0}\xspace$ pair versus that of the
$\ensuremath{K^+}\xspace\ensuremath{\pi^0}\xspace$ pair in selected \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace events (two entries per event);
(b) sum of the projections of (a) (dots, four entries per event).
The curves represent the result of the fit described in the text.
}
\label{kkstarpi0}
\end{figure}
\begin{figure}[tbh]
\includegraphics[width=1.0\linewidth]{fig21a.eps
\vspace{-0.4cm}
\caption{
The number of
$K^{*}(892)^{-}$ (a) and $K_{2}^{*}(1430)^{-}$ (b) events obtained from the fits to the
$K^- \pi^0$ invariant mass distributions for each 0.04~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace interval of $K^+\pi^0$ mass.
The curves result from the fits described in the text.
}
\label{k2stars0}
\end{figure}
A plot of the invariant mass
of the $K^-\pi^0$ pair versus that of the $K^+\pi^0$ pair is shown in
Fig.~\ref{kkstarpi0}(a)
(two entries per event) for the \ensuremath{\chi^2}\xspace signal region after removing the
$\phi(1020)$ contribution by $|m(K^+ K^-)-m(\phi)|>0.01$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
Horizontal and vertical bands corresponding to the $K^{*}(892)^{-}$ and
$K^{*}(892)^{+}$, respectively, are visible.
Figure~\ref{kkstarpi0}(b) shows as points the sum of the two projections of
Fig.~\ref{kkstarpi0}(a);
a large $K^{*}(892)^{\pm}$ signal is evident.
Fitting this distribution with the function used in
Sec.~\ref{kstarxs}, we obtain the number of events corresponding to
$K^{*}(892)^{\pm}$ ($7734\pm320$) and $K^{*}(1430)^{\pm}$ ($793\pm137$) production.
The $K^{*}(1430)^{\pm}$:$K^{*}(892)^{\pm}$ ratio is consistent with that
obtained for neutral $K^*$ production in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace channel,
but the number of $K^{*}(892)^{\pm}$ combinations in the peak
is larger than the total number of \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace events (5522).
This indicates the presence of
some number of correlated $K^{*}(892)^{+} K^{*}(892)^{-}$ pairs.
Fitting the $K^-\pi^0$ mass distribution in each 0.04~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace bin of
$K^+\pi^0$ invariant mass,
we obtain the number of $K^{*}(892)^{-}$ and $K^{*}(1430)^{-}$ events
shown in Fig.~\ref{k2stars0}(a,b). The correlated production of
$K^{*}(892)^{+} K^{*}(892)^{-}$ and $K^{*}(892)^{+}
K_2^{*}(1430)^{-}$ is clearly seen, and the fits yield $1750\pm60$ and
$140\pm49$ events, respectively.
Note that $K^{*}(892)^{+} K^{*}(892)^{-}$ accounts for about 30\% of all
\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace events, in contrast with the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace channel, where only
$548\pm263$ events (less than 1\% of the total) are found to result from the
$\kern 0.2em\overline{\kern -0.2em K}{}\xspace^{*}(892)^{0} K^{*}(892)^{0}$ pair production.
We find no evidence for resonance production
in the $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^0}\xspace$ or $K^\pm\ensuremath{\pi^0\pi^0}\xspace$ subsystems.
Since the
statistics are low in any given mass interval, we do not attempt to extract a
separate $K^{*}(892)^{+} K^{-}\pi^0 + c.c.$ cross section.
The total \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace cross section is roughly a factor of four lower than
the $K^{*}(892)^{0} K^{-}\pi^+$ cross section observed in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace
final state.
This is consistent with what is expected from isospin considerations and the
charged versus neutral $K^*$ branching fractions involving charged kaons.
\input{phif0_table2_22}
\subsection{\boldmath The $\phi(1020)\ensuremath{\pi^0\pi^0}\xspace$ Intermediate State}
\label{sec:phipipi2}
The selection of events containing $\phi(1020) \!\ensuremath{\rightarrow}\xspace\! \ensuremath{\Kp \kern -0.16em \Km}\xspace$ decays
follows that in Sec.~\ref{sec:phipipi}.
Figure~\ref{phif0_sel2}(a) shows the
plot of the invariant mass of the $\ensuremath{\pi^0\pi^0}\xspace$ pair versus that of the
\ensuremath{\Kp \kern -0.16em \Km}\xspace pair.
The $\phi$ resonance is visible as a vertical band,
whose intensity decreases with increasing \ensuremath{\pi^0\pi^0}\xspace mass except for an enhancement
in the $f_{0}(980)$ region (Fig.~\ref{phif0_sel2}(b)).
The $\phi$ signal is also visible in the $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ invariant mass
projection for events in the control region,
shown in Fig.~\ref{phif0_sel2}(c).
The relative non-$\phi$ background is smaller than in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace
mode,
but there is a large background from ISR $\phi\pi^0$, $\phi\eta$
and/or $\phi\ensuremath{\pi^0\pi^0}\xspace\ensuremath{\pi^0}\xspace$ events,
as indicated by the control region histogram (hatched) in
Fig.~\ref{phif0_sel2}(c).
The contributions from non-ISR and ISR $\ensuremath{\pi^+\pi^-}\xspace\ensuremath{\pi^0\pi^0}\xspace$ events are negligible.
Selecting $\phi$ candidate and side band events as for the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace mode
(vertical lines in Figs.~\ref{phif0_sel2}(a,c)),
we obtain the \ensuremath{\pi^0\pi^0}\xspace mass projections shown as the open and cross-hatched
histograms, respectively, in Fig.~\ref{phif0_sel2}(b).
Control region events (hatched histogram) are concentrated at low mass values
in Fig.~\ref{phif0_sel2}(b), and
a peak corresponding to the $f_{0}(980)$ is visible over a relatively
low background.
\begin{figure}[tbh]
\includegraphics[width=1.0\linewidth]{fig22.eps
\vspace{-0.4cm}
\caption{
(a) Plot of the $\ensuremath{\pi^0\pi^0}\xspace$ invariant mass versus the $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ invariant
mass for all selected \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace events;
(b) the $\ensuremath{\pi^0\pi^0}\xspace$ invariant mass projections for events in the $\phi$ peak
(open histogram), sidebands (cross-hatched), and control region
(hatched);
(c) the $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ mass projection for events in the signal (open) and
control (hatched) regions;
(d) the difference between the open histogram and sum of the
other contributions to (b).
}
\label{phif0_sel2}
\end{figure}
In Fig.~\ref{phif0_sel2}(d) we show the $\ensuremath{\pi^0\pi^0}\xspace$ mass distribution
associated with $\phi$ production after subtraction of all background
contributions. The distribution is consistent in shape with that of
Fig.~\ref{phif0_sel}(d), but with a data sample which is about six times
smaller.
We obtain the number of $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^0\pi^0}\xspace$ events in 0.04~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace
intervals of
$\phi\ensuremath{\pi^0\pi^0}\xspace$ invariant mass by fitting the $K^+ K^-$ invariant mass
projection in that interval to the $\phi$ signal,
after subtracting the non-\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace background, the same way as described in
Sec.~\ref{sec:phipipi}. The obtained cross section is shown in
Fig.~\ref{phi2pi0xs} and is very similar to that obtained from the
\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace final state shown in Fig.~\ref{phipipixs}. The errors shown
reflect not only that there are six times fewer events, but also
a much larger background level.
As before, we defer discussion to Secs.~\ref{phipipistudy} and ~\ref{phif0bump}.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig22a.eps
\vspace{-0.4cm}
\caption{
Cross section for the reaction $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi(1020)
\ensuremath{\pi^0\pi^0}\xspace$
as a function of \ensuremath{e^+e^-}\xspace c.m.\ energy obtained from the
\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace final state.
}
\label{phi2pi0xs}
\end{figure}
\subsection{\boldmath The $\phi(1020) f_0(980)$ Intermediate State}
\label{sec:phif02}
Since the background under the $f_0(980)$ peak in Figs.~\ref{phif0_sel2}(b,d)
is 25\% or less, we are able to extract the $\phi(1020) f_0(980)$
contribution.
As in Sec.~\ref{sec:phif01}, we require the dipion mass to be in the
range 0.85--1.10~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace and fit the background-subtracted $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$
mass projection in each 0.04~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace interval of \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace mass to obtain the number of
$\phi f_0$ events.
Again, some $\phi\ensuremath{\pi^0\pi^0}\xspace$ events are present in which the \ensuremath{\pi^0\pi^0}\xspace
pair is not produced through the $f_0$.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig23.eps
\vspace{-0.4cm}
\caption{
Cross section for the reaction $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi(1020)
f_{0}(980)$, $f_0\ensuremath{\rightarrow}\xspace\pi\pi$
as a function of \ensuremath{e^+e^-}\xspace c.m.\ energy obtained from the
\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace final state.
}
\label{phif0xs2}
\end{figure}
We convert the number of $f_0(980)$ events in each mass interval into a
measurement of the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi(1020) f_{0}(980)$ cross
section as described previously, and divide by the
$f_{0}(980) \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{\pi^0\pi^0}\xspace$ branching fraction of 1/3
to obtain the $f_{0}(980)\ensuremath{\rightarrow}\xspace\pi\pi$ value.
The cross section, corrected for the $\phi(1020)\ensuremath{\rightarrow}\xspace\ensuremath{\Kp \kern -0.16em \Km}\xspace$ decay rate,
is shown in Fig.~\ref{phif0xs2} as a function of
\ensuremath {\rm E_{c.m.}}\xspace and is listed in Table~\ref{phif0_tab2}.
Due to the smaller number of events, we have used larger
intervals at higher energies.
The overall shape is consistent with that obtained from the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace final state
(see Fig.~\ref{phif0xs}),
and there seems to be a sharp drop near 2.2~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace;
however, the statistical errors are large and no conclusion can be drawn
from this mode alone.
Possible interpretations are discussed in Section~\ref{phif0bump}.
\section{\boldmath The \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace Final State}
\subsection{Final Selection and Background}
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig24.eps
\vspace{-0.4cm}
\caption{
Distribution of \ensuremath{\chi^2}\xspace from the three-constraint fit for \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace
candidates in the data (points).
The open histogram is the distribution for simulated signal events,
normalized as described in the text.
The shaded histogram represents the background from
non-ISR events, estimated
as described in the text. The region defined by the dashed contour
is for simulated ISR $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+\pi^-}\xspace$ events with at least one pion
misidentified as a kaon.
}
\label{4k_chi2_all}
\end{figure}
Figure~\ref{4k_chi2_all} shows the distribution of \ensuremath {\chi^2_{4K}}\xspace for the
\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace candidates as points. The open histogram is the distribution for
simulated \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace events,
normalized to the data in the region $\ensuremath {\chi^2_{4K}}\xspace \! <\! 5$ where
the relative contributions of the backgrounds and radiative corrections are small.
The shaded histogram represents the background from non-ISR
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{q\overline q}\xspace$ events, evaluated as for the other modes.
The region defined by the dashed contour represents the background
from simulated ISR \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace events with at least one charged pion
misidentified as a kaon.
We define signal and control regions by $\ensuremath {\chi^2_{4K}}\xspace\! <\! 20$ and
$20\! <\! \ensuremath {\chi^2_{4K}}\xspace\! <\! 40$, respectively.
The signal region contains 4190 data and 14904 simulated events,
and the control region contains 877 data and 1437 simulated events.
Figure~\ref{4k_babar} shows the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace invariant mass distribution from
threshold up to 4.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace for events in the signal region as points
with errors.
The \ensuremath{q\overline q}\xspace background (shaded histogram) is small at all masses.
Since the ISR \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace background does not peak at low \ensuremath {\chi^2_{4K}}\xspace values,
we include it in the background evaluated from the control region,
according to the method explained in Sec.~\ref{sec:selection1}.
It dominates this background, which is about 20\% for 2.3-2.6~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace and
10\% or lower at all other mass values.
The total background is shown as the hatched histogram in Fig.~\ref{4k_babar}.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig25.eps
\vspace{-0.4cm}
\caption{
Invariant mass distribution for \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace candidates in the data (points).
The shaded histogram represents the non-ISR background, and
the hatched region is for the ISR background from the control region,
which is dominated by the contribution from misidentified ISR \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace events.
}
\label{4k_babar}
\end{figure}
We subtract the sum of backgrounds from the number of selected events
in each mass interval to obtain the number of signal events.
Considering the uncertainties in the cross sections for the background processes,
the normalization of events in the control region,
and the simulation statistics,
we estimate that the systematic uncertainty on the signal yield is less than
5\% in the 2--3~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace region, but it
increases to about 10\% above 3~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig26.eps
\vspace{-0.4cm}
\caption{
(a) Invariant mass distributions for simulated \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace events in the
signal (open) and control (hatched) regions (see Fig.~\ref{4k_chi2_all});
(b) net reconstruction and selection efficiency as a function of mass
obtained from this simulation;
the curves represent third-order polynomial fits for
the phase space model (solid) and the $\phi K^+ K^-$ model (dashed).
}
\label{mc_acc4}
\end{figure}
\subsection{Selection Efficiency}
The detection efficiency is determined as for the other two final states.
Figure~\ref{mc_acc4}(a) shows the simulated \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace invariant-mass
distributions in the signal and control regions from the phase space
model.
We divide the number of reconstructed events in each
mass interval by the number generated in that interval
to obtain the efficiency shown by the points in Fig.~\ref{mc_acc4}(b).
It is quite uniform, and we fit the measurements using a third-order
polynomial, which we then use to obtain the cross section.
As discussed previously, this efficiency includes
the difference between the EMC acceptance and
the region of ISR photon simulation.
A simulation assuming dominance of the $\phi\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ channel,
with the \ensuremath{\Kp \kern -0.16em \Km}\xspace pair in an angular-momentum S-wave state, gives
consistent results, as shown by the
dashed curve in Fig.~\ref{mc_acc4}(b),
and we estimate a 5\% systematic uncertainty associated with the difference.
We correct only for mis-modeling of the track-finding and
ISR-photon-detection efficiency
as in Sec.~\ref{sec:eff1}.
\subsection{\boldmath Cross Section for $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace K^+ K^- K^+ K^-$}
\label{sec:4kxs}
We calculate the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace$ cross section in
0.025~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace intervals of \ensuremath {\rm E_{c.m.}}\xspace from the analog of Eq.(\ref{xseqn}),
using the invariant mass of the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace system to determine the
c.m.\ energy.
We show the cross section in Fig.~\ref{4k_ee_babar}, and list the
measured values in
Table~\ref{4k_tab}. The cross section increases from threshold
to a peak value of about 0.1~nb near 2.7~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
then decreases slowly with increasing energy.
The only statistically significant narrow structures are the large $J/\psi$ peak and
a possible narrow structure near ~2.3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, which will be discussed in
Sec.~\ref{sec:phif03}.
Again, the differential luminosity contribution in each \ensuremath {\rm E_{c.m.}}\xspace interval
includes corrections for vacuum polarization that
should be omitted for the calculations of $a_\mu$.
This measurement supersedes our previous result~\cite{isr4pi}.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig27.eps
\vspace{-0.5cm}
\caption{
The $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace$ cross section as a function of \ensuremath{e^+e^-}\xspace
c.m.\ energy measured with ISR data at \mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}.
The errors are statistical only.
}
\label{4k_ee_babar}
\end{figure}
\input{4k_table_22}
The simulated \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace invariant mass resolution is 3.0~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace in the
2.0--2.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace range, increasing with mass to 4.7~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace in the
2.5--3.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace range, and to about 6.5~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace at higher masses.
Since the cross section has no sharp structure except for the $J/\psi$
peak, we again make no correction for resolution.
The errors shown in Fig.~\ref{4k_ee_babar} and listed in Table~\ref{4k_tab} are
statistical only.
The point-to-point systematic uncertainties are much smaller, and
the errors on the normalization are summarized in Table~\ref{error3_tab},
along with the corrections applied to the measurements.
The total correction is +4.0\%,
and the total systematic uncertainty is 9\% at low mass, linearly
increasing to 13\% above 3~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
%
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig28.eps
\vspace{-0.6cm}
\caption{
Invariant mass distribution for all \ensuremath{\Kp \kern -0.16em \Km}\xspace pairs in selected
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace$ events (open histogram), and for the
combination in each event closest to the $\phi$-meson mass (hatched).
}
\label{mkk_phi}
\end{figure}
\begin{table}[tbh]
\caption{
Summary of corrections and systematic uncertainties
for the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace$ cross section measurements.
The total correction is the linear sum of the individual corrections, and the
total uncertainty is the sum in quadrature of the separate uncertainties.
}
\label{error3_tab}
\begin{ruledtabular}
\begin{tabular}{l c r@{}l}
Source & Correction & \multicolumn{2}{c}{Uncertainty} \\
\hline
& & & \\[-0.2cm]
Rad. Corrections & -- & 1\% & \\
Backgrounds & -- & 5\% &, $\ensuremath {\rm E_{c.m.}}\xspace <3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
& & 5-10\%&, $\ensuremath {\rm E_{c.m.}}\xspace >3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
Model Dependence & -- & 5\% & \\
\ensuremath {\chi^2_{4K}}\xspace Distribution & -- & 3\% & \\
Tracking Efficiency & +3.0\% & 2\% & \\
Kaon ID Efficiency & -- & 4\% & \\
ISR-photon Efficiency & +1.0\% & 0.5\% & \\
ISR Luminosity & -- & 3\% & \\[0.1cm]
\hline
& & & \\[-0.2cm]
Total & +4.0\% & 9\% &, $\ensuremath {\rm E_{c.m.}}\xspace <3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
& & 9-13\%&, $\ensuremath {\rm E_{c.m.}}\xspace >3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ \\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{figure*}[t]
\includegraphics[width=0.25\linewidth]{fig29.eps
\includegraphics[width=0.25\linewidth]{fig30.eps
\includegraphics[width=0.25\linewidth]{fig31.eps
\includegraphics[width=0.25\linewidth]{fig32.eps
\vspace{-0.4cm}
\caption{ (a)
The invariant mass distribution for \ensuremath{\Kp \kern -0.16em \Km}\xspace pairs in events in which
the other \ensuremath{\Kp \kern -0.16em \Km}\xspace pair has mass closest to, and within 10~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace of,
the nominal $\phi$ mass (open histogram); events within $\pm 50\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace$
of the $J/\psi$ mass have been excluded.
The hatched histogram corresponds to events with \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace invariant mass in the
$J/\psi$ peak. The numbered regions of the combined histograms from (a) are
used to calculate the cross sections shown in Figs. (b), (c) and (d)
for regions 1, 2 and 3 respectively.
}
\label{mkk_notphi}
\end{figure*}
\subsection{\boldmath The $\phi(1020) \ensuremath{\Kp \kern -0.16em \Km}\xspace$ Intermediate State}
\label{sec:phif03}
Figure~\ref{mkk_phi} shows the invariant mass distribution for all
\ensuremath{\Kp \kern -0.16em \Km}\xspace pairs in the selected \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace events (4 entries per event) as the
open histogram.
A prominent $\phi$ peak is visible along with a possible excess near ~1.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
The hatched histogram is for the pair in each event with mass closest
to the nominal $\phi$ mass,
and indicates that the $\phi \ensuremath{\Kp \kern -0.16em \Km}\xspace$ channel dominates the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace final state;
we do not see any other significant contribution.
If the invariant mass of the $\ensuremath{\Kp \kern -0.16em \Km}\xspace$ pair that
is closest to the $\phi$ mass is within $\pm$10~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace of the $\phi$
peak, then we include the invariant mass of the other $\ensuremath{\Kp \kern -0.16em \Km}\xspace$ combination
in Fig.~\ref{mkk_notphi}(a).
Events with \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace mass within $\pm 50\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace$ of
the $J/\psi$ mass are excluded.
Events within $\pm 50$~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace of the $J/\psi$ mass
are shown as the
hatched histogram. The latter is in agreement with results from the BES
experiment~\cite{bes4k}, for which
the structures around 1.5, 1.7, and 2.0~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace were studied in detail.
For the dots with error bars
there is an enhancement at threshold that can be
interpreted as being due to $f_0(980)\ensuremath{\rightarrow}\xspace\ensuremath{\Kp \kern -0.16em \Km}\xspace$ decay.
This is expected in light of the $\phi f_0$ cross sections measured
above in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace and \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace final states, but a contribution from
the $a_0(980)\ensuremath{\rightarrow}\xspace\ensuremath{\Kp \kern -0.16em \Km}\xspace$ can not be excluded.
For the combined histograms of Fig.~\ref{mkk_notphi}(a), we
select events with $m(\ensuremath{\Kp \kern -0.16em \Km}\xspace)<1.06$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace (shown as region 1) and
calculate a cross
section enriched in the $\ensuremath{e^+e^-}\xspace \ensuremath{\rightarrow}\xspace \phi f_0(980)$ reaction
(Fig.~\ref{mkk_notphi}(b)).
A bump at $\ensuremath {\rm E_{c.m.}}\xspace = 2.175\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ is seen;
however, the small number of events and uncertainties in the $f_0(980) \!\ensuremath{\rightarrow}\xspace\! \ensuremath{\Kp \kern -0.16em \Km}\xspace$
line-shape do not allow a meaningful extraction of the cross section
for this $f_0(980)$ decay mode.
A clear signal corresponding to the $f'_2(1525)$ is seen in both histograms shown in
Fig.~\ref{mkk_notphi}(a). The $f'_2(1525)$ region is defined by
$1.45<m(\ensuremath{\Kp \kern -0.16em \Km}\xspace)<1.6$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace, and is indicated as region 3 in Fig.~\ref{mkk_notphi}(a).
The corresponding cross section is shown
in Fig.~\ref{mkk_notphi}(d) and exhibits a broad (about
0.10-0.15~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace) structure at 2.7~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace and a strong $J/\psi$ signal.
In Fig.~\ref{mkk_notphi}(a)(open histogram) there is an indication
of structure for \ensuremath{\Kp \kern -0.16em \Km}\xspace invariant mass in the 1.3-1.4~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace region; this
may correspond to production of the $\phi f_0(1370)$ final state.
Finally, we tried to find a region of \ensuremath{\Kp \kern -0.16em \Km}\xspace invariant mass
corresponding to the spike seen at about 2.3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace in the total $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace$
cross section shown in Fig.~\ref{4k_ee_babar}. This spike is much more
significant if we require $1.06<m(\ensuremath{\Kp \kern -0.16em \Km}\xspace)<1.2$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace,
shown as region 2 in Fig.~\ref{mkk_notphi}(a),
with corresponding cross section shown in Fig.~\ref{mkk_notphi}(c).
We have no explanation of this structure.
We observe no significant structure in the $K^+ K^- K^{\pm}$ mass distribution.
We use the $\phi\ensuremath{\Kp \kern -0.16em \Km}\xspace$ events to investigate the possibility that part of our
$\phi\ensuremath{\pi^+\pi^-}\xspace$ signal is due to $\phi\ensuremath{\Kp \kern -0.16em \Km}\xspace$ events with the two kaons interpreted
as pions.
No structure is present in the resulting \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace invariant mass
distribution.
\section{\boldmath The $\ensuremath{e^+e^-}\xspace \!\ensuremath{\rightarrow}\xspace \phi\pi\pi$ cross section}
\label{phipipistudy}
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig33.eps
\vspace{-0.3cm}
\caption{
The two-Breit-Wigner fit to the $\ensuremath{\pi^+\pi^-}\xspace$ invariant mass distribution
of Fig.~\ref{phif0_sel}(d).
The dashed curve corresponds to the inclusion of the partial width to
$K\kern 0.2em\overline{\kern -0.2em K}{}\xspace$ in the propagator of the $f_0(980)$ BW.
}
\label{pi2fit2bw}
\end{figure}
We next perform a more detailed study in the \ensuremath {\rm E_{c.m.}}\xspace region from
threshold to 3.0~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace of the
$\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\phi(1020)\pi\pi$ cross section.
For this study we use the cross section for the $\phi\ensuremath{\pi^+\pi^-}\xspace$
final state shown in Figs.~\ref{phipipixs} and
\ref{phif0xs}, after scaling by a factor of 1.5 to take into account the
$\phi\ensuremath{\pi^0\pi^0}\xspace$ contribution.
The cross section for the $\phi\ensuremath{\pi^0\pi^0}\xspace$ (see Fig.~\ref{phi2pi0xs})
final state does not help much due to large statistical errors.
There are at least two candidate resonant
structures in Fig.~\ref{phipipixs}. These are associated with the
peaks observed at 1.7~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace and at 2.1~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
As shown in Sec.~\ref{sec:phif01}, the latter is related to
$\phi(1020) f_0(980)$ production, while the best candidate for
the former may be
the $\phi(1680)$, which is a radial excitation of the $s \bar s$ state
decaying predominantly to $K^*(892) \bar K$~\cite{isrkkpi}.
This would be another confirmation of the decay of this state to
$\phi(1020)\pi\pi$, previously reported in Refs.~\cite{isr2k2pi,belle_phif0}.
As discussed in Sec.~\ref{sec:phipipi} we associate
the narrow peak in the $\ensuremath{\pi^+\pi^-}\xspace$ invariant mass distribution, shown on a
larger scale in Fig.~\ref{pi2fit2bw}, with the $f_0(980)$ (denoted as the
$f_0$ meson), and observe a broad enhancement at about 0.6~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace; the angular
distributions of Fig.~\ref{phi_angle} justify that these structures are in
an S-wave state.
This low mass bump
can not be formed by pure three-body phase space. Indeed, the
$\phi(1020)\pi\pi$ threshold is 1.3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, but the observed cross
section has a slow rise starting at 1.4~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace. This indicates that the
observed structure could be a result of
$f_0(600)$ resonance decay.
The observed two-pion-mass shape of the $f_0(600)$ (denoted as the
$\sigma$ meson) is distorted by
the $\phi(1020)\pi\pi$ final state. This is less of an issue for the
narrower $f_0(980)$. Nevertheless,
to obtain mass and width parameter values
for these states, we fit the data of Fig.~\ref{pi2fit2bw} using a
function consisting of an incoherent sum of two S-wave relativistic
BW intensity
distributions, modified to account
for the two pion phase space.
The fit values obtained are
\begin{equation}
m_{\sigma}=(0.692\pm0.030)~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace,
\Gamma_{\sigma}=(0.538\pm0.075)~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
\label{sigpar}
\end{equation}
and
\begin{equation}
m_{f_0}=(0.972\pm0.002)~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace, \Gamma_{f_0}=(0.056\pm0.011)~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
\label{f0par}
\end{equation}
and the fit result is represented by the solid curve in
Fig.~\ref{pi2fit2bw}. Note that the $f_0(980)$ parameters are consistent
with the PDG values~\cite{PDG}, indicating that interference
with the $f_0(600)$ (or $\pi\pi$ coherent continuum) is minimal. This is expected
because events with $m(\pi\pi)<0.85$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace are associated with
the resonance at 1.7~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace in the $\phi(1020)\pi\pi$ mass, while the $f_0(980)$
contributes only to a structure above 2~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace (see Fig.~\ref{phif0xs}).
To confirm this, we examine two $m(\ensuremath{\pi^+\pi^-}\xspace)$ distributions using the selections shown
in Fig.~\ref{phif0_sel}, but for events with either $m(\phi\ensuremath{\pi^+\pi^-}\xspace)<1.95$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace
or $1.95<m(\phi\ensuremath{\pi^+\pi^-}\xspace)<3.0$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
For the first case we observe only the bump at 0.6~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace of
Fig.~\ref{pi2fit2bw}, with no evidence for the $f_0(980)$.
For the second case we see a clear $f_0(980)$ signal but no
evidence for the $f_0(600)$.
We fit each distribution
the same way as the data in Fig.~\ref{pi2fit2bw}. The resulting parameters for the $f_0(600)$
and $f_0(980)$ are in agreement with those presented above.
The dashed curve of Fig.~\ref{pi2fit2bw} is obtained when the $f_0(980)\ensuremath{\rightarrow}\xspace\kern 0.2em\overline{\kern -0.2em K}{}\xspace K$ partial
width is incorporated into the BW propagator (the so called Flatt\'e
approximation used in Ref.~\cite{wa76} with parameters $c1/c2$ and
$m\cdot c1$, which correspond to the ratio of the coupling constants
$g^2_{KK}/g^2_{\pi\pi}$ and effective $f_0(980)$ width). It differs only slightly at
the top of the $f_0(980)$, but the wider shape of the Flatt\'e function
leaves less room for the remaining events and we obtain:
\begin{equation}
m_{\sigma}=(0.631\pm0.020)~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace,
\Gamma_{\sigma}=(0.472\pm0.075)~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
\label{sigpar1}
\end{equation}
The obtained Flatt\'e function parameters are in agreement with those
obtained in Ref.~\cite{wa76}: $c2/c1=2.20\pm0.67$, $m\cdot
c1=0.131\pm0.033$.
The Flatt\'e approximation gives a little better description of the
observed $\pi\pi$ mass spectrum, and so we
use it in the analysis of the structures observed
in the $\phi\pi\pi$ cross section.
It appears that the structure at $\ensuremath {\rm E_{c.m.}}\xspace\approx 2.1\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ in the
$\phi\ensuremath{\pi^+\pi^-}\xspace$ cross section (Figs.~\ref{phipipixs} and~\ref{phif0xs}) couples to the
$f_0(980)$ but not to the $f_0(600)$. This is very similar to the
behavior observed for the $\ensuremath{\pi^+\pi^-}\xspace$ system in $J/\psi\ensuremath{\rightarrow}\xspace\phi\ensuremath{\pi^+\pi^-}\xspace$
decay~\cite{bes4k} (and demonstrated with our data in
Fig.~\ref{jpsi_phipipi} of
Sec.~\ref{sec:charmonium}), and in $D_s\ensuremath{\rightarrow}\xspace\ensuremath{\pi^+\pi^-}\xspace\pi^+$ decay~\cite{D+s}. In
both instances a clear $f_0(980)$ signal is observed, while the broad
$f_0(600)$ enhancement of Fig.~\ref{pi2fit2bw} is absent. In contrast
we note that in $J/\psi\ensuremath{\rightarrow}\xspace\omega\ensuremath{\pi^+\pi^-}\xspace$ decay~\cite{besomega} exactly
the opposite behavior is observed; the $\ensuremath{\pi^+\pi^-}\xspace$ system exhibits a broad
low-mass enhancement, and there is no evidence of an $f_0(980)$ signal.
In contrast with the ``clean'' $m(\phi\ensuremath{\pi^+\pi^-}\xspace)$ distribution, obtained from the fit on the
$\phi$ peak, the $m(\ensuremath{\pi^+\pi^-}\xspace)$ distribution is obtained by the selection
of $\phi$ signal in the $K^+ K^-$ invariant mass distribution, with
background subtraction performed using the $\phi$ side bands and control
region of the \ensuremath{\chi^2}\xspace distribution (see Fig.~\ref{phif0_sel}).
To minimize these uncertainties, we use a BW description for $\sigma$
and the Flatt\'e approximation for $f_0$ to
incorporate these two states in a simple model
describing the structures in the $\phi\pi\pi$ cross section data of
Fig.~\ref{phipipixs} (after scaling by a factor of 1.5 to take into account the
$\phi\ensuremath{\pi^0\pi^0}\xspace$ contribution). The model consists of the incoherent
addition of two contributions at each value of \ensuremath {\rm E_{c.m.}}\xspace. The first
represents the decay process $\phi(1680)\ensuremath{\rightarrow}\xspace\phi f_0(600)$, with the
parameters of the $\sigma$ given by Eq.~(\ref{sigpar1}); the second
results from the coherent superposition of amplitudes describing the
processes $\phi(1680)\ensuremath{\rightarrow}\xspace\phi f_0(980)$ and $Y(2175)\ensuremath{\rightarrow}\xspace\phi
f_0(980)$, where the $Y(2175)$ BW amplitude describes the peak
observed at $\approx$2.2~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace in Fig.~\ref{phipipixs}. We note that
in Ref.~\cite{belle_phif0} the contribution from $\phi(1680)\ensuremath{\rightarrow}\xspace\phi
f_0(980)$ decay was not taken into account. We see no physical evidence to justify
doing this, and so allow the presence of this amplitude in our model.
The angular distributions of Fig.~\ref{phi_angle} are consistent with
the $\phi(1020)$ and the S-wave $\pi\pi$ system being in an S-wave
orbital angular momentum state, and so our model includes no
centrifugal barrier factor in the amplitude representations.
\begin{figure*}[tbh]
\includegraphics[width=0.32\linewidth]{fig34.eps
\includegraphics[width=0.32\linewidth]{fig35.eps
\includegraphics[width=0.32\linewidth]{fig36.eps
\vspace{-0.3cm}
\caption{
(a) The fit to the $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\phi\pi\pi$ cross section
using the model described in the text; the entire contribution due to
the $\phi(1680)$ is shown by the dashed curve. The dotted curve shows
the contribution for only $\phi f_0(980)$ decay. (b) Comparison of the
data and the curve obtained from the overall fit, with the restriction
$m(\pi\pi)<0.85$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace. (c) The $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi(1020)
f_{0}(980)$ cross section with the requirement
$0.85<m(\pi\pi)<1.1$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace; the dashed and dotted curves represent
the contributions from $\phi(1680)\ensuremath{\rightarrow}\xspace\phi(1020) f_{0}(980)$, and
$\phi(1680)\ensuremath{\rightarrow}\xspace\phi(1020) f_{0}(600)$ calculated using the parameter
values from the overall fit to the cross section data.
}
\label{phipipifit3}
\end{figure*}
\begin{table*}[tbh]
\caption{
Summary of parameter values obtained from the fits with
Eq.~(\ref{2bwysig}) described in the text.
An asterisk denotes a value that was fixed in that fit.
}
\label{fittab3}
\begin{ruledtabular}
\begin{tabular}{ c c c c }
Fit & All $m(\pi\pi)$ & $m(\pi\pi)<0.85$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace & $0.85<m(\pi\pi)<1.1$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace \\
\hline
$\sigma_{11}$ (nb) & 0.655$\pm$0.039$\pm$0.040 & 0.678$\pm$0.047$\pm$0.040 &
0.655* \\
$m_1$(\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace) & 1.742$\pm$0.013$\pm$0.012 & 1.733$\pm$0.010$\pm$0.010 &
1.742* \\
$\Gamma_1$(\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace) & 0.337$\pm$0.043$\pm$0.061 & 0.300$\pm$0.015$\pm$0.037 &
0.337* \\
$\sigma_{22}$ (nb) & 0.082$\pm$0.024$\pm$0.010 & 0.082* &
0.094$\pm$0.023$\pm$0.010 \\
$m_2$(\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace) & 2.176$\pm$0.014$\pm$0.004 & 2.176* &
2.172$\pm$0.010$\pm$0.008 \\
$\Gamma_2$(\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace) & 0.090$\pm$0.022$\pm$0.010 & 0.090* &
0.096$\pm$0.019$\pm$0.012 \\
$\sigma_{12}$(nb) & 0.152$\pm$0.034$\pm$0.040 & 0.152* &
0.132$\pm$0.010$\pm$0.010 \\
$\psi$ (rad) & -1.94$\pm$0.34$\pm$0.10 & -1.94* &
-1.92$\pm$0.24$\pm$0.12 \\
\ensuremath{\chi^2}\xspace /n.d.f. & 48/(67-9) & 46/(66-4)&
38/(46-6) \\
P(\ensuremath{\chi^2}\xspace) & 0.74 & 0.96 &
0.40 \\
\end{tabular}
\end{ruledtabular}
\end{table*}
We fit the observed $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\phi\pi\pi$ cross section
using the function
\begin{eqnarray}
\sigma(s) & = &
\frac{P_{\phi\sigma}(s)}{s^{3/2}} \cdot
\left| \frac{A_{11}(s)}{\sqrt{P_{\phi\sigma}(m_1)}}\right|^2 \label{2bwysig} \\
& + & \frac{P_{\phi f_0}(s)}{s^{3/2}} \cdot
\left| \frac{A_{12}(s)e^{i\psi}}{\sqrt{P_{\phi f_0}(m_1)}} +
\frac{A_{22}(s)}{\sqrt{P_{\phi f_0}(m_2)}} \right|^2,\nonumber
\end{eqnarray}
where
\begin{eqnarray}
A_{ij}(s) & = & \frac{\sqrt{\sigma_{ij}} m^{3/2}_{i} m_{i} \Gamma_i}
{m_{i}^2- s -i \sqrt{s} \Gamma_{i}(s)} , \nonumber
\end{eqnarray}
with \\
\\
$i=1$ for the $\phi(1680)$, $i=2$ for the $Y(2175)$,\\
$j=1$ for the $f_0(600)$, $j=2$ for the $f_0(980)$, \\
\\
so that \\
\\
$A_{11}(s)$ describes $\phi(1680)\ensuremath{\rightarrow}\xspace\phi(1020) f_0(600)$ decay, \\
$A_{12}(s)$ describes $\phi(1680)\ensuremath{\rightarrow}\xspace\phi(1020) f_0(980)$ decay, \\
$A_{22}(s)$ describes $Y(2175)\ensuremath{\rightarrow}\xspace\phi(1020) f_0(980)$ decay;\\
\\
$s=\ensuremath {\rm E_{c.m.}}\xspace^2$, $m_1$ and $\Gamma_1$ are the mass and width of the
$\phi(1680)$, $m_2$ and $\Gamma_2$ are the mass and width of the
$Y(2175)$, and the $\sigma_{ij}$ represent the peak cross section values.
The factors
$P_{\phi\sigma}(s)$ and $P_{\phi f_0}(s)$ represent quasi-two body
phase space integrated over the range of $\pi\pi$ invariant mass
available at $\ensuremath {\rm E_{c.m.}}\xspace=\sqrt{s}$, and are obtained from
\begin{equation}
P_{\phi\pi\pi}(s)=\int_{2m_\pi}^{\sqrt{s}-m_{\phi}}
BW_{\pi\pi}(m) q(s,m,m_{\phi}) dm,
\label{ps}
\end{equation}
where $BW_{\pi\pi}(m)$ is a BW function with
$f_0(980)$ parameters ($BW_{f_0}(m)$) to define $P_{\phi f_0}(s)$, or with
$f_0(600)$ parameters ($BW_{\sigma}(m)$) to define
$P_{\phi\sigma}(s)$~\cite{achasov_koz}, and $q$ is the
momentum of the particles with masses $m$ and $m_{\phi}$ in the
two-body reaction at $\ensuremath {\rm E_{c.m.}}\xspace=\sqrt{s}$.
Since the decay $\phi(1680)\ensuremath{\rightarrow}\xspace\phi(1020) f_0(980)$ is suppressed by
phase space near $\sqrt{s}=m_1$, the value of $\sigma_{12}$ is much
smaller than that of $\sigma_{11}$, but its contribution to $\sigma(s)$
increases rapidly beyond the $\phi(1020) f_0(980)$ threshold.
The $\phi(1680)$ resonance
decays mainly to $\kern 0.2em\overline{\kern -0.2em K}{}\xspace K^*(892)$ and $\phi(1020)\eta$~\cite{PDG,isrkkpi}.
We find that it has a branching fraction of about
10\% to $\phi\pi\pi$, which together with other modes
listed in PDG, leads to an energy-dependent width
that can be written as
\begin{eqnarray}
\Gamma_{1}(s) & = & \Gamma_{1}\Bigl[0.7\frac{m_1^3 P_{2K}(s)}{s^{3/2} P_{2K}(m_1^2)} \nonumber \\
& + & 0.2\frac{m_1 P_{\phi\eta}(s)}{s^{1/2} P_{\phi\eta}(m_1^2)}
+ 0.1\frac{m_1 P_{\phi\pi\pi}(s)}{s^{1/2} P_{\phi\pi\pi}(m_1^2)}\Bigr],
\end{eqnarray}
with $P_{2K}(s)=q^{3}(\sqrt{s},m_{K},m_{K^*})$, and $P_{\phi\eta}(s)=q(\sqrt{s},m_{\phi},m_{\eta})$.
For the second resonance candidate, which decays mostly to $\phi\pi\pi$,
the energy dependence of the width is written as
\begin{eqnarray}
\Gamma_{2}(s) & = & \Gamma_{2}\frac{m_2 P_{\phi\pi\pi}(s)}
{s^{1/2} P_{\phi\pi\pi}(m_2^2)}.
\end{eqnarray}
We note that the introduction of an energy dependence for each width
significantly increases the values of the resonance mass and width,
especially for broad structures.
The results of the fits are shown in Fig.~\ref{phipipifit3} and
summarized in Table~\ref{fittab3}. The first error is statistical, and the second
error represents the systematic uncertainty estimated as a difference in fitted
values for two different descriptions of the two-pion spectrum as shown in
Fig.~\ref{pi2fit2bw}.
In Fig.~\ref{phipipifit3}(a) we show the contribution from the $\phi(1680)$ for
both modes (dashed curves), and for $\phi(1680)\ensuremath{\rightarrow}\xspace\phi(1020)
f_0(980)$ only (dotted curve).
The increase of the cross section at about 2~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace is explained by
the opening of the $\phi f_0(980)$ decay channel of the $\phi(1680)$
resonance. However the fit shows that an additional relatively narrow
state is needed in order to provide a
better description of the observed data.
It is important to note that this model describes the observed data very well
independently of the $m(\pi\pi)$ region selected.
Figure~\ref{phipipifit3}(b) shows the $\phi\pi\pi$ cross section for
$m(\ensuremath{\pi^+\pi^-}\xspace)<0.85$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace for the data; the curve is obtained by using the
parameter values from the overall fit and yields
\ensuremath{\chi^2}\xspace/n.d.f. = 63/(66-1) (P(\ensuremath{\chi^2}\xspace) = 0.54).
If we fit this distribution, slightly better parameter values can
be obtained (see Table~\ref{fittab3}), but these still agree well with
those from the overall fit. We consider them as our measurement of the
$\phi(1680)$ resonance parameters. They correspond to
the product of the electronic
width, $\Gamma_{ee}$, and branching fraction to $\phi\pi\pi$, ${\ensuremath{\cal B}\xspace}_{\phi\pi\pi}$,
$$
{\ensuremath{\cal B}\xspace}_{\phi\pi\pi} \cdot \Gamma_{ee} =
\frac{\Gamma_1 \sigma_{11} m_1^2}{12\pi C } = (42\pm 2\pm 3)~\ensuremath{\rm \,e\kern -0.08em V}\ ,
$$
where we fit the product $\Gamma_1 \sigma_{11}$ to reduce correlations,
and $C$, the conversion constant, is $0.389$~mb(\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace)$^2$~\cite{PDG}.
The second error is systematic, and corresponds to the normalization
uncertainty on the cross section, and to the uncertainty in the $m(\pi\pi)$
distribution description.
If we require $0.85 < m(\pi\pi) < 1.1$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace
(Fig.~\ref{phipipifit3}(c)), then
without additional fitting the model yields \ensuremath{\chi^2}\xspace/n.d.f. = 48/(46-1)
(P(\ensuremath{\chi^2}\xspace) = 0.31), and improves
to \ensuremath{\chi^2}\xspace/n.d.f. = 38/(46-6) (P(\ensuremath{\chi^2}\xspace) = 0.40 )
by refitting using the parameter values listed
in Table~\ref{fittab3}. If we try to explain the observed cross
section only in terms of the $\phi(1680)$ without any narrow state (dashed
curve in Fig.~\ref{phipipifit3}(c)), the fit
gives \ensuremath{\chi^2}\xspace/n.d.f. = 123/(46-2) (P(\ensuremath{\chi^2}\xspace) = $10^{-7}$) and so this hypothesis
is not compatible with the data.
Note, that the contribution of $\phi f_0(600)$, shown by
dotted curve in Fig.~\ref{phipipifit3}(c), is very small.
The model described above provides an excellent description
of the observed cross section behavior, and suggests that
the $Y(2175)$ may not be a radially excited $s\bar s$ state,
since such a state would be expected to be
much wider (300-400~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace) and also should decay to $\phi f_0(600)$,
like the $\phi(1680)$.
\begin{figure}[tbh]
\includegraphics[width=0.95\linewidth]{fig37.eps
\vspace{-0.3cm}
\caption{
The $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi(1020) f_{0}(980)$ cross section measured
in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace (solid dots) and \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace (open squares) final states.
The solid (dashed) curve represents the result of the two-resonance
(one-resonance - $\phi(1680)\ensuremath{\rightarrow}\xspace\phi(1020) f_{0}(980)$)
fit using Eq.(\ref{2bwysig}), as described in the text.
The hatched area and dotted curve show the $Y(2175)$ contribution for two solutions.
}
\label{phif0xsall}
\includegraphics[width=0.95\linewidth]{fig38.eps
\vspace{-0.3cm}
\caption{
The $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi(1020) f_{0}(980)$ cross section
measurements from the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace final state from ~\mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}~(dots) and
Belle~\cite{belle_phif0}(squares).
}
\label{phif0xsbelle}
\end{figure}
\begin{figure*}[tbh]
\includegraphics[width=0.32\linewidth]{fig39.eps
\includegraphics[width=0.32\linewidth]{fig40.eps
\includegraphics[width=0.32\linewidth]{fig41.eps
\vspace{-0.3cm}
\caption{
(a) The $m(\ensuremath{\pi^+\pi^-}\xspace)$ distribution without background subtraction for
$K^+ K^-\ensuremath{\pi^+\pi^-}\xspace$ events. The vertical lines indicate the $f_0(980)$ region.
(b) All selected $K^+ K^-\ensuremath{\pi^+\pi^-}\xspace$ events (open histogram), selected
$K^+ K^- f_0(980)$ events (cross-hatched histogram), and all the rest
(hatched histogram).
(c) The $K^+ K^- f_0(980)$ events (open histogram) in comparison with the
$\phi(1020) f_0(980)$ sample (hatched histogram).
}
\label{select_2kf0ch}
\end{figure*}
\begin{figure*}[tbh]
\includegraphics[width=0.32\linewidth]{fig42.eps
\includegraphics[width=0.32\linewidth]{fig43.eps
\includegraphics[width=0.32\linewidth]{fig44.eps
\vspace{-0.3cm}
\caption{
(a) The $m(\ensuremath{\pi^0\pi^0}\xspace)$ distribution without background subtraction for
$K^+ K^-\ensuremath{\pi^0\pi^0}\xspace$ events. The vertical lines indicate the $f_0(980)$ region.
(b) All selected $K^+ K^-\ensuremath{\pi^0\pi^0}\xspace$ events (open histogram), selected
$K^+ K^- f_0(980)$ events (cross-hatched histogram), and all the rest
(hatched histogram).
(c) The $K^+ K^- f_0(980)$ events (open histogram) in comparison with the
$\phi(1020) f_0(980)$ sample (hatched histogram).
}
\label{select_2kf0nu}
\end{figure*}
\begin{figure*}[tbh]
\begin{center}
\includegraphics[width=0.32\linewidth]{fig45.eps
\includegraphics[width=0.32\linewidth]{fig46.eps
\includegraphics[width=0.32\linewidth]{fig47.eps
\vspace{-0.4cm}
\caption{
Raw invariant mass distribution for all selected events in the
charmonium region for (a)
$\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace$, (b) $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace K^+ K^- \ensuremath{\pi^0\pi^0}\xspace$, and
(c) $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace K^+ K^- K^+ K^- $; in each figure the curve
represents the result of the fit described in the text.
}
\label{jpsi}
\end{center}
\end{figure*}
\section{\boldmath $\ensuremath{e^+e^-}\xspace \!\ensuremath{\rightarrow}\xspace \phi f_0$ Near Threshold}
\label{phif0bump}
The behavior of the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi f_0$ cross section near
threshold shows a structure near 2.175~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
and we have published this result in Ref.~\cite{isr2k2pi}.
Here we provide a more detailed study of the cross section for this channel in the
1.8--3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace region with the full \mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}~ dataset.
In Fig.~\ref{phif0xsall} we superimpose the cross
sections measured in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace
and \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace final states (shown in Figs.~\ref{phif0xs} and~\ref{phif0xs2});
they are consistent with each other.
We perform a combined fit to these cross section data using Eq.(\ref{2bwysig})
with the two-pion mass restricted to the region 0.85-1.1~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
We fix the $\phi(1680)$ parameters for the $\phi(1020) f_0(600)$ decay mode
(which gives a small contribution in this mass range) and allow all
other parameters to float.
The result of the fit is shown as the solid curve in
Fig.~\ref{phif0xsall}.
As demonstrated in Ref.~\cite{belle_phif0}, the observed
pattern can be a result of a constructive or destructive interference of
the narrow structure at 2.175~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace with the coherent background.
The fit with constructive interference gives the resonance parameter values
\begin{eqnarray*}
\sigma_{22} & = & (0.093 \pm 0.021 \pm 0.010)~{\rm nb} , \\
m_2 & = & (2.180 \pm 0.008 \pm 0.008)~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace , \\
\Gamma_2 & = & (0.077 \pm 0.015 \pm 0.010)~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace , \\
\psi_2 & = & (-2.11 \pm 0.24 \pm 0.12)~{\rm rad} ,\\
\sigma_{12} & = & (0.140 \pm 0.009 \pm 0.010)~{\rm nb,}
\end{eqnarray*}
and \ensuremath{\chi^2}\xspace/n.d.f.$=\! 57/(61-6)$ (P(\ensuremath{\chi^2}\xspace) = 0.33).
The statistical precision is improved compared to that of
Ref.~\cite{isr2k2pi}, for which the analysis was
based on half as much data.
For this state we estimate the product of electronic width and branching
fraction to $\phi f_0$ as
$$
{\ensuremath{\cal B}\xspace}_{\phi f_0} \cdot \Gamma_{ee} =
\frac{\Gamma_2 \sigma_{22} m_2^2}{12\pi C } = (2.3\pm 0.3\pm 0.3)~\ensuremath{\rm \,e\kern -0.08em V}\ ,
$$
where we fit the product $\Gamma_2 \sigma_{22}$ to reduce correlations.
The second error is systematic, and corresponds to the normalization
uncertainty on the cross section.
The destructive interference yields exactly the same
overall curve with the same
parameters for the mass and width of the narrow state, but
significantly larger peak
cross section with opposite sign of the mixing angle:
$\sigma_{22} = (1.13 \pm 0.15 \pm 0.12)~{\rm nb}$, $ \psi_2 =(2.47
\pm 0.17 \pm 0.13)~{\rm rad} $. To select between two
solutions, we need more information
on the decay rates to another modes, which are not available now.
If we assume no resonance structure other than the tail from
$\phi(1680)\ensuremath{\rightarrow}\xspace\phi(1020) f_0(980)$, the fit yields
\ensuremath{\chi^2}\xspace/n.d.f.$= 150/(61-2)$ with P(\ensuremath{\chi^2}\xspace)=$8\cdot 10^{-9}$.
The result of this fit
is shown as the dashed curve in Fig.~\ref{phif0xsall}.
It is a poor fit to the region below 2.3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, but
gives a good description of the cross section behavior at higher
values of \ensuremath {\rm E_{c.m.}}\xspace.
The fit, with or without the resonance at 2.18~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace, gives a
maximum value of the $\phi(1680)\ensuremath{\rightarrow}\xspace\phi f_0$
cross section of 0.3 nb at $\ensuremath {\rm E_{c.m.}}\xspace\approx 2.1\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$.
This is of independent theoretical interest, because it can be related
to the $\phi\ensuremath{\rightarrow}\xspace f_0(980)\gamma$ decay studied at the
$\phi$-factory~\cite{phif0theory,phif0dispersion}.
The significance of the structure calculated from the change in \ensuremath{\chi^2}\xspace
between the fits with and without the resonance at 2.18~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace
is $\sqrt{150 - 61} = 9.4$ standard deviations; the \ensuremath{\chi^2}\xspace value, 61
for 61-2 n.d.f., yields the same probability as the \ensuremath{\chi^2}\xspace value 57
for 61-6 n.d.f..
The cross section measurements from the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace final state
shown in Fig.~\ref{phif0xsall} are compared to those from
Belle~\cite{belle_phif0} in Fig.~\ref{phif0xsbelle}.
There is good overall agreement between the results from the two
experiments.
Overall agreement between the results of the fits to
the ~\mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}~ and Belle data is also good.
\subsection{Structures in the $K^+ K^- f_0(980)$ final state}
We next search for other decay modes of the $Y(2175)$ state.
Figure~\ref{select_2kf0ch}(a) shows the ``raw''
(no background subtraction) two-pion mass distribution for all selected
$K^+ K^-\ensuremath{\pi^+\pi^-}\xspace$ events, and Fig.~\ref{select_2kf0nu}(a) shows the
same distribution for the $K^+ K^-\ensuremath{\pi^0\pi^0}\xspace$ sample. The $f_0(980)$ contribution is
relatively small for the charged-pion mode, and larger for the
neutral-pion mode.
If we select the region $0.85< m(\pi\pi)<1.1$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace, and plot the $K^+
K^-\pi\pi$ mass distribution, the bump at 2.175~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace is seen
much more clearly in spite of larger background
(Figs.~\ref{select_2kf0ch}(b) and ~\ref{select_2kf0nu}(b)), and a bump at
2.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace is also seen; the rest
of events have no structures at 2.175~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace or 2.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace
(Figs.~\ref{select_2kf0ch}(b) and ~\ref{select_2kf0nu}(b) hatched histograms).
The bumps are
seen only in the $K^+ K^- f_0(980)$ sample (Figs.~\ref{select_2kf0ch}(c)
and ~\ref{select_2kf0nu}(c)), but if we select the
$\phi(1020)$ region, no bumps are seen at 2.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace, as shown by the
hatched histograms in
Figs.~\ref{select_2kf0ch}(c) and ~\ref{select_2kf0nu}(c).
From these histograms we can conclude that the $Y(2175)$ resonance
has a $K^+ K^- f_0(980)$ decay mode when the $\ensuremath{\Kp \kern -0.16em \Km}\xspace$ system is not from $\phi$,
and that the decay rate is comparable to that for $\phi f_0(980)$. Also another
state at 2.5~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace seems to exist; this decays to $K^+ K^- f_0(980)$
(but seems not to couple
to $\phi f_0(980)$) with width $\approx$0.06-0.08~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace (see
Ref.~\cite{isr2k2pi}). The large background does not
allow us to clearly separate this state.
\section{\boldmath The Charmonium Region}
\label{sec:charmonium}
For the \ensuremath {\rm E_{c.m.}}\xspace region above 3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, our data can be used to measure,
or set limits on, the decay branching fractions for the
$J/\psi$ and $\psi(2S)$ (See Figs.~\ref{2k2pi_ee_babar},
~\ref{2k2pi0_ee_babar}, and ~\ref{4k_ee_babar}).
In addition, these signals allow checks of our mass
scale, and of our measurements of mass resolution.
Figure~\ref{jpsi} shows the invariant
mass distributions for the selected \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace, \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace, and \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace events,
respectively, in this region, using smaller mass intervals
than in the corresponding Figs.~\ref{2k2pi_babar},
~\ref{2k2pi0_babar}, and~\ref{4k_babar}.
We do not subtract any background from the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace and \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace distributions,
since it is small and nearly uniformly distributed, but
we use the \ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace control region to subtract part of the ISR
background from the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace distribution.
Production of the $J/\psi$ is apparent in all three distributions,
and a small, but clear, $\psi(2S)$ signal is visible in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace mode.
We fit each of these distributions using a sum of two Gaussian functions
to describe the $J/\psi$ signal and incorporate a similar
representation of a $\psi (2S)$ signal, although there is no clear
evidence of the latter in Figs.~\ref{jpsi}(b) and \ref{jpsi}(c).
In each case, a second-order-polynomial function is used to
describe the remainder of the distribution.
We take the signal function parameter values from simulation, but let
the overall mean and width values vary in the fits, together with
the coefficients of the polynomial. For the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace and \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace
modes we fix the $\psi (2S)$ mass position~\cite{PDG}, and take the
width from MC simulation.
The fits are of good quality, and are shown by the curves in
Fig.~\ref{jpsi}.
In all cases, the fitted mean value is within 1~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace of the nominal
$J/\psi$ or $\psi (2S)$ mass position~\cite{PDG}
and the width is within 10\% of the simulated resolution
discussed in Secs.~\ref{sec:xs2k2pi}, \ref{sec:2k2pi0xs}, and~\ref{sec:4kxs}.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig48.eps
\vspace{-0.6cm}
\caption{
The $K^\pm\pi^\mp$ invariant mass versus \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace invariant mass for
events with the $K^\mp\pi^\pm$ combination in one of the
$K^{*}(892)^{0}$ regions of Fig.~\ref{kkstar}(a); for events in overlap region,
only one combination is chosen.
}
\label{jpsi_kkstarvs2k2pi}
\end{figure}
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig49.eps
\vspace{-0.6cm}
\caption{
The $K^\pm\pi^\mp$ mass projection for events from Fig.~\ref{jpsi_kkstarvs2k2pi}
with \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace invariant mass within 50~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace of the nominal $J/\psi$ mass
(open histogram), and for events for which this mass value is
50--100~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace less than nominal (hatched).
}
\label{jpsi_kk2}
\end{figure}
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig50.eps
\label{jjj}
\includegraphics[width=0.9\linewidth]{fig51.eps
\vspace{-0.3cm}
\caption{
Raw invariant mass distributions in the charmonium region for
(a) candidate $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^+\pi^-}\xspace$ events (open histogram), and for
events in the $\phi$ sideband regions of Fig.~\ref{phif0_sel}(c)
(hatched); (b) candidate $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^0\pi^0}\xspace$ events (open histogram) and
events in the \ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace control region (hatched).
}
\label{jpsi_phi2pi}
\end{figure}
The fitted $J/\psi$ signals for the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace, \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace, and \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace final
states are found to contain
$3137\pm67$,
$388\pm28$, and
$287\pm24$ events, respectively.
From the number of events in each final state $f$,
$N_{J/\psi \!\ensuremath{\rightarrow}\xspace\! f}$,
we calculate the product of the $J/\psi$ branching fraction to $f$ and
the $J/\psi$ electronic width using
\begin{equation}
{\ensuremath{\cal B}\xspace}_{J/\psi \!\ensuremath{\rightarrow}\xspace\! f} \cdot \Gamma^{J/\psi}_{ee} =
\frac{N_{J/\psi \!\ensuremath{\rightarrow}\xspace\! f} \cdot m_{J/\psi}^2}
{6\pi^2 \cdot d{\cal L}/dE \cdot \epsilon_f(m_{J/\psi}) \cdot C}
~~~, \\
\label{jpsicalc}
\end{equation}
where
$d{\cal L}/dE = 173.1\pm1.7~\ensuremath{\mbox{\,nb}^{-1}}\xspace/\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$, and $\epsilon_f(m_{J/\psi})$
are the ISR luminosity and corrected selection efficiency, respectively,
at the $J/\psi$ mass,
and $C$ is the conversion constant.
We estimate $\epsilon_{\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace} \! =\! 0.198\pm0.006$,
$\epsilon_{\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace} \! =\! 0.079\pm0.004$, and $\epsilon_{\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace} \!
=\! 0.173\pm0.012$ using the corrections and errors discussed
in Secs.~\ref{sec:xs2k2pi}, \ref{sec:2k2pi0xs}, and~\ref{sec:4kxs}.
We list the values of the product of the branching fraction(s) and
$\Gamma^{J/\psi}_{ee}$ in Table~\ref{jpsitab}, and using
$\Gamma^{J/\psi}_{ee}\! =\! (5.55\pm0.14)~\ensuremath{\mathrm{\,ke\kern -0.1em V}}\xspace$~\cite{PDG},
obtain the corresponding branching fraction values and list them
together with their PDG values~\cite{PDG}.
The systematic uncertainties quoted include a 2.5\% uncertainty on $\Gamma^{J/\psi}_{ee}$.
Our measured branching fractions of \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace, \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace, and \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace are
more precise than the current PDG values,
which are dominated by our previous results
((6.6$\pm$0.5)$\times$10$^{-3}$, (2.5$\pm$0.3)$\times$10$^{-3}$ and
(7.6$\pm$0.9)$\times$10$^{-4}$, respectively~\cite{isr2k2pi}).
These fits also yield 133$\pm$21 \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace events,
17$\pm$9 \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace events and 13$\pm$6 \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace events
in the $\psi(2S)$ peak.
We expect 12 events from
$\psi(2S) \!\ensuremath{\rightarrow}\xspace\! J/\psi\ensuremath{\pi^+\pi^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace$
from the relevant branching fractions~\cite{PDG},
which is less than the statistical error.
Subtracting this contribution and using the calculation analogous
to Eq.(\ref{jpsicalc}), with $d{\cal L}/dE\! =\! 221.2\pm2.2~\ensuremath{\mbox{\,nb}^{-1}}\xspace/\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$,
we obtain the product of the branching fraction and electronic width for the decays
$\psi(2S) \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace$, $\psi(2S) \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace$, and
$\psi(2S) \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace$.
Dividing by $\Gamma^{\psi(2S)}_{ee}=2.36\pm0.04$~\ensuremath{\mathrm{\,ke\kern -0.1em V}}\xspace~\cite{PDG},
we obtain the branching fractions listed in Table~\ref{jpsitab}.
The \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace and \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace values are consistent with those in
Ref.~\cite{PDG}. There is no entry in Ref.~\cite{PDG}
for the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace decay mode of the $\psi(2S)$.
As noted in Sec.~\ref{sec:kaons} and shown in Fig.~\ref{kkstar} and
Fig.~\ref{kstar_sel},
the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace final state is dominated by the $K^{*}(892)^{0} K^- \pi^+$ channels,
with a small contribution from the $K^{*}(892)^0 \kern 0.2em\overline{\kern -0.2em K}{}\xspace_2^{*}(1430)^0$
channels.
Figure~\ref{jpsi_kkstarvs2k2pi} shows a plot of the invariant
mass of a $K^\pm\pi^\mp$ pair versus that of the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace system for events with
the mass of the $K^\mp\pi^\pm$ pair near the $K^{*}(892)^0$ mass,
i.e., within the bands in Fig.~\ref{kkstar}(a), but with only one
combination plotted in the overlap region.
There is a large concentration of entries in the $J/\psi$ band with
$K^\pm\pi^\mp$ mass values near 1.43~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace,
but a relatively small number of events in a horizontal band corresponding to
the $K_2^{*}(1430)^0$ production outside the $J/\psi$ region.
We show the $K^\pm\pi^\mp$ mass projection for the subset of events
with \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace mass within 50~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace of the nominal $J/\psi$ mass
in Fig.~\ref{jpsi_kk2} as the open histogram.
The hatched histogram is the projection for events with a \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace mass
between 50 and 100~\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace away from the nominal $J/\psi$ mass.
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig53.eps
\label{jjjj}
\includegraphics[width=0.9\linewidth]{fig52.eps
\vspace{-0.3cm}
\caption{
Raw invariant mass distribution in the charmonium region (a)
for candidate $\phi f_0$,
$f_0 \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{\pi^+\pi^-}\xspace$ events (open histogram), and for
events in the $\phi$ sideband region (hatched), and
(b) for candidate $\phi f_0$,
$f_0 \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{\pi^0\pi^0}\xspace$ events (open histogram) and for
events in the \ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace control region (hatched).
}
\label{jpsi_phif0}
\end{figure}
\begin{figure}[tbh]
\includegraphics[width=0.9\linewidth]{fig54.eps
\vspace{-0.3cm}
\caption{
The $\ensuremath{\pi^+\pi^-}\xspace$ invariant mass distribution for $\phi \ensuremath{\pi^+\pi^-}\xspace$ events from
the $J/\psi$ peak of Fig.~\ref{jpsi_phi2pi}(a) (open histogram), and
for events in the $\phi$ sideband region (hatched).
}
\label{jpsi_phipipi}
\end{figure}
\begin{table*}[tbh]
\caption{
Summary of the $J/\psi$ and $\psi(2S)$ parameters
obtained in this analysis.
}
\label{jpsitab}
\begin{ruledtabular}
\begin{tabular}{r@{$\cdot$}l r@{.}l@{$\pm$}l@{$\pm$}l
r@{.}l@{$\pm$}l@{$\pm$}l
r@{.}l@{$\pm$}l }
\multicolumn{2}{c}{Measured} & \multicolumn{4}{c}{Measured} &
\multicolumn{7}{c}{$J/\psi$ or $\psi(2S)$ Branching Fraction (10$^{-3}$)}\\
\multicolumn{2}{c}{Quantity} & \multicolumn{4}{c}{Value (\ensuremath{\rm \,e\kern -0.08em V})} &
\multicolumn{4}{c}{This work} &
\multicolumn{3}{c}{PDG2010} \\
\hline
$\Gamma^{J/\psi}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{J/\psi \ensuremath{\rightarrow}\xspace \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace}$ &
37&94&0.81& 1.10 & 6&84 & 0.15 & 0.27 & 6&6 & 0.5 \\
$\Gamma^{J/\psi}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{J/\psi \ensuremath{\rightarrow}\xspace \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace}$ &
11&75&0.81& 0.90 & 2&12 & 0.15 & 0.18 & 2&45 & 0.31 \\
$\Gamma^{J/\psi}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{J/\psi \ensuremath{\rightarrow}\xspace \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace} $ &
4&00& 0.33& 0.29 & 0&72 & 0.06 & 0.05 & 0&76 & 0.09 \\[0.4cm]
$\Gamma^{J/\psi}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{J/\psi\ensuremath{\rightarrow}\xspace K^{*0}\kern 0.2em\overline{\kern -0.2em K}{}\xspace_{0,2}^{*0}}
\cdot {\ensuremath{\cal B}\xspace}_{K^{*0}\ensuremath{\rightarrow}\xspace K^+\pi^-}
\cdot {\ensuremath{\cal B}\xspace}_{\kern 0.2em\overline{\kern -0.2em K}{}\xspace_{0,2}^{*0}\ensuremath{\rightarrow}\xspace K^-\pi^+} $ &
8&59& 0.36 & 0.27 & 6&98 & 0.29 & 0.21 & 6&0 & 0.6 \\
$\Gamma^{J/\psi}_{ee} $ & ${\ensuremath{\cal B}\xspace}_{J/\psi\ensuremath{\rightarrow}\xspace K^{*0}\kern 0.2em\overline{\kern -0.2em K}{}\xspace^{*0}}
\cdot{\ensuremath{\cal B}\xspace}_{K^{*0}\ensuremath{\rightarrow}\xspace K^+\pi^-}
\cdot{\ensuremath{\cal B}\xspace}_{\kern 0.2em\overline{\kern -0.2em K}{}\xspace^{*0}\ensuremath{\rightarrow}\xspace K^-\pi^+}$ &
0&57 & 0.15 & 0.03 & 0&23 & 0.06 & 0.01 & 0&23 & 0.07 \\
$\Gamma^{J/\psi}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{J/\psi \ensuremath{\rightarrow}\xspace \phi\ensuremath{\pi^+\pi^-}\xspace}
\cdot {\ensuremath{\cal B}\xspace}_{\phi \ensuremath{\rightarrow}\xspace \ensuremath{K^+}\xspace \ensuremath{K^-}\xspace}$ &
2&19& 0.23& 0.07 & 0&81 & 0.08 & 0.03 & 0&94 & 0.09 \\
$\Gamma^{J/\psi}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{J/\psi \ensuremath{\rightarrow}\xspace \phi\ensuremath{\pi^0\pi^0}\xspace}
\cdot {\ensuremath{\cal B}\xspace}_{\phi \ensuremath{\rightarrow}\xspace \ensuremath{K^+}\xspace \ensuremath{K^-}\xspace}$ &
1&36& 0.27& 0.07 & 0&50 & 0.10 & 0.03 & 0&56 & 0.16 \\
$\Gamma^{J/\psi}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{J/\psi \ensuremath{\rightarrow}\xspace \phi\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}
\cdot {\ensuremath{\cal B}\xspace}_{\phi \ensuremath{\rightarrow}\xspace \ensuremath{K^+}\xspace \ensuremath{K^-}\xspace}$ &
2&26& 0.26& 0.16 & 1&66 & 0.19 & 0.12 & 1&83 & 0.24
\footnote{$\phi$ is selected as $|m_{\phi}-m(\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace)|<10$~\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace, ${\ensuremath{\cal B}\xspace}_{J/\psi\ensuremath{\rightarrow}\xspace\phi\kern 0.2em\overline{\kern -0.2em K}{}\xspace K}$ obtained as $2\cdot{\ensuremath{\cal B}\xspace}_{J/\psi\ensuremath{\rightarrow}\xspace\phi\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}$.}\\
$\Gamma^{J/\psi}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{J/\psi \ensuremath{\rightarrow}\xspace \phi f_0}
\cdot {\ensuremath{\cal B}\xspace}_{ \phi \ensuremath{\rightarrow}\xspace \ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}
\cdot {\ensuremath{\cal B}\xspace}_{ f_0 \ensuremath{\rightarrow}\xspace \ensuremath{\pi^+\pi^-}\xspace} $ &
0&69& 0.11& 0.05 & 0&25 & 0.04 & 0.02 & 0&18 & 0.04
\footnote{Not corrected for the $f_0\ensuremath{\rightarrow}\xspace\ensuremath{\pi^0\pi^0}\xspace$ mode. $f_0$ selected by $0.85<m(\ensuremath{\pi^0\pi^0}\xspace)<1.1$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace} \\
$\Gamma^{J/\psi}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{J/\psi \ensuremath{\rightarrow}\xspace \phi f_0}
\cdot {\ensuremath{\cal B}\xspace}_{ \phi \ensuremath{\rightarrow}\xspace \ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}
\cdot {\ensuremath{\cal B}\xspace}_{ f_0 \ensuremath{\rightarrow}\xspace \ensuremath{\pi^0\pi^0}\xspace} $ &
0&48& 0.12& 0.05 & 0&18 & 0.04& 0.02 & 0&17 & 0.07
\footnote{Not corrected for the $f_0\ensuremath{\rightarrow}\xspace\ensuremath{\pi^+\pi^-}\xspace$ mode. $f_0$ selected by $0.85<m(\ensuremath{\pi^+\pi^-}\xspace)<1.1$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace } \\
$\Gamma^{J/\psi}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{J/\psi \ensuremath{\rightarrow}\xspace \phi f_x}
\cdot {\ensuremath{\cal B}\xspace}_{ \phi \ensuremath{\rightarrow}\xspace K^+K^-}
\cdot {\ensuremath{\cal B}\xspace}_{ f_x \ensuremath{\rightarrow}\xspace \ensuremath{\pi^+\pi^-}\xspace} $ &
0&74& 0.12& 0.05 & 0&27 & 0.04 & 0.02 & 0&72 & 0.13
\footnote{We compare our $\phi f_x, f_x\ensuremath{\rightarrow}\xspace\ensuremath{\pi^+\pi^-}\xspace$ mode, selected by
$1.1<m(\ensuremath{\pi^+\pi^-}\xspace)<1.5$ with $\phi f_2(1270)$.} \\[0.4cm]
$\Gamma^{\psi(2S)}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{\psi(2S) \ensuremath{\rightarrow}\xspace \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace} $ &
1&92& 0.30& 0.06 & 0&81 & 0.13 & 0.03 & 0&75 & 0.09 \\
$\Gamma^{\psi(2S)}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{\psi(2S) \ensuremath{\rightarrow}\xspace \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace} $ &
0&60& 0.31& 0.03 & 0&25 & 0.13 & 0.02 & \multicolumn{3}{c}{no entry} \\
$\Gamma^{\psi(2S)}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{\psi(2S) \ensuremath{\rightarrow}\xspace \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace} $ &
0&22& 0.10& 0.02 & 0&09 & 0.04 & 0.01 & 0&060 & 0.014 \\
$\Gamma^{\psi(2S)}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{\psi(2S) \ensuremath{\rightarrow}\xspace \phi\ensuremath{\pi^+\pi^-}\xspace}
\cdot {\ensuremath{\cal B}\xspace}_{\phi \ensuremath{\rightarrow}\xspace \ensuremath{K^+}\xspace \ensuremath{K^-}\xspace} $ &
0&27& 0.09& 0.02 & 0&23 & 0.08 & 0.01 & 0&117& 0.029 \\
$\Gamma^{\psi(2S)}_{ee}$ & ${\ensuremath{\cal B}\xspace}_{\psi(2S) \ensuremath{\rightarrow}\xspace \phi f_0}
\cdot {\ensuremath{\cal B}\xspace}_{\phi \ensuremath{\rightarrow}\xspace K^+K^-}
\cdot {\ensuremath{\cal B}\xspace}_{ f_0 \ensuremath{\rightarrow}\xspace \ensuremath{\pi^+\pi^-}\xspace} $ &
\hspace*{0.5cm} 0&17 & 0.06 & 0.02 \hspace{0.6cm} &
\hspace*{0.5cm} 0&15 & 0.05 & 0.01 \hspace{0.6cm} &
\hspace*{0.5cm} 0&068 & 0.024
\footnote{ ${\ensuremath{\cal B}\xspace}_{\psi(2S) \ensuremath{\rightarrow}\xspace \phi f_0}, f_0\ensuremath{\rightarrow}\xspace \ensuremath{\pi^+\pi^-}\xspace$ } \hspace{0.6cm} \\
\end{tabular}
\end{ruledtabular}
\end{table*}
The $K \pi$ distribution from the $J/\psi$ is dominated by the
$K_2^{*}(1430)^0$ and $K_0^{*}(1430)^0$
signals~\cite{PDG,beskstar}.
A small signal at the $K^{*}(892)^0$
indicates the presence of $K^{*}(892)^0\bar K^{*}(892)^0$
decay of the $J/\psi$; this is also seen
as an enhancement in the $J/\psi$ band in Fig.~\ref{jpsi_kkstarvs2k2pi}.
The enhancement at 1.9~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace of Fig.~\ref{jpsi_kk2}
may be due to the $^3 F_2$ ground state, or to the first
radial excitation of the $K^*_2(1430)$, neither of
which has been reported previously.
Subtracting the number of sideband events from the number in the
$J/\psi$ mass window,
we obtain 710$\pm$30 events with $K^\pm\pi^\mp$ mass in the range
1.2--1.7~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace,
which we take as a measure of $J/\psi$ decay into
$K^{*}(892)^0 \kern 0.2em\overline{\kern -0.2em K}{}\xspace_{0,2}^{*}(1430)^0$.
According to Ref.~\cite{beskstar}, there
is an equal contribution from $K_0^{*}(1430)^0$ and $K_2^{*}(1430)^0$, which we cannot
separate with our selection.
We obtain $47 \pm 12$ events in the
0.8--1.0~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace window for
$K^{*}(892)^0\bar K^{*}(892)^0$ decay, and
$185 \pm 21$ events for decay to $K^{*}(892)^0 K^- \pi^+$
with $m(K\pi)$ in the 1.7--2.0~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace region.
We convert these to branching fractions using Eq.(\ref{jpsicalc}), and
divide by the known branching fractions of the $K^*$ states~\cite{PDG}.
The results are listed in Table~\ref{jpsitab}, which
are more precise than those in Ref.~\cite{PDG}.
For the 1.7--2.0~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace mass region we obtain
$\Gamma^{J/\psi}_{ee} B_{J/\psi\ensuremath{\rightarrow}\xspace K^{*}(892)^0 K^- \pi^+} =
(2.24 \pm 0.25 \pm 0.15)\ensuremath{\rm \,e\kern -0.08em V}$.
We study decays into $\phi\ensuremath{\pi^+\pi^-}\xspace$ and $\phi\ensuremath{\pi^0\pi^0}\xspace$ using the
mass distributions shown in Figs.~\ref{jpsi_phi2pi}(a),(b).
The open histograms are for events with $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ mass within the
$\phi$ bands of Figs.~\ref{phif0_sel}(c) and~\ref{phif0_sel2}(c).
The hatched histogram in Fig.~\ref{jpsi_phi2pi}(a) is from the
$\phi$ sidebands of Fig.~\ref{phif0_sel}(c), and represents the
dominant background in the $\phi\ensuremath{\pi^+\pi^-}\xspace$ mode.
The hatched histogram in Fig.~\ref{jpsi_phi2pi}(b) is from the
\ensuremath {\chi^2_{2K2\ensuremath{\pi^0}\xspace}}\xspace control region, and represents the dominant background in
the $\phi\ensuremath{\pi^0\pi^0}\xspace$ mode.
Subtracting these backgrounds, and subtracting a small remaining
background using $J/\psi$ or $\psi(2S)$ sideband events,
we find 181$\pm$19 $J/\psi \!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^+\pi^-}\xspace$ events,
45$\pm$9 $J/\psi \!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^0\pi^0}\xspace$ events,
and 19$\pm$6 $\psi(2S) \!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^+\pi^-}\xspace$ events.
We convert these to branching fractions and,
after correcting for the modes other than $\phi\ensuremath{\rightarrow}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$,
list them in Table~\ref{jpsitab}.
All are consistent with current PDG values,
of which the first two are dominated
by our previous measurement.
We do not observe any evidence for $Y(4260)$ decays to these modes, nor
do we see a $Y(4260)$ signal in any other mode studied here.
Figures ~\ref{jpsi_phif0}(a)(b) show the
corresponding mass distributions for $\phi f_0(980)$ events,
i.e., the subsets of the events in Figs.~\ref{jpsi_phi2pi}(a)
and~\ref{jpsi_phi2pi}(b) with a di-pion mass in the range 0.85--1.10~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace.
Signals at the $J/\psi$ mass are visible in both cases.
From Fig.~\ref{jpsi_phif0}(b) we estimate
$16 \pm 4$ $\phi f_0$ events in the $\ensuremath{\pi^0\pi^0}\xspace$ mode.
However, $\phi f_0(980)$ is not the dominant mode
contributing to $J/\psi \!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{\pi^+\pi^-}\xspace$ decay.
The open histogram of
Fig.~\ref{jpsi_phipipi} shows the $\ensuremath{\pi^+\pi^-}\xspace$ invariant mass
distribution for events in the $J/\psi$ peak of
Fig.~\ref{jpsi_phi2pi}(a)
($|m(\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace)-m(J/\psi)| <0.05$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace); events in the
$J/\psi$ sidebands ($0.05<|m(\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace)-m(J/\psi)| <0.1$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace)
are shown by the hatched histogram.
A two-peak structure is visible that is very similar to that
studied by the BES Collaboration~\cite{bes4k} and
observed in $D^+_s\ensuremath{\rightarrow}\xspace\ensuremath{\pi^+\pi^-}\xspace\pi^+$ decay~\cite{D+s}. In both cases
the $\ensuremath{\pi^+\pi^-}\xspace$ system is believed to couple to an $s\bar s$ system;
both $\ensuremath{\pi^+\pi^-}\xspace$ distributions exhibit a clear $f_0(980)$ peak and a broad
bump in the 1.3-1.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace region. The analysis of
Refs.~\cite{bes4k, D+s}
shows that this bump is made up of $f_2(1270)$ and $f_0(1370)$
contributions; we denote this region by $f_x$.
By selecting $f_0(980)$ in the 0.85--1.10~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace range and
$f_x$ in the 1.1--1.5~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace range, shown by vertical lines in
Fig.~\ref{jpsi_phipipi}, and subtracting $J/\psi$ sideband background
we find $57\pm 9$ $J/\psi \!\ensuremath{\rightarrow}\xspace\! \phi f_0(980)$ events and
$61\pm 10$ $J/\psi \!\ensuremath{\rightarrow}\xspace\! \phi f_x$ events.
Using Eq.(\ref{jpsicalc}) and dividing by the appropriate branching
fractions,
we obtain the $J/\psi$ branching fractions listed in Table~\ref{jpsitab}.
The measurements of ${\ensuremath{\cal B}\xspace}_{J/\psi\ensuremath{\rightarrow}\xspace\phi f_0}$ in the \ensuremath{\pi^+\pi^-}\xspace and \ensuremath{\pi^0\pi^0}\xspace
decay modes of the $f_0$ are consistent with each other and with the
PDG value, and combined they have roughly the same precision as given
in Ref.\cite{PDG}.
Note that, in contrast to $\phi(1680)\ensuremath{\rightarrow}\xspace\phi\pi\pi$ decay, there is
no indication of a $J/\psi\ensuremath{\rightarrow}\xspace\phi f_0(600)$ decay mode. Only
$J/\psi\ensuremath{\rightarrow}\xspace\phi f_0(980)$ is observed, as is true for the $Y(2175)$ state.
We also observe $12\pm 4$ $\psi(2S) \!\ensuremath{\rightarrow}\xspace\! \phi f_0$, $f_0 \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{\pi^+\pi^-}\xspace$
events, which we convert to the branching fraction listed in
Table~\ref{jpsitab};
it is consistent with the value in Ref.~\cite{PDG}, assuming
${\ensuremath{\cal B}\xspace}_{f_0\ensuremath{\rightarrow}\xspace\ensuremath{\pi^+\pi^-}\xspace} = 2/3$.
The hatched histogram in Fig.~\ref{mkk_notphi}(a)
shows the $K^+ K^-$ invariant mass distribution,
when the other kaon pair is in the $\phi$ region,
for the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace events in the $J/\psi$ peak, selected by requiring
$|m(\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace)-m(J/\psi)| < 0.05$~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace. Subtracting sideband events we
find 163 $\pm$ 19 events corresponding to
$J/\psi\ensuremath{\rightarrow}\xspace\phi K^+ K^-$ decay.
Using our normalization we
obtain the branching fraction listed in Table~\ref{jpsitab},
which agrees with that in Ref.~\cite{PDG} but
has better precision.
In obtaining these values, we
have used
$B(\phi\ensuremath{\rightarrow}\xspace K^+ K^-)=0.489$~\cite{PDG}, and assume equal
rates for $J/\psi\ensuremath{\rightarrow}\xspace\phi K^+ K^-$ and $J/\psi\ensuremath{\rightarrow}\xspace\phi K^0 \bar K^0$.
\section{Summary}
\label{sec:Summary}
\noindent
We use the excellent charged-particle tracking, track identification, and
photon detection of the \mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}\ detector to fully reconstruct events
of the type
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \gamma\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \gamma\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace$,
$\gamma\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace$, and
$\gamma\ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace$,
where the $\gamma$ is radiated from the initial state $e^+$ or $e^-$.
Such events are equivalent to direct \ensuremath{e^+e^-}\xspace annihilation at a
c.m.\@ energy corresponding to the mass of the hadronic system.
Consequently, we are able to use the full ~\mbox{\sl B\hspace{-0.4em} {\small\sl A}\hspace{-0.37em} \sl B\hspace{-0.4em} {\small\sl A\hspace{-0.02em}R}}~ dataset to
study annihilation into these three final states
from their respective production thresholds up to 5~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace c.m.~energy.
The \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace, \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace and \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace measurements are consistent with, and supersede,
our previous results~\cite{isr2k2pi}.
The systematic uncertainties on the
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace$, \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace and \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace cross section values are
4\%, 7\% and 9\%, respectively, for $\ensuremath {\rm E_{c.m.}}\xspace \! <\! 3$~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace,
and increase, respectively, to 11\%, 16\% and 13\% in the 3--5~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace range.
The values
obtained are considerably more precise than previous measurements, and
cover this low-energy range completely. As such they
provide useful input to calculations of the hadronic corrections to
the anomalous magnetic moment of the muon, and of the fine structure
constant at the $Z^0$ mass.
These final states exhibit complex resonant substructures.
For the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace final state we measure the cross sections for
the specific channels
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K^{*}(892)^0\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace$, $\phi\ensuremath{\pi^+\pi^-}\xspace$, and $\phi f_0$,
and, for the first time, for the $\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K_2^{*}(1430)^0\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace$ and
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\rho(770)^0\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ reactions.
We also observe signals for the $K_1(1270)$, $K_1(1400)$, and $f_2(1270)$ resonances.
It is difficult to disentangle these contributions to the final state,
and we make no attempt to do so in this paper.
We note that
the $\rho^0$ signal is consistent with being due entirely to $K_1$ decays,
and that while
the total cross section is dominated by the $K^{*}(892)^0\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace$ channels,
only about 1\% of the events correspond to the
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K^{*}(892)^0\kern 0.2em\overline{\kern -0.2em K}{}\xspace^{*}(892)^0$ reaction.
For the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace final state we measure the cross section for
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi f_0$,
and observe signals for the $K^{*}(892)^{\pm}$ and $K_2^{*}(1430)^{\pm}$
resonances.
Again, the total cross section is dominated by the
$K^{*}(892)^+ K^-\ensuremath{\pi^0}\xspace$ channel, but about 30\% of events
are produced in the
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! K^{*}(892)^+ K^{*}(892)^-$ reaction.
For the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace final state
we note that the cross section is roughly a factor of four smaller
than that for \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace over most of the \ensuremath {\rm E_{c.m.}}\xspace range,
consistent with a factor of two isospin
suppression of the \ensuremath{\pi^0\pi^0}\xspace final state and another factor of two for the
relative branching fractions of the neutral and charged $K^*$ to
charged kaons.
With the larger data sample of the present analysis,
we perform a more detailed study of
the $\ensuremath{e^+e^-}\xspace\ensuremath{\rightarrow}\xspace\phi(1020)\pi\pi$ reaction. The $\ensuremath{\pi^+\pi^-}\xspace$ and $\ensuremath{\pi^0\pi^0}\xspace$
invariant mass distributions
both show a clear $f_0(980)$ signal,
and a broad structure at lower mass interpreted as the
$f_0(600)$. We obtain parameter values for these resonances.
The $\phi\ensuremath{\pi^+\pi^-}\xspace$ cross section measured in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{\pi^+}\xspace\ensuremath{\pi^-}\xspace}\xspace final state shows
a structure around 1.7~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace and some structures above 2.0~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
The corresponding $\phi\ensuremath{\pi^0\pi^0}\xspace$ cross section in the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\pi^0\pi^0}\xspace
final state shows similar behavior.
If the $f_0(980)$ is excluded from
the di-pion mass distribution, no structures above 2.0~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace are seen. We fit
the observed cross section with the VMD model assuming $\phi(1680)\ensuremath{\rightarrow}\xspace\phi
f_0(600)$ and $\phi(1680)\ensuremath{\rightarrow}\xspace\phi f_0(980)$ decay; the latter appears
to be responsible
for the threshold increase of the cross section at 2.0~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
Confirming our previous study~\cite{isr2k2pi}, our data require
an additional resonance at 2.175~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, which we call the $Y(2175)$, with decay to
$\phi f_0(980)$, but not to $\phi f_0(600)$.
Further investigation reveals consistent results for the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace
final state,
and clear $Y(2175)$ signals in the $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace f_0(980)$ channels, with
$f_0(980) \!\!\ensuremath{\rightarrow}\xspace\! \ensuremath{\pi^+\pi^-}\xspace$ and \ensuremath{\pi^0\pi^0}\xspace.
This structure can be interpreted as a strange partner
(with $c$-quarks replaced by $s$-quarks)
of the $Y(4260)$~\cite{y4260}, which has the analogous decay mode
$J/\psi\ensuremath{\pi^+\pi^-}\xspace$,
or perhaps as an $\ensuremath{s\overline s}\xspace\ssbar$ state that decays predominantly to $\phi f_0$.
In the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace mode we find
$\ensuremath{e^+e^-}\xspace \!\!\ensuremath{\rightarrow}\xspace\! \phi\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ to be the dominant channel.
With the current data sample we can say little about other $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$
combinations. We observe an enhancement near threshold,
consistent with the $\phi f_0$ channel and if these events are
selected we have an indication of a $Y(2175)$ signal.
Two other structures in the $\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ invariant mass spectrum are seen:
the smaller could be an indication of the $\phi f_0(1370)$ final
state, and the larger
of the $\phi f_2'(1525)$ mode. If events corresponding to
the $\phi f_2'(1525)$ final state are selected, the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace cross section
shows a resonance-like structure around 2.7~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, and a strong $J/\psi$ signal,
which has been studied in detail by the BES Collaboration~\cite{bes4k}.
In the \ensuremath {\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace}\xspace cross section we observe a sharp peak at ~2.3~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, which
corresponds to the $\phi \ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ channel with the
$\ensuremath{K^+}\xspace\ensuremath{K^-}\xspace$ invariant mass in the 1.06--1.2~\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace
region.
We also investigate charmonium decays into the
studied final states and through corresponding intermediate channels,
and measure the product of the electron
width and the corresponding branching fraction.
Some of the obtained $J/\psi$ branching fractions
listed in Table~\ref{jpsitab} are as
precise as, or more precise than, the current world averages, many of which
were obtained in our previous study~\cite{isr2k2pi}; the latter are superseded
by our new results.
We do not observe the $Y(4260)$ in any of the final states examined.
\section*{ACKNOWLEGMENTS}
\label{sec:Acknowledgments}
\input pubboard/acknowledgements
\input{biblio_2k2pi_PRD}
\end{document}
|
\section{Introduction}
The edge between classical and quantum description of a phenomenon is
related to the interactions occurring between the system under
investigation and its environment. As a consequence, if we could, in
ideal conditions, avoid irreversible interactions among them we should
observe the emergence of quantum behavior even in macroscopic systems.
As a matter of fact, the technological developments of the recent years
have made it possible to start inquiring into the quantum limit even in
mesoscopic mechanical systems and experiments have been designed which
realize a solid state analogue of cavity quantum electrodynamics. Many
of these experiments focus on detecting the quantization of vibrational
modes in a mechanical oscillator
\cite{nmr1,nmr2,nmr3,nmr4,nmr5,nmr6,nmr7,nmr8,nmr9,nmr10,nmr11}.
Experimental conditions such that a mechanical object may behave in a
quantum fashion are achieved in the low temperature regime. For example,
for a single vibrational mode of energy $\hbar \omega$ to show quantum
features, as the quantization of lattice vibrations, temperatures $T \ll
\frac{\hbar \omega}{k_B}$ are required, which for a micro-sized object
oscillating in the microwave band correspond to few mK.
\par
In this framework it has become increasingly relevant to have a precise
determination of the temperature. However, for a quantum system in
equilibrium with a thermal bath, there is no linear operator that acts
as an observable for temperature. Temperature, thought as a macroscopic
manifestation of random energy exchanges between particles, still
retains its meaning but we have lost any operational definition. This
kind of impediment often occurs in physics, and especially in quantum
mechanics, whenever one is interested in quantities which are not directly
accessible, i.e. they do not correspond to observable quantities. This
may either be due to experimental impossibilities, or be a matter of
principle, as it happens for nonlinear functions of the density
operator. In both cases, it turns out that the only way to gain some
knowledge about the quantity of interest is to measure one or more
proper observables somehow related to the parameter we are interested
in, and upon suitably processing the outcomes, to come back and infer
its value. Hence, any conceivable strategy aimed to evaluate the
quantity of interest ultimately reduces to a parameter estimation
problem. Relevant examples of this situation are given by estimation of
the quantum phase of a harmonic oscillator \cite{Mon06,HDB09,asp09,PD11}, the
amount of entanglement of a bipartite quantum state
\cite{EE08,EE10,EEL11} and the coupling constants of different kinds of
interactions \cite{Sar06,Hot06,Mon07,Fuj01,Zhe06,Boi08,ZP07,Cam10,Pat06,
mon10,mon11}. Here we
focus on the estimation of temperature \cite{Man89} and, motivated by recent
experimental achievements \cite{nmr11}, we specifically refer to schemes
where a micromechanical resonator is coupled to a superconducting
qubit, and then a measurement of the excited state population
is performed on the qubit itself. From the statistics of the
population measurement is then possible to obtain information about
the oscillator state, e.g. infer how close it is to the ground
state, and in turn its temperature.
\par
In this context an optimization problem naturally arises, aimed at
finding the most efficient inference procedure leading to minimum
fluctuations in the temperature estimate. In this paper we address
this problem in the framework of local quantum estimation theory
(QET) \cite{lqe1,lqe2,lqe3,lqe4,lqe5,lqe6}.
We solve the dynamics of the qubit-resonator coupled system and, in order
to match realistic scenarios, we also take into account an effective
model for non dissipative decoherence. Then, we evaluate the Fisher
information (FI) for the estimation of temperature via population
measurement (hereafter referred to as the FI {\em of the population
measurement}) and find both the
optimal initial qubit preparation and the smallest temperature value
that can be discriminated. Moreover, we evaluate the Quantum Fisher
Information (QFI) in terms of the symmetric logarithmic derivative in order to
calculate the ultimate bound to precision allowed by quantum mechanics.
This enable us to show that population measurement is
indeed optimal for a suitable choice of the initial preparation of the
qubit, and to provide quantum benchmarks for temperature
estimation.
\par
It is worth noting at this point that we are not discussing here
temperature fluctuations in a thermodynamical setting. Although
temperature itself may not fluctuate, as it is suggested by quantum
thermodynamical approaches \cite{qth}, we expect that fluctuations
always appear in the temperature estimates coming from indirect measurements
\cite{web07,Jan11}. Quantum estimation theory provides the tools to evaluate lower bounds to
the amount of fluctuations for a given measurement, as well as the
ultimate bounds imposed by quantum mechanics.
\par
The paper is structured as follows. In Sec. \ref{s:model} we describe
the interaction model: first we briefly review the unitary Jaynes-Cummings
dynamics for the coupled system and describe the measurements performed
on the qubit, and then we take into account the decoherence effects. In
Sec. \ref{s:QET} we show how QET techniques applies to our system,
providing explicit formulas for both the FI and the QFI. The results
are finally shown in detail in Sec. \ref{s:results} both for the unitary
and the noisy dynamics. Sec. \ref{s:out} closes the paper with some
concluding remarks.
\section{The physical model}\label{s:model}
As the temperature decreases a mechanical oscillator starts to exhibit
its quantum nature, which mainly manifests itself in quantization of the
vibrational modes. Hence, for our purposes the resonator can be regarded
as a collection of phonons in a thermal equilibrium state.
We assume that the resonator is built as to display an isolated
mechanical mode at a given frequency, so that it can be modeled, rather
than a phonon bath with some spectral distribution, as a single mode
phonon field in thermal equilibrium.
\subsection{Unitary dynamics}
Let $\mathcal{H}_R$ be the infinite dimensional Hilbert space
associated with the single mode phonon field. Upon introducing
the creation and annihilation operators $[a,a^{\dagger}]=1$
one has the number operator $N=a^{\dagger}a$,
and its eigenstates $\left\{
\ket{n}\right\}_{n=0}^{\infty}$. The field Hamiltonian reads:
\begin{equation}
H_{\scriptscriptstyle F} = \hbar\, \Omega\, a^{\dagger}a \; ,
\end{equation}
where $\Omega$ denotes the frequency of the vibrational mode.
We assume the resonator in a thermal equilibrium state, i.e.
described by the density operator
\begin{eqnarray}\label{thermal}
\varrho_{F}&=&\frac{\exp(-\beta H_{\scriptscriptstyle F})}{\tr{\exp(-\beta H_{\scriptscriptstyle F})}}
= \sum_{n=0}^{\infty} p_n (\Omega, \beta)
\ketbra{n}{n} \; , \nonumber
\end{eqnarray}
where $\beta=(k_B T)^{-1}$ and:
\begin{equation}\label{pn}
p_n (\Omega, \beta) =
e^{-\beta \hbar \Omega n}\left(1- e^{-\beta \hbar \Omega}\right).
\end{equation}
The resonator is coupled to a superconducting qubit whose initial
preparation is under control and, after a given interaction time, the
excited state population is detected. The qubit is treated as a
normalized vector in a two-dimensional complex Hilbert space $\mathcal{H}_Q$, with
$\{\ket{e},\ket{g}\}$ providing an orthonormal basis. The qubit is
initially prepared in a pure state
\begin{equation}
\ket{\psi}=\cos \frac{\vartheta}{2} \ket{e}+e^{i\varphi}\sin \frac{\vartheta}{2}
\ket{g} \, ,
\end{equation}
with $\varphi \in \left[0,2\pi\right)$ and $\vartheta \in
\left[0,\pi\right]$. Hence the qubit density operator reduces to the
projector $\varrho_{\scriptscriptstyle Q}=\ketbra{\psi}{\psi}$. Being a two-level system, by
appropriately choosing the zero energy level and denoting by $\omega$ its
transition frequency, the qubit Hamiltonian can be written as
\begin{equation*}
H_q=\frac{\hbar \omega}{2} \sigma_z \; .
\end{equation*}
The qubit-resonator interaction is the interaction between a
single-mode bosonic field and a two-level system.
In the rotating-wave approximations and for the near-resonant
case, i.e., for small values of the detuning $\delta=\omega - \Omega$
we have the Jaynes-Cummings (JC) model with Hamiltonian
\begin{eqnarray}\label{JC}
\tilde{H}_{{\hbox{\small\sc jc}}}&=&H_q + H_{\scriptscriptstyle F} + H_{int} \nonumber \\
&=&\frac{\hbar \omega}{2} \sigma_z + \hbar \Omega a^{\dagger}a
+ \hbar \lambda\left(\sigma_+a + \sigma_-{a^{\dagger}}\right) \; .
\end{eqnarray}
The unperturbed Hamiltonian $\tilde{H}_{{\hbox{\small\sc jc}}}^{(0)}=H_q + H_{\scriptscriptstyle F} $ satisfies
the eigenvalues equations
$$\tilde{H}_{{\hbox{\small\sc jc}}}^{(0)}\ket{k,n}=\hbar\left[n\Omega + \frac12\, \omega\,
(-1)^k \right]\ket{k,n}\,,$$ with $k=e,g$ and with
the correspondences $0 \leftrightarrow e$, $1 \leftrightarrow g$.
In Eq. (\ref{JC}) $\lambda \in \mathbb{R}$ represents the coupling strength,
$\sigma_+ a$ and $\sigma_-{a^{\dagger}}$ stand respectively for the
operators $\sigma_+\otimes a$, $\sigma_-\otimes {a^{\dagger}}$ acting
on the tensor product space, where $\sigma_{\pm}$ are the qubit
ladder operators. Upon choosing a suitable rotating frame one rewrites
the Hamiltonian in interaction picture $H_{{\hbox{\small\sc jc}}}$:
\begin{equation}\label{JCint}
H_{{\hbox{\small\sc jc}}}=\frac{\hbar \delta \sigma_z}{2} +
\hbar \lambda\left(\sigma_+a + \sigma_-{a^{\dagger}}\right) \; .
\end{equation}
The interaction only couples, for a given $n$, the states $\ket{e,n}$
and $\ket{g,n+1}$, and thus it is possible to study the interaction
inside the two-dimensional manifold spanned by these states leading to
a representation -- the so called dressed states basis -- where
$H_{{\hbox{\small\sc jc}}}$ is diagonal. We further assume the absence of any initial
correlations between the qubit and the oscillator, thus choosing at
time $t=0$ the following factorized density operator
$$
\varrho(0)=\varrho_{\scriptscriptstyle Q} \otimes \varrho_{{\scriptscriptstyle F}} \; ,
$$
whose dynamical evolution with respect to the JC Hamiltonian is
given by:
$$
\varrho(t)=U(t) \varrho(0){U}^{\dagger}(t) \;,
$$
with $U(t)=\exp{\left(-\frac{i}{\hbar} H_{{\hbox{\small\sc jc}}}t\right)}$.
\par
Time evolution entangles the qubit and the resonator \cite{sch10} and the
probabilities for the qubit to be found in the ground or excited state
are obtained via the Born rule as
\begin{equation}\label{prob1}
p(j|\beta)=\Tr{{\scriptscriptstyle Q}\!{\scriptscriptstyle F}}{ \varrho(t)\ketbra{j}{j}\otimes {\mathbb I}_{\scriptscriptstyle F}}
\qquad j=e,g
\end{equation}
where $p(j|\beta)$ denotes the conditional probability of obtaining the
value $j$ when the value of the temperature parameter is $\beta$.
Upon introducing the following quantum operation:
\begin{equation}\label{probe1}
\varrho_{\scriptscriptstyle Q} \stackrel{\mathcal{E}}{\longmapsto}
\varrho_{\scriptscriptstyle P} \equiv \Tr{{\scriptscriptstyle F}}{U(t)\,\varrho_{\scriptscriptstyle Q}\otimes
\varrho_{\scriptscriptstyle F}\,{U}^{\dagger}(t)} \;,
\end{equation}
where $\mathcal{E} : \mathcal{L}(\mathcal{H}_{\scriptscriptstyle Q}) \rightarrow \mathcal{L}(\mathcal{H}_{\scriptscriptstyle Q})$,
Eq.~(\ref{prob1}) can be equally rewritten at the level of the qubit
subsystem alone, namely:
\begin{equation}\label{prob2}
p(j|\beta)=\Tr{{\scriptscriptstyle Q}}{\varrho_{{\scriptscriptstyle P}}\ketbra{j}{j}} \;.
\end{equation}
In the following we will refer to $\varrho_{\scriptscriptstyle P}$ as the \textit{probe
state}: It describes the qubit subsystem at time $t$, obtained as
the partial trace over the phonon field of the overall evolved state
of the coupled system. Since it is a density operator on
$\mathcal{H}_{Q}$ it can be arranged in a 2$\times$2 density matrix. We have
\begin{eqnarray}
\varrho_{\scriptscriptstyle P}&=&\sum_{n=0}^{\infty} p_n(\Omega,\beta)
\left(
\begin{array}{cc}
\varrho_{ee} & \varrho_{eg} \\ \varrho_{ge} & \varrho_{gg}
\end{array}
\right) \;, \nonumber
\end{eqnarray}
where:
\begin{subequations}\label{matrixelement} \label{rhs}
\begin{align}
\ \varrho_{ee}&= \cos^{2}\frac{\vartheta}{2}\left[\cos^2\theta_n t +
4\, \frac{\delta^2}{\theta_n^2}\sin^2\theta_n t\right] \nonumber
\\ &\hspace{0.5cm}+ \sin^{2}\frac{\vartheta}{2}
\frac{\lambda^2n}{\theta_{n-1}^2}
\sin^2\theta_{n-1}t,\\
\varrho_{eg}&= \frac12 e^{-i\varphi}\sin \vartheta
\left[ \cos\theta_{n-1}t +i\frac{2\delta}{\theta_{n-1}}\sin\theta_{n-1}t
\right] \nonumber \\
&\hspace{0.5cm}\times \left[ \cos\theta_n t -i\frac{2\delta}{\theta_n}
\sin\theta_n t \right], \\
\ \varrho_{ge}&= \varrho_{eg}^{\ast} \quad \hbox{and} \quad
\ \varrho_{gg}= 1- \varrho_{ee},
\end{align}
\end{subequations}
with: $$\theta_n\equiv \theta_n (\delta,\lambda)
=\frac12 \sqrt{\delta^2 +4\lambda^2\left( n+1\right)}\,.$$
\subsection{Effects of decoherence}
A purely Hamiltonian dynamics doesn't match realistic features. In
real-life scenarios quantum coherence is hard to achieve in
mechanical objects, and can be maintained for relatively small times
($\approx 10^{-9} $s ). Complete Rabi oscillations between the phonon
and the qubit excitation involve only the first Rabi half periods, then
a damping of the probabilities $p(j|\beta)$ to $\frac{1}{2}$ is
observed: the most striking signature of decoherence. Hence we include
in our model the treatment of non dissipative decoherence occurring
between the qubit and the resonator. Following Ref. \cite{atomtrap}
we consider an effective model provided by adding a power-law term in the thermal
distribution, which leads to probe state matrix elements given by:
$$ \tilde\varrho_{ij}=\sum_{n=0}^{\infty}p_n (\Omega,\beta)
\left[ e^{-\gamma_nt}
\varrho_{ij} + \frac12 \left(1-e^{-\gamma_nt}\right)\right]
$$
being $\varrho_{ij}$ the matrix elements of Eq. (\ref{rhs}), as
evaluated for the unitary case,
$i,j \in \{e,g\}$ and $$\gamma_n=b(1+n)^a\,.$$ More explicitly
\begin{subequations}\label{rhsd}
\begin{align}
\tilde\varrho_{ee}&=\frac{1}{2}
\left[1+\sum_{n=0}^{\infty} p_n(\Omega,\beta)
e^{-\beta e^{-\gamma_nt}}
\left(\varrho_{ee}-\varrho_{gg}\right)\right], \\
\tilde\varrho_{eg}&= \frac{1}{2}\sum_{n=0}^{\infty}
p_n (\Omega,\beta)e^{-\gamma_nt}\varrho_{eg},\\
\tilde\varrho_{ge}&=\tilde\varrho_{eg}^{\ast}\quad \hbox{and} \quad
\tilde\varrho_{gg}= 1- \tilde\varrho_{ee}\,.
\end{align}
\end{subequations}
One can see that the dynamical evolution now drives the qubit towards
the maximally mixed state, described by the density operator $\frac{\mathbb I}{2}$.
\section{Quantum thermometry}\label{s:QET}
In this section we apply the tools of (local) quantum estimation theory
(QET) to the coupled qubit-oscillator system. An
estimation problem always consists in two steps: at first one has to choose
a measurement and then, after collecting a sample of outcomes, one
should find an estimator, i.e. a function to process data and to infer
the value of the quantity of interest.
In our case, temperature, expressed as
$\beta$, is the unknown parameter which has to be estimated from the sample
of outcomes coming from measurements performed on the qubit. The results, a
string of zeroes and ones for the case of population measurement, are
distributed according to the probabilities $p(j|\beta)\equiv\varrho_{jj}$ of
Eqs. (\ref{prob2}) and (\ref{rhs}) [or Eq. (\ref{rhsd}) in presence
of decoherence].
The Cram\'er-Rao inequality establishes that the variance
Var$(\beta)$ of any unbiased estimator is lower bounded by
\begin{equation}\label{cramer-rao}
\mbox{Var}(\beta) \geq \frac{1}{M F(\beta)} \ ,
\end{equation}
where $M$ is the cardinality of the sample, i.e., the number of
measurements, and $F(\beta)$ the so-called Fisher information (FI):
\begin{eqnarray}\label{Fisher}
F(\beta)&=& \sum_{j=e,g} p(j|\beta)\left[\partial_{\beta}\ln
p(j|\beta)\right]^2 \nonumber \\ &=& \frac{{\left[ \partial_{\beta}
p(e|\beta)\right]}^2}{p(e|\beta)} + \frac{{\left[ \partial_{\beta}
p(g|\beta)\right]}^2}{p(g|\beta)} \; .
\end{eqnarray}
Efficient estimators are those saturating the Cram\'er-Rao inequality and
their existence depends on the statistical model. However, independently
of the statistical model we have that for sufficiently large samples, i.e.,
in the asymptotic regime $M\gg 1$, maximum likelihood estimators are
always efficient.
\par
Quantum mechanically, the probability of obtaining the outcome
$j\in\{e,g\}$ from a measurement is given according to the Born rule
by $p(j|\beta)=\tr{\varrho_{\scriptscriptstyle P} \Pi_j}$, where the probe state
$\varrho_{\scriptscriptstyle P}\equiv \varrho_{\scriptscriptstyle P} (\beta)$ parametrized by the unknown
quantity $\beta$ is referred to as the quantum statistical model, and
the collection of operators $\{\Pi_j\}$, $\Pi_j\geq 0$, $\sum_j\Pi_j=
\mathbb{I}$ is the probability operator-valued measure describing the
measurement taking place on the qubit. In our
case the qubit excited state population is probed and the measurement
reduces to a projective one, $\ketbra{e}{e}$ and
$\ketbra{g}{g}=\mathbb{I}-\ketbra{e}{e}$, i.e., we are measuring
the Pauli operator $\sigma_z=\ketbra{e}{e}-\ketbra{g}{g}$.
\par
Once the observable is fixed, we optimize the estimation
procedure by maximizing the FI over
the qubit state parameters, $\vartheta$ and $\varphi$,
as well as over the parameters driving the interaction -- i.e., the detuning
$\delta$ and the interaction time $t$. In other words, by
employing the optimal qubit preparation and tuning the interaction
parameters one may find a working regime achieving the maximum
precision for that kind of measurement.
\par
On the other hand, one may also maximize the FI over all possible
quantum measurements. Upon defining the symmetric logarithmic
derivative (SLD) $L_{\beta}$ as the selfadjoint operator satisfying
the equation
\begin{equation}
\frac{L_{\beta}\varrho_{\scriptscriptstyle P} +\varrho_{\scriptscriptstyle P} L_{\beta}}{2}=
\partial_{\beta}\varrho_{\scriptscriptstyle P} \; ,
\end{equation}
it is possible to show that the Fisher information $F(\beta)$ of any quantum
measurement is upper bounded by the following quantity:
\begin{equation}\label{Ineq}
F(\beta)\leq G(\beta)\equiv \tr{\varrho_{\scriptscriptstyle P} L_{\beta}^2} \; ,
\end{equation}
which is called quantum Fisher information (QFI). QFI does not depend on
the measurement carried on the qubit---indeed being obtained by maximizing
over the possible measurement. It is rather an attribute of the family
of states $\varrho_{\scriptscriptstyle P}(\beta)$ parametrized by the temperature. Looking
back to the Cram\'er-Rao inequality Eq.(\ref{cramer-rao}) one sees that
QFI allows one to write its natural quantum version
\begin{align}
\mbox{Var}(\beta) \geq \frac1{M G(\beta)}\,.\label{QCR}
\end{align}
The above equation represents the Quantum Cram\'er-Rao bound (QCR),
i.e. the ultimate bound to the precision allowed by quantum mechanics
for a given statistical model $\varrho_{\scriptscriptstyle P}(\beta)$.
An optimal measurement, i.e. a measurement whose FI $F(\beta)=G(\beta)$
equals the QFI for the parameter $\beta$, is given by the observable
corresponding to the spectral measure of the SLD $L_\beta$. On the other
hand, other kind of measurements may achieve optimality for the whole range
of values of $\beta$ or for a subset of values. Indeed, we will see in
the following that population measurement is optimal for a suitable
choice of the initial qubit preparation. We remind that for the estimation
of a single parameter, as it is in our case, the QCR may be always attained,
and an estimator saturating Ineq. (\ref{QCR}) is called efficient.
The existence of an efficient estimator depends on the statistical
model. However, independently of the statistical model, for sufficiently
large samples, i.e., in the asymptotic regime $M\gg 1$, maximum likelihood
and Bayesian estimators are always efficient.
\par
Upon diagonalizing the probe state one achieves the decomposition
$\varrho_{\scriptscriptstyle P}=\varrho_+\ketbra{\psi_+}{\psi_+}\,
+\, \varrho_-\ketbra{\psi_-}{\psi_-} $ and is able to solve the equation for SLD
\begin{align}
L_{\beta}=&\,\frac{\bra{\psi_+}\partial_{\beta}
\varrho_{\scriptscriptstyle P}\ket{\psi_+}}{\varrho_+}\ketbra{\psi_+}{\psi_+}
\nonumber \\ & +\,
\frac{\bra{\psi_-}\partial_{\beta}\varrho_{\scriptscriptstyle P}
\ket{\psi_-}}{\varrho_-}\ketbra{\psi_-}{\psi_-}
\nonumber \\ & +\, \frac{2}{\varrho_+ +
\varrho_-}\left[\bra{\psi_+}\partial_{\beta}
\varrho_{\scriptscriptstyle P}\ket{\psi_-}\ketbra{\psi_+}{\psi_-}
\right. \nonumber \\ & +\, \left.
\bra{\psi_-}\partial_{\beta}\varrho_{\scriptscriptstyle P}
\ket{\psi_+}\ketbra{\psi_-}{\psi_+}
\right], \end{align}
finally obtaining an explicit formula for the QFI
\begin{align}\label{qfi}
G(\beta)=&\,
\frac{\left(\partial_{\beta}\varrho_{+}\right)^2}{\varrho_{+}} +
\frac{\left(\partial_{\beta}\varrho_{-}\right)^2}{\varrho_{-}}
\nonumber \\
&+\, 2\kappa\,\left[ \left| \langle
\psi_-|\partial_{\beta}\psi_{+}\rangle \right|^2 + \left| \langle
\psi_+|\partial_{\beta}\psi_{-}\rangle \right|^2 \right]
\end{align}
where
$$|\partial_{\beta}\psi_{\pm}\rangle =
\partial_{\beta}\langle e|\psi_{\pm}\rangle\,|e\rangle
+ \partial_{\beta}\langle g|\psi_{\pm}\rangle\,|g\rangle\,,$$
and
$$\kappa=\frac{\left(\varrho_+ - \varrho_-
\right)^2}{\varrho_+ + \varrho_-}=(1-2\varrho_+)^2\,.$$
Eq. (\ref{qfi}) contains a first term which
resembles the FI and a second one, truly quantum in nature, which leads
to the QCR and vanishes whenever $\ket{\psi_{\pm}}$ does not depend on
$\beta$.
\section{Dynamics of the Fisher information and optimal
working regimes}\label{s:results}
In this section we report results for the qubit-resonator coupled
system with physical parameters chosen in a range matching the
experimental setup of Ref. \cite{nmr11}. More specifically, we
present a systematic study of the FI for population measurement as a
function of the state and interaction parameters, carrying out
numerical maximization and finding the optimal working
regimes. We also evaluate the QFI of the family of states
$\varrho_{P}(\beta)$ and find the ultimate bound to
precision, i.e. a benchmark
in order to assess the performances of qubit thermometry via population measurement.
\par
Hereafter we work with dimensionless quantities by rescaling times and
frequencies in units of the coupling $\lambda$. We thus substitute
time, detuning and decoherence parameters by their rescaled counterparts
\begin{align}
t \longmapsto \tau \equiv \lambda t \notag, \quad
\delta \longmapsto \gamma \equiv \delta/\lambda \notag, \quad
b \longmapsto \tilde{b}\equiv b/\lambda\,. \notag
\end{align}
Effective detuning $\gamma$ will range in $|\gamma| \in [0, 1.5]$. Also
a dimensionless effective temperature $\tilde{\beta}$ is defined,
provided by the substitution
$$\beta \longmapsto \tilde{\beta}\equiv\beta\hbar\Omega\,.$$
For convenience, we continue to term
$\tilde{\beta}$ and $\tilde{b}$ respectively $\beta$ and $b$.
\subsection{Resonant Hamiltonian regime}
Upon using the expression of the diagonal matrix elements in Eqs.
(\ref{rhs}) we have evaluated the FI of Eq. (\ref{Fisher}).
We start the discussion by considering the resonant case, i.e
zero detuning, and analyze the effect of detuning afterward in this
Section. For convenience we
adopt the notation $F(\beta)$ for the FI, but keep in mind the
complete dependence $F(\beta;\vartheta,\tau,\gamma)$ on both the qubit
degrees of freedom and the parameters $\gamma$ and $\tau$ which drive
the coupling. Notice that $F(\beta)$ does not depend on the qubit phase $\varphi$: its
building-blocks are in fact the probabilities $p(e|\beta)$ and
$p(g|\beta)$, whereas $\varphi$ only appears in off-diagonal matrix
elements. Varying the parameter $\vartheta$ from $\pi$ to $0$ we span
the entire class of qubit preparation, starting from $\ket{1}$, going
trough a superposition and ending in $\ket{0}$.
\begin{figure}[h!]
\includegraphics[width=0.85\columnwidth]{f1a_Ftime.eps}
\includegraphics[width=0.85\columnwidth]{f1b_Ftheta.eps}
\caption{(Color online) Upper panel: FI for $\beta=10$ as a function of
the effective time $\tau$, for different $\vartheta$ values:
$\vartheta=\pi$ (dashed blue), $\vartheta=0.95\, \pi$ (dot-dashed magenta) and
$\vartheta=0$ (solid green). FI takes a pronounced global maximum at
$(\vartheta,\tau)=\left(\pi,\frac{\pi}{2}\right)$ while it
is possible to see a secondary extremely peaked maximum, which
occurs for $\tau=\pi$ and preparing the qubit in $\ket{0}$. Lower
panel: log-linear plot of the FI for $\beta=10$ as a function of $\vartheta$
for, $\tau=\frac{\pi}{2}$ (dashed blue),
$\tau=\frac{\pi}{2}+\varepsilon$ (dot-dashed magenta),
$\tau=\pi$ (solid green),
$\tau=\pi+\varepsilon$ (dotted red), with $\varepsilon = 0.01$.
\label{f:fish1}}
\end{figure}
\par
Let us now consider the system at a fixed value of the temperature,
e.g. where the resonator is supposed to be very close to the
ground state, say $\beta=10$. The probabilities
$p(j|\beta)=\varrho_{jj}$ evolve periodically in time
according to Eq. (\ref{rhs}), as the coupled system undergoes Rabi
oscillations. The corresponding behavior of the
FI is shown in the upper panel of Fig.~\ref{f:fish1}. The FI displays
a robust maximum at the optimal time $\tau_{\rm max}=\frac{\pi}{2}$ for
$\vartheta=\pi$, corresponding to prepare the qubit in its ground state.
This maximum is, at the same time, the global and the smoothest one. In
fact, as soon as $\vartheta$ is moved from $\pi$ the FI
suddenly drops to zero, except for a sharp peak centered in
$\tau_{\rm max}$, monotonically decreasing with respect to $\vartheta$, as
shown in the lower panel of Fig.~\ref{f:fish1}. Another maximum of
the same order of the global one can be found at $(\vartheta,\tau)=
\left(0, \pi \right)$ but
it is extremely peaked, thus representing a bad (unstable) choice for
a possible measurement. Upon inspecting the temporal evolution of the
excited state probability we found that $p(e|\beta)$ has a minimum at
$\tau=\tau_{\rm max}$, a fact which gives us a physical insight on the FI
behavior: since our goal is the estimation of a vanishing quantity which
carries information about thermal disorder, we expect to find the maximum
sensitivity in our predictions where the excitation is most likely
stored -- as a phonon -- in the resonator, i.e., when $p(e|\beta)$ is
minimum.
\begin{figure}[h!]
\includegraphics[width=0.85\columnwidth]{f2_Fbeta.eps}
\caption{(Color online) Log-linear plot of the FI as a function of
effective time $\tau$ for different values of $\beta$. The qubit
is prepared in the ground state $\ket{1}$ ($\vartheta=\pi$). From bottom to top
$\beta=15$ (solid blue), $\beta=10$ (dashed magenta), $\beta=5$ (dot-dashed green),
$\beta=1$ (dotted red). Upon raising the temperature the FI no longer keeps
a scale-free shape: thermal excitations modifies its profile making
it irregular. In particular the global maximum comes earlier in
time.
\label{f:fish3}} \end{figure}
\par
Let us now turn our attention to the dependence of the FI
on the temperature itself. In Fig.~\ref{f:fish3} we show, on a logarithmic
scale, the temporal evolution of the FI for different values of $\beta$.
FI varies over several orders of magnitude, matching our intuition
that the closer we are to the ground state, the harder is to achieve a
given precision in estimation of temperature. Furthermore, upon
lowering the temperature, the temporal evolution of $p(j|\beta)$ becomes
less involved, finally approaching the exactly periodic one of Rabi
oscillations, which in turn freezes the profile of the FI in a
shape independent on the temperature itself.
\par
The qubit preparation $\theta=\pi$ is universally optimal, i.e., it
leads to a maximum of the FI independently of the interaction time.
After fixing $\theta=\pi$ we have numerically maximized $F(\beta)$
with respect to $\tau$. The solid blue line of the upper panel of
Fig.~\ref{f:max} is the the log-plot of
$$F_M (\beta) = \max_\tau F(\beta)\,,$$
as a
function of $\beta$, from which it is apparent the exponential decrease
of the maximum value
achieved by the FI for increasing $\beta$. The Cram\'er-Rao inequality
immediately relates this fact to an exponential loss of sensitivity
moving towards the quantum ground state of the resonator. An other
interesting feature that emerges from the maximization is a shift in
the value of the optimal interaction time. In the lower panel of
Fig.~\ref{f:max} we can recognize the existence of a steady value for
the optimal time $\tau_{\rm max}=\frac{\pi}{2}$ when approaching the
ground state, while for smaller values of $\beta$ the optimal time
comes earlier. In fact, the temporal evolution of FI (see Fig.
\ref{f:fish3}) not only predicts
an exponential increase of the global maximum when temperatures are
raised, but also a shift of its location.
\begin{figure}[h!]
\centering
\includegraphics[width=0.91\columnwidth]{f3a_maxF.eps}
\includegraphics[width=0.854\columnwidth]{f3b_taumax.eps}
\caption{(Color online) Upper panel: log-log plot of the FI maximized
over $\tau$ as a function of $\beta$, with $\theta=\pi$ for different
values of detuning: $\gamma=0$ (solid blue), $\gamma=1$ (dashed
magenta), $\gamma=1.5$ (dot-dashed green). Bottom panel: the times
$\tau_{\rm max}$ which maximizes the FI as a function of $\beta$, with
$\theta=\pi$ for different values of $\gamma$ (same values and colors
of the upper panel).\label{f:max}}
\end{figure}
\par
\subsection{Effects of detuning}
In this section we take into account the possible existence
of a nonzero detuning $\gamma$ between the oscillator and the
qubit frequencies. This has two main consequences, which are both
illustrated in Fig.~\ref{f:max}. On the one hand, the maximum achievable
value of the FI slightly decreases and, on the other hand, the optimal
interaction time $\tau_{\rm max}$ at which the maximum takes
place anticipates. Therefore, the best working conditions to achieve
the optimal sensitivity in the estimation of $\beta$ correspond to
have the qubit and the resonator in resonance. It is also
worth to notice that $\gamma$ does not represent a critical parameter,
as the initial preparation of the qubit,
since the FI dependence on $\gamma$ is smooth. One can see this in the
upper panel of Fig.~\ref{f:max}, where we see that curves corresponding
to quite different values of the detuning are almost superposed.
\subsection{Quantum Fisher information}
In order to assess the performances of the population measurement
in the estimation of temperature we have evaluated the QFI of the family
$\varrho_{\scriptscriptstyle P}(\beta)$. The diagonalization of the probe state has to be
carried out numerically, hence in general analytical expressions of the
QFI are not available. A first fact is that $G(\beta)$ turns
out to be independent on the qubit phase $\varphi$, which then does not
represent an extra degree of freedom whereby gain more restrictive
bounds to precision on $\mbox{Var}(\beta)$. Even the optimal qubit
preparation for to the best conceivable measurement involves
control of the parameter $\vartheta$ only.
\par
As we have done for the FI, we start to inspect the QFI behavior for a fixed
value of temperature $\beta$ in the resonant case. Also for the QFI
the maximum is achieved by preparing the qubit in the state $\ket{g}$
and probing it at time $\tau_{\rm max}$. In this case the behavior of
$G(\beta)$ is identical to that of $F(\beta)$, as it is apparent by
comparing Figs. \ref{f:fish1} and \ref{f:qfi}. In other words, for a
given value of the parameter $\beta$ into the range explored, the
choice $(\vartheta,\tau)=(\pi,\tau_{\rm max})$ makes population
measurement optimal. Moreover, the QFI itself reaches its global
maximum for that choice. Thus, provided that an optimal estimator is
employed, e.g. maximum likelihood in the asymptotic regime, this
strategy provides optimality in sense that either inequality
(\ref{Ineq}) is saturated and the right-hand side of QCR is as low
as possible.
\par
This conclusion is confirmed upon a closer inspection of the probe state. When
$\vartheta=\pi$ the off-diagonal terms vanish and $\varrho_{\scriptscriptstyle P}$ is
diagonal, with eigenvalues
\begin{subequations}
\begin{align}
\varrho_+&= \sum_{n=0}^{\infty} p_n(\Omega,\beta)
\sin^2\left[\sqrt{\gamma^2+4n}\,\frac{\tau}{2}\right] \frac{n}{n+\gamma^2/4} \\
\varrho_-&= 1-\varrho_+
\end{align}
\end{subequations}
As a consequence, the QFI reduces to
$$G(\beta;\pi,\tau,\gamma)=\frac{\left(\partial_{\beta}
\varrho_{+}\right)^2}{\varrho_{+}} +
\frac{\left(\partial_{\beta}\varrho_{-}\right)^2}{\varrho_{-}}\,,$$
which coincides with the FI ruling the estimation of
$\beta$ via population measurement. \par
On the other hand, some striking difference emerges between the
performances of population measurement and that of the optimal one if
the qubit is not prepared in the optimal (ground) state.
\begin{figure}[h!]
\centering
\includegraphics[width=0.85\columnwidth]{f4a_QFtime.eps}
\includegraphics[width=0.85\columnwidth]{f4b_QFtheta.eps}
\caption{(Color online) Upper panel : QFI for
$\beta=10$ as a function of $\tau$, for $\vartheta=\pi$ (dashed blue),
$\vartheta=0.95\, \pi$ (dot-dashed magenta) and $\vartheta=0$ (solid green). QFI
behaves like FI for $\vartheta=\pi$ leading to the same maximum,
while for smaller angles it shows a smoother profile. For angles
$0<\vartheta<\pi$ one may find measurements which improve the
precision of temperature estimation.
Bottom
panel: QFI for $\beta=10$ as a function of $\vartheta$
for
$\tau=\frac{\pi}{2}$ (dashed blue),
$\tau=\frac{\pi}{2}+\varepsilon$ (dot-dashed magenta),
$\tau=\pi$ (solid green),
$\tau=\pi+\varepsilon$ (dotted red), with $\varepsilon = 0.01$. \label{f:qfi}}
\end{figure}
\par
In the lower panel of Fig.~\ref{f:qfi} we show $G(\beta)$ as a function of $\tau$ for
different values of $\vartheta$: for $\vartheta<\pi$ the decrease of
$G$ is definitely smoother than that of $F$ and thus, in principle, some
measurement may be found making the initial preparation a less critical
parameter. Moreover inspecting the cut of the QFI along
$\tau=\pi$ we note that the maximum in $\vartheta=0$ becomes more
achievable compared to the one of $F(\beta)$. All these features
suggest that for qubit preparations different from the ground state
there will
be a sensible difference between the precision provided by population
measurement and the optimal one implementable on the system.
On the other hand, being the overall maximum achievable with population
measurement, our results indicate that the
achievement of the ultimate bound to precision allowed by quantum
mechanics is in the capabilities of the current technology.
\subsection{Effects of decoherence}
In this section we discuss the solution of the reduced qubit
dynamics in the presence of dissipative decoherence, see Eq. (\ref{rhsd}),
and inspect the corresponding behavior of the FI. For the sake of
simplicity we consider zero detuning. Analogue results are obtained when
including the detuning.
\par
The probabilities $p(j|\beta)=\tilde\varrho_{jj}$ are
damped so that, waiting for a sufficient long time, whose value
depends on $a$ and $b$, we would find them to be identically $1/2$ or,
equally stated, the dynamical evolution brings the state to the maximally
mixed one. The contribution of decoherence is of the kind
exp$\left[-b(1+n)^a\tau \right]$ for every $n$, where $b$ has been
rescaled in coupling units $b\longmapsto b/\lambda$. Being a
multiplicative coefficient, as soon as $b$ is different from zero, the
exponential term will participate in killing the sums. Our calculations
show a relevant dependence of the FI on the parameter $b$, namely
values $b\approx 10^{-5}$ are sufficient to produce visible effects,
while varying $a$ in the range $(0,1)$ does not
deeply influence of FI behavior.
\par
In Fig.~\ref{f:fishdec1} we show the temporal evolution of the FI for
$\beta=10$, in the presence of decoherence and for different initial
preparations of the qubit. In the Hamiltonian regime
for large $\beta$ the resonator is close to the ground state,
the evolution of $p(j|\beta)$ is periodic and hence, due to
Eq. (\ref{Fisher}), the same is true for the FI. Upon incorporating
decoherence we see that FI decays at a rate depending on $b$ and thus
an irreversible dynamics emerges, which matches the physical evidence
of a limited coherence time. On the other hand, a clear maximum at
$\tau=\pi/2$ still appears, with a slightly decreased value of
$F(\beta)$. In the lower panel of Fig.~\ref{f:fishdec1} we show the
maximum value $F_M = \max_\tau F(\beta)$ for different values of the decoherence
parameter. As it
is apparent from the (log-log) plot for high temperature (smaller $\beta$)
the effect of decoherence is negligible, whereas for increasing $\beta$
the effect is becoming more and more relevant.
\begin{figure}[h!]
\includegraphics[width=0.85\columnwidth]{f5a_dectime.eps}
\includegraphics[width=0.9\columnwidth]{f5b_decmax.eps}
\caption{(Color online) Upper panel: Fisher information $F(\beta)$ for $\beta=10$
as a function of $\tau$ in the presence of decoherence and
for different qubit preparations. The decoherence parameters
are chosen as to $a=0.1$ and $b=10^{-5}$. Dashed blue line stands for $\vartheta=\pi$,
dot-dashed magenta for $\vartheta=0.95\,\pi$ while solid green ones for $\vartheta=0$.
Having included decoherence treatment enables us not to restrict
the evolution to the first Rabi half-period.
Lower panel: log-log plot of the Fisher information $F_M (\beta)$ maximized over the
interaction time, and in the presence of decoherence, as a function
of $\beta$ and for fixed $\vartheta=\pi$, for
$b=0$ (solid green), $b=10^{-5}$ (dashed magenta), $b=10^{-4}$ (dot-dashed blue).\label{f:fishdec1}}
\end{figure}
\par
\section{Conclusions}
\label{s:out}
The temperature of a physical object cannot be directly measurable. On
the other hand is can be regarded as a parameter whose value can be
indirectly inferred by measuring some proper observable and then
suitably processing the outcomes, an inference procedure usually
referred to as an estimation procedure. In the case of a micromechanical
oscillator with an isolated vibrational mode, effective schemes have
been suggested and realized \cite{nmr11} which rely on coupling the
resonator to a superconducting qubit and probing the latter using
population measurements. In other words, the qubit is employed as a
quantum thermometer to demonstrate that the resonator has been cooled to
its quantum ground state. In this paper we have analyzed in details
qubit thermometry in these systems, i.e., the estimation of temperature
via quantum limited measurements performed on the qubit. In the
framework of quantum estimation theory we have analyzed precision as a
function of both the qubit initial preparation and the interaction
parameters, and we have evaluated the limits to precision posed by
quantum mechanics to qubit thermometry.
\par
We have computed the FI for population measurement, which is the
appropriate figure of merit to assess the precision of estimation,
and have found that its maximum, and hence the minimum variance in
the estimated temperature, is achieved by preparing the qubit in
the ground state, and probing it at an emergent time $\tau_{\rm max}$,
which is predictable. Furthermore, we have analyzed in details how
the maximum depends on the temperature itself, on the detuning, and on the
noise parameter when one takes into account non dissipative decoherence.
In order to evaluate the ultimate bound allowed by quantum mechanics
to the sensitivity of temperature estimation, we have also computed the
quantum Fisher information. We found that QFI is maximized for the same choice of qubit
preparation and measurement time of the FI, and that for these common
values the maxima of FI and QFI coincide. We thus conclude that
population measurement is optimal for temperature estimation.
\par
The range of parameters addressed in our analysis is that
of recent experimental implementations \cite{nmr11}. We thus conclude
that optimal estimation of temperature can be done with
current technology. Since the FI of population measurement,
and the QFI of the model, both decrease with the decrease of
temperature, the estimation of lower temperature will be
intrinsically less precise. On the other hand, since the are
regimes, also in the presence of decoherence, where the maxima
of the FI and the QFI are reasonably smooth as a function of
the qubit preparation and of the interaction time we do
not expect any "no-go" theorem for temperature estimation.
In other words, we expect that optimal estimation of lower
resonator temperatures, perhaps achievable with further experimental
advances, will be still possible with population measurements. On the
other hand, ``optimality'' will correspond to an inherently less precise
procedure compared to the case of higher temperature.
\par
Our analysis shows the optimality of feasible qubit thermometry in
providing quantum benchmarks for high precision temperature measurement,
as well as an efficient operational quantification of temperature for
mechanical modes lying arbitrary close to their ground state. In other
words, achievement of the ultimate bound to precision allowed by quantum
mechanics is in the capabilities of the current technology. Our results
also confirm that QET is a useful tool for assessing and comparing
inference procedures arising in quantum limited measurements \cite{sta10},
even when mesoscopic objects are involved.
\section*{Acknowledgments}
This work has been partially supported the CNR-CNISM agreement.
|
\section{Introduction}
In this paper we present a general topos-theoretic interpretation of `Stone-type dualities'; by this term we refer, following the standard terminology, to a class of dualities or equivalences between categories of preordered structures and categories of posets, locales or topological spaces, a class which notably includes the classical Stone duality for Boolean algebras (or, more generally, for distributive lattices), the duality between spatial frames and sober spaces, the equivalence between preorders and Alexandrov spaces, the Lindenbaum-Tarski duality between sets and complete atomic Boolean algebras, and the Birkhoff's duality between finite distributive lattices and finite posets.
We introduce an abstract framework in which all of these dualities are interpreted as instances of just one topos-theoretic phenomenon, and in which several new dualities are introduced. In fact, the known dualities, as well as the new ones, all arise from the application of one `general machinery for generating dualities' to specific `sets of inputs' which vary from case to case.
In section \ref{locales} we show that, under relatively mild hypotheses, one can naturally identify, through an isomorphism of categories, the opposite of a given category of ordered structures with a subcategory of the category of locales, in general in more than one way; in fact, this result (Theorem \ref{teoabstr}) provides us with an infinite number of dualities between categories of posets and subcategories of the category of locales. Anyway, these subcategories are not in general closed under arbitrary isomorphisms of locales, and in fact, in order to obtain `intrinsic' dualities between categories of preorders and categories of locales whose objects (resp. arrows) can be characterized as the locales (resp. locale homomorphisms) which satisfy some locale-theoretic invariant, we have to enlarge the target subcategory of locales to include all the isomorphic copies of the objects and arrows in it, and to look for a functor defined on this enlarged subcategory which is inverse (up to isomorphism) to the functor from posets to locales forming one half of the original isomorphism of categories. This can be done, under some natural assumptions, by functorially transferring topos-theoretic invariants across two different sites of definition of the same topos, according to the method `toposes as bridges' introduced in \cite{OC10}. For instance, the dualities between a given category $\cal K$ of preorders and a category of locales arise from the process of assigning to each structure $\cal C$ of $\cal K$, equipped with a subcanonical Grothendieck topology $J_{\cal C}$ in such a way that the morphisms in the category $\cal K$ induce morphisms of the associated sites, the locale $Id_{J}({\cal C})$ of $J$-ideals on $\cal C$, and from the inverse process of functorially recovering $\cal C$ from the locale $Id_{J}({\cal C})$ (equivalently, from the topos $\Sh({\cal C}, J) \simeq \Sh(Id_{J}({\cal C}))$) through a topos-theoretic invariant. The covariant equivalences with categories of locales are established in a similar way; the structures need not be equipped with any Grothendieck topology, and one relies on the well-known possibility of assigning a geometric morphism $[{\cal C}, \Set]\to [{\cal D}, \Set]$ to a given functor ${\cal C} \to {\cal D}$ in a canonical way.
In section \ref{dualtop}, we give a general methodology for `enriching' a given duality (resp. equivalence) between a category $\cal K$ of preorders and a category of locales to a duality (resp. equivalence) between $\cal K$ and a category of topological spaces; this methodology relies on an appropriate choice of points of the toposes corresponding to the structures.
In the following sections of the paper, we investigate further consequences of the topos-theoretic perspective introduced in the previous sections, again in light of the method `toposes as bridges' of \cite{OC10}. For example, we apply this method to establish various results connecting properties of preordered structures with properties of the corresponding locales or topological spaces, and we obtain a number of adjunctions between categories of these kinds.
The theory developed in the present paper provides a unified perspective on the subject of Stone-type dualities, in that the well-known dualities are easily recovered as applications of it. Anyway, what we consider to be the main interest of our topos-theoretic machinery is, apart from the conceptual enlightenment that it brings into the world of classical dualities, its inherent technical flexibility. In fact, one can generate infinitely many new dualities by applying it; examples are provided in the paper to illustrate how to do this in practice, and the reader will be able to use the method to generate his or her favorite applications.
The different `ingredients' that our `machinery' for generating dualities with categories of locales or topological spaces takes as `inputs' are: the initial category $\cal K$ of preordered structures, the subcanonical Grothendieck topologies $J_{\cal C}$ on the structures $\cal C$ in $\cal K$, the topos-theoretic invariant enabling one to recover a structure $\cal C$ from the topos $\Sh({\cal C}, J_{\cal C})$ and, if a duality with topological spaces is to be generated, appropriate sets of points of the toposes $\Sh({\cal C}, J_{\cal C})$ (and functions between them). In fact, the more general approach of section \ref{generalization} provides us with an additional degree of freedom in the choice of ingredients. Given such ingredients, dualities are generated in an automatic and `uniform' way by the `machine', as different concrete instances of a unique abstract pattern; in this way, the problem of building dualities gets reduced in many important cases to the much easier one of choosing appropriate sets of ingredients for this `machine'.
In connection with the perspectives outlined in \cite{OC10}, we remark that in this paper we have just `brought to the surface' a limited number of results which can be established by means of the methods of \cite{OC10}, carefully selected by virtue of their `representativeness' in illustrating the nature and variety of the insights obtainable by applying our techniques. In fact, the paper contains many general ideas which can be applied in the context of arbitrary, rather than just preordered, categories; for example, the method of section \ref{Mordual} of building dualities or equivalences starting from Morita-equivalences (by equipping structures, regarded as categories, with subcanonical Grothendieck topologies in such a way that the `structure-preserving' maps between the structures yield morphisms of the associated sites, and recovering each of the structures functorially from the corresponding topos through a topos-theoretic invariant), is potentially applicable beyond the preordered context that we have addressed in the present paper (cf. section \ref{conclusions} for a further discussion of this point). Even in the preordered context that we have addressed in the present paper, much remains to be discovered, in the form of new dualities, representation theorems, adjunctions and characterization theorems arising from translations of properties from one side to another of a given duality or equivalence. Anyway, the careful reader will realize that much of all of this can be easily uncovered in a semi-automatic way, by using similar means to those that we have adopted in the paper to generate our examples.
\subsection{An overview of the paper}
The contents of the present work can be described more in detail as follows.
In section \ref{subterminaltop}, we introduce a general method for building topological spaces from toposes equipped with a set of points. Specifically, we show that, given any subframe $\Gamma$ of the frame of subterminals of a locally small cocomplete topos $\cal E$ and any set of points of $\cal E$ indexed by a set $I$, we can naturally define a topology on the set $I$, which we call the ($\Gamma$-)subterminal topology, and that this construction can be naturally made functorial. The interest of this notion lies in its level of generality, which encompasses that of classical topology (every topological space arises from this construction in a canonical way), as well as in its formulation as a topos-theoretic invariant admitting a `natural behaviour' with respect to sites. Indeed, as shown in section \ref{exsub}, this notion allows us to recover, with natural choices of sites of definition and of sets of points of toposes, many interesting topological spaces considered in the literature, leaving at the same time enough freedom to construct new ones with particular properties. In section \ref{Sobriety} we give a general criterion for deciding the sobriety of topological spaces built in this way.
In section \ref{general}, we present our general topos-theoretic interpretation of Stone-type dualities; first we discuss dualities with categories of locales, then we introduce a general methodology, based on the notion of subterminal topology, for `enriching' them so to obtain dualities with categories of topological spaces.
In section \ref{ex}, we discuss various examples of dualities generated by using our method; we recover the classical Stone duality for distributive lattices (and Boolean algebras), the Alexandrov duality between preorders and Alexandrov spaces, the Lindenbaum-Tarski duality, the duality between spatial frames and sober spaces, and we establish new ones, including localic and topological dualities for meet-semilattices, an equivalence between the category of posets and a category of spatial locales (equivalently, a category of sober topological spaces), a localic duality for $k$-frames (for a regular cardinal $k$), and new dualities between specific categories of preordered structures.
In section \ref{generalization} we further generalize the method of section \ref{locales} for building dualities or equivalences with categories of locales starting from Morita-equivalences of the form $\Sh({\cal C}, J) \simeq \Sh(Id_{J}({\cal C}))$ to general equivalences $\Sh({\cal C}, J)\simeq \Sh({\cal D}, K)$, where $\cal C$ and $\cal D$ are preordered categories. This allows an abstract symmetric definition of the functors yielding the dualities, and provides us with an additional degree of freedom in building dualities or equivalences between categories of preordered structures. Grothendieck Comparison Lemma turns out be an extremely fruitful source of Morita-equivalences to which we can apply our methods; we illustrate this point in section \ref{addex} by generating several new dualities or equivalences. In particular, we establish a duality which naturally generalizes Birkhoff's duality for finite distributive lattices, and a duality which generalizes the well-known duality between algebraic lattices and sup-semilattices.
In section \ref{adj}, we apply the method `toposes as bridges' of \cite{OC10} to the Morita-equivalences $\Sh({\cal C}, J) \simeq \Sh(Id_{J}({\cal C}))$ and to the other equivalences established in section \ref{generalization} to obtain adjunctions which extend the dualities obtained in the previous sections; in particular, we establish reflections from various categories of preordered structures to the category of frames, as well as reflections between categories of posets satisfying some generalized `distributive law' and full subcategories of them consisting of posets satisfying certain `topological conditions'. In this context, as another application of our general method, we establish adjunctions between categories of toposes paired with points (as defined in section \ref{subterminal}) and categories of topological spaces.
In section \ref{insights}, we prove a number of results connecting properties of preordered structures with properties of the locales or topological spaces corresponding to them via the dualities or equivalences considered in the previous sections. Again, the technique that we employ for performing these `translations' is that of using toposes as `bridges' for transferring properties between their distinct sites of definition; specifically, we consider a number of logically-motivated topos-theoretic invariants, admitting bijective site characterizations, and rephrase them in terms of the two different representations $\Sh({\cal C}, J)$ and $\Sh(Id_{J}({\cal C}))$. We also establish various other results for preordered structures by using the same method. Of course, the topos-theoretic notions completely disappear in the final formulation of the results, they are just instrumental for performing the `automatic translation' (in the sense of \cite{OC10}) of properties from one site of definition into another.
Section \ref{spacesprop} deals with the problem of giving concrete characterizations of the topological spaces arising by putting the subterminal topology on a given set of points of a localic topos (equivalently, on a given set of models of a propositional geometric theory). To this end, we introduce a general method for building analogues of the Zariski spectrum for structures which can be described as models of propositional theories; in this context, we also provide characterizations of the syntactic categories of these theories as ordered algebraic structures presented by generators and relations. We treat this latter topic in full generality by introducing a generalized notion of first-order mathematical theory and a corresponding notion of syntactic category; we then restrict our attention to the syntactic categories of these general propositional theories and show that they can be characterized as models presented by generators and relations of certain Horn theories, which we call `ordered algebraic theories'. These notions pave the way for a topos-theoretic interpretation of the problem of giving explicit descriptions of models of theories of this kind presented by generators and relations. After discussing the abstract features of this interpretation, we illustrate its effectiveness by discussing several examples, notably including the construction (in section \ref{freecomjoin}) of the free frame on a complete join-semilattice, which solves in particular an open problem stated by P. Resende and S. Vickers in 2003 (cf. \cite{RV}). In section \ref{zariski}, we analyze the classical Zariski spectrum from our general topos-theoretic perspective and extend the definition of the Zariski topology on the collection of prime ideals of a ring to arbitrary collections of subsets of the ring.
In the appendix of the paper, we provide `elementary' proofs (that is, proofs which do not rely on results in Topos Theory) of some of the central results in the paper, including those which constitute the `machinery for generating dualities' of sections \ref{general} and \ref{generalization}; in fact, although we have established most of our results through a natural combination of abstract topos-theoretic techniques in light of the philosophy `toposes as bridges' of \cite{OC10}, most of our results, as well as their proofs, can be directly reformulated in the language of Locale Theory. The comparison between the `abstract' and `concrete' proofs of our results is an interesting one; on one hand, the direct arguments might be judged preferable to the abstract ones because they can be easier to understand to the reader who is not familiar with Topos Theory (in fact, this is the main reason why we have decided to provide them in the appendix of the paper), while on the other hand it is precisely the topos-theoretic arguments that are mostly illuminating from a conceptual perspective, and which can be most naturally generalized beyond the contexts that we have specifically addressed in the paper (cf. section \ref{conclusions} below for a further discussion of this point).
\subsection{Terminology and notation}\label{terminology}
The terminology and notation used in this paper are standard and borrowed from \cite{El}, if not otherwise specified. Conventions that we will frequently employ include the following.
Given a topological space $X$, we denote by $X_{0}$ its underlying set and by ${\cal O}(X)$ its frame of open sets. For a locale $L$, we denote by ${\cal O}(L)$ its underlying frame and, for a morphism $g:L\to L'$ of locales, we denote by $g^{\ast}:{\cal O}(L')\to {\cal O}(L)$ the frame homomorphism corresponding to it. Occasionally, when there is no risk of confusion, we denote the frame underlying a locale $L$ by the same letter.
By a \emph{subframe} of a frame $A$ we mean a subset $B\subseteq A$ such that $B$ is a frame with respect to the order induced by that on $A$ (equivalently, $0_{A}, 1_{A}\in B$ and $B$ is closed under arbitrary joins and finitary meets in $A$).
By an \emph{indexing function} of a set of points $\textsc{P}$ of a topos $\cal E$ by a set $X$ we mean a surjective function $i:X \to \textsc{P}$. If $\cal E$ has arbitrary set-indexed products, any such indexing $i$ naturally identifies with a geometric morphism $[X, \Set]\to {\cal E}$, denoted by $\tilde{i}$ and defined as follows: $\tilde{i}^{\ast}(A)(x)=i(x)^{\ast}(A)$ for any object $A$ of $\cal E$ and $\tilde{i}^{\ast}(A\to B)(x)=i(x)^{\ast}(A\to B)$ for any arrow $A\to B$ in $\cal E$. For any topos $\cal E$, we call a geometric morphism $\xi:[X, \Set]\to {\cal E}$ an \emph{indexing} of points of $\cal E$ by the set $X$; any such morphism induces an indexing function $i_{\xi}$ with domain $X$ of a set of points of $\cal E$, such that $\tilde{i_{\xi}}\cong \xi$; for any $x\in X$, we denote also by $\xi_{x}$ the point $i_{\xi}(x)$ of $\cal E$.
We consider points of toposes up to isomorphisms, rather than strictly.
We will denote by \textbf{Loc} the category of locales and by \textbf{Top} the category of topological spaces.
Given a category $\cal C$, we denote by $ob({\cal C})$ the collection of the objects of $\cal C$; for any arrow $f$ in $\cal C$, we denote by $dom(f)$ its domain and by $codom(f)$ its codomain. We sometimes write $c\in {\cal C}$ to mean that $c$ is an object of $\cal C$. The operation of composition of arrows in a category will always be denoted by the symbol $\circ$.
Given a preorder $\cal P$, and an element $a\in {\cal P}$, we denote by $(a)\downarrow$ the set of elements $b\in {\cal P}$ such that $b\leq a$.
The union of a family of subobjects $\{a_{i} \mono a \textrm{ | } i\in I\}$ in a topos will be denoted by $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}\mono a$.
The powerset of a given set $A$ will be denoted by ${\mathscr{P}}(A)$. The collection of all the finite subsets of a set $A$ will be denoted by ${\mathscr{P}}_{fin}(A)$.
By an atom in a category $\cal C$ with an initial object $0_{\cal C}$ we mean an object $c\ncong 0_{{\cal C}}$ of $\cal C$ such that for any monomorphism $b\mono c$, either $b\cong 0_{\cal C}$ or $b\mono c$ is an isomorphism; by an atomic subobject in $\cal C$ we mean a subobject whose domain is an atom in $\cal C$.
When we define `concrete' categories in this paper by only specifying their objects and arrows, we tacitly assume that the definition of the identities and composition rule is straightforward, i.e. that the identity arrow on any object is given by the identity function on its underlying set and the composition of arrows in the category is given by the set-theoretic composition of the underlying functions; that is, we consider all these `concrete' categories as categories structured over the category of sets.
\section{Topos-theoretic topologies}\label{subterminaltop}
In this section we introduce a method for building topological spaces starting from toposes equipped with a set of points. The resulting notions will be central for our purposes, in that, as we shall see in section \ref{dualtop}, they will allow us to naturally `lift` a given duality (resp. equivalence) between a category $\cal K$ of preorders and a category of locales to a duality (resp. equivalence) between $\cal K$ and a category of topological spaces.
\subsection{Spaces of points of a topos}\label{subterminal}
First, let us recall the following standard definition of the \emph{space of points} of a locale (cf. for example p. 491 \cite{El}).
A point of a locale $X$ is defined to be a locale morphism $p: 1 \to X$, where $1$ denotes the locale corresponding to the one-point space; equivalently, it is a frame homomorphism $p^{\ast}: {\cal O}(X) \to {\cal O}(1)\cong \{0,1\}$.
\begin{definition}
Let $X$ be a locale. The \emph{space of points} of $X$ is the set $X_{p}$ of all the points of $X$ equipped with the topology given by the image of the frame homomorphism $\phi_{X}: {\cal O}(X) \to {\mathscr{P}}(X_{p})$ defined by:
\[
\phi_{X}(U)=\{p\in X_{p} \textrm{ | } p^{\ast}(U)=1\},
\]
for any $U\in {\cal O}(X)$.
\end{definition}
From now on we will consider $X_{p}$ as a space equipped with this topology.
We note that we can interpret this definition topos-theoretically, as follows. The set $X_{p}$ of points of $X$ corresponds bijectively with the set $\textsc{P}$ of (isomorphism classes of) points of the topos $\Sh(X)$ (cf. Proposition C1.4.5 \cite{El}), and, for any point $p$ of $X$, the map $p^{\ast}:{\cal O}(X)\to \{0,1\}$ corresponds precisely to the action on subterminals of the inverse image $f_{p}^{\ast}:\Sh(X) \to \Set$ of the geometric morphism $f_{p}:\Set \to \Sh(X)$ corresponding to the point $p$. In these terms, the frame homomorphism $\phi_{X}$ acquires the following expression:
\[
\phi_{X}(U)\cong\{f_{p} \in \textsc{P} \textrm{ | } f_{p}^{\ast}(U)\cong 1_{\Set}\}.
\]
This remark naturally leads to the following more general definition.
\begin{definition}\label{funddef}
Let $\cal E$ be a locally small cocomplete topos, $\Gamma$ be a subframe of $\Sub_{\cal E}(1)$ and $i:X \to \textsc{P}$ be an indexing function of a set $\textsc{P}$ of points of $\cal E$ by a set $X$. The \emph{$\Gamma$-subterminal} topology ${\tau}^{\cal E}_{\Gamma, i}$ on the set $X$ is the image of the function $\phi_{\Gamma, {\cal E}}: \Gamma \to {\mathscr{P}}(X)$ given by
\[
\phi_{\Gamma, {\cal E}}(u)=\{x\in X \textrm{ | } \xi(x)^{\ast}(u) \cong 1_{\Set}\}.
\]
In other words, the subsets in ${\tau}^{\cal E}_{\Gamma, i}$ are precisely those of the form $\phi_{\Gamma, {\cal E}}(u)$ where $u$ ranges among the subterminals in $\Gamma$.
\end{definition}
If $\Gamma=\Sub_{{\cal E}}(1)$, we simply call ${\tau}^{\cal E}_{\Gamma, i}$ the \emph{subterminal topology}, and denote it by ${\tau}^{\cal E}_{i}$. If $\xi:[X, \Set]\to {\cal E}$ is an indexing of points of $\cal E$ by the set $X$ and $i_{\xi}$ is the corresponding indexing function of a set of points of $\cal E$, we denote ${\tau}^{\cal E}_{\Gamma, i}$ (resp. ${\tau}^{\cal E}_{i}$) also by ${\tau}^{\cal E}_{\Gamma, \xi}$ (resp. ${\tau}^{\cal E}_{\xi}$).
Instances of this notion have occasionally been considered in the literature; for example, the construction for $\Gamma=\Sub_{\cal E}(1)$ is used in the proof of Theorem 7.25 \cite{topos} (in fact, this result is subsumed by our Theorem \ref{topol} below).
Under the hypotheses of Defininition \ref{funddef}, we say that a collection $\textsc{P}$ of points of $\cal E$ \emph{separates the subterminals in $\Gamma$} if for any non-isomorphic subterminals $u, v$ in $\Gamma$ there exists a point $p \in \textsc{P}$ such that $p^{\ast}(u) \not \cong p^{\ast}(v)$. Note that if $\textsc{P}$ is a separating set of points for $\cal E$ (i.e. the inverse images of the points in $\textsc{P}$ jointly reflect isomorphisms) then in particular $\textsc{P}$ separates the subterminals in $\Gamma$, for any $\Gamma$.
\begin{theorem}\label{topol}
Let $\cal E$ be a locally small cocomplete topos, $\Gamma \subseteq \Sub_{{\cal E}}(1)$ be a subframe of $\Sub_{{\cal E}}(1)$ and $i:X\to \textsc{P}$ be an indexing function of a set of points of $\cal E$. Then
\begin{enumerate}[(i)]
\item The collection ${\tau}^{\cal E}_{\Gamma, i}$ of subsets of $X$ defines a topology on the set $X$;
\item The collection $\textsc{P}$ of points of $\cal E$ separates the subterminals in $\Gamma$ if and only if the frame of open sets of the topology ${\tau}^{\cal E}_{\Gamma, i}$ on $X$ is isomorphic to the frame $\Gamma$ via the map $\phi_{\Gamma, {\cal E}}$. In particular, if $\textsc{P}$ separates the subterminals of $\cal E$ then $\phi_{\Sub_{\cal E}(1), {\cal E}}$ yields an isomorphism between the frame of open sets of the topology ${\tau}^{\cal E}_{i}$ and $\Sub_{{\cal E}}(1)$.
\end{enumerate}
\end{theorem}
\begin{proofs}
$(i)$ Clearly, it suffices to prove that the function $\phi_{\Gamma, {\cal E}}: \Gamma \to {\mathscr{P}}(X)$ of Definition \ref{funddef} is a frame homomorphism.
The fact that for any $u,v\in \Gamma$, $\phi_{\Gamma, {\cal E}}(u\wedge v)=\phi_{\Gamma, {\cal E}}(u)\cap \phi_{\Gamma, {\cal E}}(v)$ (where $u\wedge v$ denotes the intersection of $u\mono 1$ and $v\mono 1$ in $\Sub_{\cal E}(1)$ or, equivalently, in $\Gamma$) follows immediately from the fact that each $\xi(x)^{\ast}$ preserves intersections of subobjects (being the inverse image functor of a geometric morphism).
Similarly, the equality $\mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{i\in I}}\phi_{\Gamma, {\cal E}}(u_{i})=\phi_{\Gamma, {\cal E}}({\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}u_{i}})$ follows from the fact that each $\xi(x)^{\ast}$ preserves unions of subobjects (being the inverse image functor of a geometric morphism).
$(ii)$ The function $\phi_{\Gamma, {\cal E}}$ is, by $(i)$, a frame homomorphism and, by definition of ${\tau}^{\cal E}_{\Gamma, i}$, always surjective, so it is an isomorphism precisely when it is injective i.e. when $\textsc{P}$ separates the subterminals in $\Gamma$.
\end{proofs}
Given $\cal E$, $\Gamma$ and $\xi:X\to \textsc{P}$ as in the hypotheses of the theorem, we denote the set $X$ equipped with the topology ${\tau}^{\cal E}_{\Gamma, i}$ (resp. ${\tau}^{\cal E}_{i}$) by $X_{{\tau}^{\cal E}_{\Gamma, i}}$ (resp. $X_{{\tau}^{\cal E}_{i}}$). If $\xi$ is the identity function on a set of points $\textsc{P}$ of $\cal E$, we denote ${\tau}^{\cal E}_{i}$ (resp. $X_{{\tau}^{\cal E}_{i}}$) also by ${\tau}^{\cal E}_{\textsc{P}}$ (resp. $X_{{\tau}^{\cal E}_{\textsc{P}}}$); if moreover $\cal E$ has only a \emph{set} of points (up to isomorphism) and $\textsc{P}$ is the collection of all the points of $\cal E$ we denote ${\tau}^{\cal E}_{\textsc{P}}$ (resp. $X_{{\tau}^{\cal E}_{\textsc{P}}}$) simply by ${\tau}^{\cal E}$ (resp. $X_{{\tau}^{\cal E}}$), and we call $X_{{\tau}^{\cal E}}$ the \emph{space of points} of the topos $\cal E$.
We note that part $(ii)$ of the theorem generalizes the following well-known result from Locale Theory: if a locale $L$ has enough points then the frame of open sets of the space of points of $L$ is isomorphic to ${\cal O}(L)$.
\begin{remarks}\label{topolinvariant}
\begin{enumerate}[(a)]
\item By Theorem \ref{topol}(ii), if $\textsc{P}$ separates the subterminals of $\cal E$ then $\Sh(X_{{\tau}^{\cal E}_{\textsl{P}}})$ is (equivalent to ) the localic part of the hyperconnected-localic factorization of the unique (up to isomorphism) geometric morphism ${\cal E}\to \Set$.
\item If $\cal E$ is a localic topos with enough points and $\textsl{P}$ is the collection of all the points of $\cal E$ then the topological space $X_{{\tau}^{\cal E}_{\textsl{P}}}$ is sober (cf. Theorem \ref{sobriety} below in view of Remark \ref{topolinvariant}(a)).
\item Every topological space is of the form $X_{{\tau}^{\cal E}_{\xi}}$ for some topos $\cal E$ and indexing function $i$ of a set of points of $\cal E$. Indeed, given a topological space $X$, $X$ is homeomorphic to the space $X_{{\tau}^{\cal E}_{i_{X}}}$, where $\cal E$ is the topos $\Sh(X)$ and $i_{X}:X_{0} \to \textsl{P}$ is the indexing function sending a point $x$ of $X_{0}$ to the geometric morphism $i_{X}(x):\Set \to \Sh(X)$ whose inverse image is the stalk functor at the point $x$.
\item The ($\Gamma$-)subterminal topology is a topos-theoretic invariant; that is, if $f:{\cal E}\to {\cal F}$ is an equivalence of toposes then for any indexing function $i:X \to \textsc{P}$ of a set $\textsc{P}$ of points of $\cal E$, denoted by $i':X\to \textsc{Q}$ the indexing function of points of $\cal F$ defined by setting $i'(x)$ equal to the point of $\cal F$ given by the composite of $i(x)$ with the equivalence $f$, the spaces $X_{{\tau}^{\cal E}_{i}}$ and $X_{{\tau}^{\cal F}_{i'}}$ are homeomorphic. In particular, if $\cal E$ and $\cal F$ are two equivalent toposes with the property of having only a \emph{set} of points (for example, if $\cal E$ and $\cal F$ are localic toposes) then the spaces $X_{{\tau}^{\cal E}}$ and $X_{{\tau}^{\cal F}}$ are homeomorphic.
\end{enumerate}
\end{remarks}
The construction of the $\Gamma$-subterminal topology can be naturally made functorial, as follows.
Let us define a category $\mathfrak{Top}_{t}$ whose objects are the triples $({\cal E}, \Gamma, \xi)$ where $\cal E$ is a locally small cocomplete topos, $\Gamma$ is a subframe of $\Sub_{\cal E}(1)$ and $\xi:[X, \Set] \to {\cal E}$ is an indexing of set of points of $\cal E$, and whose arrows $({\cal E}, \Gamma, \xi)\to ({\cal F}, \Gamma', \xi')$, where $\xi:[X, \Set] \to {\cal E}$ and $\xi':[Y, \Set] \to {\cal F}$ are indexings of points respectively of $\cal E$ and of $\cal F$, are the pairs $(f, l)$ where $f:{\cal E}\to {\cal F}$ is a geometric morphism such that $f^{\ast}$ sends the subterminals in $\Gamma'$ to subterminals in $\Gamma$ and $l:X\to Y$ is a function such that, denoted by $E_{l}:[X, \Set]\to [Y, \Set]$ the geometric morphism induced by $l$ as in Example A4.1.4 \cite{El}, the diagram
\[
\xymatrix {
[X, \Set] \ar[r]^{E_{l}} \ar[d]^{\xi} & [Y, \Set] \ar[d]^{\xi'} \\
{\cal E} \ar[r]^{f} & {\cal F}}
\]\\
commutes (up to isomorphism). Identities and composition in $\mathfrak{Top}_{t}$ are defined componentwise in the obvious way.
Notice that, given an geometric morphism $f:{\cal E}\to {\cal F}$ such that $f^{\ast}$ sends the subterminals in $\Gamma'$ to subterminals in $\Gamma$, $f$ lifts to an arrow $({\cal E}, \Gamma, \xi)\to ({\cal F}, \Gamma' \xi')$ (i.e. there exists $l:X\to Y$ such that $(f, l)$ yields a morphism $({\cal E}, \Gamma, \xi)\to ({\cal F}, \Gamma' \xi')$ in $\mathfrak{Top}_{s}$), where $\xi:[X, \Set] \to {\cal E}$ and $\xi':[Y, \Set] \to {\cal F}$ are indexings of points respectively of $\cal E$ and of $\cal F$, if and only if there is a way of assigning to each $x\in X$ an element $y\in Y$ with the property that the composite $f\circ \xi_{x}:\Set \to {\cal F}$ is equal to $\xi_{y}$. Note that if $i_{\xi'}$ is bijective then there is at most one such `lifting' $(f, l)$; in particular, if $\cal F$ has only a set of points and $i_{\xi'}$ is the identity on this set of points then there is exactly one `lifting' $(f, l_{f})$ for any geometric morphism $f:{\cal E}\to {\cal F}$.
Every arrow $(f, l):({\cal E}, \Gamma, \xi)\to ({\cal F}, \Gamma', \xi')$ in $\mathfrak{Top}_{t}$ gives rise to a continuous map of topological spaces $X_{{\tau}^{\cal E}_{\Gamma_{\cal E}, \xi}}\to X_{{\tau}^{\cal E}_{\Gamma_{\cal F}, \xi'}}$ with underlying function $l$. Indeed, by the commutativity of the diagram above, for any subterminal $v$ in $\Gamma_{\cal F}$, $l^{-1}(\phi_{\Gamma_{\cal F}, {\cal F}}(v))=\{x\in X \textrm{ | } l(x) \in \phi_{\Gamma_{\cal F}, {\cal F}}(v)\}=\{x\in X \textrm{ | } \xi(x)^{\ast}(f^{\ast}(v)) \cong 1_{\Set}\}=\phi_{\Gamma_{\cal E}, {\cal E}}({f^{\ast}(v)})$. We thus have a functor
\[
\Theta_{t}: \mathfrak{Top}_{t} \to \textbf{Top}.
\]
which sends an object $({\cal E}, \Gamma, \xi)$ of $\mathfrak{Top}_{t}$ to the topological space $X_{{\tau}^{\cal E}_{\Gamma_{\cal E}, \xi}}$ and an arrow $(f, l):({\cal E}, \Gamma, \xi)\to ({\cal F}, \Gamma', \xi')$ in $\mathfrak{Top}_{t}$ to the continuous map $l:X_{{\tau}^{\cal E}_{\Gamma_{\cal E}, \xi}}\to X_{{\tau}^{\cal E}_{\Gamma_{\cal F}, \xi'}}$. We will discuss various properties of this functor in section \ref{adj}.
We denote by $\mathfrak{Top}_{p}$ the full subcategory of $\mathfrak{Top}_{t}$ on the objects of the form $({\cal E}, \xi, \Sub_{\cal E}(1))$ , and call it the category of \emph{toposes paired with points}. Note that the objects of $\mathfrak{Top}_{p}$ can be simply identified with the pairs $({\cal E}, \xi)$, where $\cal E$ is a locally small cocomplete topos and $\xi:[X, \Set] \to {\cal E}$ is an indexing of a set of points of $\cal E$.
We denote the restriction of the functor $\Theta_{t}$ to the category $\mathfrak{Top}_{p}$ simply by $\Theta_{p}:\mathfrak{Top}_{p} \to \textbf{Top}$.
For any two subframes $\Gamma$ and $\Delta$ of $\Sub_{\cal E}(1)$, if $\Gamma$ is a subframe of $\Delta$ then we have a continuous surjection of topological spaces $X_{{\tau}^{\cal E}_{\Delta}, \xi}\to X_{{\tau}^{\cal E}_{\Gamma}, \xi}$; this surjection induces a geometric surjection of toposes $\Sh(X_{{\tau}^{\cal E}_{\Delta}, \xi}) \to \Sh(X_{{\tau}^{\cal E}_{\Gamma}, \xi})$ (cf. Example A4.2.7(c) \cite{El}).
The usefulness of Theorem \ref{topol} lies in the fact that it allows us to build topological spaces from toposes through a topos-theoretic invariant which has a natural behaviour with respect to sites; indeed, by Diaconescu's equivalence, the points of a topos $\Sh({\cal C}, J)$ correspond precisely to the flat $J$-continuous functors $\cal C \to \Set$ (in particular, if $\cal E$ is the classifying topos of a geometric theory $\mathbb T$ then the points of $\cal E$ correspond precisely to the models of $\mathbb T$ in $\Set$). Moreover, the formulation of the notion of ($\Gamma$-)subterminal topology as a topos-theoretic invariant paves the way, in light of the methodologies introduced in \cite{OC10}, for an effective transfer of properties between topological spaces constructed from two different sites of definition of a given topos. We will see concrete applications of this remark in sections \ref{insights} and \ref{zariski} below.
In passing, we note that the specialization order $\leq$ on $X_{{\tau}^{\cal E}_{\Gamma, \xi}}$ can be naturally characterized as a topos-theoretic invariant: $x\leq x'$ in $X_{{\tau}^{\cal E}_{\Gamma, \xi}}$ if and only if for every subterminal $u$ in $\Gamma$, $x\in \phi_{\Gamma, {\cal E}}(u)$ implies $x'\in \phi_{\Gamma, {\cal E}}(u)$.
\subsection{Examples}\label{exsub}
Let us begin our list of examples of subterminal topologies by giving an explicit description of the spaces of points of toposes of the form $\Sh({\cal C}, J)$, where $\cal C$ is a preorder category and $J$ is a Grothendieck topology on it. To this end, we introduce the following notions.
\begin{definition}
Let $({\cal C}, J)$ be a site.
\begin{enumerate}[(a)]
\item A \emph{$J$-ideal} on $\cal C$ is a subset $I\subseteq ob({\cal C})$ such that for any arrow $f:b\to a$ in $\cal C$ if $a\in I$ then $b\in I$, and for any $J$-covering sieve $R$ on an object $c$ of $\cal C$, if $dom(f)\in I$ for every $f\in R$ then $c\in I$; we denote by $Id_{J}({\cal C})$ the set of $J$-ideals on $\cal C$, endowed with the subset-inclusion order relation. If $J$ is the trivial topology on $\cal C$, we call the $J$-ideals on $\cal C$ simply \emph{ideals}, and we denote $Id_{J}({\cal C})$ by $Id({\cal C})$.
\item Given an object $c$ of $\cal C$, we the \emph{principal $J$-ideal} $(c)\downarrow_{J}$ generated by $c$ is the smallest $J$-ideal on $\cal C$ containing the object $c$, that is the collection of all the objects $d\in {\cal C}$ such that there exists a $J$-covering sieve $R$ on $d$ with the property that for every $f\in R$ there exists an arrow $dom(f)\to c$ in $\cal C$.
\item Let $({\cal C}, \leq)$ be a preorder category. A \emph{$J$-prime filter} on $\cal C$ is a subset $F\subseteq ob({\cal C})$ such that $F$ is non-empty, $a\in F$ implies $b\in F$ whenever $a\leq b$ in $\cal C$, for any $a, b\in F$ there exists $c\in F$ such that $c\leq a$ and $c\leq b$, and for any $J$-covering sieve $\{a_{i} \to a \textrm{ | } i\in I\}$ in $\cal C$ if $a\in F$ then there exists $i\in I$ such that $a_{i}\in F$.
\end{enumerate}
\end{definition}
Notice that if $\cal C$ is cartesian (i.e. a meet-semilattice) then the second condition in the definition of $J$-prime filter can be equivalently replaced by the requirement that for any $a,b\in F$, $a\wedge b\in F$ (where $\wedge$ denotes the meet operation in $\cal C$), while the condition that $F$ should be non-empty can be replaced by the requirement that $1\in F$ (where $1$ is the top element of $\cal C$).
The notion of $J$-ideal defined above makes also sense for a (not necessarily Grothendieck) coverage $J$ on ${\cal C}$, in the sense of Definition C2.1.1 \cite{El}; in fact, we will use this more general notion in section \ref{concreteness} below.
\begin{remarks}\label{rem2}
\begin{enumerate}[(a)]
\item The $J$-ideals on $\cal C$ can be identified with the subterminals of the topos $\Sh({\cal C}, J)$ (cf. Example C1.1.16 \cite{El}); under this identification, the usual order in $\Sub_{\Sh({\cal C}, J)}(1_{\Sh({\cal C}, J)})$ corresponds to the subset-inclusion order on the set of $J$-ideals on $\cal C$, and the principal $J$-ideal $(c)\downarrow_{J}$ on an object $c$ of $\cal C$ corresponds to the $J$-closure of the subobject $S_{c}\mono 1$ in $[{\cal C}^{\textrm{op}}, \Set]$ given by the monic part of the cover-mono factorization in $[{\cal C}^{\textrm{op}}, \Set]$ of the unique arrow $Hom_{\cal C}(-,c)\to 1_{[{\cal C}^{\textrm{op}}, \Set]}$.
\item If $J$ is subcanonical then $(c)\downarrow_{J}$ is equal to the collection of the objects $d$ of $\cal C$ such that there exists an arrow $d\to c$ in $\cal C$. Indeed, since $1_{[{\cal C}^{\textrm{op}}, \Set]}$ is always a sheaf, the action $a_{J}(S_{c})$ of the associated sheaf functor $a_{J}:[{\cal C}^{\textrm{op}}, \Set]\to \Sh({\cal C}, J)$ on the subobject $S_{c}\mono 1$ coincides with its $J$-closure. In particular, if $({\cal C}, \leq)$ is a preorder then $(c)\downarrow_{J}$ is equal to the set $(c)\downarrow$ of elements $d\in {\cal C}$ such that $d\leq c$.
\end{enumerate}
\end{remarks}
We can now state the following proposition.
\begin{proposition}\label{mslattice}
Let $\cal C$ be a preorder and $J$ be a Grothendieck topology on it. Then the topological space $X_{{\tau}^{\Sh({\cal C}, J)}}$ is homeomorphic to the space which has as set of points the collection ${\cal F}^{J}_{\cal C}$ of the $J$-prime filters on $\cal C$ and as open sets the sets the form
\[
{\cal F}_{I}=\{F\in {\cal F}^{J}_{\cal C} \textrm{ | } F\cap I\neq \emptyset\},
\]
where $I$ ranges among the $J$-ideals on $\cal C$. In particular, a sub-basis for this topology is given by the sets
\[
{\cal F}_{c}=\{F\in {\cal F}^{J}_{\cal C} \textrm{ | } c\in F\},
\]
where $c$ varies among the elements of $\cal C$.
\end{proposition}
\begin{proofs}
As we observed in Remark \ref{rem}(a), the subterminals in $\Sh({\cal C}, J)$ can be identified with the $J$-ideals on $\cal C$.
The points of $X_{{\tau}^{\Sh({\cal C}, J)}}$ are, by definition of $X_{{\tau}^{\Sh({\cal C}, J)}}$, the isomorphism classes of geometric morphisms $\Set \to \Sh({\cal C}, J)$. These correspond, by Diaconescu's equivalence, to the $J$-continuous flat functors $\cal C \to \Set$, and these in turn correspond exactly to the $J$-prime filters on $\cal C$ (cf. section \ref{logical} below).
Now, if $F$ is the $J$-prime filter corresponding to a point $p:\Set \to \Sh({\cal C}, J)$ then, since every $J$-ideal is the union of the principal $J$-ideals generated by the elements belonging to it, $p^{\ast}(I)=\mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{c\in I}}p^{\ast}((c)\downarrow_{J})$ is isomorphic to $1_{\Set}$ if and only if there exists $c\in I$ such that $p^{\ast}((c)\downarrow_{J})\cong 1_{\Set}$ (equivalently, $c\in F$), from which our thesis follows.
\end{proofs}
Now that we have an explicit description, provided by Proposition \ref{mslattice}, of the spaces of points of toposes of the form $\Sh({\cal C}, J)$, where $\cal C$ is a preorder category, it is natural to wonder if we can also explicitly describe in these terms the continuous map $X_{{\tau}^{f}}:X_{{\tau}^{\Sh({\cal D}, K)}}\to X_{{\tau}^{\Sh({\cal C}, J)}}$ resulting from applying the functor $\Theta_{p}: \mathfrak{Top}_{p} \to \textbf{Top}$ to the morphism $(\dot{f}, l_{f}):(\Sh({\cal D}, K), P_{\Sh({\cal D}, K)}, \to (\Sh({\cal C}, J)), P_{\Sh({\cal D}, K)})$, where $\dot{f}:\Sh({\cal D}, K) \to \Sh({\cal C}, J)$ is the geometric morphism induced by a morphism of sites $f:({\cal C}, J)\to ({\cal D}, K)$, $P_{\Sh({\cal D}, K)}$ (resp. $P_{\Sh({\cal C}, J)}$) is the indexing induced by the identity indexing function on the set $p_{\Sh({\cal D}, K)}$ (resp. $p_{\Sh({\cal C}, J)}$) of all the points of $\Sh({\cal D}, K)$ (resp. $\Sh({\cal C}, J)$) and $l_{f}:p_{\Sh({\cal D}, K)}\to p_{\Sh({\cal C}, J)}$ is the function induced by composition with $f$. The following proposition gives a positive answer to this question.
\begin{proposition}\label{mslatticefun}
With the notation above, if we identify $X_{{\tau}^{\Sh({\cal D}, K)}}$ (resp. $X_{{\tau}^{\Sh({\cal C}, J)}}$) with the set ${\cal F}^{K}_{\cal D}$ (resp. ${\cal F}^{J}_{\cal C}$) of $K$-prime filters on $\cal D$ (resp. of $J$-prime filters on $\cal C$) endowed with the subterminal topology, as in Proposition \ref{mslattice}, the continuous map $X_{{\tau}^{f}}:X_{{\tau}^{\Sh({\cal D}, K)}}\to X_{{\tau}^{\Sh({\cal C}, J)}}$ admits the following description: for any filter $F\in X_{{\tau}^{\Sh({\cal D}, K)}}$,
\[
X_{{\tau}^{f}}(F)=f^{-1}(F).
\]
\end{proposition}
\begin{proofs}
This immediately follows from the fact that if a point $p:\Set \to \Sh({\cal D}, K)$ of the topos $\Sh({\cal D}, K)$ corresponds to a $K$-prime filter $F$ on $\cal D$ (as in Proposition \ref{mslattice}) then the point $\dot{f} \circ p$ of the topos $\Sh({\cal C}, J)$ corresponds to the $J$-prime filter $f^{-1}(F)$ on $\cal C$. This fact is in turn easily verified by using the explicit description of Diaconescu's equivalence and observing that the inverse image functor $\dot{f}^{\ast}$ satisfies the property that for any $c\in {\cal C}$, $\dot{f}^{\ast}(y(c))\cong y'(f(c))$, where $y:{\cal C} \to \Sh({\cal C}, J)$ and $y':{\cal D}\to \Sh({\cal D}, K)$ are the composites of the associated sheaf functors $[{\cal C}^{\textrm{op}}, \Set]\to \Sh({\cal C}, J)$ and $[{\cal D}^{\textrm{op}}, \Set]\to \Sh({\cal D}, K)$ respectively with the Yoneda embeddings ${\cal C}\to [{\cal C}^{\textrm{op}}, \Set]$ and ${\cal D}\to [{\cal D}^{\textrm{op}}, \Set]$.
\end{proofs}
\begin{examples}\label{exa}
\begin{enumerate}[(a)]
Let us now point out several interesting topological spaces which naturally arise from putting the subterminal topology on the domain of some indexing function of set of points of a topos.
\item \emph{The trivial topology}
Given a topos $\cal E$ and an indexing $\xi$ of a set of points of $\cal E$ by a set $X$, if $\Gamma$ is equal to the subframe $\{0_{\cal E},1_{\cal E}\} \subseteq \Sub_{\cal E}(1)$ then the topological space $X_{{\tau}^{\cal E}_{\Gamma, \xi}}$ identifies with the trivial topological space with underlying set $X$.
\item \emph{The Alexandrov topology}
Let $({\cal P}, \leq)$ be a preorder, $\cal E$ be the topos $[{\cal P}, \Set]$, and let $i:{\cal P}\to \textsl{S}$ be the indexing function sending an element $p\in {\cal P}$ to the geometric morphism $e_{p}:\Set\to [{\cal P}, \Set]$ whose inverse image $e_{p}^{\ast}:[{\cal P}, \Set]\to \Set$ is the evaluation functor at the object $p$. The subterminals in $[{\cal P}, \Set]$ can be identified with the subsets $T$ of $\cal P$ such that for any $a\leq b$ in $\cal P$, $a\in T$ implies $b\in T$; clearly, the open set of the topological space ${\cal P}_{{{\tau}^{\cal E}}, i}$ corresponding to a subterminal $T$ coincides with $T$ itself (regarded as a subset of $\cal P$). So, the topological space ${\cal P}_{{{\tau}^{\cal E}}, i}$ is precisely the Alexandrov space associated to the preorder $\cal P$ (i.e. the topological space whose underlying set is $\cal P$ and whose open sets are the upper sets in $\cal P$).
\item \emph{The Stone topology for distributive lattices}
Let $\cal D$ be a distributive lattice (regarded as a preorder coherent category) and $\cal E$ be the topos $\Sh({\cal D}, J_{coh})$, where $J_{coh}$ is the coherent topology on $\cal D$ (recall that, for any element $d\in {\cal D}$, the $J_{coh}$-covering sieves on $d$ are precisely the sieves on $d$ which contain finite families $\{d_{i}\leq d \textrm{ | } i\in I\}$ with the property that $d=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}d_{i}$).
The subterminals in $\cal E$ can be identified with the $J_{coh}$-ideals in $D$ i.e. with the usual ideals of the distributive lattice $\cal D$. The points of $\Sh({\cal D}, J_{coh})$ can be identified with the prime filters on $\cal D$. Using Proposition \ref{mslattice}, we obtain that the subterminal topology ${\tau}^{\cal E}_{X}$ on the collection $X$ of prime filters on $\cal D$ has as basis of open sets the collection of sets of the form $\{F\in X \textrm{ | } d\in F\}$, where $d$ varies among the elements of $\cal D$. We have thus recovered the classical Stone topology on the set of prime filters on $\cal D$; in particular, if $\cal D$ is a Boolean algebra then the prime filters on $\cal D$ coincide with the ultrafilters on $\cal D$ and hence we recover the Stone topology on the set of ultrafilters of the Boolean algebra $\cal D$.
\item \emph{A topology for meet-semilattices}
Let $\cal M$ be a meet-semilattice (regarded as a preorder cartesian category) and let $\cal E$ be the topos $[{\cal M}^{op}, \Set]$. The points of the topos $\cal E$ can be identified with the filters on $\cal M$. By Proposition \ref{mslattice}, the topology ${\tau}^{\cal E}_{X}$ on the set $X$ of filters on $\cal M$ has as basis of open sets the collection of sets of the form $\{F\in X \textrm{ | } d\in F\}$ where $d$ varies among the elements of $\cal M$.
\item \emph{The space of points of a locale}
Let $L$ be a locale and $\cal E$ be the localic topos $\Sh(L)$; the points of the topos $\Sh(L)$ correspond bijectively with the points of the locale $L$ and hence there is only a \emph{set} $\textsl{P}$ of such points. Now, the points of a locale $L$ can be identified with the completely prime filters on the frame ${\cal O}(L)$ corresponding to $L$, while the subterminals in $\Sh(L)$ can be identified with the elements $u$ of $L$. Proposition \ref{mslattice} thus yields that the open sets of the subterminal topology ${\tau}^{\cal E}_{X}$ on the set $X$ of completely prime filters on $L$ are precisely the sets of the form $F_{u}=\{F\in X \textrm{ | } u\in F\}$ where $u$ ranges among the elements of $L$. We have thus recovered the usual topology on the space of points of the locale (as described for example in section C1.2 of \cite{El}).
\item \emph{A logical topology}
Let $\cal E$ be the classifying topos of a geometric theory $\mathbb T$ over a signature $\Sigma$. The points of $\cal E$ can be identified with the models of $\mathbb T$ in $\Set$. Let $X$ be a collection $X$ of such models. The subterminals in $\cal E$ can be identified with the $\mathbb T$-provable equivalence classes of geometric sentences over $\Sigma$, and the open set $F_{\phi}$ corresponding in the subterminal topology ${\tau}^{\cal E}_{X}$ to such a formula $\phi$ (regarded as a subterminal in $\cal E$) is the collection of all the models in $X$ which satisfy $\phi$. Thus $X_{{\tau}^{\cal E}_{\textsl{P}}}$ yields in this case a `logical topology' on the collection $X$ of models of $\mathbb T$; if every model of $\mathbb T$ in $\Set$ occurs as an element of $X$, we call the resulting topological space the \emph{logical space} associated to the theory $\mathbb T$. Notice that, since every geometric sentence over $\Sigma$ is $\mathbb T$-provably equivalent to a disjunction of coherent sentences, the collection of sets of the form $F_{\phi}$ for $\phi$ \emph{coherent} over $\Sigma$, forms a basis for the topology $X_{{\tau}^{\cal E}}$.
\item \emph{The Zariski topology}
Let $A$ be a commutative ring with unit, and let $L(A)$ be the distributive lattice generated by symbols $D(a)$, $a \in A$, subject to the relations $D(1_{A})=1_{L(A)}$, $D(a \cdot b) = D(a) \wedge D(b)$, $D(0_{A})=0_{L(A)}$, and $D(a+ b) \leq D(a) \vee D(b)$. If we equip $L(A)$ with the coherent topology then the space $X_{{\tau}^{\Sh(L(A), J)}}$ is homeomorphic to the topological space obtained by equipping the prime spectrum $Spec(A)$ of $A$ with the Zariski topology (cf. section \ref{zariski} below).
\end{enumerate}
\end{examples}
\subsection{Sobriety}\label{Sobriety}
In this section, we present a natural application of the invariant concept of point of a topos to the investigation of the property of sobriety of a topological space; specifically, Theorem \ref{sobriety} below gives a criterion for a topological space built through the subterminal topology to be sober.
\begin{theorem}\label{sobriety}
Let $\cal E$ be a locally small cocomplete topos and let $\textsl{P}$ be a set of points of $\cal E$ which separates the subterminals of $\cal E$. Let $h:{\cal E}\to {\cal F}$ be the hyperconnected part of the hyperconnected-localic factorization of the unique geometric morphism ${\cal E}\to \Set$. Then the topological space $X_{{\tau}^{\cal E}_{\textsl{P}}}$ is sober if and only if the function from $\textsl{P}$ to the collection of points of $\cal F$ which sends a point $s$ in $\textsl{P}$ to the point of $\cal F$ given by the composite geometric morphism $h\circ s$ is a bijection. In particular, if $\cal E$ is localic then the topological space $X_{{\tau}^{\cal E}_{\textsl{P}}}$ is sober if and only if $\textsl{P}$ is the collection of all the points of $\cal E$.
\end{theorem}
\begin{proofs}
Let $X$ be a topological space and let ${\cal F}_{X}$ be the collection of all the completely prime filters on the frame ${\cal O}(X)$ of open sets of $X$. It is well-known (cf. p. 491 \cite{El}) that a topological space $X$ is sober if and only if the map $\eta_{X}:X_{0}\to {\cal F}_{X}$ which sends a point $x\in X_{0}$ to the filter $\eta_{X}(x)$ consisting of all the open sets which contain $x$ is a bijection. Now, the filters in ${\cal F}_{X}$ can be identified with the geometric morphisms $\Set \to \Sh(X)$ and hence we can reformulate the property of sobriety of $X$ by saying that the map $\xi_{X}$ from $X_{0}$ to the collection of all the points of the topos $\Sh(X)$ sending a point $x\in X_{0}$ to the geometric morphism $f_{x}:\Set \to \Sh(X)$ whose inverse image is the stalk functor at $x$ is a bijection.
Now, since $\textsl{P}$ separates the subterminals in $\cal E$, $\cal F$ is equivalent to the topos $\Sh(X_{{\tau}^{\cal E}_{\textsl{P}}})$ (by Remark \ref{topolinvariant}(a)). It is easy to verify that, under the equivalence ${\cal F}\simeq \Sh(X_{{\tau}^{\cal E}_{\textsl{P}}})$ the map $\xi_{X_{{\tau}^{\cal E}_{\textsl{P}}}}$ defined above corresponds precisely to the map sending a point $s$ in $\textsl{P}$ (regarded as an element of the underlying set of $X_{{\tau}^{\cal E}_{\textsl{P}}}$) to the point of $\cal F$ given by the composite geometric morphism $h\circ s$. From this our thesis follows immediately.
\end{proofs}
It follows at once from the theorem that the Stone spaces associated to distributive lattices, the topological spaces associated to meet-semilattices as in Example \ref{exa}(d) and the logical spaces of Example \ref{exa}(f), are all sober. The following criterion for an Alexandrov space to be sober also follows immediately from our theorem (the equivalence $(i)\biimp (iii)$ in Corollary \ref{alexsober} below is well-known).
\begin{corollary}\label{alexsober}
Let $X$ be an Alexandrov space, and let $X_{\leq}$ be the preorder obtained by equipping the underlying set of $X$ with the specialization order. Then the following conditions are equivalent:
\begin{enumerate}[(i)]
\item $X$ is sober;
\item Every flat functor $X_{\leq}^{\textrm{op}}\to \Set$ is representable;
\item Every non-empty directed ideal of $X_{\leq}$ is principal.
\end{enumerate}
\end{corollary}
\begin{proofs}
$(i)\biimp (ii)$ If $X$ is an Alexandrov space then $X$ is homeomorphic to the space ${(X_{0})}_{{{\tau}^{\cal E}}, i}$, where $i:X_{0}\to \textsl{S}$ is the indexing function of points of the topos $[X_{\leq}, \Set]$ considered in Example \ref{exa}(b). From Theorem \ref{sobriety} we thus know that $X$ is sober if and only if every point of the topos $[X_{\leq}, \Set]$ is one of the points in $\textsl{S}$. Now, recalling the equivalence between geometric morphisms $\Set \to [X_{\leq}, \Set]$ and flat functors $X_{\leq}^{\textrm{op}}\to \Set$, under which the point of $[X_{\leq}, \Set]$ whose inverse image functor is the evaluation functor at the object $x$ of $X_{\leq}$ corresponds to the flat functor $X_{\leq}^{\textrm{op}}\to \Set$ represented by $x$, we can alternatively reformulate this condition as the requirement that every flat functor $X_{\leq}^{\textrm{op}}\to \Set$ should be representable.
$(ii)\biimp (iii)$ Every flat functor $F:{\cal P} \to \Set$ from a preorder $\cal P$ to $\Set$ takes values in $\{0, 1\}\hookrightarrow \Set$ and, under the assignment $F \to F^{-1}(1)$, the flat functors $X_{\leq}^{\textrm{op}}\to \Set$ correspond bijectively with the non-empty directed ideals on $X_{\leq}$ (cf. section \ref{logical} below); in these terms, the condition for a flat functor $X_{\leq}^{\textrm{op}}\to \Set$ to be representable amounts precisely to the requirement that the corresponding ideal should be principal, which proves our thesis.
\end{proofs}
\section{The general approach to Stone-type dualities}\label{general}
In this section we present our general topos-theoretic framework for interpreting `Stone-type dualities'. For a classical treatment of these dualities, done from a locale-theoretic and categorical perspective, we refer the reader to the excellent book \cite{stone} by Johnstone, which in fact provided significant inspiration for the present work.
It will be clear from our analysis that the known Stone-type dualities are just a few of a large class of dualities that can be established through our topos-theoretic machinery.
The main ingredient of our interpretation is the following well-known result from Topos Theory (cf. Example A4.6.2(e) \cite{El}): any topos $\Sh({\cal C}, J)$ of sheaves on a site $({\cal C}, J)$ whose underlying category $\cal C$ is a preorder is localic.
Now, for any locally small cocomplete topos $\cal E$, we can consider the hyperconnected-localic factorization ${\cal E} \to {\cal F} \to \Set$ of the unique geometric morphism ${\cal E} \to \Set$. Notice that, since $\cal F$ is localic, the Comparison Lemma yields an equivalence ${\cal F}\simeq \Sh(\Sub_{{\cal F}}(1_{\cal F}))$ and hence, by Proposition A4.6.6 \cite{El}, we have an equivalence ${\cal F}\simeq \Sub_{{\cal E}}(1_{\cal E})$. In particular, if ${\cal E}$ is localic then $\cal E$ is equivalent to the localic topos $\Sh(L)$ where $L$ is the locale $\Sub_{\cal E}(1)$ of subterminals in $\cal E$. If we apply this to the topos $\Sh({\cal C}, J)$ of sheaves on a site $({\cal C}, J)$ whose underlying category $\cal C$ is a preorder we thus obtain, in view of Remark \ref{rem}(a), an equivalence $\Sh({\cal C}, J) \simeq \Sh(Id_{J}({\cal C}))$ where $Id_{J}({\cal C})$ is the \emph{locale of $J$-ideals on $\cal C$} i.e. the locale whose corresponding frame consists of the set of $J$-ideals on $\cal C$ equipped with the subset-inclusion order relation. That is, we have the following result.
\begin{theorem}\label{fund}
Let $\cal C$ be a preorder and $J$ be a Grothendieck topology on $\cal C$. Then the toposes $\Sh({\cal C}, J)$ and $\Sh(Id_{J}({\cal C}))$ are equivalent.
\end{theorem}\qed
As we shall argue below, this representation theorem, which can be read logically as a Morita-equivalence between two distinct geometric theories (cf. section \ref{logical} below), provides a general framework for analyzing the known Stone-type dualities and extracting new information about them, as well as for generating new dualities. In fact, we shall identify a set of general principles for building dualities starting from Theorem \ref{fund}, and illustrate them in action in several examples, including the classical ones.
As the reader will have the opportunity to notice, this approach represents a clear implementation of the philosophy `toposes as bridges' introduced in \cite{OC10}. Indeed, the dualities or equivalences arise precisely from the process of `functorializing' a bunch of Morita-equivalences given by Theorem \ref{fund}; for each structure we have a Morita-equivalence, and the relationship between the preordered structure and the corresponding locale or topological space is determined by the expression of a topos-theoretic invariant in terms of the two different sites of definition of the topos.
We start by making a distinction between
\begin{enumerate}
\item dualities between categories of preorders and categories of \emph{locales} and
\item dualities between categories of preorders and categories of \emph{topological spaces}.
\end{enumerate}
The first kind of dualities have an essentially constructive nature, while the second class of dualities, which notably includes the classical Stone dualities for distributive lattices and Boolean algebras, may require some form of the axiom of choice. Anyway, the two classes of dualities are strongly interconnected, in that, as we shall see below, it is often possible to extract from a duality of the second kind a duality of the first kind and viceversa.
Broadly speaking, our method for building dualities of the first kind consists in equipping each of the structures $\cal C$ in a given category of preorders with an appropriate Grothendieck topology $J_{\cal C}$ in a such a way that this assignment is `natural' in $\cal C$. Such a choice induces a functor from the category $\cal K$ or its opposite (according to whether the duality is covariant or contravariant) to the category of locales sending each structure $\cal C$ to the locale $Id_{J_{\cal C}}({\cal C})$ of $J_{\cal C}$-ideals in $\cal C$. The operation of `recovering' a structure $\cal C$ from the corresponding topos $\Sh({\cal C}, J_{\cal C})\simeq \Sh(Id_{J_{\cal C}}({\cal C}))$ through a topos-theoretic invariant gives rise to a functor going in the converse direction which yields, together with it, the desired duality or equivalence.
Once a duality of the first kind is established, we use the notion of subterminal topology introduced in section \ref{subterminaltop} to `enrich' the given duality to a duality with a category of topological spaces. This can be done in various ways, and this process of `enrichment' might require, depending on the case, some form of the axiom of choice.
We thus begin by focusing on the first kind of dualities.
\subsection{Dualities with categories of locales}\label{locales}
Let us fix a category $\cal K$ of preordered structures. Our aim is to equip each structure $\cal C$ in $\cal K$ (regarded here as a preorder category) with a Grothendieck topology $J_{\cal C}$ on $\cal C$ in such a way that the assignment ${\cal C}\to Id_{J_{\cal C}}({\cal C})$ can be made into a functor (either covariant or contravariant) from $\cal K$ to the category $\textbf{Loc}$ of locales.
The following notions will be central for our purposes.
Recall from section C2.3 of \cite{El} that a \emph{morphism of sites} $({\cal C}, J_{{\cal C}})\to ({\cal D}, J_{{\cal D}})$ is a flat functor ${\cal C}\to {\cal D}$ (i.e., a functor $F:{\cal C}\to {\cal D}$ such that for any object $d$ of $\cal D$ the category $(d\downarrow F)$ is cofiltered) which is cover-preserving, i.e. which sends every $J_{{\cal C}}$-covering sieve to a family which generates a $J_{{\cal D}}$-covering sieve. In particular, if $\cal C$ and $\cal D$ are meet-semilattices (regarded as cartesian categories) a morphism of sites $({\cal C}, J_{{\cal C}})\to ({\cal D}, J_{{\cal D}})$ is a cover-preserving meet-semilattice homomorphism ${\cal C}\to {\cal D}$.
For preorder categories $\cal C$ and $\cal D$, the following `concrete' characterization of flat functors ${\cal C}\to {\cal D}$ holds.
\begin{proposition}
Let $({\cal C}, \leq)$ and $({\cal D}, \leq')$ be preorder categories and let $F:{\cal C}\to {\cal D}$ be a functor. Then $F$ is flat if and only if both of the following conditions hold:
\begin{enumerate}[(i)]
\item For any $d\in {\cal D}$ there exists $c\in {\cal C}$ such that $d\leq' F(c)$;
\item For any object $d\in {\cal D}$ and any objects $c,c'\in {\cal C}$ such that $d\leq' F(c)$ and $d\leq' F(c')$ there exists $c''\in {\cal C}$ such that $c''\leq c$, $c''\leq c'$ and $d\leq' F(c'')$.
\end{enumerate}
\end{proposition}
\begin{proofs}
The proposition follows immediately from the definition of cofiltered category.
\end{proofs}
Coming back to our original problem of making the assignment ${\cal C}\to Id_{J_{\cal C}}({\cal C})$ into a functor (either covariant or contravariant) from $\cal K$ to the category $\textbf{Loc}$ of locales, we have to distinguish between the covariant and contravariant case.
\begin{enumerate}[(i)]
\item If we want to obtain a contravariant functor from $\cal K$ to $\textbf{Loc}$, we equip each structure $\cal C$ in $\cal K$ with a Grothendieck topology $J_{{\cal C}}$ on $\cal C$ in such a way that every arrow $f:{\cal C}\to {\cal D}$ in $\cal K$ gives rise to a morphism of sites $\hat{f}:({\cal C}, J_{{\cal C}})\to ({\cal D}, J_{{\cal D}})$;
\item If we want to obtain a covariant functor ${\cal K} \to \textbf{Loc}$, we do not equip each of the structures in $\cal K$ with any Grothendieck topology.
\end{enumerate}
In case $(i)$, we obtain a functor $A:{\cal K}^{\textrm{op}}\to \textbf{Loc}$ from the opposite of the category $\cal K$ to the category $\textbf{Loc}$ of locales, while in case $(ii)$ we obtain a functor $B:{\cal K}\to \textbf{Loc}$.
In case $(i)$, the functor $A:{\cal K}^{\textrm{op}}\to \textbf{Loc}$ is defined as follows. Given a structure $\cal C$ in $\cal K$, we put $A({\cal C})=Id_{J_{\cal C}}({\cal C})$. Given an arrow $f:{\cal C}\to {\cal D}$ in ${\cal K}$, the morphism of sites $\hat{f}:({\cal C}, J_{{\cal C}})\to ({\cal D}, J_{{\cal D}})$ corresponding to $f$ gives rise, functorially, to a geometric morphism $\dot{f}:\Sh({\cal D}, J_{{\cal D}}) \to \Sh({\cal C}, J_{{\cal C}})$ (cf. Corollary C2.3.4 \cite{El}), which corresponds, via the equivalences of Theorem \ref{fund}, to a geometric morphism $\Sh(Id_{J_{\cal D}}({\cal D})) \to \Sh(Id_{J_{\cal C}}({\cal C}))$, which in turn corresponds, by Proposition C1.4.5 \cite{El}, to a unique morphism of locales $Id_{J_{\cal D}}({\cal D}) \to Id_{J_{\cal C}}({\cal C})$; we set $A(f):Id_{J_{\cal D}}({\cal D}) \to Id_{J_{\cal C}}({\cal C})$ equal to this morphism. It is easy to see that, concretely, $A(f)$ acts, at the level of frames, as the homomorphism sending a $J_{{\cal C}}$-ideal $I$ on $\cal C$ to the smallest $J_{\cal D}$-ideal on $\cal D$ containing the image of $I$ under $f$.
In case $(ii)$, the functor $B:{\cal K}\to \textbf{Loc}$ is defined as follows. Given a structure $\cal C$ in $\cal K$, we put $B({\cal C})=Id({{\cal C}^{\textrm{op}}})$. Any arrow $f:{\cal C}\to {\cal D}$ in ${\cal K}$ gives rise, functorially, to a geometric morphism $[{\cal C}, \Set] \to [{\cal D}, \Set]$ (as in Example A4.1.4 \cite{El}), which corresponds via the equivalences of Theorem \ref{fund} to a geometric morphism $\Sh(Id({{\cal C}^{\textrm{op}}})) \to \Sh(Id({{\cal D}^{\textrm{op}}}))$, which in turn corresponds (by Proposition C1.4.5 \cite{El}) to a unique morphism of locales $Id({{\cal C}^{\textrm{op}}}) \to Id({{\cal D}^{\textrm{op}}})$; we set $B(f):Id({\cal C}^{\textrm{op}}) \to Id({{\cal D}^{\textrm{op}}})$ equal to this morphism. Concretely, $B(f)$ acts, at the level of frames, as the homomorphism sending an ideal $I$ on $\cal D$ to the inverse image $f^{-1}(I)$ of $I$ under $f$.
We note that, if we regard both ${\cal K}^{\textrm{op}}$ (resp. $\cal K$) and $\textbf{Loc}$ as preordered $2$-categories in the natural way (i.e., given two arrows $f,g:{\cal C}\to {\cal D}$ in $\cal K$ we set $f\leq g$ in $\cal K$ (or in ${\cal K}^{\textrm{op}}$), if for every $c\in {\cal C}$ $f(c)\leq g(c)$, and similarly for arrows in $\textbf{Loc}$, where the inequalities are considered in the dual category $\textbf{Frm}$), then the functor $A$ (resp. the functor $B$) becomes a $2$-functor which is covariant on $2$-cells (cf. Remark C2.3.5 and Proposition C1.4.5 \cite{El}). Concretely, this amounts precisely to saying that if $f\leq g$ in $\cal K$ then $A(f)\leq A(g)$ (resp. $B(f)\leq B(g)$) in $\textbf{Loc}$.
So far, we have described a general method for constructing a (either covariant or contravariant) functor from a given category $\cal K$ of preorders to the category $\textbf{Loc}$ of locales; to build, starting from such a functor, an equivalence or duality between $\cal K$ and a subcategory of $\textbf{Loc}$, we have to care about how to go in the other direction. The general strategy is the following (we describe it for the case $(i)$ but our arguments can be trivially adapted to work in the case $(ii)$, cf. section \ref{prealex} below): if we are able to recover a structure ${\cal C}$ in $\cal K$ (uniquely up to canonical isomorphism) from the topos $\Sh({\cal C}, J_{{\cal C}})$ (equivalently, from the locale $A({\cal C})=Id_{J_{\cal C}}({\cal C})$) by means of a topos-theoretic invariant (in the sense of \cite{OC10}) functorially in ${\cal C} \in {\cal K}^{\textrm{op}}$ then we can expect to be able to use the invariant to define a functor $I_{A}:{\cal U}\to {\cal K}^{\textrm{op}}$ on a subcategory $\cal U$ of $\textbf{Loc}$ (namely, the extended image of $A$, cf. Definition \ref{extendedimage} below) which, together with $A$, yields an equivalence of categories ${\cal K}^{\textrm{op}} \simeq {\cal U}$.
We note that in order to be able to recover a structure $\cal C$ from the topos $\Sh({\cal C}, J_{{\cal C}})$, the topology $J_{\cal C}$ must be `small enough' so that the associated sheaf functor $a_{J_{\cal C}}:[{\cal C}^{\textrm{op}}, \Set]\to \Sh({\cal C}, J_{{\cal C}})$ does not send two distinct representable functors to isomorphic objects. Clearly, if $J_{\cal C}$ is subcanonical then this problem does not subsist (in fact, as we shall see in sections \ref{ex} and \ref{Mordual} below, the classical examples of dualities, as well as our new examples, all arise when the Grothendieck topologies are subcanonical) If $\cal C$ is moreover a poset then the objects of $\cal C$ correspond bijectively with the principal $J_{{\cal C}}$-ideals on $\cal C$ (since, $J_{{\cal C}}$ being subcanonical, the latter are all of the form $(c)\downarrow$ for $c\in {\cal C}$), so the problem reduces to that of characterizing the principal ($J_{{\cal C}}$-)ideals on $\cal C$ among the $J$-ideals on $\cal C$ (i.e. the subterminals of the topos $\Sh({\cal C}, J_{{\cal C}})$) by means of a topos-theoretic invariant. We will discuss this problem in full generality in section \ref{charinv}, and will show that it has a positive solution in many cases of interest.
Recall that a Grothendieck topology $J$ on a small category $\cal C$ is said to be subcanonical if all the representable functors ${\cal C}^{\textrm{op}}\to \Set$ are $J$-sheaves. If $({\cal C}, \leq)$ is a preorder category, the condition for a Grothendieck topology $J$ to be subcanonical admits the following more concrete description: for every $J$-covering sieve $S:=\{c_{i} \leq c \textrm{ | } i\in I\}$ on an object $c\in {\cal C}$, $c$ is the supremum in $\cal C$ of the $c_{i}$ for $i\in I$ (i.e., for any object $c'$ in $\cal C$ such that for every $i\in I$ $c_{i}\leq c'$, $c\leq c'$). We can show this as follows. If $\cal C$ is a preorder then any representable functor $F:{\cal C}^{\textrm{op}}\to \Set$ is a subterminal object of the topos $[{\cal C}^{\textrm{op}}, \Set]$ and hence, the terminal object $1$ of $[{\cal C}^{\textrm{op}}, \Set]$ being always a sheaf, $F$ is a $J$-sheaf if and only if it is $J$-closed as a subobject of $1$. But this condition, for a representable ${\cal C}(-, a)$, amounts precisely to requiring that for any $J$-covering sieve $S:=\{c_{i} \leq c \textrm{ | } i\in I\}$ on an object $c\in {\cal C}$, if $c_{i}\leq a$ for all $i\in I$ then $c\leq a$, from which our thesis follows immediately.
Recall that, if a functor $F:{\cal C}\to {\cal D}$ between two categories $\cal C$ and $\cal D$ is injective on objects (i.e., for any $c,c'\in {\cal C}$, $F(c)=F(c')$ implies $c=c'$) then we have a subcategory $Im(F)$ of $\cal D$, called the \emph{image} of $F$, whose objects are those of the form $F(c)$ for an object $c$ of $\cal C$ and whose arrows are those of the form $F(f)$ for an arrow $f$ of $\cal C$. Similarly, under the hypothesis that $F$ creates isomorphisms (i.e., for any isomorphism $l:F(c)\cong F(d)$ there exists an isomorphism $u:c\cong d$ such that $F(u)=l$), we can give the following definition.
\begin{definition}\label{extendedimage}
Let $F:{\cal C}\to {\cal D}$ be a functor which creates isomorphisms. The \emph{extended image} $ExtIm(F)$ of $F$ is the subcategory of $\cal D$ having as objects the objects of $\cal D$ which are isomorphic to an object of the form $F(c)$, and as arrows the arrows $f:x\to y$ in $\cal D$ such that there exist objects $c,c'\in {\cal C}$, an arrow $u:c\to c'$ in $\cal C$ and isomorphisms $x\cong F(c)$ and $y\cong F(c')$ such that $F(u)$ is the factorization of $f$ through these isomorphisms.
\end{definition}
The functors $F:{\cal C}\to {\cal D}$ which create isomorphisms enjoy a nice property, namely the fact that whenever they have a categorical left inverse, they yield an equivalence of categories between $\cal C$ and the extended image $ExtIm(F)$ of $F$. We shall appeal to this fact, recorded in the following proposition, in this section and in the next one.
\begin{proposition}\label{extim}
Let $F:{\cal C}\to {\cal D}$ be a functor which creates isomorphisms and let $G:{\cal D}\to {\cal C}$ be a functor such that $G\circ F \cong 1_{\cal C}$. Then the functors $\dot{F}:{\cal C}\to ExtIm(F)$ and $\dot{G}:ExtIm(F) \to {\cal C}$ obtained from $F$ and $G$ by restricting the codomain (resp. the domain) of $F$ (resp. of $G$) to the category $ExtIm(F)$ are categorical inverses to each other, and hence yield an equivalence of categories ${\cal C}\simeq ExtIm(F)$.
\end{proposition}
\begin{proofs}
Let $\alpha:G\circ F \cong 1_{\cal C}$ be a natural isomorphism; then clearly $\alpha$ yields a natural isomorphism $\dot{G}\circ \dot{F} \cong 1_{\cal C}$. We can construct a natural isomorphism $\beta:\dot{F}\circ \dot{G}\to 1_{ExtIm(F)}$ as follows. Given $x\in ExtIm(F)$ there exists $c\in {\cal C}$ and an isomorphism $r:x\to F(c)$ in $\cal D$; we put $\beta(x):F(G(x))\to x$ equal to the composite $r^{-1}\circ F(\alpha(c)) \circ F(G(r))$. It is easy to see that this assignment does not depend on the choice of the object $c$ and of the isomorphism $r$, and that it is natural in $c\in ExtIm(F)$; therefore, all the arrows $\beta(x)$ being isomorphisms, this assignment defines a natural isomorphism $\beta:\dot{F}\circ \dot{G}\to 1_{ExtIm(F)}$.
\end{proofs}
Note that the proof of Proposition \ref{extim} does not require any form of the axiom of choice.
It is clear from their definitions that our functors $A$ and $B$ are injective on objects. In fact, we can also prove that they are faithful, provided that all the Grothendieck topologies $J_{\cal C}$ are subcanonical and the categories in $\cal K$ are posets. Let us show this for the functors $A$; we will discuss the faithfulness of the functor $B$ in section \ref{prealex}.
From Lemma C2.3.8 \cite{El} we know that, given a morphism of sites $\hat{f}:({\cal C}, J_{{\cal C}})\to ({\cal D}, J_{{\cal D}})$ as above, if $J_{\cal C}$ and $J_{\cal D}$ are subcanonical then the inverse image of the geometric morphism $\dot{f}:\Sh({\cal D}, J_{{\cal D}}) \to \Sh({\cal C}, J_{{\cal C}})$ factors through the Yoneda embeddings ${\cal C} \hookrightarrow \Sh({\cal C}, J_{{\cal C}})$ and ${\cal D} \hookrightarrow \Sh({\cal D}, J_{{\cal D}})$ yielding a functor ${\cal C}\to {\cal D}$ which is isomorphic to $f$. If the categories $\cal C$ and $\cal D$ are posets, we can conclude that this functor actually coincides with $f$. This argument implies that, under the hypothesis that the categories $\cal C$ are posets and the topologies $J_{\cal C}$ are subcanonical, the functor $A$ is faithful; indeed, for any arrow $f$ in ${\cal K}^{\textrm{op}}$, $f$ is equal to the restriction ${\cal C} \to {\cal D}$ of the factorization $\Sh({\cal C}, J_{{\cal C}})\to \Sh({\cal D}, J_{{\cal D}})$ of the inverse image functor of the geometric morphism $\Sh(A(f)):\Sh(A({\cal D}))\to \Sh(A({\cal C}))$ through the equivalences of Theorem \ref{fund}. Therefore, since $A$ is faithful and injective on objects, it yields an isomorphism of categories between ${\cal K}^{\textrm{op}}$ and the image of $A$.
We can summarize what we have established as follows.
\begin{theorem}\label{teoabstr}
With the above notation, if all the Grothendieck topologies $J_{\cal C}$ are subcanonical and all the categories in $\cal K$ are posets then the functor $A:{\cal K}^{\textrm{op}}\to \textbf{Loc}$ yields an isomorphism of categories between ${\cal K}^{\textrm{op}}$ and the subcategory of $\textbf{Loc}$ given by the image of $A$.
\end{theorem}
If our categories are just preorders (rather than posets), $f$ remains determined by $A(f)$ only up to isomorphism, and hence $A$ becomes faithful when we factorize the set of arrows in $\cal K$ by the equivalence relation which identifies two arrows $f,g:{\cal C}\to {\cal D}$ if and only if they are isomorphic (i.e., for any $c\in {\cal C}$, $f(c)\cong g(c)$); if we denote the resulting quotient category of $\cal K$ by ${\cal K}_{\cong}$ we thus obtain that the factorization $A_{\cong}:{\cal K}_{\cong}^{\textrm{op}} \to \textbf{Loc}$ of the functor $A$ through the quotient functor ${\cal K}^{\textrm{op}} \to {\cal K}_{\cong}^{\textrm{op}}$ (note that this factorization exists since $A$ is a $2$-functor when ${\cal K}^{\textrm{op}}$ and $\textbf{Loc}$ are regarded as $2$-categories in the natural way) is faithful and hence yields, under the hypothesis that all the Grothendieck topologies $J_{\cal C}$ be subcanonical, an isomorphism of categories between ${\cal K}_{\cong}^{\textrm{op}}$ and the subcategory of $\textbf{Loc}$ given by the image of $A_{\cong}$.
Theorem \ref{teoabstr} provides a categorical equivalence between the opposite of our original category of preorders $\cal K$ and a subcategory of $\textbf{Loc}$. Still, it would be desirable to have a duality of $\cal K$ with a subcategory of $\textbf{Loc}$ whose objects (and arrows) admit an intrinsic description in localic terms; now, since such characterizations can clearly determine the objects (and arrows) in $\textbf{Loc}$ only up to isomorphism, in order to obtain an `intrinsic' duality, we necessarily have to enlarge the target of the functor $A$ to the extended image of $A$. In fact, if we are able to define a functor $I_{A}:ExtIm(A)\to {\cal K}^{\textrm{op}}$ which allows us to recover each structure $\cal C$ in ${\cal K}^{\textrm{op}}$ from $A({\cal C})$ up to natural isomorphism (i.e. such that that the composite functor $I_{A}\circ A$ is naturally isomorphic to the identity functor on $\cal C$) then the composite functor $A\circ I_{A}$ will automatically be naturally isomorphic to the identity functor on $ExtIm(A)$ (by Proposition \ref{extim}).
From now on we will suppose for simplicity that all the structures $\cal C$ in $\cal K$ are posets and that all the Grothendieck topologies $J_{\cal C}$ are subcanonical.
Concerning the definition of the functor $I_{A}$, we note that it suffices to specify the action of $I_{A}$ on the objects, its action on the arrows being uniquely determined by it; indeed, for any locale $L$ in the extended image of $A$, $I_{A}(L)$ can be identified with a subset $B_{L}$ of $L$, and the action of $I_{A}$ on an arrow $f:L\to L'$ in $ExtIm(A)$ must be equal to the factorization $B_{L'}\to B_{L}$ of the frame homomorphism corresponding to $f$ through the inclusions $B_{L}\hookrightarrow L$ and $B_{L'}\hookrightarrow L'$.
The definition of the functor $I_{A}$ on the objects relies on the existence of an appropriate topos-theoretic invariant $U$ which enables one to identify the principal $J_{\cal C}$-ideals on $\cal C$ among the $J_{\cal C}$-ideals on $\cal C$, for any structure $\cal C$ in $\cal K$; indeed, if such an invariant exists then each structure $\cal C$ in $\cal K$ can be recovered, up to isomorphism, as the poset of subterminals of the topos $\Sh({\cal C}, J_{\cal C})$ which satisfy the invariant $U$, and $I_{A}$ is defined as the functor which sends an locale $L$ in $ExtIm(A)$ to the set of subterminals of $\Sh(L)$ (i.e., of elements of $L$) which satisfy the invariant $U$, endowed with the induced order. As we shall see in section \ref{charinv}, the problem of the existence of an appropriate invariant $U$ is strictly related to the possibility of describing the Grothendieck topologies $J_{\cal C}$ `uniformly' by means of a topos-theoretic invariant; indeed, if the topologies $J_{\cal C}$ are `uniformly induced' by an invariant $C$ then the notion of $C$-compact subterminal yields an invariant $U$ which has the required property (cf. section \ref{charinv} below and in particular Theorem \ref{construction}).
The existence of an appropriate topos-theoretic invariant $U$ is, in a sense, the only element of `non-canonicity' in our machinery building dualities; indeed, it is not \emph{a priori} guaranteed that such an invariant should exist for our category of structures $\cal K$ and for our choice of the Grothendieck topologies $J_{\cal C}$. On the other hand, our notion of Grothendieck topology induced by a topos-theoretic invariant, introduced in section \ref{charinv} below, provides us with a systematic means for building appropriate invariants for many interesting and naturally arising classes collections of Grothendieck topologies. In particular, all the classical examples of dualities fall under this framework, and in fact the level of generality of this notion is such that it allows one to easily build infinitely many new dualities through our machinery.
Let us now discuss the problem of characterizing the subcategories of $\textbf{Loc}$ which arise as the extended image of a functor $A$ (the case of the functor $B$ is analogous and our arguments below can be easily adapted to work in that case, cf. section \ref{prealex} below) in `topological' terms, as for example in the case of the classical Stone duality. Clearly, one cannot reasonably expect this kind of problems to admit `canonical' solutions. On the other hand, topos-theoretic invariants can be profitably used to establish characterizations of the objects and arrows in the extended image of the functor $A$ starting from properties of the structures in $\cal K$. The notion of basis of a locale (resp. of basis of a topological space) is a particularly convenient one for achieving these characterizations.
By a \emph{basis} of a frame $L$ (equivalently, of the locale corresponding to it) we mean a subset $B\subseteq L$ such that every element of $L$ is a join of elements in $B$; note that a basis of a topological space $X$ is precisely a basis of the frame ${\cal O}(X)$ of open sets of $X$. Below, when we say that a basis (of a frame or a topological space) is closed under some kind of operation we mean that the result of applying the operation to any set of elements of the basis belonging to its domain is again an element of the basis.
As we shall see in section \ref{charinv}, if the Grothendieck topologies $J_{\cal C}$ can be defined `uniformly' by means of a topos-theoretic invariant $C$ (technically, are all $C$-induced in the sense of Definition \ref{inducedtop} below) which satisfies a natural property (technically, the hypothesis of Theorem \ref{construction}) then the principal $J_{\cal C}$-ideals on $\cal C$ can be characterized among the $J_{\cal C}$-ideals on $\cal C$ (i.e. the subterminals of the topos $\Sh({\cal C}, J_{\cal C})$) as the `$C$-compact' ones (in the sense of Definition \ref{compact}(a) below), and the locales $L$ in the extended image of our functor $A$ can be characterized `intrinsically' as the locales which possess a basis of $C$-compact elements satisfying some specific invariant properties (cf. Theorem \ref{propext}). So, for example, if $\cal K$ consists of all the Boolean algebras, each of which equipped with the coherent topology, one requires the elements of the basis to be closed in the locale under finite meets and joins and to form, with respect to the induced order, a Boolean algebra (see section \ref{ex} below for more examples).
Concerning the problem of characterizing the arrows in $ExtIm(A)$ `intrinsically', we observe that if the arrows ${\cal C}\to {\cal D}$ in $\cal K$ coincide exactly with the morphisms of sites $({\cal C}, J_{\cal C})\to ({\cal D}, J_{\cal D})$, as in fact it happens in many cases of interest, then, if the principal ideals on any $\cal C$ in $\cal K$ can be characterized as above as the $C$-compact elements of the locale $Id_{J_{\cal C}}({\cal C})$, the arrows in $ExtIm(A)$ can be characterized precisely as the locale morphisms whose associated frame homomorphisms send $C$-compact elements to $C$-compact elements (cf. Theorem \ref{propext}). If the arrows in $\cal K$ induce morphisms of sites $({\cal C}, J_{\cal C})\to ({\cal D}, J_{\cal D})$ but not all such morphisms yield arrows in $\cal K$ then the arrows in $ExtIm(A)$ can still be characterized, although less `intrinsically', as the locale morphisms whose associated frame homomorphisms, when restricted to the relevant bases of the locales, yield a function which, when the latter bases are regarded as objects of $\cal K$, can be identified with an arrow in $\cal K$ between them (cf. Theorem \ref{propext} below).
\subsection{Dualities with categories of topological spaces}\label{dualtop}
So far we have discussed how to build dualities or equivalences between categories of preorders and categories of locales. Our purpose in this section is to show that dualities (resp. equivalences) with categories of topological spaces can be naturally obtained starting from dualities (resp. equivalences) with categories of locales, this latter class being the one which, as we have seen, can be most naturally investigated by means of our topos-theoretic approach.
Our strategy consists in `enriching' a given duality (resp. equivalence) between a category of preorders $\cal K$ and a category of locales established by the method of section \ref{locales} to a duality (resp. equivalence) between $\cal K$ and a category of topological spaces, by means of an appropriate choice of points of the toposes corresponding to the structures in $\cal K$ as in section \ref{locales}.
We denote by $U:\textbf{Top}\to \textbf{Loc}$ the usual functor sending a topological space $X$ to the locale ${\cal O}(X)$ of open sets of $X$.
Let as assume to start with a duality $A:{\cal K}^{\textrm{op}}\to \textbf{Loc}$ obtained by the method of section \ref{locales}. Suppose that we have assigned to every structure $\cal C$ in $\cal K$ an indexing $\xi_{\cal C}:[X_{\cal C}, \Set] \to \Sh({\cal C}, J_{{\cal C}})$ of a set of points of the topos $\Sh({\cal C}, J_{{\cal C}})$ which separates the subterminals in $\Sh({\cal C}, J_{{\cal C}})$, and to each arrow $f:{\cal C}\to {\cal D}$ in $\cal K$ a function $l_{f}:X_{\cal D} \to X_{\cal C}$ such that, denoted by $\dot{f}$ the geometric morphism $\Sh({\cal D}, J_{{\cal D}}) \to \Sh({\cal C}, J_{{\cal C}})$ induced by the morphism of sites $\hat{f}:({\cal C}, J_{{\cal C}}) \to ({\cal D}, J_{{\cal D}})$, the pair $(\dot{f}, l_{f})$ defines an arrow $(\Sh({\cal D}, J_{{\cal D}}), \xi_{\cal D}) \to (\Sh({\cal C}, J_{{\cal C}}), \xi_{\cal C})$ in the category $\mathfrak{Top}_{p}$ of toposes paired with points (cf. section \ref{subterminal} above). Then we can define a functor $\tilde{A}:{\cal K}^{\textrm{op}}\to \textbf{Top}$ as follows:
\[
\tilde{A}({\cal C})=\Theta_{p}((\Sh({\cal C}, J_{{\cal C}}), \xi_{\cal C}))={X_{\cal C}}_{{{\tau}^{\Sh({\cal C}, J_{{\cal C}})}_{\xi_{\cal C}}}}
\]
for ${\cal C}\in {\cal K}$, and
\[
\tilde{A}(f)=\Theta_{p}(\dot{f},l_{f})=l_{f}:X_{\cal D} \to X_{\cal C}
\]
for any arrow $f:{\cal C}\to {\cal D}$ in $\cal K$ (cf. section \ref{subterminal} for the definition of the functor $\Theta_{p}$).
By Theorem \ref{topol}(ii), we have an isomorphism of functors $U\circ \tilde{A}\cong A$. Now, if $\tilde{A}$ creates isomorphisms (note that, since $A$ creates isomorphisms, this condition is automatically satisfied when $U$ is faithful on the image of $\tilde{A}$, for example when all the topological spaces in the image of $\tilde{A}$ are sober) then the functor $\tilde{I_{A}}:ExtIm(\tilde{A}) \to {\cal K}^{\textrm{op}}$ defined as the composite of $I_{A}$ with the restriction of the functor $U$ to $ExtIm(\tilde{A})$ clearly yields an inverse (up to isomorphism) to the functor $\tilde{A}$ (cf. Proposition \ref{extim}). We thus obtain an equivalence between ${\cal K}^{\textrm{op}}$ and the subcategory of $\textbf{Top}$ given by the extended image of $\tilde{A}$, which `lifts' the equivalence between ${\cal K}^{\textrm{op}}$ and the extended image of $A$ from which we started.
Let us now consider the covariant case. Let as assume to start with a functor $B:{\cal K} \to \textbf{Loc}$ obtained by the method of section \ref{locales}. Similarly as above, suppose that we have assigned to each structure $\cal C$ in $\cal K$ an indexing $\xi_{\cal C}:[X_{\cal C}, \Set] \to [{\cal C}, \Set]$ of a set of points of the topos $[{\cal C}, \Set]$ which separates the subterminals in $[{\cal C}, \Set]$, and to each arrow $f:{\cal C}\to {\cal D}$ in $\cal K$ a function $l_{f}:X_{\cal C} \to X_{\cal D}$ such that, denoted by $\dot{f}$ the geometric morphism $[{\cal C}, \Set] \to [{\cal D}, \Set]$ induced by $f$ as in section \ref{locales} above, the pair $(\dot{f}, l_{f})$ defines an arrow $([{\cal C}, \Set], \xi_{\cal C}) \to ([{\cal D}, \Set], \xi_{\cal D})$ in the category $\mathfrak{Top}_{p}$. Then we can define a functor $\tilde{B}:{\cal K}\to \textbf{Top}$ by putting
\[
\tilde{B}({\cal C})=\Theta_{p}(([{\cal C}, \Set], \xi_{\cal C}))={X_{\cal C}}_{{{\tau}^{[{\cal C}, \Set]}_{\xi_{\cal C}}}}
\]
for ${\cal C}\in {\cal K}$, and
\[
\tilde{B}(f)=\Theta_{p}(\dot{f}, l_{f})=l_{f}:X_{\cal C} \to X_{\cal D}
\]
for any arrow $f:{\cal C}\to {\cal D}$ in $\cal K$.
We verify as above that $U\circ \tilde{B}\cong B$; and, provided that $B$ is faithful and creates isomorphisms, if $\tilde{B}$ creates isomorphisms then $\tilde{B}$ yields an equivalence of categories onto its extended image.
We note that, in the contravariant (resp. covariant) case, if for every structure $\cal C$ in $\cal K$ the indexing $\xi_{\cal C}:[X_{\cal C}, \Set] \to \Sh({\cal C}, J_{{\cal C}})$ (resp. $\xi_{\cal C}:[X_{\cal C}, \Set] \to [{\cal C}, \Set]$) of points of the topos $\Sh({\cal C}, J_{{\cal C}})$ (resp. of the topos $[{\cal C}, \Set]$) is induced by the identity indexing function on the set $\textsc{P}_{\cal C}$ of \emph{all} the points of the topos $\Sh({\cal C}, J_{{\cal C}})$ (resp. of the topos $[{\cal C}, \Set]$) then for every arrow $f$ in $\cal K^{\textrm{op}}$ (resp. in $\cal K$), the definition of $l_{f}:X_{\cal D} \to X_{\cal C}$ (resp. of $l_{f}:X_{\cal C} \to X_{\cal D}$) is automatic; indeed, $l_{f}$ must be equal to the function sending a point $p:\Set \to \Sh({\cal D}, J_{{\cal D}})$ (resp. a point $p:\Set \to [{\cal C}, \Set]$) to the composite $\dot{f}\circ p$.
We note that, while the method of section \ref{locales} provides a `canonical' way of building dualities or equivalences with subcategories of \textbf{Loc}, once the category $\cal K$ of structures and the Grothendieck topologies $J_{\cal C}$ have been fixed, the method for building dualities with categories of topological spaces that we have just explained requires the choice of appropriate sets of points of the toposes involved, and this can be done in general in several different ways, possibly leading to different dualities or equivalences between the category $\cal K$ and a subcategory of $\textbf{Top}$ (cf. section \ref{ex} below for a concrete example of this phenomenon).
In connection with this, we also remark that some form of the axiom of choice may be necessary in establishing that certain sets of points of a given topos separate the subterminals of the topos (we refer the reader to section \ref{ex} for a detailed analysis of various examples also in light of the axiom of choice).
Concerning the problem of characterizing the subcategories of $\textbf{Top}$ arising as the extended image of a functor $\tilde{A}$ (resp. of a functor $\tilde{B}$) in `topological' terms, we have a particularly satisfying answer when the topological spaces in the extended image of the functor are all sober (cf. Theorem \ref{sobriety} for a topos-theoretic criterion of sobriety); indeed, under this hypothesis the spaces in $\cal V$ can be characterized precisely as the sober topological spaces whose frames of open sets (regarded as locales) belong to the extended image of the functor $A$ (resp. of the functor $B$). Indeed, one direction is obvious while the other one follows from the fact that the assignment $X\to {\cal O}(X)$ defines an embedding of the category of sober spaces into the category of locales (cf. Corollary C1.2.3 \cite{El}). In particular, if the locales in the extended image of $A$ (resp. of $B$) can be characterized by means of a topos-theoretic invariant $C$ as in section \ref{locales} above then the spaces in $ExtIm(\tilde{A})$ (resp. in $ExtIm(\tilde{B})$) can be characterized precisely as the sober topological spaces such that the collection of their $C$-compact open sets forms a basis which, endowed with the subset-inclusion ordering, has the structure of a preorder in $\cal K$ and satisfies some specific invariant properties (as in the characterization of the locales in $ExtIm(A)$ (resp. in $ExtIm(B)$)), while the arrows in $ExtIm(\tilde{A})$ (resp. in $ExtIm(\tilde{B})$) can be characterized precisely as the continuous maps between the spaces in $ExtIm(\tilde{A})$ (resp. in $ExtIm(\tilde{B})$) such that their inverse images send $C$-compact open sets to $C$-compact open sets.
\subsection{Characterizing principal ideals through invariants}\label{charinv}
In this section, we investigate the problem of recovering a (preorder) category $\cal C$, equipped with a (subcanonical) Grothendieck topology $J$ from the topos $\Sh({\cal C}, J)$ (up to categorical equivalence) by means of a topos-theoretic invariant. Notice that any categorical equivalence between two preorder categories is an isomorphism, so our considerations will address in particular the problem of recovering a preordered structure $\cal C$ in $\cal K$ from the topos $\Sh({\cal C}, J_{{\cal C}})$ (respectively, from the topos $[{\cal C}, \Set]$) uniquely up to isomorphism.
Let us denote by $l_{{\cal C}}^{J}:{\cal C}\to \Sh({\cal C}, J)$ the composite of the Yoneda embedding $Y:{\cal C}\to [{\cal C}^{\textrm{op}}, \Set]$ with the associated sheaf functor $a_{J}:[{\cal C}^{\textrm{op}}, \Set]\to \Sh({\cal C}, J)$. The functor $l_{{\cal C}}^{J}$ maps $\cal C$ into $\Sh({\cal C}, J)$ not faithfully in general; anyway, if $J$ is subcanonical then $l_{{\cal C}}^{J}$ can be identified with the Yoneda embedding and hence it is full and faithful. It is thus natural to ask, under this hypothesis, whether one can characterize the objects of $\Sh({\cal C}, J)$ of the form $l_{{\cal C}}^{J}(c)$ for an object $c$ of $\cal C$ in terms of a topos-theoretic invariant applied to the topos $\Sh({\cal C}, J)$. If ${\cal C}$ is a preorder $({\cal P}, \leq)$ then for an object $c$ of $\cal C$ the unique arrow $l_{{\cal C}}^{J}(c)\to 1_{\Sh({\cal C}, J)}$ is a monomorphism and hence $l_{{\cal C}}^{J}(c)$ can be identified with a $J$-ideal on $\cal C$; in fact, if $J$ is subcanonical then $l_{{\cal C}}^{J}(c)$ is (isomorphic to) the the principal $J$-ideal $(c)\downarrow$ generated by $c$ (i.e. the collection of all the objects $d$ of $\cal P$ such that $d\leq c$). This leads us to investigating the problem of characterizing the subterminals in $\Sh({\cal C}, J)$ which correspond to the principal $J$-ideals on $\cal C$ through a topos-theoretic invariant applied to the topos $\Sh({\cal C}, J)$.
The following definition will be central for our purposes. Recall from section \ref{exsub} that a $J$-ideal $I$ on $\cal C$ is said to be principal if $I=(c)\downarrow_{J}$ for an object $c$ of $\cal C$.
\begin{definition}
Let $({\cal C}, J)$ be a site and $I$ be a $J$-ideal on $\cal C$. We say that $I$ is \emph{$J$-compact} if for every covering $\{I_{k} \textrm{ | } k\in K\}$ of $I$ by subterminals in $\Sh({\cal C}, J)$ there exists a $J$-covering sieve $\{f_{h}:c_{h}\to c \textrm{ | } h\in H\}$ in $\cal C$ such that for every $h\in H$, $(c_{h})\downarrow_{J}$ is contained in some $I_{k}$, and $\{(c_{h})\downarrow_{J} \textrm{ | } h\in H\}$ covers $I$ in $\Sh({\cal C}, J)$.
\end{definition}
\begin{theorem}\label{Jcomp}
Let $({\cal C}, J)$ be a site and $I$ be a $J$-ideal on $\cal C$. Then $I$ is a principal $J$-ideal if and only if it is $J$-compact.
\end{theorem}
\begin{proofs}
Let us suppose that $I=(c)\downarrow_{J}$ is a $J$-principal ideal. The union $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{k\in K}}I_{k}$ in $\Sh({\cal C}, J)$ of a family $\{I_{k} \textrm{ | } k\in K\}$ of $J$-ideals is the $J$-closure of the union $\mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{k\in K}}I_{k}$ in $[{\cal C}^{\textrm{op}}, \Set]$ of the ideals $I_{k}$ (considered as subterminals in $[{\cal C}^{\textrm{op}}, \Set]$); specifically, $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{k\in K}}I_{k}=\{d\in {\cal C} \textrm{ | } \{f:e\to d \textrm{ | } e\in \mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{k\in K}}I_{k}\} \in J(d)\}$. This implies that $I=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{k\in K}}I_{k}$ if and only if $\{f:e\to c \textrm{ | } e\in \mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{k\in K}}I_{k}\} \in J(c)$. So, if we put $S=\{f:e\to c \textrm{ | } e\in \mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{k\in K}}I_{k}\}$ then we have that for any $f:e\to c$ in $S$, $(e)\downarrow_{J}$ is contained in some $I_{k}$. To show that the sieve $S$ witnesses the fact that $I$ is $J$-compact, it remains to prove that $I=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{f\in S}}(dom(f))\downarrow_{J}$. Since each $(dom(f))\downarrow_{J}$ is contained in $I$ (being included in some $I_{k}$) then the inclusion ``$\supseteq$'' holds. The other inclusion follows from the fact that every $J$-ideal that contains each of the $(dom(f))\downarrow_{J}$ must contain $c$ (since $S$ is $J$-covering) and hence $I$.
Conversely, suppose that $I$ is $J$-compact. We want to prove that $I$ is principal. If we take as family $\{I_{k} \textrm{ | } k\in K\}$ the singleton family $\{I\}$, the fact that $I$ is $J$-compact implies that there exists a $J$-covering sieve $\{f_{h}:c_{h}\to c \textrm{ | } h\in H\}$ in $\cal C$ such that $I=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{h\in H}}(c_{h})\downarrow_{J}$. But $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{h\in H}}(c_{h})\downarrow_{J}=(c)\downarrow_{J}$, from which it follows that $I=(c)\downarrow_{J}$ and hence that $I$ is a principal $J$-ideal.
\end{proofs}
The characterization of Theorem \ref{Jcomp} represents a first step for achieving, under appropriate hypotheses, characterizations of the property of being a principal $J$-ideal on a preorder $\cal C$ as a topos-theoretic invariant on the topos $\Sh({\cal C}, J)$. To this end, we introduce the following definitions.
\begin{definition}\label{ade}
\begin{enumerate}[(a)]
\item We say that a family of subobjects $\{a_{i} \mono a \textrm{ |} i\in I\}$ of a given object $a$ in a (locally small cocomplete) topos $\cal E$ is \emph{refined} by a family $\{b_{j} \mono a \textrm{ | } j\in J\}$ of subobjects of $a$ if for every $j\in J$ there exists $i\in I$ such that $b_{j}\mono a$ factors through $a_{i}\mono a$ and the union $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $\Sub_{\cal E}(a)$ is equal to the union $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{j\in J}}b_{j}$ in $\Sub_{\cal E}(a)$.
\item Let $({\cal C}, J)$ be a site. We say that a topos-theoretic invariant property $C$ of families ${\cal F}$ of subterminals in a (locally small cocomplete) topos $\cal E$ is \emph{$({\cal C}, J)$-adequate} if, when $\cal E$ is the topos $\Sh({\cal C}, J)$ and $\cal F$ is a family of principal $J$-ideals on $\cal C$ (regarded as subterminals in $\Sh({\cal C}, J)$), $\cal F$ has a refinement which satisfies $C$ if and only if there exists a $J$-covering sieve $\{f_{h}:d_{h}\to d \textrm{ | } h\in H\}$ in $\cal C$ such that for any $h\in H$, $(d_{h})\downarrow_{J}$ is contained in some ideal in $\cal F$ and the union in $\cal E$ of the family $\{(d_{h})\downarrow_{J} \textrm{ | } h\in H\}$ is equal to the union in $\cal E$ of the family $\cal F$ (equivalently, the objects of $\cal C$ whose corresponding principal $J$-ideals are contained in some ideal in $\cal F$ are the domains of a family of arrows in $\cal C$ which generates a $J$-covering sieve).
\end{enumerate}
\end{definition}
\begin{remark}\label{grot}
If $({\cal C}, \leq)$ is a poset category and $J$ is subcanonical then the elements in $\cal F$ can be thought of as elements of $\cal C$, and the condition that there exist a $J$-covering sieve $\{f_{h}:d_{h}\to d \textrm{ | } h\in H\}$ in $\cal C$ such that for any $h\in H$, $(d_{h})\downarrow_{J}$ is contained in some ideal in $\cal F$ and the union in $\cal E$ of the family $\{(d_{h})\downarrow_{J} \textrm{ | } h\in H\}$ is equal to the union in $\cal E$ of the family $\cal F$ can be equivalently reformulated as the requirement that the supremum $s$ in $\cal C$ of the set of the objects of $\cal C$ whose corresponding principal $J$-ideals are contained in some ideal in $\cal F$ should exist and the collection of all the arrows in $\cal C$ from such objects to $s$ should generate a $J$-covering sieve.
\end{remark}
\begin{theorem}\label{equivJcomp}
Let $C$ be an invariant property of families of subterminals in a (locally small cocomplete) topos which is $({\cal C}, J)$-adequate. Then a $J$-ideal on $\cal C$ is $J$-compact (equivalently, by Theorem \ref{Jcomp}, principal) if and only if every covering family for $I$ has a refinement which satisfies $C$.
\end{theorem}
\begin{proofs}
If a $J$-ideal $I$ is $J$-compact then every covering family for $I$ has a refinement by principal $J$-ideals which satisfies $C$.
Conversely, suppose that every covering family for $I$ has a refinement which satisfies $C$. If we take as covering family the collection of all the principal $J$-ideals of the form $(c)\downarrow_{J}$ for $c\in I$ then we conclude, from the fact that $C$ is $({\cal C}, J)$-adequate, that $I$ is the join of a family of principal $J$-ideals indexed over a $J$-covering sieve, and hence it is principal (equivalently, $J$-compact), as required.
\end{proofs}
Let us now introduce some natural topos-theoretic invariants of families of subobjects in a topos, which we shall use in section \ref{ex} in the context of our examples.
Below, by atomic (resp. supercompact) subobject of a given object $a$ in a locally small cocomplete topos $\cal E$ we mean a subobject $m$ of $a$ such that $m$ is non-zero and does not contain any proper subobject of $a$ in $\cal E$ (resp. such that every covering of $m$ in the frame $\Sub_{{\cal E}}(a)$ contains the identity subobject of $a$). An element $a$ of a frame $L$ (equivalently, of the corresponding locale) is an atom (resp. supercompact) if, denoted by $y:L\to \Sh(L)$ the Yoneda embedding, $y(a)\mono 1_{\Sh(L)}$ is an atomic (resp. supercompact) subobject in the topos $\Sh(L)$.
\begin{definition}\label{refinsubcover}
Let $\cal E$ be a locally small cocomplete topos and $\cal F$ be a family of subobjects of a given object in $\cal E$.
\begin{enumerate}[(a)]
\item The family $\cal F$ is said to \emph{have a finite subcover} if there exists a finite subfamily ${\cal F}'$ of $\cal F$ such that the union of the subobjects in ${\cal F}'$ is equal to the union of the subobjects in $\cal F$;
\item The family $\cal F$ is said to \emph{have a singleton subcover} if there exists a single subobject in $\cal F$ which is the union of the subobjects in the family ${\cal F}$;
\item Given a regular cardinal $k$, the family $\cal F$ is said to \emph{have a $k$-subcover} if there exists a subfamily ${\cal F}'$ of $\cal F$ of cardinality $\lt k$ such that the union of the subobjects in ${\cal F}'$ is equal to the union of the subobjects in $\cal F$;
\item The family $\cal F$ is said to \emph{have a disjunctive refinement} (resp. to \emph{have a finite disjunctive refinement}) if there exists a family (resp. a finite family) of pairwise disjoint subobjects which refines $\cal F$;
\item The family $\cal F$ is said to \emph{have an atomic refinement} (resp. to \emph{have a finite atomic refinement} if either $\cal F$ is a singleton or there exists a family (resp. a finite family) of atomic subobjects which refines $\cal F$;
\item The family $\cal F$ is said to \emph{have a supercompact refinement} (resp. to \emph{have a finite supercompact refinement} if either $\cal F$ is a singleton or there exists a family (resp. a finite family) of supercompact subobjects which refines $\cal F$;
\item The family $\cal F$ is said to \emph{have a directed refinement} if there exists a directed family $\cal G$ (i.e. a non-empty family such that for any two subobjects in $\cal G$ there exists a subobject in $\cal G$ which contains both) which refines $\cal F$.
\end{enumerate}
\end{definition}
Notice that for $k=2$ (resp. $k=\omega$) the invariant `to have a $k$-subcover' specialize to the invariant `to have a singleton subcover' (resp. `to have a finite subcover').
\begin{remark}\label{refin}
All the invariants defined above are of the form `to have a refinement satisfying $C$' for a topos-theoretic invariant $C$ of families of subterminals in a topos.
\end{remark}
The definition above naturally leads us to introducing the following notions (the concepts of compactness and supercompactness in the definition below already appear in \cite{El}).
\begin{definition}
Let $\cal E$ be a locally small cocomplete topos and let $L$ be a locale, with corresponding Yoneda embedding $y:L\to \Sh(L)$.
\begin{enumerate}[(a)]
\item An object $A$ of $\cal E$ is said to be \emph{compact} if every covering family of $A$ in $\cal E$ has a finite subcover.
An element $a$ of a locale $L$ (equivalently, of a frame ${\cal O}(L)$) is said to be compact if the object $y(a)$ is compact in $\Sh(L)$, equivalently if whenever $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ there exists a finite subset $I_{0}\subseteq I$ such that $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I_{0}}}a_{i}$.
\item An object $A$ of $\cal E$ is said to be \emph{supercompact} if every covering family of $A$ has a singleton subcover.
An element $a$ of a locale $L$ (equivalently, of a frame ${\cal O}(L)$) is said to be supercompact if the object $y(a)$ is supercompact in $\Sh(L)$, equivalently if whenever $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ there exists an index $i\in I$ such that $a=a_{i}$.
\item For a regular cardinal $k$, an object $A$ of $\cal E$ is said to be \emph{$k$-compact} if every covering family of $A$ in $\cal E$ has a $k$-subcover.
An element $a$ of a locale $L$ (equivalently, of a frame ${\cal O}(L)$) is said to be $k$-compact if the object $y(a)$ is $k$-compact in $\Sh(L)$, equivalently if whenever $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ there exists a subset $I_{0}\subseteq I$ of cardinality $\lt k$ such that $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I_{0}}}a_{i}$.
\item An object $A$ of $\cal E$ is said to be \emph{disjunctively compact} (resp. \emph{infinitarily disjunctively compact}) if every covering family of $A$ in $\cal E$ has a finite disjunctive refinement (resp. a disjunctive refinement).
An element $a$ of a locale $L$ (equivalently, of a frame ${\cal O}(L)$) is said to be disjunctively compact (resp. infinitarily disjunctively compact) if the object $y(a)$ is disjunctively compact (resp. infinitarily disjunctively compact) in $\Sh(L)$, equivalently if whenever $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ there exists a finite family (resp. a family) $\{b_{j} \leq a \textrm{ | } j\in J\}$ of elements of $L$ such that for every $j\in J$ there exists $i\in I$ such that $b_{j}\leq a_{i}$, the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ is equal to the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{j\in J}}b_{j}$ and for any distinct $j,j'\in J$, $b_{j}\wedge b_{j'}=0$.
\item An object $A$ of $\cal E$ is said to be \emph{atomically compact} (resp. \emph{infinitarily atomically compact}) if every covering family of $A$ in $\cal E$ has a finite atomic refinement (resp. an atomic refinement).
An element $a$ of a locale $L$ (equivalently, of a frame ${\cal O}(L)$) is said to be \emph{atomically compact} (resp. \emph{infinitarily atomically compact}) if the object $y(a)$ is atomically compact (resp. infinitarily atomically compact) in $\Sh(L)$, equivalently if whenever $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ either there is $i\in I$ such that $a_{i}=a$ or there exists a finite family (resp. a family) $\{b_{j} \leq a \textrm{ | } j\in J\}$ of elements of $L$ such that the $b_{j}$ (for $j\in J$) are atoms in $L$, for every $j\in J$ there exists $i\in I$ such that $b_{j}\leq a_{i}$, and the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ is equal to the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{j\in J}}b_{j}$.
\item An object $A$ of $\cal E$ is said to be \emph{supercompactly compact} (resp. \emph{infinitarily supercompactly compact}) if every covering family of $A$ in $\cal E$ has a finite supercompact refinement (resp. a supercompact refinement).
An element $a$ of a locale $L$ (equivalently, of a frame ${\cal O}(L)$) is said to be \emph{supercompactly compact} (resp. \emph{infinitarily supercompactly compact}) if the object $y(a)$ is supercompactly compact (resp. infinitarily supercompactly compact) in $\Sh(L)$, equivalently if whenever $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ either there is $i\in I$ such that $a_{i}=a$ or there exists a finite family (resp. a family) $\{b_{j} \leq a \textrm{ | } j\in J\}$ of elements of $L$ such that the $b_{j}$ (for $j\in J$) are supercompact elements of $L$, for every $j\in J$ there exists $i\in I$ such that $b_{j}\leq a_{i}$, and the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ is equal to the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{j\in J}}b_{j}$.
\item An object $A$ of $\cal E$ is said to be \emph{directedly compact} if every covering family of $A$ in $\cal E$ has a directed refinement.
An element $a$ of a locale $L$ (equivalently, of a frame ${\cal O}(L)$) is said to be \emph{directedly compact} if the object $y(a)$ is directedly compact in $\Sh(L)$, equivalently if whenever $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ there exists a family $\{b_{j} \leq a \textrm{ | } j\in J\}$ of elements of $L$ such that for any $j, j'\in J$ there exists $j''\in J$ such that $b_{j}\leq b_{j''}$ and $b_{j'}\leq b_{j''}$, for every $j\in J$ there exists $i\in I$ such that $b_{j}\leq a_{i}$, and the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ is equal to the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{j\in J}}b_{j}$.
\end{enumerate}
\end{definition}
\begin{remark}\label{loctop}
Topos-theoretic invariants of families of subterminals in a locally small cocomplete topos can be identified with locale-theoretic invariants. Indeed, a topos-theoretic invariant $U$ of families of subterminals in a locally small cocomplete topos yields a locale-theoretic invariant property $L_{U}$ of families of elements of a locale, defined by saying that for any locale $L$, $a\in L$ satisfies $L_{U}$ if and only if $y(a)$ satisfies $U$ in the topos $\Sh(L)$, where $y:L\hookrightarrow \Sh(L)$ is the Yoneda embedding; conversely, given a locale-theoretic invariant property $Q$, we have a topos-theoretic invariant $T_{Q}$ of families of subterminals of a topos, defined by saying that a family of subterminals in a locally small cocomplete topos $\cal E$ satisfies $T_{Q}$ if and only if, regarded as a family of elements of $\Sub_{\cal E}(1)$, it satisfies $Q$.
\end{remark}
The following abstract definition unifies the notions introduced above and enlightens the link between them and the concepts of Definition \ref{refinsubcover}.
\begin{definition}\label{compact}
\begin{enumerate}[(a)]
\item Given a topos-theoretic invariant property $C$ of families of subobjects of a given object in a (locally small cocomplete) topos, an object $A$ of a topos is said to be \emph{$C$-compact} if and only if any covering family of subobjets of $A$ in $\cal E$ has a refinement by a family of subobjects satisfying the property $C$;
\item Given a topos-theoretic invariant $C$ of families of subterminals in a topos (equivalently, by Remark \ref{loctop}, a locale-theoretic invariant of families of elements of a locale), an element $l$ of a locale $L$ (equivalently, of the corresponding locale) is said to be \emph{$C$-compact} if every covering family of $l$ in $L$ has a refinement satisfying $C$; that is, whenever $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ there exists a family $\{b_{j} \leq a \textrm{ | } j\in J\}$ of elements of $L$ satisfying $C$ such that for every $j\in J$ there exists $i\in I$ such that $b_{j}\leq a_{i}$, and the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $L$ is equal to the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{j\in J}}b_{j}$ in $L$.
\end{enumerate}
\end{definition}
\begin{examples}\label{excomp}
\begin{enumerate}[(a)]
\item An object is compact if and only if it is $C$-compact where $C$ is the invariant `to be finite';
\item An object is supercompact if and only if it is $C$-compact where $C$ is the invariant `to be a singleton';
\item An object is $k$-compact (for a regular cardinal $k$) if and only if it is $C$-compact wheren $C$ is the invariant `to be of cardinality $\lt k$';
\item An object is disjunctively compact (resp. infinitarily disjunctively compact) if and only if it is $C$-compact where $C$ is the invariant `to be finite and disjoint' (resp. to be disjoint');
\item An object is atomically compact (resp. infinitarily atomically compact) if and only if it is $C$-compact where $C$ is the invariant `to be a singleton or to be finite and consisting of atomic subobjects' (resp. to be a singleton or to consist of atomic subobjects');
\item An object is supercompactly compact (resp. infinitarily supercompactly compact) if and only if it is $C$-compact where $C$ is the invariant `to be a singleton or to be finite and consisting of supercompact subobjects' (resp. to be a singleton or to consist of supercompact subobjects')
\item An object is directedly compact if and only if it is $C$-compact where $C$ is the invariant `to be directed'.
\end{enumerate}
\end{examples}
We note that on any \emph{disjunctively coherent category} $\cal C$ (i.e., regular category in which the initial object exists and the unions of finite families of pairwise disjoint subobjects exist and are stable under pullback) one can put the \emph{disjunctive topology} i.e. the Grothendieck topology $D_{\cal C}$ on $\cal C$ whose $D_{\cal C}$-covering sieves on any object $c$ of $\cal C$ are exactly the covering sieves on $c$ in $\cal C$ which contain finite families of subobjects of $c$ which are pairwise disjoint from each other. Note that on a Boolean algebra (regarded as a coherent category) the disjunctive topology coincides with the coherent topology, while on a total order (regarded as a coherent category) the disjunctive topology specializes exactly to the trivial topology.
Similarly, on any \emph{disjunctively geometric category} $\cal C$ (i.e., regular category in which the initial object exists and the unions of set-indexed families of pairwise disjoint subobjects exist and are stable under pullback) one can put the \emph{infinitary disjunctive topology}, i.e. the Grothendieck topology $G_{\cal C}$ on $\cal C$ whose $G_{\cal C}$-covering sieves on any object $c$ of $\cal C$ are exactly the covering sieves on $c$ in $\cal C$ which contain set-indexed families of subobjects of $c$ which are pairwise disjoint from each other.
Recall that, given a regular cardinal $k$, a \emph{$k$-geometric category} is a regular category in which unions of arbitrary families of $\lt k$ subobjects exist and are stable under pullback; on a $k$-geometric category $\cal C$ we can define the \emph{$k$-covering topology} as the Grothendieck topology on $\cal C$ whose covering sieves on any object $c$ of $\cal C$ are exactly the sieves on $c$ which contain covering families of subobjects of $c$ of cardinality $\lt k$. Note that the $2$-geometric categories are precisely the regular categories, while the $\omega$-geometric categories are precisely the coherent categories.
We shall call a poset a \emph{disjunctively distributive lattice} (resp. an \emph{disjunctively distributive frame}) if is is a disjunctively coherent category (resp. a disjunctively geometric category) when regarded as a preorder category. We define the category $\textbf{DJLat}$ of disjunctively distributive lattices as the category whose objects are the disjunctively distributive lattices and whose arrows are the meet-semilattices homomorphisms between them which preserve (the $0$ and) finite pairwise disjoint joins. Similarly, we define the category $\textbf{DJFrm}$ of disjunctively distributive frames as the category whose objects are the disjunctively distributive frames and whose arrows are the meet-semilattices homomorphisms between them which preserve (the $0$ and) arbitrary pairwise disjoint joins.
We shall call a poset a \emph{$k$-frame} (cf. \cite{kframes}) if it is a $k$-geometric category when regarded as a preorder category, equivalently if it is a meet-semilattice in which there exists the join of any family of $\lt k$-elements and the infinite distributive law of these joins with respect to binary meets holds. We define the category $k\textrm{-}\textbf{Frm}$ of $k$-frames as the category whose objects are the $k$-frames and whose arrows are the meet-semilattices homomorphisms between them which send joins of families of $\lt k$-elements to joins.
Recall that a \emph{preframe} is a poset with finite meets and directed joins, in which the finite meets distribute over the directed joins. We denote by $\textbf{PFrm}$ the category having as objects the preframes and as arrows the meet-semilattice homomorphisms between them which send directed joins to directed joins. Given a preframe $\cal P$, we define the \emph{directed topology} $J^{dir}_{{\cal P}}$ on $\cal P$ as the Grothendieck topology on $\cal P$ whose covering sieves on any object $p\in {\cal P}$ are the sieves on $p$ which contain a directed sieve on $p$, i.e. a sieve $S$ on $p$ such that for any arrows $a\leq p$ and $b\leq p$ in $S$ there exists $c\leq p$ in $S$ such that $a\leq c$, $b\leq c$.
Let us define a \emph{weakly atomic meet-semilattice} to be a meet-semilattice in which the bottom element $0$ exists and in which finite joins of atoms always exist and distribute over finite meets. Similarly, we define an \emph{infinitarily weakly atomic meet-semilattice} to be a meet-semilattice in which the bottom element $0$ exists and in which arbitrary joins of atoms always exist and distribute over finite meets. We define the category $\textbf{WAtMSLat}$ of weakly atomic meet-semilattices as the category whose objects are the weakly atomic meet-semilattices and whose arrows are the meet-semilattices homomorphisms between them which preserve the $0$, send atoms to atoms and preserve finite joins of atoms. Similarly, we define the category $\textbf{IWAtMSLat}$ of infinitarily weakly atomic meet-semilattices as the category whose objects are the infinitarily weakly atomic meet-semilattices and whose arrows are the meet-semilattices homomorphisms between them which preserve the $0$, send atoms to atoms and preserve arbitrary joins of atoms.
On a weakly atomic meet-semilattice $\cal M$ (regarded as a cartesian preorder category) one can define the \emph{atomically generated topology} $J^{at}_{\cal M}$, as follows. For any sieve $S$ in $\cal M$ on an object $m\in {\cal M}$, we set $S\in J^{at}_{\cal M}(m)$ if and only if either $S$ is the maximal sieve or $S$ is a covering sieve in $\cal C$ (i.e. $m$ is the supremum in $\cal M$ of all the domains of the arrows in $S$) generated by a finite family of atoms in $\cal M$. Clearly, on an infinitarily weakly atomic meet-semilattice we can define the \emph{infinitary atomically generated topology}, as the (obvious) infinitary analogue of the atomically generated topology.
If in the definition of weakly atomic meet-semilattice we replace atoms by supercompact objects we obtain a similar but more general notion: we define a \emph{weakly supercompact meet-semilattice} to be a meet-semilattice in which the bottom element $0$ exists and in which finite joins of supercompact elements always exist and distribute over finite meets. Similarly, we define an \emph{infinitarily weakly supercompact meet-semilattice} to be a meet-semilattice in which the bottom element $0$ exists and in which arbitrary joins of supercompact elements exist and distribute over finite meets. We define the category $\textbf{WScMSLat}$ of weakly supercompact meet-semilattices as the category whose objects are the weakly supercompact meet-semilattices and whose arrows are the meet-semilattices homomorphisms between them which preserve the $0$, send supercompact elements to supercompact elements and preserve finite joins of supercompact elements. Similarly, we define the category $\textbf{IWScMSLat}$ of infinitarily weakly supercompact meet-semilattices as the category whose objects are the infinitarily weakly supercompact meet-semilattice and whose arrows are the meet-semilattices homomorphisms between them which preserve the $0$, send supercompact elements to supercompact elements and preserve arbitrary joins of supercompact elements.
On a weakly supercompact meet-semilattice $\cal M$ (regarded as a cartesian preorder category) one can define the \emph{supercompactly generated topology} $J^{sc}_{\cal M}$, as follows: for any sieve $S$ in $\cal M$ on an object $m\in {\cal M}$, we set $S\in J^{sc}_{\cal M}(m)$ if and only if either $S$ is the maximal sieve or $S$ is a covering sieve in $\cal C$ (i.e. $m$ is the supremum in $\cal M$ of all the domains of the arrows in $S$) generated by a finite family $\{a_{i}\leq m \textrm{ | } i\in I\}$ where the $a_{i}$ are all supercompact elements of $\cal M$. Clearly, on an infinitarily weakly supercompact meet-semilattice we can define the \emph{infinitary supercompactly generated topology}, as the (obvious) infinitary analogue of the supercompactly generated topology.
Note that all the Grothendieck topologies defined above are subcanonical.
The following theorem shows that, for several naturally arising subcanonical sites $({\cal C}, J)$, including the ones introduced above, there are topos-theoretic invariants which are $({\cal C}, J)$-adequate.
\begin{theorem}\label{thmadequate}
Let $({\cal C}, J)$ be a site. Then
\begin{enumerate}[(i)]
\item If $\cal C$ is a coherent category and $J$ is the coherent topology on $\cal C$ then the invariant `to be finite' is $({\cal C}, J)$-adequate;
\item If $\cal C$ is a regular category and $J$ is the regular topology on $\cal C$ then the invariant `to be a singleton' is $({\cal C}, J)$-adequate;
\item If $\cal C$ is any category and $J$ is the trivial topology on $\cal C$ then the invariant `to be a singleton' is $({\cal C}, J)$-adequate;
\item If $\cal C$ is a $k$-geometric category (for a regular cardinal $k$) and $J$ is the $k$-covering topology on $\cal C$ then the invariant `to have cardinality $\lt k$' is $({\cal C}, J)$-adequate;
\item If $\cal C$ is a disjunctively distributive lattice (resp. a disjunctively distributive frame) and $J$ is the disjunctive topology (resp. the infinitary disjunctive topology) on $\cal C$ then the invariant `to be finite and disjoint' (resp. `to be disjoint') is $({\cal C}, J)$-adequate;
\item If $\cal C$ is a weakly atomic meet-semilattice (resp. an infinitarily weakly atomic meet-semilattice) and $J$ is the atomically generated topology (resp. the infinitary atomically generated topology) on $\cal C$ then the invariant `to be a singleton or to be finite and consisting of atomic subobjects' (resp. `to be a singleton or to consist of atomic subobjects') is $({\cal C}, J)$-adequate;
\item If $\cal C$ is a weakly supercompact meet-semilattice (resp. an infinitarily weakly supercompact meet-semilattice) and $J$ is the supercompactly generated topology (resp. the infinitary atomically generated topology) on $\cal C$ then the invariant `to be a singleton or to be finite and consisting of supercompact subobjects' (resp. `to be a singleton or to consist of supercompact subobjects') is $({\cal C}, J)$-adequate.
\end{enumerate}
\end{theorem}
\begin{proofs}
$(i)$ Given a family ${\cal F}=\{(c_{i})\downarrow_{J} \textrm{ | } i\in I\}$ of principal $J$-ideals on $\cal C$, if this family has a finite subcover $\{(c_{i})\downarrow_{J} \textrm{ | } i\in I_{0}\}$, where $I_{0}$ is a finite subset of $I$, we can construct a $J$-covering sieve $S$ on $\cal C$ such that the $c_{i}$ (for $i\in I_{0}$) are exactly the domains $dom(f)$ of a family of arrows $f$ generating the sieve $S$. Indeed, for any $i\in I_{0}$, consider the union $e\mono 1_{\cal C}$ of the subobjects arising as the monic parts of the cover-mono factorizations of the unique arrows $c_{i}\to 1_{\cal C}$ to the terminal object $1_{\cal C}$ of $\cal C$; we have, for each $i\in I_{0}$, a canonical arrow $c_{i}\to e$, and if we define $S$ to be the sieve on $e$ generated by these arrows $c_{i}\to e$ then $S$ is $J$-covering (by definition of coherent topology) and satisfies the condition in the definition of $({\cal C}, J)$-adequate invariant.
Conversely, if $\cal F$ satisfies the condition in the definition of $({\cal C}, J)$-adequate invariant then clearly $\cal F$ has a finite subcover.
$(ii)$, $(iii)$ and $(iv)$ follow from an entirely analogous argument to that for $(i)$.
$(v)$ We only prove the finitary version of the result, the infinitary one being entirely analogous to it. Given a family ${\cal F}$ of principal $J$-ideals on $\cal C$, if this family has a finite disjunctive refinement $\{I_{k} \textrm{ | } k\in K\}$ then for any distinct $k,k'\in K$, $I_{k}\cap I_{k'}=\{0_{\cal C}\}$, and for any $k\in K$ there exists $c_{k}$ in $\cal C$ such that $(c_{k})\downarrow_{J}$ belongs to $\cal F$ and $I_{k}\subseteq (c_{k})\downarrow_{J}$. First, we note that for every $k\in K$, $I_{k}$ is a principal $J$-ideal. Indeed, since $c_{k}\in \mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{}{\cal F}}=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{k\in K}I_{k}}$ there exists a $J$-covering sieve $S:=\{a_{w}^{k} \mono c_{k} \textrm{ | } w\in W_{k}\}$ on $c_{k}$ made of pairwise disjoint subobjects of $c_{k}$ such that every $a_{w}^{k}$ belongs to $\mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{k\in K}I_{k}}$. By definition of $J$, $c_{k}$ is equal to the disjoint join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{w\in W_{k}}a^{k}_{w}}$ in $\cal C$. Now, if $b\in I_{k}$ then $b\leq c_{k}$ (since $I_{k}\subseteq (c_{k})\downarrow_{J}=(c_{k})\downarrow$) and hence $b=b\wedge c_{k}=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{w\in W_{k}}(b\wedge a^{k}_{w})}=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{w\in W_{k} \textrm{ | } a_{w}^{k}\in I_{k}}(b\wedge a^{k}_{w})}=b\wedge ( \mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{w\in W_{k} \textrm{ | } a_{w}^{k}\in I_{k}} a^{k}_{w}})$ (since the $I_{k}$ are pairwise disjoint). But, $I_{k}$ being a $J$-ideal, the element $a_{k}:=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{w\in W_{k} \textrm{ | } a_{w}^{k}\in I_{k}} a^{k}_{w}}$ belongs to $I_{k}$, from which it follows that $I_{k}=(a_{k})\downarrow=(a_{k})\downarrow_{J}$ is a principal $J$-ideal. Now, the fact that $l_{{\cal C}}^{J}:{\cal C}\to \Sh({\cal C}, J)$ is full and faithful, preserves the initial object and finite meets, combined with the fact that $(c)\downarrow_{J}=l_{{\cal C}}^{J}(c)$ for any $c\in {\cal C}$, ensures that the sieve $S$ satisfies the condition in the definition of $({\cal C}, J)$-adequate invariant with respect to the family $\cal F$.
The fact that if a family $\cal F$ of principal $J$-ideals on $\cal C$ satisfies the condition in the definition of $({\cal C}, J)$-adequate invariant then $\cal F$ has a disjunctive refinement follows similarly from the same arguments.
$(vi)$ We only prove the finitary version of the result, the infinitary one being entirely analogous. Given a family ${\cal F}$ of principal $J$-ideals on $\cal C$, if this family has a finite atomic refinement then either $\cal F$ is a singleton $(c)\downarrow_{J}$, in which case the maximal sieve on $c$ satisfies the condition of $({\cal C}, J)$-adequate invariant with respect to the family $\cal F$, or there exists a family $\{I_{k} \textrm{ | } k\in K\}$ of $J$-ideals such that for any $k\in K$ $I_{k}$ is an atom in $\Sh({\cal C}, J)$ and $I_{k}\subseteq (c_{k})\downarrow_{J}$ for some $c_{k}\in {\cal C}$ such that $(c_{k})\downarrow_{J}$ belongs to $\cal F$. First, we observe that for every $k\in K$, $I_{k}$ is a principal $J$-ideal; indeed, this follows immediately from the fact that $I_{k}$ is an atom, since every $J$-ideal can be expressed as a union of principal ideals. Now, suppose that, for any $k\in K$, $I_{k}=(a_{k})\downarrow_{J}$. It is immediate to see (by using the fact that $l_{{\cal C}}^{J}:{\cal C}\to \Sh({\cal C}, J)$ is full and faithful and preserves the initial object) that for any object $c$ in $\cal C$, $l_{{\cal C}}^{J}(c)=(c)\downarrow_{J}$ is an atom in $\Sh({\cal C}, J)$ if and only if $c$ is an atom in $\cal C$. Thus the elements $a_{k}$ are atoms in $\cal C$; so the supremum $a$ of the $a_{k}$ in $\cal C$ exists and the sieve $\{a_{k}\mono a \textrm{ | } k\in K \}$ is $J$-covering and satisfies the condition in the definition of $({\cal C}, J)$-adequate invariant with respect to the family $\cal F$.
The converse implication is clear for the same reasons.
$(vii)$ The proof is entirely analogous to that of $(vi)$, the key point being that, under the hypotheses of the theorem, $l_{{\cal C}}^{J}(c)=(c)\downarrow_{J}$ is a supercompact object of $\Sh({\cal C}, J)$ if and only if $c$ is an supercompact object of $\cal C$.
\end{proofs}
Note that in fact part $(iv)$ of Theorem \ref{thmadequate} subsumes both parts $(i)$ and $(ii)$ of the theorem (obtained respectively as the particular cases of $(iv)$ for $k=\omega$ and $k=2$).
\begin{corollary}\label{cor}
Let $({\cal C}, J)$ be a site.
\begin{enumerate}[(i)]
\item If $\cal C$ is a coherent category and $J$ is the coherent topology on $\cal C$ then a $J$-ideal $I$ in $\cal C$ is principal if and only if it is compact in $\Sh({\cal C}, J)$;
\item If $\cal C$ is a regular category and $J$ is the regular topology on $\cal C$ then a $J$-ideal $I$ in $\cal C$ is principal if and only if it is supercompact in $\Sh({\cal C}, J)$;
\item If $\cal C$ is any category and $J$ is the trivial topology on $\cal C$ then a $J$-ideal $I$ in $\cal C$ is principal if and only if it is supercompact in $\Sh({\cal C}, J)$;
\item If $\cal C$ is a $k$-geometric category, for a regular cardinal $k$, and $J$ is the $k$-covering topology on $\cal C$ then a $J$-ideal $I$ in $\cal C$ is principal if and only if it is $k$-compact in $\Sh({\cal C}, J)$;
\item If $\cal C$ is a disjunctively distributive lattice (resp. a disjunctively distributive frame) and $J$ is the disjunctive topology (resp. the infinitary disjunctive topology) on $\cal C$ then a $J$-ideal $I$ in $\cal C$ is principal if and only if it is disjunctively compact (resp. infinitary disjunctively compact) in $\Sh({\cal C}, J)$;
\item If $\cal C$ is a a weakly atomic meet-semilattice (resp. an infinitarily weakly atomic meet-semilattice) and $J$ is the atomically generated topology (resp. the infinitary atomically generated topology) on $\cal C$ then a $J$-ideal $I$ in $\cal C$ is principal if and only if it is atomically compact (resp. infinitarily atomically compact) in $\Sh({\cal C}, J)$.
\item If $\cal C$ is a a weakly supercompact meet-semilattice (resp. an infinitarily weakly supercompact meet-semilattice) and $J$ is the atomically generated topology (resp. the infinitary supercompactly generated topology) on $\cal C$ then a $J$-ideal $I$ in $\cal C$ is principal if and only if it is supercompactly compact (resp. infinitarily supercompactly compact) in $\Sh({\cal C}, J)$.
\end{enumerate}
\end{corollary}
\begin{proofs}
This immediately follows from Theorem \ref{thmadequate} and Theorem \ref{equivJcomp} in view of Theorem \ref{Jcomp}, Remark \ref{refin} and Examples \ref{excomp}.
\end{proofs}
We remark that the particular case of Corollary \ref{cor}(i) for distributive lattices (i.e. preorder coherent categories) already appears in the proof of Proposition II 3.2 \cite{stone}.
The invariants considered in the Corollary have all the property of being $({\cal C}, J_{\cal C})$-adequate for a given class of sites $({\cal C}, J_{\cal C})$ associated to the structures in a certain category $\cal K$ as in section \ref{locales}. This motivates the following definition.
\begin{definition}
Let $\cal K$ be a category of structures as in section \ref{locales}, where each of the structures $\cal C$ in $\cal K$ is equipped with a Grothendieck topology $J_{\cal C}$, and let $C$ be an invariant property of families of subterminals in a topos.
The invariant $C$ is said to be \emph{$\cal K$-adequate} is it is $({\cal C}, J_{\cal C})$-adequate for every structure $\cal C$ in $\cal K$.
\end{definition}
\begin{examples}\label{excomp2}
\begin{enumerate}[(a)]
\item The invariant `to be finite' is $\textbf{DLat}$-adequate, where $\textbf{DLat}$ is the category of distributive lattices;
\item The invariant `to be a singleton' is $\textbf{MSLat}$-adequate, where $\textbf{MSLat}$ is the category of meet-semilattices;
\item The invariant `to be of cardinality $\lt k$' (for a regular cardinal $k$) is $k\textrm{-}\textbf{Frm}$-adequate, where $k\textrm{-}\textbf{Frm}$ is the category of $k$-frames;
\item The invariant `to be finite and disjoint' is $\textbf{DJLat}$-adequate, where\\ $\textbf{DJLat}$ is the category of disjunctively distributive lattices;
\item The invariant `to be disjoint' is $\textbf{DJFrm}$-adequate, where $\textbf{DJFrm}$ is the category of disjunctively distributive frames;
\item The invariant `to be a singleton or finite and consisting of atomic subobjects' is $\textbf{WAtMSLat}$-adequate, where $\textbf{WAtMSLat}$ is the category of weakly atomic meet-semilattices;
\item The invariant `to be a singleton or to consist of atomic subobjects' is $\textbf{IWAtMSLat}$-adequate, where $\textbf{WAtMSLat}$ is the category of infinitarily weakly atomic meet-semilattices;
\item The invariant `to be a singleton or finite and consisting of supercompact subobjects' is $\textbf{WScMSLat}$-adequate, where $\textbf{WScMSLat}$ is the category of weakly supercompact meet-semilattices;
\item The invariant `to be a singleton or to consist of supercompact subobjects' is $\textbf{IWScMSLat}$-adequate, where $\textbf{IWScMSLat}$ is the category of infinitarily weakly supercompact meet-semilattices.
\end{enumerate}
\end{examples}
Let $\cal K$ be a category of poset structures, as in section \ref{locales}.
If $C$ is a $\cal K$-adequate invariant then, by Theorem \ref{equivJcomp}, every poset structure $\cal C$ in $\cal K$ can be recovered up to isomorphism from the corresponding topos $\Sh({\cal C}, J_{\cal C})$ as the poset of subterminals which are $C$-compact. In fact, if the invariant $C$ satisfies the property that for any structure $\cal C$ in $\cal K$ and for any family $\cal F$ of principal $J_{\cal C}$-ideals on $\cal C$, $\cal F$ has a refinement satisfying $C$ (if and) only if it has a refinement satisfying $C$ made of principal $J_{\cal C}$-ideals on $\cal C$ (cf. Theorem \ref{construction} below) then we can identify a stronger invariant property which is satisfied by all the embeddings ${\cal C} \hookrightarrow Id_{J_{\cal C}}({\cal C})$ (for $\cal C$ in $\cal K$), namely the fact that every covering of a principal $J_{\cal C}$-ideal on $\cal C$ in the topos $\Sh({\cal C}, J_{\cal C})$ is refined by a covering of principal $J_{\cal C}$-ideals on $\cal C$ which satisfy $C$; this follows by an inspection of the proof of Theorem \ref{Jcomp}, recalling the definition of $({\cal C}, J_{\cal C})$-adequate invariant. Therefore, if an embedding $B_{L}\hookrightarrow L$ is isomorphic to any of the embeddings ${\cal C} \hookrightarrow Id_{J_{\cal C}}({\cal C})$, so that $B_{L}$ is the subset of $C$-compact elements of $L$, then we can conclude that it satisfies the property that every covering in $L$ of an element of $B_{L}$ is refined by a covering made of elements of $B_{L}$ which satisfies the invariant $C$. Notice also that all the embeddings $i_{\cal C}:{\cal C} \hookrightarrow Id_{J_{\cal C}}({\cal C})$ satisfy the property that the $J_{\cal C}$-covering sieves are sent by $i_{\cal C}$ to covering families in $Id_{J_{\cal C}}({\cal C})$.
Given a category $\cal K$ and a functor $A:{\cal K}^{\textrm{op}}\to \textbf{Loc}$ as in section \ref{locales}, in order to use a $\cal K$-adequate invariant $C$ for achieving an intrinsic characterization of the locales in the extended image of $A$, and to define a functor on this subcategory which is an inverse (up to isomorphism) to the functor $A$, we need $C$ to satisfy an additional property, namely the fact that for any locale $L$ with a basis $B_{L}$ of $C$-compact elements which, regarded as a preorder with the induced order, belongs to $\cal K$, and such that the embedding $B_{L}\hookrightarrow L$ possibly satisfies some invariant properties that are known to hold for the embeddings ${\cal C} \hookrightarrow Id_{J_{\cal C}}({\cal C})$, the topology $J^{L}_{can}|_{B_{L}}$ coincides with the topology $J_{B_{L}}$ with which $B$ comes equipped as a structure in $\cal K$, where $J^{L}_{can}$ is the canonical topology on $L$. Indeed, if this is the case then, regarding $B_{L}$ as full preorder subcategory of $L$, the Comparison Lemma yields an equivalence $\Sh(L)\simeq \Sh(B_{L}, J^{L}_{can}|_{B_{L}})$, and hence, $J^{L}_{can}|_{B_{L}}$ being equal to $J_{B_{L}}$, $L$ is isomorphic to the image $A(B_{L})$ of $B_{L}$ under the functor $A$.
This motivates the following definition. Below, given a locale $L$, we denote by $J_{can}^{L}$ the canonical Grothendieck topology on $L$.
\begin{definition}
Let $\cal K$ be a category of poset structures as in section \ref{locales}, where each of the structures $\cal C$ in $\cal K$ is equipped with a Grothendieck topology $J_{\cal C}$, and let $P$ be an invariant property of embeddings $B_{L}\hookrightarrow L$ of a basis $B_{L}$ of a frame $L$ into $L$ which holds for all the canonical embeddings ${\cal C}\hookrightarrow Id_{J_{\cal C}}({\cal C})$, where $\cal C$ is a structure in $\cal K$.
A topos-theoretic invariant property $C$ of families of subterminals in a locally small cocomplete topos is said to be \emph{$\cal K$-compatible relative to $P$} if for any frame $L$ with a basis $B_{L}$ of $C$-compact elements which, regarded as a poset with the induced order, belongs to $\cal K$, if the embedding $B_{L}\hookrightarrow L$ satisfies
\begin{enumerate}[(i)]
\item property $P$;
\item the property that every covering in $L$ of an element of $B_{L}$ is refined by a covering satisfying $C$ made of elements of $B_{L}$, and
\item the property that the $J_{B_{L}}$-covering sieves are sent by the embedding $B_{L}\hookrightarrow L$ into covering families in $L$, where $J_{B_{L}}$ is the Grothendieck topology with which $B_{L}$ comes equipped as a structure in $\cal K$
\end{enumerate}
then $J_{B_{L}}$ is equal to the induced Grothendieck topology $J_{can}^{L}|_{B_{L}}$ on $B_{L}$.
An invariant $C$ is said to be \emph{$\cal K$-compatible} if it is $\cal K$-compatible relative to $P$ for some invariant property $P$ as above.
\end{definition}
It is easy to verify that all the invariants considered above are compatible with respect to the corresponding categories of structures.
Anyway, there is a systematic way for building, given a category $\cal K$ of poset structures as in section \ref{locales}, invariants which are both $\cal K$-adequate and $\cal K$-compatible. This method is based on a formalization of the vague idea of a Grothendieck topology defined `through a topos-theoretic invariant', as given by the following definition.
\begin{definition}\label{inducedtop}
Let $\cal K$ be a category of poset structures as in section \ref{locales}, where each of the structures $\cal C$ in $\cal K$ is equipped with a Grothendieck topology $J_{\cal C}$, let $P$ be an invariant property of embeddings $B_{L}\hookrightarrow L$ of a basis $B_{L}$ of a frame $L$ into $L$ which holds for all the canonical embeddings ${\cal C}\hookrightarrow Id_{J_{\cal C}}({\cal C})$ (where $\cal C$ is a structure in $\cal K$), and $C$ be a topos-theoretic invariant property of families of subterminals in a locally small cocomplete topos.
Given a structure $\cal C$ in $\cal K$, the Grothendieck topology $J_{\cal C}$ is said to be \emph{$C$-induced relative to $P$} if for any $J_{can}^{L}$-dense monotone embedding $i:{\cal C} \hookrightarrow L$ into a frame $L$ which satisfies property $P$ and such that the $J_{\cal C}$-covering sieves on $\cal C$ are sent by $i$ to covering families in $L$, for any family $\cal A$ of objects in $\cal C$ there exists a $J_{\cal C}$-covering sieve on an object $c\in {\cal C}$ such that the arrows $a\mono c$ for $a\in {\cal A}$ generate $S$ if and only if the image $i(\cal A)$ of the family $\cal A$ in $L$ has a refinement satisfying $C$ made of objects of the form $i(c')$ (for $c'\in {\cal C}$).
The Grothendieck topology $J_{\cal C}$ is said to be \emph{$C$-induced} if it is $C$-induced relative to $P$ for some invariant property $P$ as above.
\end{definition}
\begin{remark}\label{chargrot}
Under the hypotheses of Definition \ref{inducedtop}, if $J_{\cal C}$ is $C$-induced then the $J_{\cal C}$-covering sieves on an object $c\in {\cal C}$ can be characterized precisely as the sieves on $c$ in $\cal C$ which contain a family of arrows $\cal A$ with codomain $c$ such that the family $\{i(dom(f)) \textrm{ | } f\in {\cal A}\}$ satisfies $C$.
\end{remark}
Let us give some examples of topologies induced (in our sense) by topos-theoretic invariants.
\begin{theorem}\label{induced}
Let $({\cal C}, J)$ be a site. Then
\begin{enumerate}[(i)]
\item If $\cal C$ is a distributive lattice and $J$ is the coherent topology on $\cal C$ then $J$ is $C$-induced where $C$ is the property `to be a finite family';
\item If $\cal C$ is any poset and $J$ is the trivial topology on $\cal C$ then $J$ is $C$-induced where $C$ is the property `to be a singleton family';
\item If $\cal C$ is a $k$-frame (for a regular cardinal $k$) and $J$ is the $k$-covering topology on $\cal C$ then $J$ is $C$-induced where $C$ is the property `to be a $k$-family (i.e. a family of cardinality $\lt k$)';
\item If $\cal C$ is a disjunctively distributive lattice (resp. a disjunctively distributive frame) and $J$ is the disjunctive topology (resp. the infinitary disjunctive topology) on $\cal C$ then $J$ is $C$-induced where $C$ is the property `to be a finite and disjont (i.e. whose elements are pairwise disjoint subobjects) family' (resp. `to be a disjoint family');
\item If $\cal C$ is a weakly atomic meet-semilattice (resp. an infinitarily weakly atomic meet-semilattice) and $J$ is the atomically generated topology (resp. the infinitary atomically generated topology) on $\cal C$ then $J$ is $C$-induced where $C$ is the property `to be a singleton family or a finite and atomic (i.e. whose elements are atoms) family' (resp. `to be singleton family or an atomic family');
\item If $\cal C$ is a weakly supercompact meet-semilattice (resp. an infinitarily weakly supercompact meet-semilattice) and $J$ is the supercompactly generated topology (resp. the infinitary atomically generated topology) on $\cal C$ then $J$ is $C$-induced where $C$ is the property `to be a singleton family or a finite and supercompact (i.e. whose elements are supercompact subobjects) family' (resp. `to be a singleton family or a supercompact family').
\end{enumerate}
\end{theorem}
\begin{proofs}
$(i)$, $(ii)$ and $(iii)$. We prove $(iii)$, since $(i)$ and $(ii)$ are particular cases of it. The topology $J$ is clearly $C$-induced relative to $P$ where $P$ is the vacuous property.
$(iv)$ The topology $J$ is easily seen to be $C$-induced relative to $P$ where $P$ is the following property of embeddings $i:B_{L}\hookrightarrow L$: `$i$ preserves finite meets' or equivalently (since the canonical embedding ${\cal C}\hookrightarrow Id_{J}({\cal C})$ satisfies the property that ${\cal C}$ is closed under finite meets in $Id_{J}({\cal C})$) the property `$i(B_{L})$ is closed in $L$ under finite meets'.
$(v)$ The topology $J$ is easily seen to be $C$-induced relative to $P$ where $P$ is the vacuous property.
$(vi)$ This is entirely analogous to $(v)$; again, $P$ can be taken to be the vacuous property.
\end{proofs}
\begin{theorem}\label{construction}
Let $\cal K$ be a category of poset structures as in section \ref{locales}. If all the Grothendieck topologies $J_{\cal C}$ associated to the structures $\cal C$ in $\cal K$ are $C$-induced (for a topos-theoretic invariant $C$ of families of subterminals in a locally small cocomplete topos) and the invariant $C$ satisfies the property that for any structure $\cal C$ in $\cal K$ and for any family $\cal F$ of principal $J_{\cal C}$-ideals on $\cal C$, $\cal F$ has a refinement satisfying $C$ (if and) only if it has a refinement satisfying $C$ made of principal $J_{\cal C}$-ideals on $\cal C$ then the invariant $C$ is both $\cal K$-adequate and $\cal K$-compatible.
\end{theorem}
\begin{proofs}
Let us first show that the invariant $C$ is $\cal K$-adequate. Given a structure $\cal C$ in $\cal K$, consider the canonical embedding ${\cal C}\hookrightarrow Id_{J_{\cal C}}({\cal C})$. Given a family $\cal F$ of principal $J_{\cal C}$-ideals on $\cal C$, we have to prove that $\cal F$ has a refinement satisfying $C$ if and only if there exists a $J$-covering sieve $\{f_{h}:d_{h}\to d \textrm{ | } h\in H\}$ in $\cal C$ such that for any $h\in H$, $(d_{h})\downarrow_{J}$ is contained in some ideal in $\cal F$ and the union in $\cal E$ of the family $\{(d_{h})\downarrow_{J} \textrm{ | } h\in H\}$ is equal to the union in $\cal E$ of the family $\cal F$. Now, by our hypothesis on $C$, $\cal F$ has a refinement satisfying $C$ if and only if $\cal F$ has a refinement satisfying $C$ made of principal $J_{\cal C}$-ideals on $\cal C$, and, $J_{{\cal C}}$ being $C$-induced, this latter condition is equivalent to the existence of a $J$-covering sieve $\{f_{h}:d_{h}\to d \textrm{ | } h\in H\}$ with the required property.
To prove that $C$ is $\cal K$-adequate, we show that $C$ is $\cal K$-adequate relative to $P$, where $P$ is the intersection of all the properties witnessing the fact that the topologies $J_{\cal C}$ are $C$-induced, so that all the Grothendieck topologies $J_{\cal C}$ are $C$-induced relative to $P$. Let $L$ be a frame with a basis $B_{L}$ of $C$-compact elements which, regarded as a preorder with the induced order, is equal to a structure $\cal C$ in $\cal K$; suppose moreover that the embedding $i:{\cal C}=B_{L}\hookrightarrow L$ satisfies property $P$, the property that every covering in $L$ of an element of $B_{L}$ is refined by a covering made of elements of $B_{L}$ which satisfies the invariant $C$, and the property that the $J_{B_{L}}$-covering sieves are sent by the embedding $B_{L}\hookrightarrow L$ into covering families in $L$ (where $J_{B_{L}}$ is the Grothendieck topology with which $B_{L}$ comes equipped as a structure in $\cal K$); we have to prove that $J_{\cal C}=J^{L}_{can}|_{\cal C}$, where $J^{L}_{can}$ is the canonical topology on $L$. Now, a sieve $S$ on an object $c\in {\cal C}$ is $J^{L}_{can}|_{\cal C}$-covering if and only if the sieve in $L$ generated by the image $i(S)$ of the arrows in $S$ under the embedding $i$ generates a $J^{L}_{can}$-covering sieve on $i(c)$ in $L$; but, by our hypotheses on the embedding $i$, this condition holds if and only if there exists a family of arrows of the form $\{i(c')\leq i(c)\}$ (for $c'\in {\cal C}$) which refines the family $\{i(dom(f))\leq i(c) \textrm{ | } f\in S\}$ and satisfies the invariant $C$. On the other hand, by Remark \ref{chargrot}, for any object $c\in {\cal C}$, the $J_{\cal C}$-covering sieves on $c$ are precisely those which contain a family of arrows $\cal A$ in $\cal C$ with codomain $c$ such that the family $\{i(dom(f)) \textrm{ | } f\in {\cal A}\}$ satisfies $C$. From this, the equality $J_{\cal C}=J^{L}_{can}|_{\cal C}$ is clear.
\end{proofs}
\begin{remark}\label{rmkprincipal}
All the invariants $C$ appearing in the statement of Theorem \ref{induced} satisfy the property in the statement of Theorem \ref{construction}. This is clear for the invariants in parts $(i)$, $(ii)$ and $(iii)$ of the theorem, and it follows from the proofs of the relevant parts of Theorem \ref{thmadequate} for the other ones.
\end{remark}
Theorem \ref{construction} is important because it assures that, whenever the subcanonical Grothendieck topologies $J_{\cal C}$ corresponding to the structures in a given category $\cal K$ are `uniformly defined through a topos-theoretic invariant' (in the sense of being $C$-induced for an invariant $C$ satisfying the hypotheses of Theorem \ref{construction}), there is a topos-theoretic invariant, namely the property of $C$-compactness, which enables us to recover each structure $\cal C$ from the corresponding topos $\Sh({\cal C}, J_{\cal C})$ up to isomorphism, and to define a functor $I_{A}$ on the extended image of the functor $A:{\cal K}^{\textrm{op}}\to \textbf{Loc}$ of section \ref{locales} which is a categorical inverse to $A$ and hence yields a duality between $\cal K$ and the subcategory of $\textbf{Loc}$ given by the extended image of $A$. Recall that $I_{A}$ is defined as the functor which acts on the objects by sending a locale $L$ in $ExtIm(A)$ to the poset of $C$-compact elements of $L$ and acts on the arrows by sending a locale morphism $f:L\to L'$ in $ExtIm(A)$ to the restriction of its associated frame homomorphism to the subsets of $C$-compact elements of $L$ and $L'$ (cf. section \ref{locales}). Summarizing, we have established the following general `duality theorem'.
\begin{theorem}\label{thmcentral}
Let $\cal K$ be a category of poset structures as in section \ref{locales}, and let $C$ be a topos-theoretic invariant of families of subterminals in a locally small cocomplete topos which satisfies the property that for any structure $\cal C$ in $\cal K$ and for any family $\cal F$ of principal $J_{\cal C}$-ideals on $\cal C$, $\cal F$ has a refinement satisfying $C$ (if and) only if it has a refinement satisfying $C$ made of principal $J_{\cal C}$-ideals on $\cal C$. If all the Grothendieck topologies $J_{\cal C}$ associated to the structures $\cal C$ in $\cal K$ are $C$-induced then the functor $A:{\cal K}^{\textrm{op}}\to \textbf{Loc}$ of section \ref{locales} admits a categorical inverse $I_{A}:ExtIm(A)\to {\cal K}^{\textrm{op}}$ sending a locale $L$ in $ExtIm(A)$ to the poset of $C$-compact elements of $L$ and a locale morphism $f:L\to L'$ in $ExtIm(A)$ to the restriction of its associated frame homomorphism to the subsets of $C$-compact elements of $L$ and $L'$.
\end{theorem}
The following result provides an `intrinsic' characterization of the category $ExtIm(A)$.
\begin{theorem}\label{propext}
Let us assume that the hypotheses of Theorem \ref{thmcentral} are satisfied, and that $C$ is $\cal K$-compatible relative to a property $P$. Then
\begin{enumerate}[(i)]
\item The locales in the extended image $ExtIm(A)$ of the functor $A:{\cal K}^{\textrm{op}}\to \textbf{Loc}$ of section \ref{locales} are precisely the locales $L$ with a basis $B_{L}$ of $C$-compact elements which, regarded as a poset with the induced order, belongs to $\cal K$, and such that the embedding $B_{L}\hookrightarrow L$ satisfies property $P$, the property that every covering in $L$ of an element of $B_{L}$ has a refinement by a covering made of elements of $B_{L}$ which satisfies the invariant $C$, and the property that the $J_{B_{L}}$-covering sieves are sent by the embedding $B_{L}\hookrightarrow L$ into covering families in $L$, where $J_{B_{L}}$ is the Grothendieck topology with which $B_{L}$ comes equipped as a structure in $\cal K$;
\item The arrows $L\to L'$ in $ExtIm(A)$ are the locale morphisms $L\to L'$ between locales in $ExtIm(A)$ whose associated frame homomorphisms send $C$-compact elements of $L'$ to $C$-compact elements of $L$ in such a way that their restriction to the subsets of $C$-compact elements of $L'$ and $L$ can be identified with an arrow in $\cal K$.
\end{enumerate}
\end{theorem}
\begin{proofs}
The characterization of the objects of $ExtIm(A)$ follows immediately from the definition of $\cal K$-compatible invariant, in light of Theorem \ref{construction}.
To characterize the arrows in $ExtIm(A)$ we observe that, by Lemma C2.3.8 \cite{El}, since the Grothendieck topologies $J_{\cal C}$ are subcanonical, the geometric morphisms $l:\Sh({\cal D}, J_{{\cal D}}) \to \Sh({\cal C}, J_{{\cal C}})$ which are induced (uniquely up to isomorphism) by a morphism of sites $({\cal C}, J_{{\cal C}}) \to ({\cal D}, J_{{\cal D}})$ can be characterized precisely as those such that the inverse image $l^{\ast}$ sends representables on $\cal C$ to representables on $\cal D$. Now, under the equivalences $\Sh({\cal C}, J_{\cal C}) \simeq \Sh(Id_{J_{\cal C}}({\cal C}))$ and $\Sh({\cal D}, J_{\cal D}) \simeq \Sh(Id_{J_{\cal D}}({\cal D}))$ of Theorem \ref{fund}, these geometric morphisms correspond exactly to the frame homomorphisms ${\cal O}(Id_{J_{\cal C}}({\cal C})) \to {\cal O}(Id_{J_{\cal D}}({\cal D}))$ which send principal ($J_{\cal C}$-)ideals on $\cal C$ to principal ($J_{\cal D}$-)ideals on $\cal D$. Therefore, if $C$ is $\cal K$-compatible then the arrows in $\cal U$ can be characterized precisely as the locale morphisms whose associated frame homomorphisms send $C$-compact elements to $C$-compact elements in such a way that their restriction to the subsets of $C$-compact elements can be identified with an arrow in $\cal K$.
\end{proofs}
Note that if, for any structures $\cal C$ and $\cal D$ in $\cal K$, the arrows ${\cal C}\to {\cal D}$ in $\cal K$ coincide exactly with the morphisms of sites $({\cal C}, J_{\cal C})\to ({\cal D}, J_{\cal D})$ then the condition `in such a way that their restriction to the subsets of $C$-compact elements of $L'$ and $L$ can be identified with an arrow in $\cal K$' in the statement of Theorem \ref{propext} can be omitted, since it is automatically satisfied (cf. section \ref{locales} above).
A similar characterization holds for the extended images of the covariant functors $B:{\cal K}\to \textbf{Loc}$ defined in section \ref{locales} (cf. section \ref{prealex} below).
\begin{remark}
A basis $B_{L}\hookrightarrow L$ satisfying the hypotheses of Theorem \ref{propext} is unique if it exists. Indeed, $C$ being $\cal K$-adequate, for any $\cal C$ in $\cal K$ the elements of $Id_{J_{\cal C}}({\cal C})$ which are principal $J_{\cal C}$-ideals on $\cal C$ are precisely those which are $C$-compact, and hence $B_{L}$ must be equal to the subset of $L$ consisting of its $C$-compact elements.
\end{remark}
The following proposition is often useful for achieving explicit descriptions of the condition in Theorem \ref{propext} that a basis $B_{L}$ of a locale $L$ in $ExtIm(A)$ should have the structure of a poset in $\cal K$ (cf. section \ref{ex} below for some concrete applications of this result).
\begin{proposition}\label{multicomposition}
Let $C$ be a topos-theoretic invariant property of families of subterminals in a locally small cocomplete topos which is multicomposition-stable in the sense that, for any family ${\cal F}:=\{a_{i} \textrm{ | } i\in I\}$ of subterminals in a locally small cocomplete topos $\cal E$ and any collection of families ${\cal F}_{i}:=\{b_{ij} \leq a_{i} \textrm{ | } j\in I_{j}\}$ of subterminals in $\cal E$ indexed over a set $I$, if $\cal F$ and all the ${\cal F}_{i}$ (for $i\in I$) satisfy $C$ then the `multicomposite' family ${\cal F} \ast \{{\cal F}_{i} \textrm{ | } i\in I\}:=\{b_{ij}, i\in I, j\in I_{j}\}$ also satisfies $C$. Then for any family $\cal G$ of subterminals in a locally small cocomplete topos $\cal E$ which satisfies $C$, if all the objects in $\cal G$ are $C$-compact then the union in $\cal E$ of all the subterminals in $\cal G$ is again $C$-compact.
\end{proposition}
\begin{proofs}
Let ${\cal G}=\{c_{i} \textrm{ | }\in I\}$ be a family of subterminals in a locally small cocomplete topos $\cal E$ such that all the $c_{i}$ (for $i\in I$) are $C$-compact, and let $c=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}c_{i}}$ be the union in $\cal E$ of all the subterminals in $\cal G$. Given a covering $\{u_{k} \textrm{ | }k\in K\}$ of $c$ in $\cal E$, we have, for any $i\in I$, $c_{i}=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{k\in K}(c_{i}\wedge u_{k})}$. This implies, since the $c_{i}$ are $C$-compact, that there exists, for each $i\in I$, a family ${\cal G}_{i}$ satisfying $C$ which refines the family $\{c_{i}\wedge u_{k} \textrm{ | } k\in K\}$. Then the fact that $C$ is multicomposition-stable implies that the family ${\cal G} \ast \{{\cal G}_{i} \textrm{ | } i\in I\}$ satisfies $C$; but this family clearly refines the family $\{u_{k} \textrm{ | }k\in K\}$, from which our thesis follows.
\end{proofs}
Note that all the invariants that we have considered in this section are multicomposition-stable.
\section{Old and new dualities}\label{ex}
In this section, we give some applications of the `machinery' for generating dualities or equivalences described in the previous sections. Specifically, we discuss how well-known dualities can be naturally recovered from our machinery and we establish new dualities and equivalences as applications of our method.
\subsection{Preorders and Alexandrov topologies}\label{prealex}
The well-known equivalence between preorders and Alexandrov spaces falls under the category of covariant equivalences obtainable by the technique of sections \ref{locales} and \ref{dualtop}.
Let $\textbf{Pro}$ be the category of preorders and monotone maps between them. The construction of covariant equivalences given in section \ref{locales} yields a functor $B:\textbf{Pro} \to \textbf{Loc}$, which assigns to a preorder $\cal P$ the frame $Id({{\cal P}^\textrm{op}})$ of upper sets in $\cal P$ and to an arrow $f:{\cal P}\to {\cal Q}$ in $\textbf{Pro}$ the frame homomorphism $B(f):Id({{\cal Q}^\textrm{op}}) \to Id({{\cal P}^\textrm{op}})$ sending an upper set $S$ in $\cal Q$ to the upper set in $\cal P$ given by the inverse image $f^{-1}(S)$ of $S$ under $f$.
We shall call the image $B({\cal P})=Id({{\cal P}^\textrm{op}})$ of a preorder $\cal P$ under the functor $B$ the \emph{Alexandrov locale} associated to $\cal P$.
By Corollary \ref{cor}(iii), any poset $\cal P$ is isomorphic in $\textbf{Pro}$ to the opposite of the set of supercompact subterminals of the topos $[{\cal P}, \Set]$ endowed with the order induced by that in $\Sub_{[{\cal P}, \Set]}(1)$. In fact, this can be made functorial, as follows. Given posets $\cal P$ and $\cal Q$, and an arrow $f:{\cal P}\to {\cal Q}$ in $\textbf{Pro}$, denoted by $E(f):[{\cal P}, \Set] \to [{\cal Q}, \Set]$ the geometric morphism induced by $f$ as in Example A4.1.4 \cite{El}, we have (cf. the proof of Lemma A4.1.5 \cite{El}) that the left adjoint to the inverse image functor $E(f)^{\ast}:[{\cal Q}, \Set] \to [{\cal P}, \Set]$ of $E(f)$ restricts to the subsets of supercompact subterminals of $[{\cal Q}, \Set]$ and of $[{\cal P}, \Set]$ yielding a map which is isomorphic to the opposite $f^{\textrm{op}}:{\cal P}^{\textrm{op}}\to {\cal Q}^{\textrm{op}}$ of the map $f$.
Let $\textbf{Pos}$ denote the full subcategory of $\textbf{Pro}$ on the posets. Then the argument above ensures that the restriction $B_{\textbf{Pos}}:\textbf{Pos} \to \textbf{Loc}$ of $B:\textbf{Pro} \to \textbf{Loc}$ is faithful (and injective on objects) and hence yields an isomorphism of categories between $\textbf{Pos}$ and a subcategory of the category $\textbf{Loc}$ of locales, namely the \emph{image} of $B_{\textbf{Pos}}$. This represents the analogue for covariant functors of Theorem \ref{teoabstr}.
Anyway, $B_{\textbf{Pos}}:\textbf{Pos} \to \textbf{Loc}$ also yields an equivalence of categories onto its extended image $ExtIm(B_{\textbf{Pos}})=ExtIm(B)$, as follows.
The inverse functor $I_{B}:ExtIm(B) \to \textbf{Pos}$ sends a locale $L$ in $ExtIm(B)$ to the opposite of the poset of supercompact elements of ${\cal O}(L)$ and a morphism $g:L\to L'$ of locales in $ExtIm(B)$ to the opposite of the restriction of the left adjoint $g_{!}:{\cal O}(L) \to {\cal O}(L')$ to $g^{\ast}:{\cal O}(L') \to {\cal O}(L)$ to the posets of supercompact elements of the frames ${\cal O}(L)$ and ${\cal O}(L')$. Note that the left adjoint $g_{!}$ has the following expression in terms of $g^{\ast}$: $g_{!}(l)=\mathbin{\mathop{\textrm{\huge $\wedge$}}\limits_{l\leq g^{\ast}(l')}l'}$.
We can read the equivalence $\textbf{Pos} \simeq ExtIm(B)$ just established more naturally by composing it with the isomorphism of categories $\textbf{Pos}\cong \textbf{Pos}$ sending every poset $\cal P$ in $\textbf{Pos}$ to the opposite poset ${\cal P}^{\textrm{op}}$, and every arrow $f:{\cal P}\to {\cal Q}$ to the opposite arrow $f^{\textrm{op}}:{\cal P}^{\textrm{op}}\to {\cal Q}^{\textrm{op}}$. The resulting equivalence $\textbf{Pos} \simeq ExtIm(B)$ sends a poset $\cal P$ in $\textbf{Pos}$ to its ideal completion $Id({\cal P})$ (i.e., the set of all the ideals of $\cal P$ with the subset-inclusion order) and a locale $L$ in $ExtIm(B)$ to the subset of supercompact elements of $L$ endowed with the induced natural order.
Let us now turn to the problem of giving an intrinsic characterization of the category $ExtIm(B)$. We shall call the locales in $ExtIm(B)$ the \emph{Alexandrov locales}.
\begin{proposition}\label{alexlocale}
Let $L$ be a locale. Then the following conditions are equivalent:
\begin{enumerate}[(i)]
\item $L$ is an Alexandrov locale;
\item $L$ has a basis of supercompact elements;
\item $\Sh(L)$ is equivalent to a presheaf topos of the form $[{\cal P}, \Set]$, where $\cal P$ is a preorder (equivalently, a poset) category.
\end{enumerate}
\end{proposition}
\begin{proofs}
$(i)\imp (ii)$ It is clear that if $L$ is an Alexandrov locale then it has a basis of supercompact elements; indeed, this property is a locale-theoretic invariant, and if $L$ is of the form $Id({{\cal P}^\textrm{op}})$ for a preorder $\cal P$ then the collection of all the principal upper sets generated by an element of $\cal P$ forms a basis of $L$.
$(ii)\imp (iii)$ If $L$ has a basis $B$ of supercompact elements then $B$, regarded as a full subcategory of $L$, is $J^{L}_{can}$-dense (where $J^{L}_{can}$ is the canonical Grothendieck topology on $L$), and hence $\Sh(L)\simeq \Sh(B, J^{L}_{can}|_{B})$ by the Comparison Lemma. But since every element of $B$ is supercompact then $J^{L}_{can}|_{B}$ is the trivial Grothendieck topology on $\cal B$ and hence $\Sh(L)\simeq \Sh(B, J^{L}_{can}|_{B})\simeq [B^{\textrm{op}}, \Set]$.
$(iii) \imp (i)$ If $\Sh(L)\simeq [{\cal P}, \Set]$ then, since $[{\cal P}, \Set]\simeq \Sh(Id({{\cal P}^\textrm{op}})$ (cf. Theorem \ref{fund}), $L$ is isomorphic to $Id({{\cal P}^\textrm{op}})$ and therefore it is an Alexandrov locale.
\end{proofs}
Concerning the arrows in $ExtIm(B)$, it is clear from the discussion preceding Proposition \ref{alexlocale} that they are precisely the morphisms $g:L\to L'$ of locales in $ExtIm(B)$ such that $g^{\ast}:{\cal O}(L') \to {\cal O}(L)$ has a left adjoint $g_{!}:{\cal O}(L) \to {\cal O}(L')$ (equivalently, $g^{\ast}$ preserves arbitrary infima) which sends supercompact elements to supercompact elements. Let us call these morphisms the complete maps. In fact, the condition that $g_{!}$ should send supercompact elements to supercompact elements is superfluous, since it is automatically satisfied (cf. section \ref{comparison} below).
Let $\textbf{AlexLoc}$ denote the category whose objects are the Alexandrov locales and whose arrows are the complete maps between them. We have thus obtained the following result.
\begin{theorem}\label{proalex}
Via the functors $B_{\textbf{Pos}}:\textbf{Pos}\to \textbf{AlexLoc}$ and\\ $I_{B}:\textbf{AlexLoc} \to \textbf{Pos}$ defined above, the category $\textbf{Pos}$ is equivalent to the category $\textbf{AlexLoc}$.
\end{theorem}
\begin{remarks}
\begin{enumerate}[(a)]
\item The functor $B_{\textbf{Pos}}$ in the statement of Theorem \ref{proalex} corresponds, under the involution $\textbf{Pos}\to \textbf{Pos}$ sending a poset to its dual, to the functor giving one half of the duality between posets and completely distributive algebraic lattices described in \cite{duality}; in fact, it follows from Theorem \ref{proalex} and the duality theorem in \cite{duality} that the category $\textbf{AlexLoc}$ defined above coincides, when considered as a category of frames, with the category of completely distributive algebraic lattices defined in \cite{duality} (since the two theorems ensure that both these categories are equal to the extended image of the same functor, namely $B_{\textbf{Pos}}:\textbf{Pos}^{\textrm{op}} \to \textbf{Frm}$); anyway, it is worth to note that our definition of the inverse functor $I_{B}:\textbf{AlexLoc} \to \textbf{Pos}$ is completely different from the construction of the inverse functor given in \cite{duality};
\item Clearly, every finite distributive lattice is a frame with a basis of supercompact elements; clearly, the posets corresponding to these frames via the equivalence of Theorem \ref{proalex} are precisely the finite ones. The equivalence of Theorem \ref{proalex} thus restricts to Birkhoff duality between finite distributive lattices and finite posets.
\end{enumerate}
\end{remarks}
The functor $B:\textbf{Pro} \to \textbf{AlexLoc}$ from preorders to Alexandrov locales can be lifted to an equivalence between $\textbf{Pro}$ and a subcategory of the category $\textbf{Top}$ of topological spaces in various ways, according to the method of section \ref{dualtop}.
Let us choose, for each preorder ${\cal P}$, as separating set of points of the topos $[{\cal P}, \Set]$, the collection of geometric morphisms $\{e_{p}:\Set\to [{\cal P}, \Set] \textrm{ | } p\in P\}$ indexed by the underlying set $P$ of $\cal P$ as in Example \ref{exa}(b). Clearly, every arrow $f:{\cal P} \to {\cal Q}$ in $\textbf{Pro}$ yields a function $l_{f}:P \to Q$ (given by the action of $f$ on the underlying sets of the preorders) which satisfies the necessary compatibility conditions of section \ref{dualtop}. As we observed in Example \ref{exa}(b), the subterminal topology induced on the underlying set $P$ of a preorder $\cal P$ is precisely the Alexandrov topology on $\cal P$; the method of section \ref{dualtop} thus yields in this case the well-known equivalence between the category $\textbf{Pro}$ of preorders and the subcategory of $\textbf{Top}$ given by the Alexandrov spaces (note that the extended functor $\tilde{B}:\textbf{Pro}\to \textbf{Top}$ is already faithful and hence it is no longer necessary, unlike in the case of Alexandrov locales, to restrict to the category $\textbf{Pos}$ of posets to obtain an equivalence of categories).
Note that the characterization of Alexandrov locales given by Proposition \ref{alexlocale} does not lift to a topological characterization of Alexandov spaces as the sober topological spaces whose locales of open sets are Alexandrov locales, since Alexandrov spaces are not in general sober (cf. Proposition \ref{alexsober}). On the other hand, we can build an equivalence of $\textbf{Pos}$ with a category of sober topological spaces, as follows.
For any preorder $\cal P$, the topos $[{\cal P}, \Set]$, being localic, has only a \emph{set} of points (up to isomorphism). By choosing, for each preorder ${\cal P}$, as set of points of the topos $[{\cal P}, \Set]$ the collection $\textsl{S}_{\cal P}$ of \emph{all} its points and, for each arrow $f:{\cal P}\to {\cal Q}$, the function $l_{f}:\textsl{S}_{\cal P}\to \textsl{S}_{\cal Q}$ sending a point $\Set \to [{\cal P}, \Set]$ to the composite of it with the geometric morphism $[{\cal P}, \Set] \to [{\cal Q}, \Set]$ induced by $f$ as in section \ref{locales} above, we obtain, by the technique of section \ref{dualtop}, a lifting of the equivalence of $\textbf{Pos}$ with the category $ExtIm(B_{\textbf{Pos}})$ of Alexandrov locales to an equivalence $\tilde{B_{\textbf{Pos}}}:\textbf{Pos} \simeq ExtIm(\tilde{B_{\textbf{Pos}}})$ between $\textbf{Pos}$ and a subcategory of $\textbf{Top}$. By Theorem \ref{sobriety}, all the spaces $X$ in $ExtIm(\tilde{B_{\textbf{Pos}}})$ are sober and have the property that their locales of open sets ${\cal O}(X)$ are Alexandrov locales; hence, by the remarks in section \ref{dualtop}, the spaces in $ExtIm(\tilde{B_{\textbf{Pos}}})$ can be characterized precisely as the sober topological spaces whose frames of open sets are, regarded as locales, Alexandrov locales, while the arrows $X\to Y$ in $ExtIm(\tilde{B_{\textbf{Pos}}})$ are exactly the continuous maps $g:X\to Y$ between such spaces.
The topological space $\tilde{B_{\textbf{Pos}}}({\cal P})$ corresponding to a poset $\cal P$ can be described concretely as the space having as set of points the collection ${\cal F}^{dir}_{\cal P}$ of all the non-empty directed ideals on $\cal P$ and as open sets the subsets of the form
\[
{\cal F}_{U}=\{F\in {\cal F}^{dir}_{\cal P} \textrm{ | } F\cap U\neq \emptyset\},
\]
where $U$ ranges among the upper sets in $\cal P$ (cf. Proposition \ref{mslattice}). Clearly, this space is precisely the sobrification of the Alexandrov space corresponding to $\cal P$. In these terms, the arrow $\tilde{B_{\textbf{Pos}}}(f)$ corresponding to an arrow $f:{\cal P}\to {\cal Q}$ in $\textbf{Pos}$ via the functor $\tilde{B_{\textbf{Pos}}}$ can be described as the map ${\cal F}^{dir}_{\cal P}\to {\cal F}^{dir}_{\cal Q}$ sending a ideal set $I$ in ${\cal F}^{dir}_{\cal P}$ to the ideal in ${\cal F}^{dir}_{\cal Q}$ generated by the image of $I$ under $f$.
Let us denote by $\textbf{SSC}$ the subcategory of $\textbf{Top}$ having as objects the sober topological spaces with a basis of supercompact open sets and as arrows the continuous maps $f:X\to Y$ between such spaces. The inverse functor $\tilde{I_{B}}:\textbf{SSC}\to \textbf{Pos}$ sends a topological space $X$ in $\textbf{SSC}$ to the opposite of the poset of its supercompact open sets and a continuous map $f:X\to Y$ in $\textbf{SSC}$ to the opposite of the restriction of the left adjoint $f_{!}:{\cal O}(X) \to {\cal O}(Y)$ to $f^{-1}:{\cal O}(Y) \to {\cal O}(X)$ to the posets of supercompact open sets of $X$ and of $Y$. The technique of section \ref{dualtop} thus yields a functor $\tilde{I_{B_{\textbf{Pos}}}}:\textbf{SSC}\to \textbf{Pos}$ which is a categorical inverse to the functor $\tilde{B_{\textbf{Pos}}}:\textbf{Pos} \to \textbf{SSC}$. Note that $\tilde{I_{B_{\textbf{Pos}}}}$ sends a space $X$ in $\textbf{SSC}$ to the opposite of the poset of supercompact open sets of $X$ and a continuous map $f:X\to Y$ in $\textbf{SSC}$ to the opposite of the restriction of the left adjoint $f_{!}:{\cal O}(X) \to {\cal O}(Y)$ to $f^{-1}:{\cal O}(Y) \to {\cal O}(X)$ to the posets of supercompact open sets of $X$ and of $Y$.
Summarizing, have the following result.
\begin{theorem}
Via the functors $\tilde{B_{\textbf{Pos}}}:\textbf{Pos} \to \textbf{SSC}$ and $\tilde{I_{B_{\textbf{Pos}}}}:\textbf{SSC}\to \textbf{Pos}$ defined above, the category $\textbf{Pos}$ is equivalent to the category $\textbf{SSC}$.
\end{theorem}
\subsection{Stone duality for distributive lattices}\label{stonedist}
This duality falls under the category of contravariant equivalences obtainable by means of the technique of sections \ref{locales} and \ref{dualtop}.
Let $\textbf{DLat}$ be the category of distributive lattices and homomorphisms (i.e. maps preserving finite meets and finite joins) between them. We can equip each lattice $\cal D$ in $\textbf{DLat}$, regarded as a preorder coherent category, with the coherent Grothendieck topology $J_{{\cal D}}$. The arrows $f:{\cal D}\to {\cal D}'$ in $\textbf{DLat}$ can be identified with the morphisms of sites $({\cal D}, J_{\cal D}) \to ({\cal D}', J_{\cal D}')$; so, by the results of section \ref{locales}, we have a functor $A:\textbf{DLat}^{\textrm{op}}\to \textbf{Loc}$ sending a lattice $\cal D$ in $\textbf{DLat}$ to the locale $Id_{J_{\cal D}}({\cal D})$. Now, by Corollary \ref{cor}(i), any $\cal D$ in $\textbf{DLat}$ can be recovered from the topos $\Sh({\cal D}, J_{\cal D})$ as the set of its compact subterminals, endowed with the induced natural order. By Theorem \ref{induced}(i) and Remark \ref{rmkprincipal}, the hypotheses of Theorem \ref{construction} are satisfied, and hence, since by Theorem \ref{induced}(i) and Remark \ref{rmkprincipal}, the hypotheses of Theorem \ref{construction} are satisfied, we have an inverse functor $I_{A}:ExtIm(A)\to \textbf{DLat}^{\textrm{op}}$, defined on the extended image of the functor $A:\textbf{DLat}^{\textrm{op}}\to ExtIm(A)$. The functor $I_{A}$ sends a locale in $ExtIm(A)$ to the lattice of compact elements in it, and an arrow $f:L\to L'$ in $ExtIm(A)$ to the restriction of its associated frame homomorphism to the subsets of compact elements of $L$ and $L'$.
By Theorem \ref{propext}, the objects in the extended image $ExtIm(A)$ of the functor $A$ can be characterized as the locales which have a basis of compact elements which form, with respect to the induced natural order, a (distributive) lattice (equivalently, as the locales such that the collection of their compact elements is closed under finite meets and forms a basis of them), while the arrows in $ExtIm(A)$ are precisely the locale morphisms whose associated frame homomorphisms send compact elements to compact elements. The category of such locales is called by Johnstone in \cite{stone} the category of \emph{coherent locales}. The method of section \ref{dualtop} thus yields in this case the well-known duality between the category $\textbf{DLat}$ of distributive lattices and the category of coherent locales (cf. Corollary II3.3 \cite{stone}).
We shall call the locale corresponding to a distributive lattice $\cal D$ via this duality the \emph{Stone locale} associated to $\cal D$.
This duality between $\textbf{DLat}$ and the category of coherent locales can be lifted (under a form of the axiom of choice, necessary to ensure that the toposes $\Sh({\cal D}, J_{\cal D})$ have enough points) to a duality between $\textbf{DLat}$ and a subcategory of the category $\textbf{Top}$ of topological spaces, by means of the technique of section \ref{dualtop} above. Indeed, we can choose, for each lattice ${\cal D}$ in $\textbf{DLat}$, as set of points $P_{\cal D}$ of the topos $\Sh({\cal D}, J_{\cal D})$ the collection of all its points, and define the action on arrows $f:{\cal D}\to {\cal D}'$ in $\textbf{DLat}$ accordingly, as specified in section \ref{dualtop}.
In this way we obtain an equivalence between $\textbf{DLat}^{\textrm{op}}$ and a subcategory $ExtIm(\tilde{A})$ of $\textbf{Top}$, whose objects can be characterized as the sober topological spaces such that the collection of their compact open sets is closed under finite intersection and forms a basis for the topology, and whose arrows are the continuous maps between these spaces such that the inverse image of any compact open set is a compact open set. The technique of section \ref{dualtop} thus yields in this case the classical Stone duality between the category $\textbf{DLat}$ of distributive lattices and the category of spectral topological spaces (cf. Example \ref{exa}(c)). We shall call the topological space corresponding to a distributive lattice $\cal D$ via this duality the \emph{Stone space} associated to $\cal D$.
If we restrict the duality between $\textbf{DLat}$ and $ExtIm(A)$ to the full subcategory $\textbf{Boole}$ of $\textbf{DLat}$ on the Boolean algebras, we obtain an equivalence between $\textbf{Boole}^{\textrm{op}}$ and a full subcategory of $ExtIm(A)$, whose objects are the locales which have a a basis of compact elements which, with respect to the induced order, forms a Boolean algebra (equivalently, on the locales such that the collection of all their complemented elements forms a basis for the locale), and whose arrows are the locale morphisms whose associated frame homomorphisms send complemented elements to complemented elements. By lifting this duality to a topological duality as above, we obtain a duality between $\textbf{Bool}$ and the subcategory of $\textbf{Top}$ whose objects are the sober topological spaces which have a basis of clopen subsets and whose arrows are the continuous maps between such spaces; that is, we recover precisely the classical Stone duality for Boolean algebras.
\subsection{The duality between spatial frames and sober spaces}
Let $\textbf{Frm}_{sp}$ be the category of spatial frames and frame homomorphisms between them. We recall that a frame $\cal F$ is spatial if there exists a topological space $X$ such that $\cal F$ is isomorphic to the frame ${\cal O}(X)$ of open sets of $X$. Obviously, the opposite of the category $\textbf{Frm}_{sp}$ trivially identifies with a subcategory $\cal U$ of $\textbf{Loc}$, and this identification has the form of a functor $A:{\textbf{Frm}_{sp}}^{\textrm{op}}\to {\cal U}=ExtIm(A)$ with inverse $I_{A}:{\cal U}\to {\textbf{Frm}_{sp}}^{\textrm{op}}$, as in section \ref{locales} above. Indeed, one can naturally equip each frame $\cal F$, regarded as a preorder geometric category, with the canonical geometric Grothendieck topology $J_{\cal F}$; morphisms of frames ${\cal F} \to {\cal F}'$ yield morphisms of sites $({\cal F}, J_{\cal F}) \to ({\cal F}', J_{{\cal F}'})$, and for any frame $\cal F$, the locale $Id_{J_{\cal F}}({\cal F})$ is precisely the locale with underlying frame $\cal F$.
Starting with this trivial duality, if we take for any frame $\cal F$ in $\textbf{Frm}_{sp}$ as set of points of the topos $\Sh({\cal F}, J_{\cal F})$ the collection of \emph{all} the points of $\Sh({\cal F}, J_{\cal F})$ (equivalently, the completely prime filters on $\cal F$), then this set of points is separating for the topos (since $\cal F$ is spatial) and hence the technique of section \ref{dualtop} yields precisely the well-known duality between the category $\textbf{Frm}_{sp}$ and the category of sober spaces and continuous maps between them (cf. Example \ref{exa}(e)).
\subsection{Lindenbaum-Tarski duality}\label{tarski}
Recall that an element $a$ of a frame $F$ is said to be an atom if $a\neq 0$ and for any $b\leq a$, either $b=0$ or $b=a$. A frame $F$ is said to be \emph{atomic} it it has a basis of atoms, i.e. if every element of $F$ can be expressed as a join of atoms; accordingly, we say that a topological space has a \emph{basis of atomic open subsets} if the frame ${\cal O}(X)$ has a basis of atoms, equivalently if every open set can be expressed as a union of non-empty open sets which do not contain any proper non-empty open set.
Let $\textbf{AtFrm}$ be the category having as objects the atomic frames and as arrows the frame homomorphisms between them. We can equip each frame $\cal C$ in $\textbf{AtFrm}$ with the canonical topology $J_{{\cal C}}$. For any $\cal C$ in $\textbf{AtFrm}$, the collection $At({\cal C})$ of atoms of $\cal C$, regarded as a full subcategory of $\cal C$, is $J_{\cal C}$-dense and the induced Grothendieck topology $J_{\cal C}|_{At({\cal C})}$ is the trivial one; the Comparison Lemma thus yields an equivalence of toposes
\[
\Sh({\cal C}, J_{\cal C})\simeq [At({\cal C}), \Set].
\]
On the other hand, the category $At({\cal C})$ is a discrete preorder and hence the upper sets in $At({\cal C})$ coincides precisely with the subsets of $At({\cal C})$; therefore, Theorem \ref{fund} yields an equivalence of toposes
\[
[At({\cal C}), \Set] \simeq \Sh({\mathscr{P}(At({\cal C}))}),
\]
where ${\mathscr{P}(At({\cal C}))}$ is the locale given by the full powerset of $At({\cal C})$ endowed with the subset-inclusion order.
Composing the two equivalences we obtain an equivalence of toposes
\[
\Sh({\cal C}, J_{\cal C})\simeq \Sh({\mathscr{P}(At({\cal C}))}).
\]
This equivalence in turn entails an isomorphism of frames
\[
{\cal C}\cong {\mathscr{P}(At({\cal C}))},
\]
which is precisely the isomorphism given by the Lindenbaum-Tarski representation theorem (generalized from atomic complete Boolean algebras to atomic frames).
We note that the opposite $\textbf{AtFrm}^{\textrm{op}}$ of the category $\textbf{AtFrm}$ is naturally identified with a subcategory of $\textbf{Loc}$; so, for any subcategory $\cal A$ of $\textbf{AtFrm}$, the restriction $i|_{\cal A}$ to $\cal A$ of the inclusion functor $i:\textbf{AtFrm}^{\textrm{op}} \hookrightarrow \textbf{Loc}$ yields a duality between $\cal A$ and the subcategory of $\textbf{Loc}$ given by the image of $i|_{\cal A}$. In fact, such identification can be seen as one induced by the method of section \ref{locales}, since every arrow $f:{\cal C}\to {\cal D}$ in $\textbf{AtFrm}$ yields a morphism of sites $({\cal C}, J_{\cal C})\to ({\cal D}, J_{\cal D})$, and for any $\cal C$ in $\textbf{AtFrm}$, the frame of $J_{\cal C}$-ideals on $\cal C$ is isomorphic to $\cal C$.
To lift this trivial duality of atomic frames with locales to a topological duality, we select as set of points of the topos $\Sh({\cal C}, J_{\cal C})\simeq [At({\cal C}), \Set]$ the collection $\xi^{A}_{\cal C}$ of the points of $[At({\cal C}), \Set]$ corresponding to the elements of $At({\cal C})$, as in section \ref{prealex}. To make this assignment functorial, we have, according to the indications given in section \ref{dualtop}, to assign to an arrow $f:{\cal C}\to {\cal D}$ in $\textbf{AtFrm}$ a function $l_{f}:At({\cal D})\to At({\cal C})$ in such a way that the pair $(\dot{f}, l_{f})$ defines an arrow $(\Sh({\cal D}, J_{{\cal D}}), \xi^{A}_{\cal D}) \to (\Sh({\cal C}, J_{{\cal C}}), \xi^{A}_{\cal C})$ in the category $\mathfrak{Top}_{p}$ of toposes paired with points (cf. section \ref{subterminal}). This condition corresponds precisely, under the equivalences $\Sh({\cal C}, J_{\cal C})\simeq [At({\cal C}), \Set]$ and $\Sh({\cal D}, J_{\cal D})\simeq [At({\cal D}), \Set]$, to the requirement that $\dot{f}:[At({\cal D}), \Set] \to [At({\cal C}), \Set]$ be induced by the arrow $l_{f}:At({\cal D})\to At({\cal C})$ as in Example A4.1.4 \cite{El}. Now, by Lemma A4.1.5 \cite{El} and the remarks preceding it (combined with the fact that every discrete category is Cauchy-complete), a geometric morphism $[At({\cal D}), \Set] \to [At({\cal C}), \Set]$ is induced by a function $At({\cal D})\to At({\cal C})$ as in Example A4.1.4 \cite{El} if and only if it is \emph{essential}, i.e. if its inverse image functor has a left adjoint (in fact, the function inducing the morphism can be recovered, up to isomorphism, from this left adjoint as its restriction to the full subcategories on the representable functors). This property is a topos-theoretic invariant, and so, following the principles of \cite{OC10}, we attempt to reformulate it in terms of the different representations of our toposes established above.
Let us start from the first representation of our morphism $\dot{f}$, as the geometric morphism $\Sh({\cal D}, J_{{\cal D}}) \to \Sh({\cal C}, J_{{\cal C}})$ induced by a frame homomorphism $f:{\cal C}\to {\cal D}$. Recall that the assignment $L\to \Sh(L)$ defines a full and faithful $2$-functor from the $2$-category of locales to the $2$-category of Grothendieck toposes (cf. Proposition C1.4.5 \cite{El}); therefore, bearing in mind the possibility of characterizing adjoint functors `equationally' in terms of their unit and counit, we conclude that $\dot{f}$ is essential if and only if $f$ has a left adjoint $f_{!}:{\cal D}\to {\cal C}$ (where $\cal C$ and $\cal D$ are regarded as poset categories); note that, by the Special Adjoint Functor Theorem, this condition holds if and only if $f$ is complete, i.e. preserves arbitrary meets as well as arbitrary joins.
This argument shows that the pair $(\dot{f}, l_{f})$ considered above defines an arrow $(\Sh({\cal D}, J_{{\cal D}}), \xi^{A}_{\cal D}) \to (\Sh({\cal C}, J_{{\cal C}}), \xi^{A}_{\cal C})$ in the category of toposes paired with points if and only if the arrow $f$ is complete.
Now, let us turn to the second representation of our morphism $\dot{f}$ as a geometric morphism $s:[At({\cal D}), \Set] \to [At({\cal C}), \Set]$. By the remarks above, $s$ is essential if and only if there exists a function $l:At({\cal D})\to At({\cal C})$ such that, denoted by $\tilde{l}:[At({\cal D}), \Set]\to [At({\cal C}), \Set]$ the geometric morphism induced by $l$ as in Example A4.1.4 \cite{El}, $s$ is isomorphic to $\tilde{l}$. Now, since the inverse image functor of $\tilde{l}$ is given by composition with $l$, this condition says precisely that $s$ corresponds, under the equivalences $\Sh({\cal C}, J_{\cal C})\simeq [At({\cal C}), \Set]$ and $\Sh({\cal D}, J_{\cal D})\simeq [At({\cal D}), \Set]$, to the geometric morphism $\Sh({\cal D}, J_{\cal D}) \to \Sh({\cal C}, J_{\cal C})$ induced by the frame homomorphism $\mathscr{P}(l):{\mathscr{P}(At({\cal C}))} \to {\mathscr{P}(At({\cal D}))}$ taking inverse images of subsets along $l$.
By putting together the two pieces of information just collected, we can conclude that a frame homomorphism $r:\mathscr{P}(At({\cal C})) \to \mathscr{P}(At({\cal D}))$ is of the form $\mathscr{P}(l)$ for some function $l:At({\cal D})\to At({\cal C})$ if and only if $r$ is complete; notice that this is the other key ingredient of Lindenbaum-Tarski duality. In fact, our discussion above provides us with a concrete description of the function $l$ in terms of the frame homomorphism $r:{\cal C}\to {\cal D}$; indeed, we know that the function $l$ inducing the morphism can be recovered as the action of the left adjoint $r_{!}:{\cal D}\to {\cal C}$ to $r$ on the atoms (the fact that this left adjoint sends atoms to atoms is a consequence of the general topos-theoretic fact that the left adjoints to the inverse image functors of essential geometric morphisms sends representables to representables). Explicitly, for an element $d\in {\cal D}$, $r_{!}(d)$ is equal to the infimum in $\cal C$ of the set $\{c\in {\cal C} \textrm{ | } r(c)\geq d\}$.
Coming back to our original aim of building a topological duality for atomic frames, we can infer from our considerations above that restricting to the subcategory $\textbf{CAtFrm}$ of $\textbf{AtFrm}$ whose objects are the atomic frames and whose arrows are the complete frame homomorphisms between them is precisely the condition which allows us to make the assignment ${\cal C}\to \xi^{A}_{\cal C}$ (of a set of points of the topos $\Sh({\cal C}, J)$ to a frame $\cal C$ in $\textbf{CAtFrm}$) functorial in the sense of section \ref{dualtop}. We thus have a `lifting' functor $\tilde{A}:\textbf{CAtFrm} \to \textbf{Top}$ which sends a frame $\cal C$ in $\textbf{CAtFrm}$ to the space obtained by equipping the set of points $At({\cal C})$ with the subterminal topology (on the topos $[At({\cal C}), \Set]$), in other words the discrete topological space with underlying set $At({\cal C})$ (cf. Example \ref{exa}(b)). In fact, since every such space is sober, the general results of section \ref{dualtop} yield the following characterization of the extended image of the functor $\tilde{A}:\textbf{CAtFrm} \to \textbf{Top}$: the objects of $ExtIm(\tilde{A})$ are the sober topological spaces which have a basis of atomic open subsets, while the arrows of $ExtIm(\tilde{A})$ are precisely the continuous maps between them.
On the other hand, we have just characterized concretely the objects in $ExtIm(\tilde{A})$ as the spaces which are homeomorphic to the discrete topological spaces on the set of atoms of an atomic frame, and it is clear, from the equivalence ${\cal C}\cong {\mathscr{P}(At({\cal C}))}$ established above, that every discrete topological space is of that form. So $ExtIm(\tilde{A})$ can alternatively be described as the subcategory of $\textbf{Top}$ whose objects are the discrete topological spaces and whose arrows are the (continuous) functions between them. Since this category is (trivially) isomorphic to the category $\Set$ of sets, we conclude that the functor $\tilde{A}$ yields a duality between the category $\textbf{CAtFrm}$ and the opposite $\Set^{\textrm{op}}$ of the category of sets. Concretely, the functor $\tilde{A}$ sends an atomic frame $\cal C$ to the set $At({\cal C})$ of its atoms and an arrow $r:{\cal C}\to {\cal D}$ in $\textbf{CAtFrm}$ to the restriction $At({\cal D})\to At({\cal C})$ of the left adjoint to $r$ to the sets of atoms of $\cal C$ and $\cal D$. The inverse $\tilde{I_{A}}:ExtIm(\tilde{A})\simeq {\Set}^{\textrm{op}} \to \textbf{CAtFrm}$ of this functor is obtained, as indicated in section \ref{dualtop}, by recovering each frame $\cal C$ in $\textbf{CAtFrm}$ from the topos $\Sh({\mathscr{P}(At({\cal C}))}) \simeq [At({\cal C}), \Set]$ (equivalently, from the set $At({\cal C})$) through a topos-theoretic invariant, functorially in $\textbf{CAtFrm}$. In this case, the topos-theoretic invariant (of families of subterminals in a topos) is the vacuous one, and hence $\tilde{I_{A}}$ sends a set $S$ to the powerset $\mathscr{P}(S)$ and a function $f:S\to T$ to the function $\mathscr{P}(f):\mathscr{P}(S) \to \mathscr{P}(T)$ sending a subset to the inverse image of it under $f$.
We have thus recovered the Lindenbaum-Tarski duality (cf. for example \cite{stone} for a classical treatment of it).
Note that, in passing, we have established the following intrinsic characterization of discrete spaces.
\begin{proposition}\label{discrete}
Let $X$ be a topological space. Then $X$ is discrete if and only if it is sober and has a basis of atomic open subsets.
\end{proposition}\qed
Finally, we remark that, in the same way we established Lindenbaum-Tarski duality for atomic frames, we can establish a more general topological duality between, on one hand, the category having as objects the frames with a basis of supercompact elements and as arrows the complete frame homomorphisms between them and, on the other hand, the category of Alexandrov spaces whose underlying preorder is a poset; this latter duality specializes to Lindenbaum-Tarski duality when sets are regarded as discrete preorders (in fact, it is essentially the same as the equivalence of Theorem \ref{proalex}).
\subsection{A duality for meet-semilattices}\label{meet}
Let $\textbf{MSLat}$ be the category of meet-semilattices and homomorphisms (i.e. maps preserving finite meets) between them.
Any meet-semilattice can be considered as a cartesian preorder category; as such, when equipped with the trivial Grothendieck topology, it gives rise to a cartesian site, and the arrows in $\textbf{MSLat}$ can be identified with the morphisms of these associated sites. The method of section \ref{locales} thus yields a faithful functor $A:\textbf{MSLat}^{\textrm{op}}\to \textbf{Loc}$ sending an object $\cal M$ of $\textbf{MSLat}$ to the frame $Id({{\cal M}})$ of lower sets in $\cal M$ and an arrow $f:{\cal M}\to {\cal N}$ to the frame homomorphism $A(f):Id({{\cal M}})\to Id({{\cal N}})$ which assigns to a lower set $S$ in $\cal M$ the lower set $A(f)(S)$ in $\cal N$ generated by the image $f(S)$ of $S$ in $\cal N$ under $f$.
By Corollary \ref{cor}(iii), every $\cal M$ in $\textbf{MSLat}$ is isomorphic to the subset of supercompact subterminals in the topos $[{\cal M}^{\textrm{op}}, \Set]$, endowed with the natural order between subterminals in $[{\cal M}^{\textrm{op}}, \Set]$. By Theorem \ref{induced}(ii) and Remark \ref{rmkprincipal}, the hypotheses of Theorem \ref{construction} are satisfied, and hence we have an inverse functor $I_{A}:\textbf{SCLoc}\to \textbf{MSLat}^{\textrm{op}}$ to $A$, where $\textbf{SCLoc}$ is the subcategory of $\textbf{Loc}$ given by the extended image of the functor $A$. By Theorem \ref{propext}, the locales in $\textbf{SCLoc}$ can be characterized precisely as those which have a basis of supercompact elements which is closed under finite meets (equivalently, as those such that the collection of their supercompact elements is closed under finite meets and forms a basis for them), while the arrows in $\textbf{SCLoc}$ are the locale morphisms whose associated frame homomorphisms send supercompact elements to supercompact elements. The functor $I_{A}:\textbf{SCLoc}\to \textbf{MSLat}^{\textrm{op}}$ sends a locale $L$ in $\textbf{SCLoc}$ to the collection of its supercompact elements (endowed with the natural order) and an arrow $f:L\to L'$ in $\textbf{SCLoc}$ to the restriction of its associated frame homomorphism to the subsets of supercompact elements of $L$ and $L'$.
We have thus obtained a duality between $\textbf{MSLat}$ and a category of locales:
\begin{theorem}\label{meetsm}
The functors
\[
A:\textbf{MSLat}^{\textrm{op}} \to \textbf{SCLoc}
\]
and
\[
I_{A}:\textbf{SCLoc}\to \textbf{MSLat}^{\textrm{op}}
\]
defined above yield a duality between the category $\textbf{MSLat}$ and the category $\textbf{SCLoc}$.
\end{theorem}
We shall call the locale corresponding to a meet-semilattice $\cal M$ via this duality the \emph{ideal locale} associated to $\cal M$.
Note that, at the level of objects, the duality of Theorem \ref{meetsm} can be read as the assertion that every meet-semilattice $\cal M$ is isomorphic to the subset of supercompact elements of the locale of ideals $Id({{\cal M}})$ of $\cal M$ (endowed with the natural induced order), and every locale with a basis of supercompact elements which is closed under finite meets is isomorphic to the locale of ideals $Id({{\cal M}})$ of the meet-semilattice $\cal M$ consisting of its supercompact elements.
The duality between $\textbf{MSLat}$ and $\textbf{SCLoc}$ can be lifted to a duality between $\textbf{MSLat}$ and a subcategory of the category $\textbf{Top}$ of topological spaces by using the method of section \ref{dualtop}. Let us choose, for each $\cal M$ in $\textbf{MSLat}$, as set of points of the topos $[{\cal M}^{\textrm{op}}, \Set]$, the set of \emph{all} the points of the topos $[{\cal M}^{\textrm{op}}, \Set]$, and define the function corresponding to the arrows in $\textbf{MSLat}$ accordingly, as specified in section \ref{dualtop}. This defines a `lifting functor' $\tilde{A}:\textbf{MSLat}^{\textrm{op}} \to \textbf{Top}$ having an inverse defined on its extended image, which we denote by $\textbf{SCTop}$. The spaces in $\textbf{SCTop}$ can be characterized precisely as the sober spaces which have a basis of supercompact open sets which is closed under finite intersections, while the arrows in $\textbf{SCTop}$ can be characterized as the continuous maps of topological spaces in $\textbf{SCTop}$ such that the inverse image of any supercompact open set is supercompact.
Concretely, by Proposition \ref{mslattice}, the extended functor $\tilde{A}:\textbf{MSLat}^{\textrm{op}} \to \textbf{Top}$ sends to a meet-semilattice $\cal M$ the topological space having as set of points the collection ${\cal F}_{\cal M}$ of filters on $\cal M$ and as open sets the sets of the form
\[
{\cal F}_{I}=\{F\in {\cal F}_{\cal M} \textrm{ | } F\cap I\neq \emptyset\},
\]
where $I$ ranges among the ideals of $\cal M$. In particular, a basis for this topological space is given by the sets
\[
{\cal F}_{m}=\{F\in {\cal F}_{\cal M} \textrm{ | } m\in {\cal M}\},
\]
where $m$ varies among the elements of $\cal M$. We shall call this topological space the \emph{ideal topological space} associated to $\cal M$.
By Proposition \ref{mslatticefun}, the continuous map ${\cal F}_{\cal N}\to {\cal F}_{\cal M}$ corresponding to a meet-semilattice homomorphism $f:{\cal M}\to {\cal N}$ via the functor $\tilde{A}$ is the map sending a filter $F$ in ${\cal F}_{\cal N}$ to the inverse image $f^{-1}(F)$.
The inverse functor $\tilde{I_{A}}:\textbf{SCTop}\to \textbf{MSLat}^{\textrm{op}}$ sends a topological space $X$ in $\textbf{SCTop}$ to the poset $\textbf{SC}({{\cal O}(X)})$ of its supercompact open sets, and a continuous map $f:X\to Y$ in $\textbf{SCTop}$ to the restriction $f^{-1}:\textbf{SC}({{\cal O}(Y)}) \to \textbf{SC}({{\cal O}(X)})$ of the inverse image $f^{-1}$ to the posets of supercompact open sets of $X$ and $Y$.
Summmarizing, we have the following result.
\begin{theorem}\label{meetdual}
Via the functors
\[
\tilde{A}:\textbf{MSLat}^{\textrm{op}} \to \textbf{SCTop}
\]
and
\[
\tilde{I_{A}}:\textbf{SCTop}\to \textbf{MSLat}^{\textrm{op}}
\]
defined above, the categories $\textbf{MSLat}$ and $\textbf{SCTop}$ are dual to each other.
\end{theorem}
\begin{remark}
\begin{enumerate}[(a)]
\item The functor $\tilde{A}:\textbf{MSLat}^{\textrm{op}} \to \textbf{SCTop}$ coincides with the functor from meet-semilattices to topological spaces giving one half of the topological duality for meet-semilattices established by M. A. Moshier and P. Jipsen in \cite{jipsen}. In fact, it follows from Theorem \ref{meetdual} and the duality theorem in \cite{jipsen} that the category $\textbf{SCTop}$ coincides with the subcategory of $\textbf{Top}$ corresponding to the category of meet-semilattices via the duality in \cite{jipsen} (since the two theorems ensure that both these categories are equal to the extended image of the same functor, namely $\tilde{A}:\textbf{MSLat}^{\textrm{op}} \to \textbf{SCTop}$); anyway, it is worth to note that our definition of the inverse functor $\tilde{I_{A}}:\textbf{SCTop}\to \textbf{MSLat}^{\textrm{op}}$ is completely different from the construction of the inverse functor given in \cite{jipsen};
\item The functor $\tilde{A}:\textbf{MSLat}^{\textrm{op}} \to \textbf{SCTop}$ can also be identified with the functor from meet-semilattices to topological meet-semilattices (regarded here as topological spaces) giving one half of the duality between meet-semilattices and compact zero-dimensional meet-semilattices established in \cite{semilattices}.
\end{enumerate}
\end{remark}
\subsection{Other dualities}\label{comparison}
We have seen in section \ref{stonedist} that the well-known Stone duality for distributive lattices is recovered through our method by equipping each distributive lattice with the coherent topology. Anyway, there are several other interesting subcanonical Grothendieck topologies which one can put on distributive lattices, or on any other kind of preordered structures; for each of these choices, we can try to use our machinery to build dualities which, although being similar in spirit to Stone's one, are different enough to capture other aspects of the structures involved. We have already presented some examples above, of both old and new dualities obtained through the application of our methodology; in this section we discuss some further ones, of more `exotic' nature, which we hope should illuminate the flexibility of our general machinery.
First, let us build a localic duality for $k$-frames (for a regular cardinal $k$), which specializes to the localic Stone duality for distributive lattices and to the duality of Theorem \ref{meetsm}. If we equip each $k$-frame ${\cal F}$ with the $k$-covering topology $J^{k}_{{\cal F}}$ we obtain that the morphisms ${\cal F}\to {\cal F}'$ of $k$-frames are precisely the morphisms of sites $({\cal F}, J^{k}_{{\cal F}}) \to ({\cal F}', J^{k}_{{\cal F}'})$. Therefore, by the technique of section \ref{locales}, we have a functor $A:{k\textrm{-}\textbf{Frm}}^{\textrm{op}}\to \textbf{Loc}$ which sends a poset $\cal F$ in ${k\textrm{-}\textbf{Frm}}$ to the locale of $J^{k}_{{\cal F}}$-ideals on $\cal F$ and an arrow $f:{\cal F}\to {\cal F}'$ in ${k\textrm{-}\textbf{Frm}}$ to the frame homomorphism $Id_{J^{k}_{{\cal F}}}({\cal F}) \to Id_{J^{k}_{{\cal F}'}}({\cal F}')$ sending an ideal $I$ in $Id_{J^{k}_{{\cal F}}}({\cal F})$ to the $J^{k}_{{\cal F}'}$-ideal on ${\cal F}'$ generated by the image $f(I)$ of $I$ under $f$.
By Theorem \ref{induced}(iii) and Remark \ref{rmkprincipal}, the hypotheses of Theorem \ref{construction} are satisfied, and hence we obtain, by the results of section \ref{locales}, that the functor $A$ admits an inverse $I_{A}$ defined on its extended image of, which we denote by $\textbf{Loc}_{k}$. This latter subcategory has as objects the locales which have a basis of $k$-compact elements which is closed under finite meets and as arrows the locale morphisms whose associated frame homomorphisms send $k$-compact elements to $k$-compact elements. The inverse functor $I_{A}: \textbf{Loc}_{k} \to {k\textrm{-}\textbf{Frm}}^{\textrm{op}}$ sends a locale in $\textbf{Loc}_{k}$ to the poset of its $k$-compact elements and a locale morphism $L\to L'$ to the restriction of its associated frame homomorphism to the subsets of $k$-compact elements of $L$ and $L'$. Summarizing, we have the following result.
\begin{theorem}
Via the functors $A:k\textrm{-}\textbf{Frm}^{\textrm{op}} \to \textbf{Loc}_{k}$ and $I_{A}: \textbf{Loc}_{k} \to {k\textrm{-}\textbf{Frm}}^{\textrm{op}}$ defined above, the categories ${k\textrm{-}\textbf{Frm}}$ and $\textbf{Loc}_{k}$ are dual to each other.
\end{theorem}
Now, let us build dualities for disjunctively distributive lattices and for weakly atomic meet-semilattices. To this end, we recall from section \ref{charinv} the definition of disjunctive topology and atomically generated topology.
Given a disjunctively distributive lattice $\cal D$, the \emph{disjunctive topology} $J_{{\cal D}}^{dj}$ on $\cal D$ is the Grothendieck topology on $\cal D$ (regarded as a preorder category) whose $J_{{\cal D}}^{dj}$-covering sieves on any element $d$ are exactly the sieves in $\cal D$ on $d$ which contain a finite family $\{d_{i} \textrm{ | } i\in I\}$ of elements $d_{i}\leq d$ such that for each pair of distinct $i, j\in I$, $d_{i}\wedge d_{j}=0$, and $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}} d_{i}=d$. Similarly, on an disjunctively distributive frame one can consider the infinitary disjunctive topology.
Given a weakly atomic meet-semilattice $\cal M$, the \emph{atomically generated top-}\\ \emph{ology} $J_{{\cal M}}^{at}$ on $\cal M$ is defined as follows: the $J_{{\cal M}}^{at}$-covering sieves on any element $d$ are the maximal sieves and the sieves in $\cal M$ on $d$ which contain a finite family $\{d_{i} \textrm{ | } i\in I\}$ of elements $d_{i}\leq d$ such that $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}} d_{i}=d$ and each $d_{i}$ is an atom in $\cal M$ (i.e. for any $e\in {\cal M}$, if $e\leq d$ then either $e=d$ or $e=0$). Similarly, on any infinitarily weakly atomic meet-semilattice one can consider the infinitary atomically generated topology.
The definition of supercompactly generated topology on a weakly supercompact meet-semilattice is of course entirely analogous.
First, let us focus on disjunctively distributive lattices.
Any arrow ${\cal D}\to {\cal D}'$ in $\textbf{DJLat}$ (cf. section \ref{charinv} for the definition of this category) yields a morphism of sites $({\cal D}, J_{{\cal D}}^{dj})\to ({\cal D}', J_{{\cal D}'}^{dj})$. So, by the technique of section \ref{locales}, we have a faithful functor $A:\textbf{DJLat}^{\textrm{op}} \to \textbf{Loc}$, which sends a poset $\cal D$ in $\textbf{DJLat}$ to the locale of $J_{{\cal D}}^{dj}$-ideals on $\cal D$ and an arrow $f:{\cal D}\to {\cal D}'$ in $\textbf{DJLat}$ to the frame homomorphism $Id_{J_{{\cal D}}^{dj}}({\cal D}) \to Id_{J_{{\cal D}'}^{at}}({\cal D}')$ which sends an ideal $I$ in $Id_{J_{{\cal D}}^{dj}}({\cal D})$ to the $J_{{\cal D}'}^{dj}$-ideal on ${\cal D}'$ generated by the image $f(I)$ of $I$ in ${\cal D}'$ under $f$.
By Corollary \ref{cor}(v), every $\cal D$ in $\textbf{DJLat}^{\textrm{op}}$ can be recovered, up to isomorphism, as the poset of disjunctively compact subterminals in the topos $\Sh({\cal D}, J_{{\cal D}}^{dj})$. By Theorem \ref{induced}(iv) and Remark \ref{rmkprincipal}, the hypotheses of Theorem \ref{construction} are satisfied, and hence we obtain, by the results of section \ref{locales}, that the functor $A$ admits an inverse $I_{A}$ defined on its extended image of, which we denote by $\textbf{Loc}_{dj}$. By Theorem \ref{propext} and Proposition \ref{multicomposition}, we can characterize the objects in $\textbf{Loc}_{dj}$ as the locales which have a basis of disjunctively compact elements which is closed under finite meets and satisfies the property that any covering of an element of the basis has a disjunctively compact refinement by elements of the basis, while the arrows in $\textbf{Loc}_{dj}$ are precisely the locale morphisms whose associated frame homomorphisms send disjunctively compact elements to disjunctively compact elements. The functor $I_{A}:\textbf{Loc}_{dj}\to \textbf{DJLat}^{\textrm{op}}$ sends a locale in $\textbf{Loc}_{dj}$ to the poset of its disjunctively compact elements and a locale morphism $L\to L'$ in $\textbf{Loc}_{dj}$ to the restriction of its associated frame homomorphism to the sets of disjunctively compact elements of $L$ and $L'$. Summarizing, we have the following result:
\begin{theorem}
Via the functors $A:\textbf{DJLat}^{\textrm{op}}\to \textbf{Loc}_{dj}$ and $I_{A}:\textbf{Loc}_{dj} \to \textbf{DJLat}^{\textrm{op}}$ defined above, the categories $\textbf{DJLat}$ and $\textbf{Loc}_{dj}$ are dual to each other.
\end{theorem}\qed
Note that the category $\textbf{DLat}$ of distributive lattices can be identified as a full subcategory of the category $\textbf{DJLat}$, so this duality restricts in particular to a duality between $\textbf{DLat}$ and a subcategory of $\textbf{Loc}_{dj}$.
Since all the toposes involved are coherent (and hence, under a form of the axiom of choice, have enough points by Deligne's theorem), this duality can be lifted to a duality with a category of topological spaces, by assigning to each $\cal D$ in $\textbf{DJLat}$ the set of points of the topos $\Sh({\cal D}, J_{{\cal D}}^{dj})$ (equivalently, the collection of \emph{disjunctive filters} $F$ on $\cal D$, i.e. the filters on $\cal D$ with the property that for any $J_{{\cal D}}^{dj}$-covering sieve $S$ on an object $d$ of $\cal D$ if $d\in F$ then there exists an element $f\in S$ such that $dom(f)$ belongs to $F$) topologized with the subterminal topology, and assigning function to arrows accordingly, as specified in section \ref{dualtop}. The functor $\tilde{A}:\textbf{DJLat}^{\textrm{op}} \to \textbf{Top}$ sends a poset $\cal D$ in $\textbf{DJLat}$ to the topological space $\tilde{A}({\cal D})$ having as underlying set the set ${\cal D}_{dj}$ of disjunctive filters on $\cal D$ and as basic open sets those of the form ${\cal F}_{c}=\{F\in {\cal D}_{dj} \textrm{ | } d\in F\}$ for $d\in {\cal D}$ (cf. Proposition \ref{mslattice}) and an arrow $f:{\cal D}\to {\cal D}'$ in $\textbf{DJLat}$ to the continuous map ${{\cal D}'}_{dj}\to {\cal D}_{dj}$ sending a filter $F$ in ${{\cal D}'}_{dj}$ to the inverse image $f^{-1}(F)$ (cf. Proposition \ref{mslatticefun}). The extended image of the functor $\tilde{A}$ is the subcategory $\textbf{Top}_{dj}$ of $\textbf{Top}$ whose objects are the sober topological spaces with a basis of disjunctively compact open sets which is closed under finite intersections and satisfies the property that any covering of a basic open set has a disjunctively compact refinement by basic open sets, and whose arrows are the continuous maps between such spaces such that the inverse image of any disjunctively compact open set is a disjunctively compact open set. The inverse functor $\tilde{I_{A}}:\textbf{Top}_{dj}\to \textbf{DJLat}^{\textrm{op}}$ sends a topological space $X$ in $\textbf{Top}_{dj}$ to the poset $\textbf{DJ}{{\cal O}(X)}$ of its disjunctively compact open sets and a continuous map $f:X\to Y$ in $\textbf{Top}_{dj}$ to the restriction $f^{-1}:\textbf{DJ}{\cal O}(Y) \to \textbf{DJ}{\cal O}(X)$ of the inverse image $f^{-1}$ to the posets of disjunctively compact open sets of $X$ and $Y$.
Summarizing, we have the following result.
\begin{theorem}
Via the functors
\[
\tilde{A}:\textbf{DJLat}^{\textrm{op}} \to \textbf{Top}_{dj}
\]
and
\[
\tilde{I_{A}}:\textbf{Top}_{dj}\to \textbf{DJLat}^{\textrm{op}}
\]
defined above, the categories $\textbf{DJLat}$ and $\textbf{Top}_{dj}$ are dual to each other.
\end{theorem}\qed
Similarly, by using Theorem \ref{induced}(v) and Remark \ref{rmkprincipal}, one obtains a duality between the category $\textbf{WAtMSLat}$ of weakly atomic meet-semilattices (cf. section \ref{charinv}) and the subcategory $\textbf{Loc}_{WAt}$ of \textbf{Loc} whose objects are the locales which have a basis of atomically compact elements which is closed under finite meets and whose arrows are the locale maps whose associated frame homomorphisms send atoms to atoms and atomically compact elements to atomically compact elements. The functor $A:\textbf{WAtMSLat}^{\textrm{op}}\to \textbf{Loc}_{WAt}$ sends a poset $\cal M$ in $\textbf{WAtMSLat}$ to the locale of $J_{{\cal M}}^{at}$-ideals on $\cal M$ and an arrow $f:{\cal M}\to {\cal N}$ in $\textbf{WAtMSLat}$ to the frame homomorphism $Id_{J_{{\cal M}}^{at}}({\cal M}) \to Id_{J_{{\cal N}}^{at}}({\cal N})$ which sends an ideal $I$ in $Id_{J_{{\cal M}}^{at}}({\cal M})$ to the $J_{{\cal N}}^{at}$-ideal on $\cal N$ generated by the image $f(I)$ of $I$ in ${\cal N}$ under $f$. The inverse functor $I_{A}:\textbf{Loc}_{WAt} \to \textbf{WAtMSLat}^{\textrm{op}}$ sends a locale $L$ in $\textbf{Loc}_{WAt}$ to the poset of atomically compact elements of $L$ and an arrow $f:L\to L'$ in $\textbf{Loc}_{WAt}$ to the restriction of its associated frame homomorphism to the subsets of atomically compact elements of $L$ and of $L'$.
Summarizing, we have the following result.
\begin{theorem}
Via the functors
\[
A:\textbf{WAtMSLat}^{\textrm{op}}\to \textbf{Loc}_{WAt}
\]
and
\[
I_{A}:\textbf{Loc}_{WAt} \to \textbf{WAtMSLat}^{\textrm{op}}
\]
defined above, the categories $\textbf{WAtMSLat}$ and $\textbf{Loc}_{WAt}$ are dual to each other.
\end{theorem}
Of course, such a duality admits an infinitary version for infinitarily weakly atomic meet-semilattices.
Since all the toposes involved are coherent (and hence, under a form of the axiom of choice, have enough points by Deligne's theorem), this duality can also be lifted to a topological duality, as in the example above. For instance, one can assign to each $\cal M$ in $\textbf{WAtMSLat}$ the set of points of the topos $\Sh({\cal M}, J_{{\cal M}}^{at})$ (i.e. the collection of \emph{atomic filters} $F$ on $\cal M$ i.e. the filters $F$ on $\cal M$ with the property that if a join of atoms belongs to $F$ then at least one of these atoms belongs to $F$) topologized with the subterminal topology, and defining the action on arrows accordingly. The resulting functor $\tilde{A}:\textbf{WAtMSLat}^{\textrm{op}} \to \textbf{Top}$ sends a poset $\cal M$ in $\textbf{WAtMSLat}$ to the topological space $\tilde{A}({\cal M})$ having as underlying set the set ${\cal M}_{at}$ of atomic filters on $\cal M$ and as basic open sets those of the form ${\cal F}_{c}=\{F\in {\cal M}_{at} \textrm{ | } d\in F\}$ for $d\in {\cal M}$ (cf. Proposition \ref{mslattice}), and an arrow $f:{\cal M}\to {\cal N}$ in $\textbf{WAtMSLat}$ to the continuous map ${{\cal N}}_{at}\to {\cal M}_{at}$ sending a filter $F$ in ${{\cal N}}_{at}$ to the inverse image $f^{-1}(F)$ (cf. Proposition \ref{mslatticefun}). The extended image of the functor $\tilde{A}$ is the subcategory $\textbf{Top}_{at}$ of $\textbf{Top}$ whose objects are the sober topological spaces with a basis of atomically compact open sets which is closed under finite intersections, and whose arrows are the continuous maps between such spaces such that the inverse image of any atomically compact open set is an atomically compact open set, and the inverse image of any atomic open set is an atomic open set. The inverse functor $\tilde{I(A)}:\textbf{Top}_{at}\to \textbf{WAtMSLat}^{\textrm{op}}$ sends a topological space $X$ in $\textbf{Top}_{at}$ to the poset $\textbf{AT}({{\cal O}(X)})$ of its atomically compact open sets and a continuous map $f:X\to Y$ in $\textbf{Top}_{at}$ to the restriction $f^{-1}:\textbf{AT}({\cal O}(Y)) \to \textbf{AT}({\cal O}(X))$ of the inverse image $f^{-1}$ to the posets of atomically compact open sets of $X$ and $Y$. Summarizing, we have the following result.
\begin{theorem}
Via the functors
\[
\tilde{A}:\textbf{WAtMSLat}^{\textrm{op}}\to \textbf{Top}_{at}
\]
and
\[
\tilde{I_{A}}:\textbf{Top}_{at} \to \textbf{WAtMSLat}^{\textrm{op}}
\]
defined above, the categories $\textbf{WAtMSLat}$ and $\textbf{Top}_{at}$ are dual to each other.
\end{theorem}
Similarly, one can obtain localic and topological dualities for weakly supercompact meet-semilattices, and a localic duality for infinitarily weakly supercompact meet-semilattices.
We have limited ourselves to presenting just a few examples of application of our machinery. Anyway, the reader should be convinced at this point that the generality and flexibility of the method described in sections \ref{locales} and \ref{dualtop} enables one to easily establish many other new dualities or equivalences. In particular, `natural' dualities, such as the ones that we have obtained above, can be generated by choosing the subcanonical topologies $J_{\cal C}$ in such a way to capture the operations defining the structure of the corresponding poset $\cal C$; for example, in the classical case of Stone duality, the joins in a distributive lattice can be directly defined in terms of the coherent topology on it, while the finite joins of pairwise disjoint elements in a disjunctively distributive lattice can be characterized in terms of the disjunctive topology on it. In connection with this, it is worth to remark that if $\cal C$ is a meet-semilattice then the Yoneda embedding $y:{\cal C} \hookrightarrow \Sh({\cal C}, J_{\cal C})\simeq \Sh(Id_{J_{\cal C}}({\cal C}))$ preserves finite meets and sends $J_{\cal C}$-covering sieves to covering families; so, meets in $\cal C$ correspond to meets in $Id_{{J_{\cal C}}}({\cal C})$, while operations on $\cal C$ defined in terms of $J$ correspond via the embedding ${\cal C}\hookrightarrow Id_{{J_{\cal C}}}({\cal C})$ to operations on $Id_{{J_{\cal C}}}({\cal C})$ involving joins (in particular, if $\cal C$ possesses a bottom element $0$ which is $J_{\cal C}$-covered by the empty sieve then $y$ sends the initial object $0$ of $\cal C$ to the initial object of $\Sh({\cal C}, J_{\cal C})$, that is to the bottom element of the frame $Id_{{J_{\cal C}}}({\cal C})$).
\section{A further generalization}\label{generalization}
We note that the fundamental ingredient that we have taken as a starting point in section \ref{locales} to generate dualities consists of bunches of Morita-equivalences which are instances of Theorem \ref{fund}.
More generally, we can expect to be able to extract representation theorems or dualities between partially ordered structures starting from general Morita-equivalences of the form $\Sh({\cal C}, J)\simeq \Sh({\cal D}, K)$, where $\cal C$ and $\cal D$ are poset categories. Indeed, if $J$ is $U$-induced for a topos-theoretic invariant $U$ of families of subterminals in a topos satisfying the hypothesis of Theorem \ref{construction}, $\cal C$ can be represented as the poset consisting of the elements of the locale $Id_{K}({\cal D})$ which are $U$-compact. And if we have a bunch of categorical equivalences $\Sh({\cal C}, J)\simeq \Sh({\cal D}, K)$ for a collection of pairs of poset structures $({\cal C}, {\cal D})$, we can expect to be able to build a duality or equivalence between categories having as objects respectively the structures of the form $\cal C$ and the structures of the form $\cal D$.
\subsection{From Morita-equivalences to dualities}\label{Mordual}
To formalize the idea described above, suppose that we have two collections ${\cal K}$ and $\cal H$ of poset structures, each of which equipped with a Grothendieck topology (we denote by $J_{\cal C}$ (resp. by $K_{\cal D}$) the Grothendieck topology associated to a structure $\cal C$ (resp. $\cal D$) in $\cal K$ (resp. in $\cal H$)), two functions $f:{\cal K}\to {\cal H}$ and $g:{\cal H}\to {\cal K}$ which are inverse to each other up to isomorphisms, and a categorical equivalence $\Sh({\cal C}, J_{\cal C})\simeq \Sh(f({\cal C}), K_{f({\cal C})})$ for each $\cal C$ in $\cal K$ (equivalently, a categorical equivalence $\Sh({\cal D}, K_{\cal D})\simeq \Sh(g({\cal D}), J_{g({\cal D})})$ for each $\cal D$ in $\cal H$).
We shall adopt the following conventions:
\begin{enumerate}
\item Given a morphism of posites $f:({\cal A}, J)\to ({\cal B}, K)$, we denote by $A(f)$ the frame homomorphism $Id_{J}({\cal A})\to Id_{K}({\cal B})$ which sends an ideal $I\in Id_{J}({\cal A})$ to the $K$-ideal on $\cal B$ generated by the image $f(I)$ of $I$ under $f$. Recall from section \ref{locales} that $A(f)$ can be identified with the restriction to the subterminals of the inverse image functor of the geometric morphism $\Sh({\cal B}, K) \to \Sh({\cal A}, J)$ induced by the morphism of sites $f$;
\item Given a functor $f:{\cal A}\to {\cal B}$ between posets, we denote by $B(f):Id({\cal B}^\textrm{op}) \to Id({\cal A}^\textrm{op})$ the frame homomorphism sending a subset $I\subseteq B$ in $Id({\cal B}^\textrm{op})$ to the inverse image $f^{-1}(I)$ of $I$ under $f$. Recall from section \ref{prealex} that $B(f)$ can be identified with the restriction to the subterminals of the inverse image functor of the geometric morphism $[{\cal A}, \Set] \to [{\cal B}, \Set]$ induced by the functor $f$ as in A4.1.4 \cite{El}.
\end{enumerate}
Let us suppose that $U$ (resp. $V$) is a topos-theoretic invariant satisfying the hypothesis of Theorem \ref{thmcentral} with respect to the category $\cal K$ (resp. the category $\cal H$), so that all the Grothendieck topologies $J_{\cal C}$ (resp. $K_{\cal D}$) are $U$-induced (resp. $V$-induced). Given a frame $L$, we denote by $J_{can}^{L}$ the canonical topology on $L$ (i.e. the Grothendieck topology on $L$ whose covering sieves are exactly those which contain a covering family in $L$) and by $U\textrm{-comp}(L)$ (resp. $V\textrm{-comp}(L)$) the collection of the elements of $L$ which are $U$-compact (resp. $V$-compact); the Grothendieck topology on $U\textrm{-comp}(L)$ (resp. on $V\textrm{-comp}(L)$) induced by the canonical topology $J_{can}^{L}$ on $L$ by regarding $U\textrm{-comp}(L)$ (resp. $V\textrm{-comp}(L)$) as a full subcategory of $L$ via the inclusion $U\textrm{-comp}(L) \hookrightarrow L$ (resp. the inclusion $V\textrm{-comp}(L) \hookrightarrow L$) (cf. p. 546 \cite{El} for the general definition of induced coverage) will be denoted by $J_{can}^{L}|_{U\textrm{-comp}(L)}$ (resp. by $J_{can}^{L}|_{V\textrm{-comp}(L)}$) or simply by $J_{can}^{L}|$ when the subcategory $U\textrm{-comp}(L)$ (resp. $V\textrm{-comp}(L)$) can be unambiguously inferred from the context. Given a frame $L$, we denote by $\textbf{SC}(L)$ the set of supercompact elements of $L$.
We distinguish between the covariant and contravariant case:
\begin{enumerate}[(i)]
\item \emph{Covariant case}
We define two categories $Ext({\cal K})$ and $Ext({\cal H})$ as follows.
The objects of $Ext({\cal K})$ are the posets which are isomorphic to a poset in $\cal K$, while the arrows ${\cal C}\to {\cal C}'$ in $Ext({\cal K})$ are the monotone maps $f:{\cal C}\to {\cal C}'$ which induce morphisms of sites $({\cal C}, J_{\cal C})\to ({\cal C}', J_{{\cal C}'})$ such that the frame homomorphism $A(f):Id_{J_{\cal C}}({\cal C})\to Id_{J_{{\cal C}'}}({\cal C}')$ sends $V$-compact elements to $V$-compact elements and its restriction
\[
A(f)|:V\textrm{-comp}(Id_{J_{\cal C}}({\cal C})) \to V\textrm{-comp}(Id_{J_{{\cal C}'}}({\cal C}'))
\]
to these posets yields a morphism of sites
\[
(V\textrm{-comp}(Id_{J_{\cal C}}({\cal C})), J_{can}^{{Id_{J_{\cal C}}({\cal C})}}|) \to (V\textrm{-comp}(Id_{J_{{\cal C}'}}({\cal C}')), J_{can}^{{Id_{J_{{\cal C}'}}({\cal C}')}}|).
\]
The definition of $Ext({\cal H})$ is perfectly symmetrical to that of $Ext({\cal K})$ ($\cal H$ playing the role of $\cal K$ and $U$ playing the role of $V$).
Let us now define two functors $D:Ext({\cal K}) \to Ext({\cal H})$ and $E:Ext({\cal H}) \to Ext({\cal K})$ which extend respectively the functions $f$ and $g$ (up to isomorphism) as follows.
The functor $D:Ext({\cal K}) \to Ext({\cal H})$ sends a poset $\cal C$ in $Ext({\cal K})$ to the poset $V\textrm{-comp}(Id_{J_{\cal C}}({\cal C}))$ of $V$-compact elements of the frame $Id_{J_{\cal C}}({\cal C})$, and an arrow $f:{\cal C}\to {\cal C}'$ in $Ext({\cal K})$ to the restriction
\[
A(f)|:V\textrm{-comp}(Id_{J_{\cal C}}({\cal C})) \to V\textrm{-comp}(Id_{J_{{\cal C}'}}({\cal C}'))
\]
of the frame homomorphism $A(f):Id_{J_{\cal C}}({\cal C})\to Id_{J_{{\cal C}'}}({\cal C}')$ to the subsets of $V$-compact elements of $Id_{J_{\cal C}}({\cal C})$ and of $Id_{J_{{\cal C}'}}({\cal C}')$.
The definition of the functor $E:Ext({\cal H}) \to Ext({\cal K})$ is perfectly symmetrical to that of the functor $D$.
\begin{theorem}\label{dualcov}
Under the hypotheses specified above, the functors $D$ and $E$ are categorical inverses to each other and hence they define a categorical equivalence between $Ext({\cal K})$ and $Ext({\cal H})$.
\end{theorem}
\begin{proofs}
The fact that the functors $D$ and $E$ are well-defined, in the sense that $D$ takes values in $Ext({\cal H})$ (and, dually, $E$ takes values in $Ext({\cal K})$), and that they are categorical inverses to each other easily follows from Theorem \ref{thmcentral} by invoking the general theory of morphisms of sites. Specifically, one appeals to the fact that, given two sites $({\cal C}, J)$ and $({\cal D}, K)$ where $J$ and $K$ are subcanonical, a geometric morphism $f:\Sh({\cal D}, K)\to \Sh({\cal C}, J)$ is induced by a morphism of sites $({\cal C}, J)\to ({\cal D}, K)$ if and only if the inverse image $f^{\ast}$ sends representables to representables (Lemma C2.3.8 \cite{El}).
\end{proofs}
From the theorem we can easily deduce the following characterization of the categories $Ext({\cal H})$ and $Ext({\cal K})$ in terms of each other, which represents the analogue of Theorem \ref{propext}.
\begin{proposition}
Let us assume that the hypotheses of Theorem \ref{dualcov} are satisfied, and that $U$ is $\cal K$-compatible relative to a property $P$. Let ${\cal F}^{U}_{P}$ denote the set of frames such that the collection of their $U$-compact elements forms a basis of them which, endowed with the induced order, has the structure of a poset in $Ext({\cal K})$ and has the property that the embedding $B_{L}\hookrightarrow L$ of it into the frame satisfies condition $P$, the property that every covering in $L$ of an element of $B_{L}$ is refined by a covering made of elements of $B_{L}$ which satisfies the invariant $U$, and the property that the $J_{B_{L}}$-covering sieves are sent by the embedding $B_{L}\hookrightarrow L$ into covering families in $L$. Then
\begin{enumerate}[(i)]
\item The objects of $Ext({\cal H})$ are precisely the posets of $V$-compact elements of the frames in ${\cal F}^{U}_{P}$;
\item The arrows in $Ext({\cal H})$ are the restrictions to the subsets of $V$-compact elements of the frame homomorphisms between frames in ${\cal F}^{U}_{P}$ which send $U$-compact elements to $U$-compact elements and $V$-compact elements to $V$-compact elements.
\end{enumerate}
\end{proposition}
Of course, the description of the category $Ext({\cal H})$ is perfectly symmetrical.
\item \emph{Contravariant case}
In the contravariant case, one of the two toposes involved in the Morita-equivalence is a presheaf topos; that is, we suppose $K_{\cal D}$ to be the trivial topology for every $\cal D$ in $\cal H$. We can therefore suppose, by Theorem \ref{induced}(i) and Remark \ref{rmkprincipal}, the invariant $V$ to be the property `to be a singleton family'.
We define two categories $Ext({\cal K})$ and $Ext({\cal H})$ as follows.
The objects of $Ext({\cal K})$ are the posets which are isomorphic to a poset in $\cal K$, while the arrows ${\cal C}\to {\cal C}'$ in $Ext({\cal K})$ are the monotone maps $f:{\cal C}\to {\cal C}'$ such that they induce morphisms of sites $({\cal C}, J_{\cal C})\to ({\cal C}', J_{{\cal C}'})$ such that the frame homomorphism $A(f):Id_{J_{\cal C}}({\cal C})\to Id_{J_{{\cal C}'}}({\cal C}')$ is complete i.e. preserves arbitrary infima (cf. section \ref{tarski}).
The objects of $Ext({\cal H})$ are the posets which are isomorphic to a poset in $\cal H$, while the arrows ${\cal D}\to {\cal D}'$ in $Ext({\cal H})$ are the monotone maps $f:{\cal D}\to {\cal D}'$ such that the frame homomorphism $B(f^{\textrm{op}}):Id({{\cal D}'}) \to Id({\cal D})$ sends $U$-compact elements to $U$-compact elements.
Let us now define two functors
\[
D:Ext({\cal K})^{\textrm{op}} \to Ext({\cal H})
\]
and
\[
E:Ext({\cal H}) \to Ext({\cal K})^{\textrm{op}}
\]
which extend respectively the functions $f$ and $g$ (up to isomorphism).
The functor $D:Ext({\cal K})^{\textrm{op}} \to Ext({\cal H})$ sends a poset $\cal C$ in $Ext({\cal K})$ to the poset of supercompact elements of the frame $Id_{J_{\cal C}}({\cal C})$, and an arrow $f:{\cal C}\to {\cal C}'$ in $Ext({\cal K})$ to the restriction of the left adjoint $A(f)_{!}:Id_{J_{{\cal C}'}}({\cal C}')\to Id_{J_{{\cal C}}}({\cal C})$ to the functor $A(f):Id_{J_{\cal C}}({\cal C})\to Id_{J_{{\cal C}'}}({\cal C}')$ to the subsets of supercompact elements of $Id_{J_{{\cal C}'}}({\cal C}')$ and of $Id_{J_{\cal C}}({\cal C})$.
The functor $E:Ext({\cal H}) \to Ext({\cal K})^{\textrm{op}}$ sends a poset $\cal D$ in $Ext({\cal H})$ to the poset of $U$-compact elements of the frame $Id({{\cal D}})$ and an arrow $f:{\cal D}\to {\cal D}'$ in $Ext({\cal H})$ to the restriction of the frame homomorphism $B(f^{\textrm{op}}):Id({{\cal D}'}) \to Id({\cal D})$ to the subsets of $U$-compact elements of $Id({\cal D}')$ and of $Id({\cal D})$.
\begin{theorem}\label{dualabstract}
Under the hypotheses specified above, the functors $D$ and $E$ are categorical inverses to each other and hence they define a duality between $Ext({\cal K})$ and $Ext({\cal H})$.
\end{theorem}
\begin{proofs}
The fact that the functors $D$ and $E$ are well-defined, in the sense that $D$ takes values in $Ext({\cal H})$ (and, dually, $E$ takes values in $Ext({\cal K})^{\textrm{op}}$), and that $D$ and $E$ are categorical inverses to each other easily follows from Theorem \ref{thmcentral} by invoking the general theory of morphisms of sites (as in the proof of Theorem \ref{dualcov}), and from the theory of geometric morphisms induced by functors as in A4.1.4 \cite{El}. Concerning the latter, one specifically uses the fact (Lemma A4.1.5 \cite{El}) that, given two Cauchy-complete categories $\cal C$ and $\cal D$, a geometric morphism $[{\cal C}, \Set]\to [{\cal D}, \Set]$ is induced by a functor $f:{\cal C}\to {\cal D}$ as in A4.1.4 \cite{El} if and only if it is essential.
\end{proofs}
Similarly to the covariant case, we can deduce from the theorem the following characterizations of the categories $Ext({\cal K})$ and $Ext({\cal H})$ in terms of each other.
\begin{proposition}
Let us assume that the hypotheses of Theorem \ref{dualcov} are satisfied, and that $U$ is $\cal K$-compatible relative to a property $P$. Let ${\cal F}^{U}_{P}$ denote the set of frames such that the collection of their $U$-compact elements forms a basis of them which, endowed with the induced order, has the structure of a poset in $Ext({\cal K})$ and has the property that the embedding $B_{L}\hookrightarrow L$ of it into the frame satisfies condition $P$, the property that every covering in $L$ of an element of $B_{L}$ is refined by a covering made of elements of $B_{L}$ which satisfies the invariant $U$, and the property that the $J_{B_{L}}$-covering sieves are sent by the embedding $B_{L}\hookrightarrow L$ into covering families in $L$. Let ${\cal F}_{sc}$ denote the set of frames such that the collection of their supercompact elements forms a basis which, endowed with the induced order, has the structure of a poset in $Ext({\cal H})$. Then
\begin{enumerate}[(i)]
\item The objects of $Ext({\cal H})$ are precisely the posets of supercompact elements of the frames in ${\cal F}^{U}_{P}$;
\item The arrows in $Ext({\cal H})$ are the restrictions to the subsets of supercompact elements of the left adjoints to the frame homomorphisms between frames ${\cal F}^{U}_{P}$ which send $U$-compact elements to $U$-compact elements;
\item The objects of $Ext({\cal K})$ are precisely the posets of $U$-compact objects of the frames in ${\cal F}_{sc}$;
\item The arrows in $Ext({\cal H})$ are the restrictions to the subsets of $U$-compact elements of the frame homomorphisms between frames in ${\cal F}_{sc}$ which are complete.
\end{enumerate}
\end{proposition}
\end{enumerate}
Note that all of this represents a clear implementation of the philosophy `toposes as bridges' of \cite{OC10}.
The process of `lifting' of dualities with categories of locales to dualities with categories of topological spaces of section \ref{dualtop} has an analogue in this more general context. Again, we have to distinguish between the covariant and contravariant cases:
\begin{enumerate}[(i)]
\item \emph{Covariant case}
Suppose that we have assigned to each structure $\cal C$ in $Ext({\cal K})$ a separating set of points of the topos $\Sh({\cal C}, J_{\cal C})$ functorially in $\cal C$, as in section \ref{dualtop}. We define a subcategory $\textbf{Top}_{\cal H}$ of $\textbf{Top}$ by taking as objects the topological spaces $X$ such that $V\textrm{-comp}({\cal O}(X))$ belongs to $Ext({\cal H})$ and as arrows $X\to Y$ are the continuous maps $f:X\to Y$ between spaces in $\textbf{Top}_{\cal H}$ with the property that $f^{-1}:{\cal O}(Y)\to {\cal O}(X)$ sends $V$-compact open sets to $V$-compact open sets in such a way that the restriction $f^{-1}|:V\textrm{-comp}({\cal O}(Y)) \to V\textrm{-comp}({\cal O}(X))$ is an arrow $V\textrm{-comp}({\cal O}(Y)) \to V\textrm{-comp}({\cal O}(X))$ in $Ext({\cal H})$. Thus we have, as in section \ref{dualtop}, a functor $\tilde{D}:Ext({\cal K})\to \textbf{Top}_{\cal H}^{\textrm{op}}$, and we can define an essentially surjective functor (both on the objects and on the arrows) $U_{\cal H}:\textbf{Top}_{\cal H}^{\textrm{op}} \to Ext({\cal H})$ such that $U_{{\cal H}}\circ \tilde{D}\cong D$; $U_{\cal H}$ sends a topological space $X$ to the poset of $V$-compact open subsets of $X$ and a continuous map $f:X\to Y$ in $\textbf{Top}_{\cal H}$ to the restriction of the inverse image $f^{-1}$ to the subsets of $V$-compact open sets of $X$ and $Y$.
Of course, we can define a subcategory $\textbf{Top}_{\cal K}$ of $\textbf{Top}$ and a functor $U_{\cal K}:\textbf{Top}_{\cal K}^{\textrm{op}} \to Ext({\cal K})$ in a perfectly symmetrical way.
\item \emph{Contravariant case}
Suppose that we have assigned to each structure $\cal C$ in $Ext({\cal K})$ a separating set of points of the topos $\Sh({\cal C}, J_{\cal C})$ functorially in $\cal C$, as in section \ref{dualtop}. Let us define a subcategory $\textbf{Top}_{\cal H}$ of $\textbf{Top}$ by taking as objects the topological spaces $X$ such that the poset $\textbf{SC}({\cal O}(X))$ of supercompact elements of ${\cal O}(X)$ belongs to $Ext({\cal H})$ and as arrows $X\to Y$ the continuous maps $f:X\to Y$ between spaces in $\textbf{Top}_{\cal H}$ such that $f^{-1}:{\cal O}(Y)\to {\cal O}(X)$ is complete and its left adjoint $f_{!}$ sends supercompact open sets to supercompact open sets in such a way that the restriction $f_{!}|:\textbf{SC}({\cal O}(X)) \to \textbf{SC}({\cal O}(Y))$ is an arrow $\textbf{SC}({\cal O}(X)) \to \textbf{SC}({\cal O}(Y))$ in $Ext({\cal H})$. Thus we have, as in section \ref{dualtop}, a functor $\tilde{D}:Ext({\cal K})^{\textrm{op}}\to \textbf{Top}_{\cal H}$, and we can define an essentially surjective functor (both on the objects and on the arrows) $U:\textbf{Top}_{\cal H} \to Ext({\cal H})$ such that $U\circ \tilde{D}\cong D$; $U$ sends a topological space $X$ to the poset of supercompact open subsets of $X$ and a continuous map $f:X\to Y$ in $\textbf{Top}_{\cal H}$ to the restriction of the left adjoint $f_{!}$ to the subsets of supercompact open sets of $X$ and $Y$.
We can make a similar construction for the category $Ext({\cal H})$, as follows. Suppose that we have assigned to each structure $\cal D$ in $Ext({\cal H})$ a separating set of points of the topos $\Sh({\cal D}, K_{\cal D})\simeq [{\cal D}^{\textrm{op}}, \Set]$ functorially in $\cal D$, as in section \ref{dualtop}. We define a subcategory $\textbf{Top}_{\cal K}$ of $\textbf{Top}$ by taking as objects the topological spaces such that $U\textrm{-comp}({\cal O}(X))$ belongs to $Ext({\cal K})$ and as arrows $X\to Y$ the continuous maps $f:X\to Y$ between spaces in $\textbf{Top}_{\cal K}$ with the property that $f^{-1}:{\cal O}(Y)\to {\cal O}(X)$ sends $U$-compact open sets to $U$-compact open sets in such a way that the restriction $f^{-1}|:U\textrm{-comp}({\cal O}(Y)) \to U\textrm{-comp}({\cal O}(X))$ is an arrow $U\textrm{-comp}({\cal O}(Y)) \to U\textrm{-comp}({\cal O}(X))$ in $Ext({\cal K})$. Thus we have, as in section \ref{dualtop}, a functor $\tilde{E}:Ext({\cal H})\to \textbf{Top}_{\cal K}$, and we can define an essentially surjective functor (both on the objects and on the arrows) $W:\textbf{Top}_{\cal K} \to Ext({\cal K})^{\textrm{op}}$ such that $W\circ \tilde{E}\cong E$; $W$ sends a topological space $X$ to the poset $U\textrm{-comp}({\cal O}(X))$ of $U$-compact open subsets of $X$ and a continuous map $f:X\to Y$ in $\textbf{Top}_{\cal K}$ to the restriction of the inverse image $f^{-1}$ to the subsets of $U$-compact open sets of $X$ and of $Y$.
\end{enumerate}
The topological spaces in the images of the `lifting functors' can be directly described in terms of the structures in the source category (cf. Proposition \ref{mslattice}); so the above constructions allow characterizations of the posets in the target categories as subsets consisting of the open sets of a topological space satisfying some generalized compactness condition.
We notice that our original framework of sections \ref{locales} and \ref{dualtop} sits inside this more general setting as the particular case in which each of the toposes associated to the structures in $\cal K$ (or in $\cal H$) is of the form $\Sh({\cal C}, J_{\cal C})$, where $\cal C$ is a locale and $J_{\cal C}$ is the canonical topology on it, and the Morita-equivalences from which the dualities or equivalences are built are instances of Theorem \ref{fund}. Note that the canonical topologies $J_{\cal C}$ are all $U$-induced where $U$ is the vacuous invariant `to be a family' (note that this invariant trivially satisfies the hypotheses of Theorem \ref{construction}); in particular, the condition of $U$-compactness is vacuous in this case.
\subsection{The role of the Comparison Lemma}
It is worth to remark the key role of Grothendieck's Comparison Lemma (cf. Theorem C2.2.3 \cite{El}) in the framework that we have developed.
The equivalence $\Sh({\cal C}, J) \simeq \Sh(Id_{J}({\cal C}))$ of Theorem \ref{fund} is, when $J$ is subcanonical, an instance of an equivalence induced by the Comparison Lemma. Indeed, if $J$ is subcanonical then $\cal C$ can be identified with a full subcategory of $\Sh(Id_{J}({\cal C}))$ which is dense with respect to the canonical coverage, and the Grothendieck topology induced by it on $\cal C$ is precisely $J$ (cf. Proposition C2.2.16(ii) \cite{El}).
So, all of the dualities established so far arise in fact from instances of the Comparison Lemma. Generalizing our original situation, we can expect to be able to obtain representation theorems for preordered structures, as well as new dualities or equivalences, starting from Morita-equivalences provided by the Lemma, following the method of section \ref{Mordual}. Without embarking on a comprehensive treatment of these more general situations (in fact, such an investigation would bring us far beyond the scope of the present paper), we limit ourselves to presenting a few examples in the next section.
\subsection{Examples}\label{addex}
In this section we give some examples of results obtained through the method of section \ref{Mordual} by applying it to Morita-equivalences arising from instances of the Comparison Lemma which are not particular cases of Theorem \ref{fund}. There are of course a great number of other dualities that can be established in a semi-automatic way by applying the technique of section \ref{Mordual}; these examples are just meant to give a flavor of the results that arise from the application of our machinery.
We have discussed in section \ref{tarski} the representation of an atomic frame as the powerset on the collection of its atoms. Similarly, one can obtain a representation theorem for \emph{atomic distributive lattices} i.e. distributive lattices in which every element is a finite join of atoms: if $\cal D$ is such a lattice and $J_{\cal D}$ is the coherent topology on it then the Comparison Lemma yields an equivalence of toposes $\Sh({\cal D}, J_{\cal D})\simeq [At({\cal D}), \Set]$, from which it follows, invoking Corollary \ref{cor}(i), that $\cal D$ is isomorphic to the the lattice of compact elements of the powerset of its collection of atoms (equivalently, to the lattice of finite subsets of the set of its atoms). If we take $\cal K$ to be the collection of atomic distributive lattices, each of which equipped with the coherent topology, and $\cal H$ to be the collection of the (discrete) posets of the form $At({\cal D})$ for some atomic distributive lattice $\cal D$, each of which equipped with the trivial Grothendieck topology, then the functions $f:{\cal K} \to {\cal H}$ and $g:{\cal H}\to {\cal K}$ defined by setting $f({\cal D})=At({\cal D})$ (for any atomic distributive lattice $\cal D$) and $g(A)={\mathscr{P}}_{fin}(A)$ (for any set $A$) are inverse to each other (up to isomorphism), and the closure under isomorphisms of $\cal H$ can be identified with the collection of all the sets (since for any set $A$, ${\mathscr{P}}_{fin}(A)$ is an atomic distributive lattice whose set of atoms is isomorphic to $A$).
Now, the category $Ext({\cal K})$ is the category $\textbf{AtDLat}$ whose objects are the atomic distributive lattices and whose arrows ${\cal D}\to {\cal D}'$ are the distributive lattices homomorphisms $f:{\cal D}\to {\cal D}'$ between them such that the frame homomorphism $A(f):Id_{J_{\cal D}}({\cal D})\to Id_{J_{\cal D}'}({\cal D}')$ which sends an ideal $I$ of $\cal D$ to the ideal of ${\cal D}'$ generated by $f(I)$ preserves arbitrary infima.
The category $Ext({\cal H})$ is the category $\Set_{f}$ whose objects are the sets and whose arrows $A\to A'$ are the functions $g:A\to A'$ such that the inverse image $g^{-1}:{\mathscr{P}}(A') \to {\mathscr{P}}(A)$ sends finite subsets of $A'$ to finite subsets of $A$.
Theorem \ref{dualabstract} thus yields two functors $D:\textbf{AtDLat}^{\textrm{op}}\to \Set_{f}$ and $E:\Set_{f} \to \textbf{AtDLat}^{\textrm{op}}$ which are categorical inverses to each other, and which therefore give a duality between $\textbf{AtDLat}$ and $\Set_{f}$. The functor
\[
D:\textbf{AtDLat}^{\textrm{op}}\to \Set_{f}
\]
sends a poset $\cal D$ in $\textbf{AtDLat}$ to the set $At({\cal D})$ of its atoms, and an arrow $f:{\cal D}\to {\cal D}'$ in $\textbf{AtDLat}$ to the restriction $At({\cal D}')\to At({\cal D})$ of the left adjoint to $A(f):Id_{J_{\cal D}}({\cal D})\to Id_{J_{\cal D}'}({\cal D}')$ to the sets of atoms of ${\cal D}'$ and of ${\cal D}$. The functor
\[
E:\Set_{f} \to \textbf{AtDLat}^{\textrm{op}}
\]
sends a set $A$ to the finite powerset ${\mathscr{P}}_{fin}(A)$ and an arrow $f:A\to A'$ in $\Set_{f}$ to the restriction ${\mathscr{P}}_{fin}(A')\to {\mathscr{P}}_{fin}(A)$ of the inverse image $f^{-1}:{\mathscr{P}}(A')\to {\mathscr{P}}(A)$.
Summarizing, we have the following result.
\begin{theorem}\label{atomfin}
Via the functors
\[
D:\textbf{AtDLat}^{\textrm{op}}\to \Set_{f}
\]
and
\[
E:\Set_{f} \to \textbf{AtDLat}^{\textrm{op}}
\]
defined above, the categories $\textbf{AtDLat}$ and $\Set_{f}$ are dual to each other.
\end{theorem}\qed
Notice that this duality restricts to a duality between the category of finite atomic distributive lattices and distributive lattice homomorphisms between them and the category of finite sets and functions between them.
One might wonder why we have not considered, instead of the class of atomic distributive lattices, the class of weakly atomic meet-semilattices in which every element is a finite join of atoms. In fact, the two classes coincide with each other: if $\cal D$ is a weakly atomic meet-semilattices in which every element is a join of atoms then, by equipping it with the atomically generated topology $J_{\cal D}$, the Comparison Lemma yields an equivalence $\Sh({\cal D}, J_{{\cal D}})\simeq [At({\cal D}), \Set]$ from which it follows, invoking Corollary \ref{cor}(v), that $\cal D$ is isomorphic to the poset of atomically compact elements of ${\mathscr{P}}(At({\cal D}))$, in other words to ${\mathscr{P}}_{fin}(At({\cal D}))$.
In fact, we can establish a more general duality, which extends that of Theorem \ref{atomfin} and represents the `finitary version' of the duality of Theorem \ref{proalex}. Given a distributive lattice $\cal D$, we say that an element $d$ of $\cal D$ is \emph{join-irreducible} if for any $a,b\in {\cal D}$, $a\vee b =d$ implies $a=d$ or $b=d$; we denote the set of join-irreducible elements of a distributive lattice $\cal D$ by $Irr({\cal D})$. We say that a distributive lattice $\cal D$ is \emph{irreducibly generated} if every element of $\cal D$ can be expressed as a finite join of join-irreducible elements. If $\cal D$ is such a lattice and $J_{\cal D}$ is the coherent topology on it then the Comparison Lemma yields an equivalence of toposes $\Sh({\cal D}, J_{\cal D})\simeq [Irr({\cal D})^{\textrm{op}}, \Set]$, from which it follows, invoking Corollary \ref{cor}(i), that $\cal D$ is isomorphic to the the lattice of compact elements of the frame $Id({Irr({\cal D})})$ of ideals on $Irr({\cal D})$.
Take $\cal K$ to be the collection of irreducibly generated distributive lattices, each of which equipped with the coherent topology, and $\cal H$ to be the collection of the posets of the form $Irr({\cal D})$ for some irreducibly generated distributive lattice, each of which equipped with the trivial Grothendieck topology; then the functions $f:{\cal K} \to {\cal H}$ and $g:{\cal H}\to {\cal K}$ defined by setting $f({\cal D})=Irr({\cal D})$ (for any atomic distributive lattice $\cal D$) and the image $g({\cal P})$ of a poset $\cal P$ in $\cal H$ under $g$ equal to the poset $Id_{comp}({\cal P})$ of compact elements of the frame $Id({\cal P})$ are inverse to each other (up to isomorphism), and the closure under isomorphisms of $\cal H$ can be identified with the collection of all the posets such that $Id_{comp}({\cal P})$ is closed under finite intersections in $Id({\cal P})$ (since for such poset $\cal P$, $Id_{comp}({\cal P})$ is an irreducibly generated distributive lattice whose poset of join-irreducible elements is isomorphic to $\cal P$). Notice that an ideal $I$ on $\cal P$ belongs to $Id_{comp}({\cal P})$ if and only if it is a finite union of principal ideals on $\cal P$, equivalently if there exists a finite set of elements $\{a_{k} \textrm{ | } k\in K\}$ of $I$ such that for any $p\in {\cal P}$, $p\in I$ if and only if $p\leq a_{k}$ for some $k\in K$. The condition that $Id_{comp}({\cal P})$ be closed under finite intersections in $Id({\cal P})$ is clearly equivalent to the requirement that the intersection in $Id({\cal P})$ of any two principal ideals on $\cal P$ be equal to the union of a finite family of principal ideals on $\cal P$, equivalently for any $a,b\in {\cal P}$ there exist a finite set of elements $\{c_{k} \textrm{ | } k\in K\}$ of $\cal P$ such that for any $p\in {\cal P}$, $p\leq a$ and $p\leq b$ if and only if $p\leq c_{k}$ for some $k\in K$. In particular, all the meet-semilattices are objects of the closure under isomorphisms of $\cal H$.
Now, the category $Ext({\cal K})$ is the category $\textbf{IrrDLat}$ whose objects are the irreducibly generated distributive lattices and whose arrows ${\cal D}\to {\cal D}'$ are the distributive lattices homomorphisms $f:{\cal D}\to {\cal D}'$ between them such that the frame homomorphism $A(f):Id_{J_{\cal D}}({\cal D})\to Id_{J_{\cal D}'}({\cal D}')$ which sends an ideal $I$ of $\cal D$ to the ideal of ${\cal D}'$ generated by $f(I)$ preserves arbitrary infima.
The category $Ext({\cal H})$ is the category $\textbf{Pos}_{comp}$ whose objects are the posets $\cal P$ such that for any $a,b\in {\cal P}$ there exists a finite set of elements $\{c_{k} \textrm{ | } k\in K\}$ of $\cal P$ such that for any $p\in {\cal P}$, $p\leq a$ and $p\leq b$ if and only if $p\leq c_{k}$ for some $k\in K$, and whose arrows ${\cal P}\to {\cal P}'$ are the monotone maps $g:{\cal P}\to {\cal P}'$ such that the inverse image $g^{-1}:Id({\cal P}') \to Id({\cal P})$ sends ideals in $Id_{comp}({\cal P}')$ to ideals in $Id_{comp}({\cal P})$, equivalently, for any $q\in {\cal P}'$, there exists a finite family $\{a_{k} \textrm{ | } k\in K\}$ of elements of $\cal P$ such that for any $p\in {\cal P}$, $g(p)\leq q$ if and only if $p\leq a_{k}$ for some $k\in K$.
Theorem \ref{dualabstract} thus yields two functors $D:\textbf{IrrDLat}^{\textrm{op}}\to \textbf{Pos}_{comp}$ and $E:\textbf{Pos}_{comp} \to \textbf{IrrDLat}^{\textrm{op}}$ which are categorical inverses to each other, and which therefore form a duality between $\textbf{IrrDLat}$ and $\textbf{Pos}_{comp}$. The functor $D:\textbf{IrrDLat}^{\textrm{op}}\to \textbf{Pos}_{comp} $ sends a poset $\cal D$ in $\textbf{IrrDLat}$ to the set $Irr({\cal D})$ of its join-irreducible elements, and an arrow $f:{\cal D}\to {\cal D}'$ in $\textbf{IrrDLat}$ to the restriction $Irr({\cal D}')\to Irr({\cal D})$ of the left adjoint to $A(f):Id_{J_{\cal D}}({\cal D})\to Id_{J_{\cal D}'}({\cal D}')$ to the sets of join-irreducible elements of ${\cal D}'$ and of ${\cal D}$. The functor $E:\textbf{Pos}_{comp} \to \textbf{IrrDLat}^{\textrm{op}}$ sends a poset $\cal P$ to $Id_{comp}({\cal P})$ and an arrow $g:{\cal P}\to {\cal P}'$ in $\textbf{Pos}_{comp}$ to the restriction $Id_{comp}({\cal P}')\to Id_{comp}({\cal P})$ of the inverse image $g^{-1}:Id({\cal P}') \to Id({\cal P})$.
Summarizing, we have the following result.
\begin{theorem}\label{birk}
Via the functors
\[
D:\textbf{IrrDLat}^{\textrm{op}}\to \textbf{Pos}_{comp}
\]
and
\[
E:\textbf{Pos}_{comp} \to \textbf{IrrDLat}^{\textrm{op}}
\]
defined above, the categories $\textbf{IrrDLat}$ and $\textbf{Pos}_{comp}$ are dual to each other.
\end{theorem}\qed
\begin{remark}
If $\cal D$ is a finite distributive lattice then \emph{a fortiori} $\cal D$ is irreducibly generated and has a finite set of join-irreducible elements; conversely, given a finite poset $\cal P$, the lattice $Id_{comp}({\cal P})$ is finite, and any monotone map ${\cal P}\to {\cal P}'$ between finite posets $\cal P$ and ${\cal P}'$ is an arrow in $\textbf{Pos}_{comp}$. We thus conclude that the duality of Theorem \ref{birk} restricts to Birkhoff's duality between finite distributive lattices and finite posets.
\end{remark}
Another example in the same style is given by disjunctively distributive frames.
Let $F$ be a disjunctively distributive frame. An element $a\in F$ is said to be \emph{indecomposable} if for any family $\{a_{i} \textrm{ | } i\in I\}$ of pairwise disjoint elements of $F$, $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}} a_{i}=a$ implies $a_{i}=a$ for some $i\in I$. Notice that, classically, an element $a\in F$ is indecomposable if and only if it is non-zero and \emph{connected} (i.e., for any elements $b,c \in F$ such that $b\wedge c=0$, $b\vee c=a$ implies that either $a=b$ or $a=c$). Indeed, to be indecomposable clearly implies to be non-zero and connected, while the converse can be proved as follows. Given a connected element $a\neq 0$, for any family $\{a_{i} \textrm{ | } i\in I\}$ of pairwise disjoint elements of $F$, $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}} a_{i}=a$ implies that for any $i\in I$, $a_{i} \vee (\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{j\in I, j\neq i}} a_{j})=a$, from which it follows, $a$ being connected, that either $a=a_{i}$ or $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{j\in I, j\neq i}} a_{j}$. But, since $a\neq 0$, an equality of the latter kind cannot hold for all $i\in I$, since $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{j\in I, j\neq i}} a_{j}$ implies $a\wedge a_{i}=0$; this shows that there is $i\in I$ such that $a=a_{i}$, as required.
Let us suppose that $F$ is a \emph{disjunctive frame}, i.e. a disjunctively distributive frame in which every element can be written as a pairwise disjoint join of indecomposable elements. Then, denoted by $J_{F}$ the infinitary disjunctive topology on $F$ and by ${\cal I}_{F}$ the full subcategory of $F$ on the indecomposable elements of $F$, we have that ${\cal I}_{F}$ is $J_{F}$-dense; hence the Comparison Lemma yields an equivalence $\Sh(F, J_{F})\simeq \Sh({\cal I}_{F}, J_{F}|_{{\cal I}_{F}})$. But, since the objects of ${\cal I}_{F}$ are indecomposable, $J_{F}|_{{\cal I}_{F}}$ is the trivial Grothendieck topology on ${\cal I}_{F}$, and hence we have an equivalence $\Sh(F, J_{F})\simeq [{\cal I}_{F}^{\textrm{op}}, \Set]$, which implies by Corollary \ref{cor}(v), that $F$ is isomorphic to the poset of infinitarily disjunctively compact elements of the frame of ideals on the poset ${\cal I}_{F}$ of indecomposable elements of $F$.
Starting from the representations $\Sh(F, J_{F})\simeq [{\cal I}_{F}^{\textrm{op}}, \Set]$, holding for any disjunctive frame $F$, we can build a duality applying Theorem \ref{dualabstract}.
We take $\cal K$ to be the collection of all the disjunctive frames, each of which equipped with the infinitary disjunctive topology, and $\cal H$ to be the collection of all the posets of the form ${\cal I}_{F}$ for a disjunctive frame $F$, each of which equipped with the trivial topology. Define $f:{\cal K}\to {\cal H}$ as the function sending a disjunctive frame $F$ to the poset ${\cal I}_{F}$ and $g:{\cal H}\to {\cal K}$ as the function which sends a poset $P$ in $\cal H$ to the frame of infinitarily disjunctively compact elements of the frame $Id(P)$ of ideals on the poset $P$. Clearly, $f$ and $g$ are inverses to each other, up to isomorphism.
Now, the category $Ext({\cal K})$ is the category $\textbf{DisFrm}$ whose objects are the disjunctive frames and whose arrows $F\to F'$ are the meet-semilattice homomorphisms $f:F\to F'$ between them which send pairwise disjoint joins to pairwise disjoint joins and have the property that the frame homomorphism $A(f):Id_{J_{F}}(F)\to Id_{J_{F'}}(F')$ sending an ideal $I$ of $F$ to the ideal of $F'$ generated by $f(I)$ preserves arbitrary infima.
The category $Ext({\cal H})$ is the category $\textbf{Pos}_{dis}$ whose objects are the posets $\cal P$ such that their frames of ideals $Id({\cal P})$ have the property that the collection of infinitarily disjunctively compact elements forms a basis of them which is closed under finite meets, and whose arrows $P\to P'$ are the monotone maps $g:{\cal P}\to {\cal P}'$ such that the inverse image $g^{-1}:Id({\cal P}') \to Id({\cal P})$ sends infinitarily disjunctively compact elements to infinitarily disjunctively compact elements.
We can describe the category $\textbf{Pos}_{dis}$ more explicitly as follows. First, we note that for any poset $\cal P$, an ideal $I$ in $Id({\cal P})$ is infinitarily disjunctively compact in $Id({\cal P})$ if and only if it is a disjoint union of principal ideals on $\cal P$, equivalently if there exists a family $\{c_{i} \textrm{ | } i\in I\}$ of elements of $I$ such that for any $p\in {\cal P}$, $p\in I$ if and only if $p\leq c_{i}$ for a unique $i\in I$; indeed, the `if' direction follows from Proposition \ref{multicomposition}, while the `only if' direction follows from the fact that if $I$ is infinitarily disjunctively compact then the covering of $I$ formed by the principal ideals on $\cal P$ generated by elements in $I$ has a disjunctive refinement by principal ideals on $\cal P$ (since the invariant `to be a disjoint family' satisfies the condition in the statement of Theorem \ref{construction}). Now, all the principal ideals on $\cal P$ are infinitarily disjunctively compact elements of $Id({\cal P})$, and hence from the characterization above it follows that the condition that the intersection of two infinitarily disjunctively compact elements of $Id({\cal P})$ should be infinitarily disjunctively compact is equivalent to the requirement that the intersection of any two principal ideals on $\cal P$ should be equal to a disjoint union of principal ideals on $\cal P$, equivalently that for any $a,b\in {\cal P}$ there should exist a family $\{c_{i} \textrm{ | } i\in I\}$ of elements of $\cal P$ such that for any $p\in {\cal P}$, $p\leq a$ and $p\leq b$ if and only if $p\leq c_{i}$ for a unique $i\in I$. Note that if $\cal P$ is a meet-semilattice than this condition is always satisfied (take the family $\{c_{i} \textrm{ | } i\in I\}$ to be equal to the singleton family $\{a\wedge b\}$). Concerning the arrows in $\textbf{Pos}_{dis}$ we note that, given a monotone map $g:{\cal P}\to {\cal P}'$, the inverse image $g^{-1}:Id({\cal P}') \to Id({\cal P})$ sends infinitarily disjunctively compact elements to infinitarily disjunctively compact elements if and only if it sends any principal ideal on ${\cal P}'$ to an infinitarily disjunctively compact elements, equivalently if for any $b\in {\cal P}'$ there exists a family $\{c_{i} \textrm{ | } i\in I\}$ of elements of $\cal P$ such that for any $p\in {\cal P}$, $g(p)\leq b$ if and only if $p\leq c_{i}$ for a unique $i\in I$.
Summarizing, the category $\textbf{Pos}_{dis}$ has as objects the posets $\cal P$ such that for any $a,b\in {\cal P}$ there exists a family $\{c_{i} \textrm{ | } i\in I\}$ of elements of $\cal P$ such that for any $p\in {\cal P}$, $p\leq a$ and $p\leq b$ if and only if $p\leq c_{i}$ for a unique $i\in I$ and as arrows ${\cal P}\to {\cal P}'$ the monotone maps $g:{\cal P}\to {\cal P}'$ such that for any $b\in {\cal P}'$ there exists a family $\{c_{i} \textrm{ | } i\in I\}$ of elements of $\cal P$ such that for any $p\in {\cal P}$, $g(p)\leq b$ if and only if $p\leq c_{i}$ for a unique $i\in I$.
Our Theorem \ref{dualabstract} thus yields two functors $D:\textbf{DisFrm}^{\textrm{op}}\to \textbf{Pos}_{dis}$ and $E:\textbf{Pos}_{dis} \to \textbf{DisFrm}^{\textrm{op}}$ which are categorical inverses to each other and hence form a duality between $\textbf{DisFrm}$ and $\textbf{Pos}_{dis}$. The functor $D:\textbf{DisFrm}^{\textrm{op}}\to \textbf{Pos}_{dis}$ sends a poset $F$ in $\textbf{DisFrm}$ to the poset ${\cal I}_{F}$ of indecomposable elements of $F$ and an arrow $f:F\to F'$ in $\textbf{DisFrm}$ to the restriction ${\cal I}_{F'} \to {\cal I}_{F}$ of the left adjoint to $A(f):Id_{J_{F}}(F)\to Id_{J_{F'}}(F')$ to the sets of indecomposable elements of $F'$ and of $F$. The functor $E:\textbf{Pos}_{dis} \to \textbf{DisFrm}^{\textrm{op}}$ sends a poset $\cal P$ in $\textbf{Pos}_{dis}$ to the set ${\cal P}_{dis}$ of ideals $I$ on $\cal P$ with the property that there exists a family $\{c_{i} \textrm{ | } i\in I\}$ of elements of $I$ such that for any $p\in {\cal P}$, $p\in I$ if and only if $p\leq c_{i}$ for a unique $i\in I$, endowed with the subset-inclusion ordering, and an arrow $g:{\cal P}\to {\cal P}'$ in $\textbf{Pos}_{dis}$ to the restriction ${{\cal P}'}_{dis}\to {\cal P}_{dis}$ of the inverse image $g^{-1}:Id({\cal P}')\to Id({\cal P})$.
In conclusion, we have the following result.
\begin{theorem}
Via the functors
\[
D:\textbf{DisFrm}^{\textrm{op}}\to \textbf{Pos}_{dis}
\]
and
\[
E:\textbf{Pos}_{dis} \to \textbf{DisFrm}^{\textrm{op}}
\]
defined above, the categories $\textbf{DisFrm}$ and $\textbf{Pos}_{dis}$ are dual to each other.
\end{theorem}\qed
Note that this duality restricts to a duality between the full subcategory of $\textbf{DisFrm}^{\textrm{op}}$ on the posets $F$ such that the meet of two indecomposable elements of $F$ is indecomposable and the full subcategory of $\textbf{Pos}_{dis}$ on the meet-semilattices.
By using the finitary version of the disjunctive topology, one can obtain a similar duality for disjunctively distributive lattices in which every element can be written as a finite pairwise disjoint join of connected elements.
For disjunctive frames with the property that the meet of any two indecomposable elements is indecomposable we can `functorialize' the representation $\Sh(F, J_{F})\simeq [{\cal I}_{F}^{\textrm{op}}, \Set]$ in a covariant way, by using Theorem \ref{dualcov}. Let $\cal K$ be the collection of such frames, each of which equipped with the infinitary disjunctive topology, and let $\cal H$ be the collection of all the posets (in fact, meet-semilattices) of the form ${\cal I}_{F}$ for a disjunctive frame $F$ in $\cal K$, each of which equipped with the trivial topology. Let $f:{\cal K}\to {\cal H}$ and $g:{\cal H}\to {\cal K}$ be defined exactly as in the example above.
Using the covariant method for `functorializing' Morita-equivalences, we obtain the categories $Ext({\cal K})$ and $Ext({\cal H})$ defined as follows.
The category $Ext({\cal K})$ coincides with the category $\textbf{DIFrm}$ having as objects the disjunctive frames $F$ such that the meet in $F$ of any two indecomposable elements of $F$ is indecomposable and as arrows $F\to F'$ the meet-semilattice homomorphisms $f:F\to F'$ between them which send pairwise disjoint joins to pairwise disjoint joins and have the property that the frame homomorphism $A(f):Id_{J_{F}}(F)\to Id_{J_{F'}}(F')$ sending an ideal $I$ in $Id_{J_{F}}(F)$ to the $J_{F'}$ ideal on $F'$ generated by $f(I)$ sends supercompact elements to supercompact elements. We note that, for any disjunctive frame $F$, since $0_{F}$ is not indecomposable in $F$, if the meet in $F$ of two indecomposable elements is indecomposable then it is non-zero; this implies that the disjunctive frames in $F$ can be identified with the meet-semilattices $F$ with a bottom element $0_{F}$ which have the property that for any $a, b\in F$, $a\wedge b=0$ implies that either $a=0$ or $b=0$; under this identification, the arrows $F\to F'$ in $\textbf{DIFrm}$ correspond precisely to the meet-semilattice homomorphisms $F\to F'$ which send $0_{F}$ to $0_{F'}$ and any non-zero element of $F$ to a non-zero element of $F'$.
Let us denote by $\textbf{MSLat}^{\ast}$ the category having as objects the meet-semilattices $\cal P$ which have a bottom element $0_{\cal P}$ and satisfy the property that for any $a, b\in F$, $a\wedge b=0$ implies that either $a=0$ or $b=0$, and as arrows ${\cal P}\to {\cal P}'$ the meet-semilattice homomorphisms ${\cal P}\to {\cal P}'$ which send $0_{{\cal P}}$ to $0_{{\cal P}'}$ and any non-zero element of ${\cal P}$ to a non-zero element of ${\cal P}'$.
The category $Ext({\cal H})$ coincides with the category $\textbf{MPos}$ which has as objects the meet-semilattices $\cal P$ such that their frames of ideals $Id({\cal P})$ have the property that the collection of infinitarily disjunctively compact elements forms a basis of them which is closed under finite meets, and whose arrows $P\to P'$ are the meet-semilattice homomorphisms $g:{\cal P}\to {\cal P}'$ such that the homomorphism $A(g):Id({\cal P}) \to Id({\cal P}')$ sends infinitarily disjunctively compact elements to infinitarily disjunctively compact elements. From the characterization of the category $\textbf{Pos}_{dis}$ obtained above, we thus conclude that the category $\textbf{MPos}$ coincides with the category $\textbf{MSLat}$ of meet-semilattices and meet-semilattice homomorphisms between them.
Theorem \ref{dualcov} thus yields two functors
\[
D:\textbf{MSLat}^{\ast}\to \textbf{MSLat}
\]
and
\[
E:\textbf{MSLat} \to \textbf{MSLat}^{\ast}
\]
which are categorical inverses to each other. The functor $D:\textbf{MSLat}^{\ast}\to \textbf{MSLat}$ sends a meet-semilattice ${\cal P}$ in $\textbf{MSLat}^{\ast}$ to the meet-semilattice ${\cal P}^{\ast}$ of its non-zero elements, and an arrow $f:{\cal P}\to {\cal P}'$ in $\textbf{MSLat}^{\ast}$ to its restriction ${\cal P}^{\ast} \to {{\cal P}'}^{\ast}$. The functor $E:\textbf{MSLat} \to \textbf{MSLat}^{\ast}$ sends a meet-semilattice $\cal P$ to the set $Id_{\ast}({\cal P})$ of the ideals on $\cal P$ which are either principal or empty, endowed with the subset-inclusion ordering, and a meet-semilattice homomorphism $g:{\cal P}\to {\cal P}'$ to the restriction $Id_{\ast}({\cal P}) \to Id_{\ast}({\cal P}')$ of $A(g):Id({\cal P})\to Id({\cal P}')$.
Summarizing, we have the following result.
\begin{theorem}
Via the functors
\[
D:\textbf{MSLat}^{\ast}\to \textbf{MSLat}
\]
and
\[
E:\textbf{MSLat} \to \textbf{MSLat}^{\ast}
\]
defined above, the categories $\textbf{MSLat}^{\ast}$ and $\textbf{MSLat}$ are equivalent to each other.
\end{theorem}\qed
As another example, we construct a duality for a natural class of preframes including all the algebraic lattices.
Given a preframe $\cal D$, we say that an element $d\in {\cal D}$ is \emph{directedly irreducible} if any directed sieve on $d$ is maximal, i.e. if for any directed family $\{a_{i} \textrm{ | } i\in I\}$ of elements of $\cal D$ such that $d=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ there exists $i\in I$ such that $d=a_{i}$; given a preframe $\cal D$, we denote by $DirIrr({\cal D})$ the poset of directedly irreducible elements of $\cal D$.
We shall call the preframes in which every element is a directed join of directedly irreducible elements the \emph{directedly generated} preframes. For any directedly generated preframe $\cal D$, the Comparison Lemma yields an equivalence of toposes $\Sh({\cal D}, J_{\cal D})\simeq [DirIrr({\cal D})^{\textrm{op}}, \Set]$, where $J_{\cal D}$ is the directed topology on $\cal D$ (cf. section \ref{charinv} above). Let us now show that the invariant `to be a directed family' satisfies the hypotheses of Theorem \ref{construction} with respect to the class of directedly generated preframes $\cal D$, each of which equipped with the directed topology $J_{\cal D}$. For any directedly generated preframe $\cal D$, the directed topology $J_{\cal D}$ is easily seen to be $C$-induced relative to $P$ where $P$ is the following property of embeddings $i:B_{L}\hookrightarrow L$: `$i$ preserves finite meets' or equivalently (since the canonical embedding ${\cal C}\hookrightarrow Id_{J}({\cal C})$ satisfies the property that ${\cal C}$ is closed under finite meets in $Id_{J}({\cal C})$) the property `$i(B_{L})$ is closed in $L$ under finite meets'. Also, since $Id_{J_{\cal D}}({\cal D})\cong Id(DirIrr({\cal D}))$ via the Comparison Lemma, any family $\cal F$ of principal $J_{\cal D}$-ideals on $\cal D$, $\cal F$ has a directed refinement (if and) only if it has a directed refinement made of principal $J_{\cal D}$-ideals on $\cal D$. Indeed, any principal $J_{\cal D}$-ideal $(d)\downarrow_{J_{\cal D}}$ on $\cal D$ corresponds via the isomorphism $Id_{J_{\cal D}}({\cal D})\cong Id(DirIrr({\cal D}))$ to the set $(d)\downarrow=\{a\in DirIrr({\cal D}) \textrm{ | } a\leq d\}$, and it is immediate to see that if a family $\cal F$ of ideals on $DirIrr({\cal D})$ of the form $(d)\downarrow$ for $d\in {\cal D}$ admits a refinement in $Id(DirIrr({\cal D}))$ by a directed family then $\cal F$ itself must be directed. Thus, if we denote by $\cal K$ the collection of all the directedly generated preframes, Theorem \ref{construction} ensures that the invariant `to be directed' is both $\cal K$-adequate and $\cal K$-compatible; this implies in particular that any directedly generated preframe $\cal D$ is isomorphic to the poset of directedly compact elements (cf. section \ref{charinv} above) of the frame $Id_{J_{\cal D}}({\cal D}) \cong Id(DirIrr({\cal D}))$.
Now, take $\cal K$ to be the collection of directedly generated preframes, each of which equipped with the directed topology, and $\cal H$ to be the collection of the posets of the form $DirIrr({\cal D})$ for some directedly generated preframe, each of which equipped with the trivial Grothendieck topology. The functions $f:{\cal K} \to {\cal H}$ and $g:{\cal H}\to {\cal K}$ defined by setting $f({\cal D})=DirIrr({\cal D})$ (for any atomic distributive lattice $\cal D$) and $g({\cal P})$ equal to the poset $Id_{dir}({\cal P})$ of directedly compact elements of the frame $Id({\cal P})$ are inverse to each other (up to isomorphism), and that the closure of $\cal H$ under isomorphisms can be identified with the collection of all the posets $\cal P$ such that the poset $Id_{dir}({\cal P})$ is closed in $Id({\cal P})$ under finite meets (since for any poset $\cal P$, $Id_{dir}({\cal P})$ is an irreducibly generated distributive lattice whose poset of join-irreducible elements is isomorphic to $\cal P$).
The category $Ext({\cal K})$ is the category $\textbf{DirIrrPFrm}$ whose objects are the directedly generated preframes and whose arrows ${\cal D}\to {\cal D}'$ are the preframe homomorphisms $f:{\cal D}\to {\cal D}'$ between them such that the frame homomorphism $A(f):Id_{J_{\cal D}}({\cal D})\to Id_{J_{{\cal D}'}}({\cal D}')$ which sends an ideal $I$ of $\cal D$ to the ideal on ${\cal D}'$ generated by $f(I)$ preserves arbitrary infima.
The category $Ext({\cal H})$ is the category $\textbf{Pos}_{dir}$ whose objects are the posets $\cal P$ such that $Id_{dir}({\cal P})$ is closed in $Id({\cal P})$ under finite meets and whose arrows ${\cal P}\to {\cal P}'$ are the monotone maps $g:{\cal P}\to {\cal P}'$ between them such that the inverse image $g^{-1}:Id({\cal P}') \to Id({\cal P})$ sends ideals in $Id_{dir}({\cal P}')$ to ideals in $Id_{dir}({\cal P})$.
We can describe the category $\textbf{Pos}_{dir}$ in more concrete terms, as follows. First, we note that for any poset $\cal P$, an ideal $I$ in $Id({\cal P})$ is directedly compact in $Id({\cal P})$ if and only if it is a directed union of principal ideals on $\cal P$, equivalently if $I$ is directed, i.e. it is non-empty and for any $a,b\in I$ there exists $c\in I$ such that $a\leq c$ and $b\leq c$. Indeed, the `if' direction follows from Proposition \ref{multicomposition}, while the `only if' direction follows from the fact that if $I$ is directedly compact then the covering of $I$ formed by the principal ideals on $\cal P$ generated by an element of $I$ has a directed refinement by principal ideals on $\cal P$ (since the invariant `to be directed' satisfies the condition in the statement of Theorem \ref{construction}, cf. above) and hence $I$ itself must be directed. Now, all the principal ideals on $\cal P$ are directedly compact elements of $Id({\cal P})$, and hence from the characterization above it follows that the condition that the intersection of two directedly compact elements of $Id({\cal P})$ should be directedly compact is equivalent to the requirement that the intersection of any two principal ideals on $\cal P$ should be a directed ideal, equivalently that for any $a,b\in {\cal P}$ there should be $c\in {\cal P}$ such that $c\leq a$ and $c\leq b$, and for any elements $d,e\in {\cal P}$ such that $d,e\leq a$ and $d,e\leq b$ there should exist $z\in {\cal P}$ such that $z\leq a$, $z\leq b$, $d,e \leq z$. Note that if $\cal P$ is a meet-semilattice than this condition is always satisfied (the intersection of $(a)\downarrow$ and $(b)\downarrow$ is equal to $(a\wedge b)\downarrow$).
Concerning the arrows in $\textbf{Pos}_{dir}$ we note that, given a monotone map $g:{\cal P}\to {\cal P}'$, the inverse image $g^{-1}:Id({\cal P}') \to Id({\cal P})$ sends directedly compact elements to directedly compact elements if and only if for any $b\in {\cal P}'$, $g^{-1}((b)\downarrow)$ is a directed ideal on $\cal P$, i.e. there exists $a\in {\cal P}$ such that $g(a)\leq b$ and for any $u,v\in {\cal P}$ such that $g(u)\leq b$ and $g(v)\leq b$ there exists $z\in {\cal P}$ such that $u,v\leq z$ and $g(z)\leq b$.
Summarizing, the category $\textbf{Pos}_{dir}$ has as objects the posets $\cal P$ such that for any $a,b\in {\cal P}$ there is $c\in {\cal P}$ such that $c\leq a$ and $c\leq b$ and for any elements $d,e\in {\cal P}$ such that $d,e\leq a$ and $d,e\leq b$ there exists $z\in {\cal P}$ such that $z\leq a$, $z\leq b$, $d,e \leq z$, and as arrows ${\cal P}\to {\cal P}'$ the monotone maps $g:{\cal P}\to {\cal P}'$ with the property that for any $b\in {\cal P}'$ there exists $a\in {\cal P}$ such that $g(a)\leq p$ and for any two $u,v\in {\cal P}$ such that $g(u)\leq b$ and $g(v)\leq b$ there exists $z\in {\cal P}$ such that $u,v\leq z$ and $g(z)\leq b$.
Our Theorem \ref{dualabstract} thus yields two functors $D:\textbf{DirIrrPFrm}^{\textrm{op}}\to \textbf{Pos}_{dir}$ and $E:\textbf{Pos}_{dir} \to \textbf{DirIrrPFrm}^{\textrm{op}}$ which are categorical inverses to each other and hence form a duality between $\textbf{DirIrrPFrm}$ and $\textbf{Pos}_{dis}$. The functor $D:\textbf{DirIrrPFrm}^{\textrm{op}}\to \textbf{Pos}_{dir}$ sends a preframe ${\cal D}$ in $\textbf{DirIrrPFrm}$ to the poset $DirIrr({\cal D})$ of directedly irreducible elements of ${\cal D}$ and an arrow $f:{\cal D}\to {\cal D}'$ in $\textbf{DirIrrPFrm}$ to the restriction $DirIrr({\cal D}) \to DirIrr({\cal D})$ of the left adjoint to $A(f):Id_{J_{{\cal D}}}({\cal D})\to Id_{J_{{\cal D}'}}({\cal D}')$ to the sets of directedly irreducible elements of ${\cal D}'$ and of ${\cal D}$. The functor $E:\textbf{Pos}_{dir} \to \textbf{DirIrrPFrm}^{\textrm{op}}$ sends a poset $\cal P$ in $\textbf{Pos}_{dir}$ to the poset $Id_{dir}({\cal P})$ of directed ideals on $\cal P$, and an arrow $g:{\cal P}\to {\cal P}'$ in $\textbf{Pos}_{dir}$ to the restriction $Id_{dir}({\cal P}')\to Id_{dir}({\cal P})$ of the inverse image $g^{-1}:Id({\cal P}')\to Id({\cal P})$.
In conclusion, we have the following result.
\begin{theorem}\label{dirirr}
Via the functors
\[
D:\textbf{DirIrrPFrm}^{\textrm{op}}\to \textbf{Pos}_{dir}
\]
and
\[
E:\textbf{Pos}_{dir} \to \textbf{DirIrrPFrm}^{\textrm{op}}
\]
defined above, the categories $\textbf{DirIrrPFrm}$ and $\textbf{Pos}_{dir}$ are dual to each other.
\end{theorem}\qed
Given a poset $\cal P$ in $\textbf{Pos}_{dir}$, if $\cal P$ has binary joins and a bottom element $0_{\cal P}$ then the condition that an ideal $I$ on $\cal P$ be directed can be reformulated as the requirement that $I$ be non-empty and that for any $a,b\in I$, $a\vee b\in I$. Moreover, it is easy to verify that for any monotone map $g:{\cal P}\to {\cal P}'$ between posets in $\textbf{Pos}_{dir}$ having a bottom element and binary joins, $g(0_{\cal P})=0_{{\cal P}'}$ and $g$ sends binary joins in $\cal P$ to binary joins in ${\cal P}'$ if and only if $g$ is an arrow ${\cal P}\to {\cal P}'$ in the category $\textbf{Pos}_{dir}$. Therefore, the category $\textbf{SSLat}$ of sup-semilattices and sup-semilattice homomorphisms between them can be identified with a full subcategory of $\textbf{Pos}_{dir}$, and hence the duality of Theorem \ref{dirirr} restricts to a duality between the full subcategory $\textbf{DirIrrPFrm}_{s}$ of $\textbf{DirIrrPFrm}$ on the preframes $\cal D$ such that $DirIrr{\cal D}$ is a sup-semilattice and the category $\textbf{SSLat}$. That is, we have the following result.
\begin{theorem}
Via the restrictions
\[
D|:\textbf{DirIrrPFrm}_{s}^{\textrm{op}}\to \textbf{SSLat}
\]
and
\[
E|:\textbf{SSLat} \to \textbf{DirIrrPFrm}_{s}^{\textrm{op}}
\]
of the functors
\[
D:\textbf{DirIrrPFrm}^{\textrm{op}}\to \textbf{Pos}_{dir}
\]
and
\[
E:\textbf{Pos}_{dir} \to \textbf{DirIrrPFrm}^{\textrm{op}}
\]
defined above, the categories $\textbf{DirIrrPFrm}_{s}$ and $\textbf{SSLat}$ are dual to each other.
\end{theorem}
\begin{remark}
The restriction $E|:\textbf{SSLat}\to \textbf{Pos}^{\textrm{op}}$ of the functor $E$ coincides with the functor from $\textbf{SSLat}$ to $\textbf{Pos}^{\textrm{op}}$ giving one half of the duality of Theorem 1.5 \cite{duality} between algebraic lattices and sup-semilattices; from this it follows that the category $\textbf{DirIrrPFrm}_{s}$ coincides with the category of algebraic lattices defined in \cite{duality} (since both categories coincide, by Theorem \ref{dirirr} and the duality theorem of \cite{duality}, with the extended image of the same functor, namely $E|:\textbf{SSLat}\to \textbf{Pos}^{\textrm{op}}$).
\end{remark}
As a final example, consider frames in which every element is a join of compact elements; by the Comparison Lemma, they can be represented as frames of ideals (i.e. lower subsets which are closed under finitary joins) on the join-semilattice of their compact elements.
Of course, these examples do not exhaust at all the possibilities of application of our techniques; they are just meant to show that a great variety of insights on preordered structures and topological spaces can be easily obtained by applying our methods. The reader is invited to build his or her favorite dualities or representation theorems applying our techniques to his or her cases of interest.
\section{Adjunctions}\label{adj}
Our approach to Stone-type dualities described so far provides us with a natural way of building adjunctions between categories of preorders and categories of locales or topological spaces which restrict, on appropriate subcategories, to the categorical equivalences established in the previous sections. Again, this represents an application of the philosophy `toposes as bridges' of \cite{OC10} in the context of the Morita-equivalence given by Theorem \ref{fund}.
\subsection{Frames presented by sites}
Starting from the equivalence
\[
\Sh({\cal C}, J)\simeq \Sh(Id_{J}({\cal C}))
\]
of Theorem \ref{fund}, we can investigate the behaviour of a particular class of invariants in relation to such equivalence, namely geometric morphisms from localic toposes to the given topos.
First, let us analyze the behavior of this invariant with respect to the first site of definition $({\cal C}, J)$ of the topos. For any locale $L$, the (isomorphism classes of) geometric morphisms $\Sh(L)\to \Sh({\cal C}, J)$ correspond exactly, by Diaconescu's equivalence, to the flat $J$-continuous functors from $\cal C$ to $\Sh(L)$; note that these latter functors always take values in the frame $L$ of subterminals in $\Sh(L)$ (cf. section \ref{logical}).
Next, let us describe the behavior of the invariant with respect to the second site of definition for the topos. For any locale $L$, the geometric morphisms $\Sh(L)\to \Sh(Id_{J}({\cal C}))$ correspond to the frame homomorphisms $Id_{J}({\cal C}) \to L$ (cf. Proposition C1.4.5 \cite{El}).
Therefore, we can conclude that the frame homomorphisms $Id_{J}({\cal C}) \to L$ correspond bijectively to the flat $J$-continuous functors ${\cal C}\to \Sh(L)$, under the bijection sending a frame homomorphism $Id_{J}({\cal C})\to L$ to the composite of it with the canonical map $l:{\cal C}\to Id_{J}({\cal C})$. The maps $f:{\cal C}\to L$ such that the composite $y\circ f:{\cal C}\to \Sh(L)$ of $f$ with the Yoneda embedding $y:L\to \Sh(L)$ is a flat functor ${\cal C}\to \Sh(L)$, which we call the \emph{filtering} maps, can be characterized explicitly.
\begin{proposition}\label{filtering}
Let $\cal C$ be a preorder, $L$ be a frame and $f:{\cal C}\to L$ be a monotone map. Then $f$ is filtering if and only if
\begin{enumerate}[(i)]
\item $1_{L}=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{c\in {\cal C}}}f(c)$;
\item For any $c,c'\in {\cal C}$, $f(c)\wedge f(c')=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{b\in B_{c, c'}}}f(b)$ where $B_{c, c'}$ is the set
\[
\{b\in {\cal C} \textrm{ | } b\leq c \textrm{ and } b\leq c'\}.
\]
\end{enumerate}
In particular, if $\cal C$ is a meet-semilattice, $f:{\cal C}\to L$ is filtering if and only if it is a meet-semilattice homomorphism ${\cal C}\to L$.
\end{proposition}
\begin{proofs}
The proposition immediately follows from the characterization of flat functors as filtering functors given by Theorem VII 10.1 \cite{MM}.
\end{proofs}
Note that a filtering map $f:{\cal C}\to L$ corresponds to a \emph{$J$-continuous} flat functor ${\cal C}\to \Sh(L)$ as above if and only if it sends $J$-covering sieves to covering families in $L$.
The argument above, combined with Proposition \ref{filtering}, thus yields the following result.
\begin{theorem}\label{freeframes}
Let $\cal C$ be a preorder and $J$ be a Grothendieck topology on $\cal C$. Then the frame $Id_{J}({\cal C})$, together with the map $\eta:{\cal C}\to Id_{J}({\cal C})$ sending an element $c\in {\cal C}$ to the principal ideal $(c)\downarrow_{J}$, satisfies the following universal property: for any map $f:{\cal C}\to L$ to a frame $L$, $f$ is filtering and sends every $J$-covering sieve to a covering family in $L$ if and only if there is a (unique) frame homomorphism $\tilde{f}:Id_{J}({\cal C})\to L$ such that $\tilde{f}\circ \eta=f$ (given by the formula $\tilde{f}(I)=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{c\in I}}f(c)$ for any $I\in Id_{J}({\cal C})$).
\end{theorem}\qed
\begin{remark}
The particular case of Theorem \ref{freeframes} when $\cal C$ is a meet-semilat-\\tice appears as Proposition II.2.11 of \cite{stone}; the reader might find it interesting to compare the simple topos-theoretic argument which proves Theorem \ref{freeframes} with the proof of the result for meet-semilattices in \cite{stone}, which consists of direct but elaborate technical arguments involving frames (cf. also section \ref{appendix} for an elementary proof of Theorem \ref{freeframes}).
\end{remark}
Provided that some natural hypotheses are satisfied, we can read the universal property of $Id_{J}({\cal C})$ stated in Theorem \ref{freeframes} as an adjunction between the category \textbf{Frm} of frames and a category having as objects the preorders $\cal C$. To this end, suppose to have, as in section \ref{general} above, a category $\cal K$ of posets $\cal C$, each of which equipped with a subcanonical Grothendieck topology $J_{{\cal C}}$. If $\cal K$ contains, among its objects, all the frames, and for any structure $\cal C$ in $\cal K$ and any frame $L$ in $\cal K$ the arrows ${\cal C}\to L$ in $\cal K$ coincide with the filtering maps ${\cal C}\to L$ which send $J_{\cal C}$-covering sieves to a covering families in $L$, then we have a forgetful functor $U_{\cal K}:\textbf{Frm}\to {\cal K}$, and the following result holds.
\begin{theorem}
Let $\cal K$ satisfy the hypotheses above. Then the functor $A:\cal K \to \textbf{Frm}$ sending an object $\cal C$ of $\cal K$ to $Id_{J}({\cal C})$ and an arrow $f:{\cal C}\to {\cal D}$ in $\cal K$ to the frame homomorphism $A(f):Id_{J_{\cal C}}({\cal C})\to Id_{J_{\cal D}}({\cal D})$ sending any ideal $I$ in $Id_{J_{\cal C}}({\cal C})$ to the $J_{\cal D}$-ideal on $\cal D$ generated by the image $f(I)$ of $I$ under $f$ is left adjoint to the forgetful functor $U_{\cal K}:\textbf{Frm}\to {\cal K}$.
\end{theorem}
\begin{proofs}
Under the hypotheses of the proposition, the arrows $f:{\cal C} \to U_{\cal K}(L)$ in $\cal K$ can be identified with the filtering maps ${\cal C}\to L$ which send $J_{\cal C}$-covering sieves to covering families in $L$, and these maps in turn correspond to the frame homomorphisms $Id_{J}({\cal C}) \to L$ (by Theorem \ref{freeframes}). It is easy to check that this bijection is natural in ${\cal C}\in {\cal K}$ and in $L\in \textbf{Frm}$ and hence yields an adjunction between $A$ and $U_{\cal K}$ in which $U_{\cal K}$ is the left adjoint. The naturality in $L$ is obvious, while the naturality in $\cal C$ follows from the fact that any arrow $f$ in $\cal K$ is naturally isomorphic to the restriction of $A(f)$ to the subsets of principal ideals on $\cal C$ and on ${\cal C}'$.
\end{proofs}
As particular cases of the theorem we recover:
\begin{enumerate}[(i)]
\item The reflection from the category of meet-semilattices to the category of frames (take $\cal K$ to be the category of meet-semilattices, each of which equipped with the trivial Grothendieck topology);
\item The reflection from the category of distributive lattices to the category of frames (take $\cal K$ to be the category of distributive lattices, each of which equipped with the coherent topology);
\item The reflection from the category of preframes to the category of frames (take $\cal K$ to be the category of preframes, each of which equipped with the directed topology, cf. section \ref{charinv}).
\end{enumerate}
As a novel application of the theorem we obtain a reflection from the category of disjunctively distributive frames to the category of frames (take $\cal K$ to be the category of disjunctively distributive frames, each of which equipped with the infinitary disjunctive topology). Also, for any regular cardinal $k$, we obtain a reflection from the category of $k$-frames (cf. section \ref{charinv} above) to the category of frames.
As another application of our usual method (of transferring invariants, in this case geometric morphisms involving localic toposes, across different sites of definition of a given topos), we establish an adjunction between the opposite of the category $\textbf{Bool}$ of Boolean algebras and the category $\textbf{Loc}$ of locales, which restricts to the equivalence between $\textbf{Bool}^{\textrm{op}}$ and the full subcategory of the category of coherent locales on the locales which have a basis of complemented elements (cf. section \ref{stonedist} above).
Our starting point is, as above, the Morita-equivalence
\[
\Sh({\cal C}, J)\simeq \Sh(Id_{J}({\cal C}))
\]
of Theorem \ref{fund}, where we suppose $\cal C$ to be a Boolean algebra (regarded as a preorder coherent category) and $J$ to be the coherent topology on it.
For any locale $L$, the (isomorphism classes of) geometric morphisms
\[
\Sh(L)\to \Sh(Id_{J}({\cal C}))
\]
correspond exactly to the frame homomorphisms $Id_{J}({\cal C})\to L$, while the (isomorphism classes of) geometric morphisms
\[
\Sh(L)\to \Sh({\cal C}, J)
\]
correspond precisely to the meet-semilattice homomorphisms ${\cal C}\to L$ which send $J$-covering sieves to covering families, in other words to the lattice homomorphisms ${\cal C}\to L$.
Now, since $\cal C$ is a Boolean algebra, every lattice homomorphism ${\cal C}\to L$ takes values in the sublattice $L_{c}$ of $L$ consisting of the complemented elements of $L$ and hence the arrows $L \to Id_{J}({\cal C})$ in \textbf{Loc} are in bijection with the Boolean algebra homomorphisms ${\cal C}\to L_{c}$; it is immediate to verify that this bijection is natural in ${\cal C}\in \textbf{Bool}$ and in $L\in \textbf{Loc}$. Thus, define $c:\textbf{Loc}\to \textbf{Bool}^{\textrm{op}}$ as the functor sending a locale $L$ to the lattice $L_{c}$ of complemented elements of $L$ and a frame homomorphism $L\to L'$ to its restriction $L_{c}\to L'_{c}$, and $Id:\textbf{Bool}^{\textrm{op}}\to \textbf{Loc}$ as the functor sending a Boolean algebra $\cal C$ to the frame $Id_{J_{\cal C}}({\cal C})$ of ideals of $\cal C$ (where $J_{\cal C}$ is the coherent topology on $\cal C$) and a morphism $f:{\cal C}\to {\cal C}'$ in $\textbf{Bool}$ to the frame homomorphism $Id_{J_{\cal C}}({\cal C})\to Id_{J_{\cal C}'}({\cal C}')$ sending an ideal $I$ in $Id_{J_{\cal C}}({\cal C})$ to the ideal in $Id_{J_{\cal C}'}({\cal C}')$ generated by the image $f(I)$ of $I$ under $f$.
We have thus established the following result.
\begin{theorem}
The functors $Id:\textbf{Bool}^{\textrm{op}}\to \textbf{Loc}$ and $c:\textbf{Loc}\to \textbf{Bool}^{\textrm{op}}$ defined above are adjoint to each other, where $c$ is the left adjoint and $Id$ is the right adjoint.
\end{theorem}\qed
\begin{remark}
Composing this adjunction with the well-known adjunction between locales and spaces yields the usual Stone adjunction between the opposite of the category of Boolean algebras and the category of topological spaces.
\end{remark}
\subsection{Disjunctive and atomic frames}
Below, we shall adopt the terminology of section \ref{addex}.
Given a disjunctively distributive frame $F$, and denoted by ${\cal I}_{F}$ the collection of its indecomposable elements, we have a map $\phi_{F}: F\to Id({\cal I}_{F})$ sending an element $a\in F$ to the set $\{b\in {\cal I}_{F} \textrm{ | } b\leq a\}$. Notice that $\phi_{F}$ takes values in the subset of infinitarily disjunctively compact elements of $Id({\cal I}_{F})$; indeed, for any $a \in F$, $\phi_{F}(a)$ is supercompact and hence in particular infinitarily disjunctively compact.
\begin{theorem}\label{disthm}
\begin{enumerate}[(i)]
\item For any disjunctively distributive frame $F$ such that the subset $Dis(Id({\cal I}_{F}))$ of infinitarily disjunctively compact elements of the frame $Id({\cal I}_{F})$ is closed in $Id({\cal I}_{F})$ under finite meets, $Dis(Id({\cal I}_{F}))$ is a disjunctive frame with the induced order, and the map $\phi_{F}:F\to Dis(Id({\cal I}_{F}))$ is a disjunctively distributive frame homomorphism which restricts to an isomorphism from ${\cal I}_{F} \subseteq F$ to the subset ${\cal I}_{Dis(Id({\cal I}_{F}))}$ of indecomposable elements of $Dis(Id({\cal I}_{F}))$;
\item A disjunctively distributive frame $F$ is disjunctive if and only if the subset $Dis(Id({\cal I}_{F}))$ of infinitarily disjunctively compact elements of the frame $Id({\cal I}_{F})$ is closed in $Id({\cal I}_{F})$ under finite meets and the map\\ $\phi_{F}:F\to Dis(Id({\cal I}_{F}))$ is an isomorphism (of disjunctively distributive frames);
\item A topological space $X$ is locally connected if and only if the frame ${\cal O}(X)$ of open sets of $X$ is isomorphic via the map $\phi_{{\cal O}(X)}$ to the poset $Dis(Id({\cal I}_{{\cal O}(X)}))$ of infinitarily disjunctively compact elements of the\\ frame of ideals on the poset of non-empty connected open subsets of $X$ and the (set-theoretic) intersection of any two infinitarily disjunctively compact elements of $Id({\cal I}_{{\cal O}(X)})$ is infinitarily disjunctively compact.
\end{enumerate}
\end{theorem}
\begin{proofs}
$(i)$ By Proposition \ref{multicomposition}, $Dis(Id({\cal I}_{F}))$ is closed in $Id({\cal I}_{F})$ under pairwise disjoint joins and, by our hypothesis, it is closed in $Id({\cal I}_{F})$ under finite meets. Therefore $Dis(Id({\cal I}_{F}))$ is, with the order induced by that of $Id({\cal I}_{F})$, a disjunctively distributive frame. The fact that $\phi_{F}:F\to Dis(Id({\cal I}_{F}))$ is a disjunctively distributive frame homomorphism is immediate to verify. Now, let us prove that $Dis(Id({\cal I}_{F}))$ is disjunctive. If $I$ is an infinitarily disjunctively compact element of the frame $Id({\cal I}_{F})$ then the covering of $I$ in $Id({\cal I}_{F})$ formed by the principal ideals generated by the elements of $I$ has a refinement consisting of pairwise disjoint elements; clearly, the elements of this refining family must all be principal ideals, from which it follows, these ideals being supercompact, and hence indecomposable, elements of $Id({\cal I}_{F})$, that $I$ can be expressed as a pairwise disjoint join in $Id({\cal I}_{F})$ of indecomposable elements of $Id({\cal I}_{F})$. Now, by Proposition \ref{multicomposition}, $Dis(Id({\cal I}_{F}))$ is closed in $Id({\cal I}_{F})$ under pairwise disjoint joins, and from the fact that $Dis(Id({\cal I}_{F}))$ is closed in $Id({\cal I}_{F})$ under finite meets it follows that two elements of $Dis(Id({\cal I}_{F}))$ are pairwise disjoint in $Dis(Id({\cal I}_{F}))$ if and only if they are pairwise disjoint in $Id({\cal I}_{F})$; therefore $I$ can be expressed as a pairwise disjoint join in $Dis(Id({\cal I}_{F}))$ of indecomposable elements of $Dis(Id({\cal I}_{F}))$. This proves that $Dis(Id({\cal I}_{F}))$ is a disjunctive frame; moreover, the argument shows that the indecomposable elements of $Dis(Id({\cal I}_{F}))$ are precisely those of the form $\phi_{F}(a)$ for $a\in Id({\cal I}_{F})$.
$(ii)$ We have already proved the `only if' direction (cf. the discussion above), so it remains to prove the `if' one. Let us suppose that $\phi_{F}:F\to Dis(Id({\cal I}_{F}))$ is an isomorphism of posets. By part $(i)$ $Dis(Id({\cal I}_{F}))$ is disjunctive and hence $F$, being isomorphic to it, is disjunctive as well, as required.
$(iii)$ This follows immediately from part $(ii)$ by recalling that a topological space is (classically) locally connected if and only if every open set is a disjoint union of non-empty connected open sets.
\end{proofs}
We now proceed to show that the construction above of a disjunctive frame consisting of the infinitarily disjunctively compact elements of a frame $Id({\cal I}_{F})$ can be naturally made into an adjunction.
Given a disjunctively distributive frames $F$ such that the subset of infinitarily disjunctively compact elements of $Id({\cal I}_{F})$ is closed in $Id({\cal I}_{F})$ under finite meets, the composite $y \circ \phi_{F}:F\to [{{\cal I}_{F}}^{\textrm{op}}, \Set]$ of the map $\phi_{F}:F\to Id({\cal I}_{F})$ with the Yoneda embedding $y:Id({\cal I}_{F}) \to \Sh(Id({\cal I}_{F}))\simeq [{{\cal I}_{F}}^{\textrm{op}}, \Set]$ preserves finite meets and sends $J_{F}$-covering sieves to covering families. Indeed, given a $J_{F}$-covering sieve $\{a_{i} \leq a \textrm{ | } \in I\}$ on an object $a$ of $F$, $\phi_{F}(\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i})=\mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{i\in I}}\phi_{F}(a_{i})$ since for any $b\in \phi_{F}(a)$, $b=b\wedge a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}(b\wedge a_{i})$ implies, $b$ being indecomposable, $b=b\wedge a_{i}$ for some $i\in I$, equivalently $b\in \phi_{F}(a_{i})$.
So, by Diaconescu's equivalence, the flat $J_{F}$-continuous functor
\[
y \circ \phi_{F}:F\to [{{\cal I}_{F}}^{\textrm{op}}, \Set]
\]
corresponds to a geometric morphism
\[
\chi_{F}:[{{\cal I}_{F}}^{\textrm{op}}, \Set] \to \Sh(F, J_{F}).
\]
Let $\textbf{DJFrm}_{dis}$ be the category whose objects are the disjunctively distributive frames $F$ such that the subset of infinitarily disjunctively compact elements of $Id({\cal I}_{F})$ is closed in $Id({\cal I}_{F})$ under finite meets, and whose arrows $F\to F'$ are the disjunctively distributive frames homomorphisms $f:F\to F'$ such that there exists a monotone map $u:{\cal I}_{F'}\to {\cal I}_{F}$ such that the diagram
\[
\xymatrix {
[{{\cal I}_{F'}}^{\textrm{op}}, \Set] \ar[r]^{E_{u}} \ar[d]^{\chi_{F'}} & [{{\cal I}_{F'}}^{\textrm{op}}, \Set] \ar[d]^{\chi_{F}} \\
\Sh(F', J_{F'}) \ar[r]^{\dot{f}} & \Sh(F, J_{F})}
\]
commutes (up to isomorphism), where $E_{u}:[{{\cal I}_{F'}}^{\textrm{op}}, \Set] \to [{{\cal I}_{F}}^{\textrm{op}}, \Set]$ is the geometric morphism induced by $u^{\textrm{op}}:{{\cal I}_{F'}}^{\textrm{op}}\to {{\cal I}_{F}}^{\textrm{op}}$ as in Example A4.1.4 \cite{El}, and the inverse image $u^{-1}:Id({\cal I}_{F})\to Id({\cal I}_{F'})$ sends the infinitarily disjunctively compact elements of $Id({\cal I}_{F})$ to infinitarily disjunctively compact elements of $Id({\cal I}_{F'})$.
Given an arrow $f:F\to F'$ in $\textbf{DJFrm}_{dis}$ there is exactly one monotone function $u:{\cal I}_{F'}\to {\cal I}_{F}$ such that the diagram
\[
\xymatrix {
[{{\cal I}_{F'}}^{\textrm{op}}, \Set] \ar[r]^{E_{u}} \ar[d]^{\chi_{F'}} & [{{\cal I}_{F}}^{\textrm{op}}, \Set] \ar[d]^{\chi_{F}} \\
\Sh(F', J_{F'}) \ar[r]^{\dot{f}} & \Sh(F, J_{F})}
\]
commutes (up to isomorphism). Indeed, there is at most one such map by definition of the category $\textbf{DJFrm}_{dis}$, while the uniqueness can be proved as follows. The commutativity of the square above (up to isomorphism) is equivalent to the (strict) commutativity of the diagram
\[
\xymatrix {
F \ar[r]^{f} \ar[d]^{\phi_{F}} & F' \ar[d]^{\phi_{F'}} \\
Id({\cal I}_{F}) \ar[r]^{u^{-1}} & Id({\cal I}_{F'})},
\]
which, in light of the fact that $u^{-1}$ preserves unions of ideals, forces $u^{-1}:Id({\cal I}_{F}) \to Id({\cal I}_{F'})$ to be equal to the map $\xi_{f}:Id({\cal I}_{F}) \to Id({\cal I}_{F'})$ sending an ideal $I$ in $Id({\cal I}_{F})$ to the union $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{p\in I}}\phi_{F'}(f(p))$ in $Id({\cal I}_{F'})$. Our claim then follows from the fact that $u$ is uniquely determined by $u^{-1}$ (cf. Theorem \ref{proalex}).
Note that the map $\xi_{f}:Id({\cal I}_{F}) \to Id({\cal I}_{F'})$ is a frame homomorphism, which implies that given a disjunctively distributive frame homomorphism $f:F\to F'$, there exists a monotone map $u:{\cal I}_{F'}\to {\cal I}_{F}$ such that the square above commutes, equivalently $u^{-1}=\xi_{f}:Id({\cal I}_{F}) \to Id({\cal I}_{F'})$, if and only if $\xi_{f}$ is complete, i.e. it preserves arbitrary infima (cf. section \ref{prealex} above).
Therefore the category $\textbf{DJFrm}_{dis}$ can be alternatively described as the category whose objects are the disjunctively distributive frames $F$ such that the subset of infinitarily disjunctively compact elements of $Id({\cal I}_{F})$ is closed in $Id({\cal I}_{F})$ under finite meets, and whose arrows $F\to F'$ are the disjunctively distributive frames homomorphisms $f:F\to F'$ such that the map $\xi_{f}:Id({\cal I}_{F}) \to Id({\cal I}_{F'})$ is complete and sends infinitarily disjunctively compact elements to infinitarily disjunctively compact elements.
Let us consider the category $\textbf{DisFrm}$ of disjunctive frames defined in section \ref{addex}; recall that the objects of $\textbf{DisFrm}$ are the disjunctive frames and the arrows $F\to F'$ of $\textbf{DisFrm}$ are the meet-semilattice homomorphisms $f:F\to F'$ between them which send pairwise disjoint joins to pairwise disjoint joins and have the property that the frame homomorphism $A(f):Id_{J_{F}}(F)\to Id_{J_{F'}}(F')$ sending an ideal $I$ of $F$ to the ideal of $F'$ generated by $f(I)$ preserves arbitrary infima.
We can define two functors
\[
i:\textbf{DisFrm}\to \textbf{DJFrm}_{dis}
\]
and
\[
L:\textbf{DJFrm}_{dis} \to \textbf{DisFrm}
\]
as follows; $i$ is the inclusion functor of $\textbf{DisFrm}$ into $\textbf{DJFrm}_{dis}$ while $L:\textbf{DJFrm}_{dis} \to \textbf{DisFrm}$ is the functor sending a poset $F$ in $\textbf{DJFrm}_{dis}$ to the poset $Dis(Id({\cal I}_{F}))$ of infinitarily disjunctively compact elements of the frame $Id({\cal I}_{F})$ and an arrow $f:F\to F'$ in $\textbf{DJFrm}_{dis}$ to the restriction of the map $\xi_{f}:Id({\cal I}_{F})\to Id({\cal I}_{F'})$ to the subsets $Dis(Id({\cal I}_{F}))$ and $Dis(Id({\cal I}_{F'}))$ of infinitarily disjunctively compact elements of $Id({\cal I}_{F})$ and of $Id({\cal I}_{F'})$ (note that the functors $i$ and $L$ are well-defined by Theorem \ref{disthm}).
\begin{theorem}\label{disadj}
The embedding
\[
i:\textbf{DisFrm}\hookrightarrow \textbf{DJFrm}_{dis}
\]
identifies $\textbf{DisFrm}$ a full reflective subcategory of the category $\textbf{DJFrm}_{dis}$, with reflector $L:\textbf{DJFrm}_{dis} \to \textbf{DisFrm}$.
\end{theorem}
\begin{proofs}
The fact that $L\circ i:\textbf{DisFrm} \to \textbf{DisFrm}$ is naturally isomorphic to the identity functor follows immediately from Theorem \ref{disthm} and the results of section \ref{Mordual}.
To prove that $L$ is left adjoint to $i$, we observe that we have the following natural correspondences, natural in $F\in \textbf{DJFrm}_{dis}$ and in $G\in \textbf{DisFrm}$, between the arrows $L(F)\to G$ in $\textbf{DisFrm}$ and the arrows $F\to i(G)$ in $\textbf{DJFrm}_{dis}$; an arrow $\alpha:L(F)\to G$ in $\textbf{DisFrm}$ corresponds to the arrow $\alpha\circ \phi_{F}:F\to i(G)$, while an arrow $\beta:F\to i(G)$ corresponds to the arrow $\phi_{G}^{-1}\circ L(\beta)$ given by the composite of $\phi_{G}^{-1}:Dis(Id({\cal I}_{G}))\to G$ with the arrow $L(\beta):L(F)\to L(i(G))$.
Note that the unit of the adjunction at an object $F$ of $\textbf{DJFrm}_{dis}$ is given by the map $\phi_{F}:F\to Dis(Id({\cal I}_{F}))$, while the counit at an object $G$ of $\textbf{DisFrm}$ is given by the isomorphism $\phi_{G}^{-1}:Dis(Id({\cal I}_{G}))\to G$.
\end{proofs}
Similarly, we can establish an adjunction based on the Lindenbaum-Tarski representation of atomic frames. The analogue of Theorem \ref{disthm} for atomic frames reads as follows.
\begin{theorem}\label{atthm}
\begin{enumerate}[(i)]
\item For any frame $F$, ${\mathscr{P}}(At(F))$ is an atomic frame (with the subset inclusion order), and the map $\psi_{F}:F\to {\mathscr{P}}(At(F))$ sending an element $p\in F$ to the subset $\{a\in At(F) \textrm{ | } a\leq p\}$ is a frame homomorphism which restricts to an isomorphism from $At(F) \subseteq F$ to the subset $At({\mathscr{P}}(At(F)))$ of atoms of ${\mathscr{P}}(At(F))$;
\item A frame $F$ is atomic if and only if the map $\psi_{F}:F\to {\mathscr{P}}(At(F))$ is a frame isomorphism;
\item A topological space $X$ is discrete if and only if it is sober and the frame ${\cal O}(X)$ of open sets of $X$ is isomorphic via the map $\psi_{{\cal O}(X)}$ to the powerset ${\mathscr{P}}(At({\cal O}(X)))$ of the set $At({\cal O}(X))$ of atoms of the frame ${\cal O}(X)$ of open sets of $X$.
\end{enumerate}
\end{theorem}
\begin{proofs}
$(i)$ It is clear that ${\mathscr{P}}(At(F))$ is an atomic frame, and it is immediate to verify that $\psi_{F}:F\to {\mathscr{P}}(At(F))$ is a frame homomorphism which restricts to an isomorphism from $At(F) \subseteq F$ to the subset $At({\mathscr{P}}(At(F)))$ of atoms of ${\mathscr{P}}(At(F))$.
$(ii)$ The `only if' direction follows from the Comparison Lemma (cf. section \ref{tarski}), so it remains to prove the `if' one. Let us suppose that $\psi_{F}:F\to {\mathscr{P}}(At(F))$ is an isomorphism. By part $(i)$ ${\mathscr{P}}(At(F))$ is atomic and hence $F$, being isomorphic to it, is atomic as well, as required.
$(iii)$ This follows immediately from part $(ii)$ by recalling that a topological space is discrete if and only if it is sober and has a basis of atomic open subsets (cf. Proposition \ref{discrete}).
\end{proofs}
Similarly to the case of disjunctively distributive frames, for any frame $F$ we have a geometric morphism $\chi_{F}:[At(F), \Set] \to \Sh(F, J_{F})$ (where $J_{F}$ is the canonical topology on $F$) corresponding to the map $\psi_{F}:F\to {\mathscr{P}}(At(F))$.
Let us define $\textbf{Frm}_{at}$ as the category having as objects the frames and as arrows $F\to F'$ the frame homomorphisms $f:F\to F'$ such that there exists a monotone map $u:At(F')\to At(F)$ making the diagram
\[
\xymatrix {
[At(F'), \Set] \ar[r]^{E_{u}} \ar[d]^{\chi_{F'}} & [At(F), \Set] \ar[d]^{\chi_{F}} \\
\Sh(F', J_{F'}) \ar[r]^{\dot{f}} & \Sh(F, J_{F})}
\]
commute (up to isomorphism), where $E_{u}:[At(F'), \Set] \to [At(F), \Set]$ the geometric morphism induced by the map $u: At(F')\to At(F)$ as in Example A4.1.4 \cite{El}.
One can easily see that the category $\textbf{Frm}_{at}$ can be alternatively described as the category whose objects are the frames $F$ and whose arrows $F\to F'$ are the frame homomorphisms $f:F\to F'$ such that the map $\xi_{f}:{\mathscr{P}}(At(F)) \to {\mathscr{P}}(At(F'))$ which sends a subset $S$ in ${\mathscr{P}}(At(F))$ to the union $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{p\in S}}\psi_{F'}(f(p))$ in ${\mathscr{P}}(At(F'))$ is complete. Note that if $f$ is complete then $\xi_{f}$ is complete; so the category of frames and complete frame homomorphisms between them sits as a subcategory of $\textbf{Frm}_{at}$.
Let $\textbf{AtFrm}$ be the category of atomic frames and complete frame homomorphisms between them, as defined in section \ref{tarski}.
We can define two functors $i':\textbf{AtFrm}\to \textbf{Frm}_{at}$ and $L':\textbf{Frm}_{dis} \to \textbf{AtFrm}$ as follows; $i'$ is the inclusion functor of $\textbf{AtFrm}$ into $\textbf{Frm}_{at}$ (note that this functor is well-defined by Theorem \ref{atthm}) while $L':\textbf{Frm}_{at} \to \textbf{AtFrm}$ sends a poset $F$ in $\textbf{Frm}_{at}$ to the poset ${\mathscr{P}}(At(F))$ and an arrow $f:F\to F'$ in $\textbf{Frm}_{at}$ to the map $\xi_{f}:{\mathscr{P}}(At(F))\to {\mathscr{P}}(At(F'))$.
\begin{theorem}
The embedding
\[
i':\textbf{AtFrm} \hookrightarrow \textbf{Frm}_{at}
\]
identifies $\textbf{AtFrm}$ a full reflective subcategory of the category $\textbf{Frm}_{at}$, with reflector $L':\textbf{Frm}_{at} \to \textbf{AtFrm}$.
\end{theorem}
\begin{proofs}
The proof is entirely analogous to that of Theorem \ref{disadj}.
The fact that $L\circ i:\textbf{AtFrm} \to \textbf{AtFrm}$ is naturally isomorphic to the identity functor follows immediately from Theorem \ref{disthm} and the results of section \ref{tarski}.
To prove that $L'$ is left adjoint to $i'$, we observe that we have the following natural correspondences, natural in $F\in \textbf{Frm}_{at}$ and in $G\in \textbf{AtFrm}$, between the arrows $L(F)\to G$ in $\textbf{DisFrm}$ and the arrows $F\to i'(G)$ in $\textbf{DJFrm}_{dis}$: an arrow $\alpha:L'(F)\to G$ in $\textbf{AtFrm}$ corresponds to the arrow $\alpha\circ \psi_{F}:F\to i'(G)$, while an arrow $\beta:F\to i'(G)$ corresponds to the arrow $\psi_{G}^{-1}\circ L'(\beta)$ given by the composite of $\psi_{G}^{-1}:{\mathscr{P}}(At(G))\to G$ with the arrow $L(\beta):L'(F)\to L'(i'(G))$.
Note that the unit of the adjunction at an object $F$ of $\textbf{Frm}_{at}$ is given by the map $\psi_{F}:F\to {\mathscr{P}}(At(F))$, while the counit at an object $G$ of $\textbf{AtFrm}$ is given by the isomorphism $\psi_{G}^{-1}:{\mathscr{P}}(At(G))\to G$.
\end{proofs}
Of course, by using the same technique of these examples, one can establish similar reflections or, more generally, adjunctions between subcategories of locales and categories of locales satisfying topological properties which can be expressed in terms of the existence of bases satisfying particular conditions; note that the topological properties of locales which we have addressed in the two examples above are local connectedness and atomicity.
\subsection{Toposes paired with points and topological spaces}
In this section we establish, as a further application of the philosophy `toposes as bridges' of \cite{OC10}, adjunctions between full subcategories of the category $\textbf{Top}$ of topological spaces and full subcategories of the category $\mathfrak{Top}_{p}$ of toposes paired with points defined in section \ref{subterminal}.
First, we note that to any topological space $X$ we can associate an object $(\Sh(X), \xi_{X})$ of $\mathfrak{Top}_{p}$ obtained by equipping $\Sh(X)$ with the set of points of $\Sh(X)$ indexed by the set $X_{0}$ of the points of $X$ (see Remark \ref{topolinvariant}(c) above). In fact, this assignment defines a functor $P:\textbf{Top}\to {\mathfrak{Top}_{p}}$, which sends a space $X$ to the object $(\Sh(X), \xi_{X})$ of $\mathfrak{Top}_{p}$ and a continuous map $f:X\to Y$ of topological spaces to the arrow $(P(f), f):(\Sh(X), \xi_{X}) \to (\Sh(Y), \xi_{Y})$ in $\mathfrak{Top}_{p}$, where $P(f)$ is the geometric morphism $\Sh(X)\to \Sh(Y)$ whose inverse image acts, at the level of \'{e}tale bundles, as the pullback functor along $f$ (note that the pair $(P(f), f)$ defines an arrow $(\Sh(X), \xi_{X}) \to (\Sh(Y), \xi_{Y})$ in ${\mathfrak{Top}_{p}}$ since for any $x\in X_{0}$, $f\circ s(x)\cong t(f(x))$, where $s(x):\Set \to \Sh(X)$ is the point of $\Sh(X)$ corresponding to the point $x\in X_{0}$ and $t(f(x)):\Set\to \Sh(Y)$ is the point of $\Sh(Y)$ corresponding to the point $f(x)\in Y_{0}$).
Let $\textbf{\cal U}$ be a full subcategory of $\textbf{Top}$ and $\mathfrak{V}$ be a full subcategory of ${\mathfrak{Top}_{p}}$ which contains all the objects of the form $P(X)$ for $X\in \textbf{\cal U}$. Suppose to have an assignment $\Gamma$ sending each pair $({\cal E}, \xi)$ in $\mathfrak{V}$ to a subframe $\Gamma_{({\cal E}, \xi)}$ of $\Sub_{\cal E}(1)$ in such a way that for any $\cal E$ in $\mathfrak{V}$, $X_{{\tau}^{\cal E}_{\Gamma_{\cal E}}, \xi}$ belongs to $\cal E$, and for any arrow $(f, l):({\cal E}, \xi)\to ({\cal F}, \xi')$ in $\mathfrak{V}$, $f^{\ast}$ sends $\Gamma_{({\cal F}, \xi')}$ to $\Gamma_{({\cal E}, \xi)}$. Then, by the results of section \ref{subterminal}, we have a functor $Q^{\Gamma}:\mathfrak{V}\to \textbf{\cal U}$ sending each pair $({\cal E}, \xi)$ in $\mathfrak{V}$ to the topological space $X_{{\tau}^{\cal E}_{\Gamma_{\cal E}}, \xi}$.
We want to investigate under which conditions we have an adjunction between (restrictions of) the functors $P$ and $Q^{\Gamma}$. The notation below is borrowed from section \ref{subterminal}.
First, we recall that an arrow $({\cal E}, \chi)\to (\Sh(X), \xi_{X})$ in $\mathfrak{Top}_{p}$ consists of a pair $(f, l)$, where $f$ is a geometric morphism ${\cal E} \to \Sh(X)$ and $l:Z\to X_{0}$ (where $Z$ is the domain of $\chi$) is a function such that the diagram
\[
\xymatrix {
[Z, \Set] \ar[r]^{E(l)} \ar[d]^{\chi} & [X_{0}, \Set] \ar[d]^{\xi_{X}} \\
{\cal E} \ar[r]^{f} & \Sh(X)}
\]
commutes (up to isomorphism).
By Diaconescu's equivalence, the geometric morphism $f:{\cal E} \to \Sh(X)$ corresponds to a flat $J$-continuous functor $F(f):{\cal O}(X)\to {\cal E}$, where ${\cal O}(X)$ is the frame of open sets of $X$ and $J$ is the canonical topology on it; note that this functor takes values in the frame of subterminals $\Sub_{\cal E}(1)$ of $\cal E$ and hence it can be equivalently regarded as a frame homomorphism ${\cal O}(X)\to \Sub_{\cal E}(1)$.
The commutativity condition of the square above amounts precisely to the requirement that for every open set $U$ of $X$, $\phi_{\Sub_{\cal F}(1), {\cal E}}(f^{\ast}(U))=l^{-1}(U)$. If ${\cal O}(X)=\Gamma_{(\Sh(X), \xi_{X})}$ (equivalently, $X=Q^{\Gamma}(P(X))$) then this condition implies that $F(f)$ takes values in $\Gamma_{(\cal E, \chi)}$ (we shall denote this image restriction of $F(f)$ to $\Gamma_{(\cal E, \chi)}$ by $F(f)|$) and $l:Z\to X_{0}$ is the underlying function of a continuous map of topological spaces $Q^{\Gamma}(({\cal E}, \chi))\to X$; in fact, $l^{-1}:{\cal O}(X)\to {\cal O}(Q^{\Gamma}(({\cal E}, \chi)))$ can be identified with the composite of $F(f)|$ with the canonical surjection $\phi_{\Gamma_{(\cal E, \chi)}, {\cal E}}:\Gamma_{(\cal E, \chi)} \to {\cal O}(Q^{\Gamma}(({\cal E}, \chi)))$.
Conversely, given a continuous map $l:Q^{\Gamma}(({\cal E}, \chi))\to X$, $l^{-1}$ defines a frame homomorphism ${\cal O}(X) \to {\cal O}(Q^{\Gamma}(({\cal E}, \chi)))$. So, if the canonical surjection $\phi_{\Gamma_{(\cal E, \chi)}, {\cal E}}:\Gamma_{(\cal E, \chi)} \to {\cal O}(Q^{\Gamma}({\cal E}, \xi))$ is injective (equivalently, bijective), we can lift this homomorphism to a frame homomorphism ${\cal O}(X) \to \Gamma_{(\cal E, \chi)} \hookrightarrow \Sub_{\cal E}(1)$. This in turn corresponds to a flat $J$-continuous functor ${\cal O}(X)\to {\cal E}$ and hence to a geometric morphism $f:{\cal E} \to \Sh(X)$ such that $(f,l)$ is an arrow $({\cal E}, \chi)\to (\Sh(X), \xi_{X})$ in $\mathfrak{Top}_{p}$.
So far we have identified two conditions:
\begin{enumerate}[(1)]
\item $X=Q^{\Gamma}(P(X))$ and
\item the injectivity (equivalently, bijectivity) of the canonical surjection\\ $\phi_{\Gamma_{(\cal E, \chi)}, {\cal E}}:\Gamma_{(\cal E, \chi)} \to {\cal O}(Q^{\Gamma}(({\cal E}, \chi)))$,
\end{enumerate}
under which we can establish a bijective correspondence between the arrows $({\cal E}, \chi)\to P(X)$ in $\mathfrak{Top}_{p}$ and the arrows $l:Q^{\Gamma}(({\cal E},\chi))\to X$ in $\textbf{Top}$, natural in $X\in \textbf{{\cal U}}$ and $({\cal E}, \chi)\in \mathfrak{V}$ .
In order to make this correspondence into an adjunction between a subcategory of $\mathfrak{V}$ and a subcategory of $\textbf{{\cal U}}$, we consider the restriction $P|:\textbf{{\cal U}}_{\Gamma} \to \mathfrak{V}$ of $P$ to the full subcategory $\textbf{{\cal U}}_{\Gamma}$ of $\textbf{{\cal U}}$ on the objects $X$ such that $X=Q^{\Gamma}(P(X))$ and the restriction $Q^{\Gamma}|:\mathfrak{V}' \to \textbf{{\cal U}}$ of $Q^{\Gamma}$ to the full subcategory $\mathfrak{V}'$ of $\mathfrak{V}$ on the objects $({\cal E}, \chi)$ such that the canonical surjection $\phi_{\Gamma_{({\cal E}, \chi)}, {\cal E}}:\Gamma_{({\cal E}, \chi)} \to {\cal O}(Q^{\Gamma}(({\cal E}, \chi)))$ is an isomorphism.
Now, since the correspondences established above are inverse to each other and natural in $({\cal E}, \chi) \in \mathfrak{V}'$ and $X\in \textbf{{\cal U}}_{\Gamma}$, to obtain an adjunction between $P|$ and $Q^{\Gamma}|$ it would be enough to show that the functor $P|$ takes values in $\mathfrak{V}'$ and the functor $Q^{\Gamma}|$ takes values in $\textbf{{\cal U}}_{\Gamma}$. The second of these conditions is always satisfied, since for any topological space $X$, $\xi_{X}$ is a separating set of points of $\Sh(X)$ (cf. Theorem \ref{topol}(ii) and Remark \ref{topolinvariant}(c)). On the other hand, for any $({\cal E}, \chi)$ in $\mathfrak{V}$, $Q^{\Gamma}(({\cal E}, \chi))$ belongs to $\textbf{{\cal U}}_{\Gamma}$ if and only if ${\cal O}(Q^{\Gamma}(({\cal E}, \chi)))=\Gamma_{\Sh(Q^{\Gamma}({\cal E}, \chi)), \xi_{Q^{\Gamma}({\cal E}, \chi)}}$ (equivalently, $Q^{\Gamma}(({\cal E}, \chi))=Q^{\Gamma}(P(Q^{\Gamma}(({\cal E}, \chi))))$). Let us denote by $(2')$ the conjunction of this latter condition with condition $(2)$ above. Clearly, for any $X\in \textbf{{\cal U}}_{\Gamma}$, $P(X)$ satisfies condition $(2')$; so, if we denote by ${\mathfrak{V}_{p}}_{\Gamma}$ the full subcategory of $\mathfrak{V}$ on the objects $({\cal E}, \chi)$ which satisfy condition $(2')$, we have the following result.
\begin{theorem}
With the above notation, the restrictions $P_{\Gamma}:\textbf{{\cal U}}_{\Gamma} \to {\mathfrak{V}_{p}}_{\Gamma}$ and $Q_{\Gamma}:{\mathfrak{V}_{p}}_{\Gamma} \to \textbf{{\cal U}}_{\Gamma}$ respectively of the functors $P$ and $Q^{\Gamma}$ define a pair of adjoint functors, with $Q_{\Gamma}$ being the left adjoint and $P_{\Gamma}$ being the right adjoint.
\end{theorem}\qed
Note that the adjunction of the theorem is in fact a reflection, since for any arrow $f:X\to Y$ in $\textbf{Top}$, $f=Q(P(f))$ (cf. condition (1) above), and hence $\textbf{{\cal U}}_{\Gamma}$ can be regarded, via $P_{\Gamma}$, as a full subcategory of ${\mathfrak{V}_{p}}_{\Gamma}$.
Finally, let us discuss a couple of applications of this result.
\begin{enumerate}[(i)]
\item If $\mathfrak{V}=\mathfrak{Top}_{p}$, $\textbf{{\cal U}}=\textbf{Top}$ and $\Gamma_{({\cal E}, \chi)}$ is the subframe $\{0,1\}$ of $\Sub_{\cal E}(1)$ for any $({\cal E}, \chi)$ in $\mathfrak{V}$ then the spaces in $\textbf{{\cal U}}_{\Gamma}$ are exactly the trivial topological spaces (so $\textbf{Top}_{\Gamma}$ is isomorphic to the category $\Set$ of sets), while the objects in ${\mathfrak{V}_{p}}_{\Gamma}$ are exactly the pairs $({\cal E}, \chi)$ such that $\cal E$ is trivial (i.e., $0\cong 1$ in $\cal E$) if and only if $dom(\chi)=\emptyset$. The functor $P_{\Gamma}:\textbf{{\cal U}}_{\Gamma} \to {\mathfrak{V}_{p}}_{\Gamma}$ sends a topological space $X$ in $\textbf{{\cal U}}_{\Gamma}$ to $(\textbf{1}, \xi_{X})$ (where $\textbf{1}\simeq \Sh(X)$ is the trivial topos) if $X$ is empty and to $({\cal S}, \xi_{X})$ if $X$ is non-empty, where ${\cal S}\simeq \Sh(X)$ is the Sierpinski topos i.e. the category of sheaves on the Sierpinski space (equivalently the topos $[\textbf{2}, \Set]$ where $\textbf{2}$ is the preorder category on the natural number $2$), while the functor $Q_{\Gamma}:{\mathfrak{Top}_{p}}_{\Gamma} \to \textbf{Top}_{\Gamma}$ sends an object $({\cal E}, \chi)$ of ${\mathfrak{Top}_{p}}_{\Gamma}$ to the topological space obtained by equipping the domain of $\chi$ with the trivial topology.
\item If $\mathfrak{V}=\mathfrak{Top}_{p}$, $\textbf{{\cal U}}=\textbf{Top}$ and $\Gamma_{({\cal E}, \chi)}$ is the whole frame $\Sub_{\cal E}(1)$ for any $({\cal E}, \chi)$ in $\mathfrak{V}$ then $\textbf{{\cal U}}_{\Gamma}=\textbf{Top}$, while a pair $({\cal E}, \chi)$ in ${\mathfrak{Top}_{p}}$ belongs to ${\mathfrak{V}_{p}}_{\Gamma}$ if and only if the canonical surjection $\phi_{\Sub_{\cal E}(1) , {\cal E}}:\Sub_{\cal E}(1)\to {\cal O}(Q(({\cal E}, \chi))$ is a bijection (that is, if and only if $\chi$ separates the subterminals in $\cal E$ i.e. for any two subterminals $U$ and $V$ in $\cal E$, if $U\ncong V$ then there exists $z\in dom(\chi)$ such that $\chi(z)^{\ast}(U)\ncong \chi(z)^{\ast}(V)$). The functor $P_{\Gamma}:\textbf{{\cal U}}_{\Gamma} \to {\mathfrak{V}_{p}}_{\Gamma}$ sends a topological space $X$ in $\textbf{{\cal U}}_{\Gamma}$ to $(\Sh(X), \xi_{X})$ while the functor $Q_{\Gamma}:{\mathfrak{V}_{p}}_{\Gamma} \to \textbf{{\cal U}}_{\Gamma}$ sends an object $({\cal E}, \chi)$ of ${\mathfrak{V}_{p}}_{\Gamma}$ to the topological space obtained by equipping the domain of $\chi$ with the subterminal topology.
\end{enumerate}
\section{Insights obtained by using invariants}\label{insights}
In this section our aim is to show that the technique of \cite{OC10} of using toposes as `bridges' for transferring information between Morita-equivalent theories (in the form of different sites of definition for the same topos), can be profitably applied in the context of our topos-theoretic interpretation of Stone-type dualities, for translating properties of preordered structures into properties of the corresponding locales or topological spaces (or, more generally, for translating properties between preordered structures $\cal C$ and $\cal D$ related by Morita-equivalences of the form $\Sh({\cal C}, J)\simeq \Sh({\cal D}, K)$).
\subsection{The logical interpretation}\label{logical}
The equivalence $\Sh({\cal C}, J)\simeq \Sh(Id_{J}({\cal C}))$ of Theorem \ref{fund} (for a preorder category ${\cal C}$ and a Grothendieck topology $J$ on ${\cal C}$) can be read as a Morita-equivalence between two distinct geometric theories: the theory of $J$-contin-\\uous flat functors on $\cal C$ and the theory of $J_{can}^{Id_{J}({\cal C})}$-continuous flat functors on $Id_{J}({\cal C})$, where $J_{can}^{Id_{J}({\cal C})}$ is the canonical topology on the frame $Id_{J}({\cal C})$ (cf. \cite{OC10}).
On the other hand, any $J$-continuous flat functor $F:{\cal C}\to {\cal E}$ from a preorder category $\cal C$ to a Grothendieck topos $\cal E$ sends every object in $\cal C$ to a subterminal object of $\cal E$. Indeed, by Diaconescu's equivalence, $F$ is isomorphic to a functor of the form $f^{\ast}\circ a_{J}\circ y$, where $f:{\cal E}\to \Sh({\cal C}, J)$ is a geometric morphism, $y:{\cal C}\to [{\cal C}^{\textrm{op}}, \Set]$ is the Yoneda embedding and $a_{J}:[{\cal C}^{\textrm{op}}, \Set]\to \Sh({\cal C}, J)$ is the associated sheaf functor; and, $\cal C$ being a preorder, the objects of $\cal C$ are sent by $y$ to subterminals in $[{\cal C}^{\textrm{op}}, \Set]$, which are in turn sent by $f^{\ast}\circ a_{J}$ to subterminals in $\cal E$.
From the characterization of flat functors ${\cal C}\to {\cal E}$ as filtering functors given in section VII.9 of \cite{MM} it thus follows that the $J$-continuous flat functors ${\cal C}\to {\cal E}$ and natural transformations between them can be identified, naturally in $\cal E$, respectively with the models in $\cal E$ and model homomorphisms between them of the propositional theory ${\mathbb T}^{{\cal C}}_{J}$ defined as follows: the signature of ${\mathbb T}^{{\cal C}}_{J}$ has no sorts and one atomic proposition $F_{a}$ for each element $a\in {\cal C}$, and the axioms of ${\mathbb T}^{{\cal C}}_{J}$ are the following:
\[
(\top \vdash \mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{c\in {\cal C}}} F_{c});
\]
\[
(F_{a} \vdash F_{b})
\]
for any $a\leq b$ in $\cal C$;
\[
(F_{a}\wedge F_{b} \vdash \mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{c\in K_{a,b}}} F_{c})
\]
for any $a, b \in {\cal C}$, where $K_{a,b}$ is the collection of all the elements $c\in {\cal C}$ such that $c\leq a$ and $c\leq b$;
\[
(F_{a} \vdash \mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}F_{a_{i}})
\]
for any $J$-covering sieve $\{a_{i} \to a \textrm{ | } i\in I\}$ in $\cal C$.
We note that the models of the theory ${\mathbb T}^{{\cal C}}_{J}$ in $\Set$ are precisely the $J$-prime filters on $\cal C$, as defined in section \ref{exsub}; accordingly, we call the theory ${\mathbb T}^{{\cal C}}_{J}$ the \emph{theory of $J$-prime filters}. In particular, if $L$ is a frame and $J$ is the canonical topology on it then the theory ${\mathbb T}^{{\cal C}}_{J}$ specializes to the propositional theory of completely prime filters on $L$ defined in section D1.1 of \cite{El}. Given a topological space $X$, we define the \emph{theory of completely prime filters on $X$} as the theory of completely prime filters on the corresponding frame ${\cal O}(X)$ of open sets of $X$.
Summarizing, we have the following result.
\begin{theorem}\label{classif}
Let $\cal C$ be a preorder category and $J$ be a Grothendieck topology on $\cal C$. Then the theory ${\mathbb T}^{{\cal C}}_{J}$ of $J$-prime filters is classified by the topos $\Sh({\cal C}, J)$.
\end{theorem}\qed
\begin{remarks}
\begin{enumerate}[(a)]
\item If ${\cal C}={\cal P}^{\textrm{op}}$, where $({\cal P}, \leq)$ is a preorder category, then the models of ${\mathbb T}^{{\cal C}}_{J}$ in $\Set$, i.e. the $J$-prime filters on $\cal C$, are precisely the non-empty directed ideals on $\cal P$;
\item If $\cal C$ is cartesian (resp. coherent, geometric) and $J$ is the trivial (resp. coherent, geometric) topology on $\cal C$, the theory ${\mathbb T}^{{\cal C}}_{j}$ specializes to the theory of filters (resp. prime filters, completely prime filters) on $\cal C$ described in section D1.1 \cite{El}.
\end{enumerate}
\end{remarks}
The equivalence $\Sh({\cal C}, J)\simeq \Sh(Id_{J}({\cal C}))$ of Theorem \ref{fund} can thus be read logically as a Morita-equivalence between the theory of $J$-prime filters on $\cal C$ and the theory of completely prime filters on $Id_{J}({\cal C})$. In particular, Stone duality for distributive lattices admits the following logical interpretation: given a locale $L$, $L$ is the Stone locale associated to a distributive lattice $\cal D$ if and only if the theory of prime filters on $\cal D$ is Morita-equivalent to the theory of completely prime filters on $L$. Of course, the other dualities of sections \ref{ex} and \ref{addex} admit similar interpretations.
As we have already remarked in section \ref{general}, the `Stone-type dualities' can be seen as arising from the process of `functorializing' a bunch of Morita-equivalences; we have one Morita-equivalence for each of the preordered structures, and these Morita-equivalences can be `merged together' to produce a `global' duality or equivalence of categories. In fact, this method of generating dualities or equivalences of categories starting from `parametrised' Morita-equivalences, which we have exploited in section \ref{Mordual}, is likely to find, because of its generality, new applications in a variety of different mathematical contexts in the near future.
From Theorem \ref{classif} it follows, by recalling the syntactic construction of classifying toposes, that for any site $({\cal C}, J)$ whose underlying category $\cal C$ is a preorder, $Id_{J}({\cal C})$ can be identified with the geometric syntactic category of the theory ${\mathbb T}^{{\cal C}}_{J}$, and hence the theory of completely prime filters on $Id_{J}({\cal C})$ is Morita-equivalent to the theory of completely prime filters on the logical space of ${\mathbb T}^{{\cal C}}_{J}$, as defined in section \ref{exsub}.
The localic dualities of section \ref{ex} can thus be read logically as `functorialized' Morita-equivalences (in the sense explained above) between propositional geometric theories and the theories of completely prime filters on their logical spaces.
Recall that the technique `toposes as bridges' introduced in \cite{OC10} consists, broadly speaking, in expressing a given topos-theoretic invariant in terms of two distinct sites of definition of a given topos; the topos acts as a `bridge' enabling one to transfer properties from one site into the other. Of course, the feasibility of this method heavily depends on how `natural' is, for the given invariant, the relationship between the topos and its different sites of definition. There is an important class of invariants for which this transfer of information is always feasible and, in a sense, even \emph{automatic}. Indeed, for logically motivated invariants such as the property of a topos to be two-valued (resp. Boolean, De Morgan), one has bijective characterizations (holding `uniformly' for any site $({\cal C}, J)$) of the kind `$\Sh({\cal C}, J)$ satisfies the invariant if and only if the site $({\cal C}, J)$ satisfies some `tractable' categorical property'. In fact, any first-order sequent $\sigma$ written in the algebraic language of the theory of Heyting algebras can be interpreted in the internal Heyting algebra $\Omega_{\cal E}$ of a topos $\cal E$ given by its subobject classifier, and as such it gives rise to a topos-theoretic invariant (namely, the interpretation of $\sigma$ in the algebra $\Omega_{{\cal E}}$) admitting a site characterization of the above kind (such characterizations can be obtained by using the canonical description of the algebra $\Omega_{\cal E}$ in terms of the given site of definition of $\cal E$, and the explicit descriptions of the interpretation of the first-order connectives and quantifiers in a topos of sheaves on a site, as given for example in Chapter III of \cite{MM}).
One might naturally wonder whether, given a sequent $\sigma$ as above, its (internal) validity in the algebra $\Omega_{\Sh(L)}$ of the topos $\Sh(L)$ is equivalent to its (external) validity in the locale $L$ (regarded as a model of the theory of Heyting algebras). This is not always the case in general, but there is an important class of sequents $\sigma$ for which these two notions of validity coincide.
\begin{proposition}\label{cartesian}
Let $\sigma$ be a cartesian (in particular, Horn) sequent in the theory of Heyting algebras. Then for any locale $L$, $\sigma$ is valid in the internal algebra $\Omega_{\Sh(L)}$ of the topos $\Sh(L)$ if and only if it is valid in $L$ (regarded as a model of the theory of Heyting algebras).
\end{proposition}
\begin{proofs}
For any locally small topos $\cal E$, from the fact the Yoneda embedding $y:{\cal E}\to [{\cal E}^{\textrm{op}}, \Set]$ is a cartesian functor, it follows that $y$ preserves and the interpretation of all the cartesian formulae and hence, given any cartesian sequent $\sigma$ in the theory of Heyting algebras, the internal Heyting algebra $\Omega_{{\cal E}}$ satisfies $\sigma$ if and only if every frame $\Sub_{\cal E}(e)\cong Hom_{\cal E}(e, \Omega)$ in $\cal E$ satisfies $\sigma$. Now, given a Grothendieck topos $\cal E$ and an object $e\in \cal E$, if $\cal C$ is a separating set for $\cal E$ then $e$ can be expressed as a quotient of a coproduct of objects in $\cal C$, that is there exists a set-indexed family $\{c_{i} \textrm{ | } i\in I\}$ of objects in $\cal C$ and an epimorphism $p:\coprod_{i\in I}c_{i}\epi e$. Since $p$ is an epimorphism, the pullback functor $p^{\ast}:\Sub_{\cal E}(e) \rightarrow \Sub_{\cal E}(\coprod_{i\in I}c_{i})\cong \prod_{i\in I}\Sub_{\cal E}(c_{i})$ is logical and conservative (cf. Example A4.2.7(a) \cite{El}); so $\Sub_{\cal E}(e)$ satisfies a first-order sequent $\sigma$ if all the $\Sub_{\cal E}(c_{i})$ do. Also, if $m:b\mono a$ is a monomorphism in $\cal E$ then the pullback functor $m^{\ast}:\Sub_{\cal E}(a)\to \Sub_{\cal E}(b)$ is logical and essentially surjective; so, if $\Sub_{\cal E}(a)$ satisfies $\sigma$ then $\Sub_{\cal E}(b)$ satisfies $\sigma$. Our thesis now follows from the combination of all these facts by recalling that, for any locale $L$, the collection of all the subterminals of $\Sh(L)$ forms a separating set of $\Sh(L)$, and $L\cong \Sub_{\Sh(L)}(1)$.
\end{proofs}
An example of this kind of invariants, which we shall consider below, is the interpretation of the sequent
\[
(\top \vdash_{x,y} (x\imp y) \vee (y\imp x)),
\]
whose validity in the algebra $\Omega_{\cal E}$ of a topos $\cal E$ amounts to saying that $\cal E$ satisfies G\"{o}del-Dummett's law.
Recall that the subobject classifier $\Omega_{\Sh({\cal C}, J)}:{\cal C}^{\textrm{op}}\to \Set$ of a topos $\Sh({\cal C}, J)$ of sheaves on a site $({\cal C}, J)$ is (isomorphic to) the functor sending any object $c$ of $\cal C$ to the collection $\Omega_{\Sh({\cal C}, J)}(c)$ of $J$-closed sieves on $c$, and an arrow $f:d\to c$ to the pullback operation $f^{\ast}:\Omega_{\Sh({\cal C}, J)}(c)\to \Omega_{\Sh({\cal C}, J)}(d)$ of sieves along $f$. Proposition \ref{cartesian} thus enables us to express `external' properties of locales $L$ formulated as cartesian sequents in the theory of Heyting algebras as `internal' properties of the corresponding localic topos $\Sh(L)$, by means of a topos-theoretic invariant which can in turn be reformulated in terms of any other site of definition $({\cal C}, J)$ of the same topos, leading to categorical characterizations involving $J$-closed sieves on the category $\cal C$. Note that this technique is bound to bring more substantial insights than a straightforward reformulation of the given property of $L$ as a topos-theoretic invariant on the frame $\Sub_{\cal E}(1)$ of subterminals of the topos ${\cal E}\simeq \Sh(L)$, since a formulation of this latter kind would merely consist in a translation of the given property of $L$ across the isomorphism $L\simeq \Sub_{\cal E}(1)$.
In the next sections, we shall provide applications of the methodology just described by using, as invariants, the property of a topos to be Boolean, to be De Morgan, to be two-valued, to satisfy G\"{o}del-Dummett logic; anyway, the reader should bear in mind that these particular examples are just meant to show the effectiveness of our technique, and notice that a whole range of new results can be `automatically' generated by applying the same method to different invariants.
In addition to the invariants discussed above, there are of course many other ones which behave `naturally' with respect to sites (cf. \cite{OC10} for an extensive discussion of these aspects) and hence are appropriate for an application of the methodology `toposes as bridges'. Two of them, which we will consider in the next section in relation to the Morita-equivalence of Theorem \ref{fund}, are the notions of point and subtopos of a given topos.
\subsection{$J$-ideals and $J$-prime filters}
Recall that, for any site $({\cal C}, J)$, the points of a topos $\Sh({\cal C}, J)$ (i.e., the geometric morphisms $\Set \to \Sh({\cal C}, J)$), can be naturally identified with the flat $J$-continuous functors on $\cal C$, while the subtoposes of $\Sh({\cal C}, J)$ (i.e. the equivalence classes of geometric inclusions into $\Sh({\cal C}, J)$) correspond precisely to the Grothendieck topologies on $\cal C$ which contain $J$.
The method `toposes as bridges' applied to the invariant notion of point and to the Morita-equivalence of Theorem \ref{fund} produces an identification between the $J$-prime filters on $\cal C$ and the completely prime filters on $Id_{J}({\cal C})$.
\begin{theorem}\label{bijectionpoints}
Let $\cal C$ be a preorder and $J$ be a Grothendieck topology on $\cal C$. Then the assignment sending a filter $F$ on $Id_{J}({\cal C})$ to the $J$-prime filter $\{c\in {\cal C} \textrm{ | } (c)\downarrow_{J}\in F\}$ on $\cal C$ defines a bijection between the completely prime filters on the frame $Id_{J}({\cal C})$ of $J$-ideals of $\cal C$ and the $J$-prime filters on $\cal C$.
In particular,
\begin{enumerate}[(i)]
\item if $\cal C$ is a meet-semilattice, we have a natural bijection between the filters on $\cal C$ and the completely prime filters on the frame of ideals (i.e. lower sets) of $\cal C$, and
\item if $\cal C$ is a distributive lattice we have a natural bijection between the prime filters on $\cal C$ and the completely prime filters on the frame of ideals (i.e. lower sets which are closed under finite joins) of $\cal C$.
\end{enumerate}
\end{theorem}
\begin{proofs}
Starting from the equivalence $\Sh({\cal C}, J)\simeq \Sh(Id_{J}({\cal C}))$ of Theorem \ref{fund}, it suffices to observe that the bijection between the points of the two toposes induced by such equivalence yields, at the level of filters, the assignment sending a filter $F$ on $Id_{J}({\cal C})$ to the filter $\{c\in {\cal C} \textrm{ | } (c)\downarrow_{J}\in F\}$ on $\cal C$.
\end{proofs}
We note that, while this result follows an a natural and immediate application of our method `toposes as bridges', lengthier and less conceptual arguments would be necessary to prove this result directly, that is without appealing to the topos-theoretic machinery (the reader might find instructive to compare our proof of the thoerem with the topological arguments used to establish the particular case of the result for distributive lattices given in Chapter II of \cite{stone}, or with the direct proof of the theorem given in Appendix \ref{appendix} below).
A further application of the philosophy `toposes as bridges' to the Morita-equivalence of Theorem \ref{fund} is the following result, obtained by considering as topos-theoretic invariant the notion of subtopos.
Recall that, given a site $({\cal C}, J)$ whose underlying category is a preorder, and any lower-set $I$ on $\cal C$, its $J$-closure $cl_{J}(I)$ is the smallest $J$-ideal on $\cal C$ which contains $I$, i.e. the ideal $cl_{J}(I)=\{c\in {\cal C} \textrm{ | } \{f:d\to c \textrm{ in $\cal C$ | } d\in I\} \in J(c)\}$.
\begin{theorem}\label{existence}
Let $\cal C$ be a preorder and $J$ be a Grothendieck topology on $\cal C$. For any surjective frame homomorphism $f:Id_{J}({\cal C})\to F$ onto a frame $F$ there exists a Grothendieck topology $J'\supseteq J$ on $\cal C$ such that $F\cong Id_{J}({\cal C})$ and $f$ corresponds, under this isomorphism, to the frame homomorphism $cl_{J'}:Id_{J}({\cal C}) \to Id_{J'}({\cal C})$ sending a $J$-ideal $I$ on $\cal C$ to its $J'$-closure. In particular, any surjective frame homomorphism whose domain is the frame $Id({\cal C})$ of lower sets on $\cal C$ is, up to isomorphism, of the form $cl_{J}:Id({\cal C}) \to Id_{J}({\cal C}) $ for some Grothendieck topology $J$.
\end{theorem}
\begin{proofs}
A surjective frame homomorphism $f:Id_{J}({\cal C})\to F$ corresponds to an embedding of the corresponding locales and hence to a geometric inclusion $\Sh(f):\Sh(F)\hookrightarrow \Sh(Id_{J}({\cal C}))$. Through the equivalence
\[
\Sh({\cal C}, J)\simeq \Sh(Id_{J}({\cal C}))
\]
of Theorem \ref{fund}, this inclusion transfers to a subtopos of $\Sh({\cal C}, J)$; and this subtopos must necessarily be, up to equivalence, of the form $i_{J'}:\Sh({\cal C}, J')\hookrightarrow \Sh({\cal C}, J)$ for a Grothendieck topology $J'$ on $\cal C$ which contains $J$, where $i_{J'}$ is the canonical geometric inclusion.
Since the equivalence of Theorem \ref{fund} is natural with respect to inclusions of Grothendieck topologies, we have a commutative diagram
\[
\xymatrix {
\Sh(Id_{J}({\cal C})) \ar[r]^{\simeq} & \Sh({\cal C}, J) \ar[r]^{\simeq} & \Sh(Id_{J}({\cal C})) \\
\Sh(Id_{J'}({\cal C})) \ar[u]^{\Sh(cl_{J'})} \ar[r]^{\simeq} & \Sh({\cal C}, J') \ar[u]^{i_{J'}} \ar[r]^{\simeq} & \Sh(F) \ar[u]^{\Sh(f)}}
\]
from which it follows that $\Sh(f)$ and $\Sh(cl_{J'})$ are equivalent as subtoposes of $\Sh(Id_{J}({\cal C}))$, equivalently $f$ is isomorphic to $cl_{J'}$, as required.
\end{proofs}
Considering as topos-theoretic invariant the notion of equivalence of sub-\\toposes, applied to the Morita-equivalence
\[
\Sh({\cal C}, J)\simeq \Sh(Id_{J}({\cal C}))
\]
of Theorem \ref{fund}, we obtain the following result.
\begin{theorem}\label{unique}
Let $\cal C$ be a preorder and let $J_{1}$ and $J_{2}$ be two Grothendieck topologies on $\cal C$. If for any lower set $I$ in $\cal C$, $I$ is a $J_{1}$-ideal if and only if it is a $J_{2}$-ideal then $J_{1}=J_{2}$.
\end{theorem}
\begin{proofs}
By Theorem \ref{fund}, for any Grothendieck topology $J$ on $\cal C$ we have a commutative diagram
\[
\xymatrix {
\Sh({\cal C}, J) \ar[r]^{\simeq} \ar[d]^{i_{J}} & \Sh(Id_{J}({\cal C})) \ar[d]^{\Sh(cl_{J})} \\
[{\cal C}^{\textrm{op}}, \Set] \ar[r]^{\simeq} & \Sh(Id({\cal C}))}
\]
where $i_{J}:\Sh({\cal C}, J)\to [{\cal C}^{\textrm{op}}, \Set]$ is the canonical inclusion, $Id({\cal C})$ is the frame of lower sets in $\cal C$ and $\Sh(cl_{J}):\Sh(Id_{J}({\cal C})) \to \Sh(Id({\cal C}))$ is the geometric morphism induced by the frame homomorphism $cl_{J}:Id({\cal C}) \to Id_{J}({\cal C})$.
Now, the condition `for any lower set $I$ in $\cal C$, $I$ is a $J_{1}$-ideal if and only if it is a $J_{2}$-ideal' can be expressed topos-theoretically by saying that there exists an equivalence $e:\Sh(Id_{J_{1}}({\cal C}))\to \Sh(Id_{J_{2}}({\cal C}))$ such that $\Sh(cl_{J_{2}}) \circ e \simeq \Sh(cl_{J_{1}})$. But transferring such an equivalence through the Morita-equivalence of Theorem \ref{fund} yields an equivalence $e':\Sh({\cal C}, J_{1}) \to \Sh({\cal C}, J_{2})$ such that $i_{J_{2}} \circ e' \simeq i_{J_{1}}$. To conclude the proof of the theorem, it suffices to recall the standard topos-theoretic fact that such an equivalence exists if and only if $J_{1}=J_{2}$.
\end{proofs}
Note that Theorem \ref{unique} assures the uniqueness of the Grothendieck topology $J'$ in the statement of Theorem \ref{existence}.
\subsection{Topological properties as topos-theoretic invariants}
In this section we apply our usual technique `toposes as bridges' (cf. \cite{OC10}) to obtain insights concerning the relationship between preordered structures and the locales or topological spaces which correspond to them under the dualities or equivalences of section \ref{ex}.
\subsubsection{Almost discreteness via the law of excluded middle}
Let us start by investigating the almost discreteness of the locale corresponding to a distributive lattice (resp. to a meet-semilattice, to a preorder) via Stone duality (resp. the duality for meet-semilattices of Theorem \ref{meetsm}, the functor $B:\textbf{Pro}\to \textbf{AlexLoc}$ of section \ref{prealex}).
If $\cal D$ is a distributive lattice, equipped with the coherent topology $J_{\cal D}$, we have two different sites of definition for the topos associated to it via the technique of section \ref{locales}: $\Sh({\cal D}, J_{\cal D}) \simeq \Sh(Id_{J_{\cal D}}({\cal D}))$. We call $Id_{J_{\cal D}}({\cal D})$ the \emph{Stone locale} associated to $\cal D$ (as in section \ref{stonedist}), and we denote it by $L_{\cal D}$.
Moreover, the Comparison Lemma yields a third site of definition of the topos, obtained by cutting down the site $({\cal D}, J_{\cal D})$ to the full subcategory ${\cal D}^{\ast}$ on the non-zero objects and equipping it with the induced Grothendieck topology $J_{\cal D}^{\ast}=J_{\cal D}|_{{\cal D}^{\ast}}$.
As we shall see below, rephrasing a given topos-theoretic invariant in terms of these three different sites of definition leads to three different expressions of it, each of them written in the `language' of the corresponding site, which are nonetheless equivalent to each other.
The following proposition provides several alternative characterizations of the property of the Stone locale associated to a distributive lattice to be almost discrete, obtained by applying this technique. The relevant topos-theoretic invariant is the property of a topos to be Boolean; in fact, it is well-known that for any locale $L$, $\Sh(L)$ is Boolean if and only if $L$ is almost discrete (i.e. every element of $L$ is complemented).
In the proofs below, the symbol $\neg$, applied to an element $l$ of a locale $L$, denotes the Heyting pseudocomplement of $l$ in $L$ (where $L$ is regarded as a Heyting algebra).
\begin{proposition}\label{boolean}
Let $\cal D$ be a distributive lattice, and let $L_{\cal D}$ (resp. $X_{\cal D}$) be its associated Stone locale (resp. Stone space). Then the following conditions are equivalent:
\begin{enumerate}[(i)]
\item $L_{\cal D}$ (equivalently, $X_{\cal D}$) is almost discrete;
\item For any ideal $I$ of $\cal D$, if $a\in {\cal D}$ is an element satisfying the property that for any non-zero element $b\leq a$ there exists a non-zero element $c\in I$ such that $c\leq b$ then $a\in I$;
\item For any collection $\{a_{i} \textrm{ | } i\in I\}$ of non-zero elements of $\cal D$ with the property that for any non-zero element $a$ there exists $i\in I$ such that $a_{i}\wedge a\neq 0$, there exists a finite subset $J\subseteq I$ such that $1=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in J}}a_{i}$;
\item Every element of $\cal D$ is complemented in $\cal D$, $\cal D$ is complete and the supremum in $\cal D$ of any subset $S$ of $\cal D$ is a finite join of elements of $S$;
\item $\cal D$ is a finite Boolean algebra.
\end{enumerate}
\end{proposition}
\begin{proofs}
It is well-known that a locale $L$ is almost discrete if and only if the topos $\Sh(L)$ is Boolean.
Now, condition $(ii)$ represents the expression of the invariant property of a topos to be Boolean in terms of the representation $\Sh(Id_{J_{\cal D}}({\cal D}))$ of the topos $\Sh(L_{\cal D})$, it being the assertion that every ideal $I$ in $Id_{J_{\cal D}}({\cal D})$ satisfies $\neg\neg I=I$ (equivalently, $\neg\neg I\subseteq I$).
Condition $(iii)$ is the expression of the invariant in terms of the representation $\Sh({\cal D}^{\ast}, J_{\cal D}^{\ast})$ of the topos $\Sh(L_{\cal D})$, as established in \cite{OC3}.
Starting from the assumption that $L_{\cal D}$ is almost discrete, condition $(iv)$ can be deduced as follows. The top element of $L_{\cal D}=Id_{J_{\cal D}}({\cal D})$ being compact, every (complemented) element of $Id_{J_{\cal D}}({\cal D})$ is compact, i.e. it is a principal ideal (cf. Corollary \ref{cor}(i)). Since every element of $L_{\cal D}$ is complemented, every element of $\cal D$ is complemented in $\cal D$. Also, since every ideal in $Id_{J_{\cal D}}({\cal D})$ is principal, $\cal D$ is isomorphic to $Id_{J_{\cal D}}({\cal D})$ and hence in particular $\cal D$ is complete. For a given subset $S$ of $\cal D$, consider the $J_{\cal D}$-ideal $I$ on $\cal D$ generated by $S$; since $I$ is principal then $I=(a)\downarrow$ for some element $a\in {\cal D}$. Clearly, by construction of $I$, $a$ is the supremum of $S$ and can be expressed as a finite join of elements in $S$. Conversely, let us suppose that condition $(iv)$ holds. Since $\cal D$ is complete and the supremum of any subset $S$ of $\cal D$ is a finite join of elements in $S$ then every ideal in $Id_{J_{\cal D}}({\cal D})$ is principal. From the fact that every element of $\cal D$ is complemented we can thus conclude that every ideal in $Id_{J_{\cal D}}({\cal D})$ is complemented, in other words that $L_{\cal D}$ is almost discrete.
Condition $(iv)$ is clearly equivalent to condition $(v)$; indeed, $(iv)$ implies $(v)$ since, the top element of $L_{\cal D}$ being compact, if every element in $\cal D$ is complemented then $\cal D$ is a finite Boolean algebra, while the fact that $(v)$ implies $(iv)$ is obvious.
\end{proofs}
The next proposition provides characterizations of the meet-semilattices (resp. preorders) whose corresponding locales via the duality for meet-semilattices of Theorem \ref{meetsm} (resp. the functor $B:\textbf{Pro}\to \textbf{AlexLoc}$ of section \ref{prealex}) are almost discrete. Again, the technique consists in using the topos as a bridge for transferring a given invariant (in this case, the property of a topos to be Boolean) across two different sites of definition for it.
\begin{proposition}\label{Boole2}
\begin{enumerate}[(i)]
\item For any meet-semilattice $\cal M$, the ideal locale $S_{\cal M}$ (equivalently, the ideal topological space) associated to $\cal M$ is almost discrete if and only if $\cal M$ is a singleton.
\item For any preorder $\cal P$, the Alexandrov locale $A_{\cal P}$ (equivalently, the Alexandov space) associated to $\cal P$ is almost discrete if and only if for any $p,q\in {\cal P}$, $p\leq q$ implies $q\leq p$.
\end{enumerate}
\end{proposition}
\begin{proofs}
We use the following well-known site characterizations for the invariant property of a topos to be Boolean:
\begin{enumerate}[(a)]
\item A presheaf topos $[{\cal C}^{\textrm{op}}, \Set]$ is Boolean if and only if the category $\cal C$ is a groupoid;
\item A localic topos $\Sh(L)$ is Boolean if and only if the locale $L$ is almost discrete.
\end{enumerate}
The thesis follows immediately from expressing the invariant property of the topos $\Sh(S_{\cal M})\simeq [{\cal M}^{\textrm{op}}, \Set]$ (resp. $\Sh(A_{\cal P})\simeq [{\cal P}, \Set]$) to be Boolean in terms of the two different sites of definition of it, according to the site characterizations reported above.
\end{proofs}
\subsubsection{Extremal disconnectedness via De Morgan's law}
The following proposition represents the analogue of Proposition \ref{boolean} for the property of a Stone locale to be extremally disconnected. Again, the proof is based on the consideration of a topos-theoretic invariant, namely the property of a topos to be De Morgan, in relation to different sites of definition of a given topos.
\begin{proposition}\label{morgan}
Let $\cal D$ be a distributive lattice, and let $L_{\cal D}$ (resp. $X_{\cal D}$) be its associated Stone locale (resp. Stone space). Then the following conditions are equivalent:
\begin{enumerate}[(i)]
\item $L_{\cal D}$ (equivalently, $X_{\cal D}$) is extremally disconnected;
\item For every ideal $I$ of $\cal D$ there exists a finite covering $1=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ of $1$ such that each $a_{i}$ either belongs to $\neg I$ (i.e., for any non-zero element $b\leq a_{i}$, $b\notin I$) or to $\neg\neg I$ (i.e., for any non-zero element $b\leq a_{i}$ there exists a non-zero element $c\in I$ such that $c\leq b$);
\item For any collection $\{a_{i} \textrm{ | } i\in I\}$ of non-zero elements of $\cal D$ there exists a finite family $\{b_{j} \textrm{ | } j\in J\}$ of non-zero elements of $\cal D$ such that $1=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{j\in J}}b_{j}$ and for each $j\in J$ either $b_{j}\wedge a_{i}=0$ for all $i\in I$ or for every non-zero element $x\leq b_{j}$ there exists $i\in I$ such that $x\wedge a_{i}\neq 0$;
\item For any ideal $I$ of $\cal D$ with the property that for any element $a\in {\cal D}$ such that for any non-zero element $b\leq a$ there exists a non-zero element $c\in I$ such that $c\leq b$, $a\in I$, there exists a complemented element $x\in {\cal D}$ such that $I=(x)\downarrow$. In fact, the lattice of complemented elements of $\cal D$ is isomorphic, via the map sending any complemented elements to the principal ideal which it generates, to the frame of ideals $I$ such that $\neg\neg I=I$.
\end{enumerate}
\end{proposition}
\begin{proofs}
It is well-known that a locale $L$ is extremally disconnected if and only if the topos $\Sh(L)$ is De Morgan.
Condition $(ii)$ represents the expression of this invariant in terms of the representation $\Sh(Id_{J_{\cal D}}({\cal D}))$ of the topos, it being the assertion that every ideal $I$ in $Id_{J_{\cal D}}({\cal D})$ satisfies $\neg I \vee \neg\neg I=L_{\cal D}$.
Condition $(iii)$ is the expression of the invariant in terms of the representation $\Sh({\cal D}^{\ast}, J_{\cal D}^{\ast})$ of the topos $\Sh(L_{\cal D})$, as established in \cite{OC3}.
The equivalence $(iv)\biimp (i)$ can be proved as follows. It is well-known that $L_{\cal D}$ is extremally disconnected if and only if every $\neg\neg$-stable element is complemented. But the complemented ideals, being compact, must all be principal, from which our thesis follows immediately.
\end{proofs}
If $\cal D$ is a Boolean algebra then the property of completeness of $\cal D$ is sufficient (as well as necessary) to ensure that condition $(iv)$ holds. We can show this as follows. If $\cal D$ is complete and $I$ is an ideal on $\cal D$ satisfying the property in condition $(iv)$, $I=(x)\downarrow$, where $x$ is the supremum of $I$ in $\cal D$. Indeed, it is clear that $I\subseteq (x)\downarrow$, while $x\in I$ because for any non-zero element $b\leq x$, since $b=b\wedge \mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{a\in I}}a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{a\in I}}(b\wedge a)$ (note that, $\cal D$ being a complete Boolean algebra, $\cal D$ is a frame and hence the infinite distributive law holds in $\cal D$), there is an element $a\in I$ such that $b\wedge a\leq b$ is non-zero and belongs to $I$ ($I$ being an ideal). We have thus recovered Lemma III.3.5 \cite{stone}.
The next proposition represents the analogue of Proposition \ref{Boole2} for meet-semilattices and preorders.
\begin{proposition}
\begin{enumerate}[(i)]
\item For any meet-semilattice $\cal M$, the ideal locale $S_{\cal M}$ (equivalently, the ideal topological space) associated to $\cal M$ is extremally disconnected;
\item For any preorder $({\cal P}, \leq)$, the Alexandrov locale $A_{\cal P}$ (equivalently, the Alexandrov space) associated to $\cal P$ is extremally disconnected if and only if $\cal P$ satisfies the amalgamation property (i.e. for any elements $a,b,c \in {\cal P}$ such that $c \leq a, b$ there exists $d\in {\cal P}$ such that $a,b \leq d$).
\end{enumerate}
\end{proposition}
\begin{proofs}
The thesis follows from a similar argument to that in the proof of Proposition \ref{Boole2}, by using the following site characterizations for the invariant property of a topos to be De Morgan (cf. \cite{El}):
\begin{enumerate}[(a)]
\item A presheaf topos $[{\cal C}^{\textrm{op}}, \Set]$ is De Morgan if and only if the category $\cal C$ satisfies the right Ore condition (i.e. the dual of the amalgamation property);
\item A localic topos $\Sh(L)$ is De Morgan if and only if the locale $L$ is extremally disconnected.
\end{enumerate}
\end{proofs}
\subsubsection{Triviality via two-valuedness}
By a two-valued locale we mean a locale $L$ such that the only two elements of $L$ are $0$ and $1$, and they are distinct from each other. Note that, for a topological space $X$, ${\cal O}(X)$ is two-valued if and only if the underlying set $X_{0}$ of $X$ is non-empty and the topology on $X$ is trivial.
Considering the invariant property of a topos to be two-valued, we obtain the following result.
\begin{proposition}
\begin{enumerate}[(i)]
\item If $\cal D$ is a distributive lattice, its associated Stone locale $L_{\cal D}$ is two-valued (equivalently, its associated Stone space is trivial and non-empty) if and only if $\cal D$ is two-valued (i.e., the only two elements of ${\cal D}$ are $0$ and $1$, and they are distinct from each other);
\item If $\cal M$ is a meet-semilattice and $S_{\cal M}$ is the ideal locale associated to $\cal M$ then $S_{\cal M}$ is two-valued (equivalently, its associated ideal topological space is trivial and non-empty) if and only if $M$ is a singleton;
\item If $({\cal P}, \leq)$ is a preorder and $A_{\cal P}$ is the Alexandrov locale associated to $\cal P$ then $A_{\cal P}$ is two-valued (equivalently, its associated Alexandrov space is trivial and non-empty) if and only if for every $p,q\in {\cal P}$, $p\leq q$ and $q\leq p$.
\end{enumerate}
\end{proposition}
\begin{proofs}
The method of proof is always the same as that employed in the proofs of the previous propositions. In this case, the invariant is the property of a topos to be two-valued, while the two different site representations for the topos are $\Sh(L_{\cal D}) \simeq \Sh({\cal D}, J_{\cal D})$ (resp. $\Sh(S_{\cal M})\simeq [{\cal M}^{\textrm{op}}, \Set]$, $\Sh(A_{\cal P})\simeq [{\cal P}, \Set]$).
The characterization of the property of two-valuedness in terms of the site $({\cal D}, J_{\cal D})$ is easily seen to yield the property of $\cal D$ to be two-valued (i.e., the condition that the only two elements of ${\cal D}$ are $0$ and $1$, and they are distinct from each other).
The other site characterizations, leading to $(i)$, $(ii)$ and $(iii)$, are the following:
\begin{enumerate}[(a)]
\item A presheaf topos $[{\cal C}^{\textrm{op}}, \Set]$ is two-valued if and only if the category $\cal C$ is strongly connected (i.e. for any two objects $a, b \in {\cal C}$, there exist arrows $a\to b$ and $b\to a$);
\item A localic topos $\Sh(L)$ is two-valued if and only if the locale $L$ is two-valued.
\end{enumerate}
\end{proofs}
\subsubsection{G\"{o}del-Dummett's law as an invariant}
Another logically-motivated topos-theoretic invariant which admits natural site characterizations is the property of a topos to satisfy G\"{o}del-Dummett's law, in the sense that the sequent
\[
(\top \vdash_{x,y} (x\imp y) \vee (y\imp x))
\]
holds in the internal Heyting algebra of the topos given by the subobject classifier. One easily calculates, by using the well-known explicit descriptions of the internal Heyting algebra operations on the subobject classifier $\Omega_{\Sh({\cal C}, J)}$ of a topos $\Sh({\cal C}, J)$ of sheaves on a site (as for example given in chapter III of \cite{MM}), that a Grothendieck topos $\Sh({\cal C}, J)$ satisfies G\"{o}del-Dummett's law if and only if for any $J$-closed sieves $R$ and $S$ on an object $c\in {\cal C}$, the sieve $\{f:d\to c \textrm{ | } f^{\ast}(R)\subseteq f^{\ast}(S) \textrm{ or } f^{\ast}(S)\subseteq f^{\ast}(R)\}$ is $J$-covering. In particular, a presheaf topos $[{\cal C}^{\textrm{op}}, \Set]$ satisfies G\"{o}del-Dummett's law if and only if $\cal C$ satisfies the following property: for any arrows $f:b\to a$ and $g:c\to a$ with common codomain, either $f$ factors through $g$ or $g$ factors through $f$ (cf. also Proposition 3.1 \cite{morgan}). On the other hand, in \cite{morgan} Johnstone established the following result: for any topological space $X$, the topos $\Sh(X)$ satisfies G\"{o}del-Dummett logic if and only if every closed subspace of $X$ is extremally disconnected.
Let us unravel the property of the topos $\Sh({\cal D}, J_{\cal D})$ of coherent sheaves on a distributive lattice $\cal D$ to satisfy G\"{o}del-Dummett's law in terms of the site $({\cal D}, J_{\cal D})$. Given two sets $A:=\{a_{i}\leq x \textrm{ | } i\in I\}$ and $B:=\{b_{j}\leq x \textrm{ | } j\in J\}$ of elements of $\cal D$, we say that $A$ \emph{refines} $B$ if the sieve in $\cal D$ on $x$ generated by $A$ is contained in the sieve in $\cal D$ on $x$ generated by $B$, equivalently if for every $i\in I$ there exists $j\in J$ such that $a_{i}\leq b_{j}$. We define a set of elements $A:=\{a_{i}\leq x \textrm{ | } i\in I\}$ to be \emph{finitely closed} if the sieve on $x$ generated by it is $J_{\cal D}$-closed (i.e., for every $b\leq a$, if there is a finite subset $I'$ of $I$ such that $b=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I'}}(a_{i}\wedge b)$ then $b\leq a_{i}$ for some $i\in I$). In these terms, the condition that $\Sh({\cal D}, J_{\cal D})$ satisfies G\"{o}del-Dummett's law rephrases as follows: for any finitely closed sets of elements $A:=\{a_{i}\leq x \textrm{ | } i\in I\}$ and $B:=\{b_{j}\leq x \textrm{ | } j\in J\}$ in $\cal D$, there exists a finite collection $\{c_{k} \textrm{ | } k\in K\}$ of elements satisfying $x=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{k\in K}}c_{k}$ such that for any $k\in K$ either $A_{c_{k}}:=\{a_{i}\wedge c_{k} \leq c_{k} \textrm{ | } i\in I\}$ refines $B_{c_{k}}:=\{b_{j}\wedge c_{k}\leq c_{k} \textrm{ | } j\in J\}$ or $B_{c_{k}}$ refines $A_{c_{k}}$. We call a distributive lattice $\cal D$ satisfying this property a G\"{o}del-Dummett distributive lattice.
Combining together all the site characterizations discussed above, the usual method `toposes as bridges' applied to the invariant `to satisfy G\"{o}del-Dumm-\\ett's law' yields the following result.
\begin{theorem}
With the notation above, we have:
\begin{enumerate}[(i)]
\item If $\cal D$ is a distributive lattice and $X_{\cal D}$ is its associated Stone space then every closed subspace of $X_{\cal D}$ is extremally disconnected if and only if $\cal D$ is a G\"{o}del-Dummett distributive lattice;
\item If $({\cal M}, \leq)$ is a meet-semilattice and $X_{\cal M}$ is the ideal topological space associated to $\cal M$ then every closed subspace of $X_{\cal M}$ is extremally disconnected if and only if for every $p,q\in {\cal M}$ such that $p,q\leq r$ for some $r$ in $\cal M$, either $p\leq q$ or $q\leq p$ (i.e., if and only if ${\cal M}$ is a forest);
\item If $({\cal P}, \leq)$ is a preorder and $A_{\cal P}$ is the Alexandrov space associated to $\cal P$ then $A_{\cal P}$ has the property that every closed subspace of $X$ is extremally disconnected if and only if for every $p,q\in {\cal P}$ such that $r\leq p,q$ for some $r$ in $\cal P$, either $p\leq q$ or $q\leq p$ (i.e., if and only if ${\cal P}^{\textrm{op}}$ is a forest).
\end{enumerate}
\end{theorem}\qed
\section{Spaces of models of propositional theories}\label{spacesprop}
We have seen in Example \ref{exa}(f) that, for any geometric theory $\mathbb T$, the space of models of $\mathbb T$ in $\Set$ can be endowed with a natural topology, of logical nature, which is a particular case of the subterminal topology introduced in section \ref{subterminaltop}.
In the following sections we focus our attention on the logical spaces of propositional geometric theories. In section \ref{subs} we give an explicit description of the logical topology on a set of models of a propositional geometric theory. Then we characterize the frames of open sets of these logical topological spaces in terms of preordered structures presented by generators and relations. To this end, we introduce in section \ref{gensyn} an abstract notion of first-order mathematical theory, and a corresponding notion of syntactic category, which is seen in sections \ref{genrel} and \ref{clex} to subsume the known ones and to provide a uniform way for building structures presented by generators and relations for certain `ordered algebraic theories'.
\subsection{Subsets and propositional theories}\label{subs}
Propositional theories are particularly convenient means for describing subsets of a given set having particular properties, for example ideals of a commutative ring (cf. section \ref{zariski}), filters on a meet-semilattice (cf. section \ref{logical}), etc.
A propositional geometric theory $\mathbb T$ can be formally defined as a geometric theory over a signature $\Sigma_{\mathbb T}$ with no sorts (cf. Part D of \cite{El}). Thus $\Sigma_{\mathbb T}$ has no function symbols or constants, and consists only of a set of $0$-ary relation symbols ${\cal R}_{\mathbb T}$. Every symbol $R$ in ${\cal R}_{\mathbb T}$ gives therefore rise to an atomic formula, and every atomic formula is of this form. Note that $\Sigma_{\mathbb T}$-structures in $\Set$ can be identified with functions $f_{M}:{\cal R}_{\mathbb T}\to \{0,1\}$, or equivalently with subsets $f^{-1}(\{1\})\subseteq {\cal R}_{\mathbb T}$ of ${\cal R}_{\mathbb T}$.
Let $\mathbb T$ be a propositional geometric theory over a signature $\Sigma_{\mathbb T}$.
Recall from Example \ref{exa}(f) that the collection $X_{\mathbb T}$ of the (isomorphism classes of) models of $\mathbb T$ in $\Set$, endowed with the subterminal topology, has as open sets exactly those of the form $F_{\phi}=\{M \in X_{\mathbb T} \textrm{ | } M \vDash \phi \}$, where $\phi$ ranges among all the geometric sentences over $\Sigma_{\mathbb T}$. In particular, since (by Lemma D1.3.8(ii)) every geometric formula over $\Sigma_{\mathbb T}$ is provably equivalent to a disjunction of conjunctions of atomic formulae in the same context, the collection of subsets of the form $F_{R}=\{M \in X_{\mathbb T} \textrm{ | } M \vDash R\}$ for an atomic formula $R$ over $\Sigma_{\mathbb T}$ forms a subbasis of the topological space $X_{\mathbb T}$; that is, the open sets of $X_{\mathbb T}$ are exactly the unions of finite intersections of sets of the form $F_{R}$ for a $0$-ary relation symbol $R$ over $\Sigma_{\mathbb T}$. Clearly, if the conjunction of any two atomic formulae over $\Sigma_{\mathbb T}$ is $\mathbb T$-provably equivalent to an atomic formula then the sets of the form $F_{R}$ (for $R\in {\cal R}_{\mathbb T}$) form a basis of the topological space $X_{\mathbb T}$ (cf. section \ref{zariski} below for a concrete example of this situation).
Using the identification of $\Sigma_{\mathbb T}$-structures in $\Set$ with subsets of ${\cal R}_{\mathbb T}$, the space $X_{\mathbb T}$ acquires the following description: its points are the subsets of ${\cal R}_{\mathbb T}$ which correspond to models of $\mathbb T$, and its open sets are the unions of subsets of the form $\{L\subseteq {\cal R}_{\mathbb T} \textrm{ | } R_{1}, \ldots, R_{n} \in L\}$ for a finite number $R_{1}, \ldots, R_{n}$ of $0$-ary relation symbols over $\Sigma_{\mathbb T}$.
As an example, take $\mathbb T$ to be the empty theory over a signature $\Sigma$ consisting of a set $A$ of $0$-ary relation symbols.
The $\Sigma$-structures in $\Set$ can be identified with the subsets of $A$, and the resulting topology on the powerset $\mathscr{P}(A)$ is given by the collection of subsets which are unions of subsets of the form $\{L\subseteq A \textrm{ | } a_{1}, \ldots, a_{n} \in L\}$ for a finite number $a_{1}, \ldots, a_{n}$ of elements of $A$. This topology is an interesting one; in fact, as we shall see in section \ref{zariski}, it specializes, under a natural bijection, precisely to the Zariski topology on the prime spectrum of a commutative ring with unit, and the frame of open sets of the resulting topological space with underlying set $A$ can be characterized as the free frame on $A$ (by the results of sections \ref{genrel} and \ref{clex} below). We will refer below to this topology as to the \emph{elemental topology}.
Given a propositional geometric theory $\mathbb T$ over a signature $\Sigma_{\mathbb T}$, we have an embedding $\xi:{\cal R}_{\mathbb T} \to {\cal C}_{\mathbb T}$ of ${\cal R}_{\mathbb T}$ into the (underlying set of the) geometric syntactic category ${\cal C}_{\mathbb T}$ of $\mathbb T$, sending a symbol $R$ in ${\cal R}_{\mathbb T}$ to the corresponding atomic formula over $\Sigma_{\mathbb T}$. The operation of taking the inverse image of subsets under $\xi$ defines a bijection between the collection of completely prime filters on ${\cal C}_{\mathbb T}$ and the collection of the subsets of ${\cal R}_{\mathbb T}$ which correspond to models of $\mathbb T$; in fact, this bijection becomes a homeomorphism of topological spaces if we endow the two sets respectively with the topology on the set of points of a locale (cf. Example \ref{exa}(e)) and with the elemental topology.
Of course, analogous results hold if we replace the geometric syntactic category ${\cal C}_{\mathbb T}$ of $\mathbb T$ with a (cartesian) syntactic category appropriate to a given fragment of logic in which the theory $\mathbb T$ lies (for example, if we replace ${\cal C}_{\mathbb T}$ with the coherent syntactic category of $\mathbb T$ in case $\mathbb T$ is coherent), and we endow the resulting space of ($J$-)filters on the corresponding syntactic category with the subterminal topology as in Proposition \ref{mslattice} above.
Note that these results arise from the expression of a particular invariant, namely the subterminal topology, in terms of different representations of the locales (equivalently, of the toposes) involved.
\subsection{General first-order theories}\label{gensyn}
In this section we introduce a general notion of first-order mathematical theory and a corresponding notion of syntactic category, which encompass all the instances of the concepts given in chapter D1 of \cite{El}. We shall present these notions in a sketchy form, since a completely formal and detailed treatment of them would bring us far beyond the scope of this paper; anyway, we are confident that the interested reader will have no trouble in filling in the details by generalizing the framework of chapter D1 of \cite{El} according to our indications.
Let $\Sigma$ be a first-order signature. Let us suppose to have a collection $\cal S$ of symbols $s$ (to be thought of as generalized connectives) each of which equipped with an arity $ar(s)$ given by a cardinal number; we treat constants as $0$-ary function symbols, and we also allow the symbols in ${\cal S}$ to have arity $\infty$, i.e. to take as inputs sets of arbitrary cardinality. Starting with atomic formulae over $\Sigma$ we can inductively (and freely) build a collection $F_{({\cal S}, \Sigma)}$ of words, which we call ${\cal S}$-formulae over $\Sigma$, as follows: $F_{({\cal S}, \Sigma)}$ is the smallest set of words such that all the atomic formulae over $\Sigma$ belong to $F_{({\cal S}, \Sigma)}$, and for any $s\in {\cal S}$ and any family of words $\{F_{i} \textrm{ | } i\in ar(s)\}$ in $F_{({\cal S}, \Sigma)}$ the word $s(\{F_{i} \textrm{ | } i\in ar(s)\})$ belongs to $F_{({\cal S}, \Sigma)}$; the notion of free variable in a word of $F_{({\cal S},\Sigma)}$ is defined by simultaneous recursion as in classical cases, by defining the set of free variables of the word $s(F_{1}, \ldots F_{a(s)})$ equal to the union of the sets of free variables of the $F_{i}$. Of course, one could define a larger fragment of words by requiring $F_{({\cal S}, \Sigma)}$ to be closed also under existential or universal quantifications, but we shall not pursue these generalizations in this paper, our interest being primarily focused on the propositional fragments of these generalized logics. We say that a ${\cal S}$-theory over a signature $\Sigma$ is \emph{propositional} if $\Sigma$ has no sorts.
Given a generalized set of connectives $\cal S$ and a first-order signature $\Sigma$, we define a \emph{${\cal S}$-theory} over $\Sigma$ to be a collection of axioms and inference rules involving sequents of the form $\phi \vdash_{\vec{x}} \psi$ where $\phi$ and $\psi$ are formulae in $F_{(\cal S, \Sigma)}$ in the context $\vec{x}$ . Note that we have a notion of provability in a $\cal S$-theory of sequents involving ${\cal S}$-formulae over its signature.
Any generalized signature $({\cal S}, \Sigma)$ determines a class of categories in which the formulae in $F_{({\cal S}, \Sigma)}$ are interpretable. Specifically, given a category $\cal C$, we interpret the $0$-ary relation symbols over $\Sigma$ as objects $c$ of $\cal C$ such that for any object $c'\in {\cal C}$ there is at most one arrow $c'\to c$ in $\cal C$; we call these objects the \emph{subterminal objects} of $\cal C$ even if $\cal C$ does not have a terminal object (in fact, in the latter case, these objects are precisely the subterminal objects in the classical sense, that is the objects $c$ such that the unique arrow $c\to 1_{\cal C}$ is monic). If for any connective $s$ in $\cal S$, there is an operation $f_{s}$ of arity $ar(s)$ on the set of subterminal objects of $\cal C$ (if $ar(s)=0$, a subterminal object of $\cal C$), the interpretation of the formulae in $F_{({\cal S}, \Sigma)}$ with no free variables can be defined inductively induction on their structure by setting the interpretation of a formula $s(\{F_{i} \textrm{ | } i\in ar(s)\})$ equal to the result of applying the operation $f_{s}$ to the subterminals given by the interpretation of the formulae $F_{i}$. Provided that $\cal C$ has at least finite limits and if for any object $c$ of $\cal C$ and any connective $s$ in $\cal S$ there is an operation $f_{s}$ of arity $ar(s)$ on the set of subobjects of $c$ in $\cal C$, the formulae with at least a free variable can be similarly interpreted in a $\Sigma$-structure in $\cal C$ (as subobjects of (products of) the underlying object(s) of the structure) by interpreting the atomic formulae over $\Sigma$ as usual, and setting the interpretation of a formula $s(\{F_{i} \textrm{ | } i\in ar(s)\})$ equal to the result of applying the operation $f_{s}$ to the subobjects given by the interpretations of the formulae $F_{i}$. We shall call a category equipped with operations on its subobjects (if $\Sigma$ has at least one sort) or subterminals (if $\Sigma$ has no sorts) which interpret the connectives in $\cal S$ a \emph{$({\cal S}, \Sigma)$-category}.
Given a functor $F:{\cal C}\to {\cal D}$ between two $({\cal S}, \Sigma)$-categories, we say that $F$ is \emph{$({\cal S}, \Sigma)$-preserving} if either $\Sigma$ has no sorts and $F$ sends subterminal objects of $\cal C$ to subterminal objects of $\cal D$ and commutes with the operations on subterminals in the two categories which interpret the connectives in $\cal S$, or, if $\Sigma$ has at least one sort, $F$ preserves finite limits and commutes with the operations on subobjects in the two categories which interpret the connectives in $\cal S$. The category of $({\cal S}, \Sigma)$-preserving functors ${\cal C} \to {\cal D}$ and natural transformations between them will be denoted by $({\cal S}, \Sigma)\textrm{-}\textbf{Fun}({\cal C}, {\cal D})$.
Given a $\cal S$-theory $\mathbb T$ over a signature $\Sigma$, and a $({\cal S}, \Sigma)$-category $\cal C$, we define a \emph{model} of $\mathbb T$ in $\cal C$ to be a $\Sigma$-structure in $\cal C$ in which all the axioms of $\mathbb T$ are valid and the inference rules of $\mathbb T$ are sound (as in \cite{El}, we say that a sequent $\phi \vdash_{\vec{x}} \psi$ is valid in the structure if the interpretation of $\phi(\vec{x})$ factors, as a subobject or subterminal, through the interpretation of $\psi(\vec{x})$, while the soundness of inference rules has the usual meaning). We remark that, unlike in Definition D1.2.1 of \cite{El}, we also consider $\Sigma$-structures in categories lacking finite products or a terminal object, provided that $\Sigma$ has no sorts (in fact, as explained above, we interpret $0$-ary relation symbols as subterminal objects in the category); this extra-generality will be crucial for our purposes in constructing syntactic categories of propositional theories which lie in fragments which are weaker than Horn logic.
We define a homomorphism of models of a $\cal S$-theory $\mathbb T$ over a signature $\Sigma$ in a $({\cal S}, \Sigma)$-category $\cal C$ to be a homomorphism of the underlying $\Sigma$-structures. Clearly, models of $\mathbb T$ in $\cal C$ and homomorphisms between them form a category, which we denote by ${\mathbb T}\textrm{-mod}_{\cal S}({\cal C})$.
If $\mathbb T$ is a propositional ${\cal S}$-theory over a signature $\Sigma$ then a model of $\mathbb T$ in a $({\cal S}, \Sigma)$-preorder category $({\cal P}, \leq)$ can be identified with a function sending every $0$-ary relation symbol over $\Sigma$ to an element of $\cal P$ in such a way its extension $f:F_{({\cal S}, \Sigma)}\to {\cal P}$ to the set of all $\cal S$-formulae over $\Sigma$ satisfies the property that whenever $F \vdash_{[]} G$ is provable in $\mathbb T$, $f(F)\leq f(G)$ in $\cal P$. Note that, if $\cal S$ consists of the usual connectives $\wedge$ and $\vee$, that is, in the case of usual first-order logic, and $\cal P$ has a top element, this notion specializes to the classical one; indeed, the elements of such a preorder $\cal P$ can be clearly identified with the subobjects of the terminal object of $\cal P$, when the latter is regarded as a preorder category.
For a $\cal S$-theory $\mathbb T$ over a signature $\Sigma$, we define the \emph{syntactic category} ${\cal C}^{({\cal S}, \Sigma)}_{\mathbb T}$ of $\mathbb T$ as follows. The objects of ${\cal C}^{({\cal S}, \Sigma)}_{\mathbb T}$ are the $\mathbb T$-provable equivalence classes of $\cal S$-formulae-in-context over $\Sigma$ (considered up to `renaming' equivalence), and the arrows of ${\cal C}^{({\cal S}, \Sigma)}_{\mathbb T}$ are the $\mathbb T$-provable equivalence classes of $\cal S$-formulae over $\Sigma$ which are $\mathbb T$-provably functional from the domain to the codomain (this is a condition which ensures that their interpretation in any model of $\mathbb T$ in a $({\cal S}, \Sigma)$-theory is the graph of an arrow from the interpretation of the $\cal S$-formula in the domain to the interpretation of the $\cal S$-formula in the codomain, cf. p. 841 \cite{El} for the details). If the signature $\Sigma$ has at least one sort and the fragment $F_{({\cal S}, \Sigma)}$ contains cartesian logic, we have a model $M_{\mathbb T}$ of $\mathbb T$ in ${\cal C}^{({\cal S}, \Sigma)}_{\mathbb T}$, defined exactly as in p. 844 \cite{El}, in which the $\cal S$-sequents over $\Sigma$ which are valid coincide precisely with those which are provable in $\mathbb T$, and, as in \cite{El} (Theorem D1.4.7 etc.), we have a categorical equivalence
\[
({\cal S}, \Sigma)\textrm{-}\textbf{Fun}({\cal C}^{({\cal S}, \Sigma)}_{\mathbb T}, {\cal C})\simeq {\mathbb T}\textrm{-mod}_{\cal S}({\cal C})
\]
for any $({\cal S}, \Sigma)$-category $\cal C$.
Note that the syntactic category ${\cal C}^{({\cal S}, \Sigma)}_{\mathbb T}$ of a propositional $\cal S$-theory $\mathbb T$ over a signature $\Sigma$ can be identified with the poset obtained by equipping the set of provable-equivalence classes of $\cal S$-formulae over $\Sigma$ with the order given by the relation of provable entailment; the model $M_{\mathbb T}$ is defined as the function sending a $0$-ary relation symbol over $\Sigma$ to the $\mathbb T$-provable equivalence class of the corresponding atomic formula, and we have a categorical equivalence $({\cal S}, \Sigma)\textrm{-}\textbf{Fun}({\cal C}^{({\cal S}, \Sigma)}_{\mathbb T}, {\cal C})\simeq {\mathbb T}\textrm{-mod}_{\cal S}({\cal C})$ exactly as above. In particular, if the axioms of $\mathbb T$ are all bisequents then the preorder ${\cal C}^{({\cal S}, \Sigma)}_{\mathbb T}$ is a discrete category (since if $F \vdash_{[]} G$ is provable in $\mathbb T$ then necessarily $F$ is provably equivalent to $G$ in $\mathbb T$).
All the notions introduced in this section are meant to provide a general notion of fragment of first-order logic which encompasses the (quantifier-free versions of the) classical ones, including Horn, regular, coherent and geometric logic (cf. Part D of \cite{El}). The usual logical connectives, as well as the inference rules which govern them, appear as instances of an abstract notion of connective (respectively, of inference rule) within the unifying framework that we have developed. Specifically, quantifier-free Horn (resp. coherent, geometric, finitary first-order) can be seen as a $\cal S$-theory, where $\cal S$ consists of the binary connective $\wedge$ and the constant $\top$ (resp. of the binary connectives $\wedge$ and $\vee$ and the constants $\top$ and $\bot$, of the binary connective $\wedge$, the infinitary connective $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{}}$ and the constants $\top$ and $\bot$, of the binary connectives $\wedge$, $\vee$, $\neg$, $\imp$ and the constants $\top$ and $\bot$), and the corresponding notion of model of a regular (resp. coherent, geometric, finitary first-order logic) theory in a regular (resp. coherent, geometric, Heyting) category is subsumed by our notion of model in a $({\cal S}, \Sigma)$-category. In fact, if we regard a (quantifier-free) Horn (resp. coherent, geometric, first-order) theory as a $\cal S$-theory then our notion of syntactic category yields a category which is equivalent to the cartesian (resp. coherent, geometric, first-order) syntactic category of the theory, as defined in chapter D1 of \cite{El}.
We note that the main technical differences between our approach and that of \cite{El} consist in allowing arbitrary inference rules in the definition of a ($\cal S$-)theory, and in extending the notion of structure in a category in such a way as to allow the interpretation in categories of fragments of logic which are weaker than Horn logic (in particular, to allow the interpretation of propositional formulae in preorders which do not possess meets or a top element).
\subsection{Generators and relations}\label{genrel}
In this section we show that the syntactic categories of propositional $\cal S$-theories can be naturally realized as structures presented by generators and relations (in the sense of universal algebra).
The following definition will be central for our purposes.
\begin{definition}
\begin{enumerate}[(a)]
\item An \emph{infinitary Horn formula} over a first-order signature $\Sigma$ is a formula built from atomic formulae by only using possibly infinitary conjunctions (including the conjunction over the empty set, which we identify with the truth formula $\top$);
\item An \emph{infinitary Horn theory} over a signature $\Sigma$ is an infinitary first-order theory over $\Sigma$ whose axioms are \emph{infinitary Horn sequents}, that is sequents of the form $\phi \vdash_{\vec{x}} \psi$ where $\phi$ and $\psi$ are infinitary Horn formulae in the same context $\vec{x}$;
\item By \emph{infinitary Horn logic} we mean the logic of infinitary Horn theories, i.e. the fragment of infinitary first-order logic in which only the structural rules and the rules for infinitary conjunctions are present;
\item An \emph{ordered algebraic theory} is an infinitary Horn theory over a one-sorted signature $\Sigma$ consisting of a set of (possibly infinitary) function symbols and a binary relation symbol $\leq$, in which the following three Horn sequents, expressing the idea that $\leq$ is a partial order, are provable (in Horn logic):
\[
(\top \vdash_{x} x\leq x),
\]
\[
(x\leq y \vdash_{x,y} y\leq x),
\]
\[
((x\leq y) \wedge (y\leq z) \vdash_{x,y,z} x\leq z).
\]
\item Following \cite{freestr}, we say that an ordered algebraic theory $\mathbb T$ over a signature $\Sigma$ is \emph{small} if for every cardinal number $k$, there is only a \emph{set} of $k$-ary terms over $\Sigma$ up to $\mathbb T$-provable equivalence.
\end{enumerate}
\end{definition}
\begin{remarks}\label{rem}
\begin{enumerate}[(a)]
\item We can suppose, without loss of generality, all the axioms of an infinitary Horn theory to be of the form $F_{1} \vdash_{\vec{x}} F_{2}$, where $F_{1}(\vec{x})$ is a (possibly infinitary) conjunction of atomic formulae and $F_{2}(\vec{x})$ is an atomic formula;
\item Any (possibly infinitary) algebraic theory $\mathbb T$ over a signature $\Sigma$ can be considered as an ordered algebraic theory whose axioms are of the form $\top \vdash_{\vec{x}} t_{1}=t_{2}$ where $t_{1}$ and $t_{2}$ are terms over $\Sigma$, and in which the sequent
\[
(x\leq y \vdash_{x,y} x=y)
\]
is provable.
\end{enumerate}
\end{remarks}
Let us now introduce the notion of model of an ordered algebraic theory presented by generators and relations.
Let $M$ be a model (in $\Set$) of an ordered algebraic theory $\mathbb A$ over a signature $\Sigma_{\mathbb A}$. Given a set $A$ of \emph{generators}, let $\Sigma_{\mathbb A}^{A}$ be the signature having, in addition to the function symbols of $\Sigma_{\mathbb A}$, a constant symbol $C_{a}$ for each element $a\in A$. Note that, given a function $\xi:A\to M$ from $A$ to (the underlying set of) a $\Sigma_{\mathbb A}$-structure $M$, $M$ can be made into a $\Sigma_{\mathbb A}^{A}$-structure, where for any $a\in A$ the interpretation of the constant symbol $C_{a}$ is given by the element $\xi(a)\in M$. Given a set $R$ of \emph{relations}, i.e. infinitary Horn sequents of the form $\phi \vdash_{[]} \psi$ where $\phi$ and $\psi$ are closed infinitary Horn formulae over the signature $\Sigma_{\mathbb A}^{A}$, we denote by ${\mathbb T}_{A, R}$ the infinitary Horn theory over $\Sigma_{\mathbb A}$ having as axioms all the axioms of $\mathbb A$ and the sequents in $R$. Clearly, ${\mathbb T}_{A, R}$ is an ordered algebraic theory over the signature $\Sigma_{\mathbb A}^{A}$.
\begin{definition}
With the notation above, given a function $\xi:A\to M$ and a set $R$ of (infinitary) Horn sequents of the form $\phi \vdash_{[]} \psi$ where $\phi$ and $\psi$ are closed infinitary Horn formulae over the signature $\Sigma_{\mathbb A}^{A}$, we say that $M$ is presented via $\xi$ by the set of \emph{generators} $A$ subject to the \emph{relations} $R$, briefly that $M$ is presented by $(A, R)$, if $M$, regarded as a $\Sigma_{\mathbb A}^{A}$-structure as specified above, is an initial object in the category ${\mathbb T}_{A, R}\textrm{-mod}(\Set)$ of ${\mathbb T}_{A, R}$-models in $\Set$ and homomorphisms between them.
\end{definition}
Intuitively, a model $M$ of $\mathbb A$ in $\Set$ is presented by the set of generators $A$ via $\xi:A\to M$ subject to relations $R$ if all the `relations' in $R$ are satisfied in $M$ when evaluated in $M$ via $\xi$ and for any function $f:A\to N$ to a model $N$ of $\mathbb A$ in $\Set$ such that all the relations in $R$ are satisfied in $N$ when evaluated in $N$ via $f$ there exists a unique $\mathbb A$-model homomorphism $g:M\to N$ such that $g\circ \xi=f$.
The following result is a natural generalization of the classical construction of free models of algebraic theories.
\begin{theorem}\label{oat}
Let $\mathbb A$ be an small ordered algebraic theory, $A$ be a set and $R$ be a set of relations over $\Sigma_{\mathbb A}^{A}$ (in the sense specified above).
Let $M_{(A, R)}$ be the set of equivalence classes $[t]$ of closed terms $t$ over $\Sigma^{A}_{\mathbb A}$ with respect to the equivalence relation $E$ defined by
\[
(t_{1}, t_{2})\in E \textrm{ if and only if } \top \vdash_{[]} t_{1}=t_{2} \textrm{ is provable in } {\mathbb T}_{A, R}.
\]
If we define the interpretation in $M_{(A, R)}$ of the symbol $\leq$ as the relation
\begin{center}
$[t_{1}]\leq [t_{2}]$ if and only if $t_{1} \vdash_{[]} t_{2}$ is provable in ${\mathbb T}_{A, R}$.
\end{center}
and we set the interpretation in $M_{(A, R)}$ of a function symbol $f$ of arity $k$ over $\Sigma_{\mathbb A}$ equal to the function $M_{(A, R)}^{n}\to M_{(A, R)}$ sending any $k$-tuple $\{[t_{i}] \textrm{ | } i\in k\}$ in $M_{(A, R)}^{k}$ to the element $[f(\{[t_{i}] \textrm{ | } i\in k\})]\in M_{(A, R)}$, $M_{(A, R)}$ becomes a model of $\mathbb A$ which is presented by the set of generators $A$ and relations $R$ via the function $\xi:A\to M_{(A, R)}$ sending any $a\in A$ to $\xi(a):=[C_{a}]$.
\end{theorem}
\begin{proofs}
First, we note that, since $\mathbb A$ is small by hypothesis, $M_{(A, R)}$ is actually \emph{set}, rather than a proper class.
It is immediate to see, by using the deduction theorem in first-order logic, that for any formula $\phi(\vec{x})$ over $\Sigma_{\mathbb A}$ and any closed terms $\vec{t}$ over $\Sigma_{\mathbb A}^{A}$ (substitutable in place of $\vec{x}$), the formula $\phi$ is valid in $M_{(A, R)}$ when evaluated in $\vec{t}$ (regarded as a string of elements of $M_{(A, R)}$) if and only if the sequent $\top \vdash_{[]} \phi([\vec{t}\slash \vec{x}])$ is provable in the theory ${\mathbb T}_{A, R}$. From this it easily follows that all the axioms of $\mathbb A$ are valid in $M_{(A, R)}$. Clearly, $M_{(A, R)}$ is a model of ${\mathbb T}_{A, R}$ in $\Set$, and is presented by $A$ via $\xi$ subject to relations $R$, since for any ${\mathbb T}_{A, R}$-model $N$, regarded a function $f:A\to N$ to a $\mathbb A$-model $N$, there exists a unique $\mathbb A$-model homomorphism $g:M_{(A, R)} \to N$ such that $g\circ \xi=f$.
\end{proofs}
Let us now show how the construction of general syntactic categories provides an alternative way for building models of small ordered algebraic theories presented by generators and relations.
Let $\mathbb A$ be an ordered algebraic theory over a signature $\Sigma_{\mathbb A}$. For each function symbol $f$ over $\Sigma_{\mathbb A}$ of arity $k$ we define a corresponding connective on formulae $s_{f}$ of arity $k$. Given a set $A$, we define a propositional signature $\Sigma_{A}$ having exactly one $0$-ary relation symbol $R_{a}$ for each element $a\in A$. We denote by $F_{{\cal S}, A}$ the collection of $\cal S$-formulae over $\Sigma_{A}$, as defined in section \ref{gensyn}.
We have a natural bijection between the closed terms over $\Sigma_{\mathbb A}^{A}$ and the $\cal S$-formulae over $\Sigma_{A}$. Indeed, we can define a correspondence between them inductively, as follows: to a constant $C_{a}$ (for $a\in A$) over $\Sigma_{\mathbb A}^{A}$ we associate the atomic formula $R_{a}$, and to a term of the form $f(t_{1}, \ldots t_{n})$, for a function symbol $f$ over $\Sigma_{\mathbb A}$ of arity $k$ and a collection of terms $\{t_{i} \textrm{ | } i\in k\}$ corresponding to $\cal S$-formulae $\{F_{i} \textrm{ | } i\in k\}$, we associate the formula $s_{f}(\{F_{i} \textrm{ | } i\in k\})$. Clearly, this correspondence is a bijection; we denote by $F_{t}$ the $\cal S$-formula over $\Sigma_{A}$ corresponding to a closed term $t$ over $\Sigma_{\mathbb A}^{A}$.
Let us build a $\cal S$-theory ${\mathbb S}_{A, R}$ over $\Sigma_{A}$, as follows.
\begin{itemize}
\item First, we put as axioms of ${\mathbb S}_{A, R}$ the usual structural rules (as described in section D1.3 \cite{El}).
\item Second, to each axiom $\phi \vdash_{\vec{x}} \psi$ of the theory $\mathbb A$, where $\phi$ is either $\top$ or a conjunction of atomic formulae over $\Sigma_{\mathbb A}$ and $\psi$ is an atomic formula (cf. Remark \ref{rem2}(b)), we associate a scheme of inference rules of ${\mathbb S}_{A, R}$, each obtained by substituting arbitrary formulae $\vec{F}$ in $F_{\cal S}$ in place of the variables $\vec{x}$ in the following way. To each atomic formula $\chi$ over $\Sigma_{\mathbb A}$ we associate a sequent $\Gamma_{\chi}$ involving $\cal S$-formulae over $\Sigma_{A}$, obtained by putting the formulae $\vec{F}$ in place of the variables of the terms occurring in the atomic formulae, replacing the function symbols $f$ over $\Sigma_{\mathbb A}$ with the corresponding connectives $s_{f}$ in $\cal S$ and replacing the relation $\leq$ (resp. the equality $=$) between terms with the implication $\vdash$ (resp. the biimplication $\dashv \vdash$) (in case $\chi$ is $\top$ then we define $\Gamma_{\chi}$ to be a tautological sequent, as for example provided by the structural rules); we then define the inference rule corresponding to an axiom $\phi \vdash_{\vec{x}} \psi$ of $\mathbb A$ to be the rule having as premises all the sequents of the form $\Gamma_{\chi}$ where $\chi$ ranges among the atomic subformulae of $\phi$, and as conclusion the sequent $\Gamma_{\psi}$. So, for example, the sequent
\[
((x \cdot (y + z) \leq (x \cdot y) + (x \cdot z) \wedge (x=y)) \vdash_{x,y,z} x=y\cdot z)
\]
in the signature of rings corresponds to the inference rule whose premis-\\es are the sequents
\[
(s_{\cdot}(F_{1}, s_{+}(F_{2}, F_{3})) \vdash s_{+}(s_{\cdot}(F_{1}, F_{2}), s_{\cdot}(F_{1}, F_{2})))
\]
and
\[
(F_{1} \dashv\vdash F_{2})
\]
and whose conclusion is the sequent
\[
(F_{1} \dashv\vdash s_{\cdot}(F_{2}, F_{3})).
\]
\item Third, to each of the (possibly infinitary) Horn sequents over the signature $\Sigma_{\mathbb A}^{A}$ corresponding to the relations in $R$, we associate an inference rule obtained by the method of the last paragraph, the only difference being that the constants $C_{a}$ (rather than the variables) arising in the terms are replaced by the corresponding $\cal S$-formulae $R_{a}$ (for each $a \in A$).
\item Finally, we add inference rules asserting the invariance under provable equivalence of the connectives; that is, for each connective $s\in {\cal S}$ of arity $k$, we add an inference rule scheme which enables to derive the bisequent $(s_{f}(\{F_{i} \textrm{ | } i\in k\}) \dashv \vdash s_{f}(\{F_{i}' \textrm{ | } i\in k\}))$ from the set of bisequents $(F_{i} \dashv \vdash F_{i}')$ (for any $\cal S$-formulae $F_{i}$ and $F_{i}'$ (for $i\in k$) over $\Sigma_{A}$).
\end{itemize}
This completes the description of the $\cal S$-theory ${\mathbb S}_{A, R}$.
Let us now show that the (underlying poset of the) syntactic category ${\cal C}^{({\cal S}, \Sigma_{A})}_{{\mathbb S}_{A, R}}$ of the $\cal S$-theory ${\mathbb S}_{A, R}$ is a model of the theory ${\mathbb T}_{A, R}$ presented by $(A, R)$.
First, we note that, since our theory ${\mathbb S}$ is small by hypothesis, the underlying preorder of the category ${\cal C}^{({\cal S}, \Sigma_{A})}_{{\mathbb S}_{A, R}}$ is actually a \emph{set}.
Next, we observe that ${\cal C}_{{\mathbb S}_{A, R}}$ can be made into a ${\mathbb T}_{A, R}$-model structure, by defining the interpretation of any generator $a\in A$ to be the ${\mathbb S}_{A, R}$-provable equivalence class $[R_{a}]$ of the atomic formula $R_{a}$. One can easily verify by induction on the structure of $\cal S$-formulae over $\Sigma_{A}$ that for any term $t$ over $\Sigma_{\mathbb A}$, the interpretation of $t$ in ${\cal C}^{({\cal S}, \Sigma_{A})}_{{\mathbb S}_{A, R}}$ is equal to $[F_{t}]$; from this it easily follows, by definition of the theory ${\mathbb S}_{A, R}$, that all the axioms of ${\mathbb T}_{A, R}$ are satisfied in ${\cal C}_{{\mathbb S}_{A, R}}$.
To prove that, with the structure just defined, ${\cal C}^{({\cal S}, \Sigma_{A})}_{{\mathbb S}_{A, R}}$ is actually an initial object of ${\mathbb T}_{A, R}\textrm{-mod}(\Set)$, we appeal to the universal property of the syntactic category ${\cal C}^{({\cal S}, \Sigma_{A})}_{{\mathbb S}_{A, R}}$ (cf. section \ref{gensyn}): for any $({\cal S}, \Sigma_{A})$-category $\cal D$, the category ${{\mathbb S}_{A, R}}\textrm{-mod}_{\cal S}({\cal D})$ of $\cal S$-models of ${\mathbb S}_{A, R}$ in $\cal D$ is equivalent to the category $({\cal S}, \Sigma_{A})\textrm{-}\textbf{Fun}({\cal C}_{{\mathbb S}_{A, R}}, {\cal D})$ of $({\cal S}, \Sigma_{A})$-preserving functors from ${\cal C}^{({\cal S}, \Sigma_{A})}_{{\mathbb S}_{A, R}}$ to ${\cal D}$, via the equivalence sending a functor $F:{\cal C}_{{\mathbb S}_{A, R}} \to {\cal D}$ to the model of ${\mathbb S}_{A, R}$ in $\cal D$ in which a generator $a\in A$ is interpreted as the image $F([R_{a}])$ of the ${\mathbb S}_{A, R}$-provable $R_{a}$ of the atomic formula $R_{a}$.
Let us apply this property to poset categories $\cal D$. Note that in this case, all of our categories being posets, the equivalence
\[
{{\mathbb S}_{A, R}}\textrm{-mod}_{\cal S}({\cal D})\simeq ({\cal S}, \Sigma_{A})\textrm{-}\textbf{Fun}({\cal C}^{({\cal S}, \Sigma_{A})}_{{\mathbb S}_{A, R}}, {\cal D})
\]
is in fact an isomorphism of categories (equivalently, of posets).
Given a ${\mathbb T}_{A, R}$-model $N$ in $\Set$, we can regard $N$ as a poset $({\cal S}, \Sigma_{A})$-category ${\cal D}_{N}$ (in which the connectives in $\cal S$ are interpreted precisely as the operations on $N$ which interpret the function symbols over $\Sigma_{\mathbb A}$) containing a ${\mathbb S}_{A, R}$-model, obtained by interpreting any $0$-ary relation symbol $R_{a}$ (for $a\in A$) as the element of $N$ which interprets the constant $C_{a}$ over $\Sigma_{\mathbb A}^{A}$ in $N$ regarded as a ${\mathbb T}_{A, R}$-model. Now, the ${\mathbb T}_{A, R}$-model homomorphisms ${\cal C}_{{\mathbb S}_{A, R}} \to N$ can be identified exactly with the $({\cal S}, \Sigma_{A})$-preserving functors ${\cal C}_{{\mathbb S}_{A, R}} \to {\cal D}_{N}$ which send any $[R_{a}]$ (for $a\in A$) to the interpretation of $C_{a}$ in $N$ (regarded as a ${\mathbb T}_{A, R}$-model); but these correspond, via the isomorphism $({\cal S}, \Sigma_{A})\textrm{-}\textbf{Fun}({\cal C}^{({\cal S}, \Sigma_{A})}_{{\mathbb S}_{A, R}}, {\cal D})\cong {\mathbb T}\textrm{-mod}_{\cal S}({\cal D})$, precisely to the ${\mathbb S}_{A, R}$-models in ${\cal D}_{N}$ in which the interpretation of the formula $R_{a}$ (for any $a\in A$) is equal to the interpretation of $C_{a}$ in $N$, and there is obviously just one such model, namely $N$. This proves our claim.
Summarizing, we have obtained the following result.
\begin{theorem}
Let $\mathbb A$ be an small ordered algebraic theory and $(A, R)$ a set of generators and relations for it. With the above notation, the poset ${\cal C}_{{\mathbb S}_{A, R}}$ is `the' $\mathbb A$-model presented by the set of generators $A$ and relations $R$ via the function $\xi:A\to {\cal C}_{{\mathbb S}_{A, R}}$ sending any $a\in A$ to the ${{\mathbb S}_{A, R}}$-provable equivalence class $[R_{a}]$ of the formula $R_{a}$.
In particular, models of small ordered algebraic theories presented by arbitrary generators and relations always exist.
\end{theorem}\qed
In fact, in light of the identification between the closed terms over $\Sigma_{\mathbb A}^{A}$ and the $\cal S$-formulae over $\Sigma_{A}$ observed above, the construction given by this theorem is isomorphic to that of Theorem \ref{oat}.
From the proof given above it is clear that the poset ${\cal C}_{{\mathbb S}_{A, R}}$ does not only satisfy the universal property of the $\mathbb A$-model presented by $(A, R)$ with respect to the \emph{$\mathbb A$-models}, but also with respect to every structure $N$ over the signature of $\mathbb A$ equipped with interpretations of the constant symbols $C_{a}$ (for $a\in A$) in such a way that the substructure of $N$ generated by these interpretations is a model of $\mathbb A$.
We have used the syntactic categories of propositional $\cal S$-theories to build models presented by generators and relations of (small) ordered algebraic theories; the question thus naturally arises whether every such syntactic category is of this form, that is if it can be regarded, in a natural way, as a model of an ordered algebraic theory presented by generators and relations. The answer to this question is positive. Given a propositional $\cal S$-theory $\mathbb T$, we take one function symbol of arity $k$ for each connective in $\cal S$ of arity $k$; this, together with the binary relation symbol $\leq$, defines a signature $\Sigma$. Take $\mathbb H$ to be the theory over $\Sigma$ having as axioms the infinitary Horn sequents over $\Sigma$ whose associated inference rule schemes (via the method above) are provably valid in $\mathbb T$. Define $A$ to be equal to the set of $0$-ary relation symbols of the signature of $\cal S$, and the set $R$ of relations involving elements in $A$ as the set of Horn sequents over $\Sigma_{\mathbb A}^{A}$ whose associated inference rule scheme (via the method above) is provably valid in $\mathbb T$. It is then clear that the $\cal S$-syntactic category of $\mathbb T$ is precisely `the' $\mathbb H$-model presented by the set of generators $A$ subject to relations $R$.
\subsection{Classical examples}\label{clex}
\begin{itemize}
\item \emph{Meet-semilattices}
The concept of meet-semilattice can be formalized as a small ordered algebraic theory, as follows. Take $\Sigma_{\mathbb A}$ to be the signature consisting of a binary relation symbol $\leq$, one constant $1$ and one binary function symbol $\wedge$. Consider the following Horn axioms over $\Sigma$:
\[
(\top \vdash_{x} (x \leq 1)),
\]
\[
(x\leq (y\wedge z) \dashv\vdash_{x,y,z} (x\leq y) \wedge (x\leq z)).
\]
Clearly, the Horn theory $\mathbb A$ over $\Sigma_{\mathbb A}$ obtained by adding to these axioms the sequents which express the fact that $\leq$ is a partial order has the property that its models in $\Set$ and homomorphisms between them can be identified with the meet-semilattices and meet-semilattice homomorphisms between them.
It is immediate to see that, for any choice of generators and relations $(A, R)$ for $\mathbb A$ such that the relations in $R$ are all of the form $\top \vdash t_{1} T t_{2}$ where $t_{1}$ and $t_{2}$ are closed terms over $\Sigma_{\mathbb A}^{A}$ and $T$ is either the relation $\leq$ or the equality relation $=$, the $\cal S$-theory $\mathbb L$ associated to the theory $\mathbb A$ via the method of the last section is precisely the Horn propositional theory $\mathbb M$ over the signature $\Gamma$ consisting of one $0$-ary relation symbol for each of the elements of $A$ (note that we do not have to add the constant $\top$ since it is already present in Horn logic), with axioms (in addition to those of Horn logic):
\[
(F_{t_{1}} \vdash_{[]} F_{t_{2}})
\]
(respectively,
\[
(F_{t_{1}} \dashv \vdash_{[]} F_{t_{2}}))
\]
for each relation in $R$ of the form $\top \vdash t_{1} \leq t_{2}$ (resp. of the form $\top \vdash t_{1}=t_{2}$), where $t_{1}$ and $t_{2}$ are closed terms over $\Sigma_{\mathbb A}^{A}$ and $F_{t_{1}}$ and $F_{t_{2}}$ are the atomic formulae over $\Gamma$ corresponding to them via the method of section \ref{genrel}. In particular, the syntactic category of the $\cal S$-theory $\mathbb L$ is equivalent to the cartesian syntactic category of the cartesian theory $\mathbb M$ (as defined in section D1 of \cite{El}).
\item \emph{Distributive lattices}
The concept of distributive lattice can also be formalized as a small ordered algebraic theory, as follows. Let $\Sigma_{\mathbb A}$ be the signature consisting of a binary relation symbol $\leq$, two constants $0, 1$ and two binary function symbols $\wedge, \vee$. Consider the following Horn axioms over $\Sigma_{\mathbb A}$:
\[
(\top \vdash_{x} (x \leq 1)),
\]
\[
(\top \vdash_{x} (0 \leq x)),
\]
\[
(x\leq (y\wedge z) \dashv\vdash_{x,y,z} (x\leq y) \wedge (x\leq z)),
\]
\[
((x\vee y) \leq z \dashv\vdash_{x,y,z} (x\leq z) \wedge (y\leq z)),
\]
\[
(\top \vdash_{x,y,z} (x\wedge (y\vee z) = (x\wedge y) \wedge (x\wedge z))).
\]
These axioms, in addition to the sequents which express the fact that $\leq$ is a partial order, define a small ordered algebraic theory $\mathbb A$ whose models in $\Set$ and homomorphisms between them can be identified with the distributive lattices and homomorphisms between them.
As with meet-semilattices, it is immediate to see that, for any choice of generators and relations $(A, R)$ for $\mathbb A$ such that the relations in $R$ are all of the form $\top \vdash t_{1} T t_{2}$ where $t_{1}$ and $t_{2}$ are closed terms over $\Sigma_{\mathbb A}^{A}$ and $T$ is either the relation $\leq$ or the equality relation $=$, the $\cal S$-theory $\mathbb L$ associated to the theory $\mathbb A$ via the method of the last section can be identified with the coherent propositional theory $\mathbb M$ over the signature $\Gamma$ consisting of one $0$-ary relation symbol for each of the elements of $A$ (note that we do not have to add the constants $\top$ and $\bot$ since they are already present in coherent logic), having as axioms (in addition to those of coherent logic) are the following the sequents of the form
\[
(F_{t_{1}} \vdash_{[]} F_{t_{2}})
\]
(respectively,
\[
(F_{t_{1}} \dashv \vdash_{[]} F_{t_{2}}))
\]
for each relation in $R$ of the form $\top \vdash t_{1} \leq t_{2}$ (resp. of the form $\top \vdash t_{1}=t_{2}$), where $t_{1}$ and $t_{2}$ are closed terms over $\Sigma_{\mathbb A}^{A}$ and $F_{t_{1}}$ and $F_{t_{2}}$ are the atomic formulae over $\Gamma$ corresponding to them via the method of section \ref{genrel}. In particular, the syntactic category of the $\cal S$-theory $\mathbb L$ is equivalent to the coherent syntactic category of the coherent theory $\mathbb M$.
\item \emph{Frames}
The notion of frame is the infinitary analogue of that of distributive lattice. Since our definition of ordered algebraic theory allows infinitary function symbols and infinitary conjunctions of atomic formulae, the arguments given above for distributive lattices straightforwardly extend to frames. Specifically, the small ordered algebraic theory formalizing the concept of frame is obtained by taking a signature $\Sigma_{\mathbb A}$ consisting of a binary relation symbol $\leq$, two constants $0, 1$, one binary function symbol $\wedge$ and an infinitary function symbol $D$. Consider the following Horn axioms over $\Sigma$:
\[
(\top \vdash_{x} (x \leq 1)),
\]
\[
(\top \vdash_{x} (0 \leq x)),
\]
\[
(x\leq (y\wedge z) \dashv\vdash_{x,y,z} (x\leq y) \wedge (x\leq z)),
\]
\[
(D(x_{i} \textrm{ | } i\in I) \leq z \dashv\vdash_{x_{i}, z} \mathbin{\mathop{\textrm{\huge $\wedge$}}\limits_{i\in I}}(x_{i}\leq z)),
\]
\[
(\top \vdash_{x_{i},y} (D(x_{i} \textrm{ | } i\in I) \wedge y = D(x_{i}\wedge y \textrm{ | } i\in I)).
\]
These axioms, in addition to the sequents which express the fact that $\leq$ is a partial order, define a small ordered algebraic theory $\mathbb A$ whose models in $\Set$ and homomorphisms between them can be identified with the frames and frame homomorphisms between them.
As in the case of distributive lattices, for any choice of generators and relations $(A, R)$ for $\mathbb A$ such that the relations in $R$ are all of the form $\top \vdash t_{1} T t_{2}$ where $t_{1}$ and $t_{2}$ are closed terms over $\Sigma_{\mathbb A}^{A}$ and $T$ is either the relation $\leq$ or the equality relation $=$, the $\cal S$-theory $\mathbb L$ associated to the theory $\mathbb T$ via the method of section \ref{genrel} can be identified with the geometric propositional theory $\mathbb M$ over the signature $\Gamma$ consisting of one $0$-ary relation symbol for each of the elements of $A$, and whose axioms (in addition to those of geometric logic) are the following:
\[
(F_{t_{1}} \vdash_{[]} F_{t_{2}})
\]
(respectively,
\[
(F_{t_{1}} \dashv \vdash_{[]} F_{t_{2}})
\]
for each relation in $R$ of the form $\top \vdash t_{1} \leq t_{2}$ (resp. of the form $\top \vdash t_{1}=t_{2}$), where $t_{1}$ and $t_{2}$ are closed terms over $\Sigma_{\mathbb A}^{A}$ and $F_{t_{1}}$ and $F_{t_{2}}$ are the atomic formulae over $\Gamma$ corresponding to them via the method of section \ref{genrel}.
\end{itemize}
\subsection{Structures presented by generators and relations}\label{concreteness}
We have seen that models of small ordered algebraic theories presented by generators and relations can always be constructed as preordered syntactic categories of generalized propositional theories.
A problem which frequently arises in practice is that of finding concrete descriptions for poset structures presented by generators and relations. The techniques elaborated in this paper, combined with the philosophy `toposes as bridges' of \cite{OC10}, provide flexible and effective tools for addressing this kind of problems. The key idea is to equip such a poset structure $\cal P$ with a Grothendieck topology $J$ such that $\cal P$ can be recovered (up to isomorphism) from the topos $\Sh({\cal P}, J)$ by means of a topos-theoretic invariant $U$ which has a `natural behaviour' with respect to sites (as we did for example in section \ref{charinv}); indeed, under these hypotheses, any other site of definition $({\cal C}, K)$ of the topos $\Sh({\cal P}, J)$ leads to a different representation for $\cal P$ as the poset of subterminals in the topos $\Sh({\cal C}, K)$ which satisfy the invariant $U$. For example, given a commutative ring with unit $(A, +, \cdot, 0_{A}, 1_{A})$, the distributive lattice generated by symbols $D(a)$, $a \in A$, subject to the relations $D(1_{A})=1_{L(A)}$, $D(a\cdot b) = D(a) \wedge D(b)$, $D(0_{A})=0_{L(A)}$, and $D(a+ b) \leq D(a) \vee D(b)$ can be characterized (up to isomorphism) as the lattice of compact elements of the frame of open sets of the prime spectrum of $A$ endowed with the Zariski topology (cf. section \ref{zariski} below).
The notion of subterminal topology can be profitably used to give `topological descriptions' of poset structures $\cal P$ (that is, descriptions in terms of frames of open sets of topological spaces) whenever the corresponding toposes $\Sh({\cal P}, J)$ have enough points. Indeed, if $U$ is a topos-theoretic invariant of families of subterminals in a topos which is $({\cal P}, J)$-adequate (in the sense of Definition \ref{ade}), so that $\cal P$ can be recovered from $\Sh({\cal P}, J)$ as the poset of $U$-compact subterminals of $\Sh({\cal P}, J)$ then, provided that $\Sh({\cal P}, J)$ has enough points, $\cal C$ is isomorphic (as a poset) to the poset of $U$-compact open sets of any topological space obtained by endowing a separating set of points of $\Sh({\cal P}, J)$ with the subterminal topology (cf. section \ref{subterminaltop}).
For example, the free frame $F(A)$ on a set $A$ can be described (up to isomorphism) as the frame of open sets of the topological space obtained by endowing the powerset of $A$ with the elemental topology (cf. section \ref{subs}). Similarly, the free meet-semilattice ${\cal M}(A)$ on a set $A$ can be regarded as the cartesian syntactic category of the empty propositional theory ${\mathbb T}_{A}$ over the signature having exactly one $0$-ary relation symbol for each element $a$ of $A$, and hence it can be described, by Corollary \ref{cor}(iii), as the poset of supercompact open sets of the topological space obtained by equipping the powerset of $A$ with the elemental topology (cf. section \ref{subs}).
In connection with the method described above, it is worth to remark that the theory of classifying toposes can be profitably exploited to obtain different representations for a given topos, regarded as a classifying topos of a geometric theory. For example, by regarding the theory ${\mathbb T}_{A}$ defined above as a cartesian theory, we immediately obtain a description of its classifying topos as the presheaf topos $[\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set), \Set]$, where $\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)$ is the category of (representatives of isomorphism classes of) finitely presentable models of ${\mathbb T}_{A}$ in $\Set$, from which it follows that the free meet-semilattice ${\cal M}(A)$ on a set $A$ is isomorphic to the opposite of the poset $\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)$; this poset can be clearly identified with the poset ${\mathscr{P}}_{fin}(A)$ of finite subsets of $A$ (with the subset-inclusion ordering), and from this description we can recover the well-known characterization of ${\cal M}(A)$ as ${\mathscr{P}}_{fin}(A)^{\textrm{op}}$. Similarly, one can recover the classical description of the free frame $F(A)$ on a set $A$ by regarding $F(A)$ as the geometric syntactic category of the theory ${\mathbb T}_{A}$; indeed, the latter can be identified with the frame of subterminals of the classifying topos $\Set[{\mathbb T}_{A}]\simeq [{\mathscr{P}}_{fin}(A), \Set]$ of ${\mathbb T}_{A}$, from which it follows that the free frame on a set $A$ is (isomorphic to) the frame of upper sets on ${\mathscr{P}}_{fin}(A)$.
\subsection{The free frame on a complete join-semilattice}\label{freecomjoin}
A particularly significant application of the general method introduced in the last section for explicitly describing structures presented by generators and relations is the solution to the problem of describing the free frame $L(A)$ on a complete join-semilattice $A$.
If we denote by $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ the join of a family of elements $\{a_{i} \textrm{ | } i\in I\}$ in a complete join-semilattice $A$ then, by the results of sections \ref{genrel} and \ref{clex}, $L(A)$ can be identified with the geometric syntactic category ${\cal C}_{{\mathbb L}_{A}}$ of the geometric theory ${\mathbb L}_{A}$ defined as follows: the signature $\Sigma_{A}$ of ${\mathbb L}_{A}$ consists of one $0$-ary relation symbol $F_{a}$ for each element $a\in A$, and the axioms of ${\mathbb L}_{A}$ are, besides those of geometric logic, all the sequents over $\Sigma_{A}$ of the form
\[
\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}F_{a_{i}} \dashv \vdash F_{a}
\]
for any family of elements $\{a_{i} \textrm{ | } i\in I\}$ in $A$ such that $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $A$.
Now, the classifying topos of ${\mathbb L}_{A}$ can be realized as a subtopos
\[
\Sh(\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)^{\textrm{op}}, J)
\]
of the classifying topos of the (cartesian) empty theory ${\mathbb T}_{A}$ over $\Sigma_{A}$, which, as we have seen above, is the presheaf topos
\[
[{\cal C}_{{\mathbb T}_{A}}^{\textrm{op}}, \Set]\simeq [\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set), \Set]\simeq [{\mathscr{P}}_{fin}(A), \Set],
\]
where ${\cal C}_{{\mathbb T}_{A}}$ is the cartesian syntactic category of ${\mathbb T}_{A}$.
The Grothendieck topology $J$ on $\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)^{\textrm{op}}\simeq {\cal C}_{{\mathbb T}_{A}}$ is defined as follows: denoted by $[F_{a}]$ the object in ${\cal C}_{{\mathbb T}_{A}}$ corresponding to the atomic formula $F_{a}$ (for an element $a\in A$), $J$ is generated by the families of sieves which contain sinks of the form
\[
\{[F_{a_{i}} \wedge F_{a}] \leq [F_{a}] \textrm{ | } i\in I\}
\]
for some family of elements $\{a_{i} \textrm{ | } i\in I\}$ in $A$ such that $a=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}$ in $A$.
Under the equivalence ${\cal C}_{{\mathbb T}_{A}} \simeq \textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)^{\textrm{op}}$, the closure under pullbacks of this family of sieves corresponds to the coverage $K$ on the category $\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)^{\textrm{op}}\simeq {\mathscr{P}}_{fin}(A)^{\textrm{op}}$ described as follows: the $K$-covering cosieves on ${\mathscr{P}}_{fin}(A)$ are those generated by families of the form
\[
\{U \cup \{a_{i}\} \supseteq U \textrm{ | } i\in I\}
\]
for any finite subset $U$ of $A$ and any family $\{a_{i} \textrm{ | } i\in I\}$ of elements of $A$ such that $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}\in U$.
Now, the geometric syntactic category ${\cal C}_{{\mathbb L}_{A}}$ of the theory ${\mathbb L}_{A}$ is isomorphic, as a poset, to the frame of subterminals of the classifying topos $\Set[{\mathbb L}_{A}]$ of ${\mathbb L}_{A}$ and hence, from the representation
\[
\Set[{\mathbb L}_{A}] \simeq \Sh(\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)^{\textrm{op}}, J)
\]
it follows that ${\cal C}_{{\mathbb L}_{A}}$ is isomorphic to the frame $Id_{J}(\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)^{\textrm{op}})$ of $J$-ideals on $\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)^{\textrm{op}}$. Note that this argument represents an application of the philosophy `toposes as bridges' of \cite{OC10}.
Since $J$ is generated by the coverage $K$ then the $J$-ideals on the category $\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)^{\textrm{op}}$ coincide exactly with the $K$-ideals on $\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)^{\textrm{op}}$. This can be proved directly (by showing that, for any $K$-ideal $I$ the collection of all the sieves $S$ on $\cal C$ such that the implication `$(dom(f)\in I$ for all $f\in S)$ entails $cod(f)\in I$' holds defines a Grothendieck topology which contains $K$) or, more conceptually, can be deduced as a consequence of the general well-known fact (cf. Proposition C2.1.9 \cite{El}) that if $J$ is the Grothendieck topology generated by a coverage $K$ on a small category $\cal C$ (i.e., the smallest Grothendieck topology $J$ on $\cal C$ such that every $K$-covering family generates a $J$-covering sieve) then the toposes $\Sh({\cal C}, J)$ and $\Sh({\cal C}, K)$ are equal, by observing that for any site $({\cal E}, W)$, the $W$-ideals on ${\cal E}$ can be (canonically) identified with the subterminals of the topos $\Sh({\cal E}, W)$.
The $K$-ideals on $\textrm{f.p.} {\mathbb T}_{A}\textrm{-mod}(\Set)^{\textrm{op}}\simeq {\mathscr{P}}_{fin}(A)^{\textrm{op}}$ can be identified precisely with the upward closed subsets ${\cal I}$ of $\mathscr{P}_{fin}(A)$ such that for every $U\in \mathscr{P}_{fin}(A)$ and every family of elements $\{a_{i} \textrm{ | } i\in I\}$ of $A$ such that $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}\in U$, if $U\cup \{a_{i}\}\in {\cal I}$ for all $i\in I$ then $U\in {\cal I}$. We thus conclude that $L(A)$ is given by the collection of all such subsets, endowed with the natural subset-inclusion ordering. The universal complete join-semilattice homomorphism $\eta_{A}:A\to L(A)$ is the map sending an element $a\in A$ to the $J$-closure of the set $I_{a}$ of all the finite subsets of $A$ which contain the element $a$. We can easily see that $I_{a}$ is $J$-closed (equivalently, $K$-closed); indeed, given $U\in \mathscr{P}_{fin}(A)$ and any family of elements $\{a_{i} \textrm{ | } i\in I\}$ of $A$ such that $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}\in U$, $a\in U\cup \{a_{i}\}$ for all $i\in I$ implies $a\in U$. The universal map $\eta_{A}:A\to L(A)$ thus sends any element $a\in A$ to the set $I_{a}$, and is therefore injective; this solves an open problem of Resende and Vickers (cf. the remarks preceding Proposition 3.7 \cite{RV}). Summarizing, we have the following result.
\begin{theorem}\label{solution}
The free frame on a complete join semilattice $A$ is (up to isomorphism) the set $L(A)$ of upward closed (with respect to the subset-inclusion ordering on $\mathscr{P}_{fin}(A)$) subsets ${\cal I}$ of $\mathscr{P}_{fin}(A)$ with the property that for every $U\in \mathscr{P}_{fin}(A)$ and every family of elements $\{a_{i} \textrm{ | } i\in I\}$ of $A$ such that $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}a_{i}\in U$, if $U\cup \{a_{i}\}\in {\cal I}$ for all $i\in I$ then $U\in {\cal I}$, endowed with the subset-inclusion ordering, while the universal complete join-semilattice homomorphism $\eta_{A}:A\to L(A)$ is the map sending any element $a\in A$ to the set $I_{a}\in L(A)$ of all the finite subsets of $A$ which contain the element $a$.
\end{theorem}\qed
We can describe the frame operations on $L(A)$ explicitly as follows: given any $I_{1}, I_{2}\in L(A)$, the meet $I_{1}\wedge I_{2}$ in $L(A)$ is simply the set-theoretic intersection $I_{1}\cap I_{2}$, while the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{j\in J}}I_{j}$ in $L(A)$ of a family $\{I_{j} \textrm{ | } j\in J\}$ of elements $I_{j}\in L(A)$ is equal to the $J$-closure of the set-theoretic union $\mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{j\in J}}I_{j}$, i.e. to the collection of all the finite subsets $U$ of $A$ with the property that there exists a family $\{a_{h} \textrm{ | } h\in H\}$ of elements of $A$ such that $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{h\in H}}a_{h}\in U$ and for every $h\in H$, $U\cup \{a_{h}\}\in I_{j}$ for some $j \in J$.
Given a frame $L$, and a homomorphism $f:A\to L$ of complete join-semilattices, the unique frame homomorphism $g:L(A)\to L$ such that $g\circ \eta_{A}=f$ is defined by the formula:
\[
g(I)=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{U\in I}} (\mathbin{\mathop{\textrm{\huge $\wedge$}}\limits_{a\in U}} f(a)),
\]
for any $I\in L(A)$.
Now that we have obtained a solution to our problem, it is natural to look for an elementary proof of the result which does not involve topos-theoretic concepts. Such a proof clearly exists, but we do not bother to write it down because, although being elementary, it involves a good amount of tedious and intricate verifications; on the other hand, we are confident that the interested reader will have no trouble in carrying out the necessary calculations by himself or herself. Rather, we emphasize that this is just an example of what can be achieved by using our methods; many similar problems can be solved by applying the same techniques, and the reader is invited to try them out in the context of his or her questions of interest.
\begin{remark}
Let $\textbf{CJSLat}$ denote the category having as objects the complete join-semilattices and as arrows the maps between them which preserve arbitrary joins, and let $\textbf{Frm}$ be the category of frames. The map from complete join-semilattices to frames which sends a complete join-semilattice $A$ to the complete frame $L(A)$ on it can be made into a functor $L:\textbf{CJSLat}\to \textbf{Frm}$ which is left adjoint to the inclusion functor $\textbf{Frm} \hookrightarrow \textbf{CJSLat}$. It is easy to see that the functor $L$ sends a complete join-semilattice $A$ to the frame $L(A)$ and an arrow $f:A\to B$ in $\textbf{CJSLat}$ to the unique frame homomorphism $L(f):L(A)\to L(B)$ such that $\xi\circ \eta_{A}=\eta_{B}\circ f$; concretely, $L(f)$ sends a subset $\cal I$ in $\mathscr{P}_{fin}(A)$ to the smallest upward closed subsets $\cal J$ of $\mathscr{P}_{fin}(B)$ (with respect to the subset-inclusion ordering on $\mathscr{P}_{fin}(B)$) which contains $f({\cal I})$ and has the property that for every $U\in \mathscr{P}_{fin}(B)$ and every family of elements $\{b_{i} \textrm{ | } i\in I\}$ of $B$ such that $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{i\in I}}b_{i}\in U$, if $U\cup \{b_{i}\}\in {\cal J}$ for all $i\in I$ then $U\in {\cal J}$.
\end{remark}
We conclude this section by presenting a couple of other examples. First, we note that the finitary analogue of the construction (that is, the version of the construction obtained by requiring all the sets $I$ to be finite) provides a way for building the free frame $F(A)$ on a join-semilattice $A$. Note that the free distributive lattice on a join-semilattice $A$ can be characterized as the lattice of compact elements of $F(A)$ (cf. Corollary \ref{cor}(i)).
The free frame $F(A)$ on a join-semilattice $(A, \vee)$ can also be described in topological terms by using the method of section \ref{concreteness}. Specifically, we get that $F(A)$ is isomorphic to the frame of open sets of the space obtained by endowing the collection of all the subsets $U$ of $A$ such that $0\notin U$ and for any pair of elements $a,b\in A$, $a\vee b \in U$ if and only if $a\in U$ or $b\in U$ with the elemental topology. Note that this topological space is homeomorphic to the space obtained by endowing the set of subsets $V$ of $A$ such that $0\in V$ and for any pair of elements $a,b\in A$, $a\vee b \in V$ if and only if $a\notin V$ and $b\notin V$ with the topology of which the sets of the form $D_{a}:=\{V \textrm{ | } a\notin A \}$ for $a\in A$ form a basis (notice the similarity between this topology and the Zariski topology, cf. section \ref{zariski} below).
\subsection{The Zariski topology}\label{zariski}
In section V.III of \cite{stone}, two different ways of building the Zariski spectrum of a commutative ring with unit are presented. In this section, we provide a topos-theoretic interpretation of these two constructions as a Morita-equivalence, in the form of two different sites of definition for the same topos. This is particularly relevant in relation to the philosophy `toposes as bridges' of \cite{OC10}; in fact, as we shall see below, one can effectively transfer information between the different sites of definition of the topos through topos-theoretic invariants, according to the indications given in \cite{OC10}.
Let us begin by discussing the constructions given in \cite{stone}.
Let $A$ be an arbitrary commutative ring with unit. Let $S(A)$ be the quotient of the multiplicative monoid of $A$ by the smallest monoid congruence $\simeq$ which identifies $a$ with $a^{2}$ for each $a \in A$; we denote by $\pi_{S}:A\to S(A)$ the canonical quotient map. Note that $S(A)$ is a commutative monoid in which every element is idempotent.
For any monoid $M$ in which every element is idempotent, we can define a partial order relation $\leq_{M}$ on $M$ as follows: for any $a,b\in M$, $a \leq_{M} b$ if and only if $a\cdot b=a$; in fact, for any $a,b\in M$, $a\cdot b$ is the greatest lower bound of $a$ and $b$ with respect to $\leq_{M}$, and hence $(M, \leq_{M})$ is a meet-semilattice.
So $S(A)$ is a meet-semilattice and hence, regarded as a category, it is a cartesian category. We can define a coverage $C$ on $S(A)$ as follows: $\emptyset \in C(\pi_{S}(0_{A}))$, and the $C$-covering sieves on an object $x\in S(A)$ are those which contain finite families of the form $\pi_{S}(a_{i})\to \pi_{S}(a)$ for $i=1, \ldots, n$, where $\pi_{S}(a_{i}) \leq_{S_{A}} \pi_{S}(a)$ and $\pi_{S}(a_{1}+ \cdots + a_{n})=\pi_{S}(a)=x$.
The Zariski spectrum ${\cal Z}_{A}$ of the ring $A$ is defined in \cite{stone} as the space of points of the locale $Id_{C}({S(A)})$ of $C$-ideals in $S(A)$.
The second approach to the construction of ${\cal Z}_{A}$ defines ${\cal Z}_{A}$ as the prime spectrum (in the sense of section II3.4 \cite{stone}) of the distributive lattice $L(A)$ generated by symbols $D(a)$, $a \in A$, subject to the relations $D(1_{A})=1_{L(A)}$, $D(a\cdot b) = D(a) \wedge D(b)$, $D(0_{A})=0_{L(A)}$, and $D(a+ b) \leq D(a) \vee D(b)$.
By the results of section \ref{clex}, $L(A)$ can be seen as the coherent syntactic category of the coherent propositional theory over the signature having a propositional symbol $P(a)$ for each element $a\in A$, whose axioms are the following:
\[
(\top \vdash P_{1_{A}});
\]
\[
(P_{0_{A}} \vdash \bot);
\]
\[
(P_{a\cdot b} \dashv\vdash P_{a} \wedge P_{b})
\]
for any $a, b$ in $A$;
\[
(P_{a + b} \vdash P_{a} \vee P_{b})
\]
for any $a, b \in A$.
The prime spectrum of $L(A)$ can thus be identified with the space of points of the locale $Id_{J}(L(A))$, where $J$ is the coherent topology on $L(A)$.
To prove that this second definition of the Zariski spectrum of $A$ coincides with the first one, one observes that the universal map $\pi_{L}: A\to L(A)$ factors through the quotient map $\pi_{S}: A\to S(A)$, and the resulting factorization is universal among the meet-semilattice homomorphisms from $S(A)$ to distributive lattices which carry the covering families in $C$ to joinly covering families. Indeed, this implies, by Theorem \ref{freeframes}, that the frames $Id_{C}({S(A)})$ and $Id_{J}({L(A)})$ are isomorphic. Note that, by the results of sections \ref{genrel} and \ref{clex}, these two isomorphic frames can also be identified with the frame $F$ generated by symbols $F(a)$, $a \in A$, subject to the relations $F(1_{A})=1_{F}$, $F(a \cdot b) = F(a) \wedge F(b)$, $F(0_{A})=0_{F}$, and $F(a+ b) \leq F(a) \vee F(b)$.
Now, since the locales $Id_{C}({S(A)})$ and $Id_{J}({L(A)})$ are isomorphic, Theorem \ref{fund} yields an equivalence of toposes
\[
\Sh(S(A), C)\simeq \Sh(L(A), J)
\]
The fact that the two definitions of the Zariski spectrum ${\cal Z}_{A}$ given above coincide then follows at once from the fact that the subterminal topology is a topos-theoretic invariant and that on a localic topos $\Sh(L)$ equipped with the collection of all its points it yields the space of points of $L$ (cf. Example \ref{exa}(e) above).
To give an explicit description of ${\cal Z}_{A}$ as a topological space, we can use either of the two topos-theoretic representations. For example, using the representation $\Sh(S(A), C)$, we see that the points of ${\cal Z}_{A}\cong X_{{\tau}^{\Sh(S(A), C)}}$ are precisely the points of the topos $\Sh(S(A), C)$. These correspond bijectively, by definition of $S(A)$, to the functions $f:A\to \{0,1\}$ such that they factor through $\pi_{S}:A\to S(A)$ and this factorization is a meet-semilattice homomorphism $S(A)\to \{0,1\}$ which sends every $C$-covering sieve to a covering family in $\{0,1\}$. Identifying such functions $f$ with the subsets $f^{-1}(1)$ of $A$, we see that the these functions correspond bijectively to the subsets $S\subseteq A$ with the property that $1\in S$, $0_{A}\notin S$, $a\cdot b\in S$ if and only if $a\in S$ and $b\in S$, and $a+b\in S$ implies that $a\in S$ or $b\in S$. These subsets are called in \cite{stone} the \emph{prime filters} on $A$, and we denote their collection by ${\cal F}_{A}$.
On the other hand, identifying the functions $f$ above with the (complementary) subsets $f^{-1}(0)$ of $A$, we get a bijection between the set of such functions $f$ and the set $Spec(A)$ of prime ideals of $A$.
By Proposition \ref{mslattice}, the subterminal topology on the set ${\cal F}_{A}\cong X_{{\tau}^{\Sh(S(A), C)}}$ has as a sub-basis of open sets the collection of sets of the form
\[
{\cal F}_{a}=\{S\in {\cal F}_{A} \textrm{ | } a\in S\},
\]
where $a$ varies among the elements of $A$.
Now, under the bijection ${\cal F}_{A}\cong Spec(A)$ sending each subset to its complement in $A$, this topology on ${\cal F}_{A}$ corresponds precisely to the \emph{Zariski topology} on $Spec(A)$, that is to the topology whose closed sets are those of the form
\[
{\cal P}_{I}=\{P\in Spec(A) \textrm{ | } P\supseteq I\},
\]
where $I$ varies among the ideals of $A$. Indeed, $Spec(A)\setminus {\cal P}_{I}=\mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{a\in I}}\{P\in Spec(A) \textrm{ | } a\in A\setminus P\}$.
Note that, the ideals in $Spec(A)$ being prime, the open subsets of $Spec(A)$ corresponding to the sub-basic open sets ${\cal F}_{a}$ of ${\cal F}_{A}$ form actually a basis of $Spec(A)$ (since for any $a,b\in A$, $\{P\in Spec(A) \textrm{ | } a\notin P\}\cap \{P\in Spec(A) \textrm{ | } b\notin P\}=\{P\in Spec(A) \textrm{ | } a\cdot b\notin P\}$).
So the topological space obtained by equipping the set $Spec(A)$ with the Zariski topology is homeomorphic to the space ${\cal Z}_{A}\simeq X_{{\tau}^{\Sh(S(A), C)}}$, equivalently to the topological space $X_{{\tau}^{\Sh(L(A), J)}}$. Note that from this latter characterization it immediately follows that ${\cal Z}_{A}$ is a spectral space.
We conclude this section by showing that the technical lemma V3.2 \cite{stone} used to establish the bijection between the ideals of the distributive lattice $L(A)$ and the radical ideals of $A$ admits an alternative (and possibly more transparent) proof obtained by translating the thesis, naturally formulated in terms of the site $(L(A), J)$, into a property of the site $(S(A), C)$. This is done by transferring an appropriate topos-theoretic invariant across the Morita-equivalence
\[
\Sh(S(A), C)\simeq \Sh(L(A), J),
\]
according to the philosophy `toposes as bridges' of \cite{OC10}.
To this end, we need the following result.
\begin{lemma}\label{msites}
Let $F:({\cal C}, J) \to ({\cal D}, K)$ be a morphism of sites. Then we have a commutative diagram (in the $2$-category of toposes)
\[
\xymatrix {
[{\cal D}^{\textrm{op}}, \Set] \ar[r]^{f'} & [{\cal C}^{\textrm{op}}, \Set] \\
\Sh({\cal D}, K) \ar[u]^{i_{K}} \ar[r]^{f} & \ar[u]_{i_{J}} \Sh({\cal C}, J),}
\]
where $i_{J}$ and $i_{K}$ are the canonical geometric inclusions, $f$ is the geometric morphism induced by $F$ as in in Corollary C2.3.4 \cite{El}, and $f'$ is the geometric morphism induced by the functor $F:{\cal C}\to {\cal D}$ as in A4.1.10 \cite{El}.
\end{lemma}
\begin{proofs}
This follows immediately from the proof of Corollary C2.3.4 \cite{El}, observing that the direct image functor of $f$ is the restriction of the direct image functor of $f'$. Indeed, by the uniqueness (up to isomorphism) of adjoints, in order to prove that two geometric morphisms are isomorphic, it suffices to establish an isomorphism between their direct image functors.
\end{proofs}
Given a commutative ring with unit $A$, let us denote by $\xi:S(A)\to L(A)$ the factorization of $\pi_{L}:A\to L(A)$ through $\pi_{S}:A\to S(A)$. Clearly, $\xi$ is a morphism of sites $(S(A), C) \to (L(A), J)$ and hence it induces a geometric morphism $f:\Sh(S(A), C) \to \Sh(L(A), J)$. It is immediate to verify that this geometric morphism is precisely the equivalence $\Sh(S(A), C) \simeq \Sh(L(A), J)$ induced by Theorem \ref{fund} as indicated above. So, by Lemma \ref{msites}, we have a commutative diagram
\[
\xymatrix {
[L(A)^{\textrm{op}}, \Set] \ar[r]^{f'} & [S(A)^{\textrm{op}}, \Set] \\
\Sh(L(A), J) \ar[u]^{i_{J}} \ar[r]^{f} & \ar[u]^{i_{S}} \Sh(S(A), C),}
\]
where $i_{S}$ and $i_{J}$ are the canonical inclusions and $f'$ is the geometric morphism induced by $\xi:S(A)\to L(A)$ as in A4.1.10 \cite{El}.
Now, the statement of Lemma V3.2 \cite{stone} reads as follows:
\emph{Let $a, b_{1}, \ldots, b_{r}$ be elements of a commutative ring $A$ with unit. Then $D(a)\leq D(b_{1})\vee \cdots \vee D(b_{r})$ in $L(A)$ if and only if there exists an integer $n \geq 1$ and elements $b_{1}, \ldots, b_{r}$ such that $a^{n}=b_{1}\cdot c_{1}+ \cdots + b_{r}\cdot c_{r}$.}
To prove the lemma, we can clearly suppose, without loss of generality, that $b_{i}$ is a multiple of $a$ in $A$ (i.e., $b_{i}=a\cdot b_{i}'$ for some $b_{i}'\in A$) for all $i=1, \ldots, r$; this condition ensures that for all $i$, $\pi_{S}(b_{i}) \leq \pi_{S}(a)$ and $\pi_{L}(b_{i}) \leq \pi_{L}(a)$.
Consider the sieve $S_{B}$ in $S(A)$ on $\pi_{S}(a)$ generated by the family
\[
\{\pi_{S}(b_{i}) \to \pi_{S}(a) \textrm{ | } i=1, \ldots, r \}
\]
in $S(A)$.
Clearly, $\xi$ sends this family of arrows to the family $\{\pi_{L}(b_{i}) \to \pi_{L}(a) \textrm{ | } i=1, \ldots, r \}$ in $L(A)$; we denote by $L_{B}$ the sieve in $L(A)$ on $\pi_{L}(a)$ generated by this family. So, if $Y_{S}:S(A)\to [S(A)^{\textrm{op}}, \Set]$ and $Y_{L}:L(A)\to [L(A)^{\textrm{op}}, \Set]$ are the Yoneda embeddings, the inverse image functor of $f'$ sends the mono-\\morphism $B_{S}\mono Y_{S}(\pi_{S}(a))$ in $[S(A)^{\textrm{op}}, \Set]$ corresponding to the sieve $S_{B}$ to the monomorphism $B_{L}\mono Y_{L}(\pi_{L}(a))$ in $[L(A)^{\textrm{op}}, \Set]$ corresponding to the sieve $L_{B}$. Now, by the commutativity of the square above and the fact that $f$ is an equivalence, the associated sheaf functor $a_{S}:[S(A)^{\textrm{op}}, \Set] \to \Sh(S(A), C)$ sends $B_{S}$ to an isomorphism (equivalently, $S_{B}$ is $\overline{C}$-covering, where $\overline{C}$ is the Grothendieck topology on $S(A)$ generated by $C$) if and only if the associated sheaf functor $a_{L}:[L(A)^{\textrm{op}}, \Set] \to \Sh(L(A), J)$ sends $B_{L}$ to an isomorphism (equivalently, $L_{B}$ is $J$-covering). We thus deduce that $S_{B}$ is $\overline{C}$-covering if and only if $L_{B}$ is $J$-covering.
Now, the hypothesis `$D(a)\leq D(b_{1})\vee \cdots \vee D(b_{r})$ in $L(A)$' in the statement of the lemma is equivalent to the statement that $L_{B}$ is $J$-covering. From this, our argument enables us to conclude that $S_{B}$ is $\overline{C}$-covering. We now show that this condition implies the second condition of the lemma. To this end, we give an explicit description of the monoid congruence $\simeq$ involved in the definition of $S(A)$. It is easy to see that $\simeq$ is the transitive closure of the binary relation $R$ on $A$ defined by saying that for any $a,b\in A$, $a R b$ if and only if there exist $c,d\in A$ and non-zero integers $n,m\geq 1$ such that $a=c^{n}\cdot d$ and $b=c^{m}\cdot d$. From this it easily follows that if $a\simeq b$ then there exists a positive integer $k\geq 1$ such that $b=a^{k}\cdot e$ for some element $e$ of $A$; below, we will refer to this condition on $a$ and $b$ as to $(\ast)$. This remark allows us to achieve an explicit description of the Grothendieck topology $\overline{C}$ on $S(A)$.
\begin{lemma}\label{topC}
Let $A$ be a commutative ring with unit. With the above notation, for any sieve $S$ in $S(A)$ on an element $x\in S(A)$, $S\in \overline{C}(x)$ if and only if $S$ contains a finite family of arrows in $S(A)$ of the form $\{\pi_{S}(a_{i})\to \pi_{S}(a) \textrm{ | } i=1, \ldots, n\}$, where $\pi_{S}(a_{i}) \leq_{S_{A}} \pi_{S}(a)$ and $a^{k}=a_{1}\cdot b_{1}+ \cdots + a_{n}\cdot b_{n}$ for some positive integer $k\geq 1$ and elements $b_{1}, \ldots, b_{n}$ in $A$.
\end{lemma}
\begin{proofs}
The `if' direction is clear from the definition of the coverage $C$; indeed, if $a^{k}=a_{1}\cdot b_{1}+ \cdots + a_{n}\cdot b_{n}$ with $\pi_{S}(a_{i}) \leq_{S_{A}} \pi_{S}(a)$ then the family of arrows $\{\pi_{S}(a_{i}\cdot b_{i})\to \pi_{S}(a) \textrm{ | } i=1, \ldots, n\}$ is $C$-covering and hence the sieve generated by the family $\{\pi_{S}(a_{i})\to \pi_{S}(a) \textrm{ | } i=1, \ldots, n\}$ is $\overline{C}$-covering, since it contains a $C$-covering family.
The `only if' direction can be deduced as a consequence of the following two facts:
\begin{enumerate}[(i)]
\item Every $C$-covering sieve satisfies the condition in the statement of the lemma;
\item The collection of the sieves on $S(A)$ which satisfy the condition in the statement of the lemma forms a Grothendieck topology;
\end{enumerate}
Fact $(i)$ follows easily from the definition of the coverage $C$ and the explicit description of the monoid congruence $\simeq$ obtained above, while Fact $(ii)$ is immediately verified.
\end{proofs}
Coming back to our original problem, the `if' direction of Lemma V3.2 \cite{stone} is trivial, so we only care to prove the `only if' one. If $S_{B}$ is $\overline{C}$-covering then, by Lemma \ref{topC}, $S_{B}$ contains a finite family of arrows of the form $\{\pi_{S}(a_{i})\to \pi_{S}(a) \textrm{ | } i=1, \ldots, n\}$, where $\pi_{S}(a_{i}) \leq_{S_{A}} \pi_{S}(a)$ and $a^{k}=a_{1}\cdot c_{1}+ \cdots + a_{n}\cdot c_{n}$ for some positive integer $k\geq 1$ and elements $c_{1}, \ldots, c_{n}$ in $A$. Now, by definition of $S_{B}$, for every $i\in \{1,\ldots, n\}$ there exists $j_{i}\in \{1, \ldots, r\}$ such that $\pi_{S}(a_{i}) \leq \pi_{S}(b_{j_{i}})$, equivalently $a_{i}\cdot b_{j_{i}}\simeq a_{i}$. From this we deduce, by invoking the characterization of the congruence $\simeq$ on $A$ obtained above, that for any sufficiently large natural number $k_{i}$, $a_{i}^{k_{i}}$ is a multiple of $b_{j_{i}}$ in $A$, which in turn implies that for a sufficiently large natural number $k$, $a^{k}$ is a sum of multiples of the $b_{i}$. This completes our proof of Lemma V3.2 \cite{stone}.
It is natural to wonder whether there is a natural way to generalize the Zariski topology on the collection of prime ideals of a ring to the collection of all the proper ideals of the ring, or to some other more general class of subsets of the ring. Thanks to the techniques developed in the paper, we are able to give a positive answer to this question. We can use propositional geometric theories to describe subsets of a ring with particular properties, such as the class of ideals of the ring; the subterminal topology then provides a way of endowing the collection of models of such a theory with a topology such that the topos of sheaves on the resulting topological space is equivalent to the classifying topos of the propositional theory; also, as we have seen above, the results of section \ref{genrel} enable us to achieve, in many cases of interest, an explicit semantic description of such classifying topos as a topos of sheaves on a poset structure presented by generators and relations with respect to some Grothendieck topology on it.
For example, consider the case of (proper) ideals of a commutative ring with unit $A$. In order to obtain the Zariski topology, we characterized the prime ideals of $A$ as the complements in $A$ of the models of a particular propositional theory, namely the theory of prime filters on $A$. Therefore, a first natural step to generalize the Zariski topology to a topology on the set of proper ideals of $A$ is to try to find a propositional theory whose models are exactly the complements in $A$ of proper ideals; we will refer to these subsets as to the \emph{op-ideals} on $A$. Explicitly, an op-ideal on $A$ is a subset $S\subseteq A$ such that for all $a, b\in A$, $a\cdot b\in S$ implies $a\in S$, $a+b\in I$ implies $a\in I$ or $b\in I$, $0_{A}\notin I$ and $1_{A}\in I$.
Op-ideals can be described as the models of the propositional coherent theory $\mathbb Q$ over the signature consisting of one $0$-ary relation symbol $P_{a}$ for each element $a\in A$, with axioms:
\[
P_{0_{A}} \vdash \bot;
\]
\[
\top \vdash P_{1_{A}};
\]
\[
(P_{a\cdot b} \vdash P_{a} \wedge P_{b})
\]
for any $a, b$ in $A$, and
\[
(P_{a + b} \vdash P_{a} \vee P_{b})
\]
for any $a, b \in A$.
The classifying topos of $\mathbb Q$, which is equivalent to the category of sheaves on the coherent syntactic category ${\cal C}_{\mathbb Q}$ of $\mathbb Q$ with respect to the coherent topology on it, can be represented as the topos of sheaves on the topological space obtained by equipping the set $Spec_{id}(A)$ of ideals on $A$ with the elemental topology (cf. section \ref{subs}). Note that ${\cal C}_{\mathbb Q}$ can be characterized as the distributive lattice $D(A)$ generated by symbols $D(a)$, $a \in A$, subject to the relations $D(0_{A})=0_{D(A)}$, $D(1_{A})=1_{D(A)}$, $D(a\cdot b) \leq D(a) \wedge D(b)$, $D(a+ b) \leq D(a) \vee D(b)$. The frame of open sets of $Spec_{id}(A)$ is thus isomorphic to the frame of ideals of the distributive lattice ${\cal C}_{\mathbb Q}$, and, by the results of sections \ref{clex} and \ref{genrel}, can also be characterized as the geometric syntactic category of the theory $\mathbb Q$, that is as the frame $F(A)$ generated by symbols $F(a)$, $a \in A$, subject to the relations $F(0_{A})=0_{F(A)}$, $F(1_{A})=1_{F(A)}$, $F(a\cdot b) \leq F(a) \wedge F(b)$ and $F(a+ b) \leq F(a) \vee F(b)$.
\section{Conclusions}\label{conclusions}
The theory developed in this paper paves the way for a vast world of new possibilities. Three natural directions that one could immediately pursue are the following.
First, one can investigate particular dualities already generated through our machinery, and described in sections \ref{ex} and \ref{addex}; this should be interesting because, as it is clear from our topos-theoretic interpretation, all of these dualities have essentially the same level of `mathematical depth' as the classical Stone duality.
Second, one can generate new dualities for particular classes of preordered structures by using our machinery. As we have already remarked in the course of the paper, the process of generation of new `Stone-type' dualities is essentially automatic, and relies on the choice of two (in the case of dualities with locales) or three (in the case of dualities with topological spaces) ingredients to give our `machine' as `inputs': the initial category $\cal K$ of preordered structures, the Grothendieck topologies $J_{\cal C}$ associated to the structures in $\cal K$, and the sets of points of the toposes $\Sh({\cal C}, J_{\cal C})$ corresponding to the structures. These choices are, although in a relation of sequential dependence each on the previous ones, essentially independent from each other, in the sense that the choice of a certain ingredient at one stage does not uniquely determine the choices of the ingredients at later stages. These two (resp. three) `degrees of freedom' in the choice of ingredients confer to our machinery a great level of techical flexibility, which allows us in particular to identify a given category of preorders with a subcategory of a category of locales or topological spaces in several non-equivalent ways (cf. section \ref{ex} for some examples of this phenomenon). In fact, the more general methodology of section \ref{Mordual} shows that there are essentially \emph{three} degrees of freedom in generating dualities (resp. equivalences) between general categories of preordered structures, with \emph{two} additional degrees of freedom if one wants to `lift' these dualities (resp. equivalences) to `topological' ones.
Third, one can perform topos-theoretic translations between between properties of preordered structures and properties of the locales or topological spaces corresponding to them via both the `Stone-type dualities', according to the technique `toposes as bridges' of \cite{OC10}; as we remarked in section \ref{insights}, this translation can be performed automatically in many cases, and, as the examples provided in section \ref{insights} show, the results generated in this way are non-trivial in general.
Overall, this paper can be read as an assay of what can be achieved by applying the methodologies of \cite{OC10}. Indeed, in \cite{OC10}, we emphasized the importance of taking Morita-equivalences as starting points of topos-theoretic investigations, and of using topos-theoretic invariants to extract information about them which is relevant for classical mathematics; the common classifying topos acts in this context as a `bridge' for transferring properties and constructions between its two different sites of definition related to each other by the Morita-equivalence. Now, in this paper, starting from the Morita-equivalence of Theorem \ref{fund}, we have proceeded to extract information about it through the use of topos-theoretic invariants of various kinds. Through the lenses of the unifying framework that we have developed, the different Stone-type dualities, as well as the other results that we have obtained in the paper, appear to be just `variations on the same theme', this theme being precisely the original Morita-equivalence (or, more generally, any equivalence of the form $\Sh({\cal C}, J)\simeq \Sh({\cal D}, K)$ as considered in section \ref{generalization}).
In order to keep our treatment self-contained, we have chosen not to discuss in detail in the paper any examples involving categories which are not preorder, but in fact there are many ideas in the paper, motivated by the principles introduced in \cite{OC10}, which can be profitably extended from the preorder context to the level of arbitrary (small) categories. For example, the idea of recovering a category $\cal C$ from a topos $\Sh({\cal C}, J)$, where $J$ is a subcanonical topology on $\cal C$, through a topos-theoretic invariant is clearly very general and as such it is liable to be applied in a variety of different contexts. In fact, whenever we have a Morita-equivalence $\Sh({\cal C}, J)\simeq \Sh({\cal D}, K)$ such that $\cal C$ can be recovered from the topos $\Sh({\cal C}, J)$ through a topos-theoretic invariant $U$, we can represent $\cal C$ in terms of the site $({\cal D}, K)$ as the category of objects of the topos $\Sh({\cal D}, K)$ which satisfy the invariant $U$, and if this equivalence holds `naturally' for $\cal C$ varying in a (large) category $\cal K$ of small categories, we can expect to be able to `functorialize' this representation of $\cal C$ in terms of $({\cal D}, K)$ so to yield a duality between $\cal K$ and a category consisting of objects from which the toposes $\Sh({\cal D}, K)$ can be `directly built' (so, for example, if the topologies $K$ on $\cal D$ are `uniformly defined' in terms of $\cal D$ by means of a topos-theoretic invariant then specifying $\cal D$ is enough to `build' the site $({\cal D}, K)$ and hence the topos $\Sh({\cal D}, K)$).
While we are on the topic of Morita-equivalences holding for large classes of toposes, we remark that a `first-order analogue' of the Morita-equivalence of Theorem \ref{fund} - which represents the topos of sheaves on a site whose underlying category is a preorder as the topos of sheaves on a locale - is the representation theorem \cite{AT} of Joyal and Tierney, which gives a representation of a general Grothendieck topos as the topos of sheaves on a localic groupoid, while an analogue of the representation of a localic topos with enough points as a topos of sheaves on a topological space is provided by Butz and Moerdijk's representation theorem \cite{BM} for Grothendieck toposes with enough points as toposes of sheaves on topological groupoids. As we have already remarked, we can expect the general methodology `toposes as bridges' recalled above to yield interesting results also in the context of these Morita-equivalences, in the form of representation theorems, adjunctions, dualities for various classes of categories or topos-theoretic translations of properties (or constructions) from one side of such a Morita-equivalence to the other. Incidentally, a result which already witnesses the effectiveness of our abstract technique also in the first-order context is the duality between the category of ($k$-small) Boolean pretoposes and Stone topological groupoids obtained in \cite{AF}, which can be seen as arising from the processes of functorializing (an adaptation of) Butz and Moerdijk's representation theorem and recovering a Boolean pretopos from the corresponding topos of coherent sheaves though a topos-theoretic invariant.
Finally, as already remarked in section \ref{concreteness}, the methods of this paper, combined with the philosophy `toposes as bridges' of \cite{OC10}, provide a topos-theoretic interpretation of the problem of finding `concrete' descriptions of models of ordered algebraic theories (in the sense of section \ref{spacesprop}) presented by generators and relations. Indeed, given such a model $\cal C$, regarded as a preorder category, for any Grothendieck topology $J$ such that $\cal C$ can be recovered, up to isomorphism, from $\Sh({\cal C}, J)$ through a topos-theoretic invariant, any alternative representation of the topos $\Sh({\cal C}, J)$ in terms of a different site $({\cal D}, K)$ yields an alternative description of $\cal C$ in terms of this latter site (applications of this methodology are given in section \ref{concreteness}); conversely, any explicit description of $\cal C$ as a preorder isomorphic to a certain structure $\cal D$ yields a Morita-equivalence $\Sh({\cal C}, J)\simeq \Sh({\cal D}, K)$, where $K$ is the Grothendieck topology on $\cal D$ given by the image of $J$ under the isomorphism ${\cal C}\simeq {\cal D}$. The problem of finding alternative descriptions of structures presented by generators and relations is thus strictly connected to the problem of finding alternative representations for toposes of sheaves on the structures with respect to appropriate subcanonical topologies. This interpretation paves the way for the use of topos-theoretic methods for addressing this kind of problems, which are in general rather hard to solve (cf. for example the problem discussed in section \ref{freecomjoin}); compelling examples of the validity of these methods are already given in the paper (specifically in section \ref{concreteness}), and the reader can easily generate new ones by following the same principles.
\newpage
\section{Appendix}\label{appendix}
In this appendix, in order to make (parts of) the theory developed in the paper intelligible to the reader who is not familiar with Topos Theory, we provide elementary proofs (i.e. proofs consisting of direct, non-topos-theoretic arguments) of some of the key results in the paper, and indicate how the arguments which constitute the general machinery for building dualities of sections \ref{locales}, \ref{dualtop} and \ref{Mordual} can be rewritten in `elementary' terms.
Let us start from the construction of frames from preorders (in general, arbitrary categories) equipped with Grothendieck topologies.
The following notions are the specializations to preorders of the well-known topos-theoretic notions of coverage and of site. For an element $c\in {\cal C}$ of a preorder $({\cal C}, \leq)$, we define the principal ideal $(c)\downarrow$ on $c$ as the set of all the elements $d$ of $\cal C$ such that $d\leq c$ in $\cal C$.
\begin{definition}
Let $({\cal C}, \leq)$ be a preorder.
\begin{enumerate}[(i)]
\item A \emph{coverage} on $\cal C$ is a function $J$ which assigns to every element $c\in {\cal C}$ a family $J(c)$ of subsets of $(c)\downarrow$ such that for any $S\in J(c)$ and any $c'\leq c$ the subset $S_{c'}=\{d\leq c' \textrm{ | } d\in S\}$ belongs to $J(c')$;
\item A coverage on $\cal C$ is said to be a \emph{Grothendieck coverage} if for any $c\in {\cal C}$, $(c)\downarrow \in J(c)$ and for any subset $S\subseteq (c)\downarrow$, if $S_{c'}\in J(c')$ for every $c'\in T$ where $T\in J(c)$ then $S\in J(c)$;
\item A \emph{site} is a pair $({\cal C}, J)$, where $\cal C$ is a preorder and $J$ is a coverage on $\cal C$;
\item A coverage $J$ on $\cal C$ is \emph{subcanonical} if for every $c\in {\cal C}$ and any subset $S\in J(c)$, $c$ is the supremum in $\cal C$ of the elements $d\in S$ (i.e., for any element $c'$ in $\cal C$ such that for every $d\in S$ $d\leq c'$, we have $c\leq c'$).
\end{enumerate}
\end{definition}
For example, on a distributive lattice $\cal C$ we can put the \emph{coherent coverage} $J$, defined by saying that for any $c\in {\cal C}$, the subsets $S$ in $J(c)$ are precisely the subsets of $(c)\downarrow$ which contain a finite subset whose join is $c$. Given a frame $L$, we define the \emph{canonical coverage} on $L$ as the coverage $J_{can}^{L}$ such that for any $c\in L$ and $S\subseteq (c)\downarrow$, $S\in J_{can}^{L}(c)$ if and only if $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{d\in S}}d=c$.
\begin{remarks}
\begin{enumerate}[(a)]
\item If $({\cal C}, \wedge)$ is a meet-semilattice and $J$ is a function which assigns to every element $c\in {\cal C}$ a family $J(c)$ of subsets of $(c)\downarrow$ such that for any $S\in J(c)$ and any $c'\leq c$, the subset $\{a\wedge c' \textrm{ | } a\in S\}$ belongs to $J(c')$ then the function $J'$ which sends an element $c$ of $\cal C$ to the collection of the subsets of $(c)\downarrow$ which contain a subset in $J(c)$ is a coverage on ${\cal C}$;
\item Given a coverage $J$ on a preorder $\cal C$, there is a smallest Grothendieck coverage $J'$ on $\cal C$ which contains $J$ (i.e. such that for any $c\in {\cal C}$, $S\in J(c)$ implies $S\in J'(c)$), called the \emph{Grothendieck coverage generated by $J$}: $J'$ can be defined by saying that for any $c\in {\cal C}$, the subsets in $J(c)$ are precisely the subsets which belong to $J'(c)$ for every Grothendieck coverage $J'$ containing $J$.
\end{enumerate}
\end{remarks}
Let us recall from section \ref{general} the following notions.
\begin{definition}
Let $({\cal C}, \leq)$ be a preorder and $J$ be a coverage on $\cal C$.
\begin{enumerate}
\item An \emph{ideal} on $\cal C$ is a subset $I\subseteq {\cal C}$ such that for any $a,b \in {\cal C}$ such that $b\leq a$ in $\cal C$, $a\in I$ implies $b\in I$. A \emph{$J$-ideal} on $\cal C$ is an ideal on $\cal C$ such that for any $R\in J(c)$, if $a\in I$ for every $a\in R$ then $c\in I$.
\item Given a subset $I$ on $\cal C$, the \emph{$J$-closure $\overline{I}^{J}$} is the smallest $J$-ideal on $\cal C$ containing $I$, equivalently the intersection of all the $J$-ideals which contain $I$.
\item Given an object $c$ of $\cal C$, we define the \emph{principal $J$-ideal} $(c)\downarrow_{J}$ generated by $c$ as the $J$-closure of the subset $(c)\downarrow$ of $\cal C$.
\end{enumerate}
\end{definition}
\begin{remark}
Note that, if $J$ is the trivial coverage on $\cal C$ (i.e., the coverage in which for any $c\in {\cal C}$, $S\in J(c)$ if and only if $S=(c)\downarrow$) then the $J$-ideals on $\cal C$ are precisely the ideals on $\cal C$.
\end{remark}
Given a preorder ${\cal C}$, equipped with a coverage $J$, we define \emph{$Id_{J}({\cal C})$} to be the set of all the $J$-ideals on $\cal C$, equipped with the subset-inclusion ordering $\subseteq$. The set of all the ideals on $\cal C$ will be denoted by $Id({\cal C})$.
\begin{remark}\label{grotcov}
\begin{enumerate}[(a)]
\item
If $J$ is a Grothendieck coverage on a preorder $({\cal C}, \leq)$ then for any ideal $I$ on $\cal C$, the $J$-closure $\overline{I}^{J}$ of $I$ is equal to the set of elements $d\in {\cal C}$ such that there exists $R\in J(d)$ with the property that for every $a\in R$, $a\in I$;
\item If $J$ is a coverage on a preorder $({\cal C}, \leq)$ and $J'$ is the Grothendieck coverage on $\cal C$ generated by $J$ then for any ideal $I$ on $\cal C$, $I$ is $J$-closed if and only if $I$ is $J'$-closed; in particular, the $J$-closure of $I$ coincides with the $J'$-closure of $I$.
\end{enumerate}
\end{remark}
\begin{theorem}
Let $\cal C$ be a preorder and $J$ be a coverage on $\cal C$. Then $(Id_{J}({\cal C}), \subseteq)$ is a frame.
\end{theorem}
\begin{proofs}
Clearly, the intersection of any two $J$-ideals on $\cal C$ is a $J$-ideal; this gives the meet in $Id_{J}({\cal C})$. The join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{k\in K}}I_{k}$ of a family $\{I_{k} \textrm{ | } k\in K \}$ of $J$-ideals on $\cal C$ is given by the $J$-closure of the set-theoretic union of the $I_{k}$; explicitly, $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{k\in K}}I_{k}=\{d\in {\cal C} \textrm{ | } \{e\leq d \textrm{ | } e\in \mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{k\in K}}I_{k}\}\supseteq S \textrm{ for some } S\in J'(d)\}$, where $J'$ is the Grothendieck topology on $\cal C$ generated by $J$ (cf. Remark \ref{grotcov}(a)).
The infinite distributive law of joins with respect to finite meets is easily verified, by using the infinite distributive law of unions with respect to finite intersections and the fact that the operation of $J'$-closure commutes with finite intersections.
\end{proofs}
Let us recall the following notion from section \ref{locales}.
\begin{definition}
Let $\cal C$ and $\cal D$ be preorders. A monotone map $f:{\cal C}\to {\cal D}$ is said to be \emph{flat} if the following two conditions hold:
\begin{enumerate}[(i)]
\item For any $d\in {\cal D}$ there exists $c\in {\cal C}$ such that $d\leq f(c)$;
\item For any element $d\in {\cal D}$ and any elements $c,c'\in {\cal C}$ such that $d\leq f(c)$ and $d\leq f(c')$ there exists $c''\in {\cal C}$ such that $c''\leq c$, $c''\leq c'$ and $d\leq f(c'')$.
\end{enumerate}
\end{definition}
\begin{remark}
If $\cal C$ and $\cal D$ are meet-semilattices then a monotone map $f:{\cal C}\to {\cal D}$ is flat if and only if it is a meet-semilattice homomorphism.
\end{remark}
If $\cal C$ and $\cal D$ are two preorders equipped respectively with coverages $J$ and $K$, we say that a flat map $f:{\cal C}\to {\cal D}$ is \emph{cover-preserving} if for any element $c\in {\cal C}$ and any $R\in J(c)$, the set of elements $d$ of $\cal D$ such that $d\leq f(c')$ for some $c'\in R$ contains a subset belonging to $K(f(c))$.
\begin{theorem}\label{concretecontrov}
Let $\cal C$ and $\cal D$ be two preorders equipped respectively with coverages $J$ and $K$, and let $f:{\cal C}\to {\cal D}$ be a flat cover-preserving map between them. Then the map
\[
A_{f}:Id_{J}({\cal C}) \to Id_{K}({\cal D})
\]
sending any $J$-ideal $I$ on $\cal C$ to the smallest $K$-ideal on $\cal D$ containing $f(I)$ is a frame homomorphism (where $Id_{J}({\cal C})$ and $Id_{K}({\cal D})$ are endowed with the subset-inclusion ordering).
\end{theorem}
\begin{proofs}
Let us begin by proving that $A_{f}$ preserves arbitrary joins. Given a family $\{I_{h} \textrm{ | } h\in H\}$ of $J$-ideals on $\cal C$, we have that $A_{f}(\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{h\in H}}I_{h})=\overline{f(\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{h\in H}}I_{h})}^{K}=\overline{f(\overline{\mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{h\in H}}I_{h}}^{J})}^{K}=\overline{\mathbin{\mathop{\textrm{\huge $\cup$}}\limits_{h\in H}}f(I_{h})}^{K}=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{h\in H}}\overline{f(I_{h})}^{K}=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{h\in H}}A_{f}(I_{h})$, where the third equality follows from the fact that $f$ is cover-preserving.
To prove that $A_{f}$ is a frame homomorphism, it remains to show that it preserves the top element and binary meets.
The fact that $A_{f}$ preserves the top element follows immediately from condition $(i)$ in the definition of flat map.
To prove that $A_{f}$ preserves finite meets, suppose that $I_{1}$ and $I_{2}$ are $J$-ideals on $\cal C$. Then, condition $(ii)$ in the definition of flat map ensures that the ideal on $\cal D$ generated by $f(I_{1})\cap f(I_{2})$ coincides with the ideal on $\cal D$ generated by $f(I_{1}\cap I_{2})$; therefore $A_{f}(I_{1}\cap I_{2})=\overline{f(I_{1}\cap I_{2})}^{K}=\overline{f(I_{1})\cap f(I_{2})}^{K}=\overline{f(I_{1})}^{K}\cap\overline{f(I_{2})}^{K}=A_{f}(I_{1})\cap A_{f}(I_{2})$.
\end{proofs}
\begin{definition}
Let $\cal C$ be a preorder and $J$ be a coverage on $\cal C$. We say that $J$ is \emph{subcanonical} if for every $c\in {\cal C}$ and any $S\in J(c)$, for any $c'\in {\cal C}$, if $d\leq c'$ for every $d\in S$ then $c\leq c'$).
\end{definition}
\begin{remark}
A coverage $J$ on a preorder $\cal C$ is subcanonical if and only if $(c)\downarrow_{J}=(c)\downarrow$ for any $c\in {\cal C}$.
\end{remark}
Given a flat cover-preserving map $f:{\cal C}\to {\cal D}$ as in the statement of Theorem \ref{concretecontrov}, and denoted by $i_{\cal C}:{\cal C}\to Id_{J}({\cal C})$ (resp. $i_{\cal D}:{\cal D}\to Id_{K}({\cal D})$) the map sending an element $c\in {\cal C}$ to the principal $J$-ideal $(c)\downarrow_{J}\in Id_{J}({\cal C})$ (resp. an element $d\in {\cal D}$ to the principal $K$-ideal $(d)\downarrow_{K}\in Id_{K}({\cal D})$), $A_{f}\circ i_{\cal C}=i_{\cal D}\circ f$. In particular, if the coverages $J$ and $K$ satisfy $(c)\downarrow_{J}=(c)\downarrow$ for any $c\in {\cal C}$ and $(d)\downarrow_{K}=(d)\downarrow$ for any $d\in {\cal D}$ (that is, if $J$ and $K$ are subcanonical) and the preorders $\cal C$ and $\cal D$ are posets then $f$ can be recovered from $A_{f}$ as its restriction to the subsets of principal ideals.
Theorem \ref{concretecontrov} provides a way for obtaining frame homomorphisms between frames of ideals on preordered structures starting from monotone maps between the structures. On the other hand, if the structures are not equipped with any coverage, we can build frame homomorphisms between the frames of ideals on them, as follows.
For any monotone map $f:{\cal C}\to {\cal D}$ between preorders, the map $B_{f}:Id({\cal D})\to Id({\cal C})$ sending an ideal $I$ on $\cal D$ to the inverse image $f^{-1}(I)$ of $I$ under $f$ is a frame homomorphism. It is natural to wonder if it is possible to recover $f$ from $B_{f}$ and characterize `intrinsically' the frame homomorphisms $Id({\cal D})\to Id({\cal C})$ of the form $B_{f}$ for some monotone map $f:{\cal C}\to {\cal D}$. The following result provides an answer to these questions. Below, for a preorder $\cal C$, we denote by $i_{\cal C}:{\cal C}\to Id({\cal C})$ the map sending an element $c\in {\cal C}$ to the principal ideal $(c)\downarrow \in Id({\cal C})$.
\begin{theorem}\label{concretecovariant}
\begin{enumerate}[(i)]
\item Let $\cal C$ and $\cal D$ be two posets. A frame homomorphism $F:Id({\cal D})\to Id({\cal C})$ is of the form $B_{f}$ for some monotone map $f:{\cal C}\to {\cal D}$, where $B_{f}:Id({\cal D})\to Id({\cal C})$ is the frame homomorphism sending an ideal $I$ on $\cal D$ to the inverse image $f^{-1}(I)$ of $I$ under $f$, if and only if $F$ preserves arbitrary infima, equivalently it has a left adjoint $F_{!}:Id({\cal C})\to Id({\cal D})$, given by the formula $F_{!}(I)=\mathbin{\mathop{\textrm{\huge $\cap$}}\limits_{I \subseteq F(I')}I'}$ (for any $I\in Id({\cal C})$);
\item For any monotone map $f:{\cal C}\to {\cal D}$, ${(B_{f})}_{!}\circ i_{\cal C}=i_{\cal D}\circ f$; in particular, $f$ can be recovered from $B_{f}$ as the restriction of its left adjoint ${(B_{f})}_{!}$ to the subsets of principal ideals.
\end{enumerate}
\end{theorem}
\begin{proofs}
$(i)$ It is clear that for any monotone map $f:{\cal C}\to {\cal D}$, $B_{f}:Id({\cal D})\to Id({\cal C})$ preserves arbitrary intersections (i.e., arbitrary infima). Conversely, suppose that $F:Id({\cal D})\to Id({\cal C})$ preserves arbitrary intersections, equivalently (by the Special Adjoint Functor Theorem) that it has a left adjoint $F_{!}:Id({\cal C})\to Id({\cal D})$ given by the formula $F_{!}(I)=\mathbin{\mathop{\textrm{\huge $\cap$}}\limits_{I \subseteq F(I')}I'}$ (for any $I\in Id({\cal C})$). Let us show that $F_{!}$ sends principal ideals to principal ideals. Given $c\in {\cal C}$, the formula for $F_{!}$ yields $F_{!}((c)\downarrow)=\mathbin{\mathop{\textrm{\huge $\cap$}}\limits_{c \in F(I')}I'}$. Now, since $F$ preserves arbitrary intersections, $T:=\mathbin{\mathop{\textrm{\huge $\cap$}}\limits_{c \in F(I')}I'}$ satisfies the property that $c\in F(T)$ and is the smallest ideal on $\cal D$ with this property. This minimality condition implies that $F_{!}((c)\downarrow)=T$ is a principal ideal on $\cal D$, since $c\in F(T)$ implies that $c\in F((t)\downarrow)$ for some $t\in T$. Now, since $\cal C$ and $\cal D$ are posets, the principal ideals on them can be identified with the elements that generate them, and hence we have a monotone map $f:{\cal C}\to {\cal D}$ such that for every $c\in {\cal C}$, $F_{!}((c)\downarrow)= (f(c))\downarrow$. This implies, since $F_{!}$ preserves unions, that for any ideal $I$ on $\cal C$, $F_{!}(I)=f(I)$, which in turn implies that the right adjoint $F$ to $F_{!}$ is equal to the inverse image map $f^{-1}:Id({\cal D})\to Id({\cal C})$, in other words to $B_{f}$.
$(ii)$ This follows immediately from the proof of part $(i)$.
\end{proofs}
The method of section \ref{locales} for constructing dualities or equivalences between a given category of posets and a category of locales can be reformulated in `concrete' terms as follows.
Let us start with the controvariant case. Given a collection of preorders $\cal C$, each of which equipped with a coverage $J_{\cal C}$, consider the category $\cal K$ having as objects the posets $\cal C$ and as arrows ${\cal C} \to {\cal D}$ the flat cover-preserving maps $({\cal C}, J_{\cal C})\to ({\cal D}, J_{\cal D})$. Then we have a functor $A:{\cal K}\to \textbf{Frm}$ sending any preorder $\cal C$ in $\cal K$ to $Id_{J_{\cal C}}({\cal C})$ and any arrow $f:{\cal C}\to {\cal D}$ in $\cal K$ to the frame homomorphism $A_{f}:Id_{J_{\cal C}}({\cal C}) \to Id_{J_{\cal D}}({\cal D})$ given by Theorem \ref{concretecontrov}.
In the covariant case, one defines a functor $B:\textbf{Pro}^{\textrm{op}}\to \textbf{Frm}$, where $\textbf{Pro}$ is the category of preorders and monotone maps between them, by sending a preorder $\cal C$ to the frame $Id({\cal C})$ of ideals on $\cal C$, and a monotone map $f$ to the frame homomorphism $B_{f}:Id({\cal D})\to Id({\cal C})$ of Theorem \ref{concretecovariant}.
In the controvariant case, if the topologies $J_{\cal C}$ are subcanonical and the preorders $\cal C$ are posets then the map $i_{\cal C}:{\cal C}\to Id_{J}({\cal C})$ is an embedding and hence we can hope to find a categorical inverse to the functor $A$, by using the technique of section \ref{charinv}.
In the covariant case, we can find a categorical inverse to the restriction to the opposite of the category $\textbf{Pos}$ of posets of the functor $B:\textbf{Pro}^{\textrm{op}}\to \textbf{Frm}$, by using the technique of section \ref{charinv}.
The proofs of the main results of section \ref{charinv} are already direct and all the concepts which appear in them can be straightforwardly reformulated in terms of frame-theoretic invariants, rather than in terms of topos-theoretic invariants of families of subterminals in a locally small cocomplete topos (cf. Remark \ref{loctop}).
Summarizing the results obtained in section \ref{charinv}, we have that if all the (Grothendieck topologies generated by the) coverages $J_{\cal C}$ are $C$-induced (in the sense of Definition \ref{inducedtop}) for a frame-theoretic invariant $C$ satisfying the property that for any structure $\cal C$ in $\cal K$ and for any family $\cal F$ of principal $J_{\cal C}$-ideals on $\cal C$, $\cal F$ has a refinement satisfying $C$ (if and) only if it has a refinement made of principal $J_{\cal C}$-ideals on $\cal C$ satisfying $C$, then we can define a functor $I_{A}$ (resp. a functor $I_{B}$) on the extended image of the functor $A$ (resp. of the functor $B$) which is a categorical inverse to the functor $A$ (resp. to the functor $B$). The notion of $C$-compactness plays a central role in obtaining `intrinsic' characterization of the extended image of the functor $A$ (resp. of the functor $B$), and in defining the functor $I_{A}$ (resp. the functor $I_{B}$). Recall that an element $l$ of a frame $L$ is said to be $C$-compact if every covering of $l$ in $L$ admits a refinement which satisfies $C$. The functor $I_{A}$ (resp. the functor $I_{B}$) sends a frame $L$ in the extended image of $A$ (resp. of $B$) to the collection of the elements of $L$ which are $C$-compact and acts on the arrows accordingly. We refer to section \ref{charinv} for the rigorous definitions and the details of this technique.
The more general methodology of section \ref{Mordual} takes as starting point a bunch of equivalences of the form $Id_{J_{\cal C}}({\cal C})\cong Id_{K_{\cal D}}({\cal D})$ and relies on the existence of invariants $C$ and $D$ of families of elements of frames which satisfy the property that for any family $\cal F$ of principal $J_{\cal C}$-ideals on $\cal C$ (resp. of $K_{\cal D}$-ideals on $\cal D$), $\cal F$ has a refinement satisfying $C$ (resp. $D$) (if and) only if it has a refinement made of principal $J_{\cal C}$-ideals on $\cal C$ (resp. of principal $K_{\cal D}$-ideals on $\cal D$)) satisfying $C$ (resp. $D$), and such that all the topologies $J_{\cal C}$ (resp. of $K_{\cal D}$) are $C$-induced (resp. $D$-induced). The technique of section \ref{Mordual} then produces equivalences between categories of posets consisting of the $C$-compact elements of some frame $L$ and categories of posets consisting of the $D$-compact elements of $L$ (cf. section \ref{Mordual} for the details). The technique of section \ref{locales} is recovered as a particular case of this more general technique when the original equivalences are of the form $Id_{J_{\cal C}}({\cal C})\cong Id_{Id_{J_{\cal C}}({\cal C})}(Id_{J_{\cal C}}({\cal C}))$, where $J^{Id_{J_{\cal C}}({\cal C})}_{can}$ is the canonical topology on the frame $Id_{J_{\cal C}}({\cal C})$.
The method of section \ref{Mordual} works for arbitrary bunches of equivalences
\[
Id_{J_{\cal C}}({\cal C})\cong Id_{K_{\cal D}}({\cal D}).
\]
An effective way for generating such equivalences is to use the following result, which is the particular case of the Comparison Lemma in Topos Theory for preorder categories (in fact, all the known Stone-type dualities considered in this paper, as well as the new ones that we generate through our methodology, arise from equivalences which are instances of the Comparison Lemma).
Let $\cal C$ be a preorder equipped with a coverage $J$, and let $\cal D$ be a subset of $\cal C$ (endowed with the induced order) which is \emph{$J$-dense}, i.e. such that for every $c\in {\cal C}$ there exists $R\in J(c)$ such that for every $b\in R$, $b\in {\cal D}$. We can define the \emph{induced coverage} $J|_{{\cal D}}$ on $\cal D$ as follows: for any $d\in {\cal D}$, $R\in J|_{{\cal D}}(d)$ if and only if $R=T\cap {\cal D}$ for some $T\in J(d)$.
\begin{theorem}
\begin{enumerate}[(i)]
\item Let $\cal C$ be a preorder equipped with a coverage $J$, and let $\cal D$ be a $J$-dense subset of $\cal C$. Then the map $\phi:Id_{J}({\cal C})\to Id_{J|_{{\cal D}}}({\cal D})$ sending a $J$-ideal $I$ on $\cal C$ to its intersection with $\cal D$ is a frame isomorphism, with inverse the map $\psi:Id_{J|_{{\cal D}}}({\cal D}) \to Id_{J}({\cal C})$ sending a subset in $Id_{J|_{{\cal D}}}({\cal D})$ to its $J$-closure in $\cal C$.
\item Any locale $L$ with a basis $B_{L}$ is isomorphic to the frame $Id_{J^{L}_{can}|_{B_{L}}}(B_{L})$ of $J_{can}|_{B_{L}}$-ideals on $B_{L}$, via the map
\[
\phi:L\to Id_{J^{L}_{can}|_{B_{L}}}(B_{L})
\]
sending an element $l\in L$ to the subset given by the intersection $B_{L}\cap (l)\downarrow$ and the map
\[
\psi:Id_{J^{L}_{can}|_{B_{L}}}(B_{L})\to L
\]
sending an ideal $I$ in $Id_{J^{L}_{can}|_{B_{L}}}$ to the supremum $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{l\in I}l}$ of $I$ in $L$.
\end{enumerate}
\end{theorem}
\begin{proofs}
$(i)$ It is immediate to verify that $\phi$ is a frame homomorphism with inverse $\psi$. Indeed, given a $J$-ideal $I$ on $\cal C$, the fact that $\cal D$ is $J$-dense ensures that $I$ is the $J$-closure of its intersection with $\cal D$; conversely, given a $J|_{{\cal D}}$-ideal $I'$ on $\cal D$, it is clear that $I'$ is equal to the intersection of $\cal D$ with its $J$-closure in $\cal C$.
$(ii)$ This can be deduced from part $(i)$ of the theorem by using the identification of $L$ with the frame of $J^{L}_{can}$-ideals on $L$. Anyway, we give a direct proof of this result for the reader's convenience.
Clearly, $\psi \circ \phi=1_{L}$. To prove that $\phi\circ \psi=1_{B_{L}}$, we have to verify that for any ideal $I$ in $Id_{J^{L}_{can}|_{B_{L}}}(B_{L})$, $I=B_{L}\cap (\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{}I})\downarrow$. The inclusion $\subseteq$ is obvious, so it remains to prove the converse one. Given $s\in B_{L}\cap (\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{}I})\downarrow$, $s=s\wedge \mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{}I}=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{a\in I}s\wedge a}$; but, $B_{L}$ being a basis of $L$, each $s\wedge a$ can be written as a join of elements of $B_{L}$ which belong to $I$ (they being less then or equal to $a$, which belongs to $I$). The fact that $I$ is a $J^{L}_{can}|_{B_{L}}$-ideal on $B_{L}$ then implies that $s$ belongs to $I$, as required.
\end{proofs}
The method of section \ref{dualtop} for `lifting' dualities or equivalences of categories of preorders with categories of frames (equivalently, with categories of locales) admits a natural frame-theoretic interpretation. Let us begin by specializing the notions of subterminal topology and the construction of the category of toposes paired with points of section \ref{subterminal} in the context of locales.
Recall that the points of the locale $L$ are the frame homomorphisms $L\to \{0,1\}$ from $L$ (regarded as a frame) to the two-element frame $\{0,1\}$.
The indexing functions $\xi:X\to \textsc{P}_{l}$ of a set of points $\textsc{P}_{l}$ of a locale $L$ by a set $X$ correspond precisely to the frame homomorphisms $L\to {\mathscr{P}}(X)$. Indeed, ${\mathscr{P}}(X)$ is the product in the category $\textbf{Frm}$ of frames of $X$-times the frame $\{0,1\}$; for any $x\in X$, we have a product projection $\overline{x}:{\mathscr{P}}(X)\to \{0,1\}$ sending a subset $S\in {\mathscr{P}}(X)$ to $1$ if $x\in S$ and to $0$ otherwise, and for any indexing function $\xi$ sending any element $x\in X$ to a frame homomorphism $\xi(x):L\to \{0,1\}$, there exists a unique frame homomorphism $g_{\xi}:L\to {\mathscr{P}}(X)$ such that for every $x\in X$, $\xi(x)=\overline{x}\circ g_{\xi}$, defined by the formula
\[
g_{\xi}(l)=\{x\in X \textrm{ | } \xi(x)(l)=1\}
\]
for any $l\in L$.
Given an indexing $\xi:X\to \textsc{P}$ of a set of points of a locale $L$, the construction of the subterminal topology on the set $X$ (cf. section \ref{subterminal}) reformulates as follows; the underlying set of the topological space is $X$, while the open sets are the subsets in the image of the frame homomorphism $g_{\xi}:L\to {\mathscr{P}}(X)$ corresponding to the indexing $\xi$ as specified above, in other words the subsets of the form $g_{\xi}(l)=\{x\in X \textrm{ | } \xi(x)(l)=1\}$ where $l$ ranges among the elements of $L$. We denote the topological space induced via this construction by an indexing $\xi:X\to \textsc{P}$ of a set of points of a locale $L$ by $X_{{\tau}^{L}_{\xi}}$.
The construction of the category of toposes paired with points of section \ref{subterminal} specializes, when restricted to the context of locales, to the following construction, of the category $\textbf{Loc}_{p}$ of \emph{locales paired with points}: the objects of $\textbf{Loc}_{p}$ are the pairs $(L, \xi)$ where $L$ is a locale and $\xi:X\to \textsc{P}$ is an indexing of a set of points $\textsc{P}$ of $L$, while the arrows $(L, \xi)\to (L', \xi')$ in $\textbf{Loc}_{p}$, where $\xi:X\to \textsc{P}$ and $\xi':Y\to \textsc{P}'$, are the pairs $(f, l)$ where $f$ is a frame homomorphism $f:L'\to L$ and $l:X\to Y$ is a function such that the diagram
\[
\xymatrix {
{\mathscr{P}}(Y) \ar[r]^{l^{-1}} & {\mathscr{P}}(X) \\
L' \ar[u]^{g_{\tilde{\xi'}}} \ar[r]^{f} & L \ar[u]^{g_{\tilde{\xi}}}}
\]
commutes.
Identities and composition in $\textbf{Loc}_{p}$ are defined in the obvious way.
Given an indexing $\xi:X\to \textsc{P}$ of a set of points of a locale $L$, we say that $\xi$ is \emph{separating} if for any $l,l'\in L$, $\xi(x)(l)=\xi(x)(l')$ for every $x\in X$ implies that $l=l'$.
In these terms, the method of \ref{dualtop} for `lifting' dualities or equivalences of categories of preorders with categories of frames reformulates as follows.
Let us first consider the contravariant case.
Let as assume to start with a duality $A:{\cal K}^{\textrm{op}}\to \textbf{Loc}$ obtained by the method of section \ref{locales}. Suppose that we have assigned to every structure $\cal C$ in $\cal K$ a separating indexing $\xi_{\cal C}:X_{\cal C}\to \textsc{P}_{\cal C}$ of a set of points $\textsc{P}_{\cal C}$ of the locale $Id_{J_{\cal C}}({\cal C})$, and to each arrow $f:{\cal C}\to {\cal D}$ in $\cal K$ a function $l_{f}:X_{\cal D} \to X_{\cal C}$ such that the pair $(A_{f}, l_{f})$ (see the statement of Theorem \ref{concretecontrov} above for the definition of the frame homomorphism $A_{f}:Id_{J_{\cal C}}({\cal C}) \to Id_{J_{\cal D}}({\cal D})$) defines an arrow $(Id_{J_{\cal D}}({\cal D}), \xi_{\cal D}) \to (Id_{J_{\cal C}}({\cal C}), \xi_{\cal C})$ in the category $\textbf{Loc}_{p}$ of locales paired with points. Then we have a functor $\tilde{A}:{\cal K}^{\textrm{op}}\to \textbf{Top}$ such that $\tilde{A}({\cal C})={X_{\cal C}}_{{{\tau}^{Id_{J_{\cal C}}({\cal C})}_{\xi_{\cal C}}}}$ for any ${\cal C}\in {\cal K}$ and $\tilde{A}(f)=l_{f}:X_{\cal D} \to X_{\cal C}$ for any arrow $f:{\cal C}\to {\cal D}$ in $\cal K$.
In the covariant case, let us assume to start with a functor $B:{\cal K} \to \textbf{Loc}$ obtained by the method of section \ref{locales}. Similarly as above, suppose that we have assigned to each structure $\cal C$ in $\cal K$ a separating indexing $\xi_{\cal C}:X_{\cal C}\to \textsc{P}_{\cal C}$ of a set $\textsc{P}_{\cal C}$ of points of the locale $Id({\cal C}^{\textrm{op}})$, and to each arrow $f:{\cal C}\to {\cal D}$ in $\cal K$ a function $l_{f}:X_{\cal C} \to X_{\cal D}$ such that the pair $(B_{f^{\textrm{op}}}, l_{f})$ (see the statement of Theorem \ref{concretecovariant} above for the definition of $B_{f^{\textrm{op}}}$) defines an arrow $(Id({\cal C}^{\textrm{op}}), \xi_{\cal C}) \to (Id({\cal D}^{\textrm{op}}), \xi_{\cal D})$ in the category $\textbf{Loc}_{p}$. Then we have a functor $\tilde{B}:{\cal K}\to \textbf{Top}$ such that $\tilde{B}({\cal C})={X_{\cal C}}_{{{\tau}^{Id({\cal C}^{\textrm{op}})}_{\xi_{\cal C}}}}$ for ${\cal C}\in {\cal K}$ and $\tilde{B}(f)=l_{f}:X_{\cal C} \to X_{\cal D}$ for any arrow $f:{\cal C}\to {\cal D}$ in $\cal K$.
We refer the reader to section \ref{dualtop} for a discussion of the properties of the functors $\tilde{A}$ and $\tilde{B}$.
Finally, let us give elementary proofs of two results in the paper concerning frames of ideals on preorders, namely Theorem \ref{freeframes} and \ref{bijectionpoints}.
Let us start with Theorem \ref{freeframes}. This result is based on the notion of filtering map, which we recall below.
\begin{definition}[cf. Proposition \ref{filtering}]
Let $\cal C$ be a preorder, $L$ be a frame and $f:{\cal C}\to L$ be a monotone map. We say that $f$ is \emph{filtering} if the following conditions hold:
\begin{enumerate}[(i)]
\item $1_{L}=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{c\in {\cal C}}}f(c)$;
\item For any $c,c'\in {\cal C}$, $f(c)\wedge f(c')=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{b\in B_{c, c'}}}f(b)$ where $B_{c, c'}$ is the set
\[
\{b\in {\cal C} \textrm{ | } b\leq c \textrm{ and } b\leq c'\}.
\]
\end{enumerate}
\end{definition}
Given a map $f:{\cal C}\to L$ and a coverage $J$ on $\cal C$, we say that $f$ is \emph{$J$-filtering} if $f$ is filtering and satisfies the property that for any $c\in {\cal C}$ and any $S\in J(c)$, $f(c)=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{d\in S}}f(d)$.
Let us now give a direct proof of Theorem \ref{freeframes}.
\begin{theorem}[cf. Theorem \ref{freeframes}]\label{freeframeselem}
Let $\cal C$ be a preorder and $J$ be a coverage on $\cal C$. Then the frame $Id_{J}({\cal C})$, together with the map $\eta:{\cal C}\to Id_{J}({\cal C})$ sending an element $c\in {\cal C}$ to the principal $J$-ideal $(c)\downarrow_{J}$, satisfies the following universal property: for any map $f:{\cal C}\to L$ to a frame $L$, $f$ is $J$-filtering if and only if there exists a (necessarily unique) frame homomorphism $\tilde{f}:Id_{J}({\cal C})\to L$ such that $\tilde{f}\circ \eta=f$ (given by the formula $\tilde{f}(I)=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{c\in I}}f(c)$ for any $I\in Id_{J}({\cal C})$).
\end{theorem}
\begin{proofs}
Given a frame homomorphism $g:Id_{J}({\cal C})\to L$, it is clear that the composite $g\circ \eta:{\cal C}\to L$ is a filtering map. Indeed, condition $(i)$ in the definition of filtering map follows from the fact that $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{c\in {\cal C}}}(g\circ \eta)(c)=g(\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{c\in {\cal C}}}(c)\downarrow_{J})=g({\cal C})=1_{L}$, while condition $(ii)$ follows from the fact that for any $c, c'\in {\cal C}$, $(g\circ \eta)(c) \wedge (g\circ \eta)(c')=g((c)\downarrow_{J} \cap (c')\downarrow_{J})=g(\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{b\in B_{c, c'}}}(b)\downarrow_{J}))=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{b\in B_{c, c'}}}g((b)\downarrow_{J})=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{b\in B_{c, c'}}}(g\circ \eta)(b)$. The fact that $g\circ \eta$ satisfies the property that for any $c\in {\cal C}$ and any $S\in J(c)$, $(g\circ \eta)(c)=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{d\in S}}(g\circ \eta)(d)$ is obvious.
Conversely, given a filtering map $f:{\cal C}\to L$, there is exactly one frame homomorphism $\tilde{f}:Id_{J}({\cal C})\to L$ such that $\tilde{f}\circ \eta=f$. Indeed, $\tilde{f}$ is forced (by the fact that it must preserve arbitrary joins) to be equal to the map sending an ideal $I$ in $Id_{J}({\cal C})$ to the join $\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{c\in I}}f(c)$. The fact that $\tilde{f}$ preserves the top element (resp. binary meets) easily follows from condition $(i)$ (resp. condition $(ii)$) in the definition of filtering map, while the fact that $\tilde{f}$ preserves arbitrary joins follows from the fact that for any $c\in {\cal C}$ and any $S\in J(c)$, $f(c)=\mathbin{\mathop{\textrm{\huge $\vee$}}\limits_{d\in S}}f(d)$.
\end{proofs}
Let us now turn our attention to Theorem \ref{bijectionpoints}. To this end, we recall from section \ref{exsub} the notion of $J$-prime filter on a preorder $\cal C$ equipped with a coverage $J$: a \emph{$J$-prime filter} on $\cal C$ is a subset $F\subseteq {\cal C}$ such that $F$ is non-empty, $a\in F$ implies $b\in F$ whenever $a\leq b$ in $\cal C$, for any $a, b\in F$ there exists $c\in F$ such that $c\leq a$ and $c\leq b$, and for any $J$-covering sieve $\{a_{i} \to a \textrm{ | } i\in I\}$ in $\cal C$ if $a\in F$ then there exists $i\in I$ such that $a_{i}\in F$. If $\cal C$ is a frame and $J$ is the canonical coverage on $\cal C$, the notion of $J$-prime filter on $\cal C$ specializes to the notion of \emph{completely prime filter} on $\cal C$.
Theorem \ref{bijectionpoints} asserts that the assignment sending a filter $F$ on $Id_{J}({\cal C})$ to the $J$-prime filter $\{c\in {\cal C} \textrm{ | } (c)\downarrow_{J}\in F\}$ on $\cal C$ defines a bijection between the completely prime filters on the frame $Id_{J}({\cal C})$ of $J$-ideals of $\cal C$ and the $J$-prime filters on $\cal C$. This result can be deduced as the particular case of Theorem \ref{freeframeselem} for $L=\{0,1\}$. Indeed, the $J$-filtering maps ${\cal C} \to \{0,1\}$ correspond exactly to the $J$-prime filters on $\cal C$, while the frame homomorphisms $Id_{J}({\cal C})\to \{0,1\}$ correspond exactly to the completely prime filters on $Id_{J}({\cal C})$.
\vspace{0.5cm}
\textbf{Acknowledgements:} I am grateful to Peter Johnstone for useful comments on a preliminary version of this paper.
|
\section{Introduction}
$N=8$ supergravity \cite{Cremmer:1979up} is undoubtedly a highly distinguished
field theory due to its high degree of symmetry and the remarkable structure of
its amplitudes that has emerged in recent work, see e.g.~\cite{ArkaniHamed:2008gz,Bern:2009kd}.
The continuous ${E}_{7(7)}$ symmetry underlying the classical field equations
has important consequences for the structure of the counterterms~\cite{Kallosh:2008rr,Brodel:2009hu,Elvang:2010kc,Bossard:2010dq,Beisert:2010jx}.
The field content of maximal supergravity is the unique $N=8$ supermultiplet
\begin{eqnarray}
\begin{tabular}{c|rrrrrrrrr}
helicity&$-2$ & $-\frac32$ & $-1$ & $-\frac12$ & $\phantom{-}0$ & $+\frac12$ &$+1$& $+\frac32$ &$+2$ \\\hline
d.o.f. &1&8&28&56&70&56&28&8&1 \\[1ex]
\end{tabular}
\label{tbl:multiplet}
\quad
\;.
\end{eqnarray}
On the other hand, the mutual couplings of the various fields are not uniquely determined,
as supersymmetry allows for the introduction of particular (non-)abelian charges and
the realization of different (non-)abelian gauge groups.
After the original version of the theory with abelian gauge fields \cite{Cremmer:1979up}
the first maximal gauged supergravity was constructed in~\cite{deWit:1982ig}
with the 28 vector fields gauging a compact $SO(8)$ subgroup of ${E}_{7(7)}$.
Non-compact versions of this theory have been constructed and classified
in~\cite{Hull:1984qz,Cordaro:1998tx}
and later been extended to other
non-semisimple gauge groups in~\cite{Andrianopoli:2002mf,Hull:2002cv}.
A general formalism for describing the gauging of subgroups in terms of an
`embedding tensor' has been established in~\cite{deWit:2002vt,deWit:2007mt}.
This constant tensor describes the embedding of the gauge group into the global ${E}_{7(7)}$ symmetry
of the ungauged theory, and parametrizes all the couplings of the gauged theory.
The aim of this paper is the construction of all possible gaugings
(and thus all possible couplings) of $N=8$ supergravity, which in particular
include a gauging of the global scaling symmetry of the theory.
Their gauge groups are embedded in the product $E_{7(7)} \times \mathbb{R}$ of the Cremmer-Julia
group $E_{7(7)}$ with the one-parameter scaling symmetry of the theory that generalizes the Weyl rescaling of general relativity and has been dubbed a `trombone' symmetry of supergravity~\cite{Cremmer:1997xj}.
Supergravity theories that include a gauging of their scaling symmetry have first been constructed
in ten dimensions by a generalized Scherk-Schwarz reduction from eleven dimensions~\cite{Howe:1997qt,Lavrinenko:1997qa}. Lower dimensional examples of such theories include \cite{Bergshoeff:2002nv} and \cite{Kerimo:2003am,Kerimo:2004md}. As a generic feature, these theories are invariant under local rescaling of the fields (including the metric) with appropriate weights upon a compensating gauge transformation on the matter fields. They do not possess an action (since they result from the gauging of an on-shell symmetry) and typically support de Sitter geometries rather than Minkowski or AdS vacua. A systematic account to the construction of these theories has been put forward in~\cite{LeDiffon:2008sh}. Based on the algebraic structure of the duality groups of the ungauged theories, the representation content and the algebraic consistency constraints for the corresponding embedding tensor have been determined for the maximal supergravities. In \cite{Riccioni:2010xx}, these structures have been shown to be naturally embedded in the framework of the very-extended infinite-dimensional Kac-Moody algebra $E_{11}$~\cite{Riccioni:2007au,Bergshoeff:2007qi}.
The general analysis reveals that four-dimensional $N=8$ supergravity admits an embedding tensor transforming in the representation ${\bf 56}+{\bf 912}$ of $E_{7(7)}$, subject to a set of bilinear algebraic consistency constraints. Gaugings defined by an embedding tensor in the irreducible ${\bf 912}$ representation describe gauge groups that entirely reside within $E_{7(7)}$ and have been constructed in \cite{deWit:2002vt,deWit:2007mt}. Additional non-vanishing components in the ${\bf 56}$ representation on the other hand define gaugings that include the trombone generator, i.e.\ theories in which local scaling invariance is part of the gauge group. These are the theories to be constructed in this paper. While the analysis of~\cite{LeDiffon:2008sh} has been purely algebraic and based on the structure of non-abelian deformations of the underlying tensor gauge algebras, it is the aim of this paper to explicitly realize these theories by constructing the full set of supersymmetric field equations. Thereby we derive the most general couplings that are compatible with $N=8$ supersymmetry in four dimensions. In particular, we confirm that the algebraic consistency constraints derived in~\cite{LeDiffon:2008sh} for the embedding tensor are indeed sufficient to ensure supersymmetry of the field equations.
This paper is organized as follows. In section~\ref{sec:gauging}, we analyze the general structure of the gauge groups induced by an embedding tensor in the ${\bf 56}+{\bf 912}$ representation. We explicitly construct the gauge group generators in terms of the embedding tensor and discuss the system of bilinear algebraic consistency constraints that the embedding tensor must satisfy. In case the gauge group includes the trombone generator, this system of constraints drastically reduces upon decomposing the embedding tensor into its $E_{6(6)}$ irreducible components, and we present a number of explicit solutions. We compute the Cartan-Killing metric of the gauge group and show that gaugings involving the local scaling symmetry are generically dyonic, i.e.\ genuinely involve electric and magnetic vector fields.
In section~\ref{sec:coset}, we review the structure of the scalar target space $E_{7(7)}/SU(8)$ and define the $T$ tensor in terms of which the couplings of the gauged theory are expressed.
Subsequently, in section~\ref{sec:susy}, we determine the modified supersymmetry transformations by verifying the closure of their algebra on the bosonic fields of the theory. Based on these results, in section~\ref{sec:eom}, we obtain the modified field equations of the gauged $N=8$ supergravity by starting from an ansatz for the fermionic field equations and calculating their transformation under supersymmetry. This allows to uniquely determine the full set of field equations in lowest order of the fermions. As a particular feature of these theories, we find that gauging of the trombone generator leads to an additional positive contribution to the effective cosmological constant. In section~\ref{subsec:masses}, we determine the conditions for extremality, i.e.\ for solutions of the field equations with constant scalar and gauge fields and give explicit formulas for the mass matrices by linearizing the field equations around these solutions. Finally, in section~\ref{sec:example}, we present a simple example of a theory with local scaling symmetry that has its higher-dimensional origin as a generalized Scherk-Schwarz reduction from five dimensions upon twisting the field with a linear combination of an $E_{6(6)}$ generator and the five-dimensional trombone symmetry. We show that this theory admits a de Sitter solution with constant scalar fields and determine its mass spectrum which seems to indicate that the solution is not stable. We conclude with an outlook on the role and the applications of these theories.
\section{Structure of gauge groups}
\label{sec:gauging}
Before explicitly constructing the full supersymmetric field equations,
in this section we will present and analyze the structure of the possible gauge algebras
that can be realized as local symmetry in maximally supersymmetric supergravity in four dimensions.
Recall, that the global symmetry group
of the ungauged maximally supersymmetric theory is given by~\cite{Cremmer:1979up}
\begin{eqnarray}
G &=& E_{7(7)} \times \mathbb{R}
\;,
\label{CJ}
\end{eqnarray}
where the second factor corresponds to the scaling (or trombone) symmetry of the equations of motion,
under which the fields transform as
\begin{eqnarray}
\delta g_{\mu\nu} &=& 2\,g_{\mu\nu} \;,\qquad
\delta {\cal A}^M_{\mu} ~=~ {\cal A}^M_{\mu} \;,\qquad
\delta \phi^i ~=~ 0
\;,
\nonumber\\
\delta \psi_\mu &=& \ft12\,\psi_\mu \;,\qquad
\delta\chi ~=~ -\ft12\,\chi
\;.
\label{scaling}
\end{eqnarray}
Here, the first line refers to the bosonic fields of spin 2, 1, and 0, while the second line gives the transformations of the spin 3/2 gravitons and the spin 1/2 matter fermions, respectively.
The $E_{7(7)}$ factor in (\ref{CJ}) in contrast only acts on vector and scalar fields, with its generators $t_\alpha$
closing into the algebra
\begin{eqnarray}
[t_\alpha, t_\beta] &=& f_{\alpha\beta}{}^\gamma\,t_\gamma
\;.
\label{E7algebra}
\end{eqnarray}
The 28 electric vector fields $A_\mu^\Lambda$ combine with their magnetic duals $A_{\mu\,\Lambda}$ into the fundamental 56-dimensional representation $A_\mu^M$ of $E_{7(7)}$ while the 70 scalar fields transform in a non-linear representation parametrizing the coset space $E_{7(7)}/SU(8)$\,.
General gaugings will also require the introduction of two-form tensor fields $B_{\mu\nu\,\alpha}$ transforming in the adjoint 133-dimensional representation of $E_{7(7)}$.
\subsection{Gauge group generators}
In this paper, we will construct the most general supersymmetric theories in which a subgroup of (\ref{CJ}) is gauged. Extending previous work~\cite{deWit:2007mt}, we will consider those theories in which
the gauge groups include the scaling symmetry, i.e.\ the second factor in (\ref{CJ}).
Let us denote by ${\bf t}_{0}$ the generator of the scaling symmetry $\mathbb{R}$, and by
${\cal A}_\mu \equiv \vartheta_M {\cal A}^M_\mu$ the linear combination of vector fields
that will be used to gauge this symmetry upon introduction of covariant derivatives.
As this symmetry also acts in the gravitational sector by scaling the metric, its gauging necessitates
a modification of the spin connection $\omega_{\mu}{}^{ab}$ and the Riemann tensor ${R}_{\mu \nu}{}^{ ab}$ according to~\cite{LeDiffon:2008sh}
\begin{eqnarray}
\widehat\omega_{\mu}{}^{ab} &=&\omega_{\mu}{}^{ab}
+2 \,e_\mu{}^{[a}\,\,e^{ b]\, \nu}{\cal A}_{\nu}
\;,
\nonumber\\[1ex]
\widehat{{\cal R}}_{\mu \nu}{}^{ab} &\equiv&
2\,\partial_{[\mu}\,\widehat\omega_{\nu]}{}^{ab} +
2\, \widehat\omega_{[\mu}{}^{ac}\:\widehat\omega_{\nu]c}{}^b
\nonumber\\[1ex]
&=& {R}_{\mu \nu}{}^{ ab}
- 4 \, e_{[\mu}{}^{[a} \nabla(\omega)_{\nu]}{\cal A}^{b]} + 4 \, e_{[\mu}{}^{[a} {\cal A}_{\nu]} {\cal A}^{b]}
- 2 \, e_{[\mu}{}^{a} e_{\nu]}{}^{b} {\cal A}_{\lambda} {\cal A}^{\lambda}
\;,
\label{modRR}
\end{eqnarray}
which are invariant under the joint transformation
\begin{eqnarray}
\delta g_{\mu\nu} &=& 2 \lambda(x) \,g_{\mu\nu}\;,
\qquad \delta {\cal A}_\mu ~=~ \partial_\mu \lambda(x)
\;.
\end{eqnarray}
Besides, they satisfy the generalized Bianchi identities
\begin{eqnarray}
\widehat{{\cal R}}_{[\mu \nu\rho]}{}^{a}
&=&
{\cal F}_{[\mu\nu}\,e_{\rho]}{}^{a}
\;.
\label{BianchiRR}
\end{eqnarray}
The most general gauging combines this symmetry with some subgroup of the $E_{7(7)}$ factor of (\ref{CJ}).
As shown in \cite{deWit:2002vt,deWit:2007mt} and \cite{LeDiffon:2008sh} for the cases without and
with the trombone factor, respectively, the parametrization of the general gauge group generators
$X_M$ allows for $56+912$ parameters spanning the `embedding tensor' $\vartheta_M$ and $\Theta_M{}^\alpha$, according to
\begin{eqnarray}
X_M &\equiv&
\vartheta_M\,{\bf t}_{0} +
\left(\Theta_M{}^\alpha + 8 \vartheta_N\,(t^\alpha)_M{}^N\right)\,t_{\alpha}
\;,
\label{generators}
\end{eqnarray}
with covariant derivatives given by $D_\mu \equiv \partial_\mu - {\cal A}^M_\mu X_M$\,.\footnote{
For transparency we have suppressed explicit coupling constants, which can at any stage
be reintroduced by rescaling $\vartheta_M \rightarrow g \vartheta_M$,\,
$\Theta_M{}^\alpha \rightarrow g \Theta_M{}^\alpha$\,.
}
Here, $(t_\alpha)_M{}^N$ are the $E_{7(7)}$ generators (\ref{E7algebra}) in the fundamental
representation,\footnote{
We raise and lower adjoint indices
with the invariant metric $\kappa_{\alpha\beta}\equiv {\rm Tr}\,[t_{\alpha}t_{\beta}]$.
Fundamental indices are raised and lowered with
the symplectic matrix $\Omega^{MN}$ using north-west south-east conventions:
$X^M=\Omega^{MN}X_N$, etc.\,.}
and the matrix $\Theta_M{}^\alpha$ is constrained by the relations
\begin{eqnarray}
\Theta_M{}^\alpha (t_\alpha)_N{}^M = 0\;,\qquad
\Theta_M{}^\alpha=-2(t_\beta\, t^\alpha)_M{}^N\,\Theta_N{}^\beta
\;,
\label{linear}
\end{eqnarray}
i.e.\ transforms in the ${\bf 912}$ representation of $E_{7(7)}$.
In absence of $\vartheta_M$, it describes the
gaugings whose gauge group entirely resides within the $E_{7(7)}$.
The relative factors in (\ref{generators}) are chosen such that
the tensor $Z^K{}_{MN} \equiv (X_{(M})_{N)}{}^K$
factors according to
\begin{eqnarray}
Z^K{}_{MN} &\equiv& (X_{(M})_{N)}{}^K ~=~
-\ft12 (t_\alpha)_{MN}\left( \Theta^{K\alpha}-16 (t^\alpha)^{KL}\vartheta_L \right)
~\equiv~ (t_\alpha)_{MN}\,Z^{K\alpha}
\;,
\label{defZZ}
\end{eqnarray}
and thus projects onto the ${\bf 133}$ representation in its indices $(MN)$.
This will be a central identity in the construction.
For convenience, we also define the projection
of the gauge group generators onto the $E_{7(7)}$ factor of (\ref{CJ}) as
\begin{eqnarray}
\check{X}_M &\equiv&
\left(\Theta_M{}^\alpha + 8 \vartheta_N\,(t^\alpha)_M{}^N\right)\,t_{\alpha}
\;,
\label{generatorsE7}
\end{eqnarray}
The gauged theory is invariant under the local symmetry
\begin{eqnarray}
\delta_\Lambda \vec\phi &=& \Lambda^M X_M \cdot \vec\phi
~\equiv~ \left(\Theta_M{}^\alpha + 8 \vartheta_N\,(t^\alpha)_M{}^N\right) \vec K_\alpha(\phi)
\;,
\nonumber\\
\delta_\Lambda {\cal A}_\mu^M &=& D_\mu \Lambda^M ~\equiv~ \partial_\mu \Lambda^M
+ {\cal A}^K_\mu (X_{K})_{N}{}^M\,\Lambda^N
\;,
\label{gaugeV1}
\end{eqnarray}
where $\vec K_\alpha(\phi)$ represent the $E_{7(7)}$ Killing vector fields on the scalar target space, and
the gauge group generators are given by evaluating (\ref{generators}) in the appropriate representation, i.e.\
\begin{eqnarray}
(X_{K})_{N}{}^M &\equiv&
- \vartheta_K \delta_N^M
+ \left(\Theta_K{}^\alpha + 8 \vartheta_L\,(t^\alpha)_K{}^L\right)\,(t_{\alpha})_N{}^M
\;.
\label{defX}
\end{eqnarray}
Finally, covariant field strengths are defined by
\begin{eqnarray}
{\cal H}^M_{\mu\nu} &\equiv&
2\partial_{[\mu} {\cal A}_{\nu]}^{M}
+ (X_{N})_{P}{}^{M}
\,{\cal A}_{[\mu}^{N} {\cal A}_{\nu]}^{P}
+ {Z}^{M\alpha}\,B_{\mu\nu\,\alpha}
\;,
\label{defH}
\end{eqnarray}
with a St\"uckelberg-type coupling to the two-forms $B_{\mu\nu\,\alpha}$ and
the (constant) intertwining tensor ${Z}^{M\alpha}$ defined in (\ref{defZZ}).
They transform covariantly under the gauge transformations (\ref{gaugeV1})
provided the two-forms transform as
\begin{eqnarray}
\delta_\Lambda B_{\mu\nu\,\alpha} &=&
-2(t_\alpha)_{MN}\,\left(\Lambda^M {\cal H}^N_{\mu\nu}
- {\cal A}_{[\mu}^M\, \delta {\cal A}_{\nu]}^N\right)
\;.
\label{gaugeV2}
\end{eqnarray}
Moreover, the covariant field strengths (\ref{defH}) are invariant under the tensor
gauge transformations\footnote{
W.r.t.\ reference~\cite{deWit:2007mt} we have rescaled the two-form fields and associated tensors as
$B_{\mu\nu\alpha}\rightarrow-B_{\mu\nu\alpha}$,
$\Xi_{\mu\alpha}\rightarrow-\Xi_{\mu\alpha}$,
$Z^{M\alpha} \rightarrow -Z^{M\alpha}$.}
\begin{eqnarray}
\delta_\Xi B_{\mu\nu\,\alpha} &=&
2{\cal D}_{[\mu} \Xi_{\nu]\,\alpha} + 2(t_\alpha)_{MN}\,{\cal A}_{[\mu}^M\, \delta {\cal A}_{\nu]}^N
\;,\nonumber\\
\delta_\Xi {\cal A}_\mu^M &=& - Z^{M\alpha}\, \Xi_{\mu\,\alpha}
\;.
\label{gaugeT}
\end{eqnarray}
The covariant field strengths (\ref{defH}) satisfy the
generalized Bianchi identities
\begin{eqnarray}
{\cal D}^{\vphantom{M}}_{[\mu}{\cal H}^M_{\nu\rho]} &=&
\ft13 {Z}^{M\alpha}\,{\cal H}_{\mu\nu\rho\,\alpha}
\label{Bianchi}
\;,
\end{eqnarray}
with the covariant non-abelian field strength ${\cal H}_{\mu\nu\rho\,\alpha}$
of the two-form tensor fields, given by:
\begin{eqnarray}
{\cal H}_{\mu\nu\rho\,\alpha}&=&3 D_{[\mu}B_{\nu\rho]\alpha}+6\,t_{\alpha PQ}\,A^P_{[\mu}\left(\partial_\nu A^Q_{\rho]}+\frac{1}{3}\,X_{RS}{}^Q\,A^R_\nu A^S_{\rho]}\right)\,,
\end{eqnarray}
where
\begin{eqnarray}
D_{[\mu}B_{\nu\rho]\alpha}&=&\partial_{[\mu}B_{\nu\rho]\alpha}+2\,t_{\alpha PQ} Z^{Q\beta}A^P_{[\mu}B_{\nu\rho]\beta}\,.
\end{eqnarray}
\subsection{Consistency constraints}
\label{concon}
The previous construction leads to a consistent (closed) gauge algebra, if the irreducible components
$\vartheta_M$, $\Theta_M{}^\alpha$ satisfy the following system of quadratic constraints~\cite{LeDiffon:2008sh}
\begin{eqnarray}
\vartheta_{M}\, \Theta^M{}^\alpha
&\stackrel!{\equiv}&16\, (t^\alpha)^{MN} \,\vartheta_{ M}\, \vartheta_{ N}
\;,
\label{Q1}\\
(t_{\gamma})_{[M}{}^{ P}\, \Theta_{N]}{}^\gamma\,\vartheta_{ P}
&\stackrel!{\equiv}&0
\;,
\label{Q2}\\
\Theta_M{}^\alpha\,\Theta^M{}^\beta &\stackrel!{\equiv}&
8 \,\vartheta_M\,\Theta_N{}^{[\alpha}\,t^{\beta]}{}^{MN}
-4\,f^{\alpha\beta}{}_{\gamma}\,\vartheta_M\,\Theta^M{}^\gamma
\;,
\label{Q3}
\end{eqnarray}
transforming in the ${\bf 133}$, the ${\bf 1539}$ and the ${\bf 133}+{\bf 8645}$,
of $E_{7(7)}$, respectively.
For $\vartheta_M=0$ they consistently reduce to the quadratic conditions of~\cite{deWit:2007mt}.
As we will show in the following, any solution to the constraints
(\ref{Q1})--(\ref{Q3}), will define a viable maximally supersymmetric gauged supergravity.
It is straightforward to show that (\ref{Q1})--(\ref{Q3}) imply several direct consequences
for the gauge group generators, such as the closure of the gauge algebra according to
\begin{eqnarray}
{}[X_M,X_N] &=& -X_{MN}{}^K\,X_K \;,
\label{XXX}
\end{eqnarray}
and orthogonality between gauge group generators and the
intertwining tensor $Z$
\begin{eqnarray}
X_{MN}{}^K\,Z^{M\alpha} ~=~ 0 ~=~
\vartheta_M\,Z^{M\alpha}
\;.
\label{QQZ}
\end{eqnarray}
The reason for the fact that the gauge transformations consistently
close into an algebra when properly extended to the two-form tensor fields
even in presence of the gauging of the scaling symmetry
is the underlying structure of a hierarchy of non-abelian tensor
gauge transformations~\cite{deWit:2005hv,deWit:2008ta} which is not based on the
existence of an action. The relative factors in (\ref{generators}) and the identity (\ref{defZZ})
are central in this construction.
What we will show explicitly in this paper is that the non-abelian deformations
defined in the previous section are precisely the ones that are moreover compatible with maximal
supersymmetry of the field equations.
\subsection{Solution to the quadratic constraints}
\label{sttqc}
In general, it is a hard task to construct solutions to the quadratic constraints of gauged supergravity.
However, it turns out that in presence of the trombone (i.e.\ non-vanishing $\vartheta_M$), the system
(\ref{Q1})--(\ref{Q3}) can be reduced to a much simpler one in terms of a reduced number of
components. The strategy for solving the quadratic constraints follows the case of pure trombone
gaugings~\cite{LeDiffon:2008sh} by decomposing all objects with respect to the $E_{6(6)}\times SO(1,1)$
subgroup of $E_{7(7)}$.
Explicitly, this means that the adjoint representation branches as
\begin{eqnarray}
t_\alpha &\rightarrow& (t_{\rm o},\; t_a,\; t_m,\; t^m)
\;,\nonumber\\
{\rm according~to}\;\;
{\bf 133} &\rightarrow& {\bf 1}^{0}+{\bf 78}^{0}+{\bf 27}^{-2}+\bar{\bf 27}^{+2}
\;,
\end{eqnarray}
while the fundamental representation breaks into
\begin{eqnarray}
\vartheta_M&\rightarrow~&
(\vartheta_\bullet,\;
\vartheta_m,\;
\vartheta^m,\;
\vartheta^\bullet)
\;,
\nonumber\\
{\rm according~to}\;\;
{\bf 56} &\rightarrow& {\bf 1}^{+3}+{\bf 27}^{+1}+\bar{\bf 27}^{-1}+{\bf 1}^{-3}
\;,
\end{eqnarray}
and the embedding tensor $\Theta_M{}^\alpha$ decomposes into
\begin{eqnarray}
\Theta_M{}^\alpha &\rightarrow~&
\left( \xi_+^{a},\;
\xi_m,\;
\xi^{mn},\;
\xi_{mn},\;
\xi^m,\;
\xi_-^{a}
\right)
\;,
\nonumber\\
{\rm according~to}\;\;
{\bf 912} &\rightarrow&
{\bf 78}^{+3}+{\bf 27}^{+1}+\overline{\bf 351}^{+1}
+{\bf 351}^{-1}+\bar{\bf 27}^{-1}+{\bf 78}^{-3}
\;,
\end{eqnarray}
with its explicit $(56\times 133)$ matrix form given in (\ref{ThetaBreak}) in the appendix.
We use indices $a, b, \dots = 1, \dots, 78$
and $m, n, \dots = 1, \dots, 27$ to label the adjoint and the fundamental representation of $E_{6(6)}$,
respectively.
In appendix~\ref{app:solve} we derive an important result: for non-vanishing $\vartheta_M$ and
up to $E_{7(7)}$ rotations, the general
solution to the system (\ref{Q1})--(\ref{Q3}) is parametrized by a real constant $\kappa$, an $E_{6(6)}$ matrix
$\Xi_m{}^n \equiv \Xi^a (t_a)_m^n$ and two real tensors $\zeta^m$, $\zeta^{[mn]}$, as follows
\begin{eqnarray}
(\vartheta_\bullet,\vartheta_n,\vartheta^n,\vartheta^\bullet)&=&
(\kappa,0,\kappa\zeta^n,0)
\;,
\nonumber\\
\left( \xi_+^{a},\;
\xi_m,\;
\xi^{mn},\;
\xi_{mn},\;
\xi^m,\;
\xi_-^{a}
\right)&=&
\left(\Xi^a, 0,\zeta^{mn}, \Xi_{[m}{}^k d_{n]kl} \zeta^l ,-\ft43 \kappa \zeta^m, 0\right)
\;,
\label{solutionemb}
\end{eqnarray}
where $d_{kmn}$ denotes the totally symmetric $E_{6(6)}$ invariant tensor.
The tensors $\zeta^m$, $\zeta^{[mn]}$ must be real eigenvectors under the action of $\Xi$ according to
\begin{eqnarray}
\delta_\Xi \zeta^m &\equiv& - \Xi_n{}^m \zeta^n ~\stackrel!{\equiv}~ \ft43\kappa \zeta^m \;,\nonumber\\
\delta_\Xi \zeta^{mn} &\equiv& 2\Xi_k{}^{[m} \zeta^{n]k} ~\stackrel!{\equiv}~ \ft23\kappa \zeta^{mn}
\;,
\label{remcon}
\end{eqnarray}
which furthermore must satisfy the following set of polynomial constraints:
\begin{eqnarray}
\zeta^k\zeta^l d_{mkl} &\stackrel!{\equiv}& 0
\;,\label{cs1}\\
\zeta^{k} \zeta^{mn} d_{kml} &\stackrel!{\equiv}& 0\;,
\label{cs2}\\
\zeta^{[k} \zeta^{mn]} &\stackrel!{\equiv}& 0 \;,
\label{cs3}\\
\left(t_a \cdot (\Xi+ \ft43 \kappa I) \cdot (\Xi-\ft23 \kappa I)\right){}\!_n{}^m \,\zeta^n &\stackrel!{\equiv}&
-\ft12 \zeta^{mk} \zeta^{ln} d_{klp} (t_a)_n{}^p
\;.
\label{cs4}
\end{eqnarray}
As we show in appendix~\ref{app:solve}, this system of equations
is equivalent to the original system of constraints~(\ref{Q1})--(\ref{Q3}).
In contrast to the original system, solutions to (\ref{remcon})--(\ref{cs4}) may easily be constructed.
A simple solution to the system (\ref{remcon})--(\ref{cs4}) is given by setting $\zeta^m=0=\zeta^{mn}$\,.
This leaves a non-trivial embedding tensor (\ref{solutionemb}) parametrized by $\kappa$
and an $E_{6(6)}$
generator $\Xi$.
This solution satisfies a stronger version of the quadratic constraints:
left and right hand sides of equations (\ref{Q1})--(\ref{Q3}) vanish separately.
In the limit $\kappa\rightarrow0$ in which $\vartheta_M$ vanishes,
this solution corresponds to the
known gaugings induced by a Scherk-Schwarz reduction~\cite{Scherk:1979zr}
from five dimensions parametrized by the choice of
an $E_{6(6)}$ generator~\cite{Andrianopoli:2002mf}.
For non-vanishing $\kappa$, the higher-dimensional origin of these
theories is a generalized Scherk-Schwarz reduction
from five dimensions in which the fields are twisted by a linear combination of the $E_{6(6)}$
generator $\Xi$ and
the five-dimensional trombone symmetry. The form of the generators (\ref{generators})
shows that even for vanishing $\Xi=0$,
switching on $\kappa$ corresponds to gauging a linear combination of the four-dimensional
trombone generator ${\bf t}_0$ and a subset of $E_{7(7)}$ generators.
More complicated solutions of the constraints involve
non-vanishing zero-modes $\zeta^m,\,\zeta^{mn}$.
While we defer the complete solution of the constraint
system (\ref{remcon})--(\ref{cs4}) to a separate publication, a typical example of such a solution will be
discussed in Section \ref{awoe}.
\subsection{Invariants of the trombone}
We can classify the inequivalent gaugings according to the ${\rm
E}_{7(7)}$-invariants constructed out of the embedding tensor. In
particular, the quadratic constraints can be regarded as conditions
on the ${\rm E}_{7(7)}$-orbits of the embedding tensor. In terms of
$\vartheta_M$ and $\Theta_M{}^\alpha$, several ${\rm
E}_{7(7)}$-invariants can be constructed of which the simplest is
the quartic invariant $I_4(\vartheta)$ depending only on the
trombone component $\vartheta_M$ according to
\begin{eqnarray}
I_4(\vartheta)&\equiv& -2\,(t_{\alpha})^{MN}(t^\alpha){}^{PQ}\:
\vartheta_M\,\vartheta_N\,\vartheta_P\,\vartheta_Q\nonumber\\[1ex]
&=&
-(\vartheta_\bullet\,\vartheta^\bullet+\vartheta_m\,\vartheta^m)^2+10
d_{mnp}\,d^{mrs}\vartheta_r\vartheta_s\,\vartheta^n\,\vartheta^p
\nonumber\\
&&{} -\ft{20}{3}\,
\vartheta^\bullet\,d^{mnp}\,\vartheta_m\vartheta_n\vartheta_p+\ft{2}{3}\,
\,\vartheta_\bullet\,d_{mnp}\,\vartheta^m\vartheta^n\vartheta^p\,.
\label{I4}
\end{eqnarray}
The different orbits of the 56-dimensional fundamental
representation of $E_{7(7)}$ are characterized via this invariant
as~\cite{Ferrara:1997uz}:
\begin{itemize}
\item[$(i)$]
{$I_4(\vartheta)>0$: the orbit is $\frac{{\rm E}_{7(7)}}{{\rm
E}_{6(2)}}$\,;}
\item[$(ii)$]
{$I_4(\vartheta)<0$: the orbit is $\frac{{\rm E}_{7(7)}}{{\rm
E}_{6(6)}}$\,;}
\item[$(iii)$]
{$I_4(\vartheta)=0,\,\frac{\partial I_4(\vartheta)}{\partial
\vartheta_M}\neq 0$:
the orbit is $\frac{{\rm E}_{7(7)}}{{\rm F}_{4(4)}\ltimes T_{26}}$\,;}
\item[$(iv)$]
{$I_4(\vartheta)=0,\,\frac{\partial I_4(\vartheta)}{\partial \vartheta_M}= 0,\,
t_{\alpha\,MN}\vartheta^M\,\vartheta^N\neq0$:
the orbit is $\frac{{\rm E}_{7(7)}}{{\rm SO}(6,5)\ltimes (T_{32}\times T_1)}$\,;}
\item[$(v)$]
{$ t_{\alpha\,MN}\vartheta^M\,\vartheta^N=0$:
the orbit is $\frac{{\rm E}_{7(7)}}{{\rm E}_{6(6)}\ltimes T_{27}}$\,.}
\end{itemize}
Inserting the solution (\ref{solutionemb}) obtained in the previous section
into (\ref{I4}), we find
\begin{eqnarray}
I_4(\vartheta)&=&\frac{2}{3}
\,\kappa^4\,d_{mnp}\,\zeta^m\zeta^n\zeta^p\,.
\end{eqnarray}
From (\ref{cs1}) it follows that
$I_4(\vartheta)=0=\,\frac{\partial I_4(\vartheta)}{\partial
\vartheta_M}$. Since $t_{\alpha\,MN}\vartheta^M\vartheta^N$ has a
non-vanishing component $(t^m)^\bullet{}_n\,\vartheta_\bullet
\vartheta^n\propto \kappa^2\,\zeta^m$, we conclude that
$\vartheta_M$ belongs to the orbit~$(iv)$ of the above classification if $\zeta^k$ is
non-vanishing, and otherwise to the orbit~$(v)$.\par In both cases
the gauge group $G_g$ will then be a subgroup of the stability
group of the corresponding orbit inside $\mathbb{R}^+\times{\rm
E}_{7(7)}$. In case $(iv)$, for instance, we should have
\begin{eqnarray}
G_g&\subset&\left[\mathbb{R}^+\times{\rm SO}(6,5)\right]\ltimes
(T_{32}\times T_1)\,,
\end{eqnarray}
where $\mathbb{R}^+$ is a suitable combination of the trombone
symmetry and the ${\rm O}(1,1)^7$ symmetry inside ${\rm E}_{7(7)}$.
\subsection{An explicit example}\label{awoe}
Here we present an example of a solution of the constraints
(\ref{remcon})--(\ref{cs4}).
The quadratic condition (\ref{cs1}) on $\zeta^m$:
\begin{eqnarray}
d_{mnp}\,\zeta^n\,\zeta^p&=&0\,,
\end{eqnarray}
can be viewed as a kind of ``$ {\rm E}_{6(6)}$-pure spinor''
constraint. It defines an orbit of the $\overline{{\bf 27}}$ with
stability group ${\rm SO}(5,5)\ltimes T_{16}$ (see
\cite{Ferrara:1997uz}). This means that there exists an ${\rm
SO}(5,5)\subset {\rm E}_{6(6)} $ with respect to which $\zeta^m$ is
a singlet. If we decompose the adjoint and the fundamental
representations of ${\rm E}_{6(6)}$ with respect to ${\rm O}(1,1)\times
{\rm SO}(5,5)\subset {\rm E}_{6(6)}$ we find:
\begin{eqnarray}
{\bf 78}&\rightarrow & {\bf 1}^0+{\bf 45}^0+{\bf 16}_c^{+3}+{\bf 16}_s^{-3}\,,\nonumber\\
\overline{{\bf 27}}&\rightarrow & {\bf 1}^{-4}+{\bf 10}^{+2}+{\bf
16}_c^{-1}\,.\label{o11o55branch}
\end{eqnarray}
The stabilizer of $\zeta^m$ is thus generated by the ${\bf
45}^0+{\bf 16}_s^{-3}$, while $\zeta^m$ corresponds to the ${\bf
1}^{-4}$. We denote by ${\bf h}$ the ${\rm O}(1,1)$ generator,
such that $\delta_{\bf h}\, \zeta^m=-4\,\zeta^m$.
Eqs.~(\ref{remcon}), on the other hand, imply that $\delta_\Xi\,
\zeta^m=-\Xi_n{}^m\,\zeta^n=\frac{4}{3}\,\kappa\,\zeta^m$. Since
$\zeta^m$ is a simultaneous eigenvector of both $\Xi$ and ${\bf h}$,
we must have $\delta_{[\Xi,\,{\bf h}]}\, \zeta^m=0$, namely that
$\Xi$ cannot have a component along the ${\bf 16}_c^{+3}$:
\begin{eqnarray}
\Xi &\in & {\bf 1}^0+{\bf 45}^0+{\bf 16}_s^{-3}\,.
\end{eqnarray}
We conclude that $\Xi$ consists of a component proportional
to ${\bf h}$ plus an element $\Xi_0$ in the algebra of the little
group of $\zeta^m$:
\begin{eqnarray}
\Xi &=& -\frac{1}{3}\,k\,{\bf h}+\Xi_0\;,\qquad
\delta_{\Xi_0} \zeta^m=0\,.\label{xisemisimple}
\end{eqnarray}
Let us consider the case in which $\Xi_0$ is a semisimple element of
$\mathfrak{so}(5,5)$ and thus can be taken as an element of its
Cartan subalgebra. One can show that in this case, taking $\zeta^{mn}=
\zeta^{[m} \eta^{n]}$, with $\eta^m$ in the ${\bf 16}_c^{-1}$, all the
constraints are satisfied. In particular the two sides of Eq.~(\ref{cs4}) are separately zero.
As we shall show in Appendix~\ref{etre}, this equation in particular implies that $\Xi_0$ should
commute with an ${\rm SO}(4,4)$ subgroup of ${\rm SO}(5,5)$.
The resulting gauge algebra $\mathfrak{g}_g$ is
$21$-dimensional and of the form:
\begin{eqnarray}
\mathfrak{g}_g&=&\mathfrak{o}(1,1)\oplus\mathfrak{so}(2,1)\oplus\mathfrak{l}^{(2\kappa)}\oplus\mathfrak{l}^{(4\kappa)}
\;,\nonumber\\[1ex]
&&{\rm dim}(\mathfrak{l}^{(2\kappa)})=16\;,\quad
{\rm dim}(\mathfrak{l}^{(4\kappa)})=1\,,\label{ga}
\end{eqnarray}
the gradings referring to the ${\rm O}(1,1)$-generator. We can
understand the embedding of the gauge group into the stability group
$\left[\mathbb{R}^+\times{\rm SO}(6,5)\right]\ltimes
(T_{32}\times T_1)$ of the
$\vartheta_M$-orbit by decomposing ${\rm SO}(6,5)$ with respect to
the ${\rm SO}(2,1)\times {\rm SO}(4,4)$. Then the generators of
$\mathbb{R}^+\times {\rm SO}(2,1)$ provide the zero-grading part of
the gauge algebra (\ref{ga}). The gauge generators can be
written in a manifestly ${\rm SO}(2,1)\times {\rm
SO}(4,4)$-covariant way. Let $A,B=1,2$ denote the ${\rm
SO}(2,1)$-doublet indices while $I,J=1,\dots, 8$ label the ${\bf
8}_s$ of ${\rm SO}(4,4)$. Then let $T_x$, $x=0,1,2,3$, be the
$\mathfrak{o}(1,1)+\mathfrak{so}(2,1)$ generators,
$\mathfrak{l}^{(2\kappa)}={\rm Span}(T_{AI})$ and
$\mathfrak{l}^{(4\kappa)}={\rm Span}(T)$. The relevant commutation
relations between the gauge generators are:
\begin{eqnarray}
[T_x,\,T_{AI}]&=&-(T_x)_A{}^B\,T_{BI}\;,
\qquad
[T_{AI},\,T_{BJ}]=\epsilon_{AB}\,C_{IJ}\,T\;,
\label{gaugecomm}
\end{eqnarray}
where $C_{IJ}$ is the symmetric invariant matrix in the product
${\bf 8}_s\times {\bf 8}_s$.
In other words, with respect to ${\rm SO}(1,1)\times{\rm
SO}(2,1)\times {\rm SO}(4,4)$ the generators $\{T_x\}$ are in the
${\bf (3,1)}^0$, $\{T_{AI}\}$ in the ${\bf (2,8_s)}^{2\kappa}$ while
$\{T\}$ is in the ${\bf (1,1)}^{4\kappa}$. In terms of the ${\rm
E}_{7(7)}$-branching with respect to the ${\rm E}_{6(6)}$-subgroup,
the $T_{AI}$ consists of $8$ generators from the $\overline{{\bf
27}}$ and $8$ from the ${\bf 78}$, while $T$ originates from the
$\overline{{\bf 27}}$.
This structure does not change either in the limit
$\zeta^{mn}\rightarrow 0$, or in the limit
$\zeta^m\rightarrow 0$. In the latter case the $\mathfrak{gl}(2)$
algebra of ${\rm SO}(1,1)\times{\rm SO}(2,1)$ contracts to a
non-semisimple algebra of the form $\mathfrak{o}(1,1)+H_3$, where
$H_3$ is a three-dimensional Heisenberg algebra spanned by nilpotent
generators. Only if both the zero-modes vanish
($\zeta^{mn}\rightarrow 0$, $\zeta^m\rightarrow 0$) the $T_{AI}$
generators which do not vanish become abelian, the last commutator in
(\ref{gaugecomm}) becomes trivial and we retrieve the first example
discussed in section~\ref{sec:example}.
\subsection{Cartan-Killing metric of the gauge group}
In the previous sections we have been discussing the general
solution to the quadratic constraints and worked out the
corresponding gauge groups in certain examples. With the general solution given
in section \ref{sttqc}, the gauge group generators may be reconstructed from (\ref{defX}),
putting together (\ref{thbreak}), (\ref{ThetaBreak}) and (\ref{solutionemb}). The explicit form of
the generators $\{X_M\} = \{X_\bullet, X_m, X^m, X^\bullet\}$ in terms of the parameters
$\kappa$, $\Xi^a$, $\zeta^m$, and
$\zeta^{mn}$ is given in (\ref{explicitgenerators}) in the appendix.
Via (\ref{XXX}) these generators also encode the structure constants
of the gauge algebra. We can compute the Cartan-Killing
metric of the gauge group as $g_{MN}\equiv{\rm Tr}(X_M\,X_N)$. Its
non-vanishing components are \begin{eqnarray} g_{\bullet \bullet} &=&
64\kappa^2+ 2\,\Xi^a \Xi_{a}
\;,\nonumber\\
g_{\bullet}{}^m &=& (96\kappa^2- 2\,\Xi^a \Xi_{a}) \,\zeta^m
\;,\nonumber\\
g_{\bullet}{}_m &=& -6\,\Xi_l{}^n d_{nmk} \zeta^{kl}
\;,\nonumber\\
g^{mn} &=& (64\kappa^2+ 2\,\Xi^a \Xi_{a}) \,\zeta^m\zeta^n
\;,\nonumber\\
g_{mn} &=& -6\, \zeta^{kl}\zeta^{pq} d_{mkp} d_{nlq}
\nonumber\\
&&{} +\ft23\,(80\kappa^2-3\,\Xi^a \Xi_{a})\,d_{mnk}\,\zeta^k +
24\,d_{kl(m}\,(\Xi^2)_{n)}{}^k \zeta^l \;.
\end{eqnarray}
For $\zeta^m=0=\zeta^{mn}$ this shows that the semisimple part
of the gauge algebra is one-dimensional in accordance with its origin as
a Scherk-Schwarz reduction from five dimensions.
If $\zeta^{mn}=0$ but $\zeta^m$ is non-vanishing, a contraction of equation (\ref{cs4})
implies that $\Xi^a \Xi_{a}=32\kappa^2$, and the Cartan-Killing form accordingly reduces to
\begin{eqnarray}
g_{\bullet \bullet} &=& 128\kappa^2
\;,\nonumber\\
g_{\bullet}{}^m &=& 32\kappa^2 \,\zeta^m
\;,\nonumber\\
g^{mn} &=& 128 \kappa^2\,\zeta^m\zeta^n
\;,\nonumber\\
g_{mn} &=& -\ft{32}3\kappa^2 \,d_{mnk}\,\zeta^k
+ 24\,d_{kl(m}\,(\Xi^2)_{n)}{}^k \zeta^l
\;.
\end{eqnarray}
\subsection{Electric/magnetic gaugings}
So far, we have discussed the structure of the gauge algebra by studying
deformations that involve vector fields from the entire 56-dimensional
fundamental representation of $E_{7(7)}$.
It is well known~\cite{Cremmer:1979up},
that only half of these vector fields are dynamical electric vector fields
while the other half is given by their magnetic duals. Accordingly, only the former half
appears in the action of the ungauged theory.
Nevertheless, the connections of a general gauging may contain magnetic
vector fields that are related by their first order duality equations to the electric fields of the theory.
In~\cite{deWit:2005ub} it has been shown how to elevate this construction to the level
of an action by introducing additional auxiliary two-form tensor fields
(which in turn are the magnetic duals to the scalar fields of the theory).
The magnetic vector fields then do not possess a standard kinetic term but rather couple via
a topological $BF$ term to the two-form tensor fields.
On the other hand, all standard gaugings of the theory~\cite{deWit:2007mt}
satisfy a symplectic locality condition that ensures the existence of a symplectic frame
in which all the vector fields involved in the gauging live in the electric sector.
In this sense even in presence of magnetic charges these theories
remain electric gaugings in disguise
which is in accordance with general results on the gauging of
electric/magnetic duality~\cite{Bunster:2010wv,Deser:2010it}.
We shall see that this is no longer the case for the gaugings considered in this paper,
related to the fact that these theories do no longer admit an action.
For the solution of the consistency constraints of the embedding tensor
discussed at the end of section~\ref{sttqc}, the
left and right hand sides of equations (\ref{Q1})--(\ref{Q3}) vanish separately.
The gauge group generators thus satisfy the symplectic locality condition
$\Omega^{MN} X_M X_N = 0$. I.e.\
as for the standard gaugings we can choose a symplectic frame
$\{X^M\} \rightarrow \{X^\Lambda, X_\Lambda\}$ such that all $X^\Lambda$ are identically zero.
Indeed, in this case the explicit form of the generators (\ref{explicitgenerators})
shows that $X_\bullet=0=X^k$\,.
Accordingly, the gauging only involves electric vector fields $\{{\cal A}_\mu^\Lambda\} = \{{\cal A}_{\mu\,\bullet}\,, {\cal A}_\mu^k\}$.
On the other hand, in the generic case the components
$\zeta^m$, $\zeta^{mn}$ in the embedding tensor are non-vanishing,
such as in the example worked out in section~\ref{awoe}. Then, equation~(\ref{Q3})
implies that $\Omega^{MN} X_M X_N \not=0$, i.e.\ there is no symplectic frame
in which the gauging involves only electric vector fields.
We conclude that the general gaugings including the trombone generator
are necessarily and genuinely dyonic!
\mathversion{bold}
\section{Scalar coset space and the $T$-tensor}
\mathversion{normal}
\label{sec:coset}
In this section, we discuss the structure of the scalar sector of the theory,
discuss its interplay with the gauging defined in the previous section,
and define the relevant scalar field dependent tensors ($T$-tensors) that enter
in the field equations of the gauged supergravity.
\mathversion{bold}
\subsection{Coset space $E_{7(7)}/SU(8)$}
\mathversion{normal}
The scalar fields of $N=8$ supergravity
can be parametrized in terms of the 56-dimensional complex vectors
${\cal V}_M{}^{ij}= ({\cal V}_\Lambda{}^{ij}, {\cal V}^{\Sigma \,ij})$ and their
complex conjugate ${\cal V}_{M\,ij}=({\cal V}_{\Lambda\,ij},{\cal V}^\Sigma{}_{ij})$,
which together constitute a $56\times 56$ matrix ${\cal V}$,
\begin{equation}
\label{VV}
{\cal V}_M{}^{\underline N} =\Big({\cal V}_M{}^{ij}, {\cal V}_M{}_{kl} \Big) =
\pmatrix{{\cal V}_\Lambda{}^{ij}&{\cal V}_{\Lambda\,kl}\cr
\noalign{\vskip 4mm}
{\cal V}^{\Sigma\,ij} & {\cal V}^\Sigma{}_{kl}\cr}\,.
\end{equation}
Indices $M, N, \dots = 1, \dots, 56$, label the fundamental representation of $E_{7(7)}$,
indices $i, j, \dots = 1, \dots, 8$ denote the fundamental complex ${\bf 8}$ of $SU(8)$.\footnote{
Earlier, in section~\ref{sttqc} we have used indices $m, n, \dots$ in a different context labeling
the 27 dimensional fundamental representation of $E_{6(6)}$. We hope that this does not
lead to extra confusion.}
The underlined indices $\underline{M}, \underline{N}, \dots = 1, \dots, 56$, are a collective label for
the ${\bf 28}$ + $\bar{\bf 28}$ of $SU(8)$\,.
The matrix ${\cal V}_M{}^{\underline N}$
transforms under rigid $E_{7(7)}$ from the left and under
local ${SU}(8)$ from the right. Strictly speaking, it does not constitute an
element of $E_{7(7)}$, but it is equal to a constant matrix (to account for
the different bases adopted on both sides) times a space-time
dependent element of $E_{7(7)}$.
We refer to~\cite{deWit:1982ig,deWit:2007mt} for further details.
In particular, the scalar matrix satisfies the properties
\begin{eqnarray}
\label{eq:VV-orthogonal}
{\cal V}_M{}^{ij} \,{\cal V}_{N\,ij} - {\cal V}_{M\,ij}\, {\cal V}_N{}^{ij} &=&
\mathrm{i}\,\Omega_{MN}\,, \nonumber\\
\Omega^{MN} \,{\cal V}_M{}^{ij} \,{\cal V}_{N\,kl} &=&
\mathrm{i}\,\delta^{ij}{}_{kl}\,, \nonumber\\
\Omega^{MN} \,{\cal V}_M{}^{ij} \, {\cal V}_N{}^{kl} &=& 0\,,
\end{eqnarray}
reflecting the fact that $E_{7(7)}$ is embedded into $Sp(56)$.
The covariant scalar currents ${\cal Q}_{\mu}{}_{i}{}^{j}$ and $\mathcal{P}_\mu{}^{ijkl}$ are defined by
\begin{eqnarray}
\partial_\mu{\cal V}_M{}^{ij}
- A_\mu^P X_{PM}{}^N \,{\cal V}_N{}^{ij}
&\equiv& - \mathcal{Q}_{\mu \,k}{}^{[i} \,{\cal V}_M{}^{j]k}
+ \mathcal{P}_\mu{}^{ijkl} \,{\cal V}_{M kl}
\;,
\label{defQP}
\end{eqnarray}
with gauge group generators from (\ref{defX}), and satisfy
\begin{eqnarray}
{\cal Q}_{\mu}{}^{i}{}_{j} = - {\cal Q}_{\mu\, j}{}^i\;,\qquad
{\cal Q}_{\mu i}{}^i=0\;,\qquad
{\cal P}_\mu{}^{ijkl} = \ft1{24}\,\varepsilon^{ijklmnpq}\, {\cal P}_{\mu\,mnpq}
\;,
\end{eqnarray}
as a consequence of ${\cal V}_M{}^{\underline N}$ being related to
an $E_{7(7)}$ element by multiplication with a constant matrix.
The integrability conditions of (\ref{defQP}) yield the
Cartan-Maurer equations,
\begin{eqnarray}
\label{eq:GECM-Q-P}
{\cal F}({\cal Q})_{\mu\nu\,i}{}^j ~\equiv~ 2\partial_{[\mu} \mathcal{Q}_{\nu] i}{}^j
+ \mathcal{Q}_{[\mu i}{}^k\,
\mathcal{Q}_{\nu]k}{}^j
& = & \ft43\,{\cal P}_{[\mu}{}^{\!jklm}
\, {\cal P}_{\nu]iklm} -\ft23 i\,\mathcal{H}_{\mu\nu}^{M} \,
(\check{X}_{M}){}^{PQ}\, {\cal V}_{P\,ik} {\cal V}_Q{}^{jk}
\nonumber \\[1ex]
{\cal D}_{[ \mu}{\cal P}_{\nu]}{}^{ijkl} &=& - \ft12
i \,\mathcal{H}_{\mu\nu}^{M} \,
(\check{X}_{M}){}^{PQ} \,{\cal V}_{P}{}^{ij} {\cal V}_{Q}{}^{kl} \,,
\label{Cartan-Maurer}
\end{eqnarray}
with the $SU(8)$ covariant derivative ${\cal D}_{\mu}$
and the covariant field strength $\mathcal{H}_{\mu\nu}^{M}$ from (\ref{defH}).
Note that its part carrying the two-forms $B_{\mu\nu\,\alpha}$ drops from
(\ref{Cartan-Maurer}) due to the orthogonality relation (\ref{QQZ}).
\mathversion{bold}
\subsection{The $T$-tensor}
\mathversion{normal}
Following~\cite{deWit:1982ig,deWit:2007mt} we define the $T$-tensor
as the gauge group generator (\ref{defX}) dressed with the scalar vielbein
\begin{eqnarray}
(T_{ij})^{klmn} &\equiv&
\ft12\,({\cal V}^{-1})_{ij}{}^M ({\cal V}^{-1})^{kl}{}^N\,(X_{M})_{N}{}^K\,{\cal V}_K{}^{mn}
\;,\qquad
\mbox{etc.}
\;.
\label{TX}
\end{eqnarray}
The various components of this tensor will show up in the modified field equations
of the gauged theory and parametrize the new couplings.
The linear constraints (\ref{linear}) satisfied by the embedding tensor can be made explicit by parametrizing
the $T$-tensor in terms of the irreducible $SU(8)$ tensors $A^{ij}$, $A_i{}^{jkl}$, $B^{ij}$, transforming in the
${\bf 36}$, ${\bf 420}$, and the ${\bf 28}$, respectively,\footnote{
I.e.\ $A^{ij}=A^{(ij)}$, $A_i{}^{jkl}=A_i{}^{[jkl]}$, $A_i{}^{ikl}=0$, $B^{ij}=B^{[ij]}$,
and complex conjugates $(A^{ij})^*=A_{ij}$, etc.
}
according to
\begin{eqnarray}
(T_{ij})_{kl}{}^{mn} &=& \ft12\delta_{[k}^{[m} A^{n]}{}_{l]ij} + \delta^{mn}_{[i[k} A_{l]j]} - \ft16 (8\,\delta^{mn}_{[i[k} B_{l]j]} + \delta^{mn}_{kl} B_{ij}) - \ft12\delta^{mn}_{kl} B_{ij}\;,\nonumber\\[.5ex]
(T_{ij})^{rs}{}_{pq} &=& - \ft12\delta_{[p}^{[r} A^{s]}{}_{q]ij} - \delta^{rs}_{[i[p} A_{q]j]} + \ft16 (8\,\delta^{rs}_{[i[p} B_{q]j]} + \delta^{rs}_{pq} B_{ij}) - \ft12\delta^{rs}_{pq} B_{ij}\;,\nonumber\\[.5ex]
(T_{ij})_{kl \, pq} &=& \ft{1}{24} \epsilon_{klpqrstu} \delta^r_{[i} A_{j]}{}^{stu}
+ \ft1{12} \epsilon_{klpqijtu} B^{tu}\;,\nonumber\\[.5ex]
(T_{ij})^{rs \; mn} &=& \delta^{[r}_{[i} A_{j]}{}^{smn]} + 2 \delta_{ij}^{[rs} B^{mn]}
\;.
\label{TAB}
\end{eqnarray}
The tensors $A^{ij}$, $A_i{}^{jkl}$
together with their complex conjugates fill the ${\bf 912}$
representation~$\Theta_M{}^\alpha$ of
the embedding tensor and carry the structure of the standard gaugings.
The tensor $B^{ij}$ is related to the new components $\vartheta_M$
of the embedding tensor according to
\begin{equation}
\vartheta_M = {\cal V}_M{}^{ij} B_{ij} + {\cal V}_{M\; ij} B^{ij}
\;,
\label{defB}
\end{equation}
and contains all the new contributions due to the gauging of the trombone generator.
Together, the tensors $A$ and $B$ will describe the scalar couplings of the gauged theory.
From their definition (\ref{TAB}) and (\ref{TX}) one derives the differential relations
\begin{eqnarray}
{\cal D}_{\mu} A{}^{ij} &=& \ft{1}{3} {A{}^{(i}}_{klm} {\cal P}_{\mu}{}^{j)klm}\;,\\
{\cal D}_{\mu} A_{i}{}^{jkl} &=&2A_{im} {\cal P}_{\mu}{}^{mjkl} + 3 {\cal P}_{\mu}{}^{mn[jk} A^{l]}{}_{imn}
+ {\cal P}_{\mu}{}^{mnp[j}\delta^k_i A^{l]}{}_{mnp}
\;,\\
{\cal D}_{\mu} B_{ij} &=& - {\cal P}_{\mu\; ijkl} B^{kl}\;,
\end{eqnarray}
where again ${\cal D}_{\mu}$ refers to the $SU(8)$ covariant derivative with the composite connection
${\cal Q}_{\mu\,i}{}^j$ from (\ref{defQP}).
For the supersymmetry algebra it will also be useful to compute the tensor
${Z}^M{}_{KL}$ upon dressing with $({\cal V}^{-1})_{kj}{}^K ({\cal V}^{-1})^{ij}{}^L$:
\begin{eqnarray}
{{Z}^M{}}_{kj}{}^{ij} &=& -\frac{3}{2} ({\cal V}^{-1\, in \, M} A_{nk} + {{\cal V}^{-1}{}}_{kl}{}^M A^{ni}) + \frac{3}{4} ({\cal V}^{-1\, mn \, M} A^i{}_{kmn} + {{\cal V}^{-1}{}}_{mn}{}^M A_k{}^{imn})\nonumber\\
&+& 4 ({\cal V}^{-1\, in \, M} B_{nk} + {{\cal V}^{-1}{}}_{kl}{}^M B^{ni}) + \frac{1}{2} \delta_k^i ({\cal V}^{-1\, mn \, M} B_{mn} + {{\cal V}^{-1}{}}_{mn}{}^M B^{mn})\;.
\nonumber\\
\label{ZM}
\end{eqnarray}
Dressing the quadratic constraints (\ref{Q1})--(\ref{Q3}) (or alternatively (\ref{XXX}))
with the scalar vielbein (\ref{VV}) induces a plethora of relations bilinear in the tensors $A$, $B$.
In appendix~\ref{app:Tids}, we have collected a number of such identities which are important in
the subsequent calculations.
Here, we only give two examples of such identities. A linear combination of the constraints
(\ref{constraints63}) transforming in the ${\bf 63}$ of $SU(8)$
shows that the traceless part of the hermitean tensor defined by
\begin{eqnarray}
\Pi^i{}_j &\equiv&
6A^{ik}A_{jk}-\ft13A^i{}_{mnk} A_j{}^{mnk}
+\ft43\left(A_j{}^{imn}B_{mn}+A^i{}_{jmn}B^{mn}\right)
-\ft{256}9 \, B^{ik}B_{jk}
\;,
\nonumber
\end{eqnarray}
vanishes
\begin{eqnarray}
\Pi^i{}_j &=& \ft18\delta^i_j \,\Pi^k{}_k
\;.
\label{QCcomb1}
\end{eqnarray}
Another useful identity is given by the self-duality equation
\begin{eqnarray}
\Pi_{ijkl} &=& \ft1{24} \epsilon_{ijklmnpq}\,\Pi^{mnpq}
\;,\nonumber\\[1ex]
\mbox{for}\quad
\Pi_{ijkl}&\equiv&
A^m{}_{[ijk} A_{l]m}
-\ft34 A^m{}_{p[ij}A^p{}_{kl]m}
+2 A^m{}_{[ijk} B_{l]m}
-8 B_{[ij}B_{kl]}
\;,
\label{QCcomb2}
\end{eqnarray}
which is obtained as a linear combination
of the constraints (\ref{constraints70}) transforming in the ${\bf 70}$ of $SU(8)$
\mathversion{bold}
\subsection{Vector fields}
\mathversion{normal}
As mentioned above, only half of the 56 vector fields ${\cal A}_\mu^M$ enter the Lagrangian of the ungauged theory. This corresponds to selecting a symplectic frame, such that the vector fields split according to $\{{\cal A}_\mu^M\} \rightarrow \{{\cal A}_\mu^\Lambda, {\cal A}_{\mu\,\Lambda}\}$ into electric and magnetic fields.
Accordingly, we define the electric field strengths
${\cal H}_{\mu\nu}^\Lambda$ via (\ref{defH}) as the curvature of ${\cal A}_\mu^\Lambda$ while their magnetic duals are defined as functions of the electric vector fields according to
\begin{eqnarray}
{\cal G}_{\mu\nu\,\Lambda}^+ &\equiv&
{\cal N}_{\Lambda\Sigma}\,{\cal H}_{\mu\nu}^{+\;\Sigma} ~+ \mbox{fermions}
\;,
\label{defGF}
\end{eqnarray}
with the complex matrix ${\cal N}_{\Lambda\Sigma}$ defined by
$\,{\cal V}^{\Sigma\, ij}\,{\cal N}_{\Lambda\Sigma}\equiv - {\cal V}_\Lambda{}^{ij}$\,,
and where the superscript~$^\pm$ refers to the (anti-)selfdual part of the
field strength.
The fermionic part of (\ref{defGF}) is explicitly given in~\cite{Cremmer:1979up,deWit:1982ig,deWit:2007mt}.
We define the full symplectic vector ${\cal G}_{\mu\nu}^M \equiv
({\cal H}_{\mu\nu}^\Lambda, {\cal G}_{\mu\nu\,\Lambda})$\,, which will
in particular enter the fermionic field equations and supersymmetry transformation rules.
By construction, it allows the decomposition
\begin{eqnarray}
{\cal G}_{\mu\nu}^M &=&
({\cal V}^{-1})^{ij\,M} {\cal G}_{\mu\nu\,ij}^{+}
+
({\cal V}^{-1})_{ij}{}^M {\cal G}_{\mu\nu}^{-\;ij}
~+ \mbox{fermions}
\;,
\end{eqnarray}
into its selfdual and anti-selfdual part.
In contrast, we introduce the field strengths ${\cal H}_{\mu\nu}^{ij}$ and
${\cal H}_{\mu\nu\,ij}$ as the dressed version
\begin{eqnarray}
{\cal H}_{\mu\nu}^M &=&
({\cal V}^{-1})^{ij\,M} {\cal H}_{\mu\nu\,ij}
+
({\cal V}^{-1})_{ij}{}^M {\cal H}_{\mu\nu}^{\;ij}
\;,
\end{eqnarray}
of the covariant non-abelian field strengths introduced in (\ref{defH}),
that combine electric and magnetic vector fields.
Note that ${\cal H}_{\mu\nu}^\Lambda={\cal G}_{\mu\nu}^\Lambda$ is identically
satisfied, whereas
${\cal H}_{\mu\nu\,\Lambda}={\cal G}_{\mu\nu\,\Lambda}$ describes the
first order duality relation between electric and magnetic vector fields.
\section{Supersymmetry algebra}
\label{sec:susy}
Before deriving the full set of supersymmetric equations of motion,
we establish the supersymmetry transformation rules by verifying the supersymmetry algebra.
Under supersymmetry, the bosonic fields transform as
\begin{eqnarray}
\delta_{\epsilon} e_{\mu}{}^{a}&=&
\bar\epsilon^{i}\gamma^{a}\psi_{\mu i} ~+~
\bar\epsilon_{i}\gamma^{a}\psi_{\mu}{}^i \;, \nonumber\\[1ex]
\delta_{\epsilon}{\cal V}_M{}^{ij} &=& 2\sqrt{2}\,{\cal V}_{M kl} \, \Big(
\bar\epsilon^{[i}\chi^{jkl]}+\ft1{24}\varepsilon^{ijklmnpq}\,
\bar\epsilon_{m}\chi_{npq}\Big) \,, \nonumber \\[1ex]
\delta_{\epsilon} {\cal A}_{\mu}^{M}
&=&
-i\,\Omega^{MN} {\cal V}_N{}^{ij}\,\Big(
\bar\epsilon^{k}\,\gamma_{\mu}\,\chi_{ijk}
+2\sqrt{2}\, \bar\epsilon_{i}\,\psi_{\mu j}\Big)~+~ {\rm c.c.}
\;, \nonumber \\[1ex]
\delta_{\epsilon} B_{\mu\nu\,\alpha} &=&\ft{2}{3} \sqrt{2} \,
(t_{\alpha})^{PQ}\, \Big( {\cal V}_{P\,ij} {\cal
V}_{Q\,kl}\,
\bar\epsilon^{[i}\,\gamma_{\mu\nu}\,\chi^{jkl]}
+ 2 \sqrt{2}\, {\cal V}_{P\,jk} {\cal V}_{Q}{}^{ik}\,
\bar\epsilon_{i}\,\gamma_{[\mu}\,\psi_{\nu]}{}^{j}
~+~ {\rm c.c.}\Big) \nonumber\\
&&{}
+2(t_{\alpha})_{MN}\,{\cal A}_{[\mu}^{M}\,\delta
{\cal A}_{\nu]}^{N} \;.
\label{susybosons}
\end{eqnarray}
while the transformation of the fermions is given by
\begin{eqnarray}
\delta_{\epsilon} \psi_{\mu}^i &=& 2 {\cal D}_{\mu} \epsilon^i + \frac{\sqrt{2}}{4} {{\cal G}}^-_{\rho \sigma}{}^{ij} \gamma^{\rho \sigma} \gamma_{\mu} \epsilon_j + \sqrt{2} A^{ij} \gamma_{\mu} \epsilon_j -2\sqrt{2} B^{ij} \gamma_{\mu} \epsilon_j\;,\nonumber\\
\delta_{\epsilon} \chi^{ijk} &=& -2 \sqrt{2} {{\cal P}}_{\mu}{}^{ijkl} \gamma^{\mu} \epsilon_l + \frac{3}{2} {{\cal G}}^-_{\mu \nu}{}^{[ij} \gamma^{\mu \nu} \epsilon^{k]} -2 A_l{}^{ijk} \epsilon^l -4 B^{[ij} \epsilon^{k]}\;,
\label{susyfermions}
\end{eqnarray}
up to higher order fermion terms.
Except for the respective last terms in the fermionic transformation rules
(carrying the tensor $B_{ij}$), these supersymmetry transformations
are known from~\cite{deWit:1982ig,deWit:2007mt}. The structure of the new terms follows from the $SU(8)$ representation content, their factors are determined from the closure of the supersymmetry algebra.
This algebra is given by
\begin{equation}
\label{eq:susy-algebra1}
{}[\delta(\epsilon_1),\delta(\epsilon_2)] = \xi^\mu \hat D_\mu +
\delta_{\rm Lor}(\Omega^{ab}) + \delta_{\rm susy}(\epsilon_3) +
\delta_{\rm SU(8)}(\Lambda^i{}_j) + \delta_{\rm gauge}(\Lambda^M) +
\delta_{\rm gauge}(\Xi_{\mu \alpha}) \,.
\end{equation}
The first term refers to a covariantized general coordinate transformation
with diffeomorphism parameter
\begin{equation}
\xi^\mu = 2\, \bar\epsilon_2{}^i \gamma^\mu \epsilon_{1\, i} + 2\, \bar\epsilon_{2\; i} \gamma^\mu \epsilon_1{}^i
\;,
\end{equation}
and including terms of order $g$ induced by the gauging.
The last two terms refer to
gauge transformations (\ref{gaugeV1}), (\ref{gaugeV2}) and (\ref{gaugeT}),
with parameters
\begin{eqnarray}
\Lambda^N &=& -4 i \sqrt{2} \; \Omega^{NP} \left({\cal V}_P{}^{mn} \bar\epsilon_{2\, m} \epsilon_{1\, n} - {\cal V}_{P\;mn} \bar\epsilon_2^m \epsilon_1^n\right)
\;,
\nonumber\\
\Xi_{\mu\alpha} &=& -\ft83 (t_\alpha)^{PQ}\,
{\cal V}_{P\,ik} {\cal V}_Q{}^{jk} \left(\bar\epsilon_2{}^i \gamma_\mu \epsilon_{1j}
+ \bar\epsilon_{2j} \gamma_\mu \epsilon_1{}^i\right)
\;,
\label{LX}
\end{eqnarray}
respectively. Up to the contributions from the new terms in the supersymmetry transformation rules,
the supersymmetry algebra
has been verified in \cite{deWit:1982ig,deWit:2007mt}. In presence of
$B_{ij}$, $B^{ij}$, the commutator (\ref{eq:susy-algebra1}) evaluated on the vielbein acquires the additional terms
\begin{eqnarray}
[\delta_{\epsilon_1}, \delta_{\epsilon_2}] \, e_{\mu}{}^a &=&
\dots \quad
-4 \sqrt{2}\, \left(B^{mn} \bar\epsilon_{2\, m} \epsilon_{1\, n} + B_{mn} \bar\epsilon_2^m \epsilon_1^n \right) e_{\mu}{}^a
\;.
\end{eqnarray}
These precisely reproduce the action
of a scaling gauge transformation with parameter (\ref{LX}) on the vielbein
\begin{eqnarray}
\delta_{\Lambda} e_{\mu}{}^a &=& \Lambda^M \vartheta_M\, {\bf t}_0 \cdot e_{\mu}{}^a \\ \nonumber
&=& -4 i \sqrt{2} \; \Omega^{MN} \vartheta_M ({\cal V}_N{}^{mn} \bar\epsilon_{2\, m} \epsilon_{1\, n} - {\cal V}_{N\;mn} \bar\epsilon_2^m \epsilon_1^n) \, e_{\mu}{}^a\\ \nonumber
&=&
-4 \sqrt{2}\, \left(B^{mn} \bar\epsilon_{2\; m} \epsilon_{1\; n} + B_{mn} \bar\epsilon_2^m \epsilon_1^n \right) e_{\mu}{}^a
\;,
\end{eqnarray}
where we have used (\ref{eq:VV-orthogonal}) and (\ref{defB}).
Similarly, one may check that the terms carrying the scalar tensors (\ref{TAB}) in the supersymmetry commutator
on the scalar fields combine into
\begin{eqnarray}\nonumber
{\cal V}^{-1\, ij \; M} \left[\delta_{\epsilon_1}, \delta_{\epsilon_2}\right] {\cal V}_M{}^{kl} &=&
\dots\quad -8\sqrt{2}\left(
T^{mn}{}^{ijkl} \,\bar\epsilon_{2\, m} \epsilon_{1\, n}+
T_{mn}{}^{ijkl} \, \bar\epsilon_2^m \epsilon_{1}^n \right)
\nonumber\\[.5ex]
&=& \dots\quad + {\cal V}^{-1\;ij \;M}\; \Lambda^N (X_N)_M{}^K\, {\cal V}_K{}^{kl}
\;,
\qquad
\mbox{etc.}
\nonumber
\end{eqnarray}
and consistently reproduce the action
of a gauge transformation with parameter (\ref{LX}).
In checking the supersymmetry algebra on the vielbein and the scalar fields,
we have fixed all the new factors in the supersymmetry transformation rules (\ref{susyfermions}).
As a consistency check, one may further verify that the algebra also closes
on the vector and the tensor gauge fields.
\section{Equations of motion}
\label{sec:eom}
\subsection{Einstein equations}
Having established the supersymmetry algebra, we can now determine the deformed equations
of motion by requiring covariance under the new supersymmetry transformation rules.
As there is no longer an action underlying the gauged theory, we have to work directly on the level
of the equations of motion.
This derivation of the supersymmetric field equations is based on reference~\cite{LeDiffon:2010}.
We will start from the gravitino equations of motion for which we use the following ansatz
\begin{eqnarray}\nonumber
0~=~ ({\cal E}_{\psi})^\mu{}_i &\equiv&
-e^{-1} \varepsilon^{\mu \nu \rho \sigma} \gamma_{\nu} {\cal D}_\rho \psi_{\sigma\, i}
- \ft{\sqrt{2}}{6} \gamma^\nu \gamma^\mu \chi^{jkl} \, {\cal P}_{\nu \, jkli}
- \ft{\sqrt{2}}{4} {\cal G}^{+\; \rho \sigma}{}_{ij} \gamma^{[\mu} \gamma_{\rho \sigma} \gamma^{\nu]} \psi_{\nu}{}^j
\nonumber\\[.5ex]
&& {}
+ \ft{1}{8} {\cal G}^{-\; \rho \sigma}{}^{jk} \gamma_{\rho \sigma} \gamma^{\mu} \chi_{jki}
+ \sqrt{2} A_{ij} \gamma^{\mu \nu} \psi_{\nu}{}^j + \ft{1}{6} A_{i}{}^{jkl} \gamma^{\mu} \chi_{jkl}
\nonumber\\[.5ex]
&& {}
+\lambda\, \sqrt{2} B_{ij} \gamma^{\mu \nu} \psi_{\nu}{}^j + \zeta \, B^{kl} \gamma^{\mu} \chi_{ikl}
\;.
\label{eomgrav}
\end{eqnarray}
Except for the last two terms, these are the equations obtained from variation of the Lagrangian~\cite{deWit:1982ig,deWit:2007mt} of the gauged theory. While the $SU(8)$ structure of these two additional terms is fully determined by the representation content, we will in the following determine their unknown coefficients $\lambda$ and $\zeta$ by compatibility with supersymmetry.
E.g.\ vanishing of the ${\cal D}_\mu \epsilon$ terms in the supersymmetry variation of (\ref{eomgrav}) imposes
\begin{eqnarray}
2\sqrt{2}\,B_{ij} \left( \lambda \gamma^{\mu\nu} - e^{-1} \varepsilon^{\mu\nu\rho\sigma} \gamma_{\rho\sigma} \right)
{\cal D}_{\nu} \epsilon^{j}
&=& 0\;,
\end{eqnarray}
from which we deduce $\lambda=-2$.
Vanishing of the terms linear in $B {\cal G}^\pm \epsilon$ further
determines $\zeta=-5/3$, but we will for the moment keep the parameters in the formulas
so as to allow for further consistency checks.
Let us concentrate on the part of the supersymmetry variation of (\ref{eomgrav}) which is bilinear in the scalar tensors $A$, $B$
which originate from variation of its last four terms. We obtain
\begin{eqnarray}
&&{}
\Big\{
6 A_{ik} A{}^{jk} - \frac{1}{3} A_{i}{}^{klm} {A{}^j}_{klm} + 12 A_{ik} B^{jk}
+ 6 \lambda B_{ik} A^{jk}\label{bil1}\\
&&{}\qquad
- \frac{2}{3} A_{i}{}^{jlm} B_{lm} - 2 \zeta {A{}^j}_{ikl} B^{kl} +( 12\lambda + \frac{8 \zeta}{3})\, B_{ik} B^{jk}
- \frac{4\zeta}{3} \delta_i^j B_{kl} B^{kl}
\Big\}\;\; \times\;\; \gamma^\mu \epsilon_j
\;.
\nonumber
\end{eqnarray}
Only the $(ij)$-trace of the braced expression can be absorbed into a modification of the Einstein equations. In particular
its anti-hermitean part must vanish. Indeed, this follows from the first of the bilinear constraint relations (\ref{constraints63})
provided that $\lambda=2\zeta+\frac43$, which is satisfied for our above choice of constants.
With this value for $\lambda$, the expression (\ref{bil1}) reduces to its hermitean part
\begin{eqnarray}
&&{}
\Big\{
6 A_{ik} A{}^{jk} - \frac{1}{3} A_{i}{}^{klm} {A{}^j}_{klm} + (10+6\zeta)( A_{ik} B^{jk}+B_{ik} A^{jk})\;\;\;
\\
&&{}\qquad
-(\ft13+\zeta) (A_{i}{}^{jlm} B_{lm} + {A{}^j}_{ikl} B^{kl}) +( 16 + \frac{80 \zeta}{3})\, B_{ik} B^{jk}
- \frac{4\zeta}{3} \delta_i^j B_{kl} B^{kl}
\Big\}\; \times\; \gamma^\mu \epsilon_j
\;.
\nonumber
\end{eqnarray}
Finally, we observe that with the above value $\zeta=-5/3$ all coefficients precisely reproduce
the linear combination appearing in the quadratic constraint (\ref{QCcomb1}), such that the
full expression reduces to its trace part
\begin{eqnarray}
\Big\{\frac{3}{4} A_{kl} A{}^{kl} - \frac{1}{24} A_{n}{}^{jkl} {A{}^n}_{jkl}
- \frac{4}{3} B_{kl} B^{kl}\Big\} \; \times\;\; \gamma^\mu \epsilon_i
\;,
\end{eqnarray}
which can be absorbed into the modified Einstein equations.
Another important ingredient in the calculation is the evaluation of the commutator
\begin{eqnarray}{}
\gamma^{\mu\nu\rho}\,[{\cal D}_\nu, {\cal D}_\rho ] \,\epsilon_i &=&
\gamma^{\mu\nu\rho}\,\left(
\ft12{\cal F}({\cal Q})_{\nu\rho\,i}{}^j\,\epsilon_j
-\ft14 \widehat{\cal R}_{\nu\rho}{}^{ab}\gamma_{ab}\,\epsilon_i
-\ft12\vartheta_M {\cal H}_{\nu\rho}^M\,\epsilon_i
\right)
\nonumber\\
&=&
\left(\ft12g^{\mu\nu}\widehat{{\cal R}}-\widehat{{\cal R}}^{(\mu \nu )}\right)
\gamma_\nu \epsilon_i
+\vartheta_M \left({\cal H}^{-\,\mu\nu}{}^M-3{\cal H}^{+\,\mu\nu}{}^M\right)
\gamma_{\nu}\epsilon_i
\nonumber\\
&&{}
+\ft12
\gamma^{\mu\nu\rho}\,
{\cal F}({\cal Q})_{\nu\rho\,i}{}^j\epsilon_j
\;,
\end{eqnarray}
with the modified Riemann tensor and the curvature of
the $SU(8)$ connection from (\ref{modRR}) and (\ref{Cartan-Maurer}), respectively
Putting all this together,
a somewhat lengthy calculation shows that the full supersymmetry
variation of the Rarita-Schwinger equation (\ref{eomgrav}) eventually takes the form
\begin{eqnarray}
\delta_\epsilon ({\cal E}_{\psi})^\mu{}_i &=&
({\cal E}_{\rm Einstein})^{\mu\nu}\,\gamma_\nu \epsilon_i
-2\sqrt{2} \,({\cal E}_{\rm vector})^{\mu}{}_{ij} \, \epsilon^j
\;,
\label{var32}
\end{eqnarray}
where $({\cal E}_{\rm Einstein})^{\mu\nu}$ and $({\cal E}_{\rm vector})^{\mu}{}_{ij}$
denote the modified Einstein and the vector field equations of motion, respectively.
In bringing the supersymmetry variation into this form, we have in particular made use
of the equations
\begin{eqnarray}
X_M\left( {\cal G}_{\mu\nu}^M - {\cal H}_{\mu\nu}^M \right)
~=~ 0 ~=~
\vartheta_M \left( {\cal G}_{\mu\nu}^M - {\cal H}_{\mu\nu}^M \right)
\;.
\label{eom:1order}
\end{eqnarray}
For a purely electric gauging, these equations are identically satisfied. In presence
of magnetic charges, these equations represent the first order duality equations
between electric and magnetic vector fields.
The second order field equations obtained in (\ref{var32}) read
\begin{eqnarray}
({\cal E}_{\rm Einstein})^{\mu\nu} &=&
\widehat{{\cal R}}^{(\mu \nu )}-\ft12g^{\mu\nu}\widehat{{\cal R}}
+\ft16\,{\cal P}^{(\mu}_{ijkl} {\cal P}^{\nu)ijkl}
-\ft1{12} g^{\mu\nu}\,{\cal P}^\rho_{ijkl} {\cal P}_\rho^{ijkl}
-{\cal G}^+{\!}_\rho{}^{(\mu}{}_{jk} \,{\cal G}^{\nu)\rho-\,jk}
\nonumber\\[1ex]
&&{}
+g^{\mu\nu} \left( \ft{3}{4} A_{kl} A{}^{kl} - \ft{1}{24} A_{n}{}^{jkl} {A{}^n}_{jkl}
- \ft{4}{3} B_{kl} B^{kl} \right)
\;.
\label{eom:Einstein}
\end{eqnarray}
and
\begin{eqnarray}
({\cal E}_{\rm vector})^{\mu}{}_{ij} &=&
{\cal D}_\nu {\cal G}^{+\nu\mu}{}_{ij} +
{\cal P}_{\nu\,ijkl}\,{\cal G}^{-\,\nu\mu\,kl}
-\ft13
\left(
A_{[i}{}^{nkl} {\cal P}^\mu{}_{j]nkl}+
4 B^{kl}{\cal P}^\mu{}_{ijkl}
\right)
\;.
\label{eom:vector}
\end{eqnarray}
The modified Einstein equations show that the presence of the tensor $B_{ij}$ induces a positive contribution to the effective cosmological constant. This is a typical feature of the theories with trombone gauging~\cite{Howe:1997qt,Lavrinenko:1997qa,Chamblin:2001dx,Bergshoeff:2002nv,Kerimo:2003am,Kerimo:2004md,LeDiffon:2008sh}.
\subsection{Scalar field equations}
We start from the following ansatz for the equations of motion for the
spin-$1/2$ fermion fields $\chi_{ijk}$
\begin{eqnarray}
0 ~=~ ({\cal E}_\chi)_{ijk} &\equiv& - \ft{1}{6} \gamma^{\mu} {\cal D}_{\mu} \chi_{ijk}
- \ft{\sqrt{2}}{6} \gamma^{\nu} \gamma^{\mu} \psi_{\nu}{}^l {\cal P}_{\mu\; ijkl}
+ \ft{1}{8} \gamma_\rho\gamma_{\mu\nu} \psi^{\rho}{}_{[k} {\cal G}^{+ \mu\nu}{}_{ij]}
\nonumber\\[.5ex]
&& {}
+ \ft{\sqrt{2}}{288} \epsilon_{ijklmnpq} \gamma^{\mu\nu} \chi^{lmn} {\cal G}^{-}{}_{\mu\nu}{}^{pq}
- \ft{1}{6} {A{}^l}{}_{ijk} \gamma^{\mu} \psi_{\mu\; l}
+ 2 A_{ijk,\, lmn} \chi^{lmn}
\nonumber\\[.5ex]
&& {}
-\ft13 \gamma^{\mu} \psi_{\mu[k} B_{ij]} + \ft{\sqrt{2}}{36} \epsilon_{ijklmnpq} \chi^{lmn} B^{pq}
\;,
\label{eom:fermions}
\end{eqnarray}
with the scalar tensor
\begin{eqnarray}
A{}_{ijk,lmn} = \ft1{144} \sqrt{2}\,
\epsilon_{ijkpqr[lm}\,A_{n]}{}^{pqr}
\;.
\end{eqnarray}
Again, up to the last two terms whose structure is determined by $SU(8)$,
equations (\ref{eom:fermions}) follow from varying
the Lagrangian of~\cite{deWit:1982ig,deWit:2007mt}. The new coefficients are
fixed by compatibility with supersymmetry and follow as in the last section
by imposing the vanishing of the linear terms of the form $B {\cal D} \epsilon$
and $B {\cal P} \epsilon$ in the supersymmetry
variation of (\ref{eom:fermions}).
Again we first focus on the part of the supersymmetry variation of $({\cal E}_\chi)_{ijk}$
which is bilinear in the scalar tensors $A$, $B$ and find
\begin{eqnarray}
&&-\; \frac{\sqrt{2}}{3} \;
\Big\{
2{A{}^r}{}_{ijk} A_{lr}
+4 {A{}^r}{}_{ijk} B_{lr}
+4 B_{[ij} A_{k]l}
-8 B_{[ij} B_{k]l}
-\frac{1}{9} \epsilon_{ijklrmnp} A_{q}{}^{rmn}B^{pq}
\nonumber\\[.5ex]
&&{}
\qquad\qquad
+\frac{1}{12} \epsilon_{ijkrmnpq} A_{l}{}^{pqs}
(A_{s}{}^{rmn} +\ft83 \delta_s^r B^{mn} )
+\frac13 \epsilon_{ijklmnpq} B^{mn} B^{pq}
\Big\} \;\times\;
\epsilon^l
\;.
\qquad\quad
\end{eqnarray}
Upon adding a proper linear combination of the two quadratic constraints (\ref{constraints378})
in the ${\bf 378}$, this expression reduces to
\begin{eqnarray}
&&-\; \frac{\sqrt{2}}{3} \;
\Big\{
2{A{}^r}{}_{[ijk} A_{l]r}
+4 {A{}^r}{}_{[ijk} B_{l]r}
-8 B_{[ij} B_{kl]}
\nonumber\\[.5ex]
&&{}
\qquad\qquad\qquad\qquad\qquad
+\frac{1}{16} \epsilon_{ijklmnpq} \left(A_{r}{}^{pqs} A_{s}{}^{mnr}
+\ft{16}3 B^{mn} B^{pq}\right)
\Big\} \;\times\;
\epsilon^l
\;,
\qquad\quad
\end{eqnarray}
which is manifestly antisymmetric in $[ijkl]$\,. Finally, the combination of quadratic constraints (\ref{QCcomb2}) can be
used to simplify this expression to the manifestly self-dual expression
\begin{eqnarray}
&&-\; \frac{\sqrt{2}}{3} \;
\Big\{\,
{\cal C}_{ijkl} + \ft1{24} \epsilon_{ijklmnpq} {\cal C}^{mnpq}\,
\Big\} \;\times\;
\epsilon^l
\;,
\qquad\quad
\end{eqnarray}
with the tensor
\begin{eqnarray}
{\cal C}_{ijkl}&=&
{A{}^m}_{[ijk} A_{l]m}
+ \ft34 {A{}^m}_{n[ij} {A^n}_{kl]m}
+ 2 {A{}^m}_{[ijk} B_{l]m}
\;.
\label{defC}
\end{eqnarray}
This expression will be part of the modified scalar field equations.
After some more calculation, and using the first order field equations
(\ref{eom:1order}), the
full supersymmetry variation of the fermionic field equation $({\cal E}_\chi)_{ijk}$
eventually takes the form
\begin{eqnarray}
\delta_\epsilon ({\cal E}_\chi)_{ijk} &=&
\ft{\sqrt{2}}3 \,({\cal E}_{\rm scalars})_{ijkl}\,\epsilon^l
-\gamma_\mu \,({\cal E}_{\rm vector})^{\mu}{}_{[ij}\,\epsilon_{k]}
\;.
\end{eqnarray}
with the vector field equations from (\ref{eom:vector}) and the full scalar field equations
given by
\begin{eqnarray}
({\cal E}_{\rm scalars})_{ijkl} &=&
{\cal D}^\mu{\cal P}_{\mu\,ijkl} -
\ft32
{\cal G}^+_{\mu\nu\,[ij}{\cal G}^{+\,\mu\nu}{}_{kl]}
-\ft1{16}\epsilon_{ijklpqrs}
{\cal G}^{-\mu\nu\,pq}{\cal G}^{-}{}_{\mu\nu}{}^{rs}
\nonumber\\
&&{}
-{\cal C}_{ijkl}-\ft1{24}\epsilon_{ijklpqrs}\, {\cal C}^{pqrs}
\;.
\label{eom:scalars}
\end{eqnarray}
We note that the term bilinear in the tensor $B_{ij}$ has dropped out from the
scalar field equations. Also this is a characteristic feature for theories with trombone gauging.\footnote{
Let us correct a misprint in reference \cite{LeDiffon:2008sh}:
the last term in equation (4.44) of that reference is in fact absent in accordance with the equation obtained
by dimensional reduction of (\ref{eom:scalars}).
}
As a consequence, for pure trombone gaugings
($\Theta_M{}^\alpha=0$, implying that $A_{ij}=0=A_i{}^{jkl}$)
a simple solution to the bosonic field equations is given by
a de Sitter geometry with all scalar and vector fields vanishing.
We shall discuss this solution in more detail below.
Let us finally note, that due to the presence of the mixed term ${A{}^m}_{[ijk} B_{l]m}$ the scalar field
equations (\ref{eom:scalars}) cannot be integrated up to a scalar potential.
\subsection{Yang-Mills equations}
The vector field equations of motion (\ref{eom:vector}) can be rewritten equivalently as
\begin{eqnarray}
{\cal V}_M{}^{ij} \, {\cal D}_{[\mu}\,{\cal G}_{\nu\rho]}{}^{M} &=&
-\ft19e\varepsilon_{\mu\nu\rho\sigma}\,
\Big\{
A^{[i}{}_{nkl} {\cal P}^{\sigma\,j]nkl}+
4 B_{kl}{\cal P}^{\sigma\,ijkl}
\Big\}
\;,
\label{eom:vector0}
\end{eqnarray}
or
\begin{eqnarray}
{\cal D}_{[\mu}\,{\cal G}_{\nu\rho]}{}^{M} &=& \ft19 e\varepsilon_{\mu\nu\rho\sigma}
Z^{M\alpha} (t_{\alpha})^{KL} {\cal V}_{K}{}^{ij}{\cal V}_{L}{}^{kl} {\cal P}^{\sigma}{}_{ijkl}
\;,
\label{eom:vector1}
\end{eqnarray}
with the tensor $Z^{M\alpha}$ from (\ref{defZZ}).
Since in the derivation of the field equations we have also come across the first order
duality equations (\ref{eom:1order}) for the vector fields, an immediate question is the
compatibility of these equations with the second order field equations. Upon contracting
equation (\ref{eom:vector1}) with $\vartheta_M$ or $X_M$, its r.h.s.\ vanishes by virtue of (\ref{QQZ})
while on the l.h.s.\ the first order duality equations (\ref{eom:1order}) allow to replace
${\cal G}_{\nu\rho}^{M}$ by the covariant field strength ${\cal H}_{\nu\rho}^{M}$. Then also
the l.h.s.\ vanishes by virtue of the Bianchi identity (\ref{Bianchi}) and (\ref{QQZ}).
Both sets of equations are thus compatible.
As another consistency check, we observe that
upon applying the operator $\varepsilon^{\mu\nu\rho\tau} {\cal D}_\tau$,
the l.h.s.\ of (\ref{eom:vector1}) reduces to
\begin{eqnarray}
- \ft12
\varepsilon^{\mu\nu\rho\tau} \,
{\cal G}^K_{\mu\nu} {\cal G}_{\rho\tau}^{L} \, X_{KL}{}^M
&=&
- Z^{M\alpha}\,(t_{\alpha})_{KL}\,
\left(
{\cal G}^{+\;K}_{\mu\nu} {\cal G}^{+\mu\nu\;L}
+{\cal G}^{-\;K}_{\mu\nu} {\cal G}^{-\mu\nu\;L}
\right)
\;,
\end{eqnarray}
whereas the r.h.s.\ contains the scalar field equation (\ref{eom:scalars}).
This provides a number of important non-trivial consistency checks on the set of bosonic
field equations that we have obtained from supersymmetry variation,
but not derived from an action.
\subsection{Maximally symmetric solutions and mass matrices}
\label{subsec:masses}
According to (\ref{eom:scalars}), a solution to the field equations with constant scalar
and vanishing vector fields requires the anti-selfduality condition
\begin{eqnarray}
{\cal C}_{ijkl}+\ft1{24}\varepsilon_{ijklpqrs}\, {\cal C}^{pqrs} &=& 0\;,
\label{extremum}
\end{eqnarray}
for the scalar field dependent
tensor ${\cal C}_{ijkl}$ from (\ref{defC}). For the standard gaugings, this is precisely
the condition for an extremal point of the scalar potential~\cite{deWit:1982ig}.
In presence of the local scaling symmetry, however, we recall that the scalar field equations
can in general not be integrated up to a scalar potential.
Any solution to (\ref{extremum}) yields a solution to the field equations with maximally
symmetric four-dimensional spacetime and cosmological constant
\begin{eqnarray}
\Lambda&=&
-\ft{3}{4} A_{kl} A{}^{kl} + \ft{1}{24} A_{n}{}^{jkl} {A{}^n}_{jkl}
+ \ft{4}{3} B_{kl} B^{kl}
\;.
\label{coco}
\end{eqnarray}
The spectrum of the theory around this solution can be obtained by linearizing
the field equations.
Using (\ref{actionE7}), linearization of
the scalar field equations (\ref{eom:scalars})
around a solution of (\ref{extremum}) yields to lowest order
\begin{eqnarray}
\Box\, \phi_{ijkl} &=&
{\cal M}_{ijkl}{}^{mnpq}\,\phi_{mnpq} + {\cal O}(\phi^2)
\;,
\end{eqnarray}
with self-dual scalar fields $\phi_{ijkl}=\ft1{24}\varepsilon_{ijklpqrs}\, \phi^{pqrs}$ and
the scalar mass matrix ${\cal M}_{ijkl}{}^{mnpq}$
whose symmetric part is given by
\begin{eqnarray}
{\cal M}_{ijkl}{}^{mnpq}\,\phi^{ijkl}\phi_{mnpq} &=&
6\left(A_{m}{}^{ijk}A^{l}_{ijn}\!-\!\ft14A_{i}{}^{jkl}A^{i}_{jmn}\!-\!
A_{m}{}^{ikl} B_{in}\!-\!A^{k}{}_{imn}B^{il}\right)\phi^{mnpq}\phi_{klpq}
\nonumber\\
&&{}
+\left(\ft5{24}\,A_{i}{}^{jkl}A^{i}{}_{jkl}-\ft12A_{ij}A^{ij}\right) \phi^{mnpq}\phi_{mnpq}
\nonumber\\
&&{}
-\ft23\,A_{i}{}^{jkl} A^{m}{}_{npq} \, \phi^{inpq}\phi_{jklm}
\;,
\label{Mscalar_sym}
\end{eqnarray}
while its antisymmetric part reads
\begin{eqnarray}
{\cal M}_{ijkl}{}^{mnpq}\,\phi_{[1}^{ijkl}\phi^{\vphantom{j}}_{2]\,mnpq} &=&
\ft83 \left(B_{rs}A_{i}{}^{rs[m}-B^{rs}A^{m}{}_{rs[i}\right)
\phi_{[1}^{ijkl}\phi^{\vphantom{j}}_{2]\,mjkl}
\nonumber\\
&&{}
-4 \left(A_{i}{}^{mnp} B_{jk}-A^{m}{}_{ijk} B^{np} \right)
\phi_{[1}^{ijkl}\phi^{\vphantom{j}}_{2]\,mnpl}
\;.
\label{Mscalar_alt}
\end{eqnarray}
The calculation of (\ref{Mscalar_sym}), (\ref{Mscalar_alt})
makes use of identities for self-dual tensors,
such as those given in~\cite{deWit:1978sh}
as well as
of the quadratic constraints derived in appendix~\ref{app:Tids}.
For $B^{ij}=0$, the mass matrix consistently reduces to the
expression given in~\cite{deWit:1983gs}
and its antisymmetric part vanishes.
For the vector field equations (\ref{eom:vector}) we find the
linearized form
\begin{eqnarray}
\Box\, {\cal A}_{\mu\,ij}
&=&
\ft23
\left(
A_{[i}{}^{nkl} -
4 \delta_{[i}^n B^{kl}
\right) (T^{pq})_{j]nkl} \, {\cal A}_\mu{\,}_{pq}
+\ft23
\left(
A_{[i}{}^{nkl} -
4 \delta_{[i}^n B^{kl}
\right)
(T_{pq})_{j]nkl}\,{\cal A}_\mu^{\,pq}
\;,
\nonumber
\end{eqnarray}
from which using (\ref{TX}), (\ref{TAB}) we read off the vector mass matrix
\begin{eqnarray}
{\cal M}_{\rm vec} &=&
\left(
\begin{array}{cc}
{\cal M}_{ij}{}^{kl} & {\cal M}_{ijkl}
\\
{\cal M}^{ijkl}&
{\cal M}^{ij}{}_{kl}
\end{array}
\right)
\;,
\label{M_vector}
\end{eqnarray}
with
\begin{eqnarray}
{\cal M}_{ij}{}^{kl}
&=&
-\ft16 A_{[i}{}^{npq} \delta_{j]}^{[k} A^{l]}{}_{npq}
+\ft12 A_{[i}{}^{pq[k}A^{l]}{}_{j]pq}
+\ft23 \delta^{[k}_{[i} A_{j]}{}^{l]pq} B_{pq}
-\ft23 A_{[i}{}^{nkl} B_{j]n}
\nonumber\\
&&
-\ft43 \delta_{[i}^{[k} A^{l]}{}_{j]pq} B^{pq}
+\ft43 A^{[k}{}_{nij} B^{l]n}
-\ft{8}9 \delta^{kl}_{ij} B^{pq} B_{pq}
-\ft{8}9 B^{kl} B_{ij}
+\ft{32}9 B^{n[k} \delta^{l]}_{[i} B_{j]n}
\;,
\nonumber\\[2ex]
{\cal M}_{ijkl} &=&
\ft1{36} A_{[i}{}^{pqr} \epsilon_{j]pqrmns[k} A_{l]}{}^{mns}
-\ft1{18} \epsilon_{klmnpqr[i} A_{j]}{}^{pqr} B^{mn}
\nonumber\\
&&
+\ft1{9}\epsilon_{ijpqrmn[k} A_{l]}{}^{pqr} B^{mn}
-\ft2{9} \epsilon_{ijklmnpq} B^{mn} B^{pq}
\;.
\end{eqnarray}
Finally, the gravitino and fermion mass matrices are
directly read off from (\ref{eomgrav}) and (\ref{eom:fermions}),
respectively and take the form
\begin{eqnarray}
{\cal M}_\psi{}^{ij} &=& \sqrt{2}\left(A^{ij}-2B^{ij}\right)
\;,\nonumber\\[.5ex]
{\cal M}_\chi{}^{ijk,lmn} &=&
\ft1{12} \sqrt{2}\,\left(
\epsilon^{ijkpqr[lm}\,A^{n]}{}_{pqr}
+2 \epsilon^{ijklmnpq} B_{pq} \right)
\;,
\label{Mferm}
\end{eqnarray}
where the first matrix carries the information
about the breaking of supersymmetry and the
latter matrix has to be evaluated after projecting out the fermions
that are eaten by the massive gravitinos.
\section{Example: de Sitter geometry and mass spectrum}
\label{sec:example}
We have in the previous section derived the full set of fermionic and bosonic field equations of the gauging in presence of the trombone generator and shown that they transform into each other under supersymmetry. In absence of an action, these equations capture the full dynamics of the theory.
As a simple example and application of the construction, in this section we analyze in more detail the gauging discussed at the end of section~\ref{sttqc}, parametrized by $\kappa$ and $\Xi^a$\,. In particular, we show that this theory admits a de Sitter solution with constant scalar fields and work out its mass spectrum by linearizing the equations of motion around the vacuum solution.
In the absence of the trombone gauging, i.e.\ for $\kappa=0$,
the theory is characterized by an $E_{6(6)}$ generator $\Xi^a$
and corresponds to the Scherk-Schwarz reduction from five dimensions
first analyzed in~\cite{Cremmer:1979uq,Sezgin:1981ac} and revisited in
the context of four-dimensional gaugings in~\cite{Andrianopoli:2002mf}.
As a first step, we calculate for this theory the value of the tensors
$A_{ij}$, $A_i{}^{jkl}$ and $B^{ij}$ at the origin, i.e.\ for all scalar fields vanishing.
Since at the origin, the group $E_{6(6)}$ is broken down to its
maximally compact subgroup, the values of these tensors will be expressed
in terms of the $USp(8)$ building blocks $(\kappa, \xi^{ij}, \xi^{ijkl})$,
transforming in the $1$, $36$, and $42$ of $USp(8)$, respectively,
of which the latter two compose the $E_{6(6)}$ generator $\Xi^a$.
The indices $i, j, \dots = 1, \dots 8,$ here label the fundamental
representation of $USp(8)$. Explicitly, these tensors satisfy the relations
\begin{eqnarray}
\xi^{ij}=\xi^{ji}\;,\quad
\xi^{ijkl}=\xi^{[ijkl]}\;,\quad
\xi^{ijkl}\omega_{kl}=0
\;,
\end{eqnarray}
with the $USp(8)$-invariant symplectic matrix $\omega_{ij}$, and the reality properties
\begin{eqnarray}
(\xi^{ij})^*= \xi_{ij}=\omega_{ik}\omega_{jl}\xi^{kl}\;,\qquad
\mbox{etc.}
\end{eqnarray}
At the origin, the scalar tensors $A_{ij}$, $A_i{}^{jkl}$ and $B^{ij}$
take the form
\begin{eqnarray}
A^{ij} = \ft{1}{\sqrt{2}}
\xi^{ij}
\;,\quad
A_i{}^{jkl} =
-\ft{3}{\sqrt{2}} \omega_{im} \xi^{m[j} \omega^{kl]}
+\omega_{im}\xi^{mjkl}
\;,\quad
B^{ij} =\ft{1}{\sqrt{2}}
\kappa\, \omega^{ij}
\;.
\label{ABk}
\end{eqnarray}
The condition for extremality (\ref{extremum})
coming from the scalar field equations
splits into the equations
\begin{eqnarray}
\kappa\,\xi^{ijkl}&=& \sqrt{2}\omega_{mn} \xi^{m[i}\xi^{jkl]n}
\;,\qquad
\xi^{ijkl}\xi_{ijkl}~=~0\;.
\end{eqnarray}
Obviously, even for non-vanishing parameter $\kappa$ these equations leave no other solution than
$\xi^{ijkl}=0$, i.e.\ induce a Scherk-Schwarz gauging with a compact generator
of $E_{6(6)}$\,. On the other hand this shows that choosing $\xi^{ijkl}=0$ suffices to guarantee that the
scalar field equations (\ref{eom:scalars}) are solved by setting all scalar fields to zero.
For the cosmological constant (\ref{coco}), we obtain
\begin{eqnarray}
\Lambda
&=&
\ft{3}{8} \xi^{ijkl} \xi_{ijkl}
+ \ft{16}{3} \kappa^2 ~=~ \ft{16}{3} \kappa^2\;,
\end{eqnarray}
i.e.\ the Einstein field equations (\ref{eom:Einstein}) are solved by
a Minkowski space for the standard gaugings and by
a de Sitter geometry for non-vanishing $\kappa$.
The fermionic mass spectrum for this solution is obtained by linearizing the fermionic
field equations (\ref{eomgrav}), (\ref{eom:fermions}) around the de Sitter background
with the mass matrices given by (\ref{Mferm}).
For the eight gravitino masses we obtain
\begin{eqnarray}
m_{\rm grav} &:& \pm\sqrt{m_i^2 +4\kappa^2}\;,\qquad
i=1, \dots, 4
\;,
\end{eqnarray}
where we have denoted by $im_i$ the eigenvalues of the anti-hermitean matrix $\xi_i{}^j$\,.
For non-vanishing $\kappa$ thus all supersymmetries are broken, as
is required by the de Sitter geometry and 8 Goldstinos are found among the
spin-1/2 fermions.
For vanishing $\kappa$ on the other hand, supersymmetry is broken according to
the number of non-vanishing eigenvalues of $\xi_i{}^j$, in accordance with the
results of~\cite{Cremmer:1979uq,Sezgin:1981ac}.
The remaining fermion masses are given by
\begin{eqnarray}
m_{\rm ferm} &:&
\pm\sqrt{m_i^2 +4\kappa^2}\;,\quad
\pm\sqrt{(m_i \pm m_j \pm m_k)^2 +4\kappa^2}\;,\;\;
(i<j<k)
\;.
\end{eqnarray}
We find that the effect of the additional trombone gauging is a shift
in all the fermion masses.
Likewise, we may calculate the scalar mass matrix (\ref{Mscalar_sym}), (\ref{Mscalar_alt}),
with (\ref{ABk}) and obtain
\begin{eqnarray}
{\cal M}_{ijkl}{}^{mnpq} &=&
\mathbb{P}_{42} \left(
\xi^r{}_s \xi^s{}_r\,\delta^{mnpq}_{ijkl}
-24 \xi^{[m}{}_{[i}\,\xi^n{}_j\,\delta^{pq]}_{kl]}
+32\,\kappa\,\xi^{[m}{}_{[i}\,\delta^{npq]}_{jkl]}
\right)
\mathbb{P}_{42}
\;,
\label{massesC4}
\end{eqnarray}
where $\mathbb{P}_{42}$ refers to the
projector onto the 42 scalars in the decomposition
\begin{eqnarray}
{\bf 70} &\longrightarrow& {\bf 1}+{\bf 27}+{\bf 42}
\;,
\end{eqnarray}
of $SU(8)$ under $USp(8)$, i.e.\
all other scalars come with zero mass.
In (\ref{massesC4}), the last term lives entirely in the antisymmetric part
of the mass matrix.
In terms of the eigenvalues of $\xi^m{}_n$, diagonalization of (\ref{massesC4})
leads to the following spectrum
\begin{eqnarray}
0&||& 30 \times\nonumber\\
(m_i \pm m_j)^2 + 4 i \kappa |m_i \pm m_j| &||& i<j \nonumber\\
(m_i \pm m_j)^2 - 4 i \kappa |m_i \pm m_j|&||& i<j \nonumber\\
(m_1\pm m_2 \pm m_3 \pm m_4)^2 + 4 i \kappa |m_1\pm m_2 \pm m_3 \pm m_4 | &||& \nonumber\\
(m_1\pm m_2 \pm m_3 \pm m_4)^2 - 4 i \kappa |m_1\pm m_2 \pm m_3 \pm m_4 | &||&
\qquad
\;,
\label{spec_scalars}
\end{eqnarray}
for the masses of the scalar fields.
For vanishing $\kappa$, we precisely reproduce the mass spectrum of \cite{Cremmer:1979uq}.
Upon switching on~$\kappa$, all non-vanishing mass-eigenvalues degenerate according to
$m^2 \rightarrow m^2 \pm 4i\kappa m$\,.
The fact that most of the mass eigenvalues come out to be
imaginary is due to the antisymmetric contributions
to the mass matrix.
Finally, the vector mass matrix (\ref{M_vector}) takes the form
\begin{eqnarray}
{\cal M}_{ij}{}^{kl} &=&
- \xi^{[k}{}_{[i} \xi^{l]}{}_{j]}
+ \xi^{[k}{}_{n} \delta^{l]}_{[i}\, \xi^n{}_{j]}
+ 4 \kappa \delta^{[k}_{[i} \xi^{l]}{}_{j]}
-\ft{16}{9} \kappa^2 \delta_{ij}^{kl}
-\ft49 \kappa^2 \,\omega_{ij} \omega^{kl}
\;,
\nonumber\\[2ex]
{\cal M}_{ijkl} &=&
\xi_{i[k}\xi_{l]j}
-\ft12\omega_{i[k} \xi^n{}_{l]} \xi_{jn}
+\ft12\omega_{j[k} \xi^n{}_{l]} \xi_{in}
-2\kappa \omega_{i[k} \xi_{l]j} + 2\kappa \omega_{j[k} \xi_{l]i}
\nonumber
\\
&&{}
-\ft{8}9 \kappa^2 \omega_{ij}\omega_{kl}
-\ft{16}9 \kappa^2 \omega_{i[k} \omega_{l]j}
\;.
\label{MMvk}
\end{eqnarray}
In terms of the eigenvalues of the matrix $\xi^m{}_n$, we find the following mass spectrum
\begin{eqnarray}
0&||& 28 \times\nonumber\\
(m_i \pm m_j)^2 + 4 i \kappa |m_i \pm m_j| -\ft{32}{9} \kappa^2 &||& i<j \nonumber\\
(m_i \pm m_j)^2 - 4 i \kappa |m_i \pm m_j| -\ft{32}{9} \kappa^2 &||& i<j \nonumber\\
-\ft{32}{9} \kappa^2 &||& 3 \times \nonumber\\
-\ft{32}{3} \kappa^2 &||& 1 \times
\;.
\end{eqnarray}
For non-vanishing $\kappa$ thus all 28 vector fields become massive.
The associated massless Goldstone bosons can be identified in the scalar spectrum
(\ref{spec_scalars}) which provides a strong consistency check of the result.
The matrix (\ref{MMvk}) has half-maximal rank in accordance with the fact
that this gauging is purely electric and involves only 28 of the vector fields.
To summarize, we have shown that for the theory discussed at the end of section~\ref{sttqc},
a de Sitter geometry with constant scalar and
vector fields provides a solution to the full set of field equations. While the
a non-vanishing $\kappa$ in the fermionic sector simply induces a shift in all the fermion
masses, we find that in the bosonic sector, most of the modes have imaginary mass square
eigenvalues. This is due to the fact that the equations of motion do not descend from an action
and may be a sign of an instability of this solution in de Sitter space.
Actually the imaginary shift in the mass squared of the bosonic fluctuations reminds of the Breit-Wigner formula\footnote{We are grateful to Riccardo D'Auria for pointing out this analogy.}
for the propagator of an unstable particle, which has a characteristic imaginary shift in the pole proportional to the particle decay width $\Gamma$:
\begin{eqnarray}
\frac{1}{p^2-m^2+im\Gamma}\,.
\end{eqnarray}
From this point of view our results seem to suggest that the bosonic fluctuations ``die off'' at some characteristic time $\delta t\sim m\Gamma/E$ proportional to the trombone parameter $\kappa$. It would be interesting to understand the implications of this feature for the stability properties of the background. \par
A particular limit of this theory is the case of a `pure trombone gauging', i.e.\ $\xi^m{}_n=0$
with vanishing mass parameters $m_i$\,. It follows from the above formulas, that in this case
all scalar fields remain massless while all vector fields appear with negative mass squares,
again a sign of an instability of the solution. It will be interesting to analyze in more detail, if
this instability is a generic feature of the theories with local scaling invariance or if some classes
of theories among the more complicated gaugings constructed in this paper eventually admit
stable vacuum solutions.
\section{Conclusions}
In this paper, we have derived the most general couplings of four-dimensional supergravity
with a maximal number of supercharges. With a gauge group embedded in the $E_{7(7)} \times \mathbb{R}$ global symmetry group of the Cremmer-Julia theory, the gauge generators are parametrized in terms of an embedding tensor, carrying $56+912$ parameters, subject to a set of bilinear algebraic consistency constraints. After suitable parametrization, we find that the latter reduces to the system (\ref{cs1})--(\ref{cs4}) which allows to construct simple solutions.
The standard gaugings whose gauge group is a subgroup of $E_{7(7)}$ correspond to an embedding tensor in the irreducible ${\bf 912}$ representation. Additional non-vanishing components in the ${\bf 56}$ representation define theories in which local scaling invariance $\mathbb{R}$ (the so-called trombone symmetry) is part of the gauge group.
We have determined the general form of the gauge groups and worked out the full set of modified field equations of these gauged $N=8$ supergravities. As a particular feature of these theories, we have found that a gauging of the trombone generator leads to an additional positive contribution to the effective cosmological constant. Moreover, it turns out that gaugings with local scaling symmetry are generically dyonic, i.e.\ involve simultaneously electric and magnetic vector fields. We have analyzed in detail the simplest example of such a theory which has its higher-dimensional origin as a generalized Scherk-Schwarz reduction from five dimensions. We have shown that this theory admits a de Sitter solution with constant scalar fields and determined its mass spectrum indicating that the solution is not stable.
While in this paper we have analyzed only a single example within the new class of theories, which describes a one-parameter deformation of the known Scherk-Schwarz gaugings from five dimensions,
it would be highly interesting to generalize this analysis to other examples and to perform a systematic study of the possibilities of deformations of the known gaugings. In particular, starting from a theory with supersymmetric AdS vacuum, the additional positive contribution to the effective cosmological constant may lift the space-time geometry to Minkowski or de Sitter upon inclusion of the trombone generator.
In this context it would be important to investigate if the instabilities that we have found in the bosonic spectrum of our example are due to the simple structure of that example or if they persist to more complex situations and represent a generic feature of these theories.
Another interesting aspect for further study is the dyonic structure of the constructed gaugings.
Whereas the appearance of magnetic vector fields is not new and has shown up in previously constructed gaugings in four dimensions~\cite{deWit:2005ub,deWit:2007mt}, in the standard theories there is always a symplectic frame in which all magnetic vector fields drop from the action and the field equations. This frame can be reached in a systematic way by integrating out the two-forms from the action.
In contrast, we have found that for the gaugings constructed in this paper there is in general no symplectic frame in which all gauge fields would be electric. These gaugings are of genuinely dyonic nature.
This does not contradict the no-go results on the gauging of electric/magnetic dualities~\cite{Bunster:2010wv,Deser:2010it}, as the resulting theories do no longer admit an action. It would be highly interesting to study the structure of such dyonic theories in more detail.
\bigskip
It is certainly remarkable that maximal supersymmetry in four dimensions
not only admits the standard gaugings with gauge groups inside $E_{7(7)}$,
described by an embedding tensor in the ${\bf 912}$ representation \cite{deWit:1982ig,deWit:2002vt,deWit:2007mt}, but moreover allows for yet another non-trivial deformation of the
field equations described by 56 additional components of the embedding tensor.
On the other hand, this may be viewed as another sign of the underlying symmetry
structure of extended supergravity theories: upon dimensional reduction to two
dimensions, the global symmetry
group of maximal ungauged supergravity is the affine group $E_{9(9)}$~\cite{Julia:1981wc}
while its gaugings are parametrized by an embedding tensor $\Theta_{\rm 2-dim}$ transforming in the
basic representation of that group~\cite{Samtleben:2007an}.
This infinite-dimensional highest-weight representation
thus captures all deformation parameters of the two-dimensional theory. Decomposition w.r.t.\
the finite-dimensional subgroup $E_{7(7)}\times SL(2)$ gives rise to its lowest level components
\begin{eqnarray}
\Theta_{\rm 2-dim} \quad&\longrightarrow&\qquad
\begin{tabular}{r|c}
$+1$&$(1,2)$\\\hline
$+2$&$(56,2)$\\\hline
$+3$&$(133,2)+(1,2)$\\\hline
$+4$&$\;\;\;(912,1)+(56,1)+(56,3)\;\;\;$\\\hline
\dots&\dots
\end{tabular}\quad
\;,
\label{basic}
\end{eqnarray}
from which the higher-dimensional origin of these theories may be inferred.
E.g.\ the theories described by parameters in the
first two rows correspond to torus reductions from four to two dimensions,
in which the KK vector field and the two-dimensional vector fields acquire non-vanishing
flux components along the two-torus, with the corresponding deformation parameters
transforming in the $(1,2)$ and the $(56,2)$, respectively.
Parameters in the third row describe Scherk-Schwarz reduction from four to two dimensions,
including twists with the four-dimensional trombone generator.
The $(912,1)$ in the fourth row corresponds to theories obtained by dimensional reduction
from the standard gaugings in four dimensions, while the $(56,1)$ describes the
dimensional reduction of the theories with local scaling symmetry constructed in this paper.
This shows that after dimensional reduction both the standard gaugings as well as trombone gaugings and combinations of the two are described on equivalent footing by parameters residing within a single irreducible representation of the affine global symmetry group.
In this sense, the new gaugings constructed in this paper may be viewed as obtained by $E_{9(9)}$ rotation from the standard gaugings in four dimensions.
Moreover, the infinite tail of higher level parameters in (\ref{basic}) still advocates the tempting
possibility of discovering yet other maximally supersymmetric couplings in four dimensions
which must however be of genuinely different nature than the present constructions.
\subsection*{Acknowledgments}
The work of H.S.\ is supported in part by the Agence Nationale de la Recherche (ANR).
M.T.\ wishes to thank Riccardo D'Auria for interesting discussions.
Part of the calculations has been facilitated by use of the computer algebra
system Cadabra~\cite{Peeters:2006kp,Peeters:2007wn}.
\newpage
\section*{Appendix}
\begin{appendix}
\section{Some algebra}
\label{app:algebra}
\subsection{Useful $E_7$ relations}
We denote by $(t_\alpha)_M{}^N$ the $E_{7(7)}$ generators in the fundamental
representation, i.e.\ the index $\alpha$ runs over $1, \dots, 133$ and $M, N = 1, \dots, 56$\,.
We raise and lower adjoint indices
with the invariant metric $\kappa_{\alpha\beta}\equiv {\rm Tr}\,[t_{\alpha}t_{\beta}]$,
which is a rescaled Cartan-Killing metric.
Fundamental indices are raised and lowered with
the symplectic matrix $\Omega^{MN}$ using north-west south-east conventions:
$X^M=\Omega^{MN}X_N$, etc.\,. We note the following two useful algebraic identities:
\begin{eqnarray}
(t^\alpha)_M{}^{K} (t_\alpha)_N{}^{L} &=&
\ft1{24}\delta_M^{K}\delta_N^{L} + \ft1{12}\delta_M^{L}\delta_N^{K}
+(t^{\alpha})_{MN}\,(t_{\alpha})^{KL}-\ft1{24}\Omega_{MN}\,\Omega^{KL}
\;,
\label{relD4}
\end{eqnarray}
and
\begin{eqnarray}
(t^\alpha)_{KL} (t_{\alpha})_{MN} &=& \ft1{12} \Omega_{K(M}\Omega_{N)L}+C_{KLMN}
\;,
\end{eqnarray}
with the quartic $E_7$ invariant $C_{KLMN}\equiv(t_\alpha)_{(KL} (t^{\alpha})_{MN)}$\,.
\subsection{Breaking $E_{7(7)}$ to $SU(8)$}
Upon breaking $E_{7(7)}$ to its maximal compact subgroup $SU(8)$, the fundamental
and the adjoint representation break according to
\begin{eqnarray}
{\bf 56} \rightarrow {\bf 28} + {\overline{\bf 28}}\;,\qquad
{\bf 133} \rightarrow {\bf 63} + {\bf 70}
\;,
\end{eqnarray}
respectively. We label the fundamental representation of $SU(8)$ by indices $i, j, \dots = 1, \dots 8$\,.
Then, the $E_{7(7)}$ generators $(t_\alpha)_M{}^N$ break according to
\begin{eqnarray}
(t_i{}^j)_{mn}{}^{kl} &=& -\delta^j_{[m} \,\delta^{kl}_{n]i}-\ft18 \delta_i^j \,\delta_{mn}^{kl}
~=~
-(t_i{}^j)^{kl}{}_{mn}
\;,\nonumber\\[1ex]
(t_{ijkl})_{mnpq} &=& \ft1{24}\,\epsilon_{ijklmnpq}
\;,\qquad
(t_{ijkl})^{mnpq} ~=~ \delta_{ijkl}^{mnpq}
\;,
\end{eqnarray}
and the rescaled Cartan-Killing metric
$\kappa_{\alpha\beta}\equiv {\rm Tr}\,[ t_\alpha t_\beta ]$ breaks into
\begin{eqnarray}
\kappa_m{}^n,\,_p{}^q &=& 3\left(\delta_m^q\delta_p^n -\ft18 \delta_m^n\delta_p^q\right)
\;,\qquad
\kappa_{ijkl,mnpq} ~=~ \ft1{12} \epsilon_{ijklmnpq}
\;.
\end{eqnarray}
\subsection{Breaking $E_{7(7)}$ to $E_{6(6)}\times SO(1,1)$}
Upon breaking $E_{7(7)}$ to its maximal subgroup $E_6\times SO(1,1)$,
its lowest dimensional representations decompose according to
\begin{eqnarray}
{\bf 56} &\rightarrow&
{\bf 1}^{+3}+{\bf 27}^{+1}+\bar{\bf 27}_{-1}+{\bf 1}^{-3}
\;,
\nonumber\\
{\bf 133} &\rightarrow&
{\bf 1}^{0}+{\bf 78}^{0}+{\bf 27}^{-2}+\bar{\bf 27}^{+2}
\;,
\nonumber\\
{\bf 912} &\rightarrow&
{\bf 78}^{+3}+{\bf 78}^{-3}+{\bf 27}^{+1}+\bar{\bf 27}^{-1}
+{\bf 351}^{-1}+\bar{\bf 351}^{+1}
\;,
\nonumber\\
{\bf 1539} &\rightarrow&
{\bf 1}^{0}+{\bf 78}^{0}+{\bf 650}^{0}+{\bf 27}^{-2}+{\bf 27}^{+4}+\bar{\bf 27}^{+2}+\bar{\bf 27}^{-4}
+{\bf 351}^{+2}+\bar{\bf 351}^{-2}
\;,
\nonumber\\
{\bf 8645} &\rightarrow&
2\cdot{\bf 78}^{0}+{\bf 650}^{0}+{\bf 2925}^{0}
+{\bf 27}^{-2}+\bar{\bf 27}^{+2}
+{\bf 351}^{+2}+{\bf 351}^{-4}
+\bar{\bf 351}^{-2}+\bar{\bf 351}^{+4}
\nonumber\\
&&{}
+{\bf 1728}^{-2}+\bar{\bf 1728}^{+2}
\;,
\label{rep_break}
\end{eqnarray}
with the superscript indicating the $SO(1,1)$ charge.
We use the explicit notation
\begin{eqnarray}
{\bf 56}&:& X_M ~\rightarrow~ (X_\bullet,\; X_m,\; X^m,\; X^\bullet)
\;,\nonumber\\
{\bf 133}&:&X_\alpha ~\rightarrow~ (X_{\rm o},\; X_a,\; X_m,\; X^m)
\;,
\end{eqnarray}
with indices $m = 1, \dots, 27$ and $a = 1, \dots, 78$ labeling the
fundamental and the adjoint representation of $E_{6(6)}$, respectively.
The symplectic matrix $\Omega_{MN}$ and the
rescaled Cartan-Killing metric $\kappa_{\alpha\beta}$
break according to
\begin{eqnarray}
\Omega_{MN} &\rightarrow&
\left(\Omega_\bullet{}^\bullet=1,\;
\Omega_m{}^n=\delta_m^n,\;
\Omega^\bullet{}_\bullet=-1,\;
\Omega^m{}_n=-\delta^m_n
\right)
\;,
\end{eqnarray}
and
\begin{eqnarray}
\kappa_{\alpha\beta} &\rightarrow&
\left(
\kappa_{\rm oo}=72\;,\qquad\kappa_{ab}=2\eta_{ab}\;,\qquad\kappa_m{}^n=12\delta_m^n
\right)
\;.
\label{CK}
\end{eqnarray}
The $E_{7(7)}$ generators $(t_\alpha)_M{}^N$ decompose as
\begin{eqnarray}
&&
(t_{\rm o})_\bullet{}^\bullet ~=~3
\;,\qquad
(t_{\rm o})_m{}^n ~=~\delta_m^n
\;,\qquad
(t_{\rm o})^m{}_n ~=~-\delta^m_n
\;,\qquad
(t_{\rm o})^\bullet{}_\bullet ~=~-3
\;,
\nonumber\\
&&
(t_a)_m{}^n ~=~ - (t_a)^n{}_m
\;,
\nonumber\\
&&
(t_m)_\bullet{}^n ~=~ -(t_m)^n{}_\bullet ~=~ \delta_m^n
\;,\qquad
(t^m)^\bullet{}_n ~=~ -(t^m)_n{}^\bullet ~=~ -\delta^m_n
\;,
\nonumber\\
&&
(t_{m})_{nk}~=~ d_{mnk}
\;,\qquad
(t^{m})^{nk}~=~ 10\,d^{mnk}
\;.
\end{eqnarray}
and the decomposition of the structure constants $f_{\alpha\beta}{}^\gamma$
can be read off from the algebra
\begin{eqnarray}
&&{}[t_{\rm o}, t_m]=2t_m\;,\qquad
[t_{\rm o},t^m]=-2t^m\;,
\nonumber\\
&&{}[t_a,t_m]=- t_{am}{}^n t_n \;,\qquad
[t_a,t^m]= t_{an}{}^m t^n\;,
\nonumber\\
&&{}[t_m, t^n] = \frac13\,\delta_m^n\,t_{\rm o} -6 (t^a){}_m{}^n\,t_a
\;.
\end{eqnarray}
Here, $d_{mnk}$ denotes the totally symmetric tensor of $E_{6(6)}$,\,
and $ t_{am}{}^n$ denotes the $E_{6(6)}$ generators in the adjoint representation.
Adjoint indices are raised and lowered with the rescaled Cartan-Killing metric
$\eta_{ab}\equiv {\rm Tr}\,[ t_a t_b ]$\,.
\subsection{Useful $E_6$ relations}
We denote by $d_{mnk}$ and $d^{mnk}$
the totally symmetric tensors of $E_{6(6)}$ in the fundamental ${\bf 27}$ and $\overline{\bf 27}$,
respectively. We choose a relative normalization such that
\begin{eqnarray}
d^{mnp}d_{mnq}&=&\delta^p_q\;.
\end{eqnarray}
In the following, we give a list of useful algebraic relations
that be be shown by various contractions and/or by
using an explicit realization of the $E_{6(6)}$ generators:
\begin{eqnarray}
d_{mrs}\,d^{spt} \, d_{tnu} \,d^{urq} &=& \ft1{10}\,\delta_{(mn)}{}^{\!\!(pq)} -
\ft25\, d_{mnr}\, d^{pqr}\,, \\[1ex]
d_{mps}\,d^{sqt} \, d_{tru} \,d^{upv}\,d_{vqw}\,d^{wrn} &=& - \ft3{10}
\,\delta_{m}{}^{n}\,,
\\[1ex]
(t^a)_m{}^{k} (t_a)_n{}^{l} &=&
\ft1{18}\,\delta_m^{k}\delta_n^{l} +\ft1{6}\,\delta_m^{l}\delta_n^{k}
-\ft53\, d_{mnp}\,d^{klp}
\;,
\label{relD5}
\\[1ex]
(t_a)_r{}^p (t_b)_s{}^q \,d^{mrs} d_{npq} &=&
-\ft1{30}\,\eta_{ab}\delta^m_n+\ft25 (t_{(a}t_{b)})_n{}^m
\;,
\\[1ex]
(t_a t_b)_r{}^q \,d^{mrs} d_{nqs} &=&
\ft1{30}\,\eta_{ab}\delta^m_n-\ft15 (t_{a}t_b)_n{}^m+\ft3{10} (t_{b}t_a)_n{}^m
\;,
\\[1ex]
d_{pqr}d^{p(kl}d^{m)qs} &=& \ft1{30} d^{klm} \delta_r^s + \ft1{10} d^{s(kl} \delta^{m)}_r
\;.
\end{eqnarray}
These play a key role in reducing the the $E_{7(7)}$ system of constraints (\ref{Q1})--(\ref{Q3})
for the embedding tensor
to the system (\ref{cs1})--(\ref{cs4}) for its $E_{6(6)}$ components.
\section{Solution of the quadratic constraints}
\label{app:solve}
Our strategy for solving the quadratic constraints
for the embedding tensor follows the analysis of~\cite{LeDiffon:2008sh}
for the pure trombone gaugings. We make use of the fact that under
breaking to $E_{6(6)}$ the tensor $\vartheta_M$ contains a singlet
which (if invertible) allows to explicitly solve all the quadratic equations.
Decomposing under $E_{6(6)}$ according to (\ref{rep_break}),
we label the components $\vartheta_M$ and $\Theta_M{}^\alpha$ of the embedding tensor
according to
\begin{eqnarray}
\vartheta_M&\rightarrow~&
(\vartheta_\bullet,\;
\vartheta_m,\;
\vartheta^m,\;
\vartheta^\bullet)
\;.
\label{thbreak}
\end{eqnarray}
and
\begin{eqnarray}
\Theta_M{}^\alpha &=&
{\scriptsize
\left(
\begin{array}{cccc}
0& \xi_+^{a}& \xi_n & 0\\
-\ft13\xi_m &
\ft32t^a{}_{m}{}^n\xi_n +3t^a{}_{p}{}^qd_{rqm}\xi^{pr} &
\ft12d_{mnp}\xi^p+\xi_{mn}&
-t_a{}_{m}{}^n \xi_+^{a}\\
-\ft13\xi^m &
\ft32t^a{}_{n}{}^m\xi^n -30t^a{}_{p}{}^qd^{rpm}\xi_{qr} &
-t_a{}_{n}{}^m \xi_-^{a}&
-5d^{mnp}\xi_p+\xi^{mn}
\\
0&\xi^{a}_-&0&\xi^n
\end{array}
\right)}
\;,
\nonumber\\
\label{ThetaBreak}
\end{eqnarray}
respectively, in terms of various $E_{6(6)}$ tensors.
The relative coefficients among the various terms within $\Theta_M{}^\alpha$ are
determined by the fact that $\Theta_M{}^\alpha$ is constrained to live in the ${\bf 912}$
representation of $E_{7(7)}$, i.e.\ satisfies the relations (\ref{linear}).
\subsection{Determining the components of the embedding tensor}
To solve the quadratic constraints in a systematic way, we start from
the equation with the highest $SO(1,1)$ grading.
From (\ref{rep_break}) is follows that this is a ${\bf 27}^{+4}$ representation
inside ${\bf 1539}$, which thus corresponds to evaluating equations (\ref{Q2}) for
$MN={}_m{}_\bullet$. Explicitly, this leads to
\begin{eqnarray}
0&\equiv& \xi_m\vartheta_\bullet +t_{am}{}^n \xi_+^a\vartheta_n \;.
\label{eqQQ1}
\end{eqnarray}
Without loss of generality we may assume $\vartheta_\bullet$ to be non-vanishing
(which can always be achieved by change of basis in case $\vartheta_M$ is not
identically zero)
and from (\ref{eqQQ1}) express $\xi_m$
(one of the ${\bf 27}^{+1}$ components of the embedding tensor)
in terms of the unconstrained parameters
$(\vartheta_\bullet, \vartheta_m, \xi_+^a)$
transforming in the ${\bf 1}^{+3}+{\bf 27}^{+1}+{\bf 78}^{+3}$\,.
For convenience, we parametrize the latter as
\begin{eqnarray}
\vartheta_\bullet \equiv \kappa\;,\qquad
\vartheta_m \equiv \kappa \lambda_m\;,
\qquad \xi^a_+ \equiv \Xi^a
\;,
\label{solL1}
\end{eqnarray}
and solve equation (\ref{eqQQ1}) as
\begin{eqnarray}
\xi_m &=& - \Xi^a t_{am}{}^n \lambda_n
~\equiv~ - \delta_{\Xi} \lambda_m
\;.
\label{solQQ1}
\end{eqnarray}
By similar computations, the remaining parts of the embedding tensor
can be determined from other components of the constraint equations.
Evaluating equation (\ref{Q3}) for $\alpha\beta={}^m{}^n$
(the $\bar{\bf 351}^{+4}$ equation) yields
\begin{eqnarray}
0&\equiv&
\vartheta_\bullet\,\xi^{mn}-3\,\Xi^a\,t_{ak}{}^{[m}\,\xi^{n]k}-{10}\,\xi_+^a\,t_{ap}{}^{[m}\,d^{n]pq}\,(\vartheta_q+\ft32\xi_q)\,,
\end{eqnarray}
which upon plugging in (\ref{solQQ1}) reduces to
\begin{eqnarray}
{\cal O}_{2/3}\cdot\xi^{mn}
&\equiv&
{10}\,\Xi^a\,t_{ak}{}^{[m}\,d^{n]kl}\: {\cal O}_{2/3}\cdot \lambda_l\,,
\label{eqQQ2}
\end{eqnarray}
where we have defined the operator
\begin{eqnarray}
{\cal O}_{2/3} &\equiv& \delta_{\Xi} -\ft23 \kappa
\;.
\end{eqnarray}
As the same operator appears on both sides of equation (\ref{eqQQ2}),
the general solution for the component $\xi^{mn}$ can be given in polynomial form
as
\begin{eqnarray}
\xi^{mn} &=& {10}\, \Xi^a t_{ak}{}^{[m}\,d^{n]kl}\,\lambda_l
+\zeta^{mn}
\;,
\label{solQQ2}
\end{eqnarray}
where $\zeta^{mn}$ denotes a (real) zero mode of the operator ${\cal O}_{2/3}$.
It corresponds to an eigenvector of the action of the $\mathfrak{e}_{6(6)}$ generator
defined by $\Xi^a$ with the particular real eigenvalue $\frac23\kappa$\,.
In particular, such zero-modes only exist for non-compact choice of $\Xi^a$ and
some very particular values of $\kappa$\,.
Going down the grading, the next constraint equations live in the
$\bar{\bf 27}^{+2}$, of which there are four different ones. The
relevant ones are obtained from (\ref{Q1}) for $\alpha={}^m$ and
from (\ref{Q2}) for $MN={}^m{}_\bullet$, respectively,
leading to
\begin{eqnarray}
(\xi^m+\ft83\vartheta^m)\,\vartheta_\bullet
+\Xi^a\,t_{an}{}^m\,\vartheta^n &=& \xi^{mn}\vartheta_n+
5\,d^{mpq}\vartheta_p\, (\xi_q+\ft83 \vartheta_q) \;,
\label{27a}
\nonumber\\[1ex]
\xi^m\vartheta_\bullet-\Xi^a\,t_{an}{}^m\,\vartheta^n &=&
\xi^{mn}\vartheta_n- 15\,d^{mpq}\vartheta_p\,\xi_q \;.
\label{27b}
\end{eqnarray}
Using the explicit form of (\ref{solQQ1}), (\ref{solQQ2}), these two
equations determine the components $\xi^m$ and $\vartheta^m$
in terms of the free parameters according to
\begin{eqnarray}
\vartheta^m &=& 5\kappa\,d^{mkl}\,\lambda_k \lambda_l + \kappa \zeta^m
\;,
\nonumber\\[1ex]
\xi^m &=& 5\,
\Xi^a t_{an}{}^{m}\,d^{nkl}\,
\lambda_k \lambda_l
+ \zeta^{mn} \lambda_n - \ft43 \kappa \zeta^m
\;,
\label{solQQ3}
\end{eqnarray}
up to the constant vector $\zeta^m$ which is a zero mode of the operator
${\cal O}_{4/3} \equiv \delta_{\Xi}-\ft43 \kappa$\,.
Again, such zero modes exist only for very particular values of $\kappa$\,.
Next, we evaluate the constraint (\ref{Q2}) for $MN={}_m{}_n$
(the equation transforming in the ${\bf 351}^{+2}$) to obtain
\begin{eqnarray}
0&\equiv&
\ft12\xi_{[m}\vartheta_{n]}-5\,d_{mkp}d^{prq}d_{qln}\,\xi^{kl}\vartheta_r
+t_{a[m}{}^kd_{n]pk}\,\Xi^a \vartheta^p-\xi_{mn}\vartheta_\bullet
\;,
\end{eqnarray}
which uniquely determines the ${\bf 351}^{-1}$ component $\xi_{mn}$
of the embedding tensor.
Explicitly, after some computation and using the relations obtained above, we find
\begin{eqnarray}
\xi_{mn} &=& 2\,\Xi^a t_{a[m}{}^k\,\lambda_{n]} \lambda_k
-5\, \Xi^a t_{a[m}{}^p d_{n]pq}d^{qkl}\,\lambda_k\lambda_l
\nonumber\\
&&{}
-5\,d_{mkp}d^{prq}d_{qln}\,\zeta^{kl}\lambda_r
+t_{a[m}{}^kd_{n]pk}\,\Xi^a\zeta^p
\;.
\label{solQQ4}
\end{eqnarray}
The singlet equation ${\bf 1}^0$ from evaluating
(\ref{Q1}) for $\alpha={\rm o}$ yields
\begin{eqnarray}
\ft43 \vartheta_\bullet \vartheta^\bullet &=&
\ft13\,(\xi^m\vartheta_m-\xi_m\vartheta^m) -\ft49\, \vartheta^m\vartheta_m
\;,
\end{eqnarray}
which allows to express the singlet ${\bf 1}^{-3}$ component $\vartheta^\bullet$ in terms of the
other fields
\begin{eqnarray}
\vartheta^\bullet &=&
-\ft{5}{3}\,\kappa \, d^{k l m} \, \lambda_k \lambda_l \lambda_m
- \kappa \zeta^m \lambda_m
\;.
\label{solL2}
\end{eqnarray}
Evaluating (\ref{Q1}) for $\alpha=a$ finally yields
\begin{eqnarray}
0&=&
\ft32 t^a{}_{m}{}^n (\xi^m\vartheta_n-\xi_n\vartheta^m)
-3 t^a{}_{p}{}^q
(d_{rqn}\xi^{pr}\vartheta^n+10 d^{rpn}\xi_{qr}\vartheta_n)
\nonumber\\[.5ex]
&&{}
+\xi_-^a\vartheta_\bullet-\Xi^a\vartheta^\bullet
+16t^a{}_{m}{}^n\vartheta^m\vartheta_n
\;.
\end{eqnarray}
which yields
the ${\bf 78}^{-3}$ component $\xi_-^a$
\begin{eqnarray}
\xi_-^a &=& 30\,\xi^b_+ (t_b t^a)_n{}^m\,d^{nkl} \,\lambda_{m}\lambda_{k}\lambda_{l}
-\ft53 \,\Xi^a\,d^{klm}\,\lambda_{m}\lambda_{k}\lambda_{l}
\nonumber\\
&&{}
+ \Xi^a \lambda_k \zeta^k
- 6 \Xi^b (t^a t_b)_k{}^l \lambda_l \zeta^k
-8 \kappa\, t^a{}_k{}^l \lambda_l \zeta^k
\nonumber\\
&&{}
-6 t^a{}_m{}^l \lambda_k\lambda_l \zeta^{km}
+15 t^a{}_n{}^q d^{klp}d_{mpq} \lambda_k \lambda_l \zeta^{mn}
\;.
\end{eqnarray}
We have thus determined all the components of the embedding tensor (\ref{thbreak}), (\ref{ThetaBreak}) in terms of the parameters $\kappa$, $\lambda_m$, $\Xi^a$, $\zeta^m$, $\zeta^{mn}$, of which the latter two are particular eigenvectors under the $E_{6(6)}$ action of $\Xi$\,.
In the following, we need to check that this solution indeed satisfies all the
constraint equations~(\ref{Q1})--(\ref{Q3}).
In particular, the remaining constraint equations may impose further restriction on these tensors.
E.g.\ evaluating equation~\ref{Q1}) for $\alpha={}_m$ implies that
\begin{eqnarray}
d_{mkl} \zeta^k \zeta^l &=& 0
\;,
\label{zetzet}
\end{eqnarray}
i.e.\ $\zeta^k$ is not only zero mode of ${\cal O}_{4/3}$ but also
satifies a `pure-spinor type' condition of $E_{6(6)}$\,.
In the following, we evaluate all the remaining constraint equations.
\subsection{Evaluating the remaining equations}\label{etre}
We can now check all the remaining equations upon using the solution obtained above.
To this end, let us first calculate the invariant $I_4(\vartheta)$ from (\ref{I4}) quartic in the
trombone parameters $\vartheta_M$ for the explicit solution (\ref{solL1}), (\ref{solQQ3}), (\ref{solL2}).
As a result, we obtain
\begin{eqnarray}
I_4(\vartheta)&=&\frac{2}{3}
\,\kappa^4\,d_{mnp}\,\zeta^m\zeta^n\zeta^p\,,
\end{eqnarray}
i.e.\ the quartic invariant does not depend on the parameters $\lambda_m$\,.
This shows that all $\lambda_m$ can be set to zero by an $E_{7(7)}$ transformation
and therefore do not induce inequivalent gaugings. For simplicity, we will thus in the following
set $\lambda_m=0$\,.
The solution found in the previous section then reduces to
the solution (\ref{solutionemb}) given in the main text. In this section
we will evaluate all remaining constraint equations for this solution.
The calculation is rather tedious and has been performed using mathematica and
cadabra~\cite{Peeters:2006kp,Peeters:2007wn}.
As a result, we find that all remaining constraint equations of the system (\ref{Q1})--(\ref{Q3})
are satisfied provided, the parameters $\zeta^k$, $\zeta^{mn}$
obey the following set of identities
\begin{eqnarray}
\zeta^k\zeta^l d_{mkl} &=& 0
\;,\label{ps1}\\
\zeta^{k} \zeta^{mn} d_{kml} &=& 0\;,
\label{ps2}\\
\zeta^{[k} \zeta^{mn]} &=& 0 \;,
\label{ps3}\\
\left(t_a \cdot (\Xi+\ft43\kappa I) \cdot (\Xi-\ft23\kappa I)\right){}\!_n{}^m \,\zeta^n &=&
-\ft12 \zeta^{mk} \zeta^{ln} d_{klp} (t_a)_n{}^p
\;,
\label{remaining_constraints}
\end{eqnarray}
with the matrix $\Xi$ given by $\Xi_m{}^n \equiv \Xi^a (t_a)_m{}^n$ and `$\cdot$'
denoting the matrix product.
The third equation
comes from the constraint (\ref{Q3}) evaluated for $[\alpha\beta]=[ab]$;
the fourth equation comes from
the same constraint evaluated for $[\alpha\beta]=\, ^{am}$.
We have explicitly verified that {\em all other constraints}
are satisfied as a consequence of the ansatz
and the relations (\ref{remaining_constraints}).
In particular, the constraint obtained from (\ref{Q3}) evaluated for $[\alpha\beta]=\, ^{a}{}_{m}$
follows after some computation from
the fourth equation of (\ref{remaining_constraints}) and the other constraints.
The fourth equation of (\ref{remaining_constraints}) is linear in $\zeta^n$.
In particular, contracting this equation with $(t^a)_m{}^q$ implies
\begin{eqnarray}
(32\,\kappa^2- \Xi^a\Xi_a)\,\zeta^m &=& \ft{15}{2}\, \zeta^{kl}\zeta^{rs} d_{krp}d_{lsq} d^{pqm}
\;,
\label{contract4}
\end{eqnarray}
i.e.\ for $\Xi^a\Xi_a\not=32\,\kappa^2$, we can express $\zeta^m$ as a bilinear in $\zeta^{kl}$\,.
Note that this is consistent, as the r.h.s.\ of (\ref{contract4}) is indeed eigenvector of $\Xi$ associated to
an eigenvalue which is twice the one of $\zeta^{kl}$\,.
Still, plugging this expression for $\xi^m$ into the four equations (\ref{remaining_constraints})
will lead to a set of nontrivial constraints polynomial in $\zeta^{mn}$\,.
\par
As illustrated in section \ref{awoe}, the ``$ {\rm E}_{6(6)}$-pure
spinor'' constraint (\ref{ps1}) on $\zeta^m$ singles out an ${\rm
O}(1,1)\times{\rm SO}(5,5)$ subgroup of ${\rm E}_{6(6)} $ in which
${\rm SO}(5,5)$ is part of the little group of its solution. In
particular $\zeta^m$ coincides with the singlet ${\bf 1}^{-4}$ in
the branching (\ref{o11o55branch}) of the $\overline{{\bf
27}}$ relative to this ${\rm O}(1,1)\times{\rm SO}(5,5)$, and $\Xi$ should have the form in Eq. (\ref{xisemisimple}).
Let us first consider the simple case in which $\Xi_0$ is a
semisimple element of $\mathfrak{so}(5,5)$ and thus can be
considered as an element of its Cartan subalgebra. Let us also
consider the case in which the two sides of Eq. (\ref{cs4}) are
separately zero. The constraint (\ref{remaining_constraints}) is
then satisfied if $\Xi_0$ commutes with an $\mathfrak{so}(4,4)$
subalgebra of $\mathfrak{so}(5,5)$ and if its norm is ${\rm
Tr}(\Xi_0\cdot \Xi_0)=24\,\kappa^2$. The last requirement is easily
understood by observing that $\Xi_0$ and $\Xi_1$ are mutually
orthogonal and that ${\rm Tr}(\Xi_1\cdot \Xi_1)=8\,\kappa^2$, since
(\ref{remaining_constraints}) implies that ${\rm Tr}(\Xi\cdot
\Xi)=32\,\kappa^2$. This fixes the normalization of $\Xi_0$. We can
branch the relevant ${\rm E}_{6(6)}$ representations with respect to
its ${\rm SO}(1,1)^2\times {\rm SO}(4,4)$ subgroup, where the ${\rm
SO}(1,1)^2$ factor is generated by $\Xi_1+\Xi_0$. The
$\overline{{\bf 27}}$ then branches as follows:
\begin{eqnarray}
\overline{{\bf 27}}&\rightarrow & {\bf 1}^{(\frac{4}{3},0)}+{\bf 1}^{(-\frac{2}{3},2)}+{\bf 1}^{(-\frac{2}{3},-2)}+{\bf 8}_v^{(-\frac{2}{3},0)}+{\bf 8}_s^{(\frac{1}{3},1)}+{\bf 8}_c^{(\frac{1}{3},-1)}\,,\nonumber\\
{{\bf 78}}&\rightarrow & ({\bf 28}+{\bf 1}+{\bf 1})^{(0,0)}+{\bf 8}_s^{(1,-1)}+{\bf 8}_c^{(1,1)}+{\bf 8}_s^{(-1,1)}+{\bf 8}_c^{(-1,-1)}+
{\bf 8}_v^{(0,2)}+{\bf 8}_v^{(0,-2)}\,,\nonumber
\end{eqnarray}
where the gradings are the eigenvalues of $\Xi_1/\kappa,\,\Xi_0/\kappa$. The ``pure spinor'' $\zeta^m$ corresponds to the ${\bf 1}^{(\frac{4}{3},0)}$ representation. Consider now the vector $\zeta\cdot t^a \equiv(\zeta^m\,t^a{}_m{}^n)$. The constraint (\ref{remaining_constraints}) reads:
\begin{eqnarray}
\left(\delta_\Xi+\frac{2}{3}\,\kappa\right)\left(\delta_\Xi-\frac{4}{3}\,\kappa\right)\zeta\cdot t^a =0\,.
\end{eqnarray}
Let us analyze the relevant cases:
\begin{itemize}\item{If $t^a\in \mathfrak{so}(5,5)+\mathfrak{so}(1,1)$, $\zeta\cdot t^a$ is still in the ${\bf 1}^{(\frac{4}{3},0)}$ and is thus annihilated by $\delta_\Xi-\frac{4}{3}\,\kappa$;}
\item{If $t^a\in {\bf 8}_s^{(1,-1)}+{\bf 8}_c^{(1,1)}$, $\zeta\cdot t^a={\bf 0}$ and the constraint is satisfied;}
\item{If $t^a\in {\bf 8}_s^{(-1,1)}+{\bf 8}_c^{(-1,-1)}$, $\zeta\cdot t^a\in {\bf 8}_s^{(\frac{1}{3},1)}+{\bf 8}_c^{(\frac{1}{3},-1)}$. The component in ${\bf 8}_s^{(\frac{1}{3},1)}$ is a zero-mode of $\delta_\Xi-\frac{4}{3}\,\kappa$, while the second component is a zero mode of $\delta_\Xi+\frac{2}{3}\,\kappa$ and the constraint is still satisfied;}
\end{itemize}
Consider now the case in which $\Xi_0$ has a nilpotent component $\Xi_{\rm n}$ in ${\bf 16}_s^{-3}$, so that $\Xi$ has a semisimple and a nilpotent component $\Xi=\Xi_{\rm ss}+\Xi_{\rm n}$. The constraint (\ref{remaining_constraints}) is still satisfied provided the following condition holds: $[\Xi_{\rm ss},\Xi_{\rm n}]=2\,\kappa\,\Xi_{\rm n}$.
\section{Gauge group generators in $E_6$ components}
Here, we give the gauge group generators $(X_M)_N{}^K$
for the general solution (\ref{solutionemb}) of the quadratic constraints as obtained from
(\ref{defX}), (\ref{thbreak}), (\ref{ThetaBreak}), and (\ref{solutionemb}).
They are parametrized by $\kappa$, $\Xi^a$, $\zeta^m$, and $\zeta^{mn}$,
which are subject to the identities (\ref{remaining_constraints}).
\begin{eqnarray}
X_\bullet &=&
{\scriptsize
\left(
\begin{array}{cccc}
0&0&0&0\\
0&\Xi_m{}^n-\ft23 \kappa \delta_m^n&0&0
\\
0&0& -\Xi_n{}^m-\ft43 \kappa \delta_n^m&0
\\
0&0&0& -2\kappa
\end{array}
\right)}
\;,
\nonumber\\[1ex]
X_k &=&
{\scriptsize
\left(
\begin{array}{cccc}
0&-\Xi_k{}^n+\ft23 \kappa \delta_k^n&0&0\\
\Xi_{[k}{}^p \zeta^q d_{m]pq}
&\ft12
\zeta^{np}d_{kpm}+5\zeta^{pq}d_{kpr}d_{qsm}d^{rsn}
&\ft23\kappa d_{kmn}-\Xi_{k}{}^pd_{pmn}
&0
\\
0&10\Xi_{[k}{}^p\zeta^q d_{r]pq}d^{rmn}&
\hspace*{-5mm}
5\zeta^{pq}d_{npr}d_{kqs}d^{rsm}-\ft12\zeta^{mp}d_{kpn}
&
\Xi_k{}^m-\ft23 \kappa \delta_k^m
\\
0&0&
-\Xi_{[k}{}^p \zeta^q d_{n]pq}
& 0
\end{array}
\right)}
\;,
\nonumber\\[1ex]
X^k &=&
{\scriptsize
\left(
\begin{array}{cccc}
0&\zeta^{kn}&0&0
\\
0& X^k{}_m{}^n & \zeta^{kp} d_{pmn} &0
\\
0&0& X^{k\,m}{}_n & \zeta^{mk}
\\
0&0&0&-2\kappa\,\zeta^m
\end{array}
\right)}
\;,
\nonumber\\[1ex]
X^\bullet &=& {\scriptsize \left(
\begin{array}{cccc}
0&-2\kappa\zeta^n&0&0\\
0&0&-2\kappa \zeta^p d_{pmn}&0
\\
0&0&0&2\kappa\zeta^m
\\
0&0&0&0
\end{array}
\right)} \;, \label{explicitgenerators} \end{eqnarray}
where
\begin{eqnarray}
X^k{}_m{}^n&=& 2\Xi_m{}^{[k}\zeta^{n]} -\ft43\kappa
\zeta^k\delta^n_m -\ft23 \kappa \zeta^n\delta^k_m+\ft{20}3 \kappa
\zeta^p d_{pqm}d^{qkn} -10 \Xi_p{}^k\zeta^q d_{mqr}d^{rpn}\;,
\nonumber\\
X^{k\,m}{}_n&=&2\Xi_n{}^{[m}\zeta^{k]}+\ft43\kappa
\delta_n^{[k}\zeta^{m]} -\ft{20}3 \kappa \zeta^p d_{pqn}d^{qkm} +10
\Xi_p{}^k\zeta^q d_{nqr}d^{rpm} \;. \nonumber
\end{eqnarray}
Via (\ref{XXX}) these generators also encode the structure constants of the
gauge algebra.
\mathversion{bold}\section{$T$-identities}
\mathversion{normal}
\label{app:Tids}
Upon dressing the quadratic constraints (\ref{Q1})--(\ref{Q3}) with the scalar vielbein
and using the definitions (\ref{TX}), (\ref{TAB}), one obtains a large number of $SU(8)$
identities bilinear in the tensors $A$ and $B$.
Here we collect those identities that are important for the calculations in the main text.
Working out all quadratic constraints
that transform in the ${\bf 63}$ of $SU(8)$, we find the following relations
\begin{eqnarray}
{\bf 1539}: \;\; 0 &=&
6 A^{ik}B_{jk}-6 B^{ik}A_{jk}+A_j{}^{imn}B_{mn}-A^i{}_{jmn}B^{mn}
\;,
\label{constraints63}
\\[1ex]
{\bf 133}:\;\; 0 &=&
2 A^{ik}B_{jk}+2 B^{ik}A_{jk}-A_j{}^{imn}B_{mn}-A^i{}_{jmn}B^{mn}
+\ft{32}{3} B^{ik}B_{jk}
~-~ \ft18 \delta^i_j \,{\rm trace}
\;,
\nonumber\\[1ex]
{\bf 133}: \;\; 0 &=&
12 A^{ik}A_{jk}-A^i{}_{mnk} A_j{}^{mnk} +3 A^k{}_{mnj}A_k{}^{mni}
\nonumber\\
&&{}
+ 12
\left(2 A^{ik}B_{jk}+2 B^{ik}A_{jk}-A_j{}^{imn}B_{mn}-A^i{}_{jmn}B^{mn}\right)
~-~ \ft18 \delta^i_j \,{\rm trace}
\;,
\nonumber\\[1ex]
{\bf 8645}: \;\; 0 &=&
-240 A^{ik}A_{jk}+11A^i{}_{mnk} A_j{}^{mnk} +21 A^k{}_{mnj}A_k{}^{mni}
\nonumber\\
&&{}
-12
\left(10 A^{ik}B_{jk}+10 B^{ik}A_{jk}+A_j{}^{imn}B_{mn}+A^i{}_{jmn}B^{mn}\right)
~-~ \ft18 \delta^i_j \,{\rm trace}
\;,\nonumber
\end{eqnarray}
descending from the various irreducible $E_{7(7)}$ contributions of (\ref{Q1})--(\ref{Q3})
as indicated. The $E_{7(7)}$ origin of these constraints can also be confirmed by calculating
the action of the quadratic Casimir operator upon using the $E_{7(7)}$ transformation properties
\begin{eqnarray}
\delta A^{ij} &=&
\ft13 A^{(i}{}_{klm} \Sigma^{j)klm}
\;,
\nonumber\\
\delta A_i{}^{jkl} &=&
2A_{im} \Sigma^{mjkl} + 3\Sigma^{mn[jk} A^{l]}{}_{imn}
+\Sigma^{mnp[j}\delta^k_i A^{l]}{}_{mnp}
\;,
\nonumber\\
\delta B^{ij} &=&
-\Sigma^{ijkl}B_{kl}
\;,
\label{actionE7}
\end{eqnarray}
with $\Sigma^{ijkl}$ satisfying
$\Sigma_{ijkl}=\frac1{24}\epsilon^{ijklmnpq}\Sigma_{mnpq}$\,.
Similarly, we can deduce
we can deduce two constraints in the
${\bf 70}$ of $SU(8)$:
\begin{eqnarray}
{\bf 133}: \;\; 0 &=&
A^m{}_{[jkl} B_{n]m}
+4B_{[jk}B_{ln]}
-\ft1{24} \epsilon_{jkln mqrs}
\left(A_p{}^{mqr} B^{sp}+4B^{mq}B^{rs} \right)
\nonumber\\[1ex]
{\bf 133}: \;\; 0 &=&
4 A^m{}_{[jkl} A_{n]m}
-3 A^m{}_{p[jk}A^p{}_{ln]m}
+16 A^m{}_{[jkl} B_{n]m}
\nonumber\\
&&{}
-\ft1{24} \epsilon_{jklnmpqr} \left(
4A_s{}^{mpq}A^{rs}
-3A_t{}^{ump} A_u{}^{qrt}
+16 A_u{}^{mpq} B^{ru}\right)\;,
\label{constraints70}
\end{eqnarray}
and two constraints in the
${\bf 378}$ of $SU(8)$:
\begin{eqnarray}
{\bf 1539}: \;\; 0 &=&
4A_{j[k}B_{ln]} + A^m{}_{j[kl} B_{n]m} + A^m{}_{kln} B_{jm}
\nonumber\\
&&{}
-\ft1{9} \epsilon_{rsmpqnkl} A_j{}^{mpq} B^{rs}
+\ft1{18} \epsilon_{mjqrsnkl} A_p{}^{qrs} B^{mp}
\;,
\nonumber\\[1ex]
{\bf 8645}: \;\; 0 &=&
-18 A^m{}_{nkl}A_{jm} - 54 A^m{}_{j[kl} A_{n]m}
+60A_{j[k}B_{ln]} -9 A^m{}_{j[kl} B_{n]m} -9 A^m{}_{kln} B_{jm}
\nonumber\\
&&{}
+ \epsilon_{mpqrsnkl} A_j{}^{urs}
(A_u{}^{mpq}
-\ft1{3} \delta_u^m B^{pq} )
-\ft34 \epsilon_{jnklpmrs} A_u{}^{rst}
(A_t{}^{mup} -\ft2{9} \delta^m_t B^{up})
\;.
\nonumber\\
\label{constraints378}
\end{eqnarray}
If the components $\Theta_M{}^\alpha$, $\vartheta_M$ are chosen such
as to satisfy the quadratic constraints (\ref{Q1})--(\ref{Q3}), the relations
(\ref{constraints63})--(\ref{constraints378}) among the scalar tensors $A$, $B$,
follow as an immediate consequence.
For $B^{ij}=0$, all these identities consistently reduce to the quadratic identities
given in~\cite{deWit:1982ig,deWit:2007mt}.
\end{appendix}
\providecommand{\href}[2]{#2}\begingroup\raggedright |
\section{Introduction}\label{section:introduction}
Studying the diameters of the graphs of polytopes and polyhedra has received a lot of attention~\cite{kalaiblog:polymath3part6} due to Santos' recent counter-example~\cite{Santos:CounterexampleHirsch} to the Hirsch Conjecture. The Hirsch Conjecture asserts that the diameter of any $d$-dimensional polytope with $n$ facets is never greater than $n-d$. Since this conjecture is now known to be false, the question of the Polynomial Hirsch Conjecture, which asserts that the diameter of any polytope with $n$ facets is polynomial in $n$, is relevant. (The dimension $d$ is not in the statement of the Polynomial Hirsch Conjecture since $n > d$.) The first step in this line of investigation is to settle the Linear Hirsch Conjecture, which asserts that the diameter is linear in the number $n$ of facets, independent of the dimension $d$.
The original Hirsch Conjecture was stated for the graphs of polyhedra. Since Klee and Walkup~\cite{Klee:d-step} showed that the Hirsch Conjecture is false for unbounded polyhedra, the Hirsch Conjecture (and the Linear and Polynomial Hirsch Conjectures, which are both still open) is usually stated for bounded polytopes. However, we should note that the Linear Hirsch or Polynomial Hirsch Conjectures would still be interesting for unbounded polyhedra.
These conjectures on the diameters of polytopes and polyhedra are interesting because of their relation to the efficiency of the simplex algorithm for linear programming. In particular, the diameter of the feasibility polyhedron is a lower bound on the number of pivot steps needed for the simplex algorithm. Thus, if the Polynomial Hirsch Conjecture is false, then no pivot rule for the simplex algorithm runs in polynomial time. For more information on the Hirsch Conjecture and its relationship to the behavior of the simplex method, see the survey~\cite{Klee:The-d-step-conjecture} or the recent survey~\cite{KimSantos:HirschSurvey} written jointly with Santos. Our terminology on polytopes follows the language in~\cite{Ziegler:Lectures}.
The study of diameters of convex polyhedra via combinatorial abstractions of polyhedral graphs was considered by Adler, Dantzig, Murty, and Saigal~\cite{Adler:AbstractPolytopesThesis, AdlerDantzigMurty:AbstractPolytopes, Adler:MaxDiamAbsPoly, AdlerSaigal:LongPathsAbstract, Murty:GraphAbstract}, Kalai~\cite{Kalai:DiameterHeight}, and Eisenbrand, H\"ahnle, Razborov, and Rothvo\ss{}~\cite{Eisenbrand:Diameter-of-Polyhedra, Hahnle:Diplomathesis}. Adler et al.{} introduced the first formally-defined combinatorial abstraction of the graphs of polytopes satisfying a collection of axioms. In~\cite{Kalai:DiameterHeight}, Kalai showed that a more general family of objects obtained by removing one of these axioms still satisfies the quasi-polynomial upper bound for the diameters of convex polyhedra proved in~\cite{Kalai:Quasi-polynomial}. A further generalization of polyhedral graphs was studied in~\cite{Eisenbrand:Diameter-of-Polyhedra} by dropping an additional axiom. Eisenbrand et al.{} prove that even for this more general class of objects, the quasi-polynomial diameter upper bound of Kalai and Kleitman in~\cite{Kalai:Quasi-polynomial} holds. On the other hand, they give a construction to prove that the diameters of some objects in this more general class are superlinear. One can say that this superlinear lower bound in~\cite{Eisenbrand:Diameter-of-Polyhedra} is evidence against the Linear Hirsch Conjecture.
In this paper, we introduce \emph{subset partition graphs} (defined in Section~\ref{section:subset-partition-graphs}), a new family of combinatorial abstractions of the graphs of polyhedra. The new abstraction is a flexible framework inspired by the combinatorial properties found in previous abstractions, and provides many variants for abstractions of polyhedra. Our main theoretical result is the construction of a family of subset partition graphs whose diameter is superpolynomial (see Theorem~\ref{theorem:general-lower}), which can be considered evidence against the Polynomial and Linear Hirsch Conjectures. We also prove a quadratic lower bound on the diameters of a special subclass of subset partition graphs, which provides further evidence against the Linear Hirsch Conjecture. Moreover, we present a strategy to disprove the Linear Hirsch Conjecture via combinatorial operations on subset partition graphs.
\vskip12pt
\noindent{\bf Outline of this paper:} In Section~\ref{section:previous-abstractions-main}, we formally define the previous abstractions and survey the known upper and lower bounds. Section~\ref{section:base-abstractions-clf} introduces the abstraction presented in~\cite{Eisenbrand:Diameter-of-Polyhedra} and Section~\ref{section:previous-abstractions} discusses combinatorial properties defining special cases. Motivated by this discussion, in Section~\ref{section:subset-partition-graphs} we define our new combinatorial abstraction, the subset partition graphs. In Section~\ref{section:bounds} we prove upper and lower bounds on the diameters of subset partition graphs that satisfy particular sets of properties. We give some final remarks in Section~\ref{section:final-remarks}, and present a strategy for disproving the Linear Hirsch Conjecture.
\section{Previous abstractions}\label{section:previous-abstractions-main}
Here we describe relevant previous combinatorial abstractions in the literature\footnote{Very recently, new abstractions unrelated to Section~\ref{section:subset-partition-graphs} introduced in~\cite{kalaiblog:polymath3part6} have formed the discussion of a web-based discussion: See \url{http://gilkalai.wordpress.com/2010/09/29/polymath-3-polynomial-hirsch-conjecture/}, \url{http://gilkalai.wordpress.com/2010/10/03/polymath-3-the-polynomial-hirsch-conjecture-2/}, \url{http://gilkalai.wordpress.com/2010/10/10/polymath3-polynomial-hirsch-conjecture-3/}, \url{http://gilkalai.wordpress.com/2010/10/21/polymath3-polynomial-hirsch-conjecture-4/}, \url{http://gilkalai.wordpress.com/2010/11/28/polynomial-hirsch-conjecture-5-abstractions-and-counterexamples/}, and \url{http://gilkalai.wordpress.com/2011/04/13/polymath3-phc6-the-polynomial-hirsch-conjecture-a-topological-approach/}.}. In all cases, the object of study is an abstract generalization of simple polyhedra. We say that a $d$-dimensional polyhedron $P$ is \emph{simple} if each of its vertices is contained in exactly $d$ of the $n$ facets of $P$. For the study of diameters of polyhedra, it is enough to consider simple polyhedra, since the largest diameter of a $d$-polyhedron with $n$ facets is found among the simple $d$-polyhedra with $n$ facets~\cite{KimSantos:HirschSurvey}.
Let $H(n,d)$ denote the maximum diameter of $d$-dimensional polytopes with $n$ facets, and let $H_u(n,d)$ denote the maximum diameter of $d$-dimensional polyhedra with $n$ facets. Since polytopes are polyhedra, we clearly have $H(n,d) \leq H_u(n,d)$. In~\cite{Kalai:Quasi-polynomial}, Kalai and Kleitman proved that $H_u(n,d) \leq n^{1+ \log_2 d}$.
\subsection{Base abstractions and connected layer families}\label{section:base-abstractions-clf}
We now describe an abstraction of Eisenbrand et al.{} introduced in~\cite{Eisenbrand:Diameter-of-Polyhedra}. Fix a finite set $S$ of cardinality $n$, called the \emph{symbol set}. (Each $s \in S$ is called a \emph{symbol}.) Let $\mathcal{A} \subseteq \binom{S}{d}$, where $\binom{S}{d}$ is the set of all $d$-element subsets of $S$. We consider connected graphs of the form $G = (\mathcal{A}, E)$ with vertex set $\mathcal{A}$ and edge set $E$ satisfying:
\begin{itemize}
\item for each $A, A' \in \mathcal{A}$, there is a path from $A$ to $A'$ in the graph $G$ using only vertices that contain $A \cap A'$.
\end{itemize}
If this occurs, we say that $G$ is a $d$-dimensional \emph{base abstraction} of $\mathcal{A}$ on the symbol set $S$. The \emph{diameter} of the base abstraction is the diameter of the graph $G$.
Note that the graphs of simple $d$-dimensional polyhedra with $n$ facets are base abstractions. Indeed, each of the $n$ facets of $P$ is associated with a symbol $s$ in $S$. Since our polyhedron $P$ is simple, each vertex of $P$ is incident to exactly $d$ facets, and so it is associated with the $d$-element subset of $S$ consisting of the corresponding symbols. The graph $G$ used in the base abstraction ``is'' the graph of the polyhedron. The defining condition is satisfied since, for every pair of vertices $y$ and $z$ on a polyhedron $P$, there is a path from $y$ to $z$ on the smallest face of $P$ containing both $y$ and $z$.
Since the graphs of polyhedra are base abstractions, we clearly have $H_u(n,d) \leq B(n,d)$, where $B(n,d)$ denotes the maximum diameter among $d$-dimensional base abstractions on a symbol set of size $n$. In~\cite{Eisenbrand:Diameter-of-Polyhedra}, Eisenbrand et al.{} prove that the Kalai-Kleitman bound of $n^{1+\log_2 d}$ is also an upper bound for $B(n,d)$. They also show that $B(n,\frac{n}{4})$ is in $\Omega(n^2 / \log n)$, i.{}e.{}, the diameters of base abstractions obey a quadratic lower bound (up to a logarithmic factor).
Their upper and lower bounds for $B(n,d)$ were proved by analyzing the diameters of a related combinatorial object. A $d$-dimensional \emph{connected layer family} of $\mathcal{A} \subseteq \binom{S}{d}$ on a set of $n = |S|$ symbols is a family $\mathcal{V} = \{ \mathcal{V}_0, \ldots, \mathcal{V}_t \}$ of non-empty sets such that:
\begin{itemize}
\item {\bf partition property:} $\mathcal{A} = \mathcal{V}_0 \cup \cdots \cup \mathcal{V}_t$,
\item {\bf disjointness property:} $\mathcal{V}_i \cap \mathcal{V}_j = \emptyset$ if $i \not= j$,
\item {\bf connectivity property:} for all $i < j < k$ and $A \in \mathcal{V}_i$, $A' \in \mathcal{V}_k$, there is an $A'' \in \mathcal{V}_j$ such that $A \cap A' \subseteq A''$.
\end{itemize}
Each individual $\mathcal{V}_i$ is called a \emph{layer}. The \emph{diameter} of the connected layer family $\mathcal{V} = \{ \mathcal{V}_0, \ldots, \mathcal{V}_t \}$ is $t$. Recall that $B(n,d)$ denotes the maximal diameter of a $d$-dimensional base abstraction on a symbol set of size $n$. We use $C(n,d)$ to denote the maximal diameter of a $d$-dimensional connected layer family on a symbol set of size $n$.
In~\cite{Eisenbrand:Diameter-of-Polyhedra}, Eisenbrand et al.{} prove $B(n,d) = C(n,d)$. The proof that $B(n,d) \geq C(n,d)$ follows from the fact that a base abstraction is obtained from a connected layer family by connecting $d$-sets $A \in \mathcal{V}_i$ and $A' \in \mathcal{V}_j$ with an edge if $|i-j| \leq 1$. The proof of $B(n,d) \leq C(n,d)$ follows from the fact that a connected layer family is obtained from a base abstraction by the following layering process: let $G = (\mathcal{A}, E)$ be a $d$-dimensional base abstraction, and fix a particular $d$-subset $Z \in \mathcal{A}$. Then, let $\mathcal{V}_i := \{ A \in \mathcal{A} : \operatorname{dist}_G(A,Z) = i\}$. Note that $\mathcal{V} = \{ \mathcal{V}_0, \ldots, \mathcal{V}_t \}$ is a $d$-dimensional connected layer family with diameter $t$ since the face path property immediately implies that the collection $\mathcal{V}$ satisfies the connectivity property. See Lemma~3.4.2 in~\cite{Hahnle:Diplomathesis} for a detailed proof. The bound of $B(n,d) \leq C(n,d)$ is obtained by choosing a base abstraction $G=(\mathcal{A},E)$ whose diameter is $B(n,d)$ and picking a pair $(Z,Z')$ of $d$-sets in $\mathcal{A}$ at distance $B(n,d)$.
\subsection{Previous abstractions satisfying additional properties}\label{section:previous-abstractions}
Our new abstraction defined in Section~\ref{section:subset-partition-graphs} is motivated by following the same layering process just described, but starting with special cases of base abstractions that were studied in~\cite{AdlerDantzigMurty:AbstractPolytopes} and~\cite{Kalai:DiameterHeight} which satisfied additional combinatorial properties. Let $G = (\mathcal{A}, E)$ be a $d$-dimensional base abstraction. If the condition ``$(A,A')$ is an edge in $E$ if and only if $|A \cap A'| = d-1$'' holds, then we say that $G$ is an \emph{ultraconnected set system}. This condition is called \emph{ultraconnectedness}. Note that ultraconnectedness holds for the graphs of polyhedra: if two vertices $y$ and $z$ of a simple $d$-polyhedron $P$ share all but one facet in common, then they are neighbors in the graph of $P$. In~\cite{Kalai:DiameterHeight}, Kalai proved that ultraconnected set systems satisfy the diameter bound of~\cite{Kalai:Quasi-polynomial}.
If the layering process described earlier is applied to a base abstraction satisfying the ultraconnectedness property, then the resulting collection $\mathcal{V} = \{ \mathcal{V}_0, \ldots, \mathcal{V}_t \}$ of layers satisfies the following adjacency property: if $A, A' \in \mathcal{A}$ and $|A \cap A'| = d-1$, then $A$ and $A'$ are in the same or adjacent layers, i.e., if $A \in \mathcal{V}_i$ and $A' \in \mathcal{V}_j$ then $|i-j| \in \{0,1\}$.
Adler, Dantzig, and Murty considered an abstraction where, in addition to ultraconnectedness, the following \emph{polytopal endpoint-count condition} must hold: if $F \in \binom{S}{d-1}$, then $\{A \in \mathcal{A} : F \subset A\}$ has cardinality either $0$ or $2$. An ultraconnected set system satisfying this additional condition is called an \emph{abstract polytope}. For polytopes, the polytopal endpoint-count condition translates into the fact that every $1$-face of a polytope is incident to exactly two $0$-faces. This condition fails for polyhedra because of $1$-faces containing only one vertex, so we consider a \emph{polyhedral endpoint-count condition}, where we allow $|\{A \in \mathcal{A} : F \subset A\}|$ to be $1$ as well. In~\cite{Adler:MaxDiamAbsPoly}, Adler and Dantzig prove that $d$-dimensional abstract polytopes with $n$ symbols satisfy the Hirsch bound if $n-d \leq 5$, the abstract analogue of~\cite{Klee:d-step}.
\section{Subset partition graphs}\label{section:subset-partition-graphs}
Now that we have seen how properties defining special classes of base abstractions imply certain structural properties on connected layer families obtained by the layering process, we are ready to define subset partition graphs, which are a generalization of base abstractions and connected layer families. As before, we have a set $S$ called the symbol set, with each $s \in S$ being a symbol.
\begin{definition}
Fix a finite set $S$ of cardinality $n$ and a set $\mathcal{A} \subseteq \binom{S}{d}$ of subsets. Let $G = (\mathcal{V},E)$ be a connected graph with vertex set $\mathcal{V} = \{\mathcal{V}_0,\ldots,\mathcal{V}_t\}$. If $\mathcal{V}$ is a partition of $\mathcal{A}$ in the sense that:
\begin{enumerate}
\item $\mathcal{A} = \mathcal{V}_0 \cup \cdots \cup \mathcal{V}_t$,
\item $\mathcal{V}_i \cap \mathcal{V}_j = \emptyset$ if $i \not= j$, and
\item $\mathcal{V}_i \not= \emptyset$ for all $i$,
\end{enumerate}
then we say that $G$ is a $d$-dimensional \emph{subset partition graph} of $\mathcal{A}$ on the symbol set $S$.
\end{definition}
Note that $d$-dimensional subset partition graphs on a symbol set $S$ of size $n$ are combinatorial abstractions of simple $d$-dimensional polyhedra with $n$ facets: each of the $n$ facets of a $d$-dimensional polyhedron $P$ corresponds to a symbol $s \in S$, and a vertex of $P$ corresponds to a $d$-set $A \in \mathcal{A}$ given by the incident facets.
As defined, the only condition on the edge set $E$ is that the graph $G$ is connected, and thus subset partition graphs do not yet give an interesting combinatorial abstraction of the graphs of polytopes and polyhedra. For this, one should require one or more of the combinatorial properties identified below, which are conditions on the set $\mathcal{A}$ of subsets or on the edge set $E$ of the graph $G$. Before identifying the properties, we need to define the following operation on subset partition graphs:
\begin{definition}[Restriction]
Let $G = (\mathcal{V},E)$ be a subset partition graph of $\mathcal{A}$ on the symbol set $S$, and let $F \subseteq S$ be a collection of symbols. We define a new subset partition graph $G_F = (\mathcal{V}_F, E_F)$ of $\mathcal{A}_F$ on the symbol set $S' := S$.
We define $\mathcal{A}_F := \{A \in \mathcal{A} : F \subseteq A\}$. That is to say, $\mathcal{A}_F$ is obtained by deleting from $\mathcal{A}$ (and the containing $\mathcal{V}_i$) any $d$-set $A$ which does not contain $F$. This deletion from the vertices in $\mathcal{V}$ may make some of them empty. The vertex set $\mathcal{V}_F$ consists of those vertices in $\mathcal{V}$ which are still non-empty, and two vertices in $\mathcal{V}_F$ are connected by an edge in $E_F$ exactly when the associated vertices were connected in $E$. The subset partition graph $G_F$ is called the \emph{restriction} of $G$ with respect to $F$.
\end{definition}
Based on the discussion in Sections~\ref{section:base-abstractions-clf} and~\ref{section:previous-abstractions}, and together with the definition of restriction, we identify the main properties that should be considered for subset partition graphs:
\begin{itemize}
\item {\bf dimension reduction:} if $F \subseteq S$ such that $|F| \leq d$, then (the underlying graph of) the restriction $G_F$ is a connected graph.
\item {\bf adjacency:} if $A, A' \in \mathcal{A}$ and $|A \cap A'| = d-1$, then $A$ and $A'$ are in the same or adjacent vertices of $G$.
\item {\bf strong adjacency:} adjacency holds and, if two vertices $\mathcal{V}_i$ and $\mathcal{V}_j$ are adjacent in $G$ then there are $d$-sets $A \in \mathcal{V}$ and $A' \in \mathcal{V}_j$ such that $|A \cap A'| = d-1$.
\item {\bf endpoint-count:} if $F \in \binom{S}{d-1}$, then $|\{A \in \mathcal{A} : F \subset A\}| \leq 2$.
\end{itemize}
Note that the graphs of polytopes satisfy dimension reduction since the graph of a face is connected. The notion of strong adjacency was proposed by H\"{a}hnle in~\cite{Hahnle:SPGs}.
\begin{example}\label{example:non-polytopal-spg}
Figure~\ref{figure:example-spg} illustrates a $3$-dimensional subset partition graph on a symbol set $S$ with $n=|S|=6$ which satisfies the dimension reduction, strong adjacency, and endpoint-count properties. The graph of $G$ has six vertices and $|\mathcal{A}| = 2 \times 2 + 4 \times 1$.
\begin{figure}[hbt]
\[
\xy
(0,0)*{\{123,126\}};
(33,0)*{\{246\}};
(0,-33)*{\{156\}};
(33,-33)*{\{345,456\}};
(22,-11)*{\{234\}};
(11,-22)*{\{135\}};
(9,0); (27,0) **\dir{-};
(6,-33); (24,-33) **\dir{-};
(0,-3); (0,-28) **\dir{-};
(33,-3); (33,-28) **\dir{-};
(30,-3); (25,-8) **\dir{-};
(3,-29.3); (8,-25) **\dir{-};
(3,-3); (9, -19) **\dir{-};
(6,-3); (19, -8) **\dir{-};
(27,-29); (12, -25) **\dir{-};
(31,-29); (23, -14) **\dir{-};
\endxy
\]
\caption{A subset partition graph $G$ with $d=3$ and $n=6$.}\label{figure:example-spg}
\end{figure}
\end{example}
We define these (and the following) properties for subset partition graphs because we want a flexible framework where we consider certain collections of properties at a time. By considering certain properties ``on'' and other properties ``off'' we hope to understand which properties are crucial in proving diameter bounds.
For instance, the class of subset partition graphs with the dimension reduction property together with the additional condition that the underlying graph of $G$ is a path is exactly the class of connected layer families. Subset partition graphs interpolate between connected layer families and polyhedral graphs by the selection of properties.
(Note that the properties presented are not necessarily independent. For instance, endpoint-count together with dimension reduction implies adjacency. However, we explicitly list the adjacency property as we will consider subset partition graphs satisfying adjacency but not satisfying dimension reduction.)
In addition to the main properties described above, there are other combinatorial properties of polytopes which translate into natural properties to consider for subset partition graphs, for example:
\begin{itemize}
\item {\bf $d$-connectedness:} the graph $G$ is $d$-connected.
\item {\bf $d$-regularity:} the graph $G$ is $d$-regular.
\item {\bf $d$-neighbors:} for every $A \in \mathcal{A}$, $|\{A' \in \mathcal{A} \setminus \{A\} : |A \cap A'| = d-1\} | = d$.
\item {\bf one-subset:} $|\mathcal{V}_i| = 1$ for each $i = 0, \ldots, t$.
\end{itemize}
The $d$-connectedness property for subset partition graphs is desirable due to Balinski's Theorem, which says that the graph of a $d$-dimensional polytope is $d$-connected~\cite{Balinski:On-the-graph-structure}. The $d$-regularity and $d$-neighbors property hold for the graphs of simple $d$-polytopes: at each vertex $v$ of a simple $d$-polytope $P$, there are $d$ edges emanating from $v$, and in a bounded polytope, each of these edges leads to another vertex of $P$. These properties do not hold for unbounded polyhedra. The one-subset property, which says that each vertex should have a single $d$-subset, holds for the graphs of polytopes: each vertex in the subset partition graph should contain the $d$-set of incident facets.
There are two easy operations one can perform on a subset partition graph $G = (\mathcal{V},E)$. Let $\mathcal{V}_i$ and $\mathcal{V}_j$ be two vertices in $\mathcal{V}$. Then:
\begin{enumerate}
\item {\bf Contraction:} If $\mathcal{V}_i$ and $\mathcal{V}_j$ are connected by an edge in $E$, contraction on the edge produces a new subset partition graph with one less vertex: the two vertices $\mathcal{V}_i$ and $\mathcal{V}_j$ are replaced with a new vertex which contains all of the $d$-sets which were in $\mathcal{V}_i$ and $\mathcal{V}_j$.
\item {\bf Edge addition:} If $\mathcal{V}_i$ and $\mathcal{V}_j$ were not connected by an edge in $E$, edge addition makes the two vertices adjacent. The resulting subset partition graph has one more edge than the original subset partition graph $G$ does.
\end{enumerate}
\begin{example}
The subset partition graph described in Example~\ref{example:non-polytopal-spg} is obtained from the natural subset partition graph for a $3$-cube after two contractions.
\end{example}
We remark that there is a clear analogue for contraction in the theory of connected layer families. We also note the following simple but potentially powerful effect of these operations, which will be important in Section~\ref{section:final-remarks}:
\begin{remark}\label{remark:effect-of-operations}
We wish to note what happens to the dimension reduction, adjacency, and endpoint-count properties for subset partition graphs under the operations of contraction and edge addition:
\begin{enumerate}
\item All three of these properties are preserved under both operations.
\item After a sufficient number of contractions and edge additions, the resulting graph will be a complete graph, and thus the dimension reduction and adjacency conditions will hold.
\end{enumerate}
\end{remark}
\section{Upper and lower bounds}\label{section:bounds}
In this section, we prove upper and lower bounds on the diameters of subset partition graphs satisfying various combinations of the main properties. First, note the following easy bound:
\begin{remark}\label{remark:cfl-one-subset-bound}
Recall that if we consider subset partition graphs with the dimension reduction property, together with the condition that the graph $G$ is a path, then this is exactly the same as studying connected layer families. If we also add the one-subset property, then the Hirsch bound of $n-d$ holds: up to permutation of symbols, the unique longest path is $\{1, \ldots, d\}, \{2, \ldots, d+1\}, \ldots, \{n-d+1, \ldots, n\}$.
\end{remark}
\subsection{Upper bound}
Here we prove that the diameters of subset partition graphs satisfying the dimension reduction property obey the Kalai-Kleitman quasi-polynomial upper bound of $n^{1+\log_2 d}$.
\begin{theorem}
Let $J(n,d)$ denote the maximal diameter among $d$-dimensional subset partition graphs on $n$ symbols which satisfy dimension reduction. Then, $J(n,d) \leq n^{1+\log_2 d}$.
\end{theorem}
\begin{proof}
Let $G = (\mathcal{V},E)$ be an arbitrary $d$-dimensional subset partition graph of $\mathcal{A}$ on $n$ symbols of maximal diameter which satisfies dimension reduction. Let $A_1$ and $A_2$ be two arbitrary subsets belonging to $\mathcal{A}$ with distance $J(n,d)$. We will construct a $d$-dimensional connected layer family with $n$ symbols whose diameter is $J(n,d)$ which proves it is bounded above by the maximal diameter $C(n,d)$ of connected layer families.
The connected layer family is obtained from the layering process. We define $\mathcal{F}_i := \{ A \in \mathcal{A} : \operatorname{dist}_G(A,A_1) = i\}$. We claim that $\mathcal{F} = \{ \mathcal{F}_0, \ldots, \mathcal{F}_{J(n,d)} \}$ is a $d$-dimensional connected layer family. To prove this, let $i < j < k$ be arbitrary and $A \in \mathcal{F}_i$, $A' \in \mathcal{F}_k$. We prove that there is an $A'' \in \mathcal{F}_j$ such that $F \subseteq A''$, where $F := A \cap A'$. For a contradiction, suppose that there is no $A'' \in \mathcal{F}_j$ containing $F$.
Since $d$-sets in adjacent vertices of $G$ are in the same or adjacent layer of $ \{ \mathcal{F}_0, \ldots, \mathcal{F}_{J(n,d)}\}$, removing all vertices (and incident edges) containing a $d$-set in $\mathcal{F}_j$ in the graph $G$ disconnects the $d$-sets in $\mathcal{F}_i$ from the $d$-sets in $\mathcal{F}_k$. Thus the vertices containing $A$ and $A'$ lie in distinct connected components of $G_F$, contradicting the dimension reduction property. Therefore $J(n,d) \leq C(n,d) \leq n^{1+\log_2 d}$.
\end{proof}
\subsection{Lower bounds}
In this section, we prove lower bounds on the diameters of subset partition graphs satisfying the adjacency and endpoint-count conditions. First, we prove a general lower bound. Then we prove a lower bound for a special subclass of subset partition graphs satisfying natural combinatorial properties coming from an interesting class of polytopes.
\begin{remark}\label{remark:spg-lower-bound-from-clf}
The construction of Eisenbrand et al.{} in~\cite{Eisenbrand:Diameter-of-Polyhedra} proves that subset partition graphs with the dimension reduction property have superlinear diameter, which can be considered evidence against the Linear Hirsch Conjecture.
\end{remark}
The following construction due to Santos (\cite{Santos:PersonalCommunication2011}, which improves the author's previously unpublished bound) gives a superpolynomial lower bound for subset partition graphs with the adjacency and endpoint-count conditions:
\begin{theorem}\label{theorem:general-lower}
Let $K(n,d)$ denote the maximum diameter of $d$-dimensional subset partition graphs with $n$ symbols satisfying the strong adjacency, endpoint-count, and one-subset properties. There is a universal constant $\kappa$ such that
\[ \frac{K(n,d)}{n^{d/4}} \geq \kappa > 0\]
for infinitely many $n$ and $d$.
\end{theorem}
\begin{proof}
Let $d \geq 8$ be a multiple of four. Let $n > d$ be even. We construct a $d$-dimensional subset partition graph $G = (\mathcal{V},E)$ with $n$ symbols. The symbol set is $S := \{1,\ldots,k\} \cup \{1',\ldots,k'\}$, where $k = \frac{n}{2}$.
Let $P$ be a $\frac{d}{2}$-dimensional cyclic polytope with $k$ vertices. The polar of $P$ is a simple $\frac{d}{2}$-polytope with $k$ facets. Let $Q$ and $Q'$ be two copies of the polar of $P$, with respective symbol sets $\Sigma = \{1,\ldots,k\}$ and $\Sigma' = \{1',\ldots,k'\}$ labeling the facets so that the involution $f : s \mapsto s'$ is a combinatorial bijection. Since $Q$ has a Hamiltonian path $\Pi$ (see~\cite{Klee:PathsII}), we order the $t$ vertices of $Q$ as $Z_1, \ldots, Z_{t}$ using the path $\Pi$. Note
\[t = \frac{n}{n-\frac{d}{2}}\binom{\frac{n}{2}-\frac{d}{4}}{\frac{d}{4}}
.\]
(Here, each $Z_i$ is a $\frac{d}{2}$-subset of $\Sigma$, thus their images $(Z_i')_{i=1}^{t}$ with $Z_i' = f(Z_i)$ trace the same Hamiltonian path in $Q'$.)
The vertex set $\mathcal{V}$ of $G$ is $\{\mathcal{V}_i : i = 1,\ldots,t\} \cup \{\mathcal{W}_{i,j} : i = 1,\ldots,t-1;\, j =1,2\}$ and each $\mathcal{W}_{i,j}$ has edges to $\mathcal{V}_i$ and $\mathcal{V}_{i+1}$. See Figure~\ref{figure:almost-path-graph}.
\begin{figure}[hbt]
\begin{center}
\[
\xy
(0,0)*+{\bullet};
(20,0)*+{\bullet};
(40,0)*+{\bullet};
(60,0)*+{\bullet};
(80,0)*+{\bullet};
(10,2)*+{\bullet};
(30,2)*+{\bullet};
(70,2)*+{\bullet};
(10,-2)*+{\bullet};
(30,-2)*+{\bullet};
(70,-2)*+{\bullet};
(0,-3.1)*+{\mathcal{V}_1};
(20,-3.1)*+{\mathcal{V}_2};
(40,-3.1)*+{\mathcal{V}_3};
(60,-3.1)*+{\mathcal{V}_{t-1}};
(80,-3.1)*+{\mathcal{V}_t};
(10,5)*+{\mathcal{W}_{1,1}};
(30,5)*+{\mathcal{W}_{2,1}};
(70,5)*+{\mathcal{W}_{t-1,1}};
(10,-5)*+{\mathcal{W}_{1,2}};
(30,-5)*+{\mathcal{W}_{2,2}};
(70,-5)*+{\mathcal{W}_{t-1,2}};
(50,0)*+{\cdots};
(0,0); (10,-2) **\dir{-};
(20,0); (30,-2) **\dir{-};
(40,0); (45,-1) **\dir{-};
(60,0); (70,-2) **\dir{-};
(20,0); (10,-2) **\dir{-};
(40,0); (30,-2) **\dir{-};
(60,0); (55,-1) **\dir{-};
(80,0); (70,-2) **\dir{-};
(0,0); (10,2) **\dir{-};
(20,0); (30,2) **\dir{-};
(40,0); (45,1) **\dir{-};
(60,0); (70,2) **\dir{-};
(20,0); (10,2) **\dir{-};
(40,0); (30,2) **\dir{-};
(60,0); (55,1) **\dir{-};
(80,0); (70,2) **\dir{-};
\endxy
\]
\end{center}
\caption{The underlying graph of $G = (\mathcal{V},E)$}\label{figure:almost-path-graph}
\end{figure}
Each vertex $\mathcal{V}_i$ and $\mathcal{W}_{i,j}$ contains exactly one $d$-set. Define $\mathcal{V}_i = \{A_i\}$, where $A_i = Z_i \cup Z_i'$. The sets $A_i$ and $A_{i+1}$ have $d-2$ elements in common, so each $D_i := A_i \vartriangle A_{i+1}$ has cardinality $4$, where $\vartriangle$ denotes symmetric difference.
\begin{figure}[hbt]
\begin{center}
\[
\xy
(-10,10)*+{1};
(-20,0)*+{2};
(-10,-10)*+{3};
(10,10)*+{1'};
(20,0)*+{2'};
(10,-10)*+{3'};
(-25,15); (22,15) **\dir{-};
(-25,-2); (22,-2) **\dir{-};
(-25,15); (-25,-2) **\dir{-};
(22,15); (22,-2) **\dir{-};%
(-27.5, 12)*+{A_i};
(-22,-15); (25,-15) **\dir{-};
(-22,2); (25,2) **\dir{-};
(-22,-15); (-22,2) **\dir{-};
(25,-15); (25,2) **\dir{-};%
(0, -18)*+{A_{i+1}};
(-12, 17); (12,17); **\dir{-};
(-12, -12); (12,-12); **\dir{-};
(-12, 17); (-12,-12); **\dir{-};
(12, 17); (12,-12); **\dir{-};%
(0, 19)*+{D_i};
(-13,13); (36, -3); **\dir{-};
(-30,0); (19, -16); **\dir{-};
(-13,13); (-30,0); **\dir{-};
(36,-3); (19,-16); **\dir{-};%
(36,-8)*+{W_{i,2}};
(13,13); (-36, -3); **\dir{-};
(30,0); (-19, -16); **\dir{-};
(13,13); (30,0); **\dir{-};
(-36,-3); (-19,-16); **\dir{-};%
(-35,-8)*+{W_{i,1}};
\endxy
\]
\end{center}
\caption{Venn diagram with $W_{i,\ell}$ for $A_i = \{1,2,1',2'\}$ and $A_{i+1} = \{2,3,2',3'\}$}\label{figure:two-W-sets}
\end{figure}
Consider the two $d$-sets $W_{i,1} = A_i \cup (A_{i+1} \cap D_i \cap \Sigma) \setminus (A_i \cap D_i \cap \Sigma)$ and $W_{i,2} = A_i \cup (A_{i+1} \cap D_i \cap \Sigma') \setminus (A_i \cap D_i \cap \Sigma')$. See Figure~\ref{figure:two-W-sets} for an example. Let $\mathcal{W}_{i,\ell}$ contain $W_{i,\ell}$. It is easy to see that $|A_i \cap A_j| \leq d-2$ if $i \not= j$. It follows that $G$ satisfies the strong adjacency and endpoint-count properties. The diameter of $G$, which is the distance from $\mathcal{V}_1$ to $\mathcal{V}_t$ is $2(t-1)$, which is $\Omega(n^{d/4})$.
\end{proof}
This proves that subset partition graphs without the dimension reduction property do not satisfy the quasi-polynomial bound in~\cite{Kalai:Quasi-polynomial}. Since the diameter of the subset partition graphs in Theorem~\ref{theorem:general-lower} is larger than the best known bound for polytopes, this class may provide a good starting point for the strategy we present to disprove the Linear Hirsch Conjecture which we discuss in Section~\ref{section:final-remarks}.
We now show that a quadratic lower bound holds for a special class of subset partition graphs inspired by spindles, which were instrumental in Santos' disproof of the Hirsch Conjecture~\cite{Santos:CounterexampleHirsch}. A \emph{spindle} is a polytope $P$ with two special vertices $A_1$ and $A_2$ (called the \emph{apices}) such that every facet of $P$ contains exactly one of the apices. The \emph{length} of a spindle $P$ is the distance in the graph of $P$ between $A_1$ and $A_2$. We say that a $d$-dimensional spindle is \emph{long} if its length exceeds $d$. In~\cite{Santos:CounterexampleHirsch}, Santos proved that long $5$-dimensional spindles exist, and moreover, the existence of a long spindle implies the existence of a long spindle with $n=2d$, thus the Hirsch Conjecture for polytopes is false. (Recall that a \emph{Dantzig figure} is an abstract polytope $G= (\mathcal{A},E)$ in the sense of~\cite{AdlerDantzigMurty:AbstractPolytopes} where $n=2d$, together with two disjoint $d$-sets $A_1,A_2 \in \mathcal{A}$. The construction of Santos is a polytopal realization of a non-Hirsch Dantzig figure.)
The property that characterizes spindles is purely combinatorial, so we make an analogous definition for subset partition graphs. We say that a subset partition graph $G$ of $\mathcal{A}$ on the symbol set $S$ \emph{satisfies the spindle property} if there are two distinguished subsets $A_1$ and $A_2$ (called the \emph{apices}) belonging to $\mathcal{A}$, such that every symbol $s \in S$ belongs to exactly one of $A_1$ or $A_2$. The \emph{length} of a subset partition graph $G$ with the spindle property is the distance in $G$ from one apex to the other.
\begin{theorem}\label{theorem:spg-spindle-length}
Let $L(n)$ denote the maximum length of $d$-dimensional subset partition graphs on $n=2d$ symbols satisfying the strong adjacency, endpoint-count, one-subset, and spindle properties.
Let $\kappa = \frac18$. For infinitely many $n$,
\[\frac{L(n)}{n^2} \geq \kappa.\]
\end{theorem}
\begin{proof}
We define a subset partition graph $G_m = (\mathcal{V},E)$ of $\mathcal{A}$ for each $m \in \mathbb{N}$. Let $S = [2m] \times \{1,2\}$, so $n = |S| = 4m$. Define the index set
\[ I := \{(a,b,c) : 0 \leq a, b < m;\, c=0,1\} \cup \{(m,0,0)\}.\]
The triples $(a,b,c) \in I$ are totally ordered using the lexicographic order $<_{\text{lex}}$. Note that since $c \in \{0,1\}$, if $a=a'$ and $b=b'$, then $(a,b,c)$ and $(a',b',c')$ are either equal or consecutive in $<_{\text{lex}}$. For $(a,b,c) \in I$, define
\begin{align*}
A_{a,b,c} \, :=
\,\, &\{ (i,j) : i = a+1, \ldots, a+m-b-1 \text{ and } j = 1, 2\} \\
\cup \,\, &\{ (i,j) : i = a+m-b \text{ and } j = c+1, \ldots, 2\} \\
\cup \,\, &\{ (i,j) : i = a+m-b+1 \text{ and } j = 1, \ldots, c\} \\
\cup \,\, &\{ (i,j) : i = a+m-b+2,\ldots, a+m+1 \text{ and } j = 1, 2\}.
\end{align*}
Figure~\ref{figure:typical-d-set} gives schematic pictures of typical sets $A_{a,b,c}$ in $\mathcal{A}$. Let $\mathcal{A} := \{A_{a,b,c} : (a,b,c) \in I\}$. For all $(a,b,c) \in I$, one has $|A_{a,b,c}| = 2m$, so the dimension of $G_m$ is $d=2m=\frac{n}{2}$.
\begin{figure}[hbt]
\begin{center}
\[
\xy
(-52,3)*+{c=0};
(-60,0); (-60,-56) **\dir{-};
(-56,0); (-56,-56) **\dir{-};
(-52,0); (-52,-56) **\dir{-};
(-60,0); (-52, 0) **\dir{-};
(-60,-4); (-52,-4) **\dir{-};
(-60,-8); (-52,-8) **\dir{-};
(-60,-12); (-52,-12) **\dir{-};
(-60,-16); (-52,-16) **\dir{-};
(-60,-20); (-52,-20) **\dir{-};
(-60,-24); (-52,-24) **\dir{-};
(-60,-28); (-52,-28) **\dir{-};
(-60,-32); (-52,-32) **\dir{-};
(-60,-36); (-52,-36) **\dir{-};
(-60,-40); (-52,-40) **\dir{-};
(-60,-44); (-52,-44) **\dir{-};
(-60,-48); (-52,-48) **\dir{-};
(-60,-52); (-52,-52) **\dir{-};
(-60,-56); (-52,-56) **\dir{-};
(-58,-18)*+{\bullet};
(-54,-18)*+{\bullet};
{\ar (-46,-18)*+!L{i=a+1}; (-50,-18)*{}};
(-58,-22)*+{\bullet};
(-54,-22)*+{\bullet};
(-58,-26)*+{\bullet};
(-54,-26)*+{\bullet};
{\ar (-46,-26)*+!L{i=a+m-b-1}; (-50,-26)*{}};
(-54,-30)*+{\bullet};
(-58,-30)*+{\bullet};
{\ar (-46,-30)*+!L{i=a+m-b}; (-50,-30)*{}};
{\ar (-46,-34)*+!L{i=a+m-b+1}; (-50,-34)*{}};
(-58,-38)*+{\bullet};
(-54,-38)*+{\bullet};
{\ar (-46,-38)*+!L{i=a+m-b+2}; (-50,-38)*{}};
(-58,-42)*+{\bullet};
(-54,-42)*+{\bullet};
(-58,-46)*+{\bullet};
(-54,-46)*+{\bullet};
{\ar (-46,-46)*+!L{i=a+m+1}; (-50,-46)*{}};
(-59.3,-58).(-56.8,-58)!C *\frm{_\}},+U*++!U\txt{$c$};
(-62,0).(-62,-16)!C *\frm{\{},+L*++!R\txt{$a$};
(-62,-36).(-62,-48)!C *\frm{\{},+L*++!R\txt{$b$};
(8,3)*+{c=1};
(0,0); (0,-56) **\dir{-};
(4,0); (4,-56) **\dir{-};
(8,0); (8,-56) **\dir{-};
(0,0); (8, 0) **\dir{-};
(0,-4); (8,-4) **\dir{-};
(0,-8); (8,-8) **\dir{-};
(0,-12); (8,-12) **\dir{-};
(0,-16); (8,-16) **\dir{-};
(0,-20); (8,-20) **\dir{-};
(0,-24); (8,-24) **\dir{-};
(0,-28); (8,-28) **\dir{-};
(0,-32); (8,-32) **\dir{-};
(0,-36); (8,-36) **\dir{-};
(0,-40); (8,-40) **\dir{-};
(0,-44); (8,-44) **\dir{-};
(0,-48); (8,-48) **\dir{-};
(0,-52); (8,-52) **\dir{-};
(0,-56); (8,-56) **\dir{-};
(2,-18)*+{\bullet};
(6,-18)*+{\bullet};
{\ar (14,-18)*+!L{i=a+1}; (10,-18)*{}};
(2,-22)*+{\bullet};
(6,-22)*+{\bullet};
(2,-26)*+{\bullet};
(6,-26)*+{\bullet};
{\ar (14,-26)*+!L{i=a+m-b-1}; (10,-26)*{}};
(6,-30)*+{\bullet};
{\ar (14,-30)*+!L{i=a+m-b}; (10,-30)*{}};
(2,-34)*+{\bullet};
{\ar (14,-34)*+!L{i=a+m-b+1}; (10,-34)*{}};
(2,-38)*+{\bullet};
(6,-38)*+{\bullet};
{\ar (14,-38)*+!L{i=a+m-b+2}; (10,-38)*{}};
(2,-42)*+{\bullet};
(6,-42)*+{\bullet};
(2,-46)*+{\bullet};
(6,-46)*+{\bullet};
{\ar (14,-46)*+!L{i=a+m+1}; (10,-46)*{}};
(0.7,-58).(3.2,-58)!C *\frm{_\}},+U*++!U\txt{$c$};
(-2,0).(-2,-16)!C *\frm{\{},+L*++!R\txt{$a$};
(-2,-36).(-2,-48)!C *\frm{\{},+L*++!R\txt{$b$};
\endxy
\]
\end{center}
\caption{Schematic drawings of the $d$-sets $A_{a,b,c} \in \mathcal{A}$, $m=7$ for $c=0$ and $c=1$}\label{figure:typical-d-set}
\end{figure}
The elements in $\mathcal{V}$ are the singleton sets $\mathcal{V}_{a,b,c} = \{A_{a,b,c}\}$ for $(a,b,c) \in I$, thus the one-subset property holds. Two vertices $\mathcal{V}_{a,b,c}$ and $\mathcal{V}_{a',b',c'}$ are connected by an edge if and only if the triples $(a,b,c)$ and $(a',b',c')$ are consecutive in $<_{\text{lex}}$, which is a total order on $I$. Therefore the graph for $G_m$ is a path, so the graph is connected. Hence, $G_m$ is a subset partition graph.
The $d$-sets $A_{0,0,0} = \{(i,j) : 1 \leq i \leq m; 1 \leq j \leq 2\}$ and $A_{m,0,0} = \{(i,j) : m+1 \leq i \leq 2m, 1 \leq j \leq 2\}$ are a disjoint partition of $S$, thus $A_{0,0,0}$ and $A_{m,0,0}$ are the apices of the spindle $G_m$.
We prove the adjacency property by showing that if the $d$-sets $A:=A_{a,b,c}$ and $A':=A_{a',b',c'}$ are not in adjacent vertices, then $|A \cap A'| < d-1$, or equivalently, $|A \vartriangle A'| \geq 2$. Since $(a,b,c) \not= (a',b',c')$, without loss of generality $(a,b,c) <_{\text{lex}} (a',b',c')$. For integers $\alpha,\beta,\gamma,\iota_1,\iota_2,\dots$ define
\[ I_{\alpha,\beta,\gamma}(\iota_1,\iota_2,\dots) := | \{ (i,j) \in A_{\alpha,\beta,\gamma} : i = \iota_1,\iota_2,\dots \} |, \]
and use $I(\iota_1,\iota_2,\dots) := I_{a,b,c}(\iota_1,\iota_2,\dots)$ and $I'(\iota_1,\iota_2,\dots) := I_{a',b',c'}(\iota_1,\iota_2,\dots)$ respectively to denote the number of elements with $i \in \{\iota_1,\iota_2,\dots\}$ in $A$ and $A'$ respectively. The proof is given in cases.
\begin{itemize}
\item Suppose $a' = a$. Since $A$ and $A'$ are not adjacent, $b' \not= b$. Either $b'-b=1$ or $b'-b>1$.
\begin{itemize}
\item
Suppose $b'-b=1$. If $c=0$, then $I(a+m-b+1)=0$ but $I'(a+m-b+1)=I'(a'+m-b'+2)=2$. If $c=1$, then non-consecutivity of $(a,b,c)$ and $(a',b',c')$ implies $c'=1$, hence both $(a+m-b-1,1)$ and $(a+m-b,2)$ are in $A$ but neither are in $A'$. In either case $|A \vartriangle A'| \geq 2$.
\item
If $b'-b>1$, then $I(a+m-b,a+m-b+1) = 2$ but $I'(a+m-b,a+m-b+1) = 4$, so $|A \vartriangle A'| \geq 2$.
\end{itemize}
\item Otherwise, $a' \not= a$. Exactly one of the following holds:
\begin{enumerate}
\item $a' - a > 1$,
\item $a' - a = 1$ and $b < m-1$,
\item $a' - a = 1$ and $b = m-1$ and $c=0$,
\item $a' - a = 1$ and $b = m-1$ and $c=1$ and $b' > 0$, or
\item $a' - a = 1$ and $b = m-1$ and $c=1$ and $b' = 0$.
\end{enumerate}
In the first three cases, $I(a+1)=2$ but $I'(a+1)=0$. In the fourth case, $I'(a'+m+1)=2$ but $I(a'+m+1)=0$. In the last case, non-consecutivity of $(a,b,c)$ and $(a',b',c')$ implies $c'=1$. Then $(a+1,2)$ and $(a+m+1,1)$ are in $A$, but neither are in $A'$. In all cases, $|A \vartriangle A'| \geq 2$.
\end{itemize}
Since non-adjacent $d$-sets $A$ and $A'$ have strictly less than $d-1$ symbols in common, the adjacency property holds. Moreover, it is clear from the definition of $A_{a,b,c}$ that if $A$ and $A'$ are adjacent then $|A \cap A'| = d-1$, so strong adjacency holds. Thus endpoint-count holds since the degree of each vertex in $G_m$ is one or two.
The underlying graph of $G_m$ is a path and its length, from $A_{0,0,0}$ to $A_{m,0,0}$, is $2m^2 = \frac18n^2$, thus $\limsup_{n \rightarrow \infty} \frac{L(n)}{n^{2}} \geq 2^{-3}$.
\end{proof}
Since the length is a lower bound for the diameter:
\begin{corollary}\label{corollary:superlinear-bound}
Let $M(n)$ denote the maximum diameter of $d$-dimensional subset partition graphs on $n=2d$ symbols satisfying the strong adjacency, endpoint-count, and spindle properties. Then,
\[ \limsup_{n \rightarrow \infty} \frac{M(n)}{n^2} \geq \kappa = \frac18.\]
\end{corollary}
Very recently, the lower bound to $M(n)$ was improved by H\"{a}hnle in~\cite{Hahnle:SPGs}.
\section{Final remarks and open problems}\label{section:final-remarks}
We saw that the Kalai-Kleitman diameter upper bound in~\cite{Kalai:Quasi-polynomial} holds for subset partition graphs satisfying dimension reduction. While we do have a lower bound for diameters of subset partition graphs with the strong adjacency and endpoint-count conditions, we ask:
\begin{problem}
Prove a non-trivial upper bound on the diameters of subset partition graphs with the strong adjacency and endpoint-count conditions.
\end{problem}
Subset partition graphs that satisfy the first main property, namely dimension reduction (see Remark~\ref{remark:spg-lower-bound-from-clf}), or the last two main properties, namely strong adjacency and endpoint-count (see Theorem~\ref{theorem:general-lower}), have superlinear diameter. Both of these results, combined with the fact that complementary sets of conditions are used, can be considered evidence against the Linear Hirsch Conjecture. Theorem~\ref{theorem:spg-spindle-length}, which presents a quadratic diameter lower bound for a special subclass provides even further evidence against the Linear Hirsch Conjecture. In light of this, we ask:
\begin{problem}
Construct a family of subset partition graphs with superlinear diameter satisfying all of the main properties.
\end{problem}
In fact, subset partition graphs provide an approach for satisfying all three properties and, moreover, an approach for disproving the Linear Hirsch Conjecture:
\begin{enumerate}
\item Start with a family of subset partition graphs satisfying at least the endpoint-count property with superlinear diameter growth, such as the family resulting from either Theorems~\ref{theorem:general-lower} or~\ref{theorem:spg-spindle-length}.
\item Gain the other main properties that do not yet hold with the contraction and edge addition operations (see Remark~\ref{remark:effect-of-operations}).
\item If the resulting family of graphs still has superlinear diameter, realize the sequence of graphs as a sequence of polytopes.
\end{enumerate}
We identify the principal difficulties with this strategy: in the second step the contraction and edge addition operations are liable to significantly reduce the diameter of the subset partition graphs, and in the third step the realization problem for polytopes and the study of polytopality of graphs is still the subject of ongoing research, e.{}g.{},~\cite{Joswig:NeighborlyCubical}, \cite{Matschke:Prodsimplicial}, \cite{Pfeifle:PolytopalityCartesian}, \cite{Pilaud:MultiPseudoTriangulationsRealization}, \cite{Ziegler:ProjectedProductsPolygons}.
In light of these difficulties, for the purpose of the above approach, we note that \emph{any} superlinear construction of subset partition graphs satisfying the endpoint-count property is useful, since the method attempts to construct polytopes using some construction of subset partition graphs as a starting point: while the bound in Theorem~\ref{theorem:general-lower} is much better than the one in Theorem~\ref{theorem:spg-spindle-length}, it could be that steps 2 and 3 above are easier to perform from the construction in Theorem~\ref{theorem:spg-spindle-length}. Thus, any superlinear construction of subset partition graphs satisfying endpoint-count is interesting.
By considering different combinations of properties, subset partition graphs provide a framework for describing which conditions are crucial in proving upper and lower bounds. It is natural to ask which combination of properties is most useful in combinatorial abstractions for superlinear lower bounds:
\begin{question}
Are there superlinear lower bounds for subset partition graphs satisfying other non-trivial combinations of properties?
\end{question}
Finally, is a certain combination of properties sufficient for proving the Polynomial Hirsch Conjecture for subset partition graphs, and thus, for polyhedra?
\section*{Acknowledgments}
I thank Jes\'us A. De Loera, Nicolai H\"ahnle, Frederik von Heymann, Gil Kalai, Vincent Pilaud, Frank Vallentin, and G\"unter M. Ziegler for their remarks. I am especially grateful to Francisco Santos for the in-depth suggestions. I express my gratitude to the anonymous referees for their helpful comments.
|
\section{Introduction}
A growing amount of experimental data
\cite{Engel:2007cr,Panitchayangkoon:2010vn,Fleming:2011tg} suggest that quantum
coherence may be at the origin of the stunning efficiency of exciton transport
in photosynthetic light harvesting, even at ambient temperatures and in a
doubtlessly very noisy environment. The simplest natural structure that exhibits
these surprising properties is the Fenna-Matthews-Olson (FMO) light harvesting
complex of green sulfur bacteria, which consists of seven (or eight
\cite{Olbrich:2010ly}) bacteriochlorophyll molecules arranged in a disordered
network \cite{Fenna:1975ys,Amerongen:2000fk}. Exciton transfer is here mediated
by dipole-dipole coupling between these different molecular sites, and is
associated with a de-excitation of the donor site from some excited to the
ground state, and an excitation of the acceptor site from the ground to some
excited state. The molecular network that mediates the exciton transfer, from
the antenna complex to the reaction center, where the excitation fuels the
organism's chemistry, is embedded into a complicated protein structure, which
seems to provide some structural stiffness, and also defines a nontrivial
spectral structure of the environment, which preserves the coherence on the FMO
network itself, on considerably longer time scales than to be expected for a
white noise environment \cite{Cheng:2009ek}. Since the prevalence of the
coherence effects in widely open systems in contact with high temperature
environments challenges our traditional understanding of what seemed the
restricted realm of quantum mechanics, these highly specialized biological
functional units let us think anew how to control efficient transport in
disordered and noisy systems, possibly by exploiting fundamental quantum
features.
The available experimental evidence, still not always fully consistent
\cite{Engel:2007cr,Panitchayangkoon:2010vn} and vividly debated, is mirrored by
a large variety of theoretical approaches, which distinguish themselves in terms
of the applied methodology as well as of the level of faithfulness of the
modeling of the actual biomolecular object under scrutiny---from advanced
quantum molecular dynamics \cite{Olbrich:2010ly}, over quantum simulations
\cite{Rebentrost:2009vn,Chin:2010fk,Wu:2010vn} based on some effective, seven
site Hamiltonian \cite{Adolphs:2006ve}, Lindblad equations
\cite{Abramavicius:2010fk} or various non-Markovian approaches derived from open
system theory
\cite{Hughes:2009hc,Thorwart:2009fk,Liang:2010nx,Sarovar:2010fk,muelken10}, to
diverse quantum optical models \cite{cai10,alicki11,wuester10} and abstract
statistical treatments \cite{scholak10,Scholak:2011uq}. In essence, there are
four principal lines of argument to explain efficient exciton transport, by (i)
noise-assisted, (ii) non-Markovian or (iii) driving, and (iv) multi-path
interference induced exciton transfer. Given the available experimental data and
the hitherto limited characterization and control of the precise microscopic
Hamiltonian and environment coupling agents that generate the experimentally
observed phenomena, it remains an open question which of these suggested
mechanisms were actually used by evolution to optimize the FMO's functionality.
Since any improvement that provides an evolutionary advantage will be
implemented, it is even not unlikely that all of them are used at some level.
However, in the light of the debate about a possible {\em quantum} enhancement
of transport in the photosynthetic complex, it is highly relevant to understand
the specific role of quantum coherence for these different mechanisms, and to
compare the achievable transfer efficiencies they allow for. It is the present
contribution's purpose to provide such comparison for noise-enhanced and
multi-path-interference transport scenarios. Large scale statistical sampling
will allow us to show that multi-path quantum interference always leads to
better results than noise-assisted, essentially classical transport processes,
though requires additional optimization of the molecular structure. Indeed, this
even holds in the presence of not too strong ambient noise, and thus identifies
yet another scenario where genuine quantum effects may define an evolutionary
advantage.
\section{Random molecular networks}
The noise-enhanced as well as the multi-path-interference scenario start out
from the same structural elements that are unambiguously given by experimental
observations: Exciton transfer occurs across a random molecular network with
local sites effectively modeled as electronic two-level systems, and further
background degrees of freedom, which may possibly exert some effective driving
\cite{cai10}, can in principle be accounted for by some non-Markovian
environment coupling \cite{Hughes:2009hc}. Experimentally observed exciton
cross-terms \cite{Engel:2007cr}, though with quite some experimental scatter
\cite{Panitchayangkoon:2010vn}, suggest excitonic coherence times which exceed
the exciton transfer time between antenna and reaction center by a factor two to
five \cite{Engel:2007cr}, even at ambient temperatures
\cite{Panitchayangkoon:2010vn}. The presently best available effective
Hamiltonian, which is inferred from experimental data and advanced quantum
chemical model calculations predicts a strong suppression of strictly coherent
transport across the FMO complex (approx. $5\ \%$ transfer efficiency
\cite{scholak10,Scholak:2011uq}), from the input to the output site, due to
predominantly destructive multi-path interference upon transmission.
However, exciton transfer in the FMO complex occurs with efficiencies larger
than $95\ \%$ \cite{Cheng:2009ek}, and there are essentially two alternative
scenarios which can explain this discrepancy between the experimental data and
the best microscopic Hamiltonian presently available. Since quantum coherence
can be destroyed by ambient noise, and since suppressed transport under purely
coherent dynamics can only be due to destructive interference effects, it is
very natural to argue in favor of noise-induced transport \cite{arndt91}. Since,
however, destructive multi-path interference is known to be very sensitive with
respect to changes of the boundary conditions and/or the Hamiltonian's coupling
matrix elements \cite{Kramer:1993yb}, and since even the best available model
Hamiltonian for the FMO complex is garnished with appreciable uncertainties for
its individual entries \cite{Adolphs:2006ve}, one may equally well argue in
favor of {\em constructive} multi-path interference as the observed efficiency's
cause. This even more so since all experimental data that so far lend support
for the coherent transport hypothesis are obtained from bulk measurements rather
than from single molecule spectroscopy, and therefore might mask much longer
coherence times by inhomogeneous broadening effects \cite{Fleming:2011tg}.
Since both alternative explanations allow to predict large transfer
efficiencies, let us have a closer look at the respective key mechanisms. We
model the energy conserving, unitary dynamics of a single
excitation injected into a molecular network alike the FMO complex by the
Hamiltonian
\begin{align}
H &= \sum_{i \neq j = 1}^7 v_{i,j} \ket{i}\bra{j}\, ,
\label{ham}
\end{align}
with $\ket{i}$ and $\ket{j}$ the electronic states where the excitation is
localized at the individual molecular sites $i$ and $j$, respectively. We assume
that initially only one site, ``in'', is excited, which is identified with the
first site of the network. The sites No.~$2$ to $6$ are referred to as the
``intermediates''. The seventh and last site is the designated output site,
``out'', where we add an energy sink, such as to mimic the irreversible exciton
absorption at the reaction center. We couple each of the seven molecular sites
to a private (\emph{i.e.}{}, there is no inter-site communication through the
environment) dephasing environment. Sink and dephasing induce some
irreversibility on the FMO degrees of freedom, which we incorporate by the
Lindblad terms
\begin{align}
L_{\mathrm{sink}}(\varrho) &= \Gamma
\bigl(\ket{0}\bra{\mathrm{out}}\varrho\ket{\mathrm{out}}\bra{0} - \tfrac{1}{2}
\{\ket{\mathrm{out}}\bra{\mathrm{out}}, \varrho\}\bigr),
\label{lind_sink}
\end{align}
where $\ket{0}$ and $\{,\}$ are the ground state of the molecular
network and an anticommutator, respectively, and
\begin{align}
L_{\mathrm{deph}}(\varrho) &= - 4 \gamma \sum_{i \neq j = 1}^7
\ket{i}\bra{i}\varrho\ket{j}\bra{j}
\label{lind_deph}
\end{align}
into the effective evolution equation of the excitonic state on the
network,
\begin{align}
\dot{\varrho}(t) &= - \mathrm{i} \, [H, \varrho(t)] +
L_{\mathrm{sink}}(\varrho(t)) + L_{\mathrm{deph}}(\varrho(t))
\label{lind_FMO}
\end{align}
(where we have set $\hbar = 1$ for convenience). To obtain a robust
comparison of the different transport mechanisms, we statistically sample the
transport efficiency over different realizations of $H$, by random sampling over
the positions of all intermediate sites, within a sphere with input and output
site placed on its north and south pole, respectively. Random positions
$\vec{r}_j$ translate into random realizations of $H$ by defining the coupling
matrix elements as a function of the intersite distances,
\begin{align}
v_{i,j} &= \alpha \, r_{i,j}^{-3}
\label{random_v}
\end{align}
with some constant $\alpha$, and $r_{i,j}=|\vec{r}_i-\vec{r}_j|$. By choosing
random positions for the intermediate sites, we in some way take into account
the experimental uncertainty concerning the actual FMO Hamiltonian
\cite{Adolphs:2006ve}. Let us stress, however, that we do not present our model
as a realistic description of a particular experiment. On the contrary, we are
rather interested in general properties of transport, as they will appear if we
allow for a large variety of possible configurations, without imposing, from the
very beginning, strict experimental boundary conditions.
The model discussed above has three intrinsic time scales which will largely
determine the expected transfer efficiencies---the direct exciton transfer time
\begin{align}
T &= \frac{\pi}{2 |v_{\mathrm{in},\mathrm{out}}|}
\label{eq:T}
\end{align}
between input and output site, in the absence of all intermediate sites, the
local dephasing rate $4\gamma$ (identical for all sites), and the sink
dissipation rate $\Gamma$, which we will fix at the value $\Gamma = 10/T$ in the
following \footnote{This implies one incoherent event on the time scale $T/10$,
hence a rather efficient drain. We have used this time scale as an efficiency
benchmark in earlier work \cite{scholak10,Scholak:2011uq}.}. Randomly placed
intermediate sites between input and output help to enhance or suppress coherent
transport \cite{scholak10,Scholak:2011uq}, by larger coupling strengths between
closer sites and appropriate phase relationships. The dephasing rate defines the
time scale $\mathfrak{T}_{\mathrm{deph}}=(4\gamma)^{-1}$ on which such phase
relationships can have a bearing on the overall transport behavior, and the sink
dissipation rate defines an optimal time scale on which population has to be
delivered to the output site, to make it immediately available for the reaction
center.
\section{Statistics of transport efficiency}
\begin{figure*}
\hfill
(a)\raisebox{-26.04681pt}{\includegraphics{topography_a}}
\hfill
(b)\raisebox{-28.42184pt}{\includegraphics{topography_b}}
\hfill
\caption{(a) Probability density $f_{\mathfrak{T}}$ of the average excitation
transfer time $\mathfrak{T}$, eq.~\eqref{eq:mathfrakT}, for $N=7$ molecular
sites and sink rate $\Gamma=10/T$, as a function of the dephasing rate
$\gamma$. The two dotted, diagonal lines are given by the dephasing time
$\mathfrak{T}_{\mathrm{deph}}= (4 \gamma)^{-1}$, and by an approximate Zeno
time $\mathfrak{T}_{\mathrm{Zeno}}\propto\gamma$, respectively. On time scales
$\mathfrak{T}>\mathfrak{T}_{\mathrm{deph}}$, the purity of the excitonic state
on the molecular network has dropped to its minimum value, hence the transport
is essentially classical. The white line shows the median
$\tilde{\mathfrak{T}}$, the dot-dashed line the minimum transfer time, and the
dashed line the transfer time of a configuration that has been optimized for
$\gamma = 0$. (b) Height profiles of the probability densities for fixed
dephasing rates $\gamma =0$ (solid line), $10^{-5}\Gamma $ (dashed line),
$1.06\Gamma $ (dotted line), $10^3\Gamma $ (dash-dotted line).}
\label{topography}
\end{figure*}
To compare the transport efficiency provided by different molecular
conformations and different dephasing rates, we define the average transfer time
\begin{align}
\mathfrak{T} &= \Gamma\int_0^\infty t \, p_{\mathrm{out}}(t) \, \mathrm{d}t \,
,
\label{eq:mathfrakT}
\end{align}
which is the expectation value of the time required to transfer the excitation
to the sink, determined by the population $p_{\mathrm{out}}(t) = \bra{0}
\dot{\varrho}(t) \ket{0} / \Gamma$ of the output site. We see here that the
ground state $\ket{0}$ can only be populated by delivering the exciton from the
output site to the sink, with rate $\Gamma$, cf.{} also eq.~\eqref{lind_sink}.
The shorter the transfer time, the more efficient the transport \footnote{Note
that $\mathfrak{T}$ is a reasonable efficiency quantifier if efficiency is
qualified as rapid and irreversible excitation transfer to the sink, but that it
does not distinguish quantum from classical transport efficiencies, since it
integrates over all times. It is however evident that the definition of any
efficiency quantifier is a matter of pragmatic choice rather than of principle,
and that all such quantifiers call for a careful interpretation}.
We are now prepared for a statistical analysis of exciton transfer times across
a molecular network alike the FMO complex, to assess the potential role of
coherent vs. noise-assisted transport mechanisms to steer transport
efficiencies. In order to draw a landscape of the exciton transport efficiency
in molecular networks as modeled by eq.~\eqref{lind_FMO}, we sample over fifty
million random and distinct conformations. For each conformation in the ensemble
the computational procedure involves the following tasks: First, the positions
of the five intermediate molecular sites are randomized, which are then used to
populate the matrix entries of the Hamiltonian $H$ according to
eq.~\eqref{random_v}. For values of $\gamma$ from $10^{-5} \Gamma$ to $10^3
\Gamma$ in 200 logarithmic steps, \emph{i.e.}{} for essentially vanishing to very strong
dephasing rates, the master equation, eq.~\eqref{lind_FMO}, is solved via exact
numerical diagonalization and the transfer times $\mathfrak{T}$ are calculated.
The last step employs logarithmic data binning to record the transfer time
histogram $f_{\mathfrak{T}}$, shown in Fig.~\ref{topography}, as a function of
$\gamma$. In the left density plot of the figure, the grey scale represents the
probability density of configurations giving rise to a certain transfer time,
under a given dephasing rate. Configurations above the left hand, decreasing
dotted line correspond to transfer times {\em longer} than
$\mathfrak{T}_{\mathrm{deph}}$, \emph{i.e.}{} the exciton transfer is here due to {\em
classical transport} across the network. For even larger dephasing rates,
transport is ``frozen'' due to a Zeno-like projection mechanism
($\mathfrak{T}_{\mathrm{Zeno}} \propto \gamma$, as indicated by the increasing
dotted line), while configurations below the left hand dotted line achieve
efficient transfer on time scales {\em shorter} than
$\mathfrak{T}_{\mathrm{deph}}$. The latter thus mediate {\em coherent exciton
transfer} and are the only ones which are eligible for claiming an unambiguous
{\em quantum enhancement} of exciton transfer in the FMO complex. Indeed, the
shortest transfer times are observed for a finite subset of these
configurations, on the lower left hand side of Fig.~\ref{topography}. If we
furthermore optimize the molecular configuration in the absence of any
dephasing, for the same sink dissipation, and then expose this optimal
configuration to finite dephasing rates, we obtain the dashed curve in
Fig.~\ref{topography}, which exhibits exciton transfer times {\em shorter} than
the classical value, associated with the highest density of configurations in
the plot, by at least a factor two, for almost all dephasing rates that allow
for coherent transfer. Note that this optimal configuration still yields very
efficient excitation transfer even for higher dephasing rates, which suggests
that the optimized coherent transport on short time scales imparts an initial
advantage, even when classical activation takes over on long time scales.
\section{Discussion}
The optimization landscape thus provides a clear picture of the possible
strategies to achieve efficient transfer across an FMO like, disordered network:
The large majority of randomly sampled configurations requires the {\em
assistance of noise} to allow for efficient transport, which, however, will be
{\em classical} in nature, and will occur on time scales clearly longer than
$\mathfrak{T}_{\mathrm{deph}}$. Indeed, the clustering of all the configurations
associated with this transport mechanism around $\mathfrak{T} \simeq 10/\Gamma
=T$ and $\gamma \simeq \Gamma$ in Fig.~\ref{topography} expresses an approximate
matching condition of dephasing and sink dissipation rates, which is an
intrinsically classical, statistical synchronization phenomenon. Excitation
transfer here stems from the noise induced destruction of quantum coherence, is
induced on time scales {\em longer} than the system inherent coherence times,
and is essentially {\em independent} of the microscopic Hamiltonian which
generates the coherent dynamics on the network (hence the large density of
configurations at these transfer times and dephasing rates). Under an
evolutionary perspective, this non-selectivity with respect to the microscopic
coupling structure can therefore neither define an evolutionary advantage by
optimizing the molecular structure---most structures will do, as also
highlighted by the pronounced minimum of the median of the distribution (white
curve) at $\gamma\simeq\Gamma$.
In contrast, Fig.~\ref{topography} also shows that a small but finite
sub-ensemble of {\em optimal configurations} mediates {\em efficient quantum
transport} of the excitation from input to output, faster than the
noise-assisted, classical transport, and due to constructive quantum
interference upon transmission, which prevails in the presence of noise,
actually in the entire range of dephasing rates considered in
Fig.~\ref{topography}. If quantum mechanics is at the origin of the
experimentally observed exciton transfer efficiencies in the FMO complex, it
therefore must stem from such type of optimal molecular configurations, or
otherwise ought to be induced by non-Markovian \cite{Thorwart:2009fk} or driving
effects \cite{cai10}. Since the optimal transfer rates mediated by such optimal
configurations define rare events in the statistical sample represented in
Fig.~\ref{topography}, it is conceivable that they define an evolutionary
advantage which was hardwired by nature. The cusp-like structure that emerges
right along the line defined by the dephasing time
$\mathfrak{T}_{\mathrm{deph}}$ highlights molecular configurations which exhibit
an eigenstate of $H$ localized on the site ``in'' and on another site $j \neq
\text{``out''}$. For these configurations, coherent transmission is suppressed
by destructive interference, and hence noise is required to assist the exciton's
delivery at the sink, with transfer times $\mathfrak{T} =
\mathfrak{T}_{\mathrm{deph}}$.
Let us conclude with a short discussion of the actual structure of optimal
configurations in the coherent regime: When investigating the molecular
conformations which exhibit transfer times shorter than the dephasing time
$\mathfrak{T}_{\mathrm{deph}}$ in Fig.~\ref{topography}, these show a wide
variability as regards their dimensionality and symmetry properties. Minimal
transfer times are achieved by near to one dimensional structures, while fully
three dimensional and prima facie disordered structures, that mediate optimal
coherent transport on closed molecular networks with vanishing sink dissipation,
achieve slightly longer transfer times $\mathfrak{T}$, though still
significantly shorter than the dephasing time. Since Fig.~\ref{topography} is
obtained for a relatively large sink dissipation rate, which corresponds to a
significant change of boundary conditions as compared to the closed network,
this is not surprising, though raises the question for the actual optimality
conditions favored by nature: Besides rapid transfer, also robustness and
multifunctionality issues define further constraints which go far beyond the
scope of our present model considerations, but define a beautiful and
fascinating area for further research. Therefore, it remains an open
question, whether nature indeed employs quantum coherence to benefit energy
transport in photosynthesis by having evolved molecular conformations
corresponding to the optimal configurations discovered here. It appears that due
to the strong sensitivity of the transport efficiency on the particular
realization of disorder this question can only be answered by a highly precise
measurement of the electronic Hamiltonian with an accuracy beyond what is
accessible today. Finally, regardless of whether the near-to-perfect efficiency
of the FMO complex is truly caused by quantum coherence or not, the existence
of optimal configurations achieving maximum performance due to constructive
quantum interference will certainly spur the design of new experiments and, in
the long run, advanced devices such as a new generation of organic solar cells
that utilize the beneficial aspects of quantum coherence.
|
\section{The Jones matrix representation of birefringent medium}
A light beam is said to be polarized whenever it is transmitted through a certain crystalline medium that allows electrical anisotropy \cite{paper7}.This change of polarization state can be written as
\begin{equation}
\varepsilon=M \varepsilon_o
\end{equation}
where $\varepsilon$ and $\varepsilon_o$ are the respective final and initial polarization state and M is polarization matrix respectively.
If the polarization of light remains unaltered after passing through any optical element then the state can be identified as the eigenvector of the optical component and in the language of matrix,Jones had shown the condition
\begin{equation}
M_i \varepsilon_i=d_i\varepsilon_i
\end{equation}
where $d_i$ is the eigenvalue corresponding to
the eigenvectors $\varepsilon_i$ of a particular polarization matrix
$M_i=\pmatrix{m_1&m_4\cr m_3&m_2}.$\\
The optical properties such as birefringence and dichroism of a homogeneous medium varies with distance.The passage of light through an optical element such as birefringent, absorbing or dichroic plate would be to change both $d_x$ and $d_y$ so that the effect may be represented by $|\psi_f>=M|\psi_i>$.
For a non-absorbing plate, there is no change in the intensity and the polarization matrix $M$ is therefore unitary $detM=1$ which makes $|\psi_f|=|\psi_i|$.
and can be studied by the differential matrix $N$ which refers to the optical element for a given infinitesimal path length within the element.For particular wavelength and direction,the evolution of a light vector $\varepsilon$ becomes
\begin{equation}
\frac{d\varepsilon}{dz}=\frac{dM}{dz}\varepsilon_0
=\frac{dM}{dz}M^{-1}\varepsilon=N\varepsilon
\end{equation}
where it is evident that $N$ is the operator that determines $dM/dz$ from $M$ as follows
\begin{equation}
N=\frac{dM}{dz}M^{-1}=\pmatrix{n_1 & n_2 \cr n_3 & n_4}
\end{equation}
When $N$ is independent of z, then on integration the dependence of polarization matrix M on z is seen from
\begin{equation}
M=M_0\exp(\int{Ndz})
\end{equation}
Jones had shown further that the eigenvectors of M are equal to that of N when it is independent of z \cite{paper8}.
Any homogeneous crystal without optical activity could be considered for normal incidence as a laminated crystal.According to the lamellar representation suggested by Jones \cite{paper7},a thin slab of a given medium is equivalent to a pile of retardation plates and partial polarizers.Eight constants are required to specify the real and imaginary parts of the four matrix elements of $2\times2$ N matrix, each possessing one and only one of the eight fundamental properties.The eight optical properties are paired \cite{paper9} and reduce to four.\\
i) Isotropic refraction and absorption\\
ii) Linear birefringence and linear dichroism along the xy coordinate axis.\\
iii) Linear birefringence and linear dichroism along the bisector of xy coordinate axes.\\
iv) Circular birefringence and circular dichroism.\\
The optical medium that has circular birefringence and linear birefringence will be our point of interest, and could have the following matrix form
\begin{eqnarray}
\theta_{cb}= \eta \pmatrix{0 & -1 \cr 1 & 0} \\
\theta_{lb}= \rho \pmatrix{0 & -i \cr i & 0}
\end{eqnarray}
These $\theta_{cb}$ and $\theta_{lb}$ matrices form the required differential matrix.
\begin{eqnarray}
N=\theta_{cb}+\theta_{lb}
=\pmatrix{0 & -\eta+i\rho \cr \eta+i\rho& 0}=\pmatrix{0 & n_2 \cr n_3 & 0}
\end{eqnarray}
where $\eta$ is the circular birefringence that measures the rotation of the plane polarized light per unit thickness and
$\rho$ is the part of linear birefringence that measures the difference between the two principal constants along the coordinate axes.
A crystal will be considered homogeneous \cite{paper18}, if any one of the optical property will be visible instead of eight (as if all the eight lamina are sandwiched).The evolution of the ray vector $\varepsilon={\varepsilon_1 \choose \varepsilon_2}$ when passes through such medium $N$,as in eq.(3) could be re-written into the components as
\begin{eqnarray}
\frac{d\varepsilon_1}{dz}=n_{1}\varepsilon_1 + n_{2}\varepsilon_2\\
\frac{d\varepsilon_2}{dz}=n_{3}\varepsilon_1 + n_{4}\varepsilon_2
\end{eqnarray}
For pure birefringent medium represented by eq.(8), one may use the evolution of ray vector
\begin{equation}
\frac{d\varepsilon_1}{dz}=n_2 \varepsilon_2,\\
\frac{d\varepsilon_2}{dz}=n_3 \varepsilon_1
\end{equation}
This shows that as the light enters into the birefringent plate,the spatial variation of component of electric vector in one direction gives the effect in the other perpendicular direction. It means that there is an exchange of optical power between the two component states of the polarized light indicating the rotation of the ray vector after entering the medium.
Geometrically this state $\varepsilon$ is a point P on the surface of the Poincare sphere that defines a position vector $\vec{p}$ in three dimensional space. The evolution of the vector $\vec{p}$ is equivalent to the cyclic change of the state vector during the passage of infinitesimal distance dz of the optical medium. Huard pointed out in his book \cite{paper18} that the spatial change of vector as it passes through the crystal becomes
\begin{equation}
\frac{d\vec{p}}{dz}=\vec{\Omega} \times \vec{p}
\end{equation}
This shows a natural twist by the instantaneous rotation vector $\Omega$ along the axis oz about which the $\vec{p}$ makes an elementary angle $d\alpha=\Omega dz$, for thickness $dz$. The magnitude and direction of the rotation vector depends on the inherent property of the medium, in other words on the element of the N matrix.
(i) {\bf The uniformly twisted crystal}\\
When an originally homogeneous crystal twisted uniformly about an axis parallel to the direction of transmission, the $N$ (z dependence) matrices are transformed upon rotation where $k$ is the angular twist of the optical medium per unit thickness \cite{paper8}.
It could be noted here that this $k$ has similar definition of $\Omega$ as
in eq.(12), having the basic difference in their space of appearance.
In fact,$k$ is associated with the external rotation of the optical element
where as $\Omega$ is the instantaneous rotation of the polarized light incident on the optical element by an angle.
The differential matrix $N$ which corresponds to the twisted crystal is related with its untwisted position matrix $N_0$ by the following relation
\begin{equation}
N=S(kz)N_0 S(-kz)
\end{equation}
where $S$ is the rotation matrix having developed after uniform rotation about an angle $\theta=kz$.
Jones realized the simultaneous rotation of the twisted state $\varepsilon'$ in the opposite direction of N matrix,
\begin{equation}
\varepsilon'=S(-kz)\varepsilon
\end{equation}
so that it satisfies the following equation
\begin{equation}
\frac{d\varepsilon'}{dz}=N'\varepsilon'
\end{equation}
with the dependence on the twisted matrix $N'$ that has independence on $z$.
It has been shown \cite{paper7} after few steps that this twisted matrix $N'$ can be expressed
\begin{equation}
N^\prime=N_0 - kS(\pi/2)
\end{equation}
in terms of $N_0=$ the matrix for untwisted crystal and $S(\pi/2)$ denotes the rotation matrix for normal incidence of light.
The solution of the above eq.(15) may be written $\varepsilon'=\exp(N'z)\varepsilon'_0$
where $\varepsilon'_0$ is the value of the vector $\varepsilon'$ at $z=0$.
(ii) {\bf The arbitrarily twisted crystal}\\
The transformation illustrated above may also be applied with an arbitrarily twisted crystal.Jones showed \cite{paper7} that the $N$ matrix of the twisted crystal is
\begin{equation}
N=S(\omega(z))N_0 S(\omega(z))
\end{equation}
where $\omega(z)$ specifies the angle of twist, that is the arbitrary function of z.
From the similar transformation of the uniformly twisted crystal, one finds
\begin{equation}
N'=N_0-(\frac{d\omega(z)}{dz})S(\pi/2)
\end{equation}
It has been pointed out by Jones that the derivative $\frac{d\omega(z)}{dz}$ is a constant for a uniformly twisted crystal.
Due to the inherent property of the birefringent optical medium (N)
a natural twist is realized by the incident polarized light. Further, external twist of the medium might develop an additional phase in fixed OAM sphere.
We realize this phase in the next section as OAM holonomy which will visualize the circular dichroism of the medium.
\section{The geometric phase by twisting homogeneous medium}
The propagation of light through an optical system was studied by Jones \cite{paper7}
in a $2x2$ matrix method.The method was based on the idea that in anisotropic media,the displacement vector D can be represented by a column vector.On similar manner Berry \cite{paper2} pointed out that a monochromatic light traveling in the z direction,the polarization state of the electric displacement vector lying in the xy plane can be written by a two component spinor
$$|\psi>={\psi_+ \choose \psi_-}$$
where $\psi_{\pm}=(d_x \pm i d_y)$ and the intensity is $I=|d_x|^2+|d_y|^2$ while the complex ratio $d_x/d_y$ defines its polarization state.
The polarization matrix $M$ satisfying $M|\psi>=1/2 |\psi>$ can be determined from the eigenvector $|\psi>$ with eigenvalue $+1/2$ using the relation $(|\psi><\psi|-1/2)$.If we consider the eigenvector by \cite{paper10}
\begin{equation}
|\psi>={\cos\theta/2 e^{i\phi} \choose \sin\theta/2}
\end{equation}
the polarization matrix in terms of polar angles $\theta$ and $\phi$ represent different point on the surface of the Poincare sphere,
\begin{equation}
\begin{array}{lcl}
M=1/2 \pmatrix{\cos\theta & \sin\theta e^{i\phi} \cr \sin\theta e^{-i\phi} & -\cos\theta \cr}
\end{array}
\end{equation}
having eigenvalues $\pm 1/2$.This eigenvalues reflect the helicity of polarized photon that has been used \cite{paper11}\cite{paper12} to evaluate the respective GP.
Here the parameters $\theta$ and $\phi$ are the coordinates of polarized light that represent a point on the Poincare sphere.
In search of the birefringent plate at a particular position $z$ the differential matrix $N$ of optical medium has been calculated from the polarization matrix $M$ as in eq.(20). Since the eigenvalues of helicities for polarized photons is
$\pm1$, we omitted the factor $1/2$ from the polarization matrix $M$ \cite{paper11} in the following equations.
\begin{equation}
N=(\frac{dM}{d\theta})(\frac{d\theta}{dz})M^{-1}
\end{equation}
Considering the angle of twist being proportional directly to the thickness z of the optical medium, $\theta=kz$ where $k$ represent the angular twist per unit thickness,
the corresponding differential matrix N \cite{paper11} becomes
\begin{equation}
N=k \pmatrix{0 & -e^{i\phi} \cr e^{-i\phi} & 0 \cr}
\end{equation}
having complex eigenvalues $\pm ik$ with eigenvectors
\begin{equation}
\left(
\begin{array}{c}
\pm i e^{i \phi } \\
1
\end{array}
\right)
\end{equation}
The nature of the optical medium could be identified
comparing the above N matrix in eq.(22) with eq.(8).It is seen that our N matrix is homogenous and possesses both circular and linear birefringence by $k\cos\phi$ and $(-k\sin\phi)$ respectively.It may be noted that in case one observes the eigenvalues of N opposite and imaginary,the optical medium possesses the property of purely circular birefringence \cite{paper7}.
Initially we consider the angular twist per unit thickness zero, $k=0$ (for $\theta=0$), then for a uniformly twisted optical medium, using eq.(16),
the twisted matrix becomes
\begin{equation}
N'=\pmatrix{0 & k \cr -k & 0 \cr}
\end{equation}
where the untwisted matrix $N_0=0$ and the rotation matrix
$S(\pi/2)=\pmatrix{0 & -1 \cr 1 & 0 \cr}$. The twisted ray will be obtained after the opposite rotation of the twisted matrix from eq.(14)
\begin{equation}
\varepsilon'=\pmatrix{\cos\theta & \sin\theta \cr -\sin\theta & \cos\theta \cr}
\left(
\begin{array}{c}
i e^{i \phi }\\
1
\end{array}
\right)
\end{equation}
in other words
\begin{equation}
\varepsilon'=
\left(
\begin{array}{c}
ie^{i\phi}\cos\theta+\sin\theta \\
-ie^{i\phi}\sin\theta+\cos\theta
\end{array}
\right)
\end{equation}.
The initial eigenvector $\varepsilon$ reappears if opposite rotation is given to $\varepsilon'$. Now considering the matrix in eq.(24) as initial matrix $N_0$,
by using $k=1$ which results the twisted matrix $N'$
\begin{equation}
N'= \pmatrix {0 & -e^{i\phi}+k \cr e^{-i\phi}-k & 0 \cr}
\end{equation}
It can be realized looking at eqs.(22),(24) and (27) that $k$, the angular twist per unit thickness in the birefringent optical medium plays the similar role as $\Omega$ for twisting a ray vector as in eq.(12).
Light having fixed polarization and helicity, if suffers the slow variation of path in real space it can be mapped on to the surface of unit sphere in the wave vector space.
The geometric phase is found to appear as the initial state $|A>$ unite with final $|A'>$.
\begin{equation}
<A|A'>=\pm \exp(i\Omega(C)/2)
\end{equation}
where $\Omega$ is the solid angle swept out by $e_k$ on its unit sphere.
Our present work is based on the consideration of polarized light passing normally
through a medium N having linear and circular birefringence. Due to the inherent property of the medium, the incident polarized light suffers a natural twist about the axis parallel to the direction of its propagation. We assume $\Upsilon$ and $\Upsilon'$ are the respective phases developed as the respective initial states $|A>=\varepsilon$ or $\varepsilon'$, are passed through the respective differential matrices $N$ or $N'$
\begin{eqnarray}
\Upsilon=\varepsilon^* \frac{d\varepsilon}{d\theta}\frac{d\theta}{dz}={\varepsilon^*}N\varepsilon\\
\Upsilon'={\varepsilon'}^* \frac{d\varepsilon'}{d\theta}\frac{d\theta}{dz}={\varepsilon'}^* N'{\varepsilon'}
\end{eqnarray}
with the consideration of $d\theta=d\theta'$.
$\Upsilon$ can be obtained,using eq.(22) and (23) and (11) in (29) in the following equation
\begin{eqnarray}
\Upsilon={\varepsilon^*}N\varepsilon
&=&{\varepsilon_1}^*n_2\varepsilon_2+{\varepsilon_2}^*n_3\varepsilon_1\nonumber\\
&=&(-ie^{-i\phi})(-ke^{i\phi})+(ke^{-i\phi})(ie^{i\phi})\nonumber\\
&=&2ik
\end{eqnarray}
Hence the phase of untwisted medium becomes $2ik$. By varying the value of $k=0,1,2..etc$, different $\Upsilon$ can be obtained.
In a similar way for twisted medium, we use the twisted matrix $N'$ eq.(24) for $k=0$ and the initial state by eq.(26) to evaluate $\Upsilon'$.
\begin{eqnarray}
\Upsilon'&=&{\varepsilon'}^*N'{\varepsilon'}
={{\varepsilon'}_1}^*{n'}_2{\varepsilon'}_2+{\varepsilon_2}^* {n'}_3 {\varepsilon'}_1\nonumber\\
&=&(-i e^{-i\phi}\cos\theta+\sin\theta)k(-ie^{i\phi}\sin\theta+\cos\theta) \nonumber\\
&+&(ie^{-i\phi}\sin\theta+\cos\theta)(-k)(ie^{i\phi}\cos\theta+\sin\theta)\nonumber\\
&=& -2ik\cos\phi
\end{eqnarray}
Any ray passing through $N$ or $N'$ will suffer a twist due to the internal dynamics of the birefringent medium. The polarized light passing through the twisted medium $N'$ will acquire the phase $\Upsilon'$,that has two parts acquired one from the dynamics and another parametric change of the medium.Thus the phase $\Upsilon'$ will contain both the dynamical and geometric phase of a uniformly twisted birefringent medium.
To grasp the geometric phase of the incident polarized state due to external twist of a birefringent medium in an initial adjustment $k=0$, we
take the difference between the two phases $\Upsilon'-\Upsilon$ that will eliminate the dynamical phase, the phase due to natural twist, resulting the outcome of geometric phase for normal incidence of the polarized light on the medium.
As a result the required geometric phase becomes
\begin{equation}
\Upsilon'-\Upsilon=-2ik\cos\phi
\end{equation}
where it is seen $\Upsilon=0$ for $k=0$.
On similar manner the required geometric phase of the other conjugate eigenvector will be
\begin{equation}
\Upsilon'-\Upsilon=2ik\cos\phi
\end{equation}
Thus it is seen the geometric phase visualize the circular birefringence of the medium by $k\cos\phi$.
If the orientation of the matrix N is changed considering $k=1$, then using the initial matrix\\
$N_0=\pmatrix{0 & -e^{i\phi} \cr e^{-i\phi} & 0 \cr}$
one can find the phase $\Upsilon=\pm 2i$ for the respective eigenvectors in eq.(23).
To calculate the corresponding twisted phase $\Upsilon'$, the twisted matrix $N'$ in eq.(27) is used that intuitively will act on the twisted light ray $\varepsilon'$ in opposite direction for making $k=1$.As result the original eigenvector $\varepsilon=\varepsilon"=S(\theta)\varepsilon'$ is visible.
In the language of mathematics the twisted GP becomes
\begin{eqnarray}
\Upsilon'&=&{{\varepsilon}_1}^*{n'}_2{\varepsilon}_2+{\varepsilon_2}^* {n'}_3 {\varepsilon}_1\nonumber\\
&=&(-i e^{-i\phi}, 1)\pmatrix {0 & -e^{i\phi}+k \cr e^{-i\phi}-k & 0 \cr}\\
\left(\begin{array}{c}
i e^{i \phi }\\
1
\end{array}
\right)
&=& 2i[1-k\cos\phi]
\end{eqnarray}
This helps us to recover again the previous form of the geometric phase
\begin{equation}
\Upsilon'-\Upsilon =2i[1-k\cos\phi]-2i= -2i(k\cos\phi)
\end{equation}
Hence again the GP in terms $k\cos\phi$ is visible where the non-zero $k$ is associated with the particular twist of the crystal. We choose the light incident at a particular angle $\theta$ on the optical medium. Now for the first choice $k=0$, the twisted matrix $N'$ gives the geometric phase which is identical with the findings for the second choice $k=1$.The geometric phase visualizes in both cases the circular birefringence of the medium.
Without considering the double rotation of light vector from $\varepsilon\longrightarrow\varepsilon'\longrightarrow\varepsilon"$,if we consider only the interaction of light ray $\varepsilon'$ in eq.(26) with the optical medium $N'$ in eq.(27) the outcome of the calculation is
\begin{eqnarray}
&&{{\varepsilon'}_1}^*{n'}_2{\varepsilon'}_2+{\varepsilon_2}^* {n'}_3 {\varepsilon'}_1\nonumber\\
&=&(-i e^{-i\phi}\cos\theta+\sin\theta)(k-e^{i\phi})(-ie^{i\phi}\sin\theta+\cos\theta)\nonumber \\
&+&(ie^{-i\phi}\sin\theta+\cos\theta)(e^{-i\phi}-k)(ie^{i\phi}\cos\theta+\sin\theta)\nonumber\\
&=& i[2-2ik\cos\phi-(1-\cos2\theta)(1-\cos2\phi)]
\end{eqnarray}
As a result one can finally find the phase $\Upsilon'$ for the initial adjustment of the optical medium at $k=1$
\begin{equation}
\Upsilon'= -i[2-2ik\cos\phi-(1-\cos2\theta)(1-\cos2\phi)]
\end{equation}
Thus for the specific twist of the medium for $k=1$, the geometric phase will be obtained from the difference of (39) and (36).
\begin{equation}
\Upsilon'-\Upsilon = 2i(k\cos\phi) + (1-\cos2\theta)(1-\cos 2\phi)]
\end{equation}
Hence it is seen that the GP appeared here eq.(40) is different from eq (37) though the twisted optical medium is same. The very cause of this difference is to consider the proper twist of the light ray passing though the medium.In all cases,the external twist of the system visualize the circular birefringence of the medium in terms of $k\cos\phi$.
It is seen from the above that the two geometric phases are in the form of the usual solid angle visualizing the circular birefringence
of the medium. If the optical medium is twisted arbitrarily, $k=d\theta/dz \neq constant$, we realize that the nature of the phase will
depend not only on the above solid angle but also on the nature of k. If $k=$periodic, a precessional type motion may be realized that could be studied further extensively.
\section{Angular momentum interpretation of polarization matrix and geometric phase}
The locus of the electric vector of polarized light when traces a circle, the light is known as circularly polarized light. When light is circularly polarized the two senses of rotation identifies two spin angular momentum (SAM) of photon given by $\sigma \hbar=\pm 1$ corresponding to left and right-handed circular polarization respectively.The photons having two spin angular momentum (SAM), can be visualized in another way \cite{paper11} by the two opposite directions of helicities.The conventional Poincare sphere is a SAM space where three kinds of polarized states are defined and where polarization changes from point to point.
Apart from the SAM, photons can also carry orbital angular momentum (OAM) arising from the inclination of the phase fronts with respect to beam's propagation axis.An important advancement was to realize the connection between topological charge and the orbital angular momentum of single photon by Allen et.al.\cite{paper19} for Laguerre-Gaussian (LG) beams with azimuthal phase $\exp(il\phi)$ for OAM $l$ per photon.For every value of $l$, there will be values of m from $-l$ to $+l$ and there shall be an infinite number of eigen states of $l$. Recently the orbital Poincare sphere has been sketched by Galves et.al. \cite{paper14}.All these helps to realize one that there will be two distinct representation of Poincare sphere by OAM and SAM respectively.
In our previous work \cite{paper11}, it has been suggested that with the change of polarization of light by some optical element,the parameter in connection with helicity changes.The behavior of chiral photon with a fixed helicity $(\pm1)$ in the polarized light is similar to a massless fermion having helicity $(\pm 1/2)$. These eigenvalues $(\pm 1/2)$ reflect the helicity of polarized photon. According to Berry \cite{paper3} photons have no magnetic moment, so it cannot be turned with a magnetic field over a closed path, but have the property of helicity to use.The above discussions help us \cite{paper10}\cite{paper11} to specify three variables $\theta,\phi$ and $\chi$ to parameterize a polarized photon. We would like to mention further that comprising these three parameters an extended Poincare sphere representation could be given considering the fixed spin vector attached at each point. In spherical geometry these $\phi$ and $\chi$ identify the operators of OAM and SAM by $i\hbar\frac{\partial}{\partial\phi}$ and $i\hbar\frac{\partial}{\partial\chi}$ respectively. The quantities $m$ and $\mu$ just represent the eigenvalues of the OAM and SAM operators.The Poincare sphere parameterized by $\theta$ and $\chi$ may be identified as the SAM sphere \cite{paper12} and comprising $\theta$ and $\phi$ parameters it is OAM sphere.
In our first work \cite{paper11},with the spinorial representation of polarized photon by spherical harmonics,the polarization matrix becomes
$$M\simeq \pmatrix{{Y_1}^0 & {Y_1}^1 \cr {Y_1}^-1 &{Y_1}^0\cr}$$
represented by each point on the Poincare sphere of OAM, parameterized by
the angles $\theta$ and $\phi$.
The elements of the polarization matrix ($2\times2$) are the product harmonics ${Y_1}^1$, ${Y_1}^{-1}$ and ${Y_1}^0$ for orbital angular momentum $l=1$.
This indicates that for higher OAM states $l=2,3..$ the respective product harmonics ${Y_l}^m$ could be used.On the other hand, if we define our polarization matrices $M,N$ on the Poincare sphere parameterized by $\theta$ and$\chi$ \cite{paper12} it is a point on the SAM sphere.
In view of this here we may pointed out that the birefringent plate represented by the matrix $N=k \pmatrix{0 & -e^{i\phi} \cr e^{-i\phi} & 0 \cr}$ lye on the OAM sphere for $l=1$. Both circular and linear birefringence,could be measured by the angular momentum of polarized photon in terms of angle $\phi$.The inherent property of the medium gives a natural twist to the incident polarized light by an angle $\Omega$ as seen in eq. (12).
The uniform twist (small) of the optical medium per unit thickness is $k$, whose value has chosen $k=0,1$ to represent the external twist of the birefringent optical medium. Intuitively $k$ behaves as $m$ because it is the eigenvalues of $N$. Hence we may consider it as the topological charge of the birefringent medium.
Geometric phase would be developed from the transfer of angular momentum in the course of variation of polarization or direction of propagation over a closed path.
The type of geometric phase acquired by the polarized photon depends only on the sphere/space where the path is traced out. The GP of Pancharatnam (PP)
developed by cyclic change of polarizations will be helicity dependent visualized by the parameter $\chi$ \cite{paper10} \cite{paper12} on the SAM sphere.In the OAM sphere the GP is arising for mode transformations \cite{paper6}.
Here we have studied the GP in the OAM sphere with use of birefringent medium.
Sanatamato \cite{paper20} pointed out that a birefringent plates affects the SAM and as well as OAM through some topological charge.He mentioned that a QP is an ordinary birefringent plate rotated at an angle $\alpha$ about the beam z-axis, with $\alpha$ given by $\alpha = \alpha(x, y)= arctan(y/x) = \phi$ where $\phi$ is the azimuthal angle in the x, y-plane.If the topological charge of the birefringent plate is q \cite{paper13}, the OAM of a light beam passing through such a "q-plate" (QP) changes by $\pm2q$ per photon.
In view of this we may comment lastly that the birefringent plate represented by the matrix $N$ here, may be identified as $k-$ plate having topological charge $k$. Independent of the choice of the initial medium,we find here that a GP depends on the final topological charge $k$ of the twisted medium. The novelty of this work also lies in the appearance of the circular birefringence $k\cos\phi$ of the optical medium through the geometric phase of the polarized light.
{\bf Acknowledgement} This work has been fully supported by the Abdus Salam International Centre for Theoretical Physics (ICTP), Trietse, Italy. Correspondence with Prof.Santamato,Napoli,Italy is gratefully acknowledged. Moreover,I am thankful for the help from Mr.Grassberger of ICTS, ICTP.
|
\section{Introduction}
The framework of Connes' noncommutative geometry \cite{connes_NCG} provides a generalization of ordinary Riemannian geometry. Within this framework, the notion of an almost commutative manifold (or an almost commutative geometry \cite{Class_IrrACG_I}) can be used to describe gauge field theories on Riemannian spin manifolds. Following a series of articles starting with \cite{CL91,CC96,CC97} this led in \cite{CCM07} to a noncommutative geometrical description of the full Standard Model of high energy physics, including the Higgs mechanism and neutrino mixing. In fact, a non-abelian $SU(N)$-Yang--Mills gauge theory can be described simply by considering matrix-valued functions on a background Riemannian manifold $M$. A key role is played by the adjoint action of the group of unitary matrices on $M$: it acts as $PSU(N)$ for $N$ the rank of the matrices.
This approach immediately raises a problem if one wishes to describe abelian gauge theories, since $PSU(N)$ is trivial if $N=1$. In fact, it was long believed to be impossible to describe abelian gauge theories within the framework of noncommutative geometry. In this paper, we show that it is very well possible, and we construct a spectral triple ({\it i.e.}\ a noncommutative manifold) that describes a $U(1)$-gauge theory and even the full theory of electrodynamics.
In \cite[Sect.~9.3]{landi} it is claimed that for commutative algebras the gauge fields (and hence the gauge group) are trivial.
The proof is based on the claim that the left and right action appearing in the adjoint action can be identified for a commutative algebra. Though this claim holds in the case of the canonical triple describing a Riemannian spin manifold, it need not be true for arbitrary commutative algebras. The almost commutative manifold given in \cref{sec:U1} below provides a counter-example.
This paper is organized as follows. We start by reviewing some definitions and results from noncommutative geometry, specializing to the case of almost commutative manifolds. We pay special attention to the form of the gauge group for such manifolds.
Then, in \cref{sec:U1} we consider the product of spacetime with a two-point space, however, from a noncommutative point of view, tracing back to the early noncommutative models \cite{CL91}.
Essentially, the Riemannian geometry of the product is the usual (commutative) one, but the spin (KO) dimension is different, very similar to \cite{CCM07}.
In \cref{sec:ED} we will show how the above example can be modified to provide a description of one of the simplest examples of a gauge theory in physics, namely electrodynamics. Because of its simplicity, it helps in gaining an understanding of the formulation of gauge theories in terms of almost commutative manifolds, and it provides a first step towards an understanding of the derivation of the Standard Model from noncommutative geometry \cite{CCM07}.
\section{Spectral triples and gauge symmetry}
\subsection{Spectral triples}
\label{sec:triple}
In this section we shall briefly recall the notion of spectral triples. We shall follow the definitions as they appear in \cite[Ch.~1, \S10]{connes-marcolli}, for more details we refer to that book and the references therein.
A {\it spectral triple} $(\mathcal{A}, \mathcal{H}, D)$ is given by a unital
$\ast$-algebra $\mathcal{A}$ represented faithfully as bounded operators on a Hilbert space $\mathcal{H}$ and a self-adjoint operator $D$ (referred to as a {\it Dirac operator}) with compact resolvent and such that all commutators $[D,a]$ are bounded for $a\in\mathcal{A}$. Note that this implies that the $\mathcal{A}$-module generated by operators of the form $a[D,b]$ ($a,b \in \mathcal{A}$) consists of bounded operators on $\mathcal{H}$. These differential one-forms will play a key role later, as they will appear as gauge fields. We set accordingly:
\begin{align*}
\Omega^1_D(\mathcal{A}) := \Big\{ \sum_j a_j[D,b_j] \mid a_j, b_j\in\mathcal{A} \Big\} .
\end{align*}
A spectral triple might have additional structure such as a $\mathbb{Z}_2$-grading $\gamma$ on $\mathcal{H}$, making $\mathcal{A}$ even, and $\Omega^1_D(\mathcal{A})$ odd. Correspondingly, the Hilbert space decomposes as $\mathcal{H} = \mathcal{H}^+ \oplus \mathcal{H}^-$ into the $\pm 1$ eigenspaces of $\gamma$. In this case, we will call the spectral triple {\it even}, otherwise it is {\it odd}.
Furthermore, an (even) spectral triple has a \emph{real structure} if there is an anti-linear isomorphism $J\colon \mathcal{H}\rightarrow\mathcal{H}$ with $J^2 = \varepsilon$, $JD = \varepsilon' DJ$ and, if the spectral triple is even, $J\gamma = \varepsilon'' \gamma J$. The signs $\varepsilon$, $\varepsilon'$ and $\varepsilon''$ determine the \emph{KO-dimension} $n$ modulo $8$ of the spectral triple, according to
\begin{align*}
\begin{array}{c|cccccccc}
n & 0 & 1 & 2 & 3 & 4 & 5 & 6 & 7 \\
\hline
\varepsilon & 1 & 1 & -1 & -1 & -1 & -1 & 1 & 1 \\
\varepsilon' & 1 & -1 & 1 & 1 & 1 & -1 & 1 & 1 \\
\varepsilon'' & 1 & & -1 & & 1 & & -1 & \\
\end{array}
\end{align*}
Moreover, the action of $\mathcal{A}$ is required to satisfy the commutation rule
\begin{align}
\label{eq:order0}
[a,b^0] = 0 \qquad\forall a,b\in\mathcal{A} ,
\end{align}
where we have defined the right action $b^0$ of $b$ on $\mathcal{H}$ by
\begin{align*}
b^0 := Jb^*J^{-1} .
\end{align*}
We also require such a commutation relation for $\Omega^1_D(\mathcal{A})$ with the right action of $\mathcal{A}$, {\it i.e.}
\begin{align}
\big[[D,a],b^0\big] = 0 \qquad\forall a,b\in\mathcal{A} .
\end{align}
\begin{example}
\label{ex:canon}
The motivating example for the definition of spectral triples is the \emph{canonical triple}. Let $M$ be a compact Riemannian spin manifold. We then define the canonical triple by
\begin{align*}
\big(\mathcal{A}, \mathcal{H}, D\big) = \big(C^\infty(M),L^2(M,S),\slashed{D}\big) ,
\end{align*}
where $S$ is the spinor bundle on $M$ and $\slashed{D}$ is the canonical Dirac operator given locally by $-i\gamma^\mu\nabla^S_\mu$. Here, $\nabla^S$ is the Levi--Civita connection lifted to the spinor bundle.
Due to the property $ [\slashed{D},a] = -i\, c(\mathrm d a)$, we can identify the differential one-forms $\Omega^1_\slashed{D}(C^\infty(M))$ with DeRham differential one-forms (via Clifford multiplication $c$).
If $M$ is even dimensional (say of dimension $m$), we have a $\mathbb{Z}_2$-grading $\gamma_{m+1}$
and an anti-linear isometry $J_M$, which is the charge conjugation operator on the spinors. The Riemannian spin manifold $M$ can be fully described by this canonical triple \cite{connes_foundation,C08}.
\end{example}
Another special case of a spectral triple is a real even finite spectral triple, given by the data
\begin{align*}
F := \left( \mathcal{A}_F, \mathcal{H}_F, D_F, \gamma_F, J_F \right) ,
\end{align*}
for a finite dimensional Hilbert space $\mathcal{H}_F$. The operators $D_F$, $\gamma_F$ and $J_F$, as well as the action of the algebra $\mathcal{A}_F$, are now simply given by matrices acting on $\mathcal{H}_F$, subject to the aforementioned (anti-)commutation relations. As a first result, we prove
\begin{lem}
\label{lem:class_real}
For any real even finite spectral triple $F$, we can write with respect to the decomposition $\mathcal{H} = \mathcal{H}^+ \oplus \mathcal{H}^-$:
\begin{align*}
\textnormal{KO-dimension }0\colon\quad J_F &= \mattwo{j_+}{0}{0}{j_-} C \quad \textnormal{for symmetric } j_\pm\in U(\mathcal{H}^\pm) ;\\
\textnormal{KO-dimension }2\colon\quad J_F &= \mattwo{0}{j}{-j^T}{0} C \quad \textnormal{for } jj^* = j^*j = \mathbb{I} ;\\
\textnormal{KO-dimension }4\colon\quad J_F &= \mattwo{j_+}{0}{0}{j_-} C \quad \textnormal{for anti-symmetric } j_\pm\in U(\mathcal{H}^\pm) ;\\
\textnormal{KO-dimension }6\colon\quad J_F &= \mattwo{0}{j}{j^T}{0} C \quad \textnormal{for } jj^* = j^*j = \mathbb{I} .
\end{align*}
\end{lem}
\begin{proof}
Let the operator $C$ denote complex conjugation. Then any anti-unitary operator $J_F$ can be written as $UC$, where $U$ is some unitary operator on $\mathcal{H}_F$. We then have $J_F^* = CU^* = U^TC$, and $J_FJ_F^* = UU^* = \mathbb{I}$.
The different possibilities for the choice of $J_F$ are characterized by the relations $J_F^2 = UCUC = U\bar U = \varepsilon$ and $J_F\gamma_F = \varepsilon''\gamma_FJ_F$. By inserting $\varepsilon,\varepsilon''=\pm1$ according to the KO-dimension, the exact form of $J_F$ can be directly computed by imposing these relations.
\end{proof}
We now combine the canonical triple for a spin manifold $M$ with the finite spectral triple $F$ to arrive at the noncommutative manifolds that are of particular interest in the context of particle physics.
\begin{defn}
A \emph{real even almost commutative (spin) manifold} $M \times F$ is described by
\begin{align*}
\left(\mathcal{A}, \mathcal{H}, D\right) := \left( C^\infty(M,\mathcal{A}_F), L^2(M,S)\otimes \mathcal{H}_F, \slashed{D}\otimes \mathbb{I} + \gamma_{m+1}\otimes D_F \right) ,
\end{align*}
together with a grading $\gamma = \gamma_{m+1}\otimes\gamma_F$ and a real structure $J = J_M \otimes J_F$.
\end{defn}
\subsection{The gauge group}
\label{sec:gauge_group}
We would like to study the notion of `symmetry' for almost commutative manifolds. The starting point is to define an equivalence of spectral triples. The symmetry is then revealed when it turns out that the bosonic and fermionic action functionals of a spectral triple are identical for equivalent spectral triples (see \cref{sec:act_funct}). We take our definition of equivalent spectral triples from \cite{C96} (cf.\ \cite[\S6.9]{landi}) but make a slight modification by incorporating the algebra isomorphism $\alpha$.
\begin{defn}
Two spectral triples $(\mathcal{A}_1, \mathcal{H}_1, D_1)$ and $(\mathcal{A}_2, \mathcal{H}_2, D_2)$, with the associated representations $\pi_j\colon\mathcal{A}_j\rightarrow B(\mathcal{H}_j)$ for $j=1,2$, are \emph{unitarily equivalent} if there exists a unitary operator $U\colon\mathcal{H}_1\rightarrow\mathcal{H}_2$, called the \emph{intertwining operator}, such that
$$
UD_1U^* = D_2; \qquad U\pi_1(a)U^* = \pi_2(\alpha(a)); \qquad (a \in \mathcal{A}),
$$
where $\alpha$ is an algebra isomorphism $\mathcal{A}_1\rightarrow\mathcal{A}_2$.
If the two triples are even with grading operators $\gamma_1$ and $\gamma_2$, one also requires that $U\gamma_1U^* = \gamma_2$. If the two triples are real with real structures $J_1$ and $J_2$, one also requires that $UJ_1U^* = J_2$.
\end{defn}
Note that for a discussion of the equivalence of spectral triples, it is good to explicitly mention the representation of the algebra on the Hilbert space, since the intertwining operator affects this representation. Let us now consider two basic examples of intertwining operators.
\begin{prop}
\label{prop:equiv_triples}
The following two spectral triples are equivalent to $(\mathcal{A}, \mathcal{H}, D, \gamma, J)$ with representation $\pi\colon\mathcal{A}\rightarrow B(\mathcal{H})$:
\begin{enumerate}
\item $(\mathcal{A}, \mathcal{H}, UDU^*, \gamma, UJU^*)$ with representation $\pi\circ\alpha_u$ for $U=\pi(u)$ with $u\in U(\mathcal{A})$;
\item $(\mathcal{A}, \mathcal{H}, UDU^*, \gamma, J)$ with representation $\pi\circ\alpha_u$ for $U=\pi(u)J\pi(u)J^*$ with $u\in U(\mathcal{A})$,
\end{enumerate}
where $\alpha_u$ is the inner automorphism of $\mathcal{A}$ given by $\alpha_u(a) := uau^*$.
\end{prop}
\begin{proof}
\begin{enumerate}
\item We only need to check that $U\pi(a)U^* = \pi\circ\alpha_u(a)$ and $U\gamma U^* = \gamma$. The latter relation is evident since the grading operator $\gamma$ commutes with the algebra. We also see that
\begin{align*}
U\pi(a)U^* = \pi(u)\pi(a)\pi(u)^* = \pi(uau^*) = \pi\circ\alpha_u(a) .
\end{align*}
\item First, we easily see from \eqref{eq:order0} that $U \equiv \pi(u) \pi(u^*)^0$ is a unitary operator.
The relation $U\gamma U^* = \gamma$ holds since $\pi(u)J\pi(u)J^*\gamma = (\varepsilon'')^2\gamma \pi(u)J\pi(u)J^*$. Since $\pi(u^*)^0$ commutes with $\pi(a)$, we find that
\begin{align*}
U\pi(a)U^* &=
\pi(u)\pi(a)\pi(u)^*
= \pi\circ\alpha_u(a) .
\end{align*}
Using the property $\pi(a)^0 J = J \pi(a^*)$ for all $a \in \mathcal{A}$, one can check that $UJU^* = J$.
\qedhere
\end{enumerate}
\end{proof}
In the first case of \cref{prop:equiv_triples}, the intertwining operator $U$ is given by left multiplication with an element of the unitary group $U(\mathcal{A})$. In the second case, the action of the operator $U$ on a vector $\xi\in\mathcal{H}$ can be written as $U\xi = u\xi u^*$, since we identify $JuJ^*$ with the right action of $u^*$. This case is especially interesting because we see that the intertwining operator has no effect on $J$. Thus, the group generated by all operators of the form $U=uJuJ^*$ characterizes all equivalent spectral triples $(\mathcal{A}, \mathcal{H}, UDU^*, \gamma, J)$, in which only the Dirac operator is affected by the unitary transformation.
\begin{defn}
\label{defn:gauge_group_NCG}
The \emph{gauge group} $\mathcal{G}(\mathcal{A})$ of a real spectral triple $(\mathcal{A}, \mathcal{H}, D, J)$ is defined by
\begin{align*}
\mathcal{G}(\mathcal{A}) := \left\{ U=uJuJ^* \mid u\in U(\mathcal{A}) \right\} .
\end{align*}
\end{defn}
Before we continue to evaluate the exact form of this gauge group, we first consider the following subalgebras of $\mathcal{A}$:
\begin{align}
\mathcal{A}_J &:= \big\{ a\in\mathcal{A} \mid aJ = Ja \big\}
,\nonumber \\
\label{eq:subalg2}
\til\mathcal{A}_J &:= \big\{ a\in\mathcal{A} \mid aJ = Ja^* \big\}
.
\end{align}
The definition of $\mathcal{A}_J$ is taken from \cite[Prop.~3.3]{CCM07} (cf.\ \cite[Prop.~1.125]{connes-marcolli}); it is a {\it real} commutative subalgebra in the center of $\mathcal{A}$. We have provided a similar but different definition for $\til\mathcal{A}_J$, since this subalgebra will turn out to be very useful for the description of the gauge group in \cref{prop:gauge_group_NCG}. Note that $aJ = Ja^*$ if and only if $a = a^0$, {\it i.e.} if and only if its left and right action on $\mathcal{H}$ coincide.
\begin{prop}
\label{prop:subalg}
For a complex algebra $\mathcal{A}$, the subalgebra $\til\mathcal{A}_J$ is an involutive commutative complex subalgebra of the center of $\mathcal{A}$.
\end{prop}
\begin{proof}
Since we must have $[a,b^0] = 0$ for any $a,b \in\mathcal{A}$, we have $[a,b] = 0$ for any $a\in\mathcal{A}$ and $b\in\til\mathcal{A}_J$, so $\til\mathcal{A}_J$ is contained in the center of $\mathcal{A}$. The requirement $a = a^0$ is complex linear, and also implies that $a^* = (a^0)^* = (a^*)^0$, so we have $a^*\in\til\mathcal{A}_J$ for $a\in\til\mathcal{A}_J$. Finally, we check that for $a,b\in\til\mathcal{A}_J$, we find $(ab)^0 = b^0a^0 = ba = ab$, so $ab\in\til\mathcal{A}_J$. \qedhere
\end{proof}
\begin{prop}
\label{prop:gauge_group_NCG}
There is a short exact sequence of groups
\begin{align*}
1\rightarrow U(\til\mathcal{A}_J)\rightarrow U(\mathcal{A})\rightarrow \mathcal{G}(\mathcal{A})\rightarrow1 ,
\end{align*}
where $\til\mathcal{A}_J$ is defined in \eqref{eq:subalg2}.
\end{prop}
\begin{proof}
The map $\Ad\colon U(\mathcal{A})\rightarrow \mathcal{G}(\mathcal{A})$ given by $u\mapsto u (u^*)^0$ is surjective by definition. The commutation relation \eqref{eq:order0} implies that $\Ad$ is a group homomorphism. Its kernel is given by elements $u \in U(\mathcal{A})$ for which $u (u^*)^0 = 1$. In other words, $\textnormal{Ker} \Ad$ consists of all unitary elements in $\tilde \mathcal{A}_J$.
\end{proof}
From \cref{prop:subalg} we know that $\til\mathcal{A}_J$ is a subalgebra of the center of $\mathcal{A}$. If we denote by $Z$ the subgroup of $U(\mathcal{A})$ that commutes with $\mathcal{A}$, then the group $U(\til\mathcal{A}_J)$ of the previous proposition is contained in $Z$. The quotient $U(\mathcal{A}) / Z$ yields the group $\textnormal{Inn}(\mathcal{A})$ of inner automorphisms of the algebra. Proposition \ref{prop:gauge_group_NCG} then implies that in general, the gauge group $\mathcal{G}(\mathcal{A})$ is larger than $\textnormal{Inn}(\mathcal{A})$. If $U(\til\mathcal{A}_J)$ is \emph{equal} to $Z$, we have in fact $\textnormal{Inn}(\mathcal{A}) \simeq \mathcal{G}(\mathcal{A})$.
\subsection{Inner fluctuations and gauge transformations}
In this section we will first define the inner fluctuations of a spectral triple. These inner fluctuations arise from considering Morita equivalences between algebras. For a detailed discussion, we refer to \cite{C96} or \cite[Ch.~1, \S10.8]{connes-marcolli}. In this section, we will simply give the definition.
Recall the Connes' differential one-forms $\Omega^1_D(\mathcal{A})$, spanned by operators of the form $a[D,b]$ (with $a,b \in \mathcal{A}$).
For a real spectral triple (endowed with a real structure $J$) we may replace $D$ by
\begin{align*}
D_A := D + A + \varepsilon'JAJ^{-1} ,
\end{align*}
for a selfadjoint $A=A^*\in\Omega^1_D$. The elements $A$ are called the \emph{inner fluctuations} of the spectral triple.
In Proposition \ref{prop:equiv_triples} we have seen that an element $U=uJuJ^*\in \mathcal{G}(\mathcal{A})$ transforms the Dirac operator as $D\rightarrow UDU^*$. Let us now consider the effect of this transformation on the fluctuated Dirac operator $D_A = D + A + \varepsilon'JAJ^*$. A direct calculation shows
that $D_A \mapsto UD_A U^*$ is equivalent to a transformation on $A$ of the form
\begin{align*}
A^u := uAu^* + u[D,u^*] \in \Omega^1_D .
\end{align*}
In other words, the transformation of a fluctuated Dirac operator can again be written in the form of a fluctuated Dirac operator. This only works because we have restricted $U(\mathcal{A})$ to the gauge group $\mathcal{G}(\mathcal{A})$, to make sure that the conjugation operator $J$ remains unchanged. The resulting transformation on the inner fluctuation $A\rightarrow A^u$ shall be interpreted in physics as the gauge transformation of the gauge field.
\subsection{The action functional}
\label{sec:act_funct}
In the previous sections, we have recalled spectral triples and their symmetries. It is now time to introduce interesting invariant functionals on them.
\begin{defn}[Chamseddine--Connes \cite{CC96}]
\label{defn:spectral_action}
Let $(\mathcal{A},\mathcal{H},D)$ be a spectral triple as above. The {\it spectral action} of a real spectral triple is defined by
\begin{align*}
S_b[A] := \textnormal{Tr} \left(f\Big(\frac{D_A}{\Lambda}\Big)\right) ,
\end{align*}
where $f$ is a positive even function, $\Lambda$ is a cut-off parameter and $D_A$ is the fluctuated Dirac operator.
\end{defn}
The spectral action describes only the action for the (bosonic) gauge fields. For the terms involving fermions and their coupling to the bosons, we need something extra. The precise form of the fermionic action depends on the KO-dimension of the spectral triple. For the purpose of this paper, we will only consider the case of KO-dimension $2$ and give the fermionic action for this case. Referring to the sign table in \cref{sec:triple}, we thus have the relations
\begin{align*}
J^2 &= -1 , & JD &= DJ , & J\gamma &= -\gamma J .
\end{align*}
We use the decomposition $\mathcal{H} = \mathcal{H}^+ \oplus \mathcal{H}^-$ by the grading $\gamma$. Following \cite{CCM07} (cf.\ \cite[Ch.~1, \S16.2-3]{connes-marcolli}) the relations above yield a natural construction of an anti-symmetric bilinear form on $\mathcal{H}^+$, given for $\xi,\xi'\in\mathcal{H}^+$ by
\begin{align*}
\mathfrak{A}_D(\xi,\xi') = \langle J\xi,D\xi'\rangle ,
\end{align*}
where $\langle\;,\;\rangle$ is the inner product on $\mathcal{H}$.
We define the set of \emph{classical fermions} corresponding to $\mathcal{H}^+$,
\begin{align*}
\mathcal{H}^+_\text{cl} := \{\til\xi \mid \xi\in\mathcal{H}^+\} ,
\end{align*}
as the set of Grassmann variables $\til\xi$ for $\xi\in\mathcal{H}^+$.
\begin{defn}
\label{defn:fermion_act}
For a real even spectral triple $(\mathcal{A}, \mathcal{H}, D, \gamma, J)$ of KO-dimension $2$ we define the full \emph{action functional} by
\begin{align*}
S[A,\xi] := S_b[A] + S_f[A,\xi] := \textnormal{Tr} \left(f\Big(\frac{D_A}{\Lambda}\Big)\right) + \frac12 \langle J\til\xi,D_A\til\xi\rangle ,
\end{align*}
for $\til\xi\in\mathcal{H}^+_\text{cl}$.
\end{defn}
The factor $1/2$ in front of the \emph{fermionic action} $S_f$ has been chosen for future convenience.
The action functional $S[A,\tilde\xi]$ is invariant under unitary transformations; in fact, it is invariant under transformations of the gauge group $\mathcal{G}(\mathcal{A})$.
Note that we have incorporated two restrictions in the fermionic action $S_f$. The first is that we restrict ourselves to even vectors in $\mathcal{H}^+$, instead of considering all vectors in $\mathcal{H}$. The second restriction is that we do not consider the inner product $\langle J\til\xi',D_A\til\xi\rangle$ for two independent vectors $\xi$ and $\xi'$, but instead use the same vector $\xi$ on both sides of the inner product. Each of these restrictions reduces the number of degrees of freedom in the fermionic action by a factor $2$, yielding a factor $4$ in total. It is precisely this approach that solves the problem of fermion doubling pointed out in \cite{lizzi} (see also the discussion in \cite[Ch.~1, \S16.3]{connes-marcolli}).
\subsubsection{The heat expansion}
For future purpose, let us recall some results on heat kernel expansions. For more details we refer the reader to \cite{Gil84}. Suppose we have a vector bundle $E$ on a compact Riemannian manifold $M$, and a second order differential operator $H\colon \Gamma(E)\rightarrow\Gamma(E)$ of the form $H = \Delta^E - Q$, where $\Delta^E$ is the Laplacian of some connection on $E$ and $Q\in\Gamma(\textnormal{End}(E))$.
For a generalized Laplacian $H$ on $E$ we have the following asymptotic expansion (as $t \to 0$), known as the \emph{heat expansion} \cite[\S1.7]{Gil84}:
\begin{align*}
\textnormal{Tr}\left(e^{-tH}\right) \sim \sum_{k\geq0} t^{\frac{k-m}2} a_k(H) ,
\end{align*}
where $m$ is the dimension of the manifold, the trace is taken over the Hilbert space $L^2(M,E)$ and the coefficients of the expansion are given by
\begin{align*}
a_k(H) := \int_M a_k(x,H) \sqrt{|g|} d^mx.
\end{align*}
The coefficients $a_k(x,H)$ are called the Seeley-DeWitt coefficients. We also state here without proof Theorem 4.8.16 from \cite{Gil84}. Note that the conventions used by \cite{Gil84} for the Riemannian curvature $R$ are such that $g^{\mu\rho}g^{\nu\sigma}R_{\mu\nu\rho\sigma}$ is negative for a sphere, in contrast to our own conventions. Therefore we have replaced $s=-R$. Furthermore, we have used that $f_{;\mu}^{\phantom{\mu};\mu} = -\Delta f$ for $f\in C^\infty(M)$.
\begin{thm}[\cite{Gil84}, Theorem 4.8.16]
\label{thm:seeley-dewitt}
For a generalized Laplacian $H = \Delta^E - Q$ the first three non-zero Seeley-DeWitt coefficients are given by
\begin{gather*}
a_0(x,H) = (4\pi)^{-\frac m2} \textnormal{Tr}(\textnormal{Id}); \qquad a_2(x,H) = (4\pi)^{-\frac m2} \textnormal{Tr}\left(\frac s6 + Q\right) \\
a_4(x,H) = (4\pi)^{-\frac m2} \frac1{360}
\textnormal{Tr}\big(-12 \Delta s
+ 5 s^2 - 2 R_{\mu\nu} R^{\mu\nu} + 2 R_{\mu\nu\rho\sigma} R^{\mu\nu\rho\sigma} \\
\hspace{5cm} + 60 sQ + 180 Q^2 - 60 \Delta Q + 30 \Omega^E_{\mu\nu} {\Omega^E}^{\mu\nu} \big) ,
\end{gather*}
where the traces are now taken over the fibre $E_x$. Here $s$ is the scalar curvature of the Levi-Civita connection $\nabla$, $\Delta$ is the scalar Laplacian and $\Omega^E$ is the curvature of the connection $\nabla^E$ corresponding to $\Delta^E$.
\end{thm}
Now, assume that the square of the fluctuated Dirac operator $D_A$ is of the form $\Delta^E - Q$ for some vector bundle $E$. Applying the heat expansion on ${D_A}^2$ then yields (as $t \to 0$):
\begin{equation*}
\textnormal{Tr}\left(e^{-t{D_A}^2}\right) \sim \sum_{k\geq0} t^{\frac{k-m}2} a_k({D_A}^2) ,
\end{equation*}
where the Seeley-DeWitt coefficients are given by \cref{thm:seeley-dewitt}. Then, on writing $f$ as a Laplace transform, we obtain in the case of a $4$-dimensional manifold asymptotically (as $\Lambda\to \infty$):
\begin{align}
\label{eq:action_expansion}
\textnormal{Tr} \left(f\Big(\frac {D_A}\Lambda\Big)\right) \sim 2 f_4\Lambda^4 a_0({D_A}^2) + 2 f_2\Lambda^2 a_2({D_A}^2) + a_4({D_A}^2) f(0) + O(\Lambda^{-1}),
\end{align}
where $f_j = \int_0^\infty f(v) v^{j-1} dv$ are the momenta of the function $f$ for $j>0$.
\begin{example}
\label{ex:canon_spec_act}
For the canonical triple of a $4$-dimensional spin manifold $M$, we obtain (see \cite[Theorem 1.158]{connes-marcolli
)
\begin{align*}
\textnormal{Tr} \left(f\Big(\frac \slashed{D}\Lambda\Big)\right) \sim \frac1{4\pi^2} \int_M \L_M(g_{\mu\nu}) \sqrt{|g|} d^4x + O(\Lambda^{-1}) ,
\end{align*}
where the gravitational Lagrangian $\L_M$ is given by
\begin{align*}
\L_M(g_{\mu\nu}) := 2f_4\Lambda^4 - \frac16 f_2\Lambda^2 s + f(0) \Big(\frac1{120} \Delta s -\frac1{80} C_{\mu\nu\rho\sigma} C^{\mu\nu\rho\sigma} + \frac{11}{1440}R^*R^* \Big) .
\end{align*}
The first two terms yield the Einstein-Hilbert action including a cosmological constant. In addition, we obtain a higher-order contribution from the Weyl gravity term $C_{\mu\nu\rho\sigma} C^{\mu\nu\rho\sigma}$, as well as a boundary term $\Delta s$ and a topological contribution from $R^*R^*$.
\end{example}
\section{The two-point space}
\label{sec:U1}
In this section we will provide a simple example of an almost commutative manifold, based on the product of a spin manifold $M$ with a two-point space $X$. The spectral triple describing this example will have a commutative algebra. As mentioned in the introduction, it has been claimed \cite[Sect.~9.3]{landi} that the inner fluctuation $A + JAJ^*$ vanishes for commutative algebras. The proof is based on the claim that the left and right action can be identified, i.e.\ $a=a^0$, for a commutative algebra. Though this claim holds in the case of the canonical triple describing a spin manifold, it need not be true for arbitrary commutative algebras. The spectral triple given in this section provides a counter-example.
What can be said for a commutative algebra, is that there exist no non-trivial inner automorphisms. It is thus an important insight here that the gauge group $\mathcal{G}(\mathcal{A})$, as defined in \cref{defn:gauge_group_NCG}, is larger than the group of inner automorphisms, so that a commutative spectral triple may still lead to a non-trivial gauge group. In fact, we will show that our example given below describes an abelian $U(1)$ gauge theory.
\subsection{A two-point space}
\label{sec:two-point}
We shall consider a finite spectral triple $F\Sub{X}$ corresponding to the two-point space $X = \{x,y\}$. A complex function on this space is simply determined by two complex numbers. The algebra of functions on $X$ is then given by $C(X) = \mathbb{C}^2$. Let us consider the \emph{even} finite spectral triple $F\Sub{X}$ given by
\begin{align*}
\left( C(X), \mathcal{H}_F, D_F, \gamma_F \right) .
\end{align*}
We require that the representation $C(X)\rightarrow B(\mathcal{H}_F)$ is faithful, which implies that the Hilbert space $\mathcal{H}_F$ must be at least $2$-dimensional. Thus, the simplest possible choice is to take $\mathcal{H}_F = \mathbb{C}^2$. We use the $\mathbb{Z}_2$-grading $\gamma_F$ to decompose $\mathcal{H}_F = \mathcal{H}_F^+ \oplus \mathcal{H}_F^- = \mathbb{C}\oplus\mathbb{C}$ into the two eigenspaces $\mathcal{H}_F^\pm = \{\psi\in\mathcal{H}_F \mid \gamma_F\psi = \pm\psi \}$. Hence, we can write
\begin{align*}
\gamma_F = \mattwo{1}{0}{0}{-1} .
\end{align*}
Because of the relations $[\gamma_F,a]=0$ and $D_F\gamma_F=-\gamma_FD_F$, the self-adjoint Dirac operator must be off-diagonal and the action of an element $a\in\mathcal{A}_F$ on $\psi\in\mathcal{H}_F$ can be written as
\begin{align}
\label{eq:rep_U1}
a\psi &= \mattwo{a_+}{0}{0}{a_-} \vectwo{\psi_+}{\psi_-} .
\end{align}
Thus, the even finite spectral triple $F\Sub{X}$ we will study in this section is given by
\begin{align}
\label{eq:triple_U1}
\left(\mathcal{A}_F, \mathcal{H}_F, D_F, \gamma_F \right) = \left( \mathbb{C}^2, \mathbb{C}^2, \mattwo{0}{t}{\bar t}{0}, \mattwo{1}{0}{0}{-1} \right) ,
\end{align}
for some complex parameter $t\in\mathbb{C}$, and with the representation of $\mathcal{A}_F$ on $\mathcal{H}_F$ given by \eqref{eq:rep_U1}.
\begin{prop}
\label{prop:zero_D_F}
The even finite spectral triple $F\Sub{X}$ given by \eqref{eq:triple_U1} can only have a real structure if $D_F = 0$.
\end{prop}
\proof
We must have ${J_F}^2 = \varepsilon$ and $J_F\gamma_F = \varepsilon''\gamma_FJ_F$, and we shall consider all possible (even) KO-dimensions separately. Thus, we apply Lemma \ref{lem:class_real} to the finite spectral triple $F\Sub{X}$ given above and, for each even KO-dimension, also impose the relations $[a,b^0]=0$ and $\big[[D_F,a],b^0\big]=0$. This gives:
\begin{description}
\item[{\it KO-dimension $0$}] \mbox{}\\
We have $J_F = \mattwo{j_+}{0}{0}{j_-} C$ for $j_\pm\in U(1)$. For $b=\mattwo{b_+}{0}{0}{b_-}$ we then obtain
$$
b^0 = \mattwo{j_+b_+\bar{j_+}}{0}{0}{j_-b_-\bar{j_-}} = b ,
$$
and see that this indeed commutes with the left action of $a\in\mathbb{C}^2$. Next, we check the order one condition
\begin{align*}
0 = \big[[D_F,a],b^0\big] &=
(a_+-a_-)(b_+-b_-)D_F .
\end{align*}
Since this must hold for all $a,b\in\mathbb{C}^2$, we conclude that we must require $D_F = 0$.
\item[{\it KO-dimension $2$}] \mbox{}\\
We have $J_F = \mattwo{0}{j}{-j}{0} C$ for $j\in U(1)$. We now obtain
$$
b^0 = \mattwo{jb_-\bar{j}}{0}{0}{jb_+\bar{j}} = \mattwo{b_-}{0}{0}{b_+} ,
$$
and see that this indeed commutes with the left action of $a\in\mathbb{C}^2$. Next, we check the order one condition
\begin{align*}
0 = \big[[D_F,a],b^0\big] &=
(a_+-a_-)(b_--b_+)D_F .
\end{align*}
Again we conclude that we must require $D_F = 0$.
\item[{\it KO-dimension $4$}] \mbox{}\\
We have $J_F$
of the same form as in KO-dimension 0, but now with $j_\pm = -j_\pm^T\in U(1)$. This implies that $j_\pm = 0$, so the given finite spectral triple cannot have a real structure in KO-dimension $4$.
\item[{\it KO-dimension $6$}] \mbox{}\\
We have $J_F = \mattwo{0}{j}{j}{0} C$ for $j\in U(1)$. We again obtain
$$
b^0 = \mattwo{jb_-\bar{j}}{0}{0}{jb_+\bar{j}} = \mattwo{b_-}{0}{0}{b_+} ,
$$
just as for KO-dimension $2$. Hence again the commutation rules are only satisfied for $D_F = 0$. \qedhere
\end{description}
\endproof
\subsection{The almost commutative manifold}
\label{sec:product_space_U1}
Let us now consider the product of the finite spectral triple $F\Sub{X}$ of the two-point space, as described by \eqref{eq:triple_U1}, with the canonical triple describing a compact Riemannian spin manifold $M$, as in \cref{ex:canon}. From here on we will take $M$ to be $4$-dimensional. Thus we consider the almost commutative manifold $M\times F\Sub{X}$ given by the data
\begin{align*}
M\times F\Sub{X} := \Big( C^\infty(M,\mathbb{C}^2), L^2(M,S)\otimes\mathbb{C}^2, \slashed{D}\otimes\mathbb{I}, \gamma_5\otimes\gamma_F, J_M\otimes J_F \Big) ,
\end{align*}
where we still need to make a choice for $J_F$. The algebra of this almost commutative manifold is given by $C^\infty(M,\mathbb{C}^2) \simeq C^\infty(M)\oplus C^\infty(M) \simeq C^\infty(M \times X)$. Thus, the underlying space $N := M\times X \simeq M \sqcup M$ consists of the disjoint union of two identical copies of the space $M$, and we can write $C^\infty(N) = C^\infty(M)\oplus C^\infty(M)$. We can also decompose the total Hilbert space as $\mathcal{H} = L^2(M,S) \oplus L^2(M,S)$. For $a,b\in C^\infty(M)$ and $\psi,\phi\in L^2(M,S)$, an element $(a,b)\in C^\infty(N)$ then simply acts on $(\psi,\phi)\in\mathcal{H}$ as $(a,b) (\psi,\phi) = (a\psi,b\phi)$.
\subsubsection{Distances}
To any spectral triple $(\mathcal{A},\mathcal{H},D)$ one can associate a distance function on the space of states on $\mathcal{A}$:
$$
d_D(p,q) = \sup \left\{ |p(a)- q(a) | \colon a\in\mathcal{A}, \|[D,a]\|\leq1 \right\} ,
$$
For the canonical triple, $\mathcal{A} = C^\infty(M)$ whose state space is homeomorphic to $M$. It turns out that $d_\slashed{D}$ is equal to the geodesic distance $d_g$ between points $p$ and $q$ on $M$.
We will use this formula as a generalized notion of distance, so on our finite spectral triple $F\Sub{X}$ we can write
\begin{align*}
d_{D_F}(x,y) = \sup \left\{ |a(x) - a(y)| \colon a\in\mathcal{A}_F, \|[D_F,a]\|\leq1 \right\} .
\end{align*}
Note that we now have only two distinct points $x$ and $y$ in the space $X$, and we shall calculate the distance between these points (cf.\ \cite[Sect.~6.8]{landi}). An element $a\in\mathbb{C}^2=C(X)$ is specified by two complex numbers $a(x)$ and $a(y)$
and its commutator with $D_F$ becomes
\begin{align*}
[D_F,a]
\big(a(y)-a(x)\big) \mattwo{0}{t}{-\bar t}{0} .
\end{align*}
The norm of this commutator is $|a(y)-a(x)|\,|t|$, so $\|[D_F,a]\| \leq 1$ if and only if $|a(y)-a(x)| \leq \frac{1}{|t|}$. We thus obtain that the distance between the two points $x$ and $y$ is given by
\begin{align*}
d_{D_F}(x,y) = \frac{1}{|t|} .
\end{align*}
If there is a real structure $J_F$, we have $t=0$ by \cref{prop:zero_D_F}, so then the distance between the two points becomes infinite.
Let $p$ be a point in $M$, and write $(p,x)$ or $(p,y)$ for the two corresponding points in $N=M\times X$. A function $a\in C^\infty(N)$ is then determined by two functions $a_x,a_y\in C^\infty(M)$, given by $a_x(p) := a(p,x)$ and $a_y(p) := a(p,y)$. Now consider the distance function on $N$ given by
\begin{align*}
d_{\slashed{D}\otimes\mathbb{I}}(n_1,n_2) = \sup \left\{ |a(n_1) - a(n_2)| \colon a\in\mathcal{A}, \|[\slashed{D}\otimes\mathbb{I},a]\|\leq1 \right\} .
\end{align*}
If $n_1$ and $n_2$ are points in the same copy of $M$, for instance if $n_1=(p,x)$ and $n_2=(q,x)$ for points $p,q\in M$, then their distance is determined by $|a_x(p) - a_x(q)|$, for functions $a_x\in C^\infty(M)$ for which $\|[\slashed{D},a_x]\|\leq1$. Thus, in this case we obtain that we recover the geodesic distance on $M$, {\it i.e.}\ $d_{\slashed{D}\otimes\mathbb{I}}(n_1,n_2) = d_g(p,q)$.
However, if $n_1$ and $n_2$ are points in a different copy of $M$, for instance if $n_1=(p,x)$ and $n_2=(q,y)$, then their distance is determined by $|a_x(p) - a_y(q)|$ for two functions $a_x,a_y\in C^\infty(M)$, such that $\|[\slashed{D},a_x]\|\leq1$ and $\|[\slashed{D},a_y]\|\leq1$. These latter requirements however yield no restriction on $|a_x(p) - a_y(q)|$, so in this case the distance between $n_1$ and $n_2$ is infinite. We thus find that the space $N$ is given by two disjoint copies of the Riemannian manifold $M$, which are separated by an infinite distance.
\subsection{\texorpdfstring{$U(1)$}{U(1)} gauge theory}
\label{sec:U1_gauge}
We would now like to derive the gauge theory that corresponds to the almost commutative manifold $M\times F\Sub{X}$. Recall that the gauge group $\mathcal{G}(\mathcal{A})$ is given by the quotient $U(\mathcal{A}) / U(\til\mathcal{A}_J)$, so if we wish to obtain a nontrivial gauge group, we need to choose $J$ such that $U(\til\mathcal{A}_J) \neq U(\mathcal{A})$. By looking at the form of $J_F$ for the different (even) KO-dimensions, as given in \cref{sec:two-point}, we conclude that $F\Sub{X}$ must have KO-dimension $2$ or $6$. In analogy with the noncommutative description of the Standard Model \cite{CCM07} we choose to work in KO-dimension 6. The almost commutative manifold $M\times F\Sub{X}$ then has KO-dimension $6+4\mod 8 = 2$. This means that we can use \cref{defn:fermion_act} to calculate the fermionic action. Therefore, we will consider the finite spectral triple $F\Sub{X}$ given by the data
\begin{align*}
F\Sub{X} := \left(\mathcal{A}_F, \mathcal{H}_F, D_F, \gamma_F, J_F \right) := \left( \mathbb{C}^2, \mathbb{C}^2, 0, \mattwo{1}{0}{0}{-1}, \mattwo{0}{C}{C}{0} \right) ,
\end{align*}
which define a real even finite spectral triple of KO-dimension $6$. Now, let us derive the gauge group.
\begin{prop}
\label{prop:gauge_group_U1}
The gauge group $\mathcal{G}(\mathcal{A}_F)$ of the two-point space is given by $U(1)$.
\end{prop}
\begin{proof}
First, note that $U(\mathcal{A}_F) = U(1) \times U(1)$. We will show that $U( (\til\mathcal{A}_F)_{J_F} ) \equiv U(\mathcal{A}_F) \cap (\til\mathcal{A}_F)_{J_F} \simeq U(1)$ so that the quotient $\mathcal{G}(\mathcal{A}_F) \simeq U(1)$ as claimed. Indeed, for $a \in \mathbb{C}^2$ to be in $(\til \mathcal{A}_F)_{J_F}$ it has to satisfy $J_F a^* J_F =a $. Since
$$
J_Fa^*J_F^* = \mattwo{0}{C}{C}{0} \mattwo{\bar a_1}{0}{0}{\bar a_2} \mattwo{0}{C}{C}{0} = \mattwo{a_2}{0}{0}{a_1} ,
$$
this is the case if and only if $a_1 = a_2$. Thus, $(\til \mathcal{A}_F)_{J_F} \simeq \mathbb{C}$ whose unitary elements form the group $U(1)$, contained in $U(\mathcal{A}_F)$ as the diagonal subgroup.
\end{proof}
We will now derive the gauge field for the almost commutative manifold $M\times F\Sub{X}$. Thus, we need to calculate the inner fluctuations of the Dirac operator. For $a,b\in C^\infty(M,\mathbb{C}^2)$, an inner fluctuation $A$ takes the form
\begin{align*}
A = a[D,b] = a[\slashed{D}\otimes\mathbb{I},b] = i\gamma^\mu \otimes a\partial_\mu b =: \gamma^\mu\otimes A_\mu ,
\end{align*}
where we have defined the hermitian field $A_\mu \in C^\infty(M,\mathbb{R}^2)$. Using the relation $J_M\gamma^\mu = -\gamma^\mu J_M$, the total inner fluctuation is then given by
\begin{align}
\label{eq:inn_fluc}
A + JAJ^* = \gamma^\mu \otimes (A_\mu - J_FA_\mu J_F^*) =: \gamma^\mu \otimes B_\mu .
\end{align}
An arbitrary hermitian field of the form $A_\mu=ia\partial_\mu b$ would be given by $\mattwo{X^1_\mu}{0}{0}{X^2_\mu}$, for two $U(1)$ gauge fields $X^1_\mu,X^2_\mu\in C^\infty(M,\mathbb{R})$. However, $A_\mu$ only appears in the combination
\begin{align*}
B_\mu = A_\mu - J_FA_\mu J_F^{-1} = \mattwo{X^1_\mu}{0}{0}{X^2_\mu} - \mattwo{X^2_\mu}{0}{0}{X^1_\mu} =: \mattwo{Y_\mu}{0}{0}{-Y_\mu} = Y_\mu\otimes\gamma_F ,
\end{align*}
where we have defined the $U(1)$ gauge field $Y_\mu := X^1_\mu-X^2_\mu \in C^\infty(M,\mathbb{R}) = C^\infty(M,i\,\mathfrak{u}(1))$. Thus, the fact that we only have the combination $A+JAJ^*$ effectively identifies the $U(1)$ gauge fields on the two copies of $M$, so that $A_\mu$ is determined by only one $U(1)$ gauge field. We summarize:
\begin{prop}
\label{prop:gauge-field}
The inner fluctuations of the almost commutative manifold $M \times F_X$ described above are parametrized by a $U(1)$-gauge field $Y_\mu$ as
$$
D \mapsto D' = D + \gamma^\mu Y_\mu \otimes \gamma_F.
$$
The action of the gauge group $\mathcal{G}(\mathcal{A}) \simeq C^\infty(M, U(1))$ on $D'$ by conjugation is implemented by
$$
Y_\mu \mapsto Y_\mu - i u \partial_\mu u^*, \qquad (u \in \mathcal{G}(\mathcal{A})).
$$
\end{prop}
So far we have seen that the almost commutative manifold $M\times F\Sub{X}$ describes a gauge theory with local gauge group $U(1)$, where the inner fluctuations of the Dirac operator provide the $U(1)$ gauge field $Y_\mu$. The question arises whether this model is suitable for a description of (classical) electrodynamics. There appear to be two problems, even before considering the fermionic action $S_f$ explicitly. First, by \cref{prop:zero_D_F}, the finite Dirac operator $D_F$ must vanish. However, we want our fermions to be massive, and for this purpose we need a finite Dirac operator that is non-zero.
Second, from \cite[Ch.7, \S5.2]{coleman}, we find the Euclidean action for a free Dirac field:
\begin{align}
\label{eq:coleman}
S = - \int i \bar\psi (\gamma^\mu\partial_\mu - m) \psi d^4x ,
\end{align}
where the fields $\psi$ and $\bar\psi$ must be considered as \emph{totally independent variables}. Thus, we require that the fermionic action $S_f$ should also yield two \emph{independent} Dirac spinors. Let us write $\left\{e, \bar e\right\}$ for the set of orthonormal basis vectors of $\mathcal{H}_F$, where $e$ is the basis element of $\mathcal{H}_F^+$ and $\bar e$ of $\mathcal{H}_F^-$. Note that on this basis, we have $J_Fe=\bar e$, $J_F\bar e=e$, $\gamma_Fe=e$ and $\gamma_F\bar e=-\bar e$. The total Hilbert space $\mathcal{H}$ is given by $L^2(M,S)\otimes\mathcal{H}_F$. Since we can also decompose $L^2(M,S) = L^2(M,S)^+ \oplus L^2(M,S)^-$ by means of $\gamma_5$, we obtain that the positive eigenspace $\mathcal{H}^+$ of $\gamma=\gamma_5\otimes\gamma_F$ is given by
\begin{align*}
\mathcal{H}^+ = L^2(M,S)^+\otimes\mathcal{H}_F^+ \oplus L^2(M,S)^-\otimes\mathcal{H}_F^- .
\end{align*}
An arbitrary vector $\xi\in\mathcal{H}^+$ can then uniquely be written as
\begin{align*}
\xi = \psi_L \otimes e + \psi_R \otimes \bar e ,
\end{align*}
for two Weyl spinors $\psi_L\in L^2(M,S)^+$ and $\psi_R\in L^2(M,S)^-$. One should note here that this vector $\xi$ is completely determined by only one Dirac spinor $\psi := \psi_L + \psi_R$, instead of the required two independent spinors. Thus, the restrictions that are incorporated into the fermionic action of \cref{defn:fermion_act} are such that the present example is in fact too restricted.
\section{Electrodynamics}
\label{sec:ED}
\subsection{The finite spectral triple}
Inspired by the previous section, which shows that one can use the framework of noncommutative geometry to describe a gauge theory with the abelian gauge group $U(1)$, we shall now attempt to describe the full theory of electrodynamics. There are two changes we need to make to the $U(1)$ gauge theory of the previous section. We need to incorporate a non-zero finite Dirac operator $D_F$ to obtain mass terms for the fermions, and we also need to obtain two independent Dirac spinors in the fermionic action. Both these changes can be simply obtained by doubling our finite Hilbert space.
We start with the same algebra $C^\infty(M,\mathbb{C}^2)$ that corresponds to the space $N = M\times X \simeq M\sqcup M$. The finite Hilbert space will now be used to describe four particles, namely both the left-handed and the right-handed electrons and positrons. We will choose the orthonormal basis $\left\{e_L, e_R, \bar{e_L}, \bar{e_R}\right\}$ for $\mathcal{H}_F = \mathbb{C}^4$, with respect to the standard inner product.
The subscript $L$ denotes left-handed particles, and the subscript $R$ denotes right-handed particles, and we take $\gamma_F e_L = e_L$ and $\gamma_F e_R = -e_R$.
We will choose $J_F$ such that it interchanges particles with their anti-particles, so $J_Fe_R = \bar{e_R}$ and $J_Fe_L = \bar{e_L}$. As in \cref{sec:U1_gauge}, we will choose the real structure such that is has KO-dimension $6$, so we have $J_F^2 = \mathbb{I}$ and $J_F\gamma_F = -\gamma_FJ_F$. This last relation implies that the element $\bar{e_R}$ is left-handed and $\bar{e_L}$ is right-handed. Hence, the grading $\gamma_F$ and the conjugation operator $J_F$ are given by
\begin{align*}
\gamma_F &= \matfour{1&0&0&0}{0&-1&0&0}{0&0&-1&0}{0&0&0&1} , & J_F &= \matfour{0&0&C&0}{0&0&0&C}{C&0&0&0}{0&C&0&0} .
\end{align*}
The grading $\gamma_F$ decomposes the Hilbert space $\mathcal{H}_F$ into $\mathcal{H}_L\oplus\mathcal{H}_R$, where the bases of $\mathcal{H}_L$ and $\mathcal{H}_R$ are given by $\{e_L,\bar{e_R}\}$ and $\{e_R,\bar{e_L}\}$, respectively. We can also decompose the Hilbert space into $\mathcal{H}_e\oplus\mathcal{H}_{\bar e}$, where $\mathcal{H}_e$ contains the electrons $\{e_L,e_R\}$, and $\mathcal{H}_{\bar e}$ contains the positrons $\{\bar{e_L},\bar{e_R}\}$.
The elements $a\in\mathcal{A}_F=\mathbb{C}^2$ are now represented on the basis $\left\{e_L, e_R, \bar{e_L}, \bar{e_R}\right\}$ as
\begin{align}
\label{eq:rep_ED}
a = \vectwo{a_1}{a_2} \mapsto \matfour{a_1&0&0&0}{0&a_1&0&0}{0&0&a_2&0}{0&0&0&a_2} .
\end{align}
Note that this representation commutes with the grading, as it should. We can also easily check that $[a,b^0] = 0$ for $b^0 := J_Fb^*J_F^*$, since both the left and the right action are given by diagonal matrices. For now, we will still take $D_F = 0$, and hence the order one condition is trivially satisfied. We have now obtained the following result:
\begin{prop}
The real even finite spectral triple
\begin{align*}
F\Sub{ED} := (\mathbb{C}^2, \mathbb{C}^4, 0, \gamma_F, J_F)
\end{align*}
as given above defines a real even finite spectral triple of KO-dimension $6$.
\end{prop}
\subsubsection{A non-trivial finite Dirac operator}
Let us now consider the possibilities for adding a non-zero Dirac operator to the finite spectral triple $F\Sub{ED}$. Since $D_F\gamma_F = -\gamma_FD_F$, the Dirac operator obtains the form
\begin{align*}
D_F = \matfour{0&d_1&d_2&0}{\bar d_1&0&0&d_3}{\bar d_2&0&0&d_4}{0&\bar d_3&\bar d_4&0} .
\end{align*}
Next, we impose the commutation relation $D_FJ_F = J_FD_F$, which implies $d_1 = \bar d_4$.
For the order one condition, we calculate
\begin{align*}
[D_F,a] &=
(a_1-a_2) \matfour{0&0&-d_2&0}{0&0&0&-d_3}{\bar d_2&0&0&0}{0&\bar d_3&0&0} .
\end{align*}
which then imposes the condition
\begin{align*}
0 = \big[[D_F,a],b^0\big]
&= (a_1-a_2)(b_2-b_1) \matfour{0&0&d_2&0}{0&0&0&d_3}{\bar d_2&0&0&0}{0&\bar d_3&0&0} .
\end{align*}
Since this must hold for all $a,b\in\mathbb{C}^2$, we must require that $d_2=d_3=0$. To conclude, the Dirac operator only depends on one complex parameter and is given by
\begin{align}
\label{eq:Dirac}
D_F = \matfour{0&d&0&0}{\bar d&0&0&0}{0&0&0&\bar d}{0&0&d&0} .
\end{align}
From here on, we will consider the finite spectral triple $F\Sub{ED}$ given by
\begin{align*}
F\Sub{ED} := (\mathbb{C}^2, \mathbb{C}^4, D_F, \gamma_F, J_F) .
\end{align*}
\subsection{The almost commutative manifold}
By taking the product with the canonical triple, our almost commutative manifold (of KO-dimension $2$) under consideration is given by
\begin{align*}
M\times F\Sub{ED} := \left( C^\infty(M,\mathbb{C}^2), L^2(M,S)\otimes\mathbb{C}^4, \slashed{D}\otimes\mathbb{I} + \gamma_5\otimes D_F, \gamma_5\otimes\gamma_F, J_M\otimes J_F \right) .
\end{align*}
As in \cref{sec:U1}, the algebra decomposes as $C^\infty(M,\mathbb{C}^2) = C^\infty(M)\oplus C^\infty(M)$, and we now decompose the Hilbert space as $\mathcal{H} = (L^2(M,S)\otimes\mathcal{H}_e)\oplus(L^2(M,S)\otimes\mathcal{H}_{\bar e})$. The action of the algebra on $\mathcal{H}$, given by \eqref{eq:rep_ED}, is then such that one component of the algebra acts on the electron fields $L^2(M,S)\otimes\mathcal{H}_e$, and the other component acts on the positron fields $L^2(M,S)\otimes\mathcal{H}_{\bar e}$.
The derivation of the gauge group for $F\Sub{ED}$ is exactly the same as in \cref{prop:gauge_group_U1}, so again we have the finite gauge group $\mathcal{G}(\mathcal{A}_F) \simeq U(1)$. The field $B_\mu := A_\mu - J_FA_\mu J_F^*$ now takes the form
\begin{align}
\label{eq:gauge_field}
B_\mu = \matfour{Y_\mu&0&0&0}{0&Y_\mu&0&0}{0&0&-Y_\mu&0}{0&0&0&-Y_\mu} \qquad\text{for } Y_\mu(x) \in \mathbb{R} .
\end{align}
Thus, we again obtain a single $U(1)$ gauge field $Y_\mu$, carrying an action of the gauge group $\mathcal{G}(\mathcal{A}) \simeq C^\infty(M, U(1))$ (as in Proposition \ref{prop:gauge-field}).
As mentioned before, our space $N$ consists of two copies of $M$, and the distance between these two copies is infinite (cf.\ \cref{sec:product_space_U1}). Now, we have introduced a non-zero Dirac operator, but it commutes with the algebra, i.e.\ $[D_F,a]=0$ for all $a\in\mathcal{A}$. Therefore, the distance between the two copies of $M$ is still infinite.
To summarize, the $U(1)$ gauge theory arises from the geometric space $N=M \sqcup M$ as follows. On one copy of $M$, we have the vector bundle $S\otimes(M\times\mathcal{H}_e)$, and on the other copy the vector bundle $S\otimes(M\times\mathcal{H}_{\bar e})$. The gauge fields on each copy of $M$ are effectively identified to each other. The electrons $e$ and positrons $\bar e$ are then both coupled to the same gauge field, and as such the gauge field provides an interaction between electrons and positrons.
\subsection{The Lagrangian}
We are now ready to explicitly calculate the Lagrangian that corresponds to the almost commutative manifold $M\times F\Sub{ED}$, and we will show that this yields the usual Lagrangian for electrodynamics (on a curved background manifold), as well as a purely gravitational Lagrangian. The action functional for an almost commutative manifold, as defined in \cref{defn:fermion_act}, consists of the spectral action $S_b$ and the fermionic action $S_f$, which we will calculate separately.
\subsubsection{The spectral action}
Before we can calculate the spectral action, we first need to study the fluctuated Dirac operator in a little more detail. As in \eqref{eq:inn_fluc}, we have $A+JAJ^* = \gamma^\mu\otimes B_\mu$, where now $B_\mu$ is given by \eqref{eq:gauge_field}. This allows us to rewrite the fluctuated Dirac operator in the form
\begin{align*}
D_A = \slashed{D}\otimes\mathbb{I} + \gamma^\mu\otimes B_\mu + \gamma_5\otimes D_F = -i\gamma^\mu\otimes\nabla^E_\mu + \gamma_5\otimes D_F ,
\end{align*}
where we have defined a new connection $\nabla^E_\mu$ by
\begin{align}
\label{eq:conn_E}
\nabla^E_\mu = \nabla^S_\mu\otimes\mathbb{I} + i \mathbb{I}\otimes B_\mu .
\end{align}
For the square of the fluctuated Dirac operator, we obtain by direct calculation that
\begin{align*}
{D_A}^2 = \Delta^E - Q ,
\end{align*}
where $\Delta^E$ is the Laplacian of the connection $\nabla^E$, and where $Q\in\Gamma(\textnormal{End}(M\times\mathcal{H}_F))$ is given by
\begin{align*}
Q= -\frac14 s\otimes\mathbb{I} - \mathbb{I}\otimes {D_F}^2 + \frac12 i \gamma^\mu\gamma^\nu\otimes F_{\mu\nu} .
\end{align*}
Here we have defined the curvature $F_{\mu\nu}$ of the field $B_\mu$ as $F_{\mu\nu} := \partial_\mu B_\nu - \partial_\nu B_\mu$.
\begin{prop}
\label{prop:spec_act_ED}
The spectral action of the almost commutative manifold
\begin{align*}
M\times F\Sub{ED} = \left( C^\infty(M,\mathbb{C}^2), L^2(M,S)\otimes\mathbb{C}^4, \slashed{D}\otimes\mathbb{I} + \gamma_5\otimes D_F, \gamma_5\otimes\gamma_F, J_M\otimes J_F \right)
\end{align*}
is given by
\begin{align*}
\textnormal{Tr} \left(f\Big(\frac {D_A}\Lambda\Big)\right) &\sim \frac1{4\pi^2} \int_M \L(g_{\mu\nu}, Y_\mu) \sqrt{|g|} d^4x + O(\Lambda^{-1}) ,
\end{align*}
for
\begin{align*}
\L(g_{\mu\nu}, Y_\mu) := 4\L_M(g_{\mu\nu}) + \L_Y(Y_\mu) + \L_H(g_{\mu\nu},d) .
\end{align*}
Here $\L_M(g_{\mu\nu})$ is defined in \cref{ex:canon_spec_act}. $\L_Y$ gives the kinetic term of the $U(1)$-gauge field $Y_\mu$ and equals
\begin{align*}
\L_Y(Y_\mu) := \frac23f(0) \mathcal{F}_{\mu\nu}\mathcal{F}^{\mu\nu} ,
\end{align*}
where we have defined the curvature $\mathcal{F}_{\mu\nu}$ of the field $Y_\mu$ as $\mathcal{F}_{\mu\nu} := \partial_\mu Y_\nu - \partial_\nu Y_\mu$.
The Higgs potential $\L_H$ (ignoring the boundary term) only gives two constant terms which add to the cosmological constant, plus an extra contribution to the Einstein-Hilbert action:
\begin{align*}
\L_H(g_{\mu\nu}) := -8f_2\Lambda^2 |d|^2 + 2 f(0) |d|^4 + \frac{1}{3} f(0) s|d|^2 .
\end{align*}
\end{prop}
\begin{proof}
Since ${D_A}^2$ is of the form $\Delta^E - Q$, we obtain the heat expansion of the spectral action from \eqref{eq:action_expansion}. Thus, we only need to calculate the Seeley-DeWitt coefficients from \cref{thm:seeley-dewitt}. The trace over the Hilbert space $\mathcal{H}_F$ yields an overall factor $4 = \dim \mathcal{H}_F$, so we obtain
\begin{align*}
a_0(x,{D_A}^2) = 4 a_0(x,\slashed{D}^2) .
\end{align*}
For the second coefficient we have
\begin{align*}
a_2(x,{D_A}^2) = 4 a_2(x,\slashed{D}^2) + \frac{1}{16\pi^2} \textnormal{Tr}\Big(-\mathbb{I}\otimes {D_F}^2 + \frac12i\gamma^\mu\gamma^\nu\otimes F_{\mu\nu}\Big) .
\end{align*}
Since $\textnormal{Tr}(\gamma^\mu\gamma^\nu) = 4g^{\mu\nu}$ and $F_{\mu\nu}$ is anti-symmetric, the trace over the last term vanishes. From \eqref{eq:Dirac} we easily see that ${D_F}^2 = |d|^2\mathbb{I}$, so we obtain
\begin{align*}
a_2(x,{D_A}^2) = 4 a_2(x,\slashed{D}^2) - \frac{|d|^2}{\pi^2} .
\end{align*}
For the last coefficient, we need the curvature $\Omega^E_{\mu\nu}$ of the connection $\nabla^E$ of \eqref{eq:conn_E}. Its square is given by
\begin{align*}
\Omega^E_{\mu\nu}{\Omega^E}^{\mu\nu} = \Omega^S_{\mu\nu}{\Omega^S}^{\mu\nu}\otimes\mathbb{I} - \mathbb{I}\otimes F_{\mu\nu}F^{\mu\nu} + 2i \Omega^S_{\mu\nu}\otimes F^{\mu\nu} ,
\end{align*}
where the last term is traceless.
We also obtain a contribution from $Q^2$, which is given by
\begin{align*}
Q^2 = \frac{1}{16} s^2\otimes\mathbb{I} + \mathbb{I}\otimes {D_F}^4 - \frac14 \gamma^\mu\gamma^\nu\gamma^\rho\gamma^\sigma\otimes F_{\mu\nu}F_{\rho\sigma} + \frac12 s\otimes{D_F}^2 + \text{ traceless terms} .
\end{align*}
We shall ignore the boundary term $\Delta Q$. The last Seeley-DeWitt coefficient is then given by
\begin{multline*}
a_4(x,{D_A}^2) = 4 a_4(x,\slashed{D}^2) + \frac{1}{16\pi^2} \frac{1}{360} \textnormal{Tr}\Big(- 60 s\otimes {D_F}^2 + 180 \big(\mathbb{I}\otimes {D_F}^4 \\
- \frac14 \gamma^\mu\gamma^\nu\gamma^\rho\gamma^\sigma\otimes F_{\mu\nu}F_{\rho\sigma} + \frac12 s\otimes{D_F}^2\big) - 30 \mathbb{I}\otimes F_{\mu\nu}F^{\mu\nu} \Big) .
\end{multline*}
Using the trace identity
\begin{align*}
\textnormal{Tr}\big(\frac14 \gamma^\mu\gamma^\nu\gamma^\rho\gamma^\sigma\big) = g^{\mu\nu}g^{\rho\sigma} - g^{\mu\rho}g^{\nu\sigma} + g^{\mu\sigma}g^{\nu\rho} ,
\end{align*}
along with the anti-symmetry of $F_{\mu\nu}$, we calculate that
\begin{align*}
\textnormal{Tr}\big(- \frac14 \gamma^\mu\gamma^\nu\gamma^\rho\gamma^\sigma\otimes F_{\mu\nu}F_{\rho\sigma}\big) = 2 \textnormal{Tr}(F_{\mu\nu}F^{\mu\nu}) = 8 \mathcal{F}_{\mu\nu}\mathcal{F}^{\mu\nu}.
\end{align*}
We thus obtain the final coefficient as
\begin{align*}
a_4(x,{D_A}^2) = 4 a_4(x,\slashed{D}^2) + \frac{1}{12\pi^2} s|d|^2 + \frac{1}{2\pi^2} |d|^4 + \frac{1}{6\pi^2} \mathcal{F}_{\mu\nu}\mathcal{F}^{\mu\nu} .
\end{align*}
The result now follows from inserting these Seeley-DeWitt coefficients into the asymptotic expansion \eqref{eq:action_expansion}, where we realize that the coefficients $a_k(\slashed{D}^2)$ yield the gravitational Lagrangian $\L_M$ of \cref{ex:canon_spec_act}.
\end{proof}
\subsubsection{The fermionic action}
We have written the set of basis vectors of $\mathcal{H}_F$ as $\left\{e_L, e_R, \bar{e_L}, \bar{e_R}\right\}$, and the subspaces $\mathcal{H}_F^+$ and $\mathcal{H}_F^-$ are spanned by $\left\{e_L, \bar{e_R}\right\}$ and $\left\{e_R, \bar{e_L}\right\}$, respectively. The total Hilbert space $\mathcal{H}$ is given by $L^2(M,S)\otimes\mathcal{H}_F$. Since we can also decompose $L^2(M,S) = L^2(M,S)^+ \oplus L^2(M,S)^-$ by means of $\gamma_5$, we obtain
\begin{align*}
\mathcal{H}^+ = L^2(M,S)^+\otimes\mathcal{H}_F^+ \oplus L^2(M,S)^-\otimes\mathcal{H}_F^- .
\end{align*}
A spinor $\psi\in L^2(M,S)$ can be decomposed as $\psi = \psi_L + \psi_R$. Each subspace $\mathcal{H}_F^\pm$ is now spanned by two basis vectors. A generic element of the tensor product of two spaces consists of sums of tensor products, so an arbitrary vector $\xi\in\mathcal{H}^+$ can uniquely be written as
\begin{align}
\label{eq:right_fermion_ED}
\xi = \chi_L\otimes e_L + \chi_R\otimes e_R + \psi_R\otimes\bar{e_L} + \psi_L\otimes\bar{e_R} ,
\end{align}
for Weyl spinors $\chi_L,\psi_L\in L^2(M,S)^+$ and $\chi_R,\psi_R\in L^2(M,S)^-$. Note that this vector $\xi\in\mathcal{H}^+$ is now completely determined by two Dirac spinors $\chi := \chi_L + \chi_R$ and $\psi := \psi_L + \psi_R$.
\begin{prop}
\label{prop:fermion_act_ED}
The fermionic action of the almost commutative manifold
\begin{align*}
M\times F\Sub{ED} = \left( C^\infty(M,\mathbb{C}^2), L^2(M,S)\otimes\mathbb{C}^4, \slashed{D}\otimes\mathbb{I} + \gamma_5\otimes D_F, \gamma_5\otimes\gamma_F, J_M\otimes J_F \right)
\end{align*}
is given by
\begin{align*}
S_f = -i\big\langle J_M\til\chi,\gamma^\mu(\nabla^S_\mu - i Y_\mu)\til\psi\big\rangle + \langle J_M\til\chi_L,\bar d\til\psi_L\rangle - \langle J_M\til\chi_R,d\til\psi_R\rangle .
\end{align*}
\end{prop}
\begin{proof}
The fluctuated Dirac operator is given by
\begin{align*}
D_A = \slashed{D}\otimes\mathbb{I} + \gamma^\mu\otimes B_\mu + \gamma_5\otimes D_F .
\end{align*}
An arbitrary $\xi\in\mathcal{H}^+$ has the form of \eqref{eq:right_fermion_ED}, and then we obtain the following expressions:
\begin{align*}
J\xi &= J_M\chi_L\otimes\bar{e_L} + J_M\chi_R\otimes\bar{e_R} + J_M\psi_R\otimes e_L + J_M\psi_L\otimes e_R, \\
(\slashed{D}\otimes\mathbb{I})\xi &= \slashed{D}\chi_L\otimes e_L + \slashed{D}\chi_R\otimes e_R + \slashed{D}\psi_R\otimes\bar{e_L} + \slashed{D}\psi_L\otimes\bar{e_R} , \\
(\gamma^\mu\otimes B_\mu)\xi &= \gamma^\mu\chi_L\otimes Y_\mu e_L + \gamma^\mu\chi_R\otimes Y_\mu e_R - \gamma^\mu\psi_R\otimes Y_\mu\bar{e_L} - \gamma^\mu\psi_L\otimes Y_\mu\bar{e_R} , \\
(\gamma_5\otimes D_F)\xi &= \gamma_5\chi_L\otimes \bar de_R + \gamma_5\chi_R\otimes d e_L + \gamma_5\psi_R\otimes d \bar{e_R} + \gamma_5\psi_L\otimes \bar d\bar{e_L} .
\end{align*}
We decompose the fermionic action into the three terms
\begin{align*}
\frac12 \langle J\til\xi,D_A\til\xi\rangle &= \frac12 \langle J\til\xi,(\slashed{D}\otimes\mathbb{I})\til\xi\rangle + \frac12 \langle J\til\xi,(\gamma^\mu\otimes B_\mu)\til\xi\rangle + \frac12 \langle J\til\xi,(\gamma_5\otimes D_F)\til\xi\rangle ,
\end{align*}
and then continue to calculate each term separately. The first term is given by
\begin{align*}
\frac12 \langle J\til\xi,(\slashed{D}\otimes\mathbb{I})\til\xi\rangle = \frac12 \langle J_M\til\chi_L,\slashed{D}\til\psi_R\rangle + \frac12 \langle J_M\til\chi_R,\slashed{D}\til\psi_L\rangle + \frac12 \langle J_M\til\psi_R,\slashed{D}\til\chi_L\rangle + \frac12 \langle J_M\til\psi_L,\slashed{D}\til\chi_R\rangle .
\end{align*}
Using the fact that $\slashed{D}$ changes the chirality of a Weyl spinor, and that the subspaces $L^2(M,S)^+$ and $L^2(M,S)^-$ are orthogonal, we can rewrite this term as
\begin{align*}
\frac12 \langle J\til\xi,(\slashed{D}\otimes\mathbb{I})\til\xi\rangle = \frac12 \langle J_M\til\chi,\slashed{D}\til\psi\rangle + \frac12 \langle J_M\til\psi,\slashed{D}\til\chi\rangle .
\end{align*}
Using the symmetry of the form $\langle J_M\til\chi,\slashed{D}\til\psi\rangle$, we obtain
\begin{align*}
\frac12 \langle J\til\xi,(\slashed{D}\otimes\mathbb{I})\til\xi\rangle = \langle J_M\til\chi,\slashed{D}\til\psi\rangle = -i \langle J_M\til\chi,\gamma^\mu\nabla^S_\mu\til\psi\rangle .
\end{align*}
Note that the factor $\frac12$ has now disappeared in the result, and this is the reason why this factor is included in the definition of the fermionic action. The second term is given by
\begin{multline*}
\frac12 \langle J\til\xi,(\gamma^\mu\otimes B_\mu)\til\xi\rangle = - \frac12 \langle J_M\til\chi_L,\gamma^\mu Y_\mu\til\psi_R\rangle - \frac12 \langle J_M\til\chi_R,\gamma^\mu Y_\mu\til\psi_L\rangle \\
+ \frac12 \langle J_M\til\psi_R,\gamma^\mu Y_\mu\til\chi_L\rangle + \frac12 \langle J_M\til\psi_L,\gamma^\mu Y_\mu\til\chi_R\rangle .
\end{multline*}
In a similar manner as above, we obtain
\begin{align*}
\frac12 \langle J\til\xi,(\gamma^\mu\otimes B_\mu)\til\xi\rangle = -\langle J_M\til\chi,\gamma^\mu Y_\mu\til\psi\rangle ,
\end{align*}
where we have used that the form $\langle J_M\til\chi,\gamma^\mu Y_\mu\til\psi\rangle$ is anti-symmetric. The third term is given by
\begin{multline*}
\frac12 \langle J\til\xi,(\gamma_5\otimes D_F)\til\xi\rangle = \frac12 \langle J_M\til\chi_L,\bar d\gamma_5\til\psi_L\rangle + \frac12 \langle J_M\til\chi_R,d\gamma_5\til\psi_R\rangle \\
+ \frac12 \langle J_M\til\psi_R,d\gamma_5\til\chi_R\rangle + \frac12 \langle J_M\til\psi_L,\bar d\gamma_5\til\chi_L\rangle .
\end{multline*}
The bilinear form $\langle J_M\til\chi,\gamma_5\til\psi\rangle$ is again symmetric, but we now have the extra complication that two terms contain the parameter $d$, while the other two terms contain $\bar{d}$. Therefore we are left with two distinct terms:
\begin{equation*}
\frac12 \langle J\til\xi,(\gamma_5\otimes D_F)\til\xi\rangle = \langle J_M\til\chi_L,\bar d\til\psi_L\rangle - \langle J_M\til\chi_R,d\til\psi_R\rangle . \qedhere
\end{equation*}
\end{proof}
\begin{remark}
\label{remark:real_mass}
It is interesting to note that the fermions acquire mass terms without being coupled to a Higgs field. However, it seems we obtain a complex mass parameter $d$, where we would desire a real parameter $m$. By simply requiring that our result should be similar to \eqref{eq:coleman}, we will choose $d:=-im$, so that
\begin{align*}
\langle J_M\til\chi_L,\bar d\til\psi_L\rangle - \langle J_M\til\chi_R,d\til\psi_R\rangle = i \big\langle J_M\til\chi,m\til\psi\big\rangle .
\end{align*}
\end{remark}
The results obtained in this section can now be summarized into the following theorem.
\begin{thm}
\label{thm:ED}
The full Lagrangian of the almost commutative manifold
\begin{align*}
M\times F\Sub{ED} = \left( C^\infty(M,\mathbb{C}^2), L^2(M,S)\otimes\mathbb{C}^4, \slashed{D}\otimes\mathbb{I} + \gamma_5\otimes D_F, \gamma_5\otimes\gamma_F, J_M\otimes J_F \right)
\end{align*}
as defined in this section, can be written as the sum of a purely gravitational Lagrangian,
\begin{align*}
\L\Sub{\text{grav}}(g_{\mu\nu}) = \frac1{\pi^2} \L_M(g_{\mu\nu}) + \frac1{4\pi^2} \L_H(g_{\mu\nu}) ,
\end{align*}
and a Lagrangian for electrodynamics,
\begin{align*}
\L\Sub{\text{ED}} = -i\Big( J_M\til\chi,(\gamma^\mu(\nabla^S_\mu - i Y_\mu)-m)\til\psi\Big) + \frac1{6\pi^2} f(0) \mathcal{F}_{\mu\nu}\mathcal{F}^{\mu\nu} .
\end{align*}
\end{thm}
\begin{proof}
The spectral action $S_b$ and the fermionic action $S_f$ are given by \cref{prop:spec_act_ED,prop:fermion_act_ED}. We shall now absorb all numerical constants into the Lagrangians as well. This immediately yields $\L\Sub{\text{grav}}$. To obtain $\L\Sub{\text{ED}}$, we need to rewrite the fermionic action $S_f$ as the integral over a Lagrangian. The inner product $\langle\;,\;\rangle$ on the Hilbert space $L^2(S)$ is given by
\begin{align*}
\langle\xi,\psi\rangle = \int_M (\xi,\psi) \sqrt{|g|} d^4x ,
\end{align*}
where the hermitian pairing $(\;,\;)$ is given by the pointwise inner product on the fibres. Choosing $d=-im$ as in \cref{remark:real_mass}, we can then rewrite the fermionic action into
\begin{align*}
S_f &= -\int_M i\Big( J_M\til\chi,\big(\gamma^\mu(\nabla^S_\mu - i Y_\mu)-m\big)\til\psi\Big) \sqrt{|g|} d^4x . \qedhere
\end{align*}
\end{proof}
\section*{Acknowledgements}
\noindent We thank Jord Boeijink and Thijs van den Broek for valuable suggestions and remarks.
\newcommand{\noopsort}[1]{}\def$'${$'$}
|
\section{Introduction}
\label{s:introduction}
Pixel detectors are an attractive choice for the inner tracking
regions of current and future particle physics detectors, as they provide
high granularity, radiation hardness, and ease of pattern recognition. A
candidate pixel ASIC which is well suited to applications at LHC,
SLHC, and future forward geometry trackers with high rate and
resolution requirements is the Timepix chip characterised in this paper.
The 55~${\rm \mu m}$ square pixel allows single sided
modules to be built, and it can supply analogue time
over threshold information in addition to the ability to associate
hits to the correct bunch crossing via time stamping.
The Timepix chip has not been extensively investigated for the
purposes of charged particle tracking, and for this reason a test
experiment was performed in a charged 120~GeV pion beam at CERN's
North Area. A particle telescope was constructed using
an array of Timepix and Medipix2 silicon sensor assemblies, and this was used to
measure the performance of a dedicated test assembly whilst varying
parameters such as silicon bias, track incident angle, and chip settings.
The telescope was also used to measure the performance of
a 3D sensor bonded to a Timepix chip, which represents an option for radiation
hard SLHC applications. The same 3D sensor was also tested in an
X-ray beam, and the results are reported in the companion
paper~\cite{3Dcompanion}.
Due to the excellent performance of
the Timepix telescope, it was possible to perform a detailed
investigation of the resolution, efficiency and charge sharing
performance of the devices under test. This paper reports for the
first time a detailed measurement of the resolution of a 55~$\mu$m
square pixel device
as a function of the angle of incident tracks to both the normal and
perpendicular directions with respect to the pixel columns.
The charge sharing characteristics and Landau distributions
obtained are discussed in detail.
Of particular interest are the plans to upgrade the LHCb silicon
vertex tracker (VELO~\cite{mvb})
for which extensive R\&D is being performed to develop a design based on pixel technology~\cite{LoIEoi,janvertex,hiroshima}. A
pixel readout chip dedicated to the VELO application (VeloPix) will be
developed based on Timepix, which will retain the advantageous
geometry and analogue ToT information, but have a significantly
increased readout bandwidth to cope with the higher data rates anticipated.
For this reason two sections are included in this paper
devoted to specific measurements relevant for LHC upgrades: the time
stamping capabilities of the Timepix ASIC and the effect of reducing
the number of bits available for the analogue charge measurements.
This paper is organised as follows. The Timepix chip is introduced in Section~\ref{s:timepixchip},
followed by a description of the telescope concept and layout in Section~\ref{s:telescope}. The
calibration of the chip, which introduces significant offsets compared
the amount of charge deposited by minimum ionising particles, is discussed in
detail. The telescope setup is described, and the extraction of the
performance, in particular the resolution at the device under test, is
given in Sections~\ref{s:datatreatment}~and~\ref{s:analysisofdeviceundertest}.
Finally, results are presented for the devices under test:
the planar sensor, which is based on well known technology providing
information on the performance of the Timepix chip, and the double-sided 3D sensor, which
is a novel device being tested together with the Timepix chip for the first
time in a pion beam, in Sections~\ref{s:clusterchars}-\ref{s:gainsection}. The results are summarised in the
Section~\ref{s:conclusions}.
This paper reports on work carried out jointly between the Medipix2 and LHCb
collaborations.
\section{The Timepix chip}
\label{s:timepixchip}
\subsection{Description of the chip}
The design of the Timepix~\cite{timepix} readout chip is derived from that of the Medipix2~\cite{medipix2} chip which was developed for
single photon counting applications. The development of the Timepix chip took place within the context of the Medipix2 collaboration
at the request of and with support from the EUDet Consortium~\cite{eudet}.
The Timepix ASIC comprises a $256 \times 256$ matrix of 55~$\mu$m square pixels, each of which contains its own analogue and
digital circuitry. A globally applied shutter signal determines when all pixels are active, switching between recording
data and transferring it off the chip.
Each pixel contains a preamplifier, a discriminator with a globally adjustable
threshold, followed by mode control logic and a 14-bit pseudorandom counter with overflow logic which stops after 11,810 counts.
The threshold applied to each pixel can be individually tuned by a four-bit in-pixel trimming circuit to compensate for variations in fabrication.
The three modes
of operation are counting, time of arrival (ToA), and time over threshold (ToT). In counting mode the Timepix pixels
behave in a similar manner to the Medipix2 pixels, incrementing the counter each time the output of the amplifier
passes the threshold. In ToA mode the pixel records the time it was first hit. The counter is started when the amplifier
first passes the threshold and is stopped when the shutter closes.
In this mode the depth of the counter and the overflow logic limit the shutter opening, a particular limitation at high clock frequencies.
Beyond 11,810 counts the counter saturates.
In ToT mode the discriminator output is used to gate the clock to the counter,
thus providing an indication of the total energy deposited.
The clock and counter are used to record
the time the amplifier signal is above threshold.
The amplifier has been designed so that this value is linear up to 50,000 electrons and its rise and fall time can be tuned respectively by adjusting the preamplifier and I$_\textbf{krum}$ DAC values. However, in ToT mode the dynamic range of the measurement extends well beyond the linear range of the preamplifier output.
The nominal rise time is ~100 ns and the fall time ~2 μs.
It is
possible to individually program the pixels to operate in different modes from their neighbours and be read out
simultaneously.
The entire chip is read out after the shutter signal goes low. Using the serial readout at 200 MHz the time to read the whole matrix is $\sim$5~ms. There is a possibility of reading out the chip using a 32-bit parallel bus. In this case the maximum clock frequency is 100 MHz and the minimum readout time $\sim$ 300~${\rm \mu s}$.
Timepix has a noise per pixel level of $\sim${100} electrons rms and a
threshold variation after tuning of $\sim${35} electrons rms, which leads to a minimum detectable signal on all the pixels of $\sim${650} electrons.
The Timepix and Medipix2 ASICs used for the telescope were bump bonded to 300~$\mu$m thick planar silicon sensor
chips which were nominally biased to 100~V. Although Timepix was designed primarily to
operate without a sensor it retains the Medipix2 threshold trimming and sensor polarity circuitry, making it
equally well suited to reading out silicon sensors.
\subsection{Use of the chip}
In the measurements described in this paper the Timepix chips were predominantly used in ToT mode as it allowed
precise track position calculations to be performed when charge was deposited in multiple pixels by the same particle.
This gave the opportunity to reconstruct the impact position of the particle with a precision better than the
binary resolution, which would correspond to the 55~$\mu$m pitch divided by $\sqrt{12}$.
As the Timepix and Medipix2 chips are read out using a shutter signal in a ``camera'' like manner and there is no
delay line or pipeline implemented in the pixels, it is not possible to use a trigger from a scintillator to
read out a selected event. However, provided the hits are sparsely distributed across the chip several hundred
tracks can be accumulated in one frame whilst the sensor is continuously sensitive and then read out when the shutter is closed.
Before the telescope assemblies can be operated it is necessary to set the threshold trim settings for each
pixel. These bits are set to values that allow the pixels across the matrix to respond to the global threshold
in a uniform way, compensating for any variation present due to inhomogenieties in fabrication or minor
radiation damage. The process of setting these bits is referred to as the ``threshold equalisation''.
Two methods of threshold equalisation are of relevance to this report, the ``noise edge'' method, which has been
used as the standard method by the Medipix group, and the `noise centre' method, which provides a good estimation
of the global noise floor of the chip.
Both methods operate by scanning the threshold level and binning the pixels according to the point at which they are noisy. Threshold scans are made with the trims set to their minimum and maximum values.
The optimum value of the trim
for each pixel is then determined and the threshold re-scanned to confirm the width of the distribution. The noise
edge and noise centre methods differ in their approach by how they locate the value of the pixels' noise point.
Using the noise mean method the threshold is scanned all the way through the noise floor and the pixels are binned according to the threshold DAC value at which the noise counts are highest. The noise edge method scans the threshold until a pixel passes a preset number of noise hits per unit time, and then uses this DAC value to bin the pixel and ceases to scan that pixel. This means that there is an offset between the absolute values of the noise position recorded. The noise mean method returns a central value which is referred to in this paper as the ``noise floor''. Typically the noise edge method returns a value with a negative offset
of about 20 THL DAC steps (the threshold decreases with increasing threshold DAC value), corresponding to roughly 500 electrons, with respect to the noise floor.
\section{The Timepix telescope}
\label{s:telescope}
The telescope, shown in Figure~\ref{fg:telescopephoto},
was constructed with four Timepix and two Medipix2 planes.
This allowed the easy integration of the Device Under Test (DUT), provided complete compatibility between the readout systems
and enabled an event rate of 200~Hz. The layout of the telescope
together with the global coordinate frame is shown in Figure~\ref{fg:telescope_schem}. Note the labels
identifying each chip, which will be used on certain plots in the remainder of this paper.
\epspicz{1}{2}{telescopephoto}{Photograph of the Timepix telescope used in the testbeam.}
\begin{figure}[h]
\begin{center}
\includegraphics[width=14cm]{Telescope_2009b.eps}
\end{center}
\caption{A diagram of the pixel sensor assemblies within the telescope, showing the angled four Timepix and two Medipix2 sensors
and the Timepix DUT with its axis of rotation. The separation of the devices shown is as measured from the centers of the chips and was
enforced by the angle and the connection to the readout systems.}
\label{fg:telescope_schem}
\end{figure}
The telescope consisted of six pixel planes, four Timepix and two Medipix2 separated from each other and the DUT by 78~mm.
As the Medipix2 planes only provide binary information, and so produce
a lower resolution, they were sandwiched between the Timepix
sensors; the Timepix sensors form the innermost and outermost stations of the telescope arms.
A particle passing through the sensor at an angle of $\tan{^{-1}(\frac{\textrm{Pitch}}{\textrm{Thickness}})}$
will always traverse more than one pixel. This is, therefore, an
approximation of the optimum angle of the sensor to the particle beam,
at which position resolution can be improved by weighting the charge
deposition in the pixels (centroiding).
When the sensor is at this angle the centroiding calculation incorporating the ToT information from the Timepix can be used for all recorded tracks,
as all the tracks will leave multi-hit clusters. Following this logic the telescope sensors were fixed to $9^{\circ}$ in both the
horizontal and vertical axes perpendicular to the beam line to optimise the spatial resolution that could be achieved by the telescope.
The DUT was mounted at the centre of a symmetric arrangement of chips to further increase the resolution that could be achieved.
To allow a resolution measurement to be taken with the DUT at as many angles as possible, and to increase precision, the DUT was mounted
on high precision rotation and translation stages~\cite{stages} driven by stepper motors that allowed it to be turned and aligned remotely.
The stages used were supplied by PI\footnote{Physik Instrumente GmbH,
D-76228 Karlsruhe} and allowed a repeatability of 2~$\mu$m and
50~$\mu$rad for the translation and rotation states respectively\footnote{The reference numbers of
the components used are M403.42S and M-060.2S}.
The Medipix2 and Timepix assemblies, including the DUT, were read out using USB driven systems provided by CTU~\cite{USB} and the Pixelman
data acquisition and control software~\cite{pixelman}. These are the standard, portable, low bandwidth readout systems used in most Medipix
applications to date. Each USB unit is attached to one chip. A signal is applied simultaneously to all the USB readout units and its rising
edge triggers local shutters to individual chips. The length of the shutter is programmable in each USB unit and it was set to be the same
for each assembly. It was optimised on a run by run basis to capture between 100 and 500 tracks per frame depending on beam conditions.
The micro-controller in the USB unit introduces a delay between the trigger being received and the shutter being sent of $4.0\pm0.5$~$\mu$s.
As shutter periods of down to 10~ms were used depending on the beam
intensity the error on the efficiency measurement introduced by this jitter should be small.
In situations where a shorter shutter is required, such as high particle flux environments, this will become a constraint on the use of the existing USB systems.
\section{Devices under test (DUTs)}
For the data taking described in this paper, two devices were
investigated in the central position of the telescope; a standard
planar sensor similar to the devices making up the telescope planes
themselves, which was used to make a thorough investigation of the
Timepix performance, and a 3D sensor, where the most interesting feature was the performance of the sensor itself in combination with the Timepix chip.
\subsection{Planar sensor}
\label{planarsensor}
The planar sensor used was one of a standard series of sensors
produced by CANBERRA\footnote{CANBERRA Industries, Semiconductor NV Belgium.} for the Medipix2 collaboration. The
substrate is n-type, and the device has $p^+$ electrodes for hole
collection. The substrate resistivity was 32~${\rm k}\Omega {\rm cm}$,
corresponding to 10~V full depletion voltage for the $300~ {\rm \mu m}$ thick
device. The leakage current of the device used in this testbeam was
less than 1~nA at 40~V applied voltage. The bump bonding assembly was
carried out at VTT\footnote{ VTT Technical Research Centre of Finland, P.O. Box 1000, FI-02044 VTT, Finland} using a
Timepix chip with no inactive channels, and there is
expected to be negligible loss due to noisy pixels.
\subsection{3D sensor}
\label{s:3dsensor}
A 3D sensor has an array of n- and p-type electrode columns passing through the
thickness of a silicon substrate, rather than being implanted on
the substrate’s surface as in a planar sensor. These electrodes are realized by a
combination of micro-machining techniques and standard
detector technologies~\cite{3DFabrication}.
By using this structure it is
possible to combine a standard substrate thickness of a few
hundred micrometers with a lateral spacing between electrodes
of a few tens of micrometers. The depletion and charge
collection distances are thereby dramatically reduced, without reducing
the sensitive thickness of the sensor. This implies that the
device has extremely fast charge collection and a low
operating voltage even after a high irradiation dose. The short
collection distance and the electric field pattern in the device
also reduce the amount trapping that takes place. These
features make 3D sensors potentially very radiation hard; in~\cite{3DIrradiation} it is shown that a $285$~$\mu$m
thick double-sided 3D sensor irradiated to $10^{16}$~1~MeV neutron equivalents/$\rm{cm}^{2}$
biased at 350~V collects 20,000 electrons.
Hence, 3D sensors are a promising technology for inner layers of vertex detectors
at the LHC upgrades, which require operation beyond this fluence.
The double-sided 3D sensor~\cite{firstDoubleSided3D} used in the
testbeam measurements described here is a modification of the traditional same-side 3D design~\cite{3DParker}.
In the same-side 3D devices the n- and p-type columns penetrate through the full thickness of the substrate, and the columns for both
electrode types are etched from the same side and a handle wafer, wafer-bonded to the sensor wafer, is required in the
processing. The double-sided device has the columns etched from opposite sides of the wafer for each type of column
doping. This removes the necessity to have a handling wafer and improves production yield and lowers fabrication costs.
The double-sided 3D sensor used in this beam test was designed by the University of Glasgow and IMB-CNM and
fabricated at IMB-CNM\footnote{IMB-CNM, Centro Nacional de Microelectr{\'o}nica. Campus Universidad
Aut{\'o}noma de Barcelona. 08193 Bellaterra (Barcelona), Spain.}.
The geometry of the sensor, with electrode columns passing through the thickness of the wafer from
the front and back sides, is illustrated in Figure~\ref{fg:3DSchematic}. The $5$~$\mu$m radius
columns are produced using deep reactive ion etching, and filled with 3~$\mu$m thickness
polysilicon around the pore walls. The columns are doped by diffusion through the polysilicon
with Boron (p-type columns) or Phosphorous (n-type columns), and the inside of the columns
passivated with Silica. The back surface is coated with metal for biasing and readout electrodes
deposited on the front-surface. Hole and electron collecting devices were fabricated. The
device used here is a hole collecting double-sided 3D n-type sensor with an n-type substrate and
p-doped columns connected to the electronic readout. The columns have a depth of $250$~$\mu$m,
whereas the substrate is $285\pm 15$~$\mu$m thick. The region between the n- and p-type columns depletes at only a few
volts. There is a lower field region in the sensor directly above a column which requires the
sensor to be biased to around 10~V to be fully active~\cite{3Dsimulation}. The device was
solder bump-bonded by VTT
to a Timepix readout chip. A lower grade Timepix readout chip was used for this R\&D project with
some inactive pixel columns (see later). The electrical characteristics of the double sided 3D
devices measured in the laboratory are reported in~\cite{3DIrradiation} and~\cite{firstDoubleSided3D},
and the device used here had a leakage current of $3.8$~$\mu$A at 20~V at room temperature.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.8\textwidth]{3DStructureSimple.eps}
\end{center}
\caption{Design of a single pixel cell in an n-type bulk double-sided 3D sensor, showing the
electrode columns partially etched from the top and bottom surfaces of the
sensor. The vertical scale has been compressed.}
\label{fg:3DSchematic}
\end{figure}
\section{Charge calibration of DUT}
\label{s:calibration}
For the data
described in this paper,
the standard threshold setting and the mean THL DAC values returned by the
equalisation scans described in Section~\ref{s:timepixchip} are shown in Table~\ref{t:THLs1}.
\begin{table}[ht]
\label{t:THLs1}
\begin{center}
\begin{tabular}{ccccc}
\hline
& Noise Floor & Noise edge & Threshold Set & Threshold Set\\
&(DAC steps) & (DAC steps) & (DAC steps) & (electrons)\\
\hline
D04-W0015 & 460 & 438 & 400 & $\sim$1520 \\
\hline
\end{tabular}\
\caption{Planar Sensor DUT Threshold settings. The nominal threshold in
electrons is given by the difference between the set threshold and
noise floor, multiplied by the expected 25.4 electrons per DAC step~\cite{LlopartCudie:1056683}.}
\end{center}
\end{table}
In order to correctly interpret the DUT data in the analyses which follow, a charge
calibration of the planar sensor DUT was performed.
The first aim of this charge
calibration was to verify that the threshold
indeed corresponded to the nominal value
shown in Table~\ref{t:THLs1}. The
second aim of the charge calibration was to be able to obtain the
relationship between raw
ToT values and charge deposited in each
pixel. As explained in Section~\ref{s:timepixchip}, the response of the Timepix is in principle a linear function
of the input charge. In practice, however, there is a deviation from
this behaviour, giving a global offset for each hit pixel with a
value dependent on the threshold. In addition, there exists a strongly
non-linear behaviour for charges within about 3000~e$^-$ of the
threshold. This behaviour was first described in~\cite{medipix-web},
where a so-called ``surrogate function" was used to parametrise the
response.
In the case of the detection of minimum ionising particles, with thresholds similar to those used in the testbeam described here, the offset is positive, causing the raw counts to overestimate the charge deposited per pixel. Also, the non linear behaviour close to threshold affects the apparent charge sharing distributions between pixels.
In addition there is a pixel-to-pixel
variation, which the calibration can in principle detect, and correct for.
The calibration was performed with testpulses and cross checked with
data taken with a radioactive source. The method used has not been
previously described, and so is explained in detail here.
For the 3D sensor DUT, a detailed charge calibration was not
performed, however the conclusions from the calibration of the planar
sensor DUT can be extrapolated to this device to infer the operational
threshold.
The first part of the calibration proceeded using the testpulse
function available in the Timepix and controlled via the
Pixelman interface.
The analog input of each pixel
features a built-in capacitor of $7.5$~fF~\cite{timepix},
and a step voltage of 1~mV
therefore corresponds to
46.9~e$^-$ injected charge~\cite{timepix}. The voltage of the
pulse sent to the chip is generated by an analog
multiplexer\footnote{Part reference MAX4534, manufactured by Maxim
Integrated Products, Sunnyvale, CA 94086, USA}
mounted on the board, which switches the output from a reference
voltage. The precise relationship between the programmed voltage and
the voltage received by the chip varies according to the USB readout system linked
to the chip in question,
and must be checked by measurement at the output of the multiplexer.
For the planar sensor DUT the following relationship is obtained:
\begin{equation}
H =H_{\rm NOM}*0.972- 26.7,
\label{eq:tpvalue}
\end{equation}
where $H$ is the true test pulse height in mV, and $H_{\rm NOM}$ is
the nominal test pulse height.
In the ideal case, the calibration would now proceed by sending one testpulse per shutter to each pixel, and measuring the ToT and efficiency numbers for each testpulse height. In the case of the Timepix chip this is not possible, as the first testpulse sent gives a lower value than expected in the pixel array, an effect which is thought to be due to the settling time of the DACs in the setup used.
This effect can be washed
out by sending a large number of testpulses, however in this case the efficiency (``counting'') measurement
and the ToT measurement must be done in two separate steps, as described below. In addition note that
it is important to send the testpulses one at a time to individual pixels within an 8 by 8 array, with a suitable
wait period between each testpulse, or the results can be misleading.
The response was first studied in
counting mode, in which case no ToT information is provided by the chip. One thousand
test pulses were sent per frame, and the mean number of counts per pixel measured.
Three thresholds were studied, at DAC values of 380,
400, and 420, corresponding to nominal thresholds of roughly 2030,
1520, and 1020 electrons. For each pixel, a curve can be constructed
of the number of counts (efficiency) as a function of the programmed
testpulse voltage. Fitting an error function to the curve gives the
threshold for each pixel at the 50~$\%$ efficiency point, as well as
the noise of that pixel. Figure~\ref{fg:efficiency_1} shows
the average efficiency of each point, with error bars indicating the
RMS spread, together with the average of the
65k curves derived and fitted in this way. The programmed
testpulse values are corrected to the true values using Equation~\ref{eq:tpvalue} and finally to electrons using
the 46.9e$^-$ per mV scaling given above. In this way the average thresholds
corresponding to 420, 400, 380 DAC counts were determined to be 1030, 1550, and
2050~e$^-$ respectively. The average pixel noise was measured to be
122 electrons, with an RMS spread of about 3 electrons. These
calibrated threshold values can be seen to be very close to the
nominal ones.
The CPU time
involved in performing these fits is considerable, however it is also possible
make a fast derivation of the average
thresholds by plotting the average efficiency of all pixels
a function of applied testpulse voltage and fitting with a single
error function. This gives results within ten electrons of
the individual pixel fits, however it should be noted that in this
case the measured noise is a combination of the individual pixel noise
and the threshold spread, which contributes about thirty electrons in
quadrature to the measured spread. These
data can also be used as an indirect cross check of the testpulse
capacitor value, due to the fact that each DAC step is expected to
correspond to 25.4~e$^-$, as reported in~\cite{LlopartCudie:1056683} from a
measurement with sources of a different Timepix chip. The differences
in thresholds between the three measurements 20 DAC steps apart
can be seen to correspond very well to this expected value. Finally,
note that at the lowest threshold an efficiency rise is visible at
very low values of applied testpulse voltage. Due to the offset shown
in Equation~\ref{eq:tpvalue}, low values of applied testpulse
voltage become negative. Each testpulse generates negative and
positive pulses in the pixel array, corresponding to the rising and
falling edge of the testpulse. When the testpulse reverses sign, the
pixels show a response corresponding to the oppositely signed edge,
and this response is mirrored about the 26.7~mV offset given in Equation~\ref{eq:tpvalue}.
\epspicz{1}{1}{efficiency_1}{Average efficiency curves
obtained with the counting mode calibration. The points show the
experimental data, with the applied testpulse voltage on the x
axis, and the error bars indicate the
RMS scatter. The curves
are the averaged curves of the individual pixel fits, as described
in the text.}
The second part of the calibration proceeded with the chip in ToT mode, in order
to derive the relationship between the input charge and the ToT
value. Note that in this mode the counting function is not available, so the efficiency curves described above cannot be simultaneously extracted.
We applied fifty testpulses per shutter and plotted the resulting ToT
value of each pixel, divided by fifty, as a function of applied testpulse voltage.
As described in \cite{medipix-web} the curve is well described with
the following ``surrogate function'':
\begin{equation}
{\rm
f(x) = ax+b-\frac{c}{x-t}},
\end{equation}
where $\rm{a}$ and $\rm{b}$ represent the slope and intercept of the line describing the behavior well above
threshold, while $\rm{t}$ and $\rm{c}$ parameterize the non-linear behavior close to threshold~\cite{surrogate-function}.
In practice, electronic noise, in combination with the fact that fifty
pulses are sent per shutter, smoothes the sharpness of this
turn-on. We therefore fitted the measured calibration data with
a convolution of the surrogate function with a Gaussian whose $\sigma$ is constrained to the one extracted for
the corresponding pixel from the efficiency curves derived in
counting mode. Figure~\ref{fg:gain_convolute_2} shows the
calibration curves of the DUT for the three thresholds
studied. Each data point is averaged over the 65k pixels.
Near the origin of the graph one can see the edge of the response for negative
pulses. It is also possible to extract the so called
``minimum detectable charge'', corresponding to the intercept of the
surrogate fits with the \textbf{x} axis, corrected according to
Equation~\ref{eq:tpvalue} and converted into electrons.
This number is close to the
threshold values quoted above, but is expected to be slightly lower due to the
fact that many pulses per shutter are sent.
Table~\ref{tab:fitpara}
shows the averages of the individual pixel parameters describing the fitted
surrogate functions. The parameters take into account the correction
described in Equation~\ref{eq:tpvalue}.
\begin{table}[hbt]
\begin{center}
\caption{Summary of the surrogate function fit parameters from the ``ToT'' testpulse scans
at three different thresholds, averaged over the individual pixel fits}\label{tab:fitpara}
\begin{tabular}{ccccccc}
\hline
THL~~ & $a$ & $b$ & $c$ & $t$ & Minimum Detectable Charge (e$^-$)\\\hline
420~~ & 0.194 & 34.4 & 195 &15.3 & 957 \\
400 ~~& 0.201 & 24.8 & 174 & 25.9 & 1480 \\
380 ~~ & 0.206& 17.6&150 & 36.4 & 1980 \\ \hline
\end{tabular}
\end{center}
\end{table}
As in the case of the counting
mode measurements, the average parameters can also be obtained in a
fast way by fitting the same function to the curve of the average ToT value as a
function of testpulse. The numbers obtained are very close to the
numbers shown in Table~\ref{tab:fitpara}.
Apart from the threshold behaviour, the main characteristic exhibited by these fits is a positive offset of the linear part of
the curve, which results in a strong difference seen in the pixel spectra depending on the number of pixels per cluster.
The calibration data can be used to calculate the detected charge in
units of electrons, given a measured ToT, using the following
``inverse surrogate function'':
\begin{equation}
{\rm
q_{in}(e^-)=46.9*\frac{t*a+ToT-b+\sqrt{(b+t*a-ToT)^2+4*a*c}}{2*a}.
}
\end{equation}
which simplifies in the linear regime (above the threshold region) to:
\begin{equation}
{\rm
q_{in}(e^-)=46.9*\frac{ToT-b}{a}.
}
\end{equation}
The parameters used can be those of the individual pixels, or the
average fitted curves. In the latter case, there is an additional
spread of less than 5$\%$, for charge deposits around the value
expected from a minimum ionising particle.
As a cross check of the studies described above, the response of the DUT to a $^{55}$Fe source was measured,
which features a 5.9~keV X-ray line, producing a charge deposition of 1640~e$^-$ in Silicon. The DUT was exposed
to this source and the response measured with the device configured in counting mode is shown in Figure ~\ref{fg:ironformarina}.
The main curve shows the measured efficiency curve, while the insert shows its derivative. The latter curve exhibits a
clear peak at a threshold of 403 DAC counts. From this study it can be inferred
that a threshold of 400 corresponds to 1560~e$^-$,
consistent with the result obtained from the test pulse calibration.
\epspicz{1}{1}{gain_convolute_2}{Average test pulse response of the DUT in ToT mode, for three different thresholds. The points represent show the
pulse response averaged over the 65k pixels in the DUT, as a function of the programmed testpulse voltage, with error bars
indicating the RMS scatter. The curves represent averages of the individual pixel fits to the surrogate function, as described
in the text. The I$_{\textrm{krum}}$ setting, which determines the conversion gain, was 5.}
\epspicz{1}{1}{ironformarina}{Average efficiency curve obtained by operating the Timepix ASIC in counting mode exposed to a $^{55}$Fe source in counting mode.}
\section{Treatment of telescope data}
\label{s:datatreatment}
\subsection{Clustering algorithm}
\label{s:clustering}
\subsubsection{Cluster finding}
A cluster is defined as a group of adjacent hits surrounded by empty pixels. For two hits to be considered adjacent, they must simply
have a difference of $1$ in their \textbf{x} or \textbf{y}
positions.
The cluster maker works by making a cluster out of the first hit on the sensor and recursively searching for adjacent hits; each time
an adjacent hit is found it is added to the cluster and the search for adjacent hits repeated. The search stops when no
further adjacent hits can be added to the cluster.
The position of the cluster in local coordinates is assigned by making
a charge weighted average of the individual pixel positions in the \textbf{row}
and \textbf{column} directions.
\subsubsection{Telescope cluster properties}
The distribution of ToT values in the individual pixels, and the
distribution summed over the clusters, are shown in
Figure~\ref{fg:simpleclustercleaning} for a typical telescope plane.
There is a small amount of noise below the Landau-like cluster peak
which occurs due to interactions from showers in the North Area pion beam; if
it is required that clusters are within 100~$\mu$m of a reconstructed track (where that device is
excluded from the track fit) this noise is removed, as can be seen on
the right hand plot.
In Figure~\ref{fg:landaus} the total cluster charge of
one, two, three, and four pixel clusters is shown for a sample telescope plane, angled at
$9^\circ$ in the horizontal and perpendicular directions.
It can be seen that the ToT of the clusters increases incrementally with each additional pixel, with most
probable values for this plane of 119,145,168, and 203 ToT counts for the
one, two, three, and four pixel clusters respectively.
This effect is due to the behaviour of the Timepix chip which gives a
ToT value which is proportional to the charge plus an offset, and
a turn-on behaviour in the threshold region, as discussed in
Section~\ref{s:calibration}. A more detailed analysis of the
cluster characteristics was performed for the DUT and is described in
Section~\ref{s:analysisofdeviceundertest}.
For the purposes of the telescope performance
this effect is not expected to have a large impact on the pointing
resolution, and for the analyses described in this paper it remains
uncorrected in the telescope data.
For the purposes of the tracking, the candidate telescope clusters have to pass only one cut, namely that the ToT is greater than 80 and less than
350 counts.
\epspicz{1}{1}{simpleclustercleaning}{The signals in individual pixels (dashed), and the signals in clusters (solid)
after the clustering algorithm, shown on the left plot for a sample telescope plane.
There is a small amount of noise below the Landau peak around 50. If the clusters are required to be within
100~$\mu$m of a track this noise is removed, as shown on the right
hand plot. The curves are normalised to unit area.}
\epspicz{1}{1}{landaus}{ToT of clusters from a sample plane in
the telescope. The curves are normalised to each other for
illustrative purposes. The difference in the MPV between each
population is due to the positive offset introduced in each pixel by
the Timepix gain curve.
The relative population of one, two, three, and four pixel clusters is dependent on the device angle and is discussed in Section~\ref{s:analysisofdeviceundertest}. The
solid black vertical lines show the positions of the applied cuts.}
\subsection{Tracking}
\label{s:patternrecognition}
The track reconstruction proceeds in two stages, a pattern recognition
followed by a track fit. A global event cut of 5000 hits
in the whole dataset is applied to reject a very small sample of
saturated events.
A reference sensor is chosen towards the centre of the telescope and the global \textbf{x} and \textbf{y} of clusters in the
other planes are compared with this. The resulting distribution for all clusters is shown in Figure~\ref{fg:inspectglobalalignment}
for an adjacent plane and for the most remote plane. It can be seen that even in the most remote plane, the track related clusters
are all contained within a window of $\pm 100$~$\mu$m. Clusters are
selected within this window, and it is required that every track contains a cluster from every telescope plane.
As the occupancy is around 0.2$\%$ the background picked up by this procedure is negligible.
The clusters are then fitted with a straight line track fit, which takes into account the full rotation matrix of
each pixel plane and is able to assign individual errors to each plane. For the purposes of this paper one error is
applied to the Timepix planes and another to the Medipix2 planes. Most data were taken with a fairly diffuse beam
covering the full surface of each sensor. Due to alignment effects there is an acceptance of about $85\%$ in the
telescope. In a typical run of 1000 frames, there are about 100k clusters reconstructed in each plane, and approximately
64k tracks. Note that due to time alignment effects with the edges of
the SPS spills some frames are empty.
Typically the number of tracks reconstructed per spill was about 750.
\epspicz{1}{1}{inspectglobalalignment}{The global \textbf{x} and \textbf{y} of clusters in a given telescope plane compared to the \textbf{x} and \textbf{y} of clusters
in the reference plane. The left hand plots show these for a telescope plane next to the reference, the right hand plots show them for
the telescope plane farthest from the reference.}
\subsection{Telescope alignment}
\label{alignMM}
The telescope was not surveyed prior to the testbeam due to the fact that the chips were not precisely placed on
the chipboards, which were not designed for this purpose. Instead, a software alignment was used,
with the only input being the rough \textbf{z} positions and inclinations of the planes. The coordinate system is the
right handed one introduced in Section~\ref{s:telescope}: \textbf{x} is perpendicular to the beamline, parallel
to the floor, \textbf{y} is perpendicular to the beamline, perpendicular to the floor, and \textbf{z} is along the beamline.
The free parameters adjusted for the alignment are then
(\textbf{x},\textbf{y},\textbf{z}, $\theta_{\rm{\textbf{x}}}$, $\theta_{\rm{\textbf{y}}}$ {\rm{and}} $\theta_{\rm{\textbf{z}}}$) where
$\theta_{\rm{\textbf{x},\textbf{y},\textbf{z}}}$ are rotations about the \textbf{x},\textbf{y},\textbf{z} axes. Physically the planes of the telescope are positioned
with the sensors away from the beam, rotated about the \textbf{x} and \textbf{y} axes by $9^\circ$, e.g. (0,0,\textbf{z},$9^\circ$,$9^\circ$,0).
A multistep offline procedure is used to refine the sensor positions. The first step roughly aligns the planes
by minimising the global \textbf{x} and \textbf{y} positions of the clusters in each plane relative to a reference plane; allowing
only \textbf{x}, \textbf{y} and $\theta_{\rm{\textbf{z}}}$ to vary. With the sensors now roughly aligned, tracks are fitted to the clusters,
only accepting tracks with a slope less than 0.005 radians. A $\chi^2$ is formed for each track from the sum of
the residuals divided by the expected error in each plane. The minimisation proceeds by varying the positions of
all telescope planes individually in all coordinates except \textbf{z}, and iterating three times. In a final step, the procedure is repeated but
the telescope planes are also allowed to move in \textbf{z}. New tracks are produced
and extrapolated to the DUT position. The DUT is always excluded from the pattern recognition and the track fitting,
to ensure a completely independent measurement. In the final step of the procedure the DUT is aligned by selecting
all clusters within a cut, typically 100~$\mu$m, and adjusting all 6 free parameters of the DUT. The quality
of the alignment can be probed by plotting the means of the resulting unbiased residuals at the DUT as a function
of various parameters. The most sensitive parameters are found to be the orthogonal coordinate of the track to the
residual plotted. The distribution is shown in Figure~\ref{fg:alignmentquality}
for two sample runs with the DUT at angles of -2$^\circ$ and 8$^\circ$: these distributions are
typical. The variation of the mean of the residual across the width of the chip is below $\pm$1~$\mu$m; for
most of the residuals measured the effect is negligible, and for the most precise measurement angles of the DUT
the resolutions extracted in Section~\ref{s:analysisofdeviceundertest} are approximately 0.2~$\mu$m
worse than what could be achieved with an optimal alignment.
\epspicz{1}{1}{alignmentquality}{An illustration of the quality of the alignment for two different runs. The unbiased residual at
the DUT is shown, where the DUT is excluded from the pattern recognition and from the track fitting.
The variation of the mean of the residuals across the length and height of the chip is illustrated, for
a run close to perpendicular, where the DUT resolution was close to 9~$\mu$m, and a run close to the
optimal angle, where the DUT resolution was close to 4~$\mu$m.}
\subsection{Track pointing precision}
\label{s:pointingresolution}
After the pattern recognition a straight line track fit is performed on the clusters selected in the telescope.
For the majority of the analyses, it is required that each track has a cluster from each of the six planes.
Due to the fact that the planes equipped with Medipix2 sensors are expected to have lower resolution than the Timepix
planes, the track fit takes into account the errors in order to weight the clusters appropriately in the fit.
The distributions of the biased residuals are studied to extract the resolution of the individual planes and the
pointing resolution of the telescope at the position of the DUT. Simplifying the expression such that the telescope
planes are arranged symetrically around $\rm{\textbf{z}}=0$, it can be
shown, following the formalism developed in~\cite{Papadelis:1186697}, that the predicted position of the track at plane $n$ is given by
\begin{equation}
{\rm {\rm y_{pred(n)}} = \displaystyle\sum_i y_i \cdot A(z_{n,i})},
\end{equation}
where
\begin{equation}
{\rm A(z_{n,i}) = \frac {\displaystyle\sum_j \frac{(z_j)^2 + z_n \cdot z_i}{ (\sigma_j)^2 \cdot (\sigma_i)^2}}{\displaystyle\sum_j \frac{1}{\sigma_j^2} \displaystyle\sum_j \frac{z_j^2}{\sigma_j^2}}},
\end{equation}
and $\rm z_n$ is the \textbf{z} position at plane $n$, and ${\rm \sigma_n}$ the error assigned to plane $n$.
The expected track pointing error is then given by:
\begin{equation}
{\rm {\rm \sigma_{pred(z_n)}} = \sqrt{\displaystyle\sum_i (\sigma_i \cdot A(z_{n,i}))^2}}.
\end{equation}
The quality of the residuals in the telescope is illustrated in Figure~\ref{fg:telescoperesidualsforpaper},
which shows an example of the fitted residuals in the \textbf{x} coordinate for the six
telescope planes. It can be seen that the shapes are
quite gaussian. The biased residuals extracted in each plane in
\textbf{x} and \textbf{y} for a typical run are shown in
table~\ref{t:biasedresiduals}.
The distribution of the biased residuals can be used together with the formulas above to extract
the expected pointing resolution at the centre of the telescope.
\begin{table}[ht]
\begin{center}
\begin{tabular}{lcccccc}
\hline
Device name & C03-W15 & K05-W19 & D09-W15 & M06-W15 & I02-W19 &
E05-W15 \\
\hline
$\sigma_{\textbf{x}}$ (${\rm \mu m}$) & 3.3 & 9.4 & 3.9 & 3.8 & 9.3
& 3.2 \\
$\sigma_{\textbf{y}}$ (${\rm \mu m}$) & 3.4 & 9.4 & 4.0 & 3.8 & 9.2
& 3.2 \\
\hline
\end{tabular}
\caption{Biased residuals measured in the telescope for a typical
run. Variations between runs are of the order of less than 0.1~$\mu$m.}
\label{t:biasedresiduals}
\end{center}
\end{table}
There is an additional error introduced by multiple scattering, although the impact is small enough that
it is still acceptable to perform a straight line fit. Each station contains roughly 2$\%$ in radiation length of material,
made up of the 700~$\mu$m thick ASIC, the 300~$\mu$m thick sensor, 1.6~mm of epoxy in the PCB and a
total of 87~$\mu$m of copper layers in the PCB. The effect of this is assessed by simulating the setup, applying an
extra scattering to the points, and performing a straight line fit. The result of the simulation is shown in
Figure~\ref{fg:simulationforpaper}. The simulation gives a reasonable fit to the data, and predicts that the pointing
resolution at the position of the DUT is 2.3~$\mu$m. The dominant uncertainty on this number comes from the
fact that the data are not a perfect match to the simulation. This was investigated by varying the resolution of the
telescope until the simulation matched perfectly the internal and external planes. The biggest deviation induced in
the pointing resolution was 0.1~$\mu$m. Additional uncertainties coming from remaning misalignments and the
lack of perfect knowledge of the material of each plane are negligible in comparison.
\epspicz{1}{1}{telescoperesidualsforpaper}{Fitted residuals in the \textbf{x} coordinate for the six telescope planes. The middle column shows the
residuals in the two Medipix2 sensors. Each plane is angled at
$9^\circ$ in the horizontal and perpendicular direction, and all
telescope planes are included in the track fit, with weighted errors.}
\epspicz{1}{2}{simulationforpaper}{The simulation of the straight line track fit in telescope planes 1-3 (downstream of the device under test)
and planes 5-7 (upstream of the device under test) shown as a solid line, and compared to the biased residuals measured
in the data for the \textbf{x} (open circles) and \textbf{y} (open squares) projections.}
\section{Analysis of the DUT}
\label{s:analysisofdeviceundertest}
\subsection{Description of datasets}
In total about 200 runs were taken over a two week data taking period,
each lasting about 50~minutes. The majority of the runs had 1000
frames, with the shutter open for
0.01~s, and a mean number of tracks of approximately 70 per frame.
Nominal settings were defined, with a threshold DAC of 400 and Ikrum 5,
and in general one parameter was varied at a time during each investigation.
The beam intensity and the shutter open time were varied for
specialised investigation of particular effects, such as alignment or
time-walk. After major telescope interventions, and before taking any
data set where the DUT was nominally
perpendicular to the beam, a quick ``angle scan'' was taken. The
cluster widths were analysed online in a similar way to that described in
Section~\ref{s:perpendicularpoint} to ensure
that the DUT was indeed close to perpendicular.
The principal datasets taken with the planar sensor:
\begin{itemize}
\item{\textbf{Angle scans:}
Using the motion stage the DUT was rotated about the \textbf{y} axis, causing
the clusters to spread in the \textbf{x} direction along the pixel columns.
Approximately 30 different angles were taken, concentrating on the
region between zero and the optimal angle of around 9$^\circ$, which
is most relevant for LHCb, increasing to a largest angle of 18$^\circ$, and including negative angles to investigate the symmetry
around the perpendicular point. As a final step the DUT was rotated
by 45$^\circ$ about the \textbf{z} axis, and a further angle scan performed, in
order to investigate the resolution of the sensor between pixels
offset by one in both the column and row direction, with a constant
telescope extrapolation precision.}
\item{\textbf{Shaping scans:}
The I$_{\textrm{krum}}$ parameter was varied from 5 up to 80 in order to investigate
the effect of the shaping time on the cluster landau distributions and
the DUT resolution.}
\item{\textbf{Threshold scans:}
The threshold DAC was varied from a lowest value of around 750 electrons through
to the highest possible value of around 6000 electrons. These data were used to
investigate the pulse shape and the DUT efficiency. A similar scan
was performed with the beam off, to investigate the noise as a
function of threshold DAC.}
\item{\textbf{HV scans:}
The HV was varied between 5 and 200~V to investigate the cluster
characteristics resolution of the DUT. These scans were performed at
three different angles: $0^{\circ}$, $10^{\circ}$, and $18^{\circ}$.}
\item{\textbf{Time-walk:}
In order to investigate the time-walk the shutter open time was
decreased to 250~$\mu$s and long data runs were acquired with the Timepix in ToA mode.}
\item{\textbf{Specialised Runs:}
A number of specialised runs were taken to cross check various
effects.
These included runs with electrons, runs with the DUT in counting mode
to check the binary resolution performance of the sensor and sparse
data runs for fine tuning the alignment.}
\end{itemize}
For the 3D sensor a subset of these datasets were taken, principally
angle scans in the normal direction, and data at three different bias
voltages.
\subsection{Establishing the orientation of the DUT in the beam}
\label{s:perpendicularpoint}
Although the rotation stage can move and reproduce the angle of the DUT with very high accuracy
(see Section~\ref{s:telescope}), there is an uncertainty on the absolute calibration of the angle
due to the uncertainty of placing the entire telescope perpendicularly in the beam line.
Even though the translations and rotations of the planes relative to one another can be determined very
precisely with the alignment, the overall angle with respect to the beam remains a weakly constrained parameter.
For this reason a wide range of angles were scanned on both sides of the nominal perpendicular position of the DUT.
Cluster characteristics independent of the tracking and alignment such as row width, column width, and fraction of 1 pixel
clusters were then used to determine by symmetry arguments the position of the true perpendicular point relative
to the nominal perpendicular point of the rotation stage. In order
to improve the accuracy of these measurements the cluster cleaning
described in Section~\ref{s:clustering} was applied.
\epspicz{1}{1}{RatioFractionAngle}{The left plot shows the
distribution of the ratio of row width to column width in the $55 \times 55$~$\mu$m$^2$ planar sensor DUT; right plot shows percentage of
various sizes of clusters as a function of nominal angle.}
\begin{table}[ht]
\begin{center}
\begin{tabular}{cc}\hline
Scan & Minimum Angle\\ \hline
\hline
Cluster Size & -0.205\\
\hline
Cluster Row Width & -0.306\\
\hline
Fraction of 1-pixel Cluster & -0.273\\
\hline
Ratio (Row/Column) & -0.366\\
\hline
\end{tabular}
\caption{The DUT angle at which each quantity is minimized.}
\label{t:MinAngles}
\end{center}
\end{table}
Distributions of these quantities as a function of nominal angle are
shown in Figure~\ref{fg:RatioFractionAngle}.
The plots show the ratio of the row width of the clusters (i.e. size in rows) to their
column width (i.e. size in columns), and the fraction of various sizes of clusters as a function of angle. The plots show clear
minima and maxima around the perpendicular position, and by fitting with polynomials the true position of the perpendicular can
be extracted.The fits are performed in the central region of $\pm 5^{\circ}$, and
the results of the fits are shown in Table~\ref{t:MinAngles}. The dispersion of the beam was measured to be $0.007^{\circ}$ in the
direction of rotation in the stage, so adds a negligible uncertainty to the determination of the DUT angle.
From these studies, the angle of the beam relative to
the nominal perpendicular point of the rotation stage was
\begin{equation}
{\rm {\rm \theta_{\rm{true~angle}}} = \theta_{\rm{nominal~angle}} -0.29^{\circ} \pm 0.06^{\circ} },
\end{equation}
and this correction is applied in subsequent plots.
The error on this angle comes from the
errors on the polynomial functions fitted to Figure~\ref{fg:RatioFractionAngle}, in which
the errors on the individual measured points are around the permille level and therefore negligible.
Essentially the entire error comes from the limited number of angles for which the measurement was performed.
\subsection{Extracting the resolution of DUT}
\label{extractingresolution}
In the sections which follow the resolution of the DUT is quoted for
varying conditions (track angle, HV). The resolution is always
extracted in the same manner. The tracks are fitted through the
telescope, as described in Sections~\ref{s:patternrecognition} and
Section~\ref{s:pointingresolution} with the DUT excluded from the
pattern recognition and from the track fit. The tracks are
extrapolated to the plane of the DUT and a global residual formed
between the tracks and all clusters in the DUT.
In order to investigate the resolution in \textbf{x}, clusters with a
residual of less than 100~$\mu$m in \textbf{y} are selected, and vice versa.
The resulting histograms are fitted with single Gaussians. The
resolution is extracted by subtracting in quadrature the 2.3~$\mu$m
contribution from the track pointing precision.
The error on the precision was estimated by dividing the datasets into two and comparing the resolution,
by varying the binning of the histograms, and by varying the cut on
the \textbf{y} of the cluster.
In addition there is slightly larger error for the perpendicular fits,
where the histograms become less Gaussian due to the more binary
nature of the resolution.
It should be noted that the 0.1~$\mu$m error on the track pointing
resolution is not
the dominant contribution to the error on the resolution measurements, which are
in all cases 4~${\rm \mu m}$ or above.
\section{Cluster characteristics of the sensors}
\label{s:clusterchars}
Some basic cluster distributions are shown in Figure~\ref{fg:clustercharacteristics}
for tracks in the DUT at perpendicular incidence and for a bias voltage of 100~V in the DUT.
Most of the clusters are formed from four pixels or less with a tail to larger clusters.
The distribution of charge in the individual and in clustered pixels is shown before and after
the charge calibration is applied. The sharp peak in the individual pixel charge as well as the
importance of pixels with a low amount of charge is apparent. The track intercept point for
associated clusters are shown as a function of position within the pixel cell for one, two, three, and four
pixel clusters in Figure~\ref{fg:clusterpositions}. The four pixel clusters, which are slightly
favoured over three pixel clusters, occupy the corners of the pixel cell, with the three pixel clusters
slightly inset.
\epspicz{4}{5}{clustercharacteristics}{Cluster characteristics
in the 300${\rm \mu m}$ thick planar sensor DUT, at perpendicular
incidence. From top to bottom:
Number of pixels per cluster, charge in individual and clustered
pixels before pixel calibration, charge in individual and clustered
pixels after charge calibration. The individual pixel charge has a
peak close to the expected minimum ionising particle deposited
charge, corresponding to one pixel cluster, and a long tail
extending to the threshold cut-off corresponding to clusters with a
large amout of charge sharing.}
\epspicz{1}{1}{clusterpositions}{Cluster position
distributions within the pixel cell. From left to right, the track
incident point is shown in the case of associated one, two, three, and four pixel clusters.}
\subsection{Planar sensor Landau distributions}
\label{planarsensorlandau}
The extensive literature related to predictions of energy loss in thin silicon
detector is reviewed by Bichsel in \cite{bichsel}. It is well known that to
describe the width of the observed straggling function the effects of
atomic binding must be included. Several methods have been proposed
to do this. One approach is to make a n-fold convolution of the single collision
cross-section derived from data on optical absorption and other
measurements. An example of such an approach is discussed in detail in
\cite{bichsel}. No analytic form for this function exists. However, in
a subsequent paper Bichsel has provided tabulated data for the
cumulative probability distribution as a function of $\beta \gamma$
which allows the expected distribution to be generated using a Monte
Carlo simulation.
An alternative approach are the so called 'mixed' methods. In these a
straggling function, obtained from a simple model of the
collision spectrum, is convolved with a distribution that accounts for
the atomic binding \cite{blunck}, \cite{hall}, \cite{hancock}. The simplest example of
this method is to convolve a Landau distribution with a Gaussian. The
resulting function is sometimes referred to as the Blunck-Leisegang
function in the literature. The Landau MPV and scale $\xi$ are constrained by the theoretical relation
\begin{equation}
MPV (keV) = \xi [12.325+\ln{\frac{\xi}{I}}],
\label{eqMPV}
\end{equation}
where
\begin{equation}
\xi = 0.017825 \times t\ (\mu{\rm m}),
\label{eqxi}
\end{equation}
and $t$ is the detector thickness in micron \cite{bichsel}. These formulae include
density effects which lead to an $8\%$ increase in the MPV for highly
relativistic particles compared to a MIP of 77.6~keV. The $\sigma$ of the
convolution Gaussian is predicted to be \cite{blunck}, \cite{hancock}
\begin{equation}
\sigma_{B} = 5.72 keV.
\label{eqsigma}
\end{equation}
Bichsel \cite{bichsel} has questioned the
validity of this approach and shown that depending on the assumptions
made in the calculation values of $\sigma_{B}$ up to a factor larger
can be obtained. To obtain good agreement between the Blunck-Leisegang and
Bichsel functions $\sigma_{B}$ should be set to 7.3~keV.
In Fig.~\ref{fig:bichs} the charge collection distribution for all cluster sizes is compared to the
results of three fits:
\begin{figure}[htb]
\centering
\label{fig:bichs}
\includegraphics[width=0.7\linewidth]{bichs.eps}
\caption{Distribution of collected charge after charge calibration. The results of the fits
described in the text are superimposed.}
\end{figure}
\begin{itemize}
\item A fit to the Bichsel function described above. Since the model
itself has no free parameters only the overall normalization is left
free in the fit. Since the tail of the distribution is expected to
be dominated by the effect of $\delta$-rays and not well described
by theory it is chosen to normalize to the height of the distribution.
\item A fit to a Landau convolved with a Gaussian with the parameters MPV, $\xi$
and $\sigma_B$ fixed to the values given by Equation \ref{eqMPV} - \ref{eqsigma}. As in the fit to the Bichsel function only the
normalization is left free.
\item A parametric fit to a Landau convolved with a Gaussian with all
parameters left free. The resulting fitted parameters, compared with the predicted values from
Equations~\ref{eqMPV}~to~\ref{eqsigma} are displayed in Table~\ref{tab:landau}.
\end{itemize}
Additional broadening effects due to electronic noise and
imperfections in the gain calibration are small and can be ignored. To
convert the theoretical estimates from keV to electrons a value of
3.62~eV for the energy necessary to create an electron-hole pair in Si
\cite{canali} was used. It can be seen that though the Bichsel
function describes the shape of the core of the distribution well it
underestimates the MPV by around $3 \%$. The tail of the distribution
is relatively poorly described. On the other hand the Blunck-Leisegang
function agrees well in terms of the MPV but underestimates the observed
width of the distribution. To describe properly the $\sigma_{B}$
of the Gaussian needs to be increased from 1580 to 1830
electrons. This is in agreement with the value for $\sigma_{B}$
proposed in \cite{bichsel}.
\begin{table}
\begin{center}
\caption{Predicted and fitted Landau parameters.}\label{tab:landau}
\begin{tabular}{ccc}
\hline
Parameter& Predicted value [$e^{-}$] & Fitted [$e^{-}$] \\ \hline
MPV & 23400 & 23410 $\pm 10$ \\
$\xi$ & 1690 & $1820 \pm 10 $ \\
$\sigma_{B}$ & 1580 & $1830 \pm 10$ \\ \hline
\end{tabular}
\end{center}
\end{table}
Studies have also been performed dividing the data according to the
cluster size. Fits to the parametric form described above are shown in
Fig~\ref{fig:clustersizefit}. The MPV returned by the fit are given in
Table~\ref{tab:clussizetab}. It can be seen that the MPV increases slightly with
increasing cluster size. This is most likely explained by the relatively high readout threshold DAC of $1500
e^-$ which was applied and means that some charge is lost. This
hypothesis is supported by the fact that if the track information is
used to select one-strip clusters where the impact point point is predicted to be
close to the centre of the pixel, the deposited charge increases from
24070 to $24479 e^-$ in better agreement with the numbers seen for
multistrip clusters.
\begin{figure}[t]
\centering
\label{fig:clustersizefit}
\includegraphics[width=0.45\linewidth]{clusters.eps}
\includegraphics[width=0.45\linewidth]{clusters_norm.eps}
\caption{Distribution of collected charge after charge calibration for different cluster
sizes. The results of fits to the parametric form described in the
text are described. The right hand plot shows the same distributions
but normalized to unit area, and with a zoomed x-axis.}
\end{figure}
\begin{table}
\begin{center}
\caption{Deposited charge for different cluster sizes.}.\label{tab:clussizetab}
\begin{tabular}{cc}
\hline
Cluster size & MPV [$e^{-}$] \\ \hline
1 & 24029 \\
2 & 24145 \\
3 & 24214 \\
4 & 24975\\ \hline
\end{tabular}
\end{center}
\end{table}
In addition, as the cluster size increases the distribution of the
deposited charge is observed to broaden. This can be explained by
the fact that multi-strip clusters are more likely to occur if a high
energy $\delta$-ray is produced.
Finally, it should be noted that if an averaged calibration is applied
to the pixels instead of the individual calibration, there is no significant change to the results.
\subsection{Double sided 3D sensor Landau Distributions}
The Landau distributions in the 3D sensor show a more complex shape due to the geometry of the
structures within the sensor. The raw ToT distributions are illustrated in Figure~\ref{fig:3dlandau} for
tracks at perpendicular and angled incidence. For tracks at perpendicular incidence there is a
clear double peak structure, originating from tracks which pass between the columns and deposit
charge in the full silicon thickness, and tracks which pass through the length of a column and
deposit charge in a small fraction of the thickness below. For tracks passing through the sensor
at larger angles of incidence, the distribution represents a more uniform Landau shape.
In both cases the distribution is seen to be separated from the noise floor,
but closer to the threshold than in the case of the planar sensor. The detailed
characteristics of the Landau distribution are discussed more extensively in the companion paper~\cite{3Dcompanion}.
\begin{figure}[htb]
\centering
\label{fig:3dlandau}
\includegraphics[width=0.45\linewidth]{3DLanGau_0.eps}
\includegraphics[width=0.45\linewidth]{3DLanGau_10.eps}
\caption{Raw ToT distributions in the 3D double sided sensor for perpendicular tracks (left) and tracks with 10 degrees incident angle (right).}
\end{figure}
\section{Efficiency and noise as a function of threshold}
\label{s:hamish}
To measure the efficiency of the DUT, tracks were reconstructed using all six telescope planes.
The intersection point of each track with the plane of the DUT was calculated, and clusters within
0.2~mm of this point were assumed to result from that track. However, if there was another
track within 0.6~mm of the track in question, both tracks were ignored to avoid associating clusters to the wrong track.
The range of $0.2$~mm is at least 10 times the width of most
residual distributions, chosen so it is highly unlikely that genuine clusters were omitted.
\subsection{Planar sensor efficiency and noise results}
The efficiency was calculated as a function of threshold for three different angles of the DUT, and is
shown in Figure~\ref{fg:EffAsFcnTHL_301110}. For most thresholds the efficiency is measured to be $99.5\%$, with a
significant drop at higher thresholds.
At low threshold levels analogue noise can cause individual pixels to be momentarily insensitive to external
signals as the digital part of the pixel only responds to positive transitions across the threshold during the
period when the shutter is open. Thus a high noise rate will increase the chance of the discriminator level already
being above threshold when the shutter is opened, causing the overall efficiency of the device to fall.
The possibility of associating clusters to tracks incorrectly,
and thereby overestimating the efficiency, was considered. The probability of associating a random cluster to a track
where no real cluster existed was evaluated. It dominates the uncertainty in Figure~\ref{fg:EffAsFcnTHL_301110} for
low thresholds, but falls rapidly and becomes negligible at thresholds above 1000~electrons. The plot in
Figure~\ref{fg:EffAsFcnTHL_301110} has statistical uncertainties and this systematic uncertainty added in quadrature.
The noise was also evaluated as a function of threshold for runs taken without beam, and is plotted in
Figure~\ref{fg:NoisevsTHL_041110}. It is high for very low thresholds, as expected, then falls to zero, or almost zero, for most thresholds under test.
\twoepspicx{EffAsFcnTHL_301110}{Efficiency (fraction of tracks
with associated clusters) of the $55 \times 55$~$\mu$m$^2$ planar DUT as a function of threshold.}
{NoisevsTHL_041110}{Noise (mean number of hits per event in runs without beam) as a function of threshold.}
\subsection{3D sensor efficiency results}
As described in Section~\ref{s:3dsensor}, the Timepix chip used for
the manufacture of the the 3D sensor was of a lower grade and had two
non-responding columns (a total of 512 dead pixels). Furthermore, the bump bonding of the sensor was not perfect
and introduced further dead or noisy pixels. For this reason,
extra cuts were applied in the 3D sensor analysis in order to remove
dead pixels. A map of dead and
noisy pixels was produced exposing the sensor for 20 minutes to an X-ray source.
An average of 1000 counts was obtained per pixel, and those pixels
more than four standard deviations from the mean were flagged as dead
or noisy.
This identified a total of 640 pixels (including the two
dead columns). This map was used in the analysis and all extrapolated tracks within
three pixels of a dead or noisy pixel on the 3D sensor were excluded from the analysis.
Furthermore, all extrapolated tracks were required to be within the active
area of the 3D sensor; the extrapolated track position was required to
be seven columns from the edge of the sensor.
The double sided 3D sensor was biased to 20~V. This voltage
ensures that both the inter-column region and the regions beneath the
columns and the front and back-planes of the sensor are biased~\cite{firstDoubleSided3D}. The Timepix chip threshold was set to
approximately 1000 electrons.
Figure~\ref{fg:3DEfficiency} shows the efficiency of the 3D sensor
as a function of the rotation angle of the sensor.
The uncertainties on the efficiencies
were estimated from the mis-identification rate. At a rotation
angle of $10^\circ$ the sensor is fully efficient, with an overall
efficiency measured as $99.8 \%$.
As the efficiency variation across a unit pixel cell of the 3D sensor
at perpendicular incidence is the main topic of the companion
paper~\cite{3Dcompanion}, only a brief description is provided here.
As discussed in the introduction, the 3D sensor has a
series of inactive columns, and for a perpendicular beam no charge would
be collected in the centre of these columns. However, the columns in
the double sided 3D sensor have $50$~$\mu$m of silicon above them
which will collect charge of around 4000 electrons when the sensor is biased at more than
a few volts. As the threshold of the sensor in this study is 1000 electrons,
the columns in the centre of the sensor are expected to be
reasonably efficient, as seen in Figure~\ref{fg:3DEfficiency}, where
the efficiency in a region of $6.25$~$\mu$m radius around the centre columns
is shown. However, for the columns at the corner of a pixel cell, the
charge is shared between neighbouring pixels and the efficiency is
lower, as seen in Figure~\ref{fg:3DEfficiency} for a selection of the
same size region around the corner electrodes. As expected the
remaining region of the cell away from the corners or centre
remains highly efficient. The sensor is aligned using the weighted
time over threshold information in the cluster positions, which leads to the pixel cell
positions being defined at the mid-plane of the silicon.
The non-uniformity effect for the corners and centres is removed as the sensor is rotated, as
seen in Figure~\ref{fg:3DEfficiency}. For this reason it is expected that the 3D sensors
will be tilted when used in a barrel detector, whereas a forward
detector geometry leads naturally to an average track angle.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.5\textwidth]{3DEfficiencyAngle.eps}
\end{center}
\caption{Efficiency of a double sided 3D pixel sensor rotated
around the \textbf{y} axis. The overall efficiency is shown alongside that
for three different regions of each pixel cell: the regions around
the corner and centre electrodes, and the region away from the
electrodes.}
\label{fg:3DEfficiency}
\end{figure}
\section{Angle dependence}
\label{s:anglescans}
\subsection{Introduction}
\label{s:angleintro}
As described in Section~\ref{s:analysisofdeviceundertest} the incident angles of the tracks were
varied in two directions: the ``normal'' direction, in which the chip was mounted perpendicular to the
floor and rotated about the global \textbf{y} axis, and the ``diagonal'' direction, where it was first rotated by 45$^\circ$
about the \textbf{z} axis and subsequently rotated about the global \textbf{y} axis.
The coordinate system is shown in Figure~\ref{fg:marco} and the measurements performed for the planar
and 3D sensor are described in the following sections.
\epspicz{1}{2}{marco}{Coordinate system for rotation in the ``normal'' direction (left figure) and the ``diagonal'' direction (right figure).}
\subsection{Corrections due to non-linear charge sharing}
\label{s:etacorrections}
Due to the fact that the pixel pitch is large compared to the diffusion width of drifting holes in 300~$\mu$m silicon,
the charge sharing between cells is not perfect and the simple weighted charge described in Section~\ref{s:clustering} is not
expected to reproduce accurately the track position. This effect is a strong function of angle. In order to study this effect
the precise pointing resolution of the tracks was used to scan the cluster characteristics across the pixel cells, and the
results are shown in Figures~\ref{fg:clustercontributions}~and~\ref{fg:etadistribution} for a selection of runs where the
DUT was oriented at $0^{\circ}$, $5^{\circ}$, $8^{\circ}$ and $18^{\circ}$ relative to the perpendicular. In Figure~\ref{fg:clustercontributions} the track
position is used to determine the inter-pixel distance, and for each position the contribution of one, two, and three column pixel clusters
is displayed. For the perpendicular tracks there is a contribution of two column clusters from the area centered between the pixels,
and close to the pixel centres only one column pixels contribute. As the angle is increased, the charge sharing region increases,
until an optimum close to $9^{\circ}$. After this point the contribution of three column clusters increases.
By comparing the weighted charge position to the track position the effect can be corrected in the data, as displayed in
Figure~\ref{fg:etadistribution}. Each cluster is used to define the relevant pixel cell, and for this pixel cell the track
position is plotted relative to the weighted charge position. The resulting histogram is fitted with a fifth order polynomial,
and the result is used to correct the weighted charge position in the data.
For angles above $12^{\circ}$ a separate correction is applied to the two column and greater than two column clusters.
This correction, referred to in the rest of this paper as ``eta
correction'', results in an improved resolution, as discussed in Section~\ref{s:resolution}.
\epspicz{1}{1}{clustercontributions}{Contribution of one, two, and three column width pixel clusters in the dut,
for a selection of runs with the DUT oriented at $0^{\circ}$, $5^{\circ}$, $8^{\circ}$ and $18^{\circ}$.}
\epspicz{1}{1}{etadistribution}{Eta corrections to the data. The comparison of the track position to the charge weighted cluster centre
is shown for four sample angles, before and after correction. A separate correction is applied to two and three column width clusters for the 18$^\circ$ plots.}
\subsection{Planar sensor normal angle scans}
\label{s:resolution}
Figure~\ref{fg:resbeforeaft} shows the residual distributions
as a function of incident track angle
together with the contributions of the one, two, and three column pixel
clusters, before and after eta correction.
These data are referred to as ``Normal Angle Scan'' as the track
incident angle is varied across the pixel columns, as opposed to the
diagonal angle scan described in the next section.
It can be seen that for the perpendicular tracks there is a good improvement in the double column clusters
after the eta correction. As the track angle increases the improvement gets smaller, giving negligible improvement
close to the optimal 9 degree angle. For large angles there are two main categories of cluster, two and three column clusters.
The three column clusters give fairly good resolution in the regions closer to the pixels, where the spatial information
is dominated by two external pixel hits. For the tracks centered between the pixels, the clusters are predominantly
two column clusters. The resolution here is slightly worse, as the two columns tend to have a similar pulse height.
For these larger angles the eta correction gives only a modest
improvement.
The resolution of the DUT is extracted from the single Gaussian fit,
and the $2.3 {\rm \mu m}$ contribution of the track
(as discussed in Section~\ref{s:pointingresolution}) is subtracted.
The angle of the track is corrected by $0.29^\circ$ according to the analysis described in
Section~\ref{s:perpendicularpoint}. The raw resolution of the planar sensor DUT is shown for the tilting and
non-tilting direction in Figure~\ref{fg:tiltnotilt}. In the
tilting direction a dramatic improvement is seen as a function of
track angle, with a best resolution of 4.0~$\mu$m for tracks oriented at 8-$10^{\circ}$ and a worst resolution of $\approx$11~$\mu$m.
In the non-tilting direction there is a small
degradation as the angle increases, due to the fact that the information is shared over more
pixels, which may fall below threshold.
The improvement brought by
the eta correction described in Section~\ref{s:etacorrections} is also illustrated in
Figure~\ref{fg:tiltnotilt}.
The best improvement is achieved when the tracks impinge perpendicularly on the
sensor; for the optimum resolution the eta correction does not bring
an improvement.
For the data shown here the full individual pixel gain calibration
described in Section~\ref{s:calibration} has
applied, however note that in terms of the best acheivable
resolution this calibration has only a small impact;
after eta correction the individual pixel gain calibrated data have a
better resolution by approximately 0.2~$\mu$m compared to the
uncalibrated eta corrected data.
The reason for
this lies in the shape of the eta distribution itself: the offset
introduced by the Timepix partially compensates the intrinsic skew in
the eta distribution from the silicon response itself. When this
offset is removed by applying the calibration the resolution worsens. After eta correction, which
absorbs most of the effect, the resolution in both cases is very comparable.
\epspicz{1}{1}{resbeforeaft}{Examples of the residual fitting to the DUT before and after eta correction, for four sample orientations of
the DUT ($0^{\circ}$, $5^{\circ}$, $8^{\circ}$ and $18^{\circ}$ ).
The residuals of all clusters are shown, and the residuals of one, two, and three column clusters are shown separately.}
\begin{figure}[p]
\centering
\includegraphics[width=0.8\linewidth]{tiltnotilt.eps}
\caption{Resolution of the $300~{\rm \mu m}$
thick planar sensor
$55~{\rm \mu m} \times 55~{\rm \mu m}^2$ pixel DUT
as a function
of track angle. The 2.3~$\mu$m track
pointing resolution has been subtracted. From top to bottom : in the non-tilting direction, in the non-tilting
compared to in the tilting direction, before and after eta correction in the tilting direction.}
\label{fg:tiltnotilt}
\end{figure}
\subsection{Planar sensor diagonal angle scans}
Given an equivalent ``normal'' resolution of the pixel device in the \textbf{row} and \textbf{column} direction,
it is expected that the ``diagonal'' resolution, i.e. between pixels displaced by one in column and row,
should also be equivalent. However, in this direction the shapes of the pixel clusters
are different, which could lead to differences in the results given by the charge weighted centre method, and an
eta correction is required in both the \textbf{x} and \textbf{y} direction, which could also affect the result. In addition, the
effective pixel pitch in this direction is greater by $\sqrt(2)$ which can lead to a different rotation angle
for the minimum resolution. This was explicitly checked in the
system by rotating the DUT by $45^{\circ}$ around the \textbf{z} axis,
such that the ``x'' (\textbf{column}) and ``y'' (\textbf{row})
measurements in the telescope were probing the ``u'' and ``v'' directions in the DUT, rotated by $45^{\circ}$
relative to the \textbf{row} and \textbf{column}
directions. For each run, the raw resolution after pixel correction is plotted,
and an eta correction is determined in the same way as in Section~\ref{s:etacorrections} for both the
\textbf{row} and \textbf{column} direction.
The resolution is then redetermined after applying the eta correction. Each measurement is based on approximately
100,000 residual measurements in the DUT. The results are shown in Figure~\ref{fg:diagonaleta}.
The left plot shows
the raw results, with no eta correction applied. The resolution is shown as a function of the changing track angle
in the diagonal direction. For reference the equivalent result obtained for the ``normal'' resolution measured
as a function of track angle in the normal direction, is shown. For perpendicular tracks the diagonal resolution
is equivalent to the ``normal'' resolution, although the shape of the residual histogram (not shown here) is
different, due to the contribution of different cluster shapes. As the track angle increases in the diagonal
direction, there is an improvement seen for normal tracks both in the direction of rotation, but also in the
perpendicular direction, which might not be naively expected. This can be understood by considering the pixel
layout: as the sensor rotates information is picked up from the next, staggered row of pixels, so new information
contributes in both coordinates.
Improvements in resolution as a function of the track diagonal angle are
also seen for the ``normal'' resolutions. It can be seen that the shapes of the curves differ, and in the
diagonal direction the resolutions are not as good as that achieved for the best ``normal'' resolutions.
The right plot shows the same curves after eta corrections
are applied in the local \textbf{row} and \textbf{column} directions. It can be seen that for the diagonal data the eta
corrections give similar improvements to the normal data.
\begin{figure}[htb]
\centering
\includegraphics[width=0.45\linewidth]{diagonalnoeta.eps}
\includegraphics[width=0.45\linewidth]{diagonaleta.eps}
\caption{The resolution in the diagonal and normal directions as a function of
track angle variation in the diagonal direction (left), and the resolution after the eta correction
is applied (right). The resolution improvement is a strong function of the
angle, giving a best improvement of around 2${\rm \mu m}$ close to perpendicular.}
\label{fg:diagonaleta}
\end{figure}
\subsection{3D sensor normal angle scans}
The resolution of the double-sided 3D sensor is shown in
Figure~\ref{fg:3DResolution} as a function of the angle of rotation of
the sensor. As shown in~\cite{3Dcompanion}, the 3D sensor has
relatively little charge sharing compared with a planar sensor. This
is expected from the self-shielding electric-field geometry of the 3D
sensor. In a planar sensor charge drifts through the thickness of the
sensor to the collection electrodes and diffusion naturally leads to
charge sharing. In a 3D sensor the charge drifts to the collection
electrode column and hence tends to remain bounded inside a pixel.
As the angle between the sensor and the tracks is varied, the
geometry leads to the charge being deposited across multiple
pixels. The fraction of single and multiple pixel clusters is shown in
Figure~\ref{fg:3DFractionClusters}. At perpendicular incidence, the measured 3D Sensor
resolution is $15.8$~$\mu$m, which is compatible with the expected binary
resolution for a $55$~$\mu$m pixel. The optimal resolution obtained was
at a rotation angle of $10^\circ$, where a resolution of $9.2$~$\mu$m
was measured. This angle corresponds to the entry and exit points of
the track in the silicon being separated by approximately one pixel,
and hence maximises the two pixel fraction. No eta correction has
been applied in this analysis, due to the limited amount of diffusion seen.
\twoepspicx{3DResolution}{Resolution of a the double sided 3D
pixel DUT as a function of track incidence angle.}
{3DFractionClusters}{Fraction of single and multiple pixel
clusters in the double sided 3D pixel DUT as a function of the rotation angle.}
\section{Planar sensor bias voltage scans}
\label{s:HVResolution}
The effects of bias voltage on the DUT performance were investigated by taking a
number of measurements with the voltages ranging from 5 to 200V. Each
scan was performed by ramping the HV both up and down, to isolate
affects solely associated with the bias conditions of the sensor. As
described in Section~\ref{planarsensor} the nominal depletion voltage
of the device was 10V. This was confirmed by exposing
the device backside to $\alpha$ source and measuring the voltage at which the signal from the energy deposited by
the $\alpha$ particle reaches the collecting electrodes.
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\linewidth]{Charge_vs_HV.eps}
\caption{ Most probable value of the collected charge for tracks at normal incidence as a function of HV.}
\label{fig:Landau}
\end{figure}
Figure~\ref{fig:Landau} shows the evolution of the measured most probable value of the cluster charge for tracks at normal
incidence. The ToT to charge conversions were performed as described
in the same way as described in Section~\ref{planarsensorlandau}, and
the most probable values are determined with fits to Landau curves
convoluted with a Gaussian described before. Above the depletion
voltage the curve rapidly reaches the maximum value of a
about 23,500 e$^-$. This is expected, as the peaking time of the
electronics (around 100~ns) is much bigger than the typical collection
time in 300~$\mu$m planar sensors.
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\linewidth]{Row_Col_Cluster.eps
\caption{ Cluster width along the tilted (row) and non-tilted (column) as a function of HV. Note that the maximum cluster
size is achieved when the sensor reaches full depletion. In the non-tilted direction, the cluster size decreases
significantly as the sensor becomes more and more overdepleted.}
\label{fig:clustersize}
\end{figure}
Figure \ref{fig:clustersize} shows the evolution of the cluster size as a function of the HV, in the bending plane
(column projection) and non-bending plane (row projection). In the projection where the
particle is at normal incidence, the cluster size is
maximum when the depletion voltage is reached, and then it decreases as the bias voltage increases because of the
faster collection time. Eventually the charge drift speed reaches saturation and the cluster size reaches an asymptotic value.
In the projection where the track is inclined, and thus the cluster
size is dominated by the geomtrical factor, there is no change in the
cluster width above the depletion voltage.
\begin{figure}[p]
\centering
\includegraphics[width=1.0\linewidth]{Row_Col_Resolution_with_eta_10deg.eps}
\caption{ Spatial resolution as a function of the applied bias voltage with the DUT oriented at 10$^\circ $ with respect
to the beam axis along the column direction. On the left the resolution for the coordinate which is sensitive to the
sensor angle is shown before (circles and diamonds), and after
(triangles and squares) eta corrections are applied; on the right
figure the corresponding curves along the non-tilted directions are
shown. The effect of the larger diffusion radius in improving the
charge sharing at low bias voltage is clearly visible for the
perpendicular incidence tracks; for the tracks impinging at 10$^\circ$
the geometrical factor dominates and the resolution is independent of
applied voltage.}
\label{fig:reso}
\end{figure}
Figure \ref{fig:reso} shows the corresponding trends in spatial
resolution, extracted in the same way as described in
Section~\ref{extractingresolution}.
Note that
in the angled projection, the resolution reaches a maximum when the sensor is fully depleted, and then it remains
essentially unchanged. Conversely, in the projection at almost normal incidence, the additional charge diffusion made
possible by the slower collection time, provides a significant improvement in the resolution.
This effect is illustrated in Fig. \ref{fig:anglecomparison} which compares the resolution as a function of angle when
the DUT is biased at 10~V with the angle scan results presented above, which was taken with sensor over-depleted
(V$_{bias}=100$~V). The improvement at low angles from the greater charge spread due to diffusion is evident.
\begin{figure}[p]
\centering
\includegraphics[width=0.7\linewidth]{reso-comp.eps}
\caption{Comparison between the spatial resolution achieved with the sensor biased at depletion voltage
(10~V) and overdepleted (100~V), as a function of track incident angle.}
\label{fig:anglecomparison}
\end{figure}
\section{Time-walk}
\label{sec:timewalk}
During the testbeam several runs were taken with the DUT in Time-of-Arrival (ToA) mode.
These runs were taken to determine the spread in arrival time of pixels within the same cluster.
In view of the VeloPix developments the spread in the ToA should be smaller than one LHC bunch crossing
time (25~ns), as this will simplify the collection of the event fragment belonging to the same bunch crossing\footnote{The
measured time of arrival is highly correlated to the time over threshold. Hence correction of ToA using ToT information is
possible but unwanted because it adds complexity to the processing of data in the off-detector electronics.}.
A source of this time spread is the timewalk of the discriminator which is largely due to the non-zero rise-time of the pre-amplifier output signal.
The maximum timewalk can be defined as the time difference between the earliest possible discriminator output signal (for
large energy depositions) and that for a signal which just crosses threshold. An ideal discriminator will have a maximum
timewalk equal to the rise-time of its input signal.
However, for the Timepix chip the measured difference in arrival time of hits within the same cluster can be as large
as 275~ns, as shown left plot of Figure~\ref{fig:tw1}.
\begin{figure}[p]
\begin{center}
\includegraphics[width=6in]{tw1.eps}
\caption{The left plot shows the measured difference in ToA the hits in a 2-pixel cluster (histogram)
and the calculated ToA difference (open circles) using the ToT distribution shown in the right hand plot.}
\label{fig:tw1}
\end{center}
\end{figure}
This large timewalk, which is significantly larger than the specified rise-time of the preamplifier of ~100 ns, is also due to a lack of overdrive of the discriminator input.
For a small voltage difference between signal input and threshold input the high frequency gain of the discriminator is not
sufficient to make the discriminator output swing from power rail to power rail. As a consequence the discriminator output will
only reach its logical switching level at a later time\footnote{The output will swing eventually if the input signal is over
threshold for a long enough time because the gain of the discriminator is much higher at low frequencies.}.
The question is whether the observed difference in ToA is compatible with timewalk of the Timepix chip obtained
from testpulse measurements~\cite{timepix}.
Unfortunately, the Timepix can only measure either the ToA or ToT, not both at the same time. Hence checking the spread
in ToA is done using data from a ToT run in identical running conditions and converting this data using a parametrization of the timewalk from~\cite{timepix}.
\\
The right plot of Figure~\ref{fig:tw1} shows the ToT for the two pixels in a 2-pixel cluster for a run with the threshold set to 400.
For each bin in this plot, the two ToT values are converted to charge using the surrogate function as described in Section~\ref{s:calibration}.
Next the corresponding time delay of the discriminator is calculated using the charge versus delay curves of the Timepix shown in Figure~\ref{fig:tw2} (from~\cite{timepix}).
The difference in delay time is modified by adding a 25~ns offset for $50\%$ of the statistics. This offset is needed because the timing clock is inverted from pixel to pixel.
Hence one of the (neighbouring) pixels in the 2-pixel cluster is delayed by half a clock cycle.
As the ToA is measured in units of clock cycles, the difference of a half a cycle is rounded to either 0 or 1 depending of arrival time of the particle w.r.t. the phase of the timing clock.
This clock inversion causes the peak at 1 of the ToA difference distribution.
The open circles in Figure~\ref{fig:tw1} are the result of the predicted timewalk using ToT data.
\begin{figure}[p]
\begin{center}
\includegraphics[width=4in]{tw_tp.eps}
\caption{Delay time of Timepix front-end as function of input charge for different thresholds (colors), from\cite{timepix}.}
\label{fig:tw2}
\end{center}
\end{figure}
The calculated ToA distribution has a shorter tail than the measured distribution.
However, the tail of the ToA distribution is due to hits with very small energy depositions. For these hits both the uncertainty in ToT and the uncertainty in delay time of the discriminator (and also the error in its parametrization) are largest.
\section{Pulse-shape}
As explained in Section~\ref{s:timepixchip} the response of the
preamplifier can be approximated to a triangular pulse,
with a rise time of the order of 100~ns for the settings used in the
experimental setup, and a fall time of the order of several
microseconds. Using the threshold scan data it is possible to
investigate qualitatively the shape of the pulse falling edge by
plotting the raw ToT response as a function of the threshold setting.
For each
threshold setting clusters were selected in the device under test
within 100~$\mu$m of an impinging track. The ToT for each threshold
setting was determined by fitting a Landau convoluted with a Gaussian,
as described in Section~\ref{s:HVResolution}, and extracting the MPV.
The values were plotted against the threshold setting in electrons,
for all clusters, and for subsamples of one pixel clusters and two
pixel clusters. The results are shown in Figure~\ref{fg:pulseshape}.
It can be seen that, for the single pixel clusters, there is a
relatively linear behaviour down to thresholds of 3000 electrons or
fewer, after which the non-linear part of the surrogate function is
reflected in a change of shape. For the double pixel clusters,
at low thresholds and high thresholds the curve lies on opposite sides
of the single pixel curve. The reason for this can be again
understood by considering the offset of the surrogate function. At
low threshold values there is a positive offset, so two pixel clusters
exhibit a higher total ToT value than single pixel clusters, whereas
at high thresholds the effect is reversed, as the offset decreases and
eventually becomes negative. For a perfectly working Timepix, two
pixel clusters should always give slightly smaller ToT, corresponding
to twice the threshold value. This plot emphasises the importance of a
good understanding of the surrogate function for the appropriate
threshold in order to extract correct calibrations from Timepix data.
\epspicz{2}{3}{pulseshape}{Timepix pulseshape falling edge, for all clusters, single pixel clusters, and double pixel clusters. The errors on the measured values are negligible.}
\section{The influence of charge digitization on the spatial resolution}
\label{s:gainsection}
For future readout chip designs based on Timepix (such as the proposed VeloPix),
it is important to understand the degradation of the spatial resolution due to limited dynamic
range and charge resolution within the charge digitisation circuit. If the dynamic range spans
less than the full landau spectrum, charge beyond the end of range will accumulate in the highest
digitised count and distort the position reconstruction. A limited charge resolution will also lead
to greater errors in position reconstruction.
Data sets with lower range or resolution were produced by scaling the measured ToT
values and saturating at the highest readout value (given by 2$^n$ - 1 where n is the
number of bits). The original raw ToT spectrum is shown in Figure~\ref{fg:raw_TOT} and the spectrum
of a transformed data set, with a lower dynamic range and smaller readout counter
(four bit) is shown in Figure~\ref{fg:raw_TOT_scaled4bit}. The spectrum is truncated at the highest count and all
higher charges are assigned to the highest count. Non-linear calibration effects and
eta corrections were not considered, and are expected not to significantly alter the results.
Two main effects were investigated. In the first, the dynamic range was kept constant ($\sim$22.5~ke$^-$)
and the number of bits in the readout counter was varied, effectively changing the charge resolution.
The effect on spatial resolution is shown in Figure~\ref{fig:bitplots} for different angles.
Since the position of one pixel
clusters does not depend on their charge, there is no expected effect from the number of bits. For
multi-pixel clusters however, it is clear that at least a three bit counter is required to maintain good
spatial resolution.
In the second set of plots the dynamic range is varied while keeping the number of readout
bits constant. The resulting spatial resolution is shown for three and four bit readout in Figure~\ref{fig:bitplots}, again at
several angles. It can be seen that a range of at least 1 MIP must be covered, with lower ranges significantly degrading the results.
In summary, the charge conversion cover a range of least 1 MIP with three bit resolution to maintain the best spatial resolution.
\twoepspicx{raw_TOT}{Raw ToT spectrum.}{raw_TOT_scaled4bit}{Scaled ToT spectrum.}
\begin{figure}[htb]
\centering
\includegraphics[width=0.3\linewidth]{bit_plot.eps}
\includegraphics[width=0.3\linewidth]{3bit_gain.eps}
\includegraphics[width=0.3\linewidth]{4bit_gain.eps}
\caption{From left to right: resolution as a function of
the number of bits read out, resolution versus gain relative to I$_{\textrm{krum}}$=5 for three bit readout,
and resolution versus gain relative to I$_{\textrm{krum}}$=5 for four bit readout.}
\label{fig:bitplots}
\end{figure}
\section{Conclusions}
\label{s:conclusions}
This paper describes the characterisation of Timepix as a charged
particle tracking device. A particle telescope has been constructed
using $1.4 \times 1.4$ ~cm square sensitive planes consisting of
Timepix and Medipix2 ASICs bonded to planar silicon sensors.
The pixel pitch was $55$~$\mu$m
square. This telescope is shown to achieve
a track pointing precision of $2.3 \pm 0.1$~$\mu$m and was
operated at a track reconstruction rate of 200~Hz. Two different
devices were tested at the centre of the telescope. A Timepix bonded
to a planar sensor was exhaustively tested, and a procedure was
developed to calibrate the sensors with testpulses. This calibration is particularly
important for charge deposits resulting from the trajectories of
minimum ionising charged particles, which fall at the lower edge of
the sensitivity of the Timepix front-end. Using the calibrations a precise comparison has
been made between the landau distributions obtained and the expected
values from literature. The test device was shown to have an
efficiency of greater than $99.5 \%$, and a charge calibration in agreement
with expectations. The resolution was
extracted as a function of track incident angle, in both the normal and
diagonal directions. A best resolution of $4.0 \pm 0.1$~$\mu$m
was measured at the optimum angle of $\sim 10^{\circ}$. For tracks at
perpendicular incidence the resolution was found to be a strong
function of bias voltage, reaching $4.4 \pm 0.1$~$\mu$m at 10~V,
the depletion voltage of the device under test, and $10 \pm 0.2$~$\mu$m
at a depletion voltage of 100~V, corresponding to typical
voltages at which the device would have to be operated in LHC
environments. The performance was also investigated as a function of
the number of bits used in the analogue readout, and the charge
sharing properties described. The time-stamping performance of the
Timepix was measured and compared to expectations. A double-sided 3D sensor bonded to a Timepix was also
characterised and the telescope precision used to make detailed
efficiency maps of the pixel cell. For angled tracks the 3D sensor
achieved equivalent efficiency to the planar sensor. The resolution
was also extracted and the sensor was seen to produce less charge
sharing than the planar device.
These studies demonstrate
comprehensively that Timepix is a very suitable device for charged
particle tracking, and its adaptation for use for the LHCb upgrade will be actively
pursued. The prototype telescope will be upgraded to boost the data acquisition
rate and time stamping peformance and the performance will be
described in a subsequent paper.
\section{Acknowledgments}
The authors would like to thank
Lukas Tlustos,
Winnie Wong,
Timo Tick,
Raphael Ballabriga, and
Sami V\"{a}h\"{a}nen for many useful discussions and their
invaluable contributions to this paper. We are very grateful to
Ian McGill, of the CERN Wire Bonding and Reliability Testing lab, for very
useful discussions and the wire bonding of the Timepix and Medipix assemblies.
The authors are particularly grateful for the support of Stansilav
Pospisil and his group at the IEAP, CTU, Prague, for providing the USB
readouts and Pixelman software. We would like to thank the operation
team of the CERN SPS for excellent support and delivery of the pion beam.
This work was partially supported and financed by the ICTS(Integrated
Nano-Microelectronics Clean Room)access within the GICSERV Programme
and by the Spanish project FPA2009-13896-C02-02.
One of the authors (R.Plackett) gratefully acknowledges support from
the ACEOLE Marie-Curie FP7 Fellowship scheme.
|
\section*{Introduction}
Consider a nonlinear diffusion process of the following for
\begin{equation}
dX\left( t\right) =\Delta \ln \left( X\left( t\right) +1\right) dt
\label{proces}
\end{equation
where $X\left( t,\xi \right) $ is the density for the time - space
coordinates $\left( t,\xi \right) .$ This equation describes the process
that has been observed during experiments using Wisconsin toroidal octupole
plasma containment device (see \cite{ehrhardt}). Kamimura and Dawson
predicted in \cite{Kamimura}\ this process for cross-field conservative
diffusion of plasma including mirror effects.
The same equation describes the expansion of a thermalized electron cloud
and arises also in studies of the central limit approximation to Carleman's
model of the Boltzmann equation (see \cite{carleman} and \cite{Kurtz}). The
asymptotic behavior of this equation was studied in \cite{berryman}. Most of
the natural phenomena exhibit variability which cannot be modeled by using
deterministic approaches. More accurately, natural systems can be
represented as stochastic models and the deterministic description can be
considered as the subset of the pertinent stochastic models.
The purpose of this paper is to analyze such equations within the framework
of stochastic evolution equations with (\ref{proces}) as underlying
motivating example. Let us now introduce the suitable framework for this
problem.
\textbf{Notation}
Let $\mathcal{O}$ be a bounded open interval of $\mathbb{R}$. Recall the
distribution spaces $H_{0}^{1}\left( \mathcal{O}\right) $ on $\mathcal{O}$\
and it's dual $H^{-1}\left( \mathcal{O}\right) $ with the scalar product and
the norm given by
\begin{equation*}
\left\langle u,v\right\rangle _{-1}=\left( \left( -\Delta \right)
^{-1}u,v\right) _{2},\quad \left\vert u\right\vert _{-1}=\left( \left(
-\Delta \right) ^{-1}u,u\right) _{2}^{1/2},
\end{equation*
respectively, where $\left( \cdot ,\cdot \right) _{2}$ is the pairing
between $H_{0}^{1}\left( \mathcal{O}\right) $ and $H^{-1}\left( \mathcal{O
\right) $ and the scalar product of $L^{2}\left( \mathcal{O}\right) .$ The
norm in $L^{p}\left( \mathcal{O}\right) $, $1\leq p\leq \infty $ will be
denoted by $\left\vert \cdot \right\vert _{p}.$
Given a Hilbert space $U$, the norm of $U$ will be denoted by $\left\vert
\cdot \right\vert _{U}$ and the scalar product by $\left( \cdot ,\cdot
\right) _{U}.$ By $C\left( \left[ 0,T\right] ;U\right) $ we shall denote the
space of $U-$ valued continuous functions on $\left[ 0,T\right] $ and by
C_{W}\left( \left[ 0,T\right] ;L^{2}\left( \Omega ,\mathcal{F},\mathbb{P
;U\right) \right) $ the space of all $U$-valued adapted stochastic processes
with respect to filtration $\mathcal{F}$ of the probability space, which are
mean square continuous.
\textbf{Formulation of the problem and hypotheses}
The main result is an existence and uniqueness theorem for the following
stochastic nonlinear diffusion equations in $H^{-1}\left( \mathcal{O}\right)
$ with additive nois
\begin{equation}
\left\{
\begin{array}{ll}
dX\left( t\right) -\Delta ~sign\left( X\left( t\right) \right) \ln \left( |
X\left( t\right)| +1\right) dt=\sqrt{Q}dW\left( t\right) ,~ & in~\left(
0,T\right) \times \mathcal{O}, \\
sign\left( X\left( t\right) \right) \ln \left( | X\left( t\right)| +1\right)
=0, & on~\left( 0,T\right) \times \partial \mathcal{O}, \\
X\left( 0\right) =x, & in~\mathcal{O}
\end{array
\right. \label{equ}
\end{equation
where $\mathcal{O}$\ is an open bounded interval of $\mathbb{R}$, $x$ is an
initial datum and
\begin{equation*}
sign\left( x\right) =\left\{
\begin{array}{l}
\dfrac{x}{\left\vert x\right\vert },\quad \text{if \ }x\neq 0; \\
\left[ -1,1\right] ,\quad \text{if \ }x=0
\end{array
\right.
\end{equation*}
Here $W\left( t\right) $ is a cylindrical Wiener process on $L^{2}\left(
\mathcal{O}\right) $ of the form
\begin{equation*}
W\left( t\right) =\sum_{k=1}^{\infty }\beta _{k}\left( t\right) e_{k},\quad
t\geq 0
\end{equation*
for $\left\{ \beta _{k}\right\} $\ a sequence of independent standard
Brownian motion on\ a filtered probability space $\left( \Omega ,\mathcal{F
,\left\{ \mathcal{F}_{t}\right\} _{t\geq 0},\mathbb{P}\right) $ and $\left\{
e_{k}\right\} $ is a complete orthonormal system in $L^{2}\left( \mathcal{O
\right) $ of eigenfunctions of $-\Delta $ with Dirichlet homogeneous
boundary conditions. We denote by $\left\{ \lambda _{k}\right\} $ the
corresponding sequence of eigenvalues. The operator $Q\in L\left(
L^{2}\left( \mathcal{O}\right) \right) $ defined by
\begin{equation*}
\sqrt{Q}h=\sum_{k=1}^{\infty }\gamma _{k}\left\langle h,e_{k}\right\rangle
_{2}e_{k},\quad \forall ~h\in L^{2}\left( \mathcal{O}\right)
\end{equation*
where $\left\{ \gamma _{k}\right\} $\ is a sequence of positive numbers, is
symmetric, self-adjoint and nonnegative. Then the random forcing term is
\begin{equation*}
\sqrt{Q}dW\left( t\right) =\sum_{k=1}^{\infty }\gamma _{k}e_{k}d\beta
_{k}\left( t\right) ,\quad t\geq 0.
\end{equation*}
Because $\Psi :\mathbb{R} \rightarrow
\mathbb{R}
,$ $\Psi \left( x\right) =sign\left( x \right) \ln \left(| x| +1\right) $ is
a maximal monotone operator, we can see that the operator defined by
A\left( x\right) =-\Delta \Psi \left( x\right) ,$ for all $x\in D\left(
A\right) ,$ where $D\left( A\right) =\left\{ x\in H^{-1}\left( \mathcal{O
\right) \cap L^{1}\left( \mathcal{O}\right) ;~\Psi \left( x\right) \in
H_{0}^{1}\left( \mathcal{O}\right) \right\} $ is maximal monotone in
H^{-1}\left( \mathcal{O}\right) \times H^{-1}\left( \mathcal{O}\right) $
(see \cite{brezis.operatori}).
In this paper we shall assume that the sequence $\left\{ \gamma _{k}\right\}
$ is such that
\begin{itemize}
\item[$\left( H_{1}\right) $] $\sum_{k=1}^{\infty }\gamma _{k}^{2}\lambda
_{k}^{2}<\infty $ and $\sum_{k=1}^{\infty }\gamma _{k}\lambda
_{k}^{3}<\infty $;
\end{itemize}
and
\begin{itemize}
\item[$\left( H_{2}\right) $] $\sqrt{Q}W\in C\left( \left[ 0,T\right] \times
\overline{\mathcal{O}}\right) $ $\mathbb{P}-a.s..$
\end{itemize}
Note that, for every $\omega $ fixed, we have, for some constant $C,$ tha
\begin{equation}
\underset{s\in \left[ 0,T\right] }{\sup }\left\vert \sqrt{Q}W\left( s\right)
\right\vert _{L^{\infty }\left( \mathcal{O}\right) }\leq C. \label{spatiu}
\end{equation}
A similar result was proven in \cite{parab} for semilinear parabolic
stochastic equations and in \cite{strong} for porous media stochastic
equations.\bigskip
Denote by $g:\mathbb{R} \rightarrow \mathbb{R},\
\begin{equation*}
g\left( x\right) =\left( |x|+1\right) \ln \left( |x|+1\right) -|x|,
\end{equation*
and note that, in our case, $\partial g=\Psi .$
\begin{definition}
\label{def}An adapted stochastic process
\begin{equation*}
X\in C_{W}\left( \left[ 0,T\right] ;H^{-1}\left( \mathcal{O}\right) \right)
\cap L^{p}\left( \left( 0,T\right) \times \mathcal{O\times }\Omega \right)
,\quad p\geq 4,
\end{equation*
is said to be a solution to equation (\ref{equ}) i
\begin{eqnarray*}
&&\frac{1}{2}\left\vert X\left( t\right) -\sqrt{Q}W\left( t\right) -Z\left(
t\right) \right\vert _{-1}^{2}+\int_{0}^{t}\int_{\mathcal{O}}g\left( X\left(
s\right) \right) d\xi ds\medskip \\
&&\quad \quad \quad \quad +\int_{0}^{t}\int_{\mathcal{O}}\left( -\Delta
\right) ^{-1}\frac{\partial }{\partial s}Z\left( s\right) \left( X\left(
s\right) -\sqrt{Q}W\left( s\right) -Z\left( s\right) \right) d\xi ds\medskip
\\
&\leq &\frac{1}{2}\left\vert x-Z\left( 0\right) \right\vert
_{-1}^{2}+\int_{0}^{t}\int_{\mathcal{O}}g\left( Z\left( s\right) +\sqrt{Q
W\left( s\right) \right) d\xi ds,\quad \mathbb{P-}a.s.
\end{eqnarray*
for every starting point $x\in L^{p}\left( \mathcal{O}\right) $ and for all
adapted stochastic processes
\begin{equation*}
Z\in C_{W}\left( \left[ 0,T\right] ;H^{-1}\left( \mathcal{O}\right) \right)
,
\end{equation*
such that, for every $\omega \in \Omega $ fixed, satisfies
\begin{itemize}
\item[i)] $Z\in C\left( \left[ 0,T\right] ;L^{2}\left( \mathcal{O}\right)
\right) ;\medskip $
\item[ii)] $Z^{\prime }\in L^{2}\left( 0,T;H^{-1}\left( \mathcal{O}\right)
\right) ;\medskip $
\item[iii)] $g\left( Z+\sqrt{Q}W\right) \in L^{1}\left( \left( 0,T\right)
\times \mathcal{O}\right) .\medskip $
\end{itemize}
\end{definition}
Definition 1 resembles the classical definition of a mild (integral)
solution to deterministic variational inequality (see, e.g., \cite{analys}).
For stochastic differential equations a slightly different version was used
in \cite{BDPR} and \cite{AR}. We can easily see that a solution in the sense
of [\cite{concise}, Definition 4.2.1] is also a solution in the sense of
Definition \ref{def} above.
\textbf{Context}
Existence results for equation
\begin{equation*}
\left\{
\begin{array}{ll}
dX\left( t\right) -\Delta \Psi \left( X\left( t\right) \right) dt=\sqrt{Q
dW\left( t\right) ,~ & in~\left( 0,T\right) \times \mathcal{O}, \\
\Psi \left( X\left( t\right) \right) =0, & on~\left( 0,T\right) \times
\partial \mathcal{O}, \\
X\left( 0\right) =x, & in~\mathcal{O}
\end{array
\right.
\end{equation*
were obtained in \cite{nonneg} for$\ \Psi $ monotonically increasing,
continuous, with $\Psi \left( 0\right) =0,$ and satisfying the following
growth conditions
\begin{equation*}
\Psi ^{\prime }\left( r\right) \leq \alpha _{1}\left\vert r\right\vert
^{m-1}+\alpha _{2}\text{ and }\int_{0}^{r}\Psi \left( s\right) ds\geq \alpha
_{3}\left\vert r\right\vert ^{m+1}+\alpha _{4},
\end{equation*
for all $r\in \mathbb{R}$, where $\alpha _{1},\alpha _{2},\alpha _{4}\geq
0,~\alpha _{3}>0$ and $m\geq 1.$ This result was generalized in \cit
{criticality}. Note that our case is not covered by those hypotheses.
An other existence result was proved in \cite{fast} for the operator
\begin{equation*}
\Psi \left( r\right) =sign\left( r\right) \left\vert r\right\vert ^{\theta
-1}\left( \log \left( \left\vert r\right\vert +1\right) \right) ^{s},
\end{equation*
$r\in \mathbb{R}$ and $\theta \in \left( 1,\infty \right) ,$ $s\in \left[
1,\infty \right) $ (see \cite{fast} Example 3.5). \newline
In the present paper we are considering the critical case $\theta =1$ which
was not covered, by using a different approach and a different definition of
the solution.
\section{The main result}
The main result of this work is the following
\begin{theorem}
\label{theorem} For all $x\in L^{p}\left( \mathcal{O}\right) ,$ $p\geq 4,$
equation (\ref{equ}) has an unique solution in the sense of Definition \re
{def}.
\end{theorem}
In order to prove this result we need some estimates that will be used for
both existence and uniqueness.
\textbf{A priori Estimates}
Denote by $\Psi :\mathbb{R}\rightarrow 2^{\mathbb{R}},~\Psi \left( x\right)
=sign\left( x\right) \ln \left( |x|+1\right) .$ Since $\Psi $ is a maximal
monotone operator we can consider the following approximating equatio
\begin{equation}
\left\{
\begin{array}{ll}
dX_{\varepsilon }\left( t\right) -\Delta \overline{\Psi }_{\varepsilon
}\left( X_{\varepsilon }\left( t\right) \right) dt=\sqrt{Q}dW\left( t\right)
,~ & in~\left( 0,T\right) \times \mathcal{O}, \\
\overline{\Psi }_{\varepsilon }\left( X_{\varepsilon }\left( t\right)
\right) =0, & on~\left( 0,T\right) \times \partial \mathcal{O}, \\
X_{\varepsilon }\left( 0\right) =x, & in~\mathcal{O}
\end{array
\right. \label{approx1}
\end{equation
where $\overline{\Psi }_{\varepsilon }\left( x\right) =\Psi _{\varepsilon
}\left( x\right) +\varepsilon x,~$for all $x\in \mathbb{R},$ and $\Psi
_{\varepsilon }$ is the Yosida approximation of $\Psi ,$ i.e.
\begin{equation*}
\Psi _{\varepsilon }\left( x\right) =\frac{1}{\varepsilon }\left( 1-\left(
1+\varepsilon \Psi \right) ^{-1}\right) \left( x\right) =\Psi \left( \left(
1+\varepsilon \Psi \right) ^{-1}x\right)
\end{equation*
for all $\varepsilon >0.$ We can take this approximation since $\Psi $ is a
maximal monotone operator (see e.g., \cite{analys}, \cite{new}).\newline
For each $\varepsilon >0$ fixed, equation (\ref{approx1}) has an unique
solution in the sense of \cite{infinite} or [\cite{concise}, Definition
4.2.1] (see Example 4.1.11 from \cite{concise}). Note that solution
X_{\varepsilon }$ to the approximation equation (\ref{approx1}) is in our
case a path-wise continuous, $H^{-1}\left( \mathcal{O}\right) $ - valued,
\left( \mathcal{F}_{t}\right) $ - adapted stochastic process. Clearly, this
is also solution in the sense of Definition \ref{def}.
Setting
\begin{equation*}
Y_{\varepsilon }\left( t\right) =X_{\varepsilon }\left( t\right) -\sqrt{Q
W\left( t\right) ,
\end{equation*
we may rewrite (\ref{approx1}) as a random equatio
\begin{equation}
\left\{
\begin{array}{ll}
dY_{\varepsilon }\left( t\right) -\Delta \overline{\Psi }_{\varepsilon
}\left( Y_{\varepsilon }\left( t\right) +\sqrt{Q}W\left( t\right) \right)
dt=0, & in~\left( 0,T\right) \times \mathcal{O}, \\
\overline{\Psi }_{\varepsilon }\left( Y_{\varepsilon }\left( t\right) +\sqrt
Q}W\left( t\right) \right) =0, & on~\left( 0,T\right) \times \partial
\mathcal{O}, \\
Y_{\varepsilon }\left( 0\right) =x, & in~\mathcal{O}
\end{array
\right. \label{approx2}
\end{equation
For each $\omega \in \Omega $ fixed, by classical existence theory for
nonlinear equation we have that equation (\ref{approx2}) has a unique
solution $Y_{\varepsilon }\in C\left( \left[ 0,T\right] ;L^{2}\left(
\mathcal{O}\right) \right) ,$ with $Y_{\varepsilon }^{\prime }\in
L^{2}\left( 0,T;H^{-1}\left( \mathcal{O}\right) \right) $ (see \cite{lions}
for the general result and \cite{strong} for a similar case).
By the It\^{o} formula with the function $x\mapsto \left\vert x\right\vert
_{2}^{2^{{}}}$\ we get from (\ref{approx1}) tha
\begin{equation*}
\mathbb{E}\left\vert X_{\varepsilon }\left( t,x\right) \right\vert
_{2}^{2}\leq \left\vert x\right\vert _{2}^{2}+t\sum_{k=1}^{\infty }\lambda
_{k}^{2}\gamma _{k}^{2}
\end{equation*
and then, we get that, for each $\omega \in \Omega $ fixed, we hav
\begin{equation}
\left\vert X_{\varepsilon }\left( t\right) \right\vert _{L^{2}\left(
\mathcal{O}\right) }^{2}\leq C\left( \omega \right) , \label{ito1}
\end{equation
for all $t\in \left[ 0,T\right] ,$ with $C$ independent of $\varepsilon $.
By assumption \textbf{H}$_{2},$ we can assume the same estimate holds for
Y_{\varepsilon },$ i.e.
\begin{equation}
\left\vert Y_{\varepsilon }\left( t\right) \right\vert _{L^{2}\left(
\mathcal{O}\right) }^{2}=\left\vert X_{\varepsilon }\left( t\right) -\sqrt{Q
W\left( t\right) \right\vert _{L^{2}\left( \mathcal{O}\right) }^{2}\leq C,
\label{ito2}
\end{equation
for all $t\in \left[ 0,T\right] ,$ with $C$ independent of $\varepsilon .$
\begin{lemma}
\label{lemma}There exists a constant $C$ (independent of $\varepsilon $)
such that for all $\omega \in \Omega $ fixed, we have tha
\begin{equation*}
\int_{0}^{T}\int_{\mathcal{O}}\left\vert \Psi _{\varepsilon }\left(
Y_{\varepsilon }\left( s\right) +\sqrt{Q}W\left( s\right) \right)
\right\vert d\xi ds\leq C,
\end{equation*
for all $t\in \left[ 0,T\right] .$
\end{lemma}
\begin{proof}[Proof of Lemma \protect\ref{lemma}]
Recall that $g:\mathbb{R} \rightarrow \mathbb{R},$ is defined b
\begin{equation*}
g\left( x\right) =\left( |x|+1\right) \ln \left( |x|+1\right) -|x|
\end{equation*
and $g_{\varepsilon }$ is the Moreau-Yosida approximation of $g$. Since
\partial g=\Psi $ we have by [\cite{analys}, Theorem 2.2] that $\nabla
g_{\varepsilon }=\Psi _{\varepsilon }.$
From the definition of the subdifferential we have, for all $0<\lambda <1$
fixed and for all $\theta ,$ such that $\left\vert \theta \right\vert
\dfrac{\lambda }{2},$ the following inequalit
\begin{eqnarray*}
&&\Psi _{\varepsilon }\left( Y_{\varepsilon }+\sqrt{Q}W\right) \left(
Y_{\varepsilon }+\sqrt{Q}W-\theta -\lambda \right) \\
&\geq &g_{\varepsilon }\left( Y_{\varepsilon }+\sqrt{Q}W\right)
-g_{\varepsilon }\left( \theta +\lambda \right) \\
&\geq &g\left( J_{\varepsilon }\left( Y_{\varepsilon }+\sqrt{Q}W\right)
\right) -g\left( \theta +\lambda \right) ,
\end{eqnarray*
$a.e.$ on $\left[ 0,T\right] \times \mathcal{O}$.
Taking into account that
\begin{equation*}
0\leq g\left( x\right) \leq x^{2},\text{ for all }x\in \mathbb{R},\
\end{equation*
we obtain tha
\begin{eqnarray*}
\Psi _{\varepsilon }\left( Y_{\varepsilon }+\sqrt{Q}W\right) \left(
Y_{\varepsilon }+\sqrt{Q}W-\theta -\lambda \right) &\geq &-\left( \theta
+\lambda \right) ^{2} \\
&>&-\frac{9\lambda ^{2}}{4}>-\frac{9}{4},
\end{eqnarray*
$a.e.$ on $\left[ 0,T\right] \times \mathcal{O}.$\newline
By taking
\begin{equation*}
\theta =\frac{\lambda }{2}\frac{\Psi _{\varepsilon }\left( Y_{\varepsilon }
\sqrt{Q}W\right) }{\left\vert \Psi _{\varepsilon }\left( Y_{\varepsilon }
\sqrt{Q}W\right) \right\vert }
\end{equation*
it follows tha
\begin{equation*}
\left\vert \Psi _{\varepsilon }\left( Y_{\varepsilon }+\sqrt{Q}W\right)
\right\vert \leq \frac{2}{\lambda }\Psi _{\varepsilon }\left( Y_{\varepsilon
}+\sqrt{Q}W\right) \left( Y_{\varepsilon }+\sqrt{Q}W-\lambda \right) +\frac{
}{2\lambda }
\end{equation*
$a.e.$ on $\left[ 0,T\right] \times \mathcal{O}.$
Consequently we obtain tha
\begin{eqnarray}
&&\int_{0}^{T}\int_{\mathcal{O}}\left\vert \Psi _{\varepsilon }\left(
Y_{\varepsilon }(s)+\sqrt{Q}W\right) (s)\right\vert d\xi ds\quad \quad \quad
\quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad
\label{estimm} \\
&\leq &\frac{2}{\lambda }\int_{0}^{T}\int_{\mathcal{O}}\Psi _{\varepsilon
}\left( Y_{\varepsilon }+\sqrt{Q}W\right) \left( Y_{\varepsilon }+\sqrt{Q
W-\lambda \right) d\xi ds+\frac{9}{2\lambda }\left\vert \mathcal{O
\right\vert T \notag \\
&\leq &\frac{2}{\lambda }\int_{0}^{T}\int_{\mathcal{O}}\left[ \left( -\Delta
\right) ^{-1}\left( -\frac{\partial }{\partial s}Y_{\varepsilon }\right)
-\varepsilon \left( Y_{\varepsilon }+\sqrt{Q}W\right) \right] \notag \\
&&\quad \quad \quad \quad \quad \quad \quad \quad \times \left(
Y_{\varepsilon }+\sqrt{Q}W-\lambda \right) d\xi ds+\frac{9}{2\lambda
\left\vert \mathcal{O}\right\vert T \notag \\
&\leq &\frac{2}{\lambda }\int_{0}^{T}\int_{\mathcal{O}}\left[ \left( -\Delta
\right) ^{-1}\left( -\frac{\partial }{\partial s}Y_{\varepsilon }\right)
\right] \left( Y_{\varepsilon }+\sqrt{Q}W-\lambda \right) d\xi ds \notag \\
&&\quad \quad \quad \quad \quad \quad -\frac{2\varepsilon }{\lambda
\int_{0}^{T}\int_{\mathcal{O}}\left( Y_{\varepsilon }+\sqrt{Q}W\right)
^{2}d\xi ds \notag \\
&&\quad \quad \quad \quad \quad \quad +2\varepsilon \int_{0}^{T}\int_
\mathcal{O}}\left( Y_{\varepsilon }+\sqrt{Q}W\right) d\xi ds+\frac{9}
2\lambda }\left\vert \mathcal{O}\right\vert T \notag \\
&\leq &\frac{2}{\lambda }\int_{0}^{T}\int_{\mathcal{O}}\left[ \left( -\Delta
\right) ^{-1}\left( -\frac{\partial }{\partial s}Y_{\varepsilon }\right)
\right] \left( Y_{\varepsilon }+\sqrt{Q}W-\lambda \right) d\xi ds+C. \notag
\end{eqnarray
Now it is sufficient to show boundedness fo
\begin{eqnarray}
&&\frac{2}{\lambda }\int_{0}^{T}\int_{\mathcal{O}}\left( -\Delta \right)
^{-1}\left( -\frac{\partial }{\partial s}Y_{\varepsilon }\right) \left(
Y_{\varepsilon }+\sqrt{Q}W-\lambda \right) d\xi ds \label{estim} \\
&=&\frac{2}{\lambda }\int_{0}^{T}\int_{\mathcal{O}}\left( -\Delta \right)
^{-1}\left( -\frac{\partial }{\partial s}Y_{\varepsilon }\right)
Y_{\varepsilon }d\xi ds \notag \\
&&+\frac{2}{\lambda }\int_{0}^{T}\int_{\mathcal{O}}\left( -\Delta \right)
^{-1}\left( -\frac{\partial }{\partial s}Y_{\varepsilon }\right) \sqrt{Q
Wd\xi ds \notag \\
&&-2\int_{0}^{T}\int_{\mathcal{O}}\left( -\Delta \right) ^{-1}\left( -\frac
\partial }{\partial s}Y_{\varepsilon }\right) d\xi ds\overset{denote}{=
I_{1}+I_{2}+I_{3.} \notag
\end{eqnarray}
Firstly, we have tha
\begin{eqnarray*}
I_{1} &=&\frac{2}{\lambda }\int_{0}^{T}\int_{\mathcal{O}}\left( -\Delta
\right) ^{-1}\left( -\frac{\partial }{\partial s}Y_{\varepsilon }\right)
Y_{\varepsilon }d\xi ds=\int_{0}^{T}\left\langle -\frac{\partial }{\partial
}Y_{\varepsilon },Y_{\varepsilon }\right\rangle _{-1}ds \\
&=&-\frac{1}{2}\left( \left\vert Y_{\varepsilon }\left( T\right) \right\vert
_{_{-1}}^{2}-\left\vert Y_{\varepsilon }\left( 0\right) \right\vert
_{_{-1}}^{2}\right) \leq C.
\end{eqnarray*
For the second ter
\begin{equation*}
I_{2}=\frac{2}{\lambda }\int_{0}^{T}\int_{\mathcal{O}}\left( -\Delta \right)
^{-1}\left( -\frac{\partial }{\partial s}Y_{\varepsilon }\right) \sqrt{Q
Wd\xi ds,
\end{equation*
we may choose, for $\alpha >0,$ small enough, a decomposition
0=t_{0}<t_{1}<...<t_{i}<t_{i+1}<...<t_{N}=T$ such that, for all $t,s\in
\left[ t_{i},t_{i+1}\right] ,$ we hav
\begin{equation*}
\left\vert \sqrt{Q}W\left( t\right) -\sqrt{Q}W\left( s\right) \right\vert
_{L^{\infty }\left( \mathcal{O}\right) }<\alpha .
\end{equation*
Consequently, we may writ
\begin{eqnarray*}
I_{2} &=&\frac{2}{\lambda }\sum_{i=0}^{N-1}\int_{t_{i}}^{t_{i+1}}\int_
\mathcal{O}}\left( -\Delta \right) ^{-1}\left( -\frac{\partial }{\partial s
Y_{\varepsilon }\right) \sqrt{Q}Wd\xi ds \\
&=&\frac{2}{\lambda }\sum_{i=0}^{N-1}\int_{t_{i}}^{t_{i+1}}\int_{\mathcal{O
}\left( -\Delta \right) ^{-1}\left( -\frac{\partial }{\partial s
Y_{\varepsilon }\right) \left( \sqrt{Q}W\left( s\right) -\sqrt{Q}W\left(
t_{i}\right) \right) d\xi ds \\
&&\quad \quad +\frac{2}{\lambda }\sum_{i=0}^{N-1}\int_{t_{i}}^{t_{i+1}}\int_
\mathcal{O}}\left( -\Delta \right) ^{-1}\left( -\frac{\partial }{\partial s
Y_{\varepsilon }\right) \sqrt{Q}W\left( t_{i}\right) d\xi ds \\
&\leq &\frac{2\alpha }{\lambda }\int_{0}^{T}\left\vert \overline{\Psi
_{\varepsilon }\left( Y_{\varepsilon }\left( s\right) +\sqrt{Q}W\left(
s\right) \right) \right\vert _{L^{1}\left( \mathcal{O}\right) }ds \\
&&\quad \quad -\frac{2}{\lambda }\sum_{i=0}^{N-1}\int_{t_{i}}^{t_{i+1}}\lef
\langle \frac{\partial }{\partial s}Y_{\varepsilon }\left( s\right) ,\sqrt{Q
W\left( t_{i}\right) \right\rangle _{-1}ds \\
&\leq &\frac{2\alpha }{\lambda }\int_{0}^{T}\left\vert \Psi _{\varepsilon
}\left( Y_{\varepsilon }\left( s\right) +\sqrt{Q}W\left( s\right) \right)
\right\vert _{L^{1}\left( \mathcal{O}\right) }ds+C.
\end{eqnarray*
Finally we hav
\begin{eqnarray*}
I_{3} &=&2\int_{0}^{T}\int_{\mathcal{O}}\left( -\Delta \right) ^{-1}\left(
\frac{\partial }{\partial s}Y_{\varepsilon }\right) d\xi ds \\
&=&2\int_{\mathcal{O}}\left( \left( -\Delta \right) ^{-1}Y_{\varepsilon
}\left( T\right) -\left( -\Delta \right) ^{-1}Y_{\varepsilon }\left(
0\right) \right) d\xi \\
&\leq &2\left( \left\vert \left( -\Delta \right) ^{-1}Y_{\varepsilon }\left(
T\right) \right\vert _{L^{1}\left( \mathcal{O}\right) }+\left\vert \left(
-\Delta \right) ^{-1}Y_{\varepsilon }\left( 0\right) \right\vert
_{L^{1}\left( \mathcal{O}\right) }\right) \leq C,
\end{eqnarray*
since $H_{0}^{1}\left( \mathcal{O}\right) \subset L^{1}\left( \mathcal{O
\right) $ and $\left\vert \left( -\Delta \right) ^{-1}x\right\vert
_{H_{0}^{1}\left( \mathcal{O}\right) }=\left\vert x\right\vert _{-1},$ for
all $x\in H^{-1}\left( \mathcal{O}\right) .$
Going back to (\ref{estim}) we get tha
\begin{eqnarray*}
&&\frac{2}{\lambda }\int_{0}^{T}\int_{\mathcal{O}}\left( -\Delta \right)
^{-1}\left( -\frac{\partial }{\partial s}Y_{\varepsilon }\right) \left(
Y_{\varepsilon }+\sqrt{Q}W-\lambda \right) d\xi ds \\
&\leq &\frac{2\alpha }{\lambda }\int_{0}^{T}\left\vert \Psi _{\varepsilon
}\left( Y_{\varepsilon }\left( s\right) +\sqrt{Q}W\left( s\right) \right)
\right\vert _{L^{1}\left( \mathcal{O}\right) }ds+C
\end{eqnarray*
and then, for $\alpha $ small enough and $\lambda $ fixed, we get from (\re
{estimm}) tha
\begin{equation*}
\int_{0}^{T}\int_{\mathcal{O}}\left\vert \Psi _{\varepsilon }\left(
Y_{\varepsilon }\left( s\right) +\sqrt{Q}W\left( s\right) \right)
\right\vert d\xi ds\leq C.
\end{equation*
The proof of the lemma is now complete.
\end{proof}
\begin{proof}[Proof of the main result]
\textbf{Existence}
We have to prove existence of the limit for $\left\{ Y_{\varepsilon
}\right\} _{\varepsilon }$ as $\varepsilon \rightarrow 0$ and consequently
we'll get existence of the solution for equation (\ref{equ}) in the sense of
Definition \ref{def}.\newline
From Lemma \ref{lemma} we get, for all $\omega \in \Omega $ fixed, tha
\begin{equation}
\left\vert \frac{\partial }{\partial t}\left( -\Delta \right)
^{-1}Y_{\varepsilon }\right\vert _{L^{1}\left( 0,T;L^{1}\left( \mathcal{O
\right) \right) }\leq C.
\end{equation
Since for $\mathcal{O}\subset \mathbb{R}$ we have $L^{1}\left( \mathcal{O
\right) \subset H^{-1}\left( \mathcal{O}\right) $, this leads to
\begin{eqnarray}
V_{0}^{T}\left( \left( -\Delta \right) ^{-1}Y_{\varepsilon }\right) &\leq
&\left\vert \frac{\partial }{\partial t}\left( -\Delta \right)
^{-1}Y_{\varepsilon }\right\vert _{L^{1}\left( 0,T;H^{-1}\left( \mathcal{O
\right) \right) } \label{helly1} \\
&\leq &\left\vert \frac{\partial }{\partial t}\left( -\Delta \right)
^{-1}Y_{\varepsilon }\right\vert _{L^{1}\left( 0,T;L^{1}\left( \mathcal{O
\right) \right) }\leq C, \notag
\end{eqnarray
for $C$ independent of $\varepsilon .$ We denoted by
\begin{equation*}
V_{0}^{T}\left( f\right) =\sup \sum\limits_{i=1}^{n}\left\vert f\left(
t_{i}\right) -f\left( t_{i-1}\right) \right\vert _{-1}
\end{equation*
where the supremum is taken over all partitions $D=\left\{
0=t_{0}<t_{1}<...<t_{n}=T\right\} $ of $\left[ 0,T\right] .$
On the other hand, by classical deterministical arguments we have that, for
each $\omega \in \Omega $ fixed
\begin{equation*}
\underset{t\in \left[ 0,T\right] }{\sup }\left\vert Y_{\varepsilon }\left(
t\right) \right\vert _{-1}^{2}\leq C\left( \omega \right)
\end{equation*
which leads t
\begin{equation}
\underset{t\in \left[ 0,T\right] }{\sup }\left\vert \left( -\Delta \right)
^{-1}Y_{\varepsilon }\left( t\right) \right\vert _{H_{0}^{1}\left( \mathcal{
}\right) }^{2}\leq C\left( \omega \right) . \label{ito}
\end{equation
Then, since $\left\{ \left( -\Delta \right) ^{-1}Y_{\varepsilon }\left(
t\right) \right\} _{\varepsilon }$ is bounded in $H_{0}^{1}\left( \mathcal{O
\right) $ and $H_{0}^{1}\left( \mathcal{O}\right) \subset H^{-1}\left(
\mathcal{O}\right) $ compactly, we have that
\begin{equation}
\left\{ \left( -\Delta \right) ^{-1}Y_{\varepsilon }\left( t\right) \right\}
_{\varepsilon }\text{ is compact in }H^{-1}\left( \mathcal{O}\right) ,\text{
for all }t\in \left[ 0,T\right] . \label{helly2}
\end{equation
From (\ref{helly1}) and\ (\ref{helly2}) it follows, via Helly-Foia\c{s}
theorem (see Theorem 3.5 and Remark 3.2 from \cite{BPr} or page 238 from
\cite{MN}), that, on a subsequence, we have
\begin{equation*}
\left( -\Delta \right) ^{-1}Y_{\varepsilon }\left( t\right) \rightarrow
G\left( t\right) \text{ strongly in }H^{-1}\left( \mathcal{O}\right) ,\text{
for all }t\in \left[ 0,T\right] .
\end{equation*
Because $H_{0}^{1}\left( \mathcal{O}\right) \subset L^{2}\left( \mathcal{O
\right) \subset H^{-1}\left( \mathcal{O}\right) $ we have tha
\begin{eqnarray*}
\left\vert \left( -\Delta \right) ^{-1}Y_{\varepsilon }\left( t\right)
-G\left( t\right) \right\vert _{L^{2}\left( \mathcal{O}\right) } &\leq
&\varepsilon \left\vert \left( -\Delta \right) ^{-1}Y_{\varepsilon }\left(
t\right) -G\left( t\right) \right\vert _{H_{0}^{1}\left( \mathcal{O}\right) }
\\
&&+C\left( \varepsilon \right) \left\vert \left( -\Delta \right)
^{-1}Y_{\varepsilon }\left( t\right) -G\left( t\right) \right\vert _{-1}
\end{eqnarray*
for all $t\in \left[ 0,T\right] $ (see \cite{lions}, p. 58)\ and therefore
\begin{equation*}
\left( -\Delta \right) ^{-1}Y_{\varepsilon }\left( t\right) \rightarrow
G\left( t\right) \text{ strongly in }L^{2}\left( \mathcal{O}\right) ,\text{
for all }t\in \left[ 0,T\right] .
\end{equation*
Using (\ref{ito2}) we obtain tha
\begin{equation}
\left( -\Delta \right) ^{-1}Y_{\varepsilon }\left( t\right) \rightarrow
\left( -\Delta \right) ^{-1}Y\left( t\right) \text{ strongly in }L^{2}\left(
\mathcal{O}\right) ,\text{ for all }t\in \left[ 0,T\right] . \label{tare2}
\end{equation
On the other hand we have
\begin{eqnarray*}
&&\left\vert Y_{\varepsilon }\left( t\right) -Y_{\lambda }\left( t\right)
\right\vert _{-1}^{2}\medskip \\
&=&\left\langle \left( -\Delta \right) ^{-1}Y_{\varepsilon }\left( t\right)
-\left( -\Delta \right) ^{-1}Y_{\lambda }\left( t\right) ,Y_{\varepsilon
}\left( t\right) -Y_{\lambda }\left( t\right) \right\rangle _{L^{2}\left(
\mathcal{O}\right) }\medskip \\
&\leq &\left\vert \left( -\Delta \right) ^{-1}Y_{\varepsilon }\left(
t\right) -\left( -\Delta \right) ^{-1}Y_{\lambda }\left( t\right)
\right\vert _{L^{2}\left( \mathcal{O}\right) }\left\vert Y_{\varepsilon
}\left( t\right) -Y_{\lambda }\left( t\right) \right\vert _{L^{2}\left(
\mathcal{O}\right) }.\medskip
\end{eqnarray*
Using (\ref{ito2}) and (\ref{tare2}) we get tha
\begin{equation}
Y_{\varepsilon }\left( t\right) \rightarrow Y\left( t\right) \text{ strongly
in }H^{-1}\left( \mathcal{O}\right) ,\text{ for all }t\in \left[ 0,T\right]
\label{convergenta}
\end{equation
and, since $X_{\varepsilon }=Y_{\varepsilon }+\sqrt{Q}W,$ with $\sqrt{Q}W\in
C\left( \left[ 0,T\right] \times \overline{\mathcal{O}}\right) $, we have
\begin{equation}
X_{\varepsilon }\left( t\right) \rightarrow X\left( t\right) \text{ strongly
in }H^{-1}\left( \mathcal{O}\right) ,\text{ for all }t\in \left[ 0,T\right] .
\label{c1}
\end{equation}
On the other hand, from (\ref{ito1}) we have, for every $\omega \in \Omega $
fixed, that
\begin{eqnarray*}
\int_{0}^{t}\int_{\mathcal{O}}g\left( \left( 1+\varepsilon \Psi \right)
^{-1}X_{\varepsilon }\left( t\right) \right) d\xi &\leq &\int_{0}^{t}\int_
\mathcal{O}}\left\vert \left( 1+\varepsilon \Psi \right) ^{-1}X_{\varepsilon
}\left( t\right) \right\vert ^{2}d\xi \\
&\leq &\int_{0}^{t}\int_{\mathcal{O}}\left\vert X_{\varepsilon }\left(
t\right) \right\vert ^{2}d\xi \leq C\left( \omega \right) ,\text{ for all
~t\in \left[ 0,T\right] ,
\end{eqnarray*
where $\left( 1+\varepsilon \Psi \right) ^{-1}$ is the resolvent of $\Psi .$
Sinc
\begin{equation*}
\underset{|x|\rightarrow \infty }{\lim }\frac{g\left( x\right) }{\left\vert
x\right\vert }=\infty
\end{equation*
we have that $\left\{ \left( 1+\varepsilon \Psi \right) ^{-1}\left(
X_{\varepsilon }\left( t\right) \right) \right\} _{\varepsilon }$ is bounded
and equi-integrable in $L^{1}\left( \mathcal{O\times }\left( 0,T\right)
\right) .$ Then, by the Dunford-Pettis theorem, we get that the sequence is
weakly compact in $L^{1}\left( \mathcal{O\times }\left( 0,T\right) \right) .$
Hence, along a subsequence, again denoting by $\varepsilon $, we obtain that
\begin{equation}
\left( 1+\varepsilon \Psi \right) ^{-1}\left( X_{\varepsilon }\left(
t\right) \right) \rightharpoonup X\left( t\right) ,\text{ \ weakly in
L^{1}\left( \mathcal{O\times }\left( 0,T\right) \right) , \label{c2}
\end{equation
as $\varepsilon \rightarrow 0.$
We know that $X_{\varepsilon }=Y_{\varepsilon }+\sqrt{Q}W$ is also a
solution to equation (\ref{approx1}) in the sense of our definition, $i.e.,
\begin{eqnarray}
&&\frac{1}{2}\left\vert X_{\varepsilon }\left( t\right) -\sqrt{Q}W\left(
t\right) -Z\left( t\right) \right\vert _{-1}^{2}+\int_{0}^{t}\int_{\mathcal{
}}g_{\varepsilon }\left( X_{\varepsilon }\left( s\right) \right) d\xi
ds\medskip \label{def-aprox} \\
&&\quad \quad \quad \quad +\int_{0}^{t}\int_{\mathcal{O}}\left( -\Delta
\right) ^{-1}\frac{\partial }{\partial s}Z\left( s\right) \left(
X_{\varepsilon }\left( s\right) -\sqrt{Q}W\left( s\right) -Z\left( s\right)
\right) d\xi ds\medskip \notag \\
&\leq &\frac{1}{2}\left\vert x-Z\left( 0\right) \right\vert _{H^{-1}\left(
\mathcal{O}\right) }^{2}+\int_{0}^{t}\int_{\mathcal{O}}g_{\varepsilon
}\left( Z\left( s\right) +\sqrt{Q}W\left( s\right) \right) d\xi ds,\quad
\mathbb{P-}a.s.\medskip \notag
\end{eqnarray
for all$~t\in \left[ 0,T\right] .$
We intend to take the $liminf$ for $\varepsilon \rightarrow 0$ in (\re
{def-aprox}).
Convergence of the first term is a direct consequence of (\ref{c1}) and for
the last term we only need to use classical properties of the Moreau-Yosida
approximation i.e.,\ $g\left( \left( 1+\varepsilon \Psi \right) ^{-1}\left(
x\right) \right) \leq g_{\varepsilon }\left( x\right) \leq g\left( x\right) $
for all $x\in \mathbb{R}$.
We shall discuss now the second and the third term of the left hand side
\newline
Since $\varphi :L^{1}\left( \mathcal{O\times }\left( 0,T\right) \right)
\rightarrow \mathbb{R}$ defined by $\varphi \left( x\right)
=\int_{0}^{t}\int_{\mathcal{O}}g\left( x\left( \xi ,s\right) \right) d\xi ds$
is weakly $l.s.c.$ on $L^{1}\left( \mathcal{O\times }\left( 0,T\right)
\right) $ (see \cite{analys} Proposition 2.9 and Proposition 2.12 from
Chapter 2) we get from (\ref{c2}) that
\begin{equation*}
\underset{\varepsilon \rightarrow 0}{\lim \inf }\int_{0}^{t}\int_{\mathcal{O
}g_{\varepsilon }\left( X_{\varepsilon }\left( s\right) \right) d\xi ds\geq
\int_{0}^{t}\int_{\mathcal{O}}g\left( X\left( s\right) \right) d\xi ds,\text{
\ for all}~t\in \left[ 0,T\right] ,
\end{equation*
and then we can pass to the $liminf$ for $\varepsilon \rightarrow 0$ in the
second term.\newline
The third term of the left hand side can be written a
\begin{equation*}
\int_{0}^{t}\left\langle \frac{\partial }{\partial s}Z\left( s\right)
,Y_{\varepsilon }\left( s\right) -Z\left( s\right) \right\rangle _{-1}ds.
\end{equation*
From (\ref{convergenta}) we have tha
\begin{equation*}
\left\langle \frac{\partial }{\partial s}Z,Y_{\varepsilon }-Z\right\rangle
_{-1}\rightarrow \left\langle \frac{\partial }{\partial s}Z,Y-Z\right\rangle
_{-1},\text{ }a.e.\text{ on }\left[ 0,T\right] .
\end{equation*
On the other hand we can easily see tha
\begin{eqnarray*}
\left\langle \frac{\partial }{\partial s}Z\left( s\right) ,Y_{\varepsilon
}\left( s\right) -Z\left( s\right) \right\rangle _{-1} &\leq &\left\vert
\frac{\partial }{\partial s}Z\left( s\right) \right\vert _{-1}\left\vert
Y_{\varepsilon }\left( s\right) -Z\left( s\right) \right\vert _{-1} \\
&\leq &\left\vert \frac{\partial }{\partial s}Z\left( s\right) \right\vert
_{-1}\underset{s\in \left[ 0,T\right] }{ess\sup }\left\vert Y_{\varepsilon
}\left( s\right) -Z\left( s\right) \right\vert _{-1} \\
&\leq &C\left\vert \frac{\partial }{\partial s}Z\left( s\right) \right\vert
_{-1}\in L^{2}\left( 0,T\right) ,\text{ }a.e.\text{ on }\left[ 0,T\right] .
\end{eqnarray*
Now, by using Lebesgue's dominated convergence theorem we obtain tha
\begin{equation*}
\underset{\varepsilon \rightarrow 0}{\lim }\int_{0}^{t}\left\langle \frac
\partial }{\partial s}Z\left( s\right) ,Y_{\varepsilon }\left( s\right)
-Z\left( s\right) \right\rangle _{-1}ds=\int_{0}^{t}\left\langle \frac
\partial }{\partial s}Z\left( s\right) ,Y\left( s\right) -Z\left( s\right)
\right\rangle _{-1}ds.
\end{equation*
At this point we can take the $\lim \inf $ \ for $\varepsilon \rightarrow 0$
in (\ref{def-aprox}) and get that
\begin{eqnarray*}
&&\frac{1}{2}\left\vert X\left( t\right) -\sqrt{Q}W\left( t\right) -Z\left(
t\right) \right\vert _{-1}^{2}+\int_{0}^{t}\int_{\mathcal{O}}g\left( X\left(
s\right) \right) d\xi ds\medskip \\
&&\quad \quad \quad \quad +\int_{0}^{t}\int_{\mathcal{O}}\left( -\Delta
\right) ^{-1}\frac{\partial }{\partial s}Z\left( s\right) \left( X\left(
s\right) -\sqrt{Q}W\left( s\right) -Z\left( s\right) \right) d\xi ds\medskip
\\
&\leq &\frac{1}{2}\left\vert x-Z\left( 0\right) \right\vert
_{-1}^{2}+\int_{0}^{t}\int_{\mathcal{O}}g\left( Z\left( s\right) +\sqrt{Q
W\left( s\right) \right) d\xi ds,\quad \mathbb{P-}a.s.
\end{eqnarray*
for all $t\in \left[ 0,T\right] .$
By applying the It\^{o} formula to equation (\ref{approx1}) with the
function $x\mapsto \left\vert x\right\vert _{-1}^{2}$ and from (\ref{c1}) we
get that
\begin{equation*}
X_{\varepsilon }\rightarrow X,\text{ weakly in }L_{W}^{2}\left( \left[ 0,
\right] ;L^{2}\left( \Omega ;H^{-1}\left( \mathcal{O}\right) \right) \right)
,\text{ as }\varepsilon \rightarrow 0.
\end{equation*}
Then, arguing as in \cite{concise}, we may replace $X$ by a $H^{-1}\left(
\mathcal{O}\right) -$ continuous version and follows that the solution is
also an $\left( \mathcal{F}_{t}\right) $- adapted stochastic process.
On the other hand we can easily see that $X\in L^{p}\left( \Omega \times
\mathcal{O}\times \left[ 0,T\right] \right) $ arguing as in Lemma 3.1 from
\cite{criticality} and that conclude the proof of the existence.
\textbf{Uniqueness}
Consider $\overline{X}$ an arbitrary solution to equation (\ref{equ}) in the
sense of Definition \ref{def}.
The main idea of the proof is to take $Z=\left( 1-\mu \Delta \right)
^{-1}Y_{\varepsilon }=J_{\mu }Y_{\varepsilon }$ in Definition \ref{def} ,
for $Y_{\varepsilon }$ the solution to equation (\ref{aprox3}) below\ and
J_{\mu }$ the resolvent of the Laplacian.
Consider, for each $\omega \in \Omega $ fixed, the following approximating
equatio
\begin{equation}
\left\{
\begin{array}{ll}
dY_{\varepsilon }\left( t\right) -\Delta \overline{\Psi }_{\varepsilon
}\left( Y_{\varepsilon }\left( t\right) +\sqrt{Q}W\left( t\right) \right)
dt=0, & in~\left( 0,T\right) \times \mathcal{O}, \\
\overline{\Psi }_{\varepsilon }\left( Y_{\varepsilon }\left( t\right) +\sqrt
Q}W\left( t\right) \right) =0, & on~\left( 0,T\right) \times \partial
\mathcal{O}, \\
Y_{\varepsilon }\left( 0\right) =x, & in~\mathcal{O}
\end{array
\right. \label{aprox3}
\end{equation
where $\overline{\Psi }_{\varepsilon }\left( x\right) =\Psi _{\varepsilon
}\left( x\right) +\varepsilon x,~~$for all $x\in \mathbb{R} $
and $\Psi _{\varepsilon }$ is the Yosida approximation of $\Psi ,$ for every
$\varepsilon >0.$ By classical existence theory, equation (\ref{aprox3}) has
a unique solution
\begin{equation*}
Y_{\varepsilon }\in C\left( \left[ 0,T\right] ;L^{2}\left( \mathcal{O
\right) \right) \cap L^{2}\left( 0,T;H_{0}^{1}\left( \mathcal{O}\right)
\right) ,
\end{equation*
with $Y_{\varepsilon }^{\prime }\in L^{2}\left( 0,T;H^{-1}\left( \mathcal{O
\right) \right) .$
For $\mu >0$ fixed we consider the resolvent of the Laplacian $J_{\mu
}:L^{2}\left( \mathcal{O}\right) \rightarrow L^{2}\left( \mathcal{O}\right)
, $ $J_{\mu }\left( x\right) =\left( 1-\mu \Delta \right) ^{-1}\left(
x\right) ,$ for all $x\in L^{2}\left( \mathcal{O}\right) .$
Since $J_{\mu }$ is differentiable we may denote by $DJ_{\mu }$ the Gateaux
differential and we see that
\begin{equation}
\frac{d }{d t}J_{\mu }\left( Y_{\varepsilon }\left( t\right) \right)
=DJ_{\mu }\left( Y_{\varepsilon }\left( t\right) \right) \frac{\partial }
\partial t}Y_{\varepsilon }\left( t\right) =J_{\mu }\left( \frac{\partial }
\partial t}Y_{\varepsilon }\left( t\right) \right) \label{comut}
\end{equation
for all $t\in \left[ 0,T\right] .$
Now we can prove that $Z=J_{\mu }Y_{\varepsilon }$ satisfies $i),~...,~iii)$
from Definition \ref{def}.
We can easily see that $J_{\mu }Y_{\varepsilon }\in C\left( \left[ 0,T\right]
;L^{2}\left( \mathcal{O}\right) \right) $ and then $i)$ is satisfied.
Concerning $ii)$ we know that, for every $\varepsilon >0$ fixed, we hav
\begin{equation*}
\int_{0}^{t}\left\vert J_{\mu }\frac{\partial }{\partial s}Y_{\varepsilon
}\left( s\right) \right\vert _{-1}^{2}ds\leq \int_{0}^{t}\left\vert \frac
\partial }{\partial s}Y_{\varepsilon }\left( s\right) \right\vert
_{-1}^{2}ds
\end{equation*
an
\begin{equation*}
\frac{\partial }{\partial s}Y_{\varepsilon }\in L^{2}\left( 0,T;H^{-1}\left(
\mathcal{O}\right) \right) .
\end{equation*
Hence, by (\ref{comut}), we get tha
\begin{equation*}
\frac{\partial }{\partial t}Z=\frac{d }{d t}J_{\mu }\left( Y_{\varepsilon
}\right) \in L^{2}\left( 0,T;H^{-1}\left( \mathcal{O}\right) \right) .
\end{equation*}
Property $iii)$ is a direct consequence of $0\leq g\left( x\right) \leq
x^{2},$ for all $x\in \left( -1,\infty \right) .$ Indeed, we have for every
\varepsilon $ and $\mu $ fixed that
\begin{equation*}
\int_{0}^{t}\int_{\mathcal{O}}g\left( J_{\mu }Y_{\varepsilon }\left(
s\right) +\sqrt{Q}W\right) d\xi ds\leq \int_{0}^{t}\int_{\mathcal{O
}\left\vert J_{\mu }Y_{\varepsilon }\left( s\right) +\sqrt{Q}W\right\vert
^{2}d\xi ds\leq \infty
\end{equation*
because $J_{\mu }Y_{\varepsilon }\in C\left( \left[ 0,T\right] ;L^{2}\left(
\mathcal{O}\right) \right) $ and $\sqrt{Q}W\in C\left( \left[ 0,T\right]
\times \overline{\mathcal{O}}\right) .$
Since $J_{\mu }Y_{\varepsilon }$ satisfies i), ii) and iii) we can write
Definition \ref{def} for the solution $\bar{X}$ with $Z=J_{\mu
}Y_{\varepsilon },$ i.e.,
\begin{eqnarray}
&&\ \ \ \frac{1}{2}\left\vert \bar{X}\left( t\right) -\sqrt{Q}W\left(
t\right) -J_{\mu }Y_{\varepsilon }\left( t\right) \right\vert
_{-1}^{2}+\int_{0}^{t}\int_{\mathcal{O}}g\left( \bar{X}\left( s\right)
\right) d\xi ds \label{def_unic} \\
&&\quad +\int_{0}^{t}\int_{\mathcal{O}}\left( -\Delta \right) ^{-1}\frac{d }
d s}J_{\mu }Y_{\varepsilon }\left( s\right) \left( \bar{X}\left( s\right)
\sqrt{Q}W\left( s\right) -J_{\mu }Y_{\varepsilon }\left( s\right) \right)
d\xi ds \notag \\
&\leq &\frac{1}{2}\left\vert x-J_{\mu }Y_{\varepsilon }\left( s\right)
\left( 0\right) \right\vert _{-1}^{2}+\int_{0}^{t}\int_{\mathcal{O}}g\left(
J_{\mu }Y_{\varepsilon }\left( s\right) +\sqrt{Q}W\left( s\right) \right)
d\xi ds,\quad \mathbb{P-}a.s.. \notag
\end{eqnarray
Applying $J_{\mu }$ (which is linear) to (\ref{aprox3}) we get tha
\begin{equation*}
\frac{d }{d t}\left[ J_{\mu }Y_{\varepsilon }\left( t\right) \right] +J_{\mu
}\left( -\Delta \right) \left( \overline{\Psi }_{\varepsilon }\left(
Y_{\varepsilon }\left( t\right) +\sqrt{Q}W\left( t\right) \right) \right)
=0.
\end{equation*
By Proposition VII 2, $a_{1})$ and $a_{2})$ from \cite{brezis.analiza} we
can rewrite this equation as follow
\begin{equation}
\frac{d }{d t}\left[ J_{\mu }Y_{\varepsilon }\left( t\right) \right] -\Delta
\left( J_{\mu }\overline{\Psi }_{\varepsilon }\left( Y_{\varepsilon }\left(
t\right) +\sqrt{Q}W\left( t\right) \right) \right) =0. \label{ecc}
\end{equation
Using (\ref{ecc}) in the third term of the left-hand side of (\ref{def_unic
) we can rewrite i
\begin{eqnarray*}
&&\int_{0}^{t}\int_{\mathcal{O}}\left( -\Delta \right) ^{-1}\frac{d }{d s
J_{\mu }Y_{\varepsilon }\left( s\right) \left( \bar{X}\left( s\right) -\sqrt
Q}W\left( s\right) -J_{\mu }Y_{\varepsilon }\left( s\right) \right) d\xi
ds\medskip \\
&=&\int_{0}^{t}\int_{\mathcal{O}}J_{\mu }\overline{\Psi }_{\varepsilon
}\left( Y_{\varepsilon }+\sqrt{Q}W\right) \left( J_{\mu }Y_{\varepsilon
}\left( s\right) -\bar{X}\left( s\right) +\sqrt{Q}W\left( s\right) \right)
d\xi ds\medskip \\
&=&\int_{0}^{t}\int_{\mathcal{O}}\overline{\Psi }_{\varepsilon }\left(
Y_{\varepsilon }+\sqrt{Q}W\right) \left( Y_{\varepsilon }+\sqrt{Q}W-J_{\mu
}\left( \overline{Y}+\sqrt{Q}W\right) \right) d\xi ds\medskip \\
&&+\int_{0}^{t}\int_{\mathcal{O}}\overline{\Psi }_{\varepsilon }\left(
Y_{\varepsilon }+\sqrt{Q}W\right) \left( J_{\mu }\sqrt{Q}W-\sqrt{Q}W\right)
d\xi ds\medskip \\
&&+\int_{0}^{t}\int_{\mathcal{O}}\overline{\Psi }_{\varepsilon }\left(
Y_{\varepsilon }+\sqrt{Q}W\right) \left( \left( 1-\mu \Delta \right)
^{-2}Y_{\varepsilon }-Y_{\varepsilon }\right) d\xi ds\medskip \\
&=&T_{1}+T_{2}+T_{3},
\end{eqnarray*
where $\overline{Y}=\bar{X}-\sqrt{Q}W.$
Since $\overline{\Psi }_{\varepsilon }\left( x\right) =\Psi _{\varepsilon
}\left( x\right) +\varepsilon x,$ where $\Psi _{\varepsilon }$ is the Yosida
approximation of $\Psi ,$ and $\overline{g}_{\varepsilon }\left( x\right)
=g_{\varepsilon }\left( x\right) +\varepsilon \frac{x^{2}}{2},$ where
g_{\varepsilon }$ is the Moreau-Yosida approximation of $g,$ we have that
\overline{g}_{\varepsilon }^{\prime }=\overline{\Psi }_{\varepsilon }$ and
then, by the definition of the subdifferential, we get that
\begin{equation*}
\overline{\Psi }_{\varepsilon }\left( x\right) \left( x-y\right) \geq
\overline{g}_{\varepsilon }\left( x\right) -\overline{g}_{\varepsilon
}\left( y\right) .
\end{equation*}
This leads t
\begin{eqnarray*}
T_{1} &=&\int_{0}^{t}\int_{\mathcal{O}}\overline{\Psi }_{\varepsilon }\left(
Y_{\varepsilon }+\sqrt{Q}W\right) \left( Y_{\varepsilon }+\sqrt{Q}W-J_{\mu
}\left( \overline{Y}+\sqrt{Q}W\right) \right) d\xi ds \\
&\geq &\int_{0}^{t}\int_{\mathcal{O}}\left( \overline{g}_{\varepsilon
}\left( Y_{\varepsilon }+\sqrt{Q}W\right) -\overline{g}_{\varepsilon }\left(
J_{\mu }\left( \overline{Y}+\sqrt{Q}W\right) \right) \right) d\xi ds.
\end{eqnarray*
We also have
\begin{eqnarray*}
-T_{2} &=&\int_{0}^{t}\int_{\mathcal{O}}\overline{\Psi }_{\varepsilon
}\left( Y_{\varepsilon }+\sqrt{Q}W\right) \left( \sqrt{Q}W-J_{\mu }\sqrt{Q
W\right) d\xi ds \\
&\leq &\int_{0}^{t}\left\vert \overline{\Psi }_{\varepsilon }\left(
Y_{\varepsilon }+\sqrt{Q}W\right) \right\vert _{-1}\left\vert \sqrt{Q
W-J_{\mu }\sqrt{Q}W\right\vert _{H_{0}^{1}\left( \mathcal{O}\right) }ds \\
&\leq &\left\vert \overline{\Psi }_{\varepsilon }\left( Y_{\varepsilon }
\sqrt{Q}W\right) \right\vert _{L^{1}\left( \left( 0,T\right) \times \mathcal
O}\right) }\left\vert \sqrt{Q}W-J_{\mu }\sqrt{Q}W\right\vert _{L^{\infty
}\left( 0,T;H_{0}^{1}\left( \mathcal{O}\right) \right) } \\
&\leq &C\left\vert \sqrt{Q}W-J_{\mu }\sqrt{Q}W\right\vert _{L^{\infty
}\left( 0,T;H_{0}^{1}\left( \mathcal{O}\right) \right) },
\end{eqnarray*
using Lemma \ref{lemma} and the fact that $L^{1}\left( \mathcal{O}\right)
\subset H^{-1}\left( \mathcal{O}\right) ,$ for $\mathcal{O}\in \mathbb{R}$
\newline
From (\ref{aprox3}) follows that
\begin{eqnarray*}
T_{3} &=&\int_{0}^{t}\int_{\mathcal{O}}\left( -\Delta \right) ^{-1}\frac
\partial }{\partial s}Y_{\varepsilon }\left( s\right) \left( Y_{\varepsilon
}\left( s\right) -\left( 1-\mu \Delta \right) ^{-2}Y_{\varepsilon }\left(
s\right) \right) d\xi ds\medskip \\
&=&\int_{0}^{t}\left\langle \frac{\partial }{\partial s}Y_{\varepsilon
}\left( s\right) ,Y_{\varepsilon }\left( s\right) \right\rangle
_{-1}ds-\int_{0}^{t}\left\langle \frac{d }{d s}J_{\mu }Y_{\varepsilon
}\left( s\right) ,J_{\mu }Y_{\varepsilon }\left( s\right) \right\rangle
_{-1}ds\medskip \\
&=&\int_{0}^{t}\frac{\partial }{\partial s}\frac{1}{2}\left\vert
Y_{\varepsilon }\left( s\right) \right\vert _{-1}^{2}ds-\int_{0}^{t}\frac
\partial }{\partial s}\frac{1}{2}\left\vert J_{\mu }Y_{\varepsilon }\left(
s\right) \right\vert _{-1}^{2}ds\medskip \\
&=&\frac{1}{2}\left( \left\vert Y_{\varepsilon }\left( t\right) \right\vert
_{-1}^{2}-\left\vert J_{\mu }Y_{\varepsilon }\left( t\right) \right\vert
_{-1}^{2}\right) -\frac{1}{2}\left( \left\vert Y_{\varepsilon }\left(
0\right) \right\vert _{-1}^{2}-\left\vert J_{\mu }Y_{\varepsilon }\left(
0\right) \right\vert _{-1}^{2}\right) \medskip \\
&\geq &-\frac{1}{2}\left( \left\vert Y_{\varepsilon }\left( 0\right)
\right\vert _{-1}^{2}-\left\vert J_{\mu }Y_{\varepsilon }\left( 0\right)
\right\vert _{-1}^{2}\right) =-\frac{1}{2}\left( \left\vert x\right\vert
_{-1}^{2}-\left\vert J_{\mu }x\right\vert _{-1}^{2}\right)
\end{eqnarray*
since $\left\vert J_{\mu }Y_{\varepsilon }\left( t\right) \right\vert
_{-1}^{2}$ $\leq \left\vert Y_{\varepsilon }\left( t\right) \right\vert
_{-1}^{2}$ \ for all $t\in \left[ 0,T\right] $ and $x$ is the starting point
of the problem.\newline
Going back to (\ref{def_unic}) we get that
\begin{eqnarray}
&&\quad \frac{1}{2}\left\vert \overline{Y}\left( t\right) -J_{\mu
}Y_{\varepsilon }\left( t\right) \right\vert _{-1}^{2}+\int_{0}^{t}\int_
\mathcal{O}}g\left( \overline{X}\right) d\xi ds+\int_{0}^{t}\int_{\mathcal{O
}\overline{g}_{\varepsilon }\left( X_{\varepsilon }\right) d\xi ds
\label{dddd} \\
&\leq &\frac{1}{2}\left\vert x-J_{\mu }Y_{\varepsilon }\left( 0\right)
\right\vert _{-1}^{2}+\int_{0}^{t}\int_{\mathcal{O}}g\left( J_{\mu
\overline{X}\right) d\xi ds+\frac{\varepsilon }{2}\int_{0}^{t}\int_{\mathcal
O}}\left( J_{\mu }\overline{X}\right) ^{2}d\xi ds \notag \\
&&\quad +\int_{0}^{t}\int_{\mathcal{O}}g\left( J_{\mu }Y_{\varepsilon }
\sqrt{Q}W\right) d\xi ds \notag \\
&&\quad +C\left\vert \sqrt{Q}W-J_{\mu }\sqrt{Q}W\right\vert _{L^{\infty
}\left( 0,T;H_{0}^{1}\left( \mathcal{O}\right) \right) } \notag \\
&&\quad +\frac{1}{2}\left( \left\vert x\right\vert _{-1}^{2}-\left\vert
J_{\mu }x\right\vert _{-1}^{2}\right) ,\quad \mathbb{P-}a.s.. \notag
\end{eqnarray
We firstly pass to the $liminf$ for $\varepsilon \rightarrow 0,$ with $\mu
>0 $ fixed, as follows.\newline
Arguing as we did in the proof of existence we have tha
\begin{equation*}
\underset{\varepsilon \rightarrow 0}{\lim \inf }\int_{0}^{t}\int_{\mathcal{O
}\overline{g}_{\varepsilon }\left( X_{\varepsilon }\left( s\right) \right)
d\xi ds\geq \int_{0}^{t}\int_{\mathcal{O}}g\left( X\left( s\right) \right)
d\xi ds,\text{ \ for all }t\in \left[ 0,T\right] .
\end{equation*
We shall now pass to the $liminf$ for $\varepsilon \rightarrow 0,$ with $\mu
>0 $ fixed, in
\begin{equation*}
\int_{0}^{t}\int_{\mathcal{O}}g\left( J_{\mu }Y_{\varepsilon }\left(
s\right) +\sqrt{Q}W\left( s\right) \right) d\xi ds.
\end{equation*
We know by (\ref{ito2}) that $\left\{ Y_{\varepsilon }\right\} _{\varepsilon
}$ is bounded in $C\left( \left[ 0,T\right] ;L^{2}\left( \mathcal{O}\right)
\right) $ and considering (\ref{convergenta}) we get that
\begin{equation*}
Y_{\varepsilon }\left( t\right) \rightarrow Y\left( t\right) ,\text{ weakly
in }L^{2}\left( \mathcal{O}\right) ,\text{\ \ for all }t\in \left[ 0,T\right]
,
\end{equation*
as $\varepsilon \rightarrow 0.$ On the other hand, we know that, for every
\mu >0$ fixed, we have that $J_{\mu }:L^{2}\left( \mathcal{O}\right)
\rightarrow L^{2}\left( \mathcal{O}\right) $ is compact and then
\begin{equation*}
J_{\mu }Y_{\varepsilon }\left( t\right) \rightarrow J_{\mu }Y\left( t\right)
,\text{ strongly in }L^{2}\left( \mathcal{O}\right) ,\text{\ \ for all }t\in
\left[ 0,T\right] ,
\end{equation*
as $\varepsilon \rightarrow 0.$\newline
Since $g:\mathbb{R} \rightarrow \mathbb{R}$ is continuous and $\sqrt{Q}W\in
C\left( \left[ 0,T\right] \times \overline{\mathcal{O}}\right) $ we have tha
\begin{equation}
g\left( J_{\mu }Y_{\varepsilon }+\sqrt{Q}W\right) \rightarrow g\left( J_{\mu
}Y+\sqrt{Q}W\right) ,~a.e.\text{ on }\left[ 0,T\right] \times \mathcal{O}.
\label{lebesgue1}
\end{equation
We also know that $J_{\mu }:L^{2}\left( \mathcal{O}\right) \rightarrow
C\left( \overline{\mathcal{O}}\right) $ is continuous and then $\left\{
J_{\mu }Y_{\varepsilon }+\sqrt{Q}W\right\} _{\varepsilon }$ is bounded in
C\left( \left[ 0,T\right] \times \overline{\mathcal{O}}\right) .$ We obtain
that
\begin{equation}
g\left( J_{\mu }Y_{\varepsilon }+\sqrt{Q}W\right) \leq \left\vert J_{\mu
}Y_{\varepsilon }+\sqrt{Q}W\right\vert ^{2}\leq C,~a.e.\text{ on }\left[ 0,
\right] \times \mathcal{O}. \label{lebesgue2}
\end{equation
Consequently, by Lebesgue's dominated convergence theorem, we get from (\re
{lebesgue1}) and (\ref{lebesgue2}) tha
\begin{equation*}
\underset{\varepsilon \rightarrow 0}{\lim }\int_{0}^{t}\int_{\mathcal{O
}g\left( J_{\mu }Y_{\varepsilon }+\sqrt{Q}W\right) d\xi ds=\int_{0}^{t}\int_
\mathcal{O}}g\left( J_{\mu }Y+\sqrt{Q}W\right) d\xi ds.
\end{equation*
Going back to (\ref{dddd}) and passing to the $liminf$ for $\varepsilon
\rightarrow 0,$ with $\mu >0$ fixed, we get that
\begin{eqnarray}
&&\quad \frac{1}{2}\left\vert \overline{Y}\left( t\right) -J_{\mu }Y\left(
t\right) \right\vert _{-1}^{2}\medskip +\int_{0}^{t}\int_{\mathcal{O
}g\left( X\right) d\xi ds+\int_{0}^{t}\int_{\mathcal{O}}g\left( \overline{X
\right) d\xi ds \label{defff} \\
&\leq &\frac{1}{2}\left\vert x-J_{\mu }x\right\vert
_{-1}^{2}+\int_{0}^{t}\int_{\mathcal{O}}g\left( J_{\mu }\overline{X}\right)
d\xi ds+\int_{0}^{t}\int_{\mathcal{O}}g\left( J_{\mu }Y+\sqrt{Q}W\right)
d\xi ds\medskip \notag \\
&&\quad +C\left\vert \sqrt{Q}W-J_{\mu }\sqrt{Q}W\right\vert _{L^{\infty
}\left( 0,T;H_{0}^{1}\left( \mathcal{O}\right) \right) } \notag \\
&&\quad +\frac{1}{2}\left( \left\vert x\right\vert _{-1}^{2}-\left\vert
J_{\mu }x\right\vert _{-1}^{2}\right) ,\quad \mathbb{P-}a.s.. \notag
\end{eqnarray
We shall now pass to the $liminf$ for $\mu \rightarrow 0$.
Firstly we discuss the following term
\begin{equation*}
\int_{0}^{t}\int_{\mathcal{O}}\left[ g\left( J_{\mu }\overline{X}\left(
s\right) \right) -g\left( \overline{X}\left( s\right) \right) \right] d\xi
ds.
\end{equation*
Since
\begin{equation*}
J_{\mu }\overline{X}\left( s\right) \rightarrow \overline{X}\left( s\right)
\text{ strongly in }L^{2}\left( \mathcal{O}\right) ,\ \text{\ for all }s\in
\left[ 0,T\right]
\end{equation*
as $\mu \rightarrow 0$ and since $g:\mathbb{R} \rightarrow \mathbb{R}$ is
continuous, we have tha
\begin{equation*}
g\left( J_{\mu }\overline{X}\left( s\right) \right) \rightarrow g\left(
\overline{X}\left( s\right) \right) ,\text{ }a.e.\text{ on }\mathcal{O},\
\text{\ for all }s\in \left[ 0,T\right]
\end{equation*
as $\mu \rightarrow 0.$\newline
By the Egorov theorem we get that, for every $\delta >0,$ there exists a set
$E_{\delta }$ with the Lebesgue's measure $\left\vert E_{\delta }\right\vert
<\delta ,$ such that
\begin{equation*}
g\left( J_{\mu }\overline{X}\right) -g\left( \overline{X}\right) \rightarrow
0,\text{ uniformly on }\mathcal{O}\backslash E_{\delta }
\end{equation*
as $\mu \rightarrow 0,$ for all $t\in \left[ 0,T\right] $ and then
\begin{equation*}
\underset{\mu \rightarrow 0}{\lim }\int_{\mathcal{O}}\left[ g\left( J_{\mu
\overline{X}\right) -g\left( \overline{X}\right) \right] d\xi =\underset{\mu
\rightarrow 0}{\lim }\int_{E_{\delta }}\left[ g\left( J_{\mu }\overline{X
\right) -g\left( \overline{X}\right) \right] d\xi
\end{equation*
for all $t\in \left[ 0,T\right] .$
On the other hand we hav
\begin{eqnarray*}
&&\int_{E_{\delta }}\left[ g\left( J_{\mu }\overline{X}\right) -g\left(
\overline{X}\right) \right] d\xi \\
&\leq &\left( \int_{E_{\delta }}1d\xi \right) ^{1/2}\left( \int_{E_{\delta
}} \left[ g\left( J_{\mu }\overline{X}\right) -g\left( \overline{X}\right)
\right] ^{2}d\xi \right) ^{1/2} \\
&\leq &2\left\vert E_{\delta }\right\vert ^{1/2}\left( \int_{E_{\delta }
\overline{X}^{4}d\xi \right) ^{1/2}\leq C\delta ,~a.e.\text{ on }\left[ 0,
\right] .
\end{eqnarray*
(Indeed, we can easily see that $\left\vert J_{\mu }\overline{X}\right\vert
_{L^{4}\left( E_{\delta }\right) }\leq \left\vert \overline{X}\right\vert
_{L^{4}\left( E_{\delta }\right) }$ by using the same argument as in \cit
{nonneg} to obtain $3.25$).\newline
Then, we get that
\begin{equation*}
\underset{\mu \rightarrow 0}{\lim }\int_{\mathcal{O}}\left[ g\left( J_{\mu
\overline{X}\right) -g\left( \overline{X}\right) \right] d\xi =0,~a.e.\text{
on }\left[ 0,T\right] .
\end{equation*
We also know from (\ref{ito1}) that
\begin{eqnarray*}
\int_{\mathcal{O}}\left[ g\left( J_{\mu }\overline{X}\left( t\right) \right)
-g\left( \overline{X}\left( t\right) \right) \right] d\xi &\leq &2\int_
\mathcal{O}}\left\vert \overline{X}\left( t\right) \right\vert ^{2}d\xi \\
&\leq &2\underset{t\in \left[ 0,T\right] }{\sup }\int_{\mathcal{O
}\left\vert \overline{X}\left( t\right) \right\vert ^{2}d\xi \leq C,~a.e
\text{ on }\left[ 0,T\right] .
\end{eqnarray*
Using Lebesgue's dominated convergence theorem we get that
\begin{equation*}
\underset{\mu \rightarrow 0}{\lim }\int_{0}^{t}\int_{\mathcal{O}}\left[
g\left( J_{\mu }\overline{X}\left( s\right) \right) -g\left( \overline{X
\left( s\right) \right) \right] d\xi ds=0.
\end{equation*}
Using the same argument we get that
\begin{equation*}
\underset{\mu \rightarrow 0}{\lim }\int_{0}^{t}\int_{\mathcal{O}}\left[
g\left( J_{\mu }Y\left( s\right) +\sqrt{Q}W\left( s\right) \right) -g\left(
X\left( s\right) \right) \right] d\xi ds=0.
\end{equation*}
In order to conclude the proof we only need to mention that, for each
\omega \in \Omega $\ fixed, we hav
\begin{equation}
\underset{\mu \rightarrow 0}{\lim }\left\vert \sqrt{Q}W-J_{\mu }\sqrt{Q
W\right\vert _{L^{\infty }\left( 0,T;H_{0}^{1}\left( \mathcal{O}\right)
\right) }=0, \label{cccc}
\end{equation
which is a consequence of the fact that $\sqrt{Q}W \in L^{\infty }\left(
0,T;H_{0}^{1}\left( \mathcal{O}\right) \right)$
Going back to (\ref{defff})\ we can pass to the $liminf$ for $\mu
\rightarrow 0 $ and get that, for each $\omega \in \Omega $ fixed, we have
\begin{equation*}
\left\vert \overline{X}\left( t\right) -X\left( t\right) \right\vert
_{-1}=0,
\end{equation*
for all $t\in \left[ 0,T\right] ,$ and that assure the uniqueness of the
solution.
\end{proof}
\begin{acknowledgement}
The author thanks the referee for the constructive comments and
suggestions.\bigskip
\end{acknowledgement}
|
\section{Introduction}
In laboratory plasma devices like tokamaks or in astrophysical plasmas, two-dimensional reduced fluid models have been used to investigate various instabilities and their impact on turbulent transport. These reduced models are very useful in practice since their theoretical and numerical analyses are simpler than for complete models. At the same time they display important features of the underlying turbulent phenomena with more clarity than the complete models. The standard procedure used for the derivation of these reduced models is to start from the equations of motion of a parent model (e.g., that given by the continuity and the momentum equations) and to perform an ordering with respect to one or several small parameters on the variables of the system (e.g., density and potential fluctuations) and on the electromagnetic field.
This ordering is suggested by the physics and the geometry of the phenomenon under consideration (e.g., one might want to filter out irrelevant time or length scales).
It is desirable that the reduced model preserves the main ingredients of the parent model under this reduction procedure, as for instance the Hamiltonian symmetry. Typical parent models have indeed this Hamiltonian symmetry, e.g., the collisionless Maxwell-Vlasov equations~\cite{Morrison80PLA,Marsden82} or ideal magnetohydrodynamics~\cite{Morrison80PRL}. However, in such a derivation procedure obtained by working at the level of the equations of motion, neglecting terms can produce ``mutilations''~\cite{Morrison05} of the original fluid equations by breaking this symmetry. These mutilations, which introduce incorrect dissipative terms, have a drastic impact on the properties of the physical system. The loss of the Hamiltonian structure indeed generates a more intricate interpretation of the numerical results, and in general clouds the relation with the parent model and its properties. In fact, as shown for example in models describing magnetic reconnection~\cite{Grasso01,Tassi10}, the Hamiltonian structure helps the interpretation and explanation of complex physical processes observed in numerical simulations. Among the other advantages of the Hamiltonian structure, there is the possibility of identifying systematically invariants of motion (e.g., Casimir invariants or conserved quantities linked to continuous symmetries), and applying tools of Hamiltonian perturbation theory and methods to investigate the stability of equilibria.\\
However, for some reduced models, the important property of possessing a Hamiltonian structure has not been shown. For instance, this is the case of the two-dimensional fluid model taken from Ref.~\cite{Haz85_86}. This four-field model includes Finite Larmor Radius (FLR) effects as in Refs.~\cite{Hsu86,Brizard92}, the drift velocity ordering as in Ref.~\cite{Mikhailovskii71} and the gyroviscous terms as in Refs.~\cite{Hinton71,Scott03}. Moreover, magnetic inhomogeneities are kept in the continuity equation as in Refs.~\cite{Haz85_86,Haz87,Eic96,Gar01}. For this particular model, the Hamiltonian structure has not been found even though the model has been shown to be energy conserving. In this paper, we consider the main obstacle associated with the search for the Hamiltonian structure of this four-field model which is already encapsulated in a two-field model for the density and the vorticity fields. This two-field model is obtained from the four-field model by suppressing the parallel dynamics (along the magnetic field), and the poloidal magnetic and parallel flow fluctuations, and reads
\begin{eqnarray}
\displaystyle
\frac{\partial n}{\partial t} &=& - \left[ \phi , n \right] - \left[ \phi , \frac{1}{B} \right] + \left[ n , \frac{1}{B} \right],\label{eq:TOKAM:continuity} \\
\nonumber\\
\displaystyle
\frac{\partial \Delta\phi}{\partial t} &=& - \left[ \phi , \Delta\phi \right] + (1+T) \left[ n , \frac{1}{B} \right] - T \bm{\nabla} \cdot \left[ n , \bm{\nabla}\phi \right],
\label{eq:TOKAM:vorticity}
\end{eqnarray}
where $n=n(x,y,t)$ is the logarithm of the normalized density fluctuations, $\phi=\phi(x,y,t)$ is the normalized electrostatic potential, $B=B(x,y)$ is the normalized magnetic field, $T$ is the constant ion temperature normalized by the electron temperature and $\left[ f , g \right] = \mathbf{\hat{z}} \cdot \bm{\nabla}f \times \bm{\nabla}g$ is the canonical bracket in the plane across the magnetic field $\mathbf{B} = B \mathbf{\hat{z}}$. All the derivatives are defined on the perpendicular plane. The bracket $\left[ \phi , n \right]$ corresponds to the dynamics due to the $\mathbf{E} \times \mathbf{B}$ drift, and the bracket $\left[ n , 1/B \right]$ corresponds to the term driving the interchange instability.\\
When addressing the question of the Hamiltonian structure of the model~(\ref{eq:TOKAM:continuity})-(\ref{eq:TOKAM:vorticity}), it is easy to show that, provided that boundary terms vanish, a conserved quantity for the model is
\begin{eqnarray}
\displaystyle
I = \frac{1}{2}\int d^2x \Bigg( (1+T) n^{2} + \vert\bm{\nabla}\phi\vert^{2} \Bigg),
\label{eq:TOKAM:Inv}
\end{eqnarray}
where the first part corresponds to the potential energy for ions and electrons and the second part is the kinetic energy. This conserved quantity $I$ is interpreted as the total energy of the system and is a natural candidate for the Hamiltonian. Nevertheless, in spite of having a conserved quantity, the model (\ref{eq:TOKAM:continuity})-(\ref{eq:TOKAM:vorticity}) does not seem to possess a Hamiltonian formulation, at least of the Lie-Poisson type, which is the most common form for reduced fluid models. More precisely, Poisson brackets between functionals $F$ and $G$ of the form
\begin{eqnarray} \label{brack}
\{ F , G \} = \sum_{i,j,k=1}^2 \int d^2x \Bigg( \left( V_{ij}^k \xi_i + \frac{\alpha_{j}^k}{B} \right) \left[ \frac{\delta F}{\delta \xi_j} , \frac{\delta G}{\delta \xi_k} \right] + \left( W_{ij}^k \xi_i + \frac{\alpha_{j}^k}{B} \right) \left[ \bm{\nabla} \frac{\delta F}{\delta \xi_j} ; \bm{\nabla} \frac{\delta G}{\delta \xi_k} \right] \Bigg),\nonumber
\end{eqnarray}
where the vector bracket is defined by $\left[ \mathbf{A} ; \mathbf{B} \right] = \sum_i \left[ A_i , B_i \right]$, the terms $\alpha_{j}^k$, $V_{ij}^k$ and $W_{ij}^k$ are constant, $\xi_1 = n$ , $\xi_2 = \Delta \phi$ and $\delta F/\delta \xi$ is the functional derivative of the functional $F$ with respect to $\xi$, fail to give the desired equations of motion~(\ref{eq:TOKAM:continuity})-(\ref{eq:TOKAM:vorticity}) when applied to $I$.
Even if a bracket giving Eqs.~(\ref{eq:TOKAM:continuity})-(\ref{eq:TOKAM:vorticity}) with the invariant $I$ defined by Eq.~(\ref{eq:TOKAM:Inv}) has been found, it does not satisfy Jacobi identity and thus is not a Poisson bracket.\\
The reason for the difficulty to find a Hamiltonian structure resides in the co-existence of compressibility terms (those containing $1/B$ in the continuity equation of the electrons) with the ion-gyroviscous term (in the vorticity equation). The compressibility terms are often retained in the continuity equation (see, e.g., Refs.~\cite{Haz85_86,Haz87,Eic96,Gar01}) in order to account for the fact that the velocity field is not incompressible in the presence of an inhomogeneous magnetic field. If these compressibility terms are neglected in the two-field model, the resulting system is Hamiltonian~\cite{Haz87,Dagnelund05}. On the other hand, it has been shown that the system obtained by keeping these terms and eliminating the ion-gyroviscous terms is also Hamiltonian~\cite{Haz87}. It is thus the simultaneous presence of the two contributions which seems to complicate the search for a Hamiltonian structure. This leads to the problem of finding a Hamiltonian model accounting for both compressibility terms and FLR corrections.\\
An elegant and effective way to introduce FLR corrections to a cold-ion model, while preserving the Hamiltonian structure, was introduced in Refs.~\cite{Morrison84,Haz87} and was referred to as the gyromap. The gyromap procedure was rigorously found from the Braginskii's closure for the stress tensor~\cite{Braginskii65} only for the three-field model described in Refs.~\cite{Hsu86,Haz87} which does not contain the compressibility terms in the continuity equation. In this paper, we apply the gyromap procedure to the cold-ion limit of the two-field model with the compressibility terms. We show that all the FLR correction terms produced by this method are obtained from the Braginskii's closure for the stress tensor, by means of an expansion based on a physically sound ordering.\\
The paper is organized as follows: In Sec.~2, we apply the gyromap procedure to the cold-ion version of the two-field model (\ref{eq:TOKAM:continuity})-(\ref{eq:TOKAM:vorticity}). In Sec.~3, we show that the resulting model is directly obtained from a systematic expansion.
\section{The gyromap}
The gyromap algorithm was introduced in Ref.~\cite{Haz87} in order to introduce FLR corrections to a cold-ion model, while preserving the Hamiltonian structure. The model with FLR corrections obtained through this procedure possesses the same bracket of the original cold-ion model, but different dynamical variables and Hamiltonian. The fact that the Poisson bracket is inherited from the cold-ion Hamiltonian model ensures that the resulting model is Hamiltonian. As a by-product, the Casimir invariants of the resulting model with FLR corrections become easily available. We also recall that, as mentioned in Ref.~\cite{Haz87}, the gyromap procedure possesses the desirable features that the cold-ion limit ($T=0$) of the post-gyromap model gives the initial cold-ion model, and that the diamagnetic effects predicted by the kinetic theory are conserved at the first order.
In what follows we use dimensionless quantities, i.e. we rescale space variables by the sonic Larmor radius $\rho_s$, the density by the equilibrium density $n_0$, the charge by the electric charge $e$, the electron and ion temperatures by the electron temperature $T_{\rm e}$, the magnetic field $B$ by the spatial mean value $B_0$ and time by the inverse of the ion cyclotron frequency $\omega_{c,{\rm i}}= e B_0/m_{\rm{i}}$ where $m_{\rm{i}}$ is the ion mass.\\
We start with the cold-ion version of Eqs.~(\ref{eq:TOKAM:continuity})-(\ref{eq:TOKAM:vorticity}) which describes the dynamical evolution of the plasma density $n$ and the vorticity $\Delta \phi$
\begin{eqnarray}
\displaystyle
\frac{\partial n}{\partial t} &=& - \left[ \phi , n \right] - \left[ \phi , \frac{1}{B} \right] + \left[ n , \frac{1}{B} \right], \label{eq:TOKAM-2D:n}\nonumber\\
\nonumber\\
\displaystyle
\frac{\partial \Delta\phi}{\partial t} &=& - \left[ \phi , \Delta\phi \right] + \left[ n , \frac{1}{B} \right].
\label{eq:TOKAM-2D:phi}\nonumber
\end{eqnarray}
We consider the system whose field variables are $n(x,y)$ and $\phi(x,y)$. Here we do not specify the time dependence of the variables which is implicitly assumed. In the algebra of observables which are functionals of $n$ and $\phi$, the Hamiltonian structure is defined by the Hamiltonian
\begin{eqnarray}
H(n,\phi) = \frac{1}{2}\int d^2x \Bigg( n^{2} + \vert\bm{\nabla}\phi\vert^{2} \Bigg),
\label{eq:TOKAM-2D:Ham}\nonumber
\end{eqnarray}
and by the Poisson bracket
\begin{eqnarray}
\displaystyle
\{ F , G \} = \int d^2x \Bigg( \left(n+\frac{1}{B}\right) \left( \left[ \frac{\delta F}{\delta n} , \frac{\delta G}{\delta n} \right] + \left[ \frac{\delta F}{\delta n} , \frac{\delta G}{\delta \Delta\phi} \right] + \left[ \frac{\delta F}{\delta \Delta\phi} , \frac{\delta G}{\delta n} \right] \right) + \Delta\phi \left[ \frac{\delta F}{\delta \Delta\phi} , \frac{\delta G}{\delta \Delta\phi} \right] \Bigg).
\label{eq:TOKAM-2D:Lie-Poisson}\nonumber
\end{eqnarray}
This bracket verifies all the required properties for a Poisson bracket like bilinearity, antisymmetry, Jacobi identity and Leibniz identity.\\
The gyromap procedure is initiated from the Hamiltonian defined by
\begin{eqnarray}
H(n,\Phi) = \frac{1}{2}\int d^2x \Bigg( (1+T)n^{2} + \vert\bm{\nabla}\Phi\vert^{2} \Bigg),
\label{eq:Gyromap:Ham}
\end{eqnarray}
where $\Phi$ is the stream function for the FLR model. The choice of Eq.~(\ref{eq:Gyromap:Ham}) as Hamiltonian is motivated by the requirement of adding the ion internal energy to the cold-ion Hamiltonian and having a kinetic energy term whose relation with the cold-ion kinetic energy is retrieved a posteriori. We introduce an auxiliary variable $\xi$ defined by $\Delta \Phi = \xi + T\Delta n/2$. The shift of $T\Delta n/2$ corresponds to a shift of half the magnetization velocity. This is the required transformation in order to yield the proper FLR corrections~\cite{Haz87}. The next step is to use the same Poisson bracket written in the new variables $(N,\xi)$,
\begin{eqnarray}
\displaystyle
\{ F , G \} = \int d^2x \Bigg( \left(N+\frac{1}{B}\right) \left( \left[ \frac{\delta F}{\delta N} , \frac{\delta G}{\delta N} \right] + \left[ \frac{\delta F}{\delta N} , \frac{\delta G}{\delta \xi} \right] + \left[ \frac{\delta F}{\delta \xi} , \frac{\delta G}{\delta N} \right] \right) + \xi \left[ \frac{\delta F}{\delta \xi} , \frac{\delta G}{\delta \xi} \right] \Bigg).
\label{eq:Gyromap:Lie-Poisson_N-xi}
\end{eqnarray}
The change of variables, $N=n$ and $\xi = \Delta \Phi - T \Delta n / 2$, includes a change of functional derivatives,
\begin{eqnarray}
\displaystyle
\frac{\delta}{\delta N} &=& \frac{\delta}{\delta n} + \frac{T}{2} \Delta \frac{\delta}{\delta \Delta \Phi},
\label{eq:Gyromap:fct_der_2}\nonumber\\
\nonumber\\
\displaystyle
\frac{\delta}{\delta \xi} &=& \frac{\delta}{\delta \Delta \Phi}. \label{eq:Gyromap:fct_der_1}\nonumber
\end{eqnarray}
Hence the bracket expressed in the variables $(n,\Phi)$, which is still a Poisson bracket, is the following,
\begin{eqnarray}
\displaystyle
\{ F , G \} &=& \int d^2x \Bigg( \left(n+\frac{1}{B}\right) \left( \left[ \frac{\delta F}{\delta n} + \frac{T}{2}\Delta \frac{\delta F}{\delta \Delta \Phi} , \frac{\delta G}{\delta \Delta \Phi} \right] + \left[ \frac{\delta F}{\delta \Delta \Phi} , \frac{\delta G}{\delta n} + \frac{T}{2}\Delta \frac{\delta G}{\delta \Delta \Phi} \right] \right)\Bigg. \nonumber\\
\displaystyle
&&\Bigg. + \left(n+\frac{1}{B}\right) \left[ \frac{\delta F}{\delta n} + \frac{T}{2}\Delta \frac{\delta F}{\delta \Delta \Phi} , \frac{\delta G}{\delta n} + \frac{T}{2}\Delta \frac{\delta G}{\delta \Delta \Phi} \right] + \left( \Delta \Phi - \frac{T}{2} \Delta n \right) \left[ \frac{\delta F}{\delta \Delta \Phi} , \frac{\delta G}{\delta \Delta \Phi} \right] \Bigg).
\label{eq:Gyromap:Lie-Poisson}
\end{eqnarray}
With the Hamiltonian defined by Eq.~(\ref{eq:Gyromap:Ham}), the equation of motion for the density is
\begin{eqnarray}
\displaystyle
\frac{\partial n}{\partial t} = \{ n , H \} = - \left[ \Phi + \frac{T}{2} \Delta \Phi , n + \frac{1}{B} \right] + (1+T) \left[ n , \frac{1}{B} \right]. \label{eq:Gyromap:continuity}
\end{eqnarray}
We assume that the continuity equation for electrons is unaffected by FLR corrections. Therefore comparing Eq.~(\ref{eq:TOKAM:continuity}) with Eq.~(\ref{eq:Gyromap:continuity}) gives the following relation between $\Phi$ and $\phi$:
\begin{eqnarray}
\displaystyle
\left( 1+\frac{T}{2}\Delta \right) \Phi = \phi + T n. \label{eq:Gyromap:Phi-vs-phi}
\end{eqnarray}
On the other hand, the equation for the generalized vorticity $\Delta\Phi$ is given by
\begin{eqnarray}
\displaystyle
\frac{\partial \Delta \Phi}{\partial t} = \{ \Delta \Phi , H \} = - \left[ \Phi , \Delta \Phi \right] + \left(1+T\right)\left(1+\frac{T}{2}\Delta\right)\left[ n , \frac{1}{B} \right] + T \bm{\nabla} \cdot \left[ n+\frac{1}{B} , \bm{\nabla} \Phi \right] - \frac{T^2}{4}\Delta \left[ \Delta \Phi , n+\frac{1}{B} \right].
\label{eq:Gyromap:vorticity}
\end{eqnarray}
Equations~(\ref{eq:Gyromap:continuity})-(\ref{eq:Gyromap:vorticity}) are the final result of the gyromap procedure. They include FLR corrections and the resulting system has a Hamiltonian structure since the bracket~(\ref{eq:Gyromap:Lie-Poisson}) is obtained from the Poisson bracket~(\ref{eq:Gyromap:Lie-Poisson_N-xi}) by a change of variables.\\
By expanding the operator $(1+T\Delta/2)^{-1}$ for small $T$ and making use of Eq.~(\ref{eq:Gyromap:continuity}), it is possible to obtain an expression for Eq.~(\ref{eq:Gyromap:vorticity}) valid for small ion temperature $T$ (in terms of $n$ and $\phi$),
\begin{eqnarray}
\displaystyle
\frac{\partial \Delta\phi}{\partial t} &=& - \left[ \phi , \Delta\phi \right] + \left[ n , \frac{1}{B} \right]
+ T \left( \left[ n , \frac{1}{B} \right] - \bm{\nabla} \cdot \left[ n , \bm{\nabla}\phi \right] - \left[ \bm{\nabla} \phi \mathbf{;} \bm{\nabla} \Delta \phi \right] \right) \nonumber\\
\displaystyle
&&+ T^2 \left( \left[ \bm{\nabla}\Delta n \mathbf{;} \bm{\nabla}\phi \right] - \left[ \bm{\nabla} n \mathbf{;} \bm{\nabla}\Delta \phi \right] - 2 \left[ n , \Delta n \right] + 2 \Delta \left(1+\frac{\Delta}{2}\right) \left[ n , \frac{1}{B} \right] \right. \nonumber\\
\displaystyle
&&\left.
- \frac{\Delta}{4} \left[ \Delta \phi , n + \frac{1}{B} \right] + \frac{1}{2} \left[ \bm{\nabla}\phi \mathbf{;} \bm{\nabla}\Delta \phi \right] + \frac{1}{4} \left[ \Delta \phi , \Delta^2 \phi \right] \right) + O(T^3).
\label{eq:Gyromap:vorticity_2}
\end{eqnarray}
This expression is useful, because it allows one to directly compare the leading order terms of the vorticity equation~(\ref{eq:Gyromap:vorticity_2}), with the vorticity equation as given in Ref.~\cite{Dagnelund05}~[see also Eq.~(\ref{eq:TOKAM:vorticity})] which refers to a model where the only FLR term is $- T \bm{\nabla} \cdot \left[ n , \bm{\nabla}\phi \right]$. We notice that already at the order $O(T)$, the two equations differ from an FLR correction term given by $-T \left[ \bm{\nabla} \phi ; \bm{\nabla} \Delta \phi \right]$. From Eq.~(\ref{eq:Gyromap:vorticity_2}), it is also straightforward to see that, in the limit $T \rightarrow 0$, the FLR model is correctly reduces to the original cold-ion model, as expected.\\
In summary, in order to obtain a Hamiltonian structure with the ion temperature, the gyromap procedure starts from the Hamiltonian structure without ion temperature and then considers FLR effects using a change of variables.\\
In what follows, we consider Braginskii's closure for the stress tensor in order to derive a model with FLR terms and compare it with the above model given by Eqs.~(\ref{eq:Gyromap:continuity})-(\ref{eq:Gyromap:vorticity}).
\section{Derivation from Braginskii's closure}
We first consider the dynamics of a plasma composed of two species, ions and electrons. In a simplified magnetic geometry where the magnetic field is constant (but non-uniform) and its direction is fixed, we restrict the plasma dynamics to the plane transverse to the magnetic field lines, i.e. all dynamical variables only depend on the two coordinates $x$ and $y$. The dynamical variables are the ion and electron densities, $n_{\rm{i}}(x,y)$ and $n_{\rm{e}}(x,y)$, and the ion and electron velocity fields ${\bf v}_{\rm i}(x,y)$ and ${\bf v}_{\rm e}(x,y)$. We assume a quasi-neutrality hypothesis which allows us to reduce the number of variables, i.e. we consider that $n_{\rm i}(x,y)=n_{\rm e}(x,y)$ which is denoted $n(x,y)$ in what follows. The dynamical equation for the density is given by the continuity equation for the electrons,
\begin{equation}
\frac{\partial n}{\partial t} + \bm{\nabla} \cdot \left( n {\bf v}_{\rm e} \right) = 0.
\label{eq:conserv_density}\nonumber
\end{equation}
Due to the quasi-neutrality assumption, the same equation holds for ions, which translates into
\begin{equation}
\label{eq:qnj}
\bm{\nabla} \cdot \left( n({\bf v}_{\rm e}-{\bf v}_{\rm i})\right)=0.
\end{equation}
The dynamical equation for the velocity field ${\bf v}_s(x,y)$ of the species $s$, where $s$ refers to ions or electrons, is given by
\begin{equation}
\mu_s n \left( \frac{\partial}{\partial t} + {\bf v}_s \cdot \bm{\nabla} \right) {\bf v}_s = n \left( {\bf E} + {\bf v}_s \times {\bf B} \right) - e_s T_s \bm{\nabla} n - e_s \bm{\nabla} \cdot \overline{\overline{{\bm \pi}_s}},
\label{eq:conserv_motion}
\end{equation}
where ${\bf E}$ is the electric field, $T_s$ is the dimensionless temperature of the species $s$ (i.e. $T_{\rm i}=T$ and $T_{\rm e}=1$), $\mu_s$ its dimensionless mass (i.e. $\mu_{\rm i}=1$ and $\mu_{\rm e} = m_{\rm e}/m_{\rm i} \ll \mu_{\rm i}$ because the mass of the electron $m_{\rm e}$ is negligible compared to the one of the ions $m_{\rm i}$) and $e_s$ its dimensionless charge (i.e. $e_{\rm i}=1$ and $e_{\rm e}=-1$). The stress tensor associated with the species $s$, denoted $\overline{\overline{{\bm \pi}_s}}$, is taken from Ref.~\cite{Braginskii65}. For each species, it is composed of viscosity terms identified by viscosity coefficients of various orders. Within a strong magnetic field approximation, we only consider the two higher order viscosity terms labeled by $\eta_0$ and $\eta_3$ with $\eta_0 \gg \eta_3$. The viscosity coefficients for ions are $\eta_0^{\rm{i}} = 0.96 n T \tau_{\rm i} \gg \eta_3^{\rm{i}} = n T/2$ where $\tau_s$ is the normalized collisional time for the species $s$.\\
In what follows, we only use the part of the stress tensor perpendicular to the magnetic field lines. Given these approximations, the divergence of the stress tensor for ions becomes
\begin{equation}
\bm{\nabla} \cdot \overline{\overline{{\bm \pi}_{\rm i}}} = - \alpha_{\rm i} \tau_{\rm i} T \bm{\nabla} \left( n \bm{\nabla} \cdot {\bf v}_{\rm i} \right) + \frac{T}{2} \left[ n\hat{\mathbf{z}}\times \Delta {\bf v}_{\rm i} - (\hat{\mathbf{z}} \times \bm{\nabla} n \cdot \bm{\nabla}) {\bf v}_{\rm i} + (\bm{\nabla} n\cdot \bm{\nabla})(\hat{\mathbf{z}}\times{\bf v}_{\rm i}) \right],\nonumber
\end{equation}
where $\alpha_{\rm{i}} = 0.32$. The viscosity coefficients for electrons are $\eta_0^{\rm{e}} = 0.73 n \tau_{\rm e} \gg \eta_3^{\rm{e}} = -n \mu_{\rm{e}}/2$ where $\tau_{\rm e}$ is the collisional time for the electrons. We consider the following approximation for the divergence of the stress tensor for the electrons
\begin{equation}
\bm{\nabla} \cdot \overline{\overline{{\bm \pi}_{\rm e}}} = - \alpha_{\rm e} \tau_{\rm e} \bm{\nabla} \left( n \bm{\nabla} \cdot {\bf v}_{\rm e}\right),\nonumber
\end{equation}
where $\alpha_{\rm{e}} = 0.24$. The dimensionless equation for the velocity field of each species becomes
\begin{eqnarray}
\mu_s n \left( \frac{\partial}{\partial t} + {\bf v}_s \cdot \bm{\nabla} \right) {\bf v}_s &=& - n \bm{\nabla} \phi + n B {\bf v}_s \times \hat{\mathbf{z}} - e_s T_s \bm{\nabla} n + \alpha_s\tau_s e_s T_s \bm{\nabla} \left( n\bm{\nabla}\cdot {\bf v}_s \right) \nonumber \\
&& - \mu_s \frac{e_s T_s}{2} \left[ n \hat{\mathbf{z}} \times \Delta {\bf v}_s - (\hat{\mathbf{z}}\times\bm{\nabla} n \cdot \bm{\nabla}) {\bf v}_s + (\bm{\nabla} n\cdot \bm{\nabla}) (\hat{\mathbf{z}}\times {\bf v}_s) \right].
\label{eq:conserv_motion_dimless}
\end{eqnarray}
We apply the cross product with $\mathbf{\hat{z}}/nB$ to Eq.~(\ref{eq:conserv_motion_dimless}) in order to obtain the velocity as the sum of the ${\bf E}\times {\bf B}$, the diamagnetic, the polarization drifts and some gyroviscous contributions from the stress tensor,
\begin{eqnarray}
{\bf v}_s &=& \frac{\hat{\mathbf{z}} \times \bm{\nabla} \left( \phi + e_sT_s \log n \right) }{B} + \mu_s \frac{1}{B} \left( \frac{\partial}{\partial t} + {\bf v}_s \cdot \bm{\nabla} \right) \hat{\mathbf{z}} \times {\bf v}_s - \frac{\alpha_s\tau_s e_s T_s}{nB} \hat{\mathbf{z}} \times \bm{\nabla} ( n \bm{\nabla} \cdot {\bf v}_s ) \nonumber \\
&& - \mu_s \frac{e_s T_s}{2B}\left[ \Delta {\bf v}_s + (\hat{\mathbf{z}}\times \bm{\nabla} \log n \cdot \bm{\nabla}) (\hat{\mathbf{z}}\times {\bf v}_s) + (\bm{\nabla} \log n \cdot \bm{\nabla}) {\bf v}_s \right].\label{eq:velocity_s}
\end{eqnarray}
We expand all dimensionless quantities in the following way
\begin{eqnarray}
1/B &=& 1 + \varepsilon/\tilde{B}, \label{eq:ordering_B}\nonumber\\
\phi &=& \varepsilon \tilde{\phi}, \label{eq:ordering_phi}\nonumber\\
n &=& 1 + \varepsilon \tilde{n}, \label{eq:ordering_n}\nonumber\\
{\bf v}_s &=& \varepsilon {\bf v}_s^{(1)} + \varepsilon^2 {\bf v}_s^{(2)}. \label{eq:ordering_v}\nonumber
\end{eqnarray}
The small parameter $\varepsilon \ll 1$ denotes the amplitude of the fluctuations of the electromagnetic field as well as that of the dynamical variables $n$ and ${\bf v}_s$. Moreover, the temporal variations of these quantities are small so that the temporal variation is denoted $\varepsilon \partial/\partial t$. In what follows, we omit the tildas over the fluctuating quantities for simplicity.
Here we consider the model for weak collisionality (i.e.\ $\mu_s / \tau_s \ll 1$) such that $\tau_{\rm{i}}$ is of order $1/\varepsilon$ and $\tau_{\rm{e}} \ll 1/\varepsilon$. In addition $T$ is of order one, and the parameters $\alpha_s$ are of order $\varepsilon$ (i.e.\ $\alpha_{\rm{i}} \tau_{\rm{i}} T$ is of order $1$ and $\alpha_{\rm{e}} \tau_{\rm{e}} \ll 1$). The leading order of Eq.~(\ref{eq:velocity_s}) (terms proportional to $\varepsilon$) gives
$$
\left( 1 + \mu_s \frac{e_s T_s}{2}\Delta \right) {\bf v}_s^{(1)} = \hat{\mathbf{z}} \times \bm{\nabla} \left( \phi + e_s T_s n \right) - \alpha_s \tau_s e_s T_s \hat{\mathbf{z}} \times \bm{\nabla} ( \bm{\nabla} \cdot {\bf v}_s^{(1)} ).
$$
From the previous equation, we directly obtain that $\bm{\nabla} \cdot {\bf v}_{\rm e}^{(1)}=0$. From the expansion of Eq.~(\ref{eq:qnj}), we also conclude that $\bm{\nabla} \cdot {\bf v}_{\rm i}^{(1)}=0$. Therefore the equation for ${\bf v}_s^{(1)}$ becomes
\begin{equation}
\label{eq:v1mu}
\left( 1 + \mu_s \frac{e_s T_s}{2}\Delta \right) {\bf v}_s^{(1)} = \hat{\mathbf{z}} \times \bm{\nabla} \left( \phi + e_s T_s n \right),
\end{equation}
so that the first order velocity is the sum of the $\mathbf{E} \times \mathbf{B}$ velocity and the diamagnetic velocity with FLR corrections on the left hand side.
The second order of the velocity ${\bf v}_s$ is given by
\begin{eqnarray}
\left( 1 + \mu_s \frac{e_s T_s}{2}\Delta \right) \mathbf{v}_s^{(2)} &=& \frac{1}{B}{\bf v}_s^{(1)} + \mu_s \left( \frac{\partial}{\partial t} + {\bf v}_s^{(1)} \cdot \bm{\nabla} \right) \hat{\mathbf{z}} \times {\bf v}_s^{(1)} - eT_s n \hat{\mathbf{z}} \times \bm{\nabla} n \nonumber\\
&&- \alpha_s \tau_s e_s T_s \hat{\mathbf{z}} \times \bm{\nabla} \left( \bm{\nabla} \cdot {\bf v}_s^{(2)} \right) - \frac{\mu_s e_s T_s}{2}\left[ (\hat{\mathbf{z}}\times\bm{\nabla} n\cdot\bm{\nabla}) (\hat{\mathbf{z}}\times {\bf v}_s^{(1)}) + (\bm{\nabla} n\cdot\bm{\nabla}) {\bf v}_s^{(1)} \right],
\label{eq:v2mu}
\end{eqnarray}
where we have used Eq.~(\ref{eq:v1mu}) for ${\bf v}_s^{(1)}$.
We notice that we have also used the expansion $\bm{\nabla} \log(1+\varepsilon n)=\varepsilon \bm{\nabla} n - \varepsilon^2 n \bm{\nabla} n + O(\varepsilon^3)$.\\
As a first step, for the electrons, we see that from Eq.~(\ref{eq:v1mu}), the first order of the electron velocity is the sum of the $\mathbf{E} \times \mathbf{B}$ velocity and the electron diamagnetic velocity
\begin{equation}
{\bf v}_{\rm e}^{(1)} = \hat{\mathbf{z}} \times \bm{\nabla} \left( \phi - n \right),
\label{eq:ve1}
\end{equation}
and the second order of the electron velocity is
\begin{equation}
{\bf v}_{\rm e}^{(2)} = \frac{1}{B} {\bf v}_{\rm e}^{(1)} + n \hat{\mathbf{z}} \times \bm{\nabla} n + \alpha_{\rm e} \tau_{\rm e} \hat{\mathbf{z}} \times \bm{\nabla} \left( \bm{\nabla} \cdot {\bf v}_{\rm e}^{(2)} \right).
\label{eq:ve2}\nonumber
\end{equation}
From this equation, we get that the divergence of ${\bf v}_{\rm e}^{(2)}$ is equal to the divergence of ${\bf v}_{\rm e}^{(1)}/B$, and using Eq.~(\ref{eq:ve1}), it becomes
\begin{equation}
\bm{\nabla} \cdot {\bf v}_{\rm e}^{(2)} = \left[ \phi - n , \frac{1}{B} \right].
\label{eq:ve2_div}
\end{equation}
The continuity equation of the electron density at the second order in $\varepsilon$ becomes
$$
\frac{\partial n}{\partial t} + \bm{\nabla} \cdot {\bf v}_{\rm e}^{(2)} + {\bf v}_{\rm e}^{(1)} \cdot \bm{\nabla} n = 0,
$$
or equivalently, using Eqs.~(\ref{eq:ve1})-(\ref{eq:ve2_div}),
\begin{equation}
\frac{\partial n}{\partial t} = - \left[ \phi , n + \frac{1}{B} \right] + \left[ n , \frac{1}{B} \right].
\label{eq:Braginskii:continuity_e}\nonumber
\end{equation}
In fact, the first order of the ion velocity given by Eq.~(\ref{eq:v1mu}) suggests a change of variables $( 1+\frac{T}{2}\Delta ) \Phi = \phi + T n$ with which we can write the first order of the ion velocity into
\begin{equation}
{\bf v}_{\rm i}^{(1)} = \hat{\mathbf{z}} \times \bm{\nabla} \Phi.
\label{eq:vi1_1Phi}\nonumber
\end{equation}
We notice that this change of variables is also the one suggested by the gyromap [see Eq.~(\ref{eq:Gyromap:Phi-vs-phi})].
Using this variable $\Phi$, the continuity equation becomes
\begin{equation}
\label{eq:cee}
\frac{\partial n}{\partial t} = - \left[ \left( 1 + \frac{T}{2} \Delta \right) \Phi , n + \frac{1}{B} \right] + \left( 1 + T \right) \left[ n , \frac{1}{B} \right].
\end{equation}
As a second step, we work with the continuity equation for ions which is
\begin{equation}
\label{eq:cei}
\frac{\partial n}{\partial t} + \bm{\nabla} \cdot {\bf v}_{\rm i}^{(2)} + {\bf v}_{\rm i}^{(1)} \cdot \bm{\nabla} n = 0.
\end{equation}
First we obtain the following formula for the divergence of ${\bf v}^{(2)}_{\rm i}$ from Eq.~(\ref{eq:v2mu}),
\begin{equation}
\left( 1 + \frac{T}{2}\Delta \right) \bm{\nabla} \cdot {\bf v}_{\rm i}^{(2)} = - \frac{\partial \Delta \Phi}{\partial t} - \left[ \Phi , \Delta \Phi \right] + \left[ \Phi , \frac{1}{B} \right] - \frac{T}{2} \left[ \Delta \Phi , n \right] - T \left[ \bm{\nabla} \Phi ; \bm{\nabla} n \right].
\label{eq:vi2_div}
\end{equation}
Next we multiply Eq.~(\ref{eq:cei}) by $\left(1+T \Delta / 2 \right)$ in order to use the Eq.~({\ref{eq:vi2_div}) and we insert Eq.~(\ref{eq:cee}) so as to obtain
\begin{equation}
\frac{\partial \Delta \Phi}{\partial t} = - \left[ \Phi , \Delta \Phi \right] + \left( 1 + T \right) \left( 1 + \frac{T}{2}\Delta \right) \left[ n , \frac{1}{B} \right] + T \bm{\nabla} \cdot \left[ n + \frac{1}{B} , \bm{\nabla} \Phi \right] - \frac{T^2}{4} \Delta \left[ \Delta \Phi , n + \frac{1}{B} \right].
\label{eq:Braginskii:vorticity}
\end{equation}
We notice that Eqs.~(\ref{eq:cee})-(\ref{eq:Braginskii:vorticity}) coincide with Eqs.~(\ref{eq:Gyromap:continuity})-(\ref{eq:Gyromap:vorticity}), which were obtained by applying the gyromap. Therefore, we have shown that the terms generated from the gyromap are obtained consistently from Braginskii's closure for the stress tensor by making use of an appropriate ordering.\\
\\
In summary, it is possible to construct a model with FLR corrections and its Hamiltonian structure from a cold-ion model which possesses a Hamiltonian structure by applying a gyromap procedure which generates all the relevant FLR terms at the leading order.
The change of variables introduced by the gyromap is directly given by the definition of the stream function at the first order of the ion velocity. We have shown that the two-field reduced model~(\ref{eq:Gyromap:continuity})-(\ref{eq:Gyromap:vorticity}) is obtained using the Braginskii's closure for the stress tensor by considering an apt ordering on the dynamical variables.
\section*{Acknowledgements}
We acknowledge financial support from the Agence Nationale de la Recherche.
This work was supported by the European Community under the contract of Association between EURATOM, CEA and the French Research Federation for fusion studies. The views and opinions expressed herein do not necessarily reflect those of the European Commission.
|
\section{Introduction}
The Milky Way provides astronomers with a unique opportunity to explore the
formation and evolution of large spiral galaxies, as well as the nature of their
stellar populations and recognized structures. The key to this understanding
comes from the availability, for large numbers of individual stars, of the
powerful combination of six-dimensional phase-space information (location and
velocity) and chemical abundances. Metal-poor stars, in particular, shed light
on the early stages of galaxy formation and chemical evolution, as they
represent the fossil record of the first generations of stars that formed
shortly after the Big Bang.
Although theory suggests that the bulge of the Galaxy may harbor numerous
ancient (though not necessarily the most metal-poor) stars (e.g., Tumlinson
2010), the vast majority of presently recognized metal-poor stars are found in
the halo system of the Galaxy. According to Carollo et al. (2007; C07) and
Carollo et al. (2010; C10), the inner and outer halos possess different peak
metallicities ([Fe/H]$_{inner}$ $\sim -1.6$; [Fe/H]$_{outer}$ $\sim -2.2$), as
well as different spatial distributions, with the inner halo exhibiting a
flatter density profile than the nearly spherical outer halo. Their
analysis of the kinematics of a local sample of calibration stars from the
Sloan Digital Sky Survey (SDSS: York et al. 2000; Gunn et al. 2006)
indicated that the transition from dominance by the inner halo to the outer
halo occurs in the range 15-20 kpc from the Sun. A similar transition range
has been inferred from analysis of the ``vertical'' photometric stripes (de
Jong et al. 2010), obtained during the Sloan Extension for Galactic
Understanding and Exploration (SEGUE) sub-survey of SDSS-II (Yanny et al.
2009).
The papers by C07 and C10 also demonstrated that the inner-halo population
is essentially non-rotating, with V$_{\phi}$ = 7 $\pm$ 4 km~s$^{-1}$, while
the outer-halo population exhibits a significant retrograde signature, with
V$_{\phi}$ = $-$80 $\pm$ 13 km~s$^{-1}$ (where V$_{\phi}$ is the
Galactocentric rotational velocity). The velocity ellipsoids of these
populations differ as well, such that ($\sigma_{V_{R}}$,
$\sigma_{V_{\phi}}$, $\sigma_{V_{Z}})$ = (150 $\pm$ 2, 95 $\pm$ 2, 85 $\pm$
1) km~s$^{-1}$ for the inner halo and (159 $\pm$ 4, 165 $\pm$ 9, 116 $\pm$
3) km~s$^{-1}$ for the outer halo, evaluated in a Galactocentric
cylindrical reference frame.
The observed differences in the nature of the spatial distributions and
kinematics of the stellar populations associated with the inner- and outer-halo
components suggests that, in the context of modern hiearchical cosmogonies,
their progenitor mini-halos and subsequent merging and accretion scenarios
differed as well. If the inner halo formed from a limited number of moderately
massive mini-halos (see, e.g., Bullock \& Johnston 2005; Schlaufman et al. 2009,
2011), while the outer halo resulted from the accretion of more numerous, but
less massive ones (C07; Frebel et al. 2010; Norris et al. 2010a,b,c), one might
expect to find chemical signatures associated with the presently observed
stellar populations that reflect these differences. Previous studies have
provided hints that this may indeed be the case. For example, studies of the
[Mg/Fe] abundance ratios of stars thought to be associated with the inner halo
appear different (generally ~0.1 dex higher) than those associated with the
outer halo (Roederer 2009). This same study also demonstrated that stars
associated with the inner halo exhibit considerably lower star-to-star abundance
scatter for both the iron-peak element ratio [Ni/Fe] and the neutron-capture
element ratio [Ba/Fe] than found for stars of the outer halo. The recent study
by Nissen \& Schuster (2010) demonstrated that nearby dwarfs with halo
kinematics could be separated into two groups based on [$\alpha$/Fe]. They proposed that
the high-$\alpha$\ stars may have been born in the disk or bulge of the Milky Way
and heated to halo kinematics by merging satellite galaxies, or else were simply
members of the early generations of halo stars born during the collapse of a
proto-Galactic gas cloud, while the low-$\alpha$\ stars may have been accreted from
dwarf galaxies. Schlaufman et al. (2009) report detections of numerous elements
of cold halo sububstructure (ECHOS) in the inner halo, essentially
overdensities in radial-velocity space along the SEGUE sightlines. The ECHOS are
systematically more Fe-rich, but less $\alpha$-enhanced than the
kinematically smooth component of the inner halo. The ECHOS are also
chemically distinct from other Milky Way components; they are more Fe-poor
than typical thick-disk stars, and both more Fe-poor and $\alpha$-enhanced
than typical thin-disk stars. See Schlaufman et al. (2011) for a more
detailed discussion.
Chemically peculiar stars, such as the $\alpha$-element-enhanced very metal-poor
star BS~16934-002 (Aoki et al. 2007a; [Fe/H] $= -2.7$), the low [Mg/Fe]
($-0.1$), high [Ca/Fe] ($+$1.1) extremely metal-poor star SDSS~J2347+0108 (Lai
et al. 2009; [Fe/H] $= -3.2$), and the low [Si/Fe] ($-1.0$), low [Ca/Fe]
($-0.6$) extremely metal-poor star HE 1424-0241, identified by Cohen et al.
(2007; [Fe/H] $\sim -4.0$), all have inferred distances (and metallicities) that
suggest membership in the outer-halo population. All three of the previously recognized
ultra metal-poor stars ([Fe/H] $\leq -4.0$, HE~0557-4840; Norris et al. 2007) or
hyper metal-poor stars ([Fe/H] $\leq -5.0$, HE~0107-5240; Christlieb et al.
2002, and HE~1327-2326; Frebel et al. 2005) are similarly thought to be members
of the outer-halo population. Likewise, the newly discovered hyper metal-poor star
SDSS~J102915+172927 appears to have an orbit consistent with outer-halo
membership (Caffau et al. 2011). These results may all be related to the star
formation histories in the progenitor populations, their accretion histories, or
both.
In the present paper we focus on another possibly useful indicator of chemical
differences between the inner- and outer-halo components, the carbon-to-iron
ratio, [C/Fe], which we refer to as the ``carbonicity.'' In particular, we make
use of the SDSS/SEGUE calibration-star sample from SDSS DR7 (Abazajian et al.
2009) in order to search for possible contrasts between the frequency and degree
of carbon enhancement for the carbon-enhanced metal-poor (CEMP) stars from this
sample that can be kinematically associated with these two halo components.
The CEMP stars were originally defined as the subset of very metal-poor stars
([Fe/H] $\leq -$2) that exhibit elevated carbon relative to iron, [C/Fe] $>$
+1.0 (Beers \& Christlieb, 2005)\footnote{Other authors have used slightly
different criteria, e.g., [C/Fe] $> +0.7$ (Aoki et al. 2007b).}. In the last two
decades it has been recognized, primarily from spectroscopic follow-up of
metal-poor candidates selected from objective-prism surveys (e.g., Beers et al.
1985, 1992; Christlieb 2003), that roughly 20\% of stars with [Fe/H] $< -$2.0
exhibit enhanced carbonicity, up to several orders of magnitude larger than the
solar ratio (Marsteller et al. 2005; Lucatello et al. 2006). Some recent studies
(e.g., Cohen et al. 2005; Frebel et al. 2006), have claimed that this fraction
is a little lower (14\% and 9\%, respectively, for [Fe/H] $< -2.0$). The Frebel
et al. (2006) study is of particular interest, as the authors argued that the
relative fraction of CEMP stars appears to increase substantially with distance
above the Galactic plane, suggesting a possible connection with changes in the
underlying stellar populations. In any case, the fraction of CEMP stars rises to
30\% for [Fe/H] $< -3.0$, 40\% for [Fe/H] $< -3.5$, and 75\% for [Fe/H] $< -4.0$
(Beers \& Christlieb 2005; Frebel et al. 2005; Norris et al. 2007; Caffau et al.
2011 -- the new [Fe/H] $= -5.0$ star does not exhibit carbon enhancement);
definitive explanations for the origin of this increase have yet to be offered.
Regardless of the ultimate reason, these results indicate that significant
amounts of carbon were produced in the early stages of chemical evolution in the
universe.
There exist a number of classes of CEMP stars, some of which have been
associated with proposed progenitor objects. The CEMP-s stars (those with
$s-$process-element enhancement), for example, are the most commonly observed type to date.
High-resolution spectroscopic studies have revealed that around 80\% of CEMP
stars exhibit $s-$process-element enhancement (Aoki et al. 2007b). The favored
mechanism invoked to account for these stars is mass transfer of carbon-enhanced
material from the envelope of an asymptotic giant-branch (AGB) star to its
binary companion; it is this surviving companion that is now observed as a
CEMP-s star (e.g., Herwig 2005; Sneden et al. 2008; Bisterzo et al. 2011).
The class of CEMP-no stars (which exhibit no strong neutron-capture-element
enhancements) is particularly prevalent among the lowest metallicity stars (Fe/H
$< -$2.5; Beers \& Christlieb 2005; Aoki et al. 2007b). Possible progenitors for
this class include massive, rapidly rotating, mega metal-poor ([Fe/H] $< -6.0$)
stars, which models suggest have greatly enhanced abundances of CNO due to
distinctive internal burning and mixing episodes, followed by strong mass loss
(Hirschi et al. 2006; Meynet et al. 2006, 2010a,b). Another suggested mechanism
for the production of the material incorporated into CEMP-no stars is pollution
of the interstellar medium by so-called faint supernovae associated with the
first generations of stars, which experience extensive mixing and fallback
during their explosions (Umeda \& Nomoto 2003, 2005; Tominaga et al. 2007). This
model well reproduces the observed abundance pattern of the CEMP-no star
BD+44$^{\circ}$493, the ninth-magnitude [Fe/H] $= -3.7$ star (with [C/Fe] =
+1.3, [N/Fe] = +0.3, [O/Fe] = +1.6) discussed by Ito et al. (2009). The recently
reported high redshift ($z = 2.3$), extremely metal-poor Damped Lyman-$\alpha$
(DLA) system by Cooke et al. (2011; [Fe/H] $\sim -3.0$) exhibits enhanced
carbonicity ([C/Fe] $= +1.5$) and other elemental abundance signatures that
Kobayashi et al. (2011) also associate with production by faint supernovae.
It is also of interest that Matsuoka et al. (2011) have reported evidence for
strong carbon production in the early universe, based on their analysis of the
optical spectrum of the most distant known radio galaxy, TN J0924-2201, with $z
= 5.19$.
Below we seek to test if the increasing frequency of CEMP stars with declining
metallicity, and the suggested increase of the fraction of CEMP stars with
increasing distance from the Galactic plane, can be explained in the context of
an inner/outer halo dichotomy and the dominance of {\it different
carbon-production mechanisms} (the processes associated with the progenitors of the
CEMP-s and CEMP-no stars) being linked to these two populations.
This paper is outlined as follows. Section 2 describes the techniques used to
estimate the atmospheric parameters ($T_{\rm eff}$ , $\log~g$ , and [Fe/H]) and carbonicity
([C/Fe]) from the low-resolution SDSS spectra, compares our estimates with a
sample of very low-metallicity stars with available high-resolution
spectroscopic determinations, and obtains first-pass estimates of the fractions
of CEMP stars for various cuts on [Fe/H]. Section 3 summarizes the
calibration-star sample from SDSS DR7 we examine here, presents the
``as-observed'' distributions of metallicity and carbonicity for this sample as
functions of height above the Galactic plane, and carries out a comparison of
the distance and rotational velocity distributions for CEMP and non-CEMP stars.
Section 4 explores the global fraction of CEMP stars of this sample in the
low-metallicity regime, describes our adopted technique for derivation of
stellar population membership probabilities for the SDSS/SEGUE DR7 calibration
stars, and obtains estimates of the fractions of CEMP stars associated with the
inner- and outer-halo populations. Finally, Section 5 summarizes our main
results and considers their implications for the formation and evolution of the
Galactic halo populations.
\section{Atmospheric Parameter Estimates and [C/Fe] Ratios \\
for the SDSS/SEGUE DR7 Calibration-Star Sample}
\subsection{Atmospheric Parameter Estimates}
Estimates of the atmospheric parameters for our program stars were obtained from
the SEGUE Stellar Parameter Pipeline (SSPP papers I-V: Lee et al 2008a,b; Allende
Prieto et al. 2008; Smolinski et al. 2011; Lee et al. 2011). Typical internal
errors for stars in the temperature range that applies to the majority of the
calibration-star sample are $\sigma (T_{\rm eff}) \sim$ 125 K, $\sigma ({\log
g}) \sim $ 0.25 dex, and $\sigma {\rm [Fe/H]} \sim$ 0.20 dex. The external
errors in these determinations are of similar size, as discussed in the SSPP
references listed above.
In C10, a correction was applied for the metallicity determinations of the SSPP,
which we adopt here as well:
\begin{equation}
{\rm [Fe/H]}_C = -0.186 + 0.765*{\rm [Fe/H]}_A - 0.068*{\rm [Fe/H]}_A^2
\end{equation}
\noindent where [Fe/H]$_A$ is the adopted metallicity from the SSPP, and
[Fe/H]$_C$ is the corrected metallicity. This polynomial has little effect on
stars with metallicity greater than about [Fe/H] = $-2.5$, but lowers the
estimated metallicities for stars below this abundance by 0.1 to 0.2 dex, an
offset that was shown to exist between the DR7 SSPP-derived metallicities and
previous high-resolution spectroscopic measurements.
\subsection{Estimation of Carbon Abundance Ratios}
Carbon-to-iron abundance ratios ([C/Fe]) are estimated from the CH G-band
at $\sim$ 4300~{\AA} by matching the observed SDSS spectra near this
feature with an extensive grid of synthetic spectra. In order to construct
the grid we employed the NEWODF models of of Castelli \& Kurucz (2003).
Synthetic spectra were generated using the \texttt{turbospectrum} synthesis
code (Alvarez \& Plez 1998), which employs line broadening according to the
prescription of Barklem \& O'Mara (1998) and Barklem \& Aspelund-Johansson
(2005). The molecular species CH and CN are provided by B. Plez (private
communication; Plez \& Cohen 2005). The other linelists used are the same
as in Sivarani et al. (2006). For the purpose of this exercise we adopted
the solar abundances of Asplund et al. (2005).
The synthetic spectra cover wavelengths between 3600{\AA} and 4600{\AA}, with
original resolution of $\Delta$$\lambda$ = 0.005{\AA}, smoothed to the SDSS
resolving power ($R = 2000$) and re-binned to linear 1{\AA} pixels. The stellar
parameters of the grid cover the ranges 3500~K $\le $ $T_{\rm eff}$\ $\le 9750$~K (steps
of $250$~K), $0.0 \le$ $\log~g$\ $\le 5.0$ (steps of 0.5 dex) and $-2.5 \le$ [Fe/H]
$\le $ 0.0 (steps of $0.5$ dex). For stars with [Fe/H] $< -2.5$, models with
[Fe/H] $= -2.5$ were adopted. At the time this analysis was carried
out, lower metallicity models from this grid were not available. However, we did
have sparsely-spaced carbon-enhanced models generated by B.Plez, which extended
down to [Fe/H] $= -5.0$, for 4000~K $\le$ $T_{\rm eff}$\ $\le$ 6000~K. We carried out a
number of tests of our use of the [Fe/H] $= -2.5$ models extrapolated to lower
metallicities, which produced results in good agreement with Plez's models. Our
program stars include objects with temperatures above 6000~K, where any effects
due to extrapolation to lower metallicities will be lower still. Indeed Masseron
et al. (2005) points out that models with $T_{\rm eff}$\ $>$ 6000~K are not affected by
enhanced carbon. In any event, our past experience using models of lower
metallicity from other sources has indicated that very little changes when
dropping below [Fe/H] $= -2.5$. This is also indicated by the generally
excellent agreement of our [C/Fe] determinations with the high-resolution
results discussed below. The carbon abundance in the grid goes from [C/H] =
[Fe/H] $-$ 0.5 to [C/H] = +0.5 (the upper limit is in agreement with AGB models)
. For example, at [Fe/H] = $-$2.5, the grid covers the range from [C/H] = $-$3.0
to $+$0.5, which corresponds to the range of carbonicity $-0.5 <$ [C/Fe] $<$
$+$3.0.
Once constructed, we linearly interpolate within this grid, which is sufficient
for the size of the steps in the parameters. We have checked this by taking a
worst-case scenario, generating test synthetic spectra at low temperatures
($T_{\rm eff}$\ $= 3500$~K) and over the above ranges of gravity, metallicity, and
carbon abundance. The linearly interpolated grid was able to recover the input
parameters to within a few tenths of dex, which is consistent with our expected
errors in the method. Estimation of carbon abundance was accomplished using
chi-square minimization of the deviations between the observed and synthetic
spectra in the wavelength region between 4285 {\AA} and 4320 \AA, as carried out
by the IDL routine AMOEBA (a down-hill Simplex search procedure). The initial
value for [C/H] for the the global grid search was set to the same value as the
input stellar [Fe/H] ([C/Fe] = 0.0). During the search, the carbon abundances
are allowed to vary; all other stellar parameters are kept constant. In order to
provide some protection from falling into local minima, separate searches were
performed with lower and higher ranges of [C/Fe] considered. In almost all cases
these converged to the same minima as found for the global search. When not, we
took the value that resulted in the best match in the region of the CH G-band,
as judged from a correlation coefficient for the resulting match.
Figure 1 provides an example of the spectral matching process for determination
of [C/Fe] for a warm CEMP star in our sample. The upper panel shows the input
optical spectrum (black line) superposed with a synthetic spectrum with [C/Fe] =
0 (red line). The middle panel shows the best matches to the CH G-band obtained
from the three different ranges of [C/Fe] considered. The lower panel shows the
final adopted match.
Note that our procedures are not traditional synthesis analyses, but are based
on spectral matching. As part of this approach, the continuum-flattened observed
spectra must be registered to match similarly flattened synthetic spectra.
Because rectification of the stellar (and synthetic) continua is sometimes
imperfect, small deviations over localized regions of spectral range can
occasionally appear. Careful inspection of the bottom panel of Figure 1 reveals,
for example, a slight mismatch near the red end. We have taken care to minimize
the occurrence of these mismatches to the extent feasible, with particular
effort made in the range that is explored for performing the match to the CH
G-band. In any case, since the atmospheric parameters are set before conducting
the matching, slight registration offsets outside of the CH G-band region have
no effect on our derived [C/Fe]. We note that most of the weak features in the
region of the CH G-band are real, and not due to noise.
Similar procedures have been employed in previously published
papers (Beers et al. 2007; Kennedy et al. 2011), to which we refer the reader for
additional discussion of these techniques. We have also pursued a modest sanity
check by carrying out full syntheses for a small number of spectra using the
approach employed by Norris et al. (2010c), based on the synthesis code of
Cottrell \& Norris (1978). These comparisons indicate that we are able to
replicate derived [C/Fe] by our spectral matching approach to within 0.1-0.2
dex, at the one-sigma level in the precision for the majority of our
program stars.
Estimates of the errors in the final determinations of [C/Fe] were obtained
based on a set of noise-injection experiments for stars over a range of
temperatures and S/N ratios. These experiments indicated that, for stars with a
minimum S/N $= 15/1$ in the region of the G-band, [C/Fe] could be measured over
the $T_{\rm eff}$\ range of our sample with a maximum error of 0.5 dex, decreasing to
0.05 dex for the highest S/N spectra (S/N $> 50/1$). We assigned final errors to
the estimate of [C/Fe] using a linear function in S/N between these extremes.
Spectra not achieving the minimum S/N level (or which suffered from anomalies
such as pixel dropouts) were considered non-observations (i.e., they were
dropped from the sample). In addition, in order to claim a detection we required
that the equivalent width of the CH G-band obtain a minimum value of 1.2 \AA, a
value again chosen based on inspection of the noise-injection
experiments.\footnote{The CH G-band equivalent width measured in this study
covers a wider wavelength region than that defined by Beers et al.
(1999): 26 {\AA} (this paper) vs. 15 {\AA} (Beers et al. 1999), but
centered on the same wavelength (4305 \AA). It also makes use of a global fit to
the continuum, rather than the fixed sidebands employed previously. Thus,
although the equivalent widths are similar, they are not identical.} In cases
where this condition was not met, we consider the measured [C/Fe] an upper
limit.
\subsection{Comparison with High-Resolution Spectroscopic Estimates}
We obtain a check on our determinations of atmospheric parameters and
[C/Fe], for at least a subset of our stars, based on available high-resolution
follow-up spectra. There are three sources for our comparisons: Aoki et al.
(2008), Behara et al. (2010), and Aoki et al. (2011, in prep.). Since these
high-resolution spectroscopic programs were mostly interested in very and
extremely metal-poor stars, our comparison sample is dominated by stars with
[Fe/H] $< -2.5$. Table 1 lists the derived atmospheric parameters and carbon
abundance ratios for the stars in common (the label HIGH indicates the
high-resolution results, while the label SSPP refers to our analysis of the
low-resolution SDSS spectra). There are 23 unique objects listed, although one
star has only upper limits for [C/Fe] determined. Note that four stars in this
table were reported on by two sets of authors, which provides some feeling for
the level of systematic differences in the derived parameters. The Aoki et al.
(2011) paper, which supplies information for most of our comparison
sample, concluded that the temperature estimates provided by the SSPP were
sufficiently good that they simply adopted them for their analysis (although
they tested alternative methods, they could not improve upon the SSPP results).
As a result, a comparison of $T_{\rm eff}$\ determinations for the high-resolution and
low-resolution spectroscopic analyses cannot be fairly carried out. In the case
of stars that were observed by multiple groups, straight averages of the listed
parameters are used (except when the Aoki et al. 2011 {\it assigned} value of $\log~g$\ $=
+4.0$, as described below, can be replaced by a {\it measured} value from the other
sources).
Figure 2 shows a comparison of the derived parameter estimates. Robust
estimates of the zero-point offsets and rms scatter indicate good agreement
in metallicity ($\langle \Delta {\rm [Fe/H]}\rangle = -0.09$~dex;
$\sigma({\rm [Fe/H]}) = 0.27$~dex), close to the random errors expected for
carbon-normal stars from the SSPP (0.2 dex). Surface gravity exhibits an
acceptably small zero-point offset but a larger scatter ($\langle \Delta
(\log g)\rangle = -0.27$~dex; $\sigma(\log g) = 0.66$~dex). Larger
differences in the $\log~g$\ determinations might be expected for several
reasons. First, the comparison spectra for the majority of our program
stars (18 of 23) are based, at least in part, on ``snapshot''
high-resolution spectra (i.e., lower S/N) spectra reported by Aoki et al.
(2011). For these spectra, estimates of $\log~g$\ were fixed to $\log~g$\ = 4.0
for all stars with $T_{\rm eff}$\ $>$ 5500 K. This was done because the usual
procedure of matching Fe abundances based on Fe I and Fe II lines was not
uniformly possible for the warmer stars, due to the weakness of the Fe II
lines and the less than ideal S/N. According to Aoki et al. (2011), the
true surface gravities for such stars could lie anywhere in the range
$\log~g$\ = 3.5 to $\log~g$\ = 4.5. Note that the reported surface-gravity
offset and scatter above does not include the three Aoki et al. (2011)
stars that did not have reported estimates from other sources.
Surface gravity estimates from the high-resolution spectra of giants with $T_{\rm eff}$
$ < 5500$~K are determined by Aoki et al. (2011) based on analysis of the Fe I
and Fe II lines, in the usual manner. Note that Aoki et al. also identified four
stars in this temperature range to be cool main-sequence stars, rather than
giants, based on the weak lines of their ionized species. For these stars,
gravity estimates are obtained by matching to isochrones for metal-poor
main-sequence stars (a similar process to that carried out by Aoki et al. 2010).
It is expected that the high-resolution analysis estimates of $\log~g$\ for the cooler
stars are precise at a level no better than 0.3 dex. Secondly, difficulty in
estimation of $\log~g$\ for CEMP stars might not be suprising, given that molecular
carbon bands (in particular for later type stars) can easily confound
gravity-sensitive features in a low-resolution spectrum. However, in our
analysis we have taken care to avoid regions of the spectrum where molecular
carbon bands have corrupted the gravity-sensitive features. In reality, this
problem is only encountered for the coolest stars with the strongest molecular
bands, which are a distinct minority of our sample. Our sample only includes
stars in the temperature range 4500~K $<$ $T_{\rm eff}$\ $<$ 7000 ~K, and among the 813
carbon-rich stars considered in our analysis below, none of them have $T_{\rm eff}$\ $<$
5000~K, and only 10\% have $T_{\rm eff}$\ $<$ 5750~K. Taking a global approximation that
the high-resolution determinations of surface gravity contribute 0.4 dex to the
rms scatter comparison with the low-resolution estimates, this indicates that
the SSPP estimates for these stars have an external error of determination of
$\sqrt{(0.66^2 - 0.4^2)}$ = 0.52 dex. We also note that the great majority of
the stars analysed in our sample have [Fe/H] $>$ $-$2.5, for which surface
gravity estimates should be better determined, due to the increasing strength of
their gravity-sensitive metallic features.
The agreement in estimates of the carbonicity is quite good ($\langle \Delta
{\rm [C/Fe]}\rangle = +0.05$~dex; $\sigma({\rm [C/Fe]}) = 0.29$~dex), since we
expect even the high-resolution determinations to be precise to no better than
about 0.15-0.20 dex. This suggests that our external errors for estimates of
[C/Fe] are on the order of 0.25 dex. Note that this level of agreement between
[C/Fe] determinations based on the high-resolution and SSPP analyses would not
be possible if the rough estimates of $\log~g$\ had a strong influence on our
procedures. However, at the suggestion of an anonymous referee, we have carried
out explicit tests of the effect of incorrect surface-gravity determinations on
the derived [C/Fe].
In order to test the impact on the derived [C/Fe] from the adoption of an incorrect
surface gravity, we have used synthetic spectra (with known atmospheric
parameters and [C/Fe]) from our grid covering three fixed metallicities, [Fe/H]
= $-1.0$, $-1.5$, and $-2.5$, three fixed gravities, $\log~g$\ = 2.0, 3.0, and 4.0,
which spans the range of the majority of our program stars, and two levels of
carbonicity, [C/Fe] = 0.0 and [C/Fe] = +1.5. We then intentionally perturbed their
input $\log~g$\ values by $-1.0$ dex, $-0.5$ dex, $+0.5$ dex, and $+1.0$ dex, and
derived estimates of their [C/Fe] following the procedures described above. As
can be seen from inspection of Figure 3 (which shows the case for [C/Fe] = 0.0),
the effects on estimates of [C/Fe] never exceed 0.5 dex (and then only in the
most extreme cases of $\pm$ 1.0 dex variation in $\log~g$\ ), and are generally much
smaller than that. Not surprisingly, the largest variations occur for the
warmest stars, which have weaker CH G-band features. However, our sample
included only a small fraction of stars with $T_{\rm eff}$\ $>$ 6500~K ($\sim 3$ \%), so
this is not expected to have a major effect. The mean zero-point offsets and rms
variations in the derived [C/Fe], relative to the known value across all $T_{\rm eff}$\
and $\log~g$\ considered, are shown for each panel in Figure 3, and are acceptably
small. Similar results apply to the case when we fix the input [C/Fe] = +1.5.
Figure 3 considered the solar [C/Fe] case, since we are more concerned with
false positives that would cause us to count a star as carbon-enhanced when it
is not. We conclude that, while some sensitivity to estimates of [C/Fe] may
exist due to errors in estimates of $\log~g$\ , its impact on our results should be
minimal, except for truly extreme cases.
\subsection {Detections, Upper Limits, and First-Pass Frequencies of CEMP Stars}
In this work there are 31187 unique stars for which estimates of [C/Fe]
were carried out (other objects were either repeats, in which case they
were averaged, had insufficient spectral S/N ratios, were clear cases of
QSOs, cool white dwarfs, or very late-type stars, or had some spectral
defect in the region of the CH G-band which prevented measurements being
obtained -- all such stars are dropped from the subsequent analysis). The
remaining sample can be divided into two categories: stars that have a
measured [C/Fe] (with the G-band detected; N$^{D}$ = 25647), and stars for
which only an upper limit on [C/Fe] has been obtained (G-band undetected;
N$^{L}$ = 5540). We call the associated two sets of stars subsample D and
subsample L, respectively. Figure 4 shows [C/Fe] as a function of
metallicity for these two categories. The ``ridge lines" in both of the
panels are due to grid effects in the chi-square matching procedure. It is
worth noting that most of the stars with measured [C/Fe] (subsample D)
exhibit carbonicity below [C/Fe] = +0.7, which indeed appears to be a
natural dividing threshold between carbon-normal stars and carbon-rich
stars. This limit was also established by Aoki et al. (2007b), who also
included in their analysis expected evolutionary effects on the definition
of CEMP stars. We adopt this value of [C/Fe] to define the carbon-rich stars in
our sample, without making any luminosity correction adjustment, since our
sample includes very few cool giants. There is an evident correlation
between [C/Fe] and [Fe/H] seen in this figure, in particular for stars with
[Fe/H] $< -$ 1.0. The relative number of stars with carbon excesses
increases as the metallicity decreases, as does the level of carbonicity,
as reported previously by several studies based on much smaller samples of
stars.
Adopting the definition of carbon-rich stars as above, we distinguish two
subcategories within subsamples D and L: stars with [C/Fe] $\leq$ +0.7 (C-norm),
and stars with [C/Fe] $>$ +0.7 (C-rich). In the case of subsample L, the carbon
status for a star can be assessed with certainty only when [C/Fe] $\leq$ +0.7,
and remains unknown for stars having [C/Fe] $>$ +0.7. Indeed, suppose that the
limit assigned to a given star is [C/Fe]$_{lim}$ = +1.5. Then, all values below
[C/Fe]$_{lim}$ can still be accepted for that star, including carbonicity
well below [C/Fe] = +0.7. When the upper limit is below [C/Fe] = +0.7, however,
the carbon status of the star can be assessed as C-norm. Not surprisingly, there
is a strong temperature effect in the assessment of the carbon status for stars
in our sample. For example, stars with higher $T_{\rm eff}$\ (above $T_{\rm eff}$\ $\sim
6250$~K) would require quite high [C/Fe] for the CH G-band to be detected; stars
without a detected CH G-band are included in the L subsample. Thus, the fractions
of CEMP stars we derive in this paper are lower limits to the true fractions.
In a future paper we plan to obtain an explicit correction function to account
for the ``missing'' stars due to this temperature effect, in order to assess its
impact on the derived frequencies of CEMP stars.
Taking into account the above definitions, the fraction of C-rich stars can
be formulated as:
\begin{equation}
F_{C-rich} = \frac{N^{D}_{C-rich}}{N^{D}_{C-rich} + N^{D}_{C-norm} + N^{L}_{C-norm}}
\end{equation}
\medskip
\noindent where $N^{D}_{C-norm}$ and $N^{D}_{C-rich}$ are stars belonging to subsample D
and having [C/Fe] $\leq$ +0.7 and [C/Fe] $>$ +0.7, respectively, while
N$^{L}_{C-norm}$ are stars in subsample L and having [C/Fe] $\leq$ +0.7. Stars with
unknown carbon status are not included in the above definition.
Table 2 reports the number of stars belonging to the various categories for
different ranges of metallicity, as well as the total fractions of C-rich
stars. Note that, when comparing the reported fractions in this table with other
fractions reported in this paper, the Table 2 fractions have no restriction
on whether or not acceptable proper motion and radial velocity measurements were
available for a given star.
\section{The Nature of the Carbon-Rich Star Sample}
\subsection{Radial Velocities, Distance Estimates, and
Definition\\ of the Extended and Local Samples}
We begin with the 31187 unique DR7 calibration stars for which estimates of
[C/Fe] exist, and with distances slightly revised from those presented in C07
and C10, as described below. Details concerning the nature of the
calibration-star sample can be found in C10; here we recall a few facts
concerning these stars that are of importance for the present analysis.
Spectra of the SDSS/SEGUE calibration stars were obtained to perform
spectrophotometric corrections and to calibrate and remove the night-sky
emission and absorption features (telluric absorption) from SDSS spectra. The
spectrophotometric calibration stars cover the apparent magnitude range 15.5 $<
g_{0} <$ 17.0, and satisfy the color ranges 0.6 $< (u-g) _{0} <$ 1.2 ; 0.0 $<
(g-r) _{0} <$ 0.6.\footnote{The subscript 0 in the magnitudes and colors
indicates that they are corrected for the effects of interstellar absorption and
reddening, based on the dust maps of Schlegel et al. (1998).} The telluric
calibration stars cover the same color ranges as the spectrophotometric
calibration stars, but at fainter apparent magnitudes, in the range 17.0 $<
g_{0} <$ 18.5.
The C10 paper describes the radial velocity estimates (which have a precision of 5-20
km~s$^{-1}$, depending on the S/N ratio of the spectrum, and with a negligible
zero-point offset), as well as photometric distance estimates, obtained by using
the SSPP surface-gravity estimate for luminosity classification, then following
the procedures of Beers et al. (2000). Since the estimated $\log~g$\ is used only
for classification, this means that for stars redder than than main-sequence
turnoff (MSTO), the expected large differences in surface gravity for dwarfs and
giants make even approximate $\log~g$\ estimates sufficient -- errors of 1 to 2 dex
in $\log~g$\ would have to be routinely made in order to confound this procedure.
Close to the MSTO, any method of photometric distance estimation becomes more
problematic, but the difference in the derived distances for stars just above or
just below the MSTO decreases the closer they are to it, which mitigates against
these effects.
Sch\"onrich et al. (2011) have criticized the Beers et al. (2000) method of
distance determination, and the results of the C07 and C10 that relied on it.
They claimed that the counter-rotating halo found in C07 and C10 is a result of
biases in distance estimates for main-sequence dwarfs, and furthermore that the
distance derivation is influenced by sorting a subset of the stars into
incorrect positions in the color-magnitude diagram of an old, metal-poor
population. In a rebuttal paper, Beers et al. (2011) demonstrated that the
the Sch\"onrich et al. claims concerning dwarf distances are incorrect (due to
their adoption of the wrong main-sequence absolute magnitude relationship from
Ivezi\'c et al. 2008). Furthermore, the claimed retrograde tail in the
rotation-velocity distribution was shown to arise from the measured
asymmetric proper motions, and was not the result of the propagation of
distance errors. In any event, we have applied the procedure suggested in
the rebuttal paper to reassign intermediate-gravity stars that were
originally classified as main-sequence turnoff stars into dwarf or
subgiant/giant luminosity classifications, in cases where their derived
$T_{\rm eff}$\ were substantially lower than expected for the turnoff. Revised
distances for these stars have been adopted as well. A reassessment of the
kinematics indicates that the retrograde signal for the subsample at very
low metallicity ([Fe/H] $< -$2.0) remains. The interested reader can find
more details in Beers et al. (2011). Based on the discussion in that paper,
we believe that our distances should be accurate to on the order of
15-20\%.
The C10 procedures applied a series of cuts to their sample designed to better
enable measurement of the kinematic and orbital properties of the various
stellar populations considered. This produced a subsample of stars in the local
volume (distance from the Sun $d < 4$ kpc, and with 7 kpc $< R <$ 10 kpc, where
$R$ is the Galactocentric distance projected onto the plane of the Galaxy),
which they referred to as the ``Local Sample" (N $\sim$ 17000). These
restrictions were made in order to mitigate against the increase in the errors
in the derived transverse velocities, which scale with distance from the Sun,
and to improve the applicability of the simple models for the adopted form of
the Galactic potential. For our present analysis we do not need to redetermine
the velocity parameters of the underlying stellar populations, so we can relax
these cuts somewhat. As described below, in order to increase the numbers of
stars in our sample with measured [C/Fe], we have changed the constraint on the
distance from the Sun, from $d < $ 4 kpc to $d <$ 10 kpc, and removed the
constraint on the projected distance $R$. With these relaxed cuts the total
number of stars climbs to 30874. We refer to this new sample as the ``Extended
Sample."
The upper panel of Figure 5 shows the as-observed metallicity distribution
function (MDF) for the DR7 calibration stars belonging to the Extended Sample
(black histogram), as well as for stars in the C10 Local Sample, as described
above (red histogram). We use the term ``as-observed'' in order to call
attention to the fact that the selection functions for the calibration stars
were not intended to return a fair sample of stars, suitable for an unbiased
analysis of the distribution of metallicity for stars in the Galaxy. Rather, the
calibration stars were selected to emphasize the numbers of moderately
low-metallicity stars that might serve to best constrain the spectrophotometry
and telluric line corrections carried out as part of the SDSS spectroscopic
pipeline reductions. Thus, these MDFs are a ``sample of convenience,''
one that is still useful for providing guidance as to the presence of various
stellar populations in metallicity space, but not for obtaining estimates of
their relative normalizations (for which other samples drawn from SDSS are more
suitable). No selection for or against carbon-enhanced stars was carried out in
the selecton of the calibration-star sample.
The MDFs of the two samples are clearly very similar, and comprise stars with
metallicities that sample all of the primary stellar components of the Galaxy
(with the exception of the bulge). The lower panel of Figure 5 is the
distribution of [C/Fe] (estimated as described above), which we refer to as the
carbonicity distribution function (CarDF) for the Extended Sample (black
histogram), and for the Local Sample (red histogram) \footnote{We employ the
term ``carbonicity distribution function'' to emphasize that we are describing
the carbon-to-iron abundance ratio, rather than the carbon abundance
distribution itself.}. The inset in the panel shows a rescaling appropriate for
the high [C/Fe] tail of this distribution. As in the case of the MDFs, the shape
of the two CarDFs are similar, with a strong peak at [C/Fe] $\sim$ +0.2 to +0.3, and two
tails, a weak one at low carbonicity ([C/Fe] $<$ 0), and a strong
one that extends to high carbonicity, +0.5 $<$ [C/Fe] $<$ +3.0. The total number
of stars with high carbonicity is significantly different in the two samples.
Indeed, for the Extended Sample we find 728 stars with [C/Fe] $>$ +0.7, while
for the Local Sample the number is reduced to 318.
\subsection{As-Observed MDF and CarDF of the Extended Sample as a Function of
Distance from the Galactic Plane}
We now examine the MDFs and the CarDFs of the Extended Sample of calibration
stars for different intervals in $|$Z$|$, with cuts chosen to ensure there
remain adequate numbers of stars in each interval.
In Figure 6, the first (left-hand) column and the third column show the MDFs,
while the second column and the fourth (right-hand) column are the CarDFs. In
the first and third columns, the red arrows indicate the peak metallicities of
the various stellar populations considered by C10. In the second and fourth
column, the blue arrows show the location of the solar carbon-to-iron ratio
([C/Fe] = 0.0), and the location of the natural threshold that divides
carbon-normal stars from carbon-rich stars ([C/Fe] = +0.7), as identified above.
Examination of the first column of panels in Figure 6 shows how the MDF changes
from the upper-left panel, in which there are obvious contributions from the
thick-disk, the metal-weak thick disk (MWTD), and inner-halo components in the
cuts close to the Galactic plane, to the lower-left panel, with an MDF dominated
primarily by inner-halo stars. In the third column of panels, with distances
from the plane greater than 5 kpc, the transition from inner-halo dominance to a
much greater contribution from outer-halo stars is clear. This demonstration is,
by design, independent of any errors that might arise from derivation of the
kinematic parameters, and provides confirmation of the difference in the
chemical properties of the inner- and outer-halo populations originally
suggested by C07. The second and fourth columns show the results of the same
exercise for the CarDFs. Close to the Galactic plane (second column; up to
$|$Z$|$ = 3 kpc), where the thick disk and MWTD are the dominant components, the
CarDF is strongly peaked at values between [C/Fe] = 0.0 and [C/Fe] = +0.3. The
CarDFs of the thick disk and MWTD will be explored in a future paper. Here we
simply note that in the regions close to the plane, where these components
dominate the sample, there are not many stars populating regions of
high carbonicity. At larger distances from the Galactic plane, where the inner
halo begins to be the dominant component, there appears a tail in the
carbonicity distribution towards higher values, [C/Fe] $>$ +0.5. In the fourth
column of panels, where the distance from the Galactic plane is $|$Z$|$ $>$ 5
kpc, and where we expect to see the beginning of the transition from inner-halo
dominance to outer-halo dominance, the tails towards high [C/Fe] values become
even more evident. The CarDF exhibits a strong peak at [C/Fe] $\sim$ +0.3 and a long
tail towards high values of carbonicity, up to [C/Fe] = +3.0. The nature of the
CarDF is likely to be influenced by the change of the MDF as a function of the
distance from the Galactic plane, due the well-known trend of increasing [C/Fe]
with declining [Fe/H]. However, as discussed below, there is evidence that the
observed changes may reflect {\it real differences} in the chemistry of the
inner- and outer-halo populations, even at a given (low) metallicity.
By adopting the threshold of [C/Fe] $> +0.7$, the fraction of SDSS/SEGUE
calibration stars with high carbonicity in the subsample at $|$Z$|$ $>$ 9
is 20\%, (which, as argued above, is a lower limit), in line with previous
estimates for stars with [Fe/H] $< -2.0$.
\subsection{Comparisons of Distance and Rotational Velocity
Distributions\\ for CEMP and non-CEMP Stars}
We now consider how the nature of our Extended Sample differs for the
low-metallicity ([Fe/H] $< -1.5$) CEMP ([C/Fe] $> +0.7$) and non-CEMP ([C/Fe] $<$
+0.7]) stars. Figure 7 shows two columns of panels, corresponding to the stars
in our sample with different cuts on Z$_{max}$ (the maximum distance of a
stellar orbit above or below the Galactic plane); the left-hand column includes
stars at all Z$_{\rm max}$, while the right-hand column only include stars
satisfying Z$_{\rm max}$ $>$ 5 kpc. The top panels show the distributions of
distance, $d$, while the bottom panels are the distribution of the
Galactocentric rotational velocity, V$_{\phi}$. For all panels, CEMP stars are
shown as red dot-dashed histograms; non-CEMP stars are shown
as black solid histograms.
As can be seen in the top panel of each column, the peak of the distance
distribution in both cases is, for the non-CEMP stars, $d \sim 2$ kpc, while for
the CEMP stars, a softer peak is seen around $d \sim$ 2.5-4 kpc (closer
to $d \sim$ 4 kpc for the higher Z$_{\rm max}$ cut), with long tails extending
to the 10 kpc cutoff of the Extended Sample. At all distances beyond $d \sim$ 3
kpc the relative fraction of CEMP stars exceeds that of the non-CEMP stars,
while closer than $d \sim$ 3 kpc the relative fraction of non-CEMP stars is
greater than that of the CEMP stars. A Kolmogorov-Smirnov (K-S) test of the
distance distributions indicates that the hypothesis that the CEMP and non-CEMP
stars are drawn from the same parent population is rejected at high significance
($p < 0.001$) for both cuts on Z$_{\rm max}$.
For the case of Z$_{max} > 0$ kpc, stars classified as dwarfs represent 72\% of
the non-CEMP sample and 42\% of the CEMP sample, while subgiants and giants
represent 24\% of the non-CEMP sample and 50\% of the CEMP sample. For both the
non-CEMP and CEMP samples, main-sequence turnoff stars represent less than 10\%
of the samples. For the case of Z$_{max} > 5$ kpc, dwarfs comprise 52\% of the
non-CEMP sample and 29\% of the CEMP sample, while subgiants/giants represent
44\% of the non-CEMP sample and 63\% of the CEMP sample. Again, main-sequence
turnoff stars comprise less the 10\% of both samples. Thus, the difference in
distance distribution is not solely dependent on the luminosity classes
associated with the populations split on carbonicity; dwarfs and
subgiants/giants are both significantly represented in each of the non-CEMP and
CEMP samples.
From inspection of the bottom panels of Figure 7, the non-CEMP stars exhibit a
slightly asymmetric distribution of rotational velocities centered close to
V$_{\phi} \sim 0$ km s$^{-1}$, and a weak retrograde tail. By contrast, the CEMP
stars exhibit a rather strong asymmetry extending to large retrograde
velocities. A K-S test of the rotational-velocity distributions indicates that
the hypothesis that the CEMP and non-CEMP stars are drawn from the same parent
population is rejected at high significance ($p < 0.001$) for both cuts on
Z$_{\rm max}$.
It is interesting to consider the distribution of luminosity classes for the
split on [C/Fe] for the retrograde stars with V$_{\phi} < -100 $ km s$^{-1}$.
For the case of Z$_{\rm max} > 0$ kpc, stars in the retrograde tail classified as
dwarfs represent 66\% of the non-CEMP sample and 20\% of the CEMP sample, while
subgiants/giants represent 30\% of the non-CEMP sample and 73\% of the CEMP sample. For
both non-CEMP and CEMP samples main-sequence turnoff stars represent less than
10\% of the samples. For the case of Z$_{\rm max} > 5$ kpc, the retrograde stars
classified as dwarfs comprise 48\% of the non-CEMP sample and 18\% of the CEMP
sample, while subgiants/giants represent 47\% of the non-CEMP sample and 75\% of
the CEMP sample. Again, main-sequence turnoff stars comprise less than 10\% of
both samples. Significant fractions of dwarfs and giants are present in the
retrograde tail for both the non-CEMP and CEMP samples. Thus, the retrograde
signature is unlikely to be due to a preponderance of stars with aberrant
distance estimates owing to luminosity misclassifications.
If we specialize to the highly retrograde tails, we find that for $Z_{\rm max}$\ $> 0$
kpc, the portion of the non-CEMP sample in this tail is only 8\% for $V_{\rm \phi}$\ $<
-150$ km s$^{-1}$\ and 4\% for $V_{\rm \phi}$\ $< -200$ km s$^{-1}$\ . For the CEMP stars, these
portions are 17\% for $V_{\rm \phi}$\ $< -150$ km s$^{-1}$\ and 12\% for $V_{\rm \phi}$\ $< -200$ km s$^{-1}$\ ,
respectively. For the case of $Z_{\rm max}$\ $> 5$ kpc, the portion of the non-CEMP
sample is 12\% for $V_{\rm \phi}$\ $< -150$ km s$^{-1}$\ and 8\% for $V_{\rm \phi}$\ $< -200$ km s$^{-1}$\ . For
the CEMP stars, these portions are 24\% for $V_{\rm \phi}$\ $< -150$ km s$^{-1}$\ and 19\% for
$V_{\rm \phi}$\ $< -200$ km s$^{-1}$\ , respectively. Thus, in both cases, a split on the level
of carbon enhancement leads directly to rather different relative population of
the retrograde tails, compared to the full distributions of non-CEMP and CEMP
stars. Identifying the highly retrograde tails with the outer-halo population,
we can already infer that the outer-halo component appears to harbor a greater
fraction of CEMP stars than the inner-halo component.
In addition, before considering a more complete discussion below, we can
make use of the $V_{\rm \phi}$\ distribution to obtain an estimate of the approximate
fraction of CEMP stars in the outer halo. We proceed by asserting that the stars
in the highly retrograde tail, with $V_{\rm \phi}$\ $< -200$ km s$^{-1}$\ , are very likely to be
members of the outer-halo population. This follows because the
dispersion in $V_{\rm \phi}$\ for an essentially non-rotating inner halo is on the order
of 100 km s$^{-1}$\ , and placing a cut at two sigma below the mean rotation of the
inner halo excludes all but 2.5\% of likely inner-halo stars. With this
assumption, we find that 11\% of the stars in the highly retrograde tail (and
with $Z_{\rm max}$\ $>$ 5) kpc) are CEMP stars. This differs from the calculations
immediately above, in that we are considering the fractions of CEMP stars
relative to the total number of stars (including those from the L subsample, but
not those with unknown carbon status; i.e., we are using Eqn. 2). Recall that
this calculation applies only for stars with [Fe/H] $< -1.5$. For this
subsample we find a mean carbonicity of $\langle$[C/Fe]$\rangle$ = +1.47 $\pm$
0.07, where the error is the standard error of the mean. This value can be
taken as a first-pass estimate of the mean outer-halo carbonicity.
Figure 8 shows the observed distributions of the measured proper motions in
the right ascension and declination directions for the low-metallicity ([Fe/H] $<
-1.5$) stars in our sample, for two cuts on $Z_{\rm max}$, represented by small blue
dots. The same panels show, represented as red stars, the objects that populate
the highly retrograde tails of the distribution of rotational velocity shown in
Figure 7. Note that cuts at $V_{\rm \phi}$\ $< -$150 km s$^{-1}$\ and $V_{\rm \phi}$\ $< -$200 km s$^{-1}$\ are
shown in the upper and lower rows of panels, respectively. As can be appreciated
from inspection of this figure, the proper motions in both directions for the
stars assigned to the highly retrograde tail are asymmetrically distributed, and
explore much larger values, relative to the rest of the sample. This supports
the reality of the highly retrograde signature seen in Figure 7, and indicates
that it is due primarily to the proper motions of the participating stars. See
Beers et al. (2011) for additional details concerning the veracity and
interpretation of SDSS proper motions for the SDSS calibration-star sample.
\section{Extreme Deconvolution and Membership Probabilities}
In the relatively nearby volume explored by the SDSS/SEGUE calibration-star
sample, C07 and C10 have shown that the two stellar components of the Galactic
halo are strongly overlapped in their spatial distribution, velocity ellipsoids,
and MDFs. Thus, to explore possible differences in the frequencies and mean
carbonicities of CEMP stars, it is essential to assign inner- and
outer-halo membership probabilities to each star in the sample that is likely to
belong to the halo system. We describe how this is accomplished in the sections
below.
\subsection{Basic Parameters}
In the solar neighborhood, the set of parameters that best identify the presence
of the main structures are the distance from the Galactic plane, the rotational
velocity of a star in a cylindrical frame with respect to the Galactic center,
and the stellar metallicity (see C10). For consideration of the disk system of
the Galaxy, the distance from the Galactic plane is best represented by $|$Z$|$
(the present distance of a star above or below the Galactic plane), while for
the halo components, a more suitable distance is Z$_{\rm max}$, which depends on the
adopted gravitational potential. The choice of Z$_{\rm max}$ for the Galactic halo
is necessary because of the much larger spatial extent of its two primary
components, relative to that of the disk components.
Before seeking a deconvolution of the inner- and outer-halo components, we need
to check for possible correlations between all of the basic parameters we wish to
employ. In C10 we have shown that Z$_{\rm max}$ has a significant correlation with
V$_{R}$ and V$_{Z}$, but not with V$_{\phi}$, other than that expected from the
presence of the thick-disk population at high positive rotation velocity and the
halo at lower rotational velocity. A very similar behavior is confirmed for the
Extended Sample as well (Figure 9). We conclude that the Galactocentric
rotational velocity, V$_{\phi}$, and the vertical distance, Z$_{\rm max}$, when
combined with metallicity, can be used to obtain useful information on the
different stellar populations present in the Extended Sample.
The primary chemical parameter of the present analysis is the carbonicity,
[C/Fe], so it is important to check for its possible correlations with the
basic parameters defined above. The result of this exercise is shown in
Figure 10. Here, the upper panel shows the Galactocentric rotational
velocity as a function of metallicity for the Extended Sample. The gray
dots represent the stars in the sample with [C/Fe] $< +0.7$, while the red
dots denote the stars with enhanced carbonicity, [C/Fe] $> +0.7$. It is
worth noting that the stars with [C/Fe] $>$ +0.7 are \emph{almost all}
located in the halo components of the Galaxy, with few exceptions. This is
perhaps not surprising, as the halo components of the Milky Way are very
metal poor. The middle panel of Figure 10 shows the Galactocentric
rotational velocity, $V_{\rm \phi}$\ , as a function of [C/Fe]. This plot shows no
evidence of correlation between V$_{\phi}$ and [C/Fe]. Finally, the lower
panel of Figure 10 shows [Fe/H] as a function of [C/Fe]. Here, there is
clear evidence of a correlation between [C/Fe] and [Fe/H] -- the
carbonicity increases as the metallicity decreases. A similar trend was
already noticed in past works, such as Rossi et al. (2005) and Lucatello et
al. (2006). We consider this result in more detail below.
\subsection{CEMP Fractions in the Low Metallicity Regime: Global Behavior}
The left panel of Figure 11 shows the fraction of CEMP stars in the
Extended Sample, as a function of [Fe/H], in the metal-poor regime ([Fe/H]
$<$ $-$1.5), and at vertical distance Z$_{\rm max}$ $>$ 5 kpc (chosen to
avoid thick disk and MWTD contamination). Here, each bin in metallicity
corresponds to an interval of $\Delta$[Fe/H] = 0.2 dex, with the exception
of the last bin, where stars are selected in the range [Fe/H] $<$ $-$2.6.
The fractions of CEMP stars in each bin are obtained by selecting objects
with [C/Fe] $>$ +0.7, and application of Eqn. 2. Errors on the CEMP
fraction are evaluated through the jackknife approach. This technique is
similar to bootstrapping, but instead of sampling with replacement, it
recomputes the statistical estimate leaving out one observation at a time
from the sample (Wall \& Jenkins 2003). The dash-dotted line is a
second-order polynomial fit to the data.
The increase of the fraction of CEMP stars with declining metallicity
pertains to the global behavior of the data in the metal-poor regime. If
metallicity alone is the driver of the carbon-enhancement phenomenon, one
might wonder if the strong increase in the fraction of CEMP stars at [Fe/H]
$< -$2.0 could be due to the increasing importance of the outer-halo
component, which has a peak of metallicity at [Fe/H] $\sim$ $-$2.2, and a
long tail extending to lower metallicity.
We have evaluated the as-observed fractions of carbon-rich stars in our
Extended Sample, for Z$_{\rm max}$ $>$ 5 kpc, over several bins in
metallicity. With the adopted definition of carbon-rich stars ([C/Fe] $>$
+0.7), and following Eqn. 2, we find that 2\%, 7\%, and 20\% of stars in
the intervals $-1.5 <$ [Fe/H] $< -0.5$, $-2.5 <$ [Fe/H] $< -1.5$, and
[Fe/H] $< -$2.5, are carbon-rich, respectively. For ease of comparison with
previous determinations of the CEMP fractions in the halo, we note that the
global fraction of CEMP stars in the halo system with $Z_{\rm max}$\ $>$ 5 kpc and
[Fe/H] $< -1.5$ is 8\%, for [Fe/H] $< -2.0$ it is 12\%, and for [Fe/H] $<
-2.5$ it is 20\%.
The right panel of Figure 11 shows the mean carbonicity,
$\langle$[C/Fe]$\rangle$, as a function of [Fe/H]. Obviously, it is a
strong function of metallicity, although there may be some sign of it
leveling off at the lowest metallicities. Larger samples, in particular of
lower metallicity CEMP stars, are required to be certain.
Figure 12 shows the fraction of CEMP stars in the Extended Sample, as a
function of distance from the Galactic plane, $|$Z$|$, for stars satisfying
$-$2.0 $<$ [Fe/H] $< -$1.5 and [Fe/H] $< -$2.0, respectively. The intervals
for the cuts on height above or below the plane have widths of
$\Delta$$|$Z$|$ = 4 kpc, and the CEMP star fractions in each bin are
obtained by applying the same criteria and definitions as used for the left
panel of Figure 11. In the range of metallicity $-$2.0 $<$ [Fe/H] $< -$1.5,
significant contamination of the sample from thick-disk stars is not
expected, while the MWTD could be still present (with metallicity peak
around [Fe/H] $\sim -$1.3), but only in the region close to the Galactic
plane (0 kpc $<$ $|$Z$|$ < 4 kpc); the MWTD would not be expected to
contribute for the metallicity range [Fe/H] $< -$2.0 even close to the
plane. Inspection of this figure indicates a clear dependence of the CEMP
star fraction on distance from the Galactic plane. Close to the plane, this
dependence is due to the combined presence of the possible MWTD and
inner-halo populations, while at $|$Z$|$ $>$ 4 kpc, the observed fractions
must essentially pertain to the halo system alone. Far from the Galactic
plane, the observed increase of the CEMP star fractions with $|$Z$|$ would
be difficult to understand if the halo system comprises a single
population. In such a case, one might expect the CEMP star fractions to be
roughly constant, as a function of $|$Z$|$, for any given cut in
metallicity. This is clearly {\it not} what the data show.
Interestingly, we note that, for the same intervals in $|$Z$|$, the mean
carbonicity remains approximately contstant, at a value
$\langle$[C/Fe]$\rangle$ $\sim +1.0$ for $-$2.0 $<$ [Fe/H] $< -$1.5 and
$\langle$[C/Fe]$\rangle$ $\sim +1.5$ for [Fe/H] $< -2.0$. Of course, the
contribution from the inner- and outer-halo components is shifting as one
progresses from low to high $|$Z$|$, and this may be smoothing out any real
variations associated with the individual components. We return to this
question below.
A similar trend of increasing CEMP star fraction with height above the
Galactic plane was previously suggested by Frebel et al. (2006), based on a
much smaller sample of stars from the Hamburg/ESO survey (Wisotzki et al.
2000; Christlieb 2003), and proportionately larger error bars. Our expanded
data set now clearly indicates the existence of a strong spatial variation
of the CEMP star frequency {\it within the halo system}, and suggests that
the observed carbon enhancement is unlikely to be purely driven by
metallicity alone. Rather, it points to a real difference in CEMP star
fractions associated with the inner- and outer-halo populations, and opens
the possibility for different carbon-production mechanisms and/or different
chemical-evolution histories within their progenitors. We return to this
question below, after considering a method to probabilistically classify
individual stars as likely inner- or outer-halo members.
\subsection{The Extreme Deconvolution Technique}
Inference of the distribution function of an observable given only a
finite, noisy set of measurements of that distribution is a problem of
significant interest in many areas of science, and in astronomy in
particular. The observed distribution of a parameter is just the starting
point, but what is desired is knowledge of the distribution that we would
have in the case of very small uncertainties of the data and with all of
the dimensions of the parameter measured; in other words, the closest
representation of the underlying distribution. Usually, the data never have
these two properties, and it is then challenging to find the underlying
distribution without taking into account the uncertainty of the data (Bovy,
Hogg, \& Roweis, 2011; hereafter BHR11). The Extreme Deconvolution (XD)
technique of BHR11 confronts all of these issues, and provides an accurate
description of the underlying distribution of a
\emph{d}-dimensional quantity by taking into account the potentially
large and heterogeneous observational uncertainties, as well as missing
dimensions.
The BHR11 paper generalized the well known mixtures-of-Gaussians
density-estimation method to the case of noisy, heterogeneous, and
incomplete data. In this method, the underlying distribution of a
quantity {\bf v} is modeled as a sum of $K$ Gaussian distributions
\begin{eqnarray}
p(\textbf{v}) = \sum_{j=1}^K \alpha_{j}N(\textbf{v|m$_{j}$},\textbf{V$_{j}$})\,,
\end{eqnarray}
where the function N(\textbf{v|m$_{j}$},\textbf{V$_{j}$}) is the
\emph{d}-dimensional Gaussian distribution with mean \textbf{m} and
variance tensor \textbf{V} and $\alpha_{j}$ are the amplitudes,
normalized to sum to unity (all of these parameters are grouped
together as $\theta$ in what follows). The data {\bf w$_i$} are
assumed to be noisy samples from this distribution
\begin{eqnarray}
{\bf w_{i} = v_{i}} + \mathrm{noise}\,,
\end{eqnarray}
where the noise is drawn from a normal distribution with zero mean and
known covariance matrix {\bf S$_{i}$}. Here and in what follows, we
ignore the projection matrices {\bf R$_i$} of BHR11, since the data we
will apply this technique to are complete.
The likelihood of the model for each data point is given by the model
convolved with the uncertainty distribution of that data point. Since
a Gaussian distribution convolved with another Gaussian distribution
is again a Gaussian, the likelihood for each data point is a sum of
Gaussian distributions
\begin{eqnarray}
p(\textbf{w$_{i}$}|\theta) = \sum_{j=1}^K \alpha_{j}N(\textbf{w$_{i}$}|\textbf{m$_{j}$},\textbf{T$_{ij}$})\,
\end{eqnarray}
where
\begin{eqnarray}
\textbf{T$_{ij}$} = \textbf{V$_{j}$} + \textbf{S$_{i}$}\,.
\end{eqnarray}
The objective function is the total likelihood, obtained by simply
multiplying the individual likelihoods together for the various data
points
\begin{eqnarray}
\ln \mathcal{L} = \sum_{i}\ln p(\textbf{w$_{i}$}|\theta) = \sum_{i}\ln\sum_{j=1}^K \alpha_{j}N(\textbf{w$_{i}$}|\textbf{m$_{j}$},\textbf{T$_{ij}$}).
\end{eqnarray}
The optimization of this objective function provides the maximum likelihood
estimate of the distribution, or its parameters. In BHR11, the authors developed
a fast and robust algorithm to optimize the likelihood, based on an adaptation
of the expectation-maximization algorithm\footnote{Code implementing this
algorithm is available at \url{http://code.google.com/p/extreme-deconvolution/}\,.}
(Dempster et al. 1977).
The XD technique provides the best-fit values of the amplitude, mean,
and standard deviation of each Gaussian component, as well as the
so-called {\it posterior probability} that the observed data point
\textbf{w$_{i}$} is drawn from the component \emph{j}. The posterior
probability is given by
\begin{eqnarray}
p_{ij} = \frac{\alpha_{j}N(\textbf{w$_{i}$}|\textbf{m$_{j}$},\textbf{T$_{ij}$})}{\sum_{k}\alpha_{k}N(\textbf{w$_{i}$}|\textbf{m$_{k}$},\textbf{T$_{ik}$})}\,,
\end{eqnarray}
(see BHR11 for the derivation of this formula). The posterior
probability is a powerful statistical tool to perform probabilistic
assignments of stars to a Gaussian component in the model
distribution. In Galactic studies, the Gaussian component could
represent a primary structural component such as a disk or halo, a moving group,
or a spatial or velocity overdensity.
In many respects the XD technique described above is similar to the maximum
likelihood technique adopted in C10, but is more general, because it takes
into account the uncertainties of the measurements and provides the
membership probabilities.
\subsection{Application to the SDSS/SEGUE DR7 Calibration Stars}
The basic parameters for application of the XD approach for the analysis at
hand are the Galactocentric rotational velocity, V$_{\phi}$, the maximum
vertical distance, Z$_{\rm max}$, and the metallicity, [Fe/H]. We employ
the XD technique to determine the underlying distribution of the rotational
velocity of halo stars and to determine the posterior membership
probabilities for each star. In practice, the entries for the XD algorithm
are the Galactocentric rotational velocity and its uncertainty for each
star, V$_{\phi,i}$ and $\varepsilon_{V_{\phi},i}$, respectively\footnote
{Note that errors in the rotational velocity depend in turn on errors in
the distances and proper motions; these are carried forward into the
analysis automatically.}. These two parameters, together with all of the other
kinematic and orbital quantities, are derived using the same procedures
employed by C07 and C10 (but with revised distances for the reassigned
main-sequence turnoff stars, as discussed above).
Our plan is to make use of a sample in which the constraint on the distance
from the Sun, $d$, and on the projected Galactic distance, $R$, are relaxed
(in order to take advantage of the larger numbers of CEMP stars in the
Extended Sample). As a consequence, the data set becomes noisier then the
Local Sample, because uncertainties on the transverse velocities increase
with the distance $d$. In the terminology of the XD approach, the analysis
of the Extended Sample falls into the case where all the observables are
known, but some of the derived parameters have large uncertainties.
The Galactic halo is assumed to be a two-component structure, comprising the inner
and the outer halo, as discussed in C07 and C10.
Thus, the general expression for the likelihood takes the form:
\begin{eqnarray}
\ln \mathcal{L} = \displaystyle\sum_{i} \ln [\alpha_{in}\cdot N^{i}_{in} + \alpha_{out}\cdot N^{i}_{out}]\,,
\end{eqnarray}
where $\alpha_{in}$ and $\alpha_{out}$ are the amplitude of the
inner halo and outer halo, respectively. The velocity distributions are
assumed to be Gaussian, thus
\begin{eqnarray}
N^i_{in/out} = N(V_{\phi,i} | V_{\phi,in/out},\sigma^2_{\phi,in/out}+\varepsilon^2_{V_{\phi},i})\,,
\end{eqnarray}
The membership probabilities then take the form:
\begin{eqnarray}
p_{i, in} = \frac{\alpha_{in}N_{in}^{i}}{\alpha_{in}N_{in}^{i} + \alpha_{out}N_{out}^{i}}
\end{eqnarray}
and
\begin{eqnarray}
p_{i, out} = \frac{\alpha_{out}N_{out}^{i}}{\alpha_{in}N_{in}^{i} + \alpha_{out}N_{out}^{i}}
\end{eqnarray}
for the \emph{i}th star.\\
\subsubsection{Extreme Deconvolution Results for the Extended Sample}
In this section the posterior probabilities, derived through the
application of the XD, are obtained for the Extended Sample ($d <$ 10 kpc,
no constraints on $R$), selected in the range of metallicity [Fe/H] $<
-2.0$, and with vertical distance Z$_{\rm max}$ $>$ 5 kpc. The results are
shown in Figure 13. The upper panel of this figure shows the distribution
of the observed rotation velocity V$_{\phi}$ (black histogram), and the
green and red curves represent the resulting Gaussian velocity
distributions for the inner- and outer-halo components. The middle panel
shows the derived velocity distribution for the two components, obtained by
weighting each star with its membership probability, and normalized such
that the total area corresponds to unity. The black histogram is the
observed distribution as in the top panel, but normalized to unity as well.
In the lower panel, the posterior probability as a function of the
rotational velocity, V$_{\phi}$, is shown. This probability has been
obtained using Eqns. 11 and 12. For each star in the sample, the XD
technique assigns two probability values, the first associated with the
inner halo, and the second related to the outer halo. Since in our model a
star is required to belong to one or the other component (there are no
orphans allowed), and the summed inner- and outer-halo posterior
probabilities are forced to unity at each point, the two curves simply
complement one another.
The values obtained for the mean rotational velocity and its dispersion
are, for the inner halo, V$_{\phi}$ = 56 $\pm$ 11 km s$^{-1}$, and
$\sigma_{V_{\phi}}$ = 93 $\pm$ 35 km s$^{-1}$; for the outer halo,
V$_{\phi}$ = $-$141 $\pm$ 31 km s$^{-1}$, and $\sigma_{V_{\phi}}$ = 138
$\pm$ 58 km s$^{-1}$. The errors on the velocity and its dispersion have
been evaluated by employing the jackknife method. A K-S test of the null
hypothesis that the velocity distributions of stars belonging to the inner-
and outer-halo components could be drawn from the same parent population is
rejected at a high level of statistical significance (\emph{p} $< 0.001$).
\subsection{Contrast of CEMP Stellar Fractions in the Two Halo Components}
We now consider the distribution of carbonicity ([C/Fe]) for the Extended
Sample with metallicity [Fe/H] $< -$2.0 and at Z$_{\rm max}$ $>$ 5 kpc.
These selections ensure the presence of essentially all halo stars in the
sample, with little or no contamination expected from the thick disk or
MWTD. The top panel of Figure 14 shows the CarDF for the Extended
Sample\footnote{Note that the number of stars shown in Figure 14 differs
from the numbers shown in Figure 13, as Figure 14 includes only stars with
detected CH G-band features.}. There are two clear peaks that emerge, one
at [C/Fe] $\sim +0.3$ to $+0.5$, and the second softer peak around [C/Fe]
$\sim +1.8$, along with a tail extending towards higher values of
carbonicity. The bottom panel of Figure 14 shows the weighted CarDFs for
the inner- and outer-halo components. The underlying distribution of each
population has been obtained by weighting the values of the CarDF with the
membership probabilities of each star. As usual, the green distribution
denotes the inner halo, while the red distribution represents the outer
halo. A K-S test of the null hypothesis that the CarDF of stars belonging
to the inner- and outer-halo components could be drawn from the same parent
population is rejected at high statistical significance (\emph{p} $<$
0.001), clearly indicating that the carbonicity distribution functions of
the inner- and outer-halo components are distinct.
The contrast of the CEMP star fractions in the inner- and outer-halo
components can be investigated by applying the XD analysis to subsamples of
stars in restricted ranges of metallicity. The stars employed are those
belonging to the Extended Sample, selected in the low metallicity range
([Fe/H] $<$ $-$1.5) and with Z$_{\rm max}$ $>$ 5 kpc.
A first attempt was made by selecting stars in intervals of metallicity
such that $\Delta$[Fe/H] = 0.2 dex, and applying the XD technique to the
rotational velocity distribution of each subsample. This experiment was not
successful, due to the low number of stars in each bin. A much better
result has been obtained by choosing a larger interval on metallicity,
$\Delta$[Fe/H] = 0.5 dex.
The left panel of Figure 15 shows the previously determined global trend of
the CEMP star fraction, as a function of metallicity (dashed curve),
overplotted with the derived inner- and outer-halo CEMP star fractions. The
blue filled stars represent the predicted values of the CEMP star fractions
for each bin of metallicity, i.e., $-$2.0 $<$ [Fe/H] $<-$ 1.5, $-$2.5 $<$
[Fe/H] $<-$ 2.0, and $-$3.0 $<$ [Fe/H] $<-$ 2.5 (hereafter bin1, bin2, and
bin3). The predicted values are derived using the global trend of CEMP star
fractions vs. [Fe/H] (second-order polynomial; Fig. 11), and the posterior
probability for each star in the subsamples associated with each bin of
metallicity, such that:
\begin{equation}
F^{k}_{CEMP_{pred}} = \frac{1}{\sum_{i}(p_{ij})}\cdot\sum_{i}(p_{ij}f^{k})
\end{equation}
\noindent where p$_{ij}$ is the posterior probability, derived by applying the XD
analysis to the subsample of stars selected in bin1, bin2, and bin3,
respectively. The parameter $f^{k}$ represents the CEMP star fractions
obtained in each bin of metallicity from application of the second order
polynomial (\emph{j} and \emph{k} denote the halo component and the bin of
metallicity, respectively).
Note that the expression in Eqn. 13 provides an estimate of the CEMP
fraction that is driven by the global trend of F$_{CEMP}$ as a function of
the metallicity (Fig. 11, left panel). Eqn. 13 makes use of the full shape
of the posterior probability, and thus, a considerable overlap between the
kinematic distributions of the two components is still present. Therefore,
we are not expecting to find a significant difference in the predicted
values for the two components in each bin of [Fe/H].
The predicted value of CEMP star fractions in bin1 is F$_{CEMP_{pred,in,
1}}$ = (5.8 $\pm$ 0.03)\%, F$_{CEMP_{pred,out, 1}}$ = (5.9 $\pm$ 0.03)\%,
for the inner halo and outer halo, respectively. The same calculations for
bin2 provide F$_{CEMP_{pred,in, 2}}$ = (11.9 $\pm$ 0.1)\% and
F$_{CEMP_{pred,out, 2}}$ = (12.0 $\pm$ 0.1)\%. For the the third bin we
find F$_{CEMP_{pred,in, 3}}$ = (22.9 $\pm$ 0.6)\% and F$_{CEMP_{pred,out,
3}}$ = (22.7 $\pm$ 0.6)\%. The error on each fraction is evaluated by
employing the jackknife method. These fractions are sufficiently close to
one another that only a single set of blue stars is shown in the left panel
of Figure 15 to represent both results.
Not surprisingly, the above predicted values of the CEMP star fractions in
the inner and outer halo are very similar in all bins of metallicity. The
presence of a contrast in the CEMP star fraction for the two components can
be better investigated by reducing the overlap in the kinematic
distributions. This can be done by employing the
\emph{hard-cut-in-probability} method. For each halo component, the lower
limit of the membership probability is chosen to be $p_{lim}$ = 0.7. This
limit is set reasonably high to reduce the contamination between the halo
components. We have selected two subsamples of stars such that p$_{i,inner}
>$ $p_{lim}$, and p$_{i,outer} >$ $p_{lim}$ (\emph{i} denotes the star),
and for each subsample, the CEMP star fraction is evaluated following Eqn.
2.
The result of this exercise is shown in the left panel of Figure 15. Here,
the green (inner) and the red (outer) filled circles denote the values of
the CEMP star fractions obtained by employing the hard-cut-in-probability
method. The observed CEMP star fractions in bin1 are F$_{CEMP_{obs, in,
1}}$ = (5.2 $\pm$ 0.6)\% and F$_{CEMP_{obs,out, 1}}$ = (7.7 $\pm$ 1.0)\%,
for the inner and outer halo, respectively. In the case of bin2, we find
F$_{CEMP_{obs, in, 2}}$ = (9.4 $\pm$ 0.8)\% and F$_{CEMP_{obs,out, 2}}$ =
(20.3 $\pm$ 2.3)\%. Finally, in the third bin, the observed fractions are
F$_{CEMP_{obs,in, 3}}$ = (20.5 $\pm$ 2.4)\% and F$_{CEMP_{obs, out, 3}}$ =
(30.4 $\pm$ 4.8)\%.
In the highest metallicity bin, $-2.0 <$ [Fe/H] $< -1.5$, the CEMP star
fractions for the two halo components are close to the expected value
($\sim$ 6\%), with overlapping error bars. Note that the inner halo is the
dominant component in this range of metallicity ([Fe/H]$_{peak,in}$ =
$-$1.6), and we were not expecting to find a high contrast in the CEMP star
fractions. At lower metallicity, $-2.5 <$ [Fe/H] $< -2.0$, where the
dominant component is the outer halo ([Fe/H]$_{peak,out} = -2.2$), there is
evidence of a significant contrast in CEMP fractions between the inner- and
outer-halo components, such that F$_{CEMP_{obs, out, 2}}$ $\sim$
2$\times$F$_{CEMP_{obs, in, 2}}$. This contrast is confirmed also in the
lowest bin of metallicity, $-3.0 <$ [Fe/H] $< -2.5$, even though it is less
remarkable, F$_{CEMP_{obs, out, 3}}$ $\sim$ 1.5$\times$F$_{CEMP_{obs, in,
3}}$, and has a much larger error bar, due to the smaller numbers of stars
involved.
It is important to note that most of the stars in the lowest metallicity
bin belong to the category for which only an upper limit of [C/Fe] has been
provided (classified as subsample L, see Section 2.4). A significant number
of stars in the lowest metallicity bin have high $T_{\rm eff}$\ ($\sim$ 65\% above
6250~K), and thus only quite high values of [C/Fe] are expected to be
detected. Thus, it is our expectation that the quoted frequencies for CEMP
star fractions, at least at low metallicity, represent lower limits.
Morever, the SDSS/SEGUE DR7 calibration-star sample has a limited number of
stars with [Fe/H] $< -2.5$. The extremely metal-poor regime will be further
investigated in a future paper using the much larger sample of such stars
in the SDSS/SEGUE DR8 release (Aihara et al. 2011).
Interestingly, the right panel of Figure 15 shows that the mean
carbonicity, $\langle$[C/Fe]$\rangle$, remains similar for both the inner-
and outer-halo components as a function of declining metallicity. There is
a hint of a split emerging for the lowest metallicity bin, $-3.0 <$ [Fe/H]
$< -2.5$, with the outer-halo $\langle$[C/Fe]$\rangle$ being slightly
higher than that of the inner-halo value. However, the error bars overlap,
indicating that the sample size is simply too small to be certain.
\section{Summary and Discussion}
We have analyzed the SDSS/SEGUE DR7 calibration stars in order to explore
the global properties of the carbonicity distribution function ([C/Fe];
CarDF) of the halo components of the Milky Way, and the possible contrast
of the CEMP star fractions and mean carbonicity between the inner halo and
the outer halo. Carbon-to-iron abundance ratios (or limits) have been
obtained for 31187 stars, based on matches to a grid of synthetic spectra
to the CH G-band at 4305 {\AA}, with an external error for [C/Fe] on the
order of 0.25 dex. Kinematic and orbital parameters were derived employing
the methodologies described in C10. We have considered the nature of the
samples of stars considered CEMP ([C/Fe] $> +0.7$) and non-CEMP ([C/Fe] $<$
+0.7), and shown that, at low metallicity ([Fe/H] $< -1.5$), their
distributions on distance and rotational velocity differ significantly from
one another. For the stars with a detectable CH G-band, a deconvolution of
the inner- and outer-halo components, along with a derivation of the
membership probabilities, has been obtained by applying the Extreme
Deconvolution (XD) analysis to the Extended Sample of calibration stars,
selected in the low metallicity range ([Fe/H] $< -$1.5) and with Z$_{max}
>$ 5 kpc. Contrasts of the CEMP star fractions and mean carbonicity have
been obtained by employing the hard-cut-in-probability method to subsamples
of stars selected in several bins of metallicity.
\subsection{Summary of Main Results}
Our main results are the following:
\begin{itemize}
\item The as-observed distribution of [C/Fe], which we refer to as the
carbonicity distribution function (CarDF), has been derived at various
intervals on distance above or below the Galactic plane. At $|$Z$|$ $>$ 5
kpc, where we expect to see the beginning of the transition from inner-halo
dominance to outer-halo dominance among halo stars, the CarDF exhibits a
strong peak at [C/Fe] $\sim$ +0.3, and a long tail towards high values of
carbonicity, up to [C/Fe] = +3.0. The fraction of CEMP stars (defined here
to mean [C/Fe] $>$ $+$0.7) for the subsample at $|$Z$|$ $>$ 9 kpc (where
the dominant component is the outer halo) is 20\%, in agreement with
several previous estimates for stars at similarly very low metallicities,
[Fe/H] $<-$2.0.
\item At low metallicities ([Fe/H] $< -1.5$), the distribution of derived
distances and rotational velocities for the CEMP and non-CEMP stars differ
significantly from one another.
\item \emph{Almost all} of the CEMP stars are located in the halo components
of the Milky Way.
\item We observe a significant increase of the fraction of CEMP
stars with declining metallicity in the halo system. Following Eqn. 2, we
find that 2\%, 7\%, and 20\% of stars in the metallicity intervals $-1.5 <$
[Fe/H] $< -0.5$, $-2.5 <$ [Fe/H] $< -1.5$, and [Fe/H] $< -$2.5, are
carbon-rich, respectively. For ease of comparison with previous estimates,
the global frequency of CEMP stars in the halo system for
[Fe/H]\ $< -1.5$ is 8\%; for [Fe/H]\ $< -2.0$ it is 12\%; for [Fe/H]\ $<-2.5$ it is 20\%.
For reasons described above, we believe these fractions should be
considered lower limits.
\item In the low-metallicity regime ([Fe/H] $< -$1.5), and at vertical
distances Z$_{max} >$ 5 kpc, a continued increase of the fraction of CEMP
stars with declining metallicity is found (Fig. 11, left panel). A
second-order polynomial provides a good fit to the CEMP star fraction as a
function of [Fe/H].
\item In the same metallicity regime the mean carbonicity,
$\langle$[C/Fe]$\rangle$, increases with declining metallicity (Fig. 11,
right panel).
\item In the low-metallicity regime, for both $-2.0 <$ [Fe/H] $< -$1.5
and [Fe/H] $< -$2.0, we find a clear increase in the CEMP star fraction
with distance from the Galactic plane, $|$Z$|$ (Fig. 12). At these low
metallicities, significant contamination from the thick disk and MWTD
populations is unlikely, and would only have a small effect for
the bin with 0 kpc $<$ $|$Z$|$ $<$ 4 kpc and $-2.0 <$ [Fe/H] $< -$1.5; this
is an observed property of the halo system. The mean carbonicity remains
roughly constant as a function of $|$Z$|$, taking a value of
$\langle$[C/Fe]$\rangle$ $\sim +1.0$ for $-2.0 <$ [Fe/H] $< -$1.5, and
$\langle$[C/Fe]$\rangle$ $\sim +1.5$ for [Fe/H]\ $< -2.0$.
\item The Extreme Deconvolution technique of BHR11 has
been applied to subsamples of stars selected in three bins of metallicity,
$-$2.0 $<$ [Fe/H] $<$ $-$1.5, $-$2.5 $<$ [Fe/H] $<$ $-$2.0, and $-$3.0 $<$
[Fe/H] $<$ $-$2.5, and with Z$_{\rm max}$ $>$ 5 kpc. We have successfully
decomposed the inner and outer halo in all three bins, and obtained the
inner- and outer-halo membership probabilities for each star.
\item At low metallicity, $-2.0 <$ [Fe/H] $< -1.5$, where the dominant component
is the inner halo ([Fe/H]$_{peak,in} = -1.6$), there is no significant
difference in the CEMP star fraction btween the inner and outer halo. In
contrast, at very low metallicity, $-2.5 <$ [Fe/H] $< -2.0$, where the
dominant component is the outer halo ([Fe/H]$_{peak,out} = -2.2$), there is
evidence for a significant difference in the CEMP star fraction between the
inner and the outer halo (Figure 15, left panel), such that F$_{CEMP_{obs,
out, 2}}$ $\sim$ 2$\times$F$_{CEMP_{obs, in, 2}}$. This difference is also
confirmed in the lowest metallicity bin, $-3.0 <$ [Fe/H] $< -2.5$, even
though it is less remarkable, F$_{CEMP_{obs, out, 3}}$ $\sim$
1.5$\times$F$_{CEMP_{obs, in, 3}}$, and has a larger error bar. We conclude
that the difference in CEMP frequency at very low metallicity is not driven
by metallicity itself, but rather, by the stellar populations present.
\item The mean carbonicity, $\langle$[C/Fe]$\rangle$, remains similar for both the
deconvolved inner- and outer-halo components as a function of declining
metallicity. Larger samples of low-metallicity CEMP stars are required to
see if a split in mean carbonicity emerges.
\end{itemize}
\subsection{Implications for Galaxy Formation}
A possible scenario for the formation of the inner- and the outer-halo
components has been described in C07. The fact that the outer-halo
component of the Milky Way exhibits a net retrograde rotation (and a
different distribution of overall orbital properties), clearly indicates
that the formation of the outer halo is distinct from that of the inner
halo and the disk components. In C07, it was suggested that the outer-halo
component formed, not through a dissipational, angular-momentum conserving
contraction, but rather through dissipationless chaotic merging of smaller
subsystems within a pre-existing dark matter halo. These subsystems would
be expected to be of much lower mass, and subject to tidal disruption in
the outer part of a dark matter halo, before they fall farther into the
inner part. As candidate (surviving) counterparts for such subsystems, one
might consider the low luminosity dwarf spheroidal galaxies surrounding the
Milky Way. In the case of the inner halo, C07 argued that the low-mass
sub-Galactic fragments, formed at an early stage, rapidly merge into several
more massive clumps, which themselves eventually dissipationally merge
(owing of the presence of gas that had yet to form stars). The essentially
radial merger of the few resulting massive clumps gives rise to the
dominance of high-eccentricity orbits for stars that we assign to
membership in the inner halo. Star formation within these massive clumps
(both pre- and post-merger) would drive the mean metallicity to higher
abundances. This would be followed by a stage of adiabatic compression
(flattening) of the inner-halo component owing to the growth of a massive
disk, along with the continued accretion of the gas onto the Galaxy. This
general picture is supported, at least qualitatively, by the most recent
numerical simulations of the formation and evolution of large disk galaxies
that include prescriptions to handle the presence of gas, and which also
include approximate star-formation schemes and follow the evolving
chemistry (e.g., Zolotov et al. 2010; Font et al. 2011; McCarthy et al.
2011; Tissera et al. 2011).
The large fractions of CEMP stars in both halo components indicates that
significant amounts of carbon were produced in the early stages of chemical
evolution in the universe. The observed contrast of CEMP star fractions
between the inner halo and outer halo strengthen the picture that the halo
components had different origins, and supports a scenario in which the
outer-halo component has been assembled by the accretion of small
subsystems, as discussed below. In this regard, it is interesting that the
MDF of the inner halo (peak metallicity at [Fe/H] $\sim -$ 1.6) is in the
metallicity regime associated with the CEMP-s stars, which are primarily
found with [Fe/H] $> -2.5$, while the MDF of the outer halo (metallicity
peak at [Fe/H] $\sim$ $-$2.2) might be associated with the metallicity
regime of the CEMP-no stars, which are primarily found with [Fe/H] $< -2.5$
(Aoki et al. 2007b).
The fact that, in the metal-poor regime, the outer halo exhibits a fraction
of CEMP stars that is larger than the inner halo (CEMP$_{outer}$ $\sim$
2$\cdot$CEMP$_{inner}$), suggests that multiple sources of carbon, besides
the nucleosynthesis of AGB stars in binary systems, were present in the
pristine environment of the outer-halo progenitors (lower mass sub-halos).
These sources could be the fast massive rotators and/or the faint
supernovae mentioned in the Introduction. Karlsson (2006) also explored the
possibility that primordial gas was pre-enriched in heavy metals by less
massive SNe (13 $<$ M/M$_{\odot}$ $<$ 30), whose ejecta underwent
substantial mixing and fallback, while the C (and N, O) originated from
massive rotating stars with M $\geq$ 40 M$_{\odot}$. If the CEMP stars in
the outer halo are predominantly CEMP-no stars (which has yet to be
established), it might suggest that non-AGB-related carbon production took
place in the primordial mini-halos. The predominance of CEMP-s stars in the
inner halo, if found, would suggest that the dominant source of carbon was
the nucleosynthesis by AGB stars in binary systems. This would place
important constraints on the primordial IMF of the sub-systems responsible
for the formation of the two halo components.
Recent efforts to model, from the population synthesis standpoint, the
fractions of observed CEMP stars in the halo have not managed to reproduce
results as high as 15-20\% for metallicities [Fe/H] $< -2.0$ (e.g., Izzard
et al. 2009; Pols et al. 2010). However, these predictions are
based solely on carbon production by AGB stars. While such calculations may
prove meaningful for the inner-halo population, they may not be telling the
full story for carbon production associated with the progenitors of the
outer-halo population. Indeed, the observed increase of the CEMP star
fraction with $|$Z$|$ (in particular, far from the Galactic plane) we have
found would be difficult to understand if the halo system represents a
single population. In any event, the lower CEMP fraction in the inner halo
may relieve some of the tension with current model predictions.
The fact that the CEMP star fraction exhibits a clear increase with $|$Z$|$
suggests that the relative numbers of CEMP stars in a stellar population is
not driven by metallicity alone. The proposed coupling of the cosmic
microwave background to the initial mass function (CMB-IMF hypothesis;
Larson 1998, 2005; Tumlinson 2007) is one mechanism for imposing a temporal
dependence on the IMF. This effect, coupled with chemical evolution models,
predicts that the CEMP fraction would be expected to increase as the
metallicity decreases, but with similar metallicity regimes forming carbon
according to the expected yields of the predominant mass range available at
that time. In the hierarchical context of galaxy formation, star-forming
regions are spatially segregated, and their chemical evolution can proceed
at different rates, with stars at the same metallicity forming at different
times. Depending on the source(s) of carbon, this trend could lead to a
spatial variation of the CEMP fraction at the same metallicity, increasing
in older populations and decreasing in younger ones.
The larger fraction of CEMP stars associated with the outer-halo component
could be due to the nature of the progenitor low-mass mini-halos (mass,
density, etc.) in which they formed. A natural place to look for surviving
examples are the ultra-faint dwarf spheroidal galaxies surrounding the
Milky Way discovered in the course of the SDSS (Willman et al. 2005;
Belokurov et al. 2006a,b; Zucker et al. 2006, and many others since). Early
hints of the possible association of CEMP stars with the ultra-faint
galaxies came from the recognition that a (serendipitously)
spectroscopically targeted star from SDSS in the direction of the Canes
Venatici ultra-faint dwarf, SDSS~J1327+3335, is a carbon-rich giant with a
radial velocity and inferred distance commensurate with this very
metal-poor satellite galaxy (Zucker et al. 2006). The existence of CEMP
stars in ultra-faint dwarf spheroidal galaxies has now been definitively
established (Norris et al. 2010a,b; Lai et al. 2011). Norris et al. 2010b
reported the discovery of an extremely carbon-rich red giant, Segue 1-7, in
the Segue 1 system. This star exhibits a metallicity [Fe/H] = $-$3.5,
carbonicity [C/Fe] $= +2.3$, and a low barium abundance ratio ([Ba/Fe] $<
-$1.0), which place it in the CEMP-no category. This discovery is
consistent with the idea that the CEMP-no stars may indeed be associated
with the ultra-faint dwarf spheroidal galaxies; further similar
investigations should prove of great interest.
\acknowledgments
D.C. gratefully acknowledges funding from RSAA ANU to pursue her research.
T.C.B. and Y.S.L. acknowledge partial funding of this work from grants PHY
02-16783 and PHY 08-22648: Physics Frontier Center/Joint Institute for Nuclear
Astrophysics (JINA), awarded by the U.S. National Science Foundation. J.B. was
partially supported by NASA (grant NNX08AJ48G) and the NSF (grant AST 09-08357).
Studies at ANU of the most metal-poor populations of the Milky Way are supported
by Australian Research Council grants DP0663562 and DP0984924, which are
gratefully acknowledged by J.~E.~N. and D.~C.
\\
{\it Facilities:} \facility{SDSS}.
|
\section{Introduction}
\label{sec:intro}
Stellar oscillations provide a powerful tool for studying the
interiors of the stars since the mode frequencies depend on the
properties of the star and give strong constraints on stellar models
and hence evolution theories. However, the observations of stellar
pulsations require extensive data sets in order to achieve accurate
frequencies and to avoid the side-lobes in the amplitude spectrum
caused by the daily cycle. Long time series are usually obtained
from the ground by means of multisite observations.
A good scenery to carry out seismic studies are short period
pulsating stars in open clusters. Since the cluster members have
been formed simultaneously in the past they share similar stellar
properties in the present as age, chemical composition and distance.
By means of isochrone fitting it is possible to fix the age and
stellar masses. These constraints imposed by the cluster are very
useful in computing seismic models [e.g. \citet{fox4,fox3,suarez}].
This has motivated, for instance, a number of observational studies
on the Pleiades $\delta$ Scuti stars. In particular,
six $\delta$ Scuti variables have been discovered
in the Pleiades cluster until now and most of them have
been intensively observed in recent years namely: V647 Tauri
\citep{liu}, HD 23628 \citep{li1}, V534 Tauri \citep{fox5, li2}, V624
Tauri \citep{fox1, fox2}, HD 23194 \citep{fox1, fox2} and V650 Tauri
\citep{kim}.
The target star V650 Tau (HD 23643, $V=7^{\rm m}.79$, A7) was
identified as a short-period pulsating variable by \citet{breger1}.
One-site CCD
photometric observations carried out by \citet{kim} in
November-December 1993, detected four frequencies.
With the aim at detecting more pulsation frequencies that may be helpful
in constructing new seismic model for the star \citep{fox6} we have organized
a multisite campaing on V650 Tauri.
\begin{figure*}[!ht]
\centering
\subfigure[]{\includegraphics[width=9cm]{curva.eps}}
\subfigure[]{\includegraphics[width=7cm]{espec_v650tau.eps}}
\caption{(a) Example of the differential light curve V650 Tauri - Comparison. The solid line represents a fit to the
observed data of
the nine frequency detected in V650 Tauri. (b) Amplitude spectrum derived from the light curve V650 Tau - Comparison. The amplitude is in mag and
the frequencies in c/d.}
\label{fig:curva}
\end{figure*}
\begin{table}[!t]\centering
\caption{Frequency peaks detected in a preliminary analisys of the light curve V650 Tauri - Comp. S/N
is the signal-to-noise ratio in amplitude after the prewhitening
process.} \label{tab:frec}
\begin{tabular}{lccr}
\hline
Freq.& A & $\varphi$/($2\pi$) & $S/N$ \\
(c/d)&(mag)&&\\
\hline
32.623471& 0.005075 &0.910606&34.3\\
35.565042 &0.002194 & 0.777990&17.6 \\
17.023812&0.002064 & 0.305423&8.9\\
18.460618&0.001216 & 0.659008&5.7\\
15.030028 &0.000832 &0.529173&4.2\\
32.705001 & 0.000761 & 0.605440&5.2\\
3.4312070 &0.000646 & 0.249472&4.0\\
24.177849 &0.000621 & 0.466309&4.6\\
32.862220 & 0.000672 &0.144418&4.6\\
4.1768560 & 0.000598& 0.383873&2.9\\
\hline
\end{tabular}
\end{table}
\section{Observations and data reduction}
\label{sec:obs}
Three site observations were obtained with a 0.50-m telescope of
Xing Long Station of National Astronomical Observatory of China, the
0.84-m telescope of
Observatorio San Pedro M\'artir in Mexico and the 0.80-m telescope
IAC80 of Teide Observatory in Spain. For all these observations CCD
cameras were used. The photometric data at Teide observatory and
Xing Long Station were acquired through a Johnson $V$ filter. In San
Pedro M\'artir a Str\"omgren $y$ filter was used instead. More than
16000 frames on 21 nights were obtained. All data were reduced using
standard IRAF routines. Aperture photometry was implemented to
extract the instrumental magnitudes of the stars. The differential
magnitudes were normalized by subtracting the mean of differential
magnitudes for each night. An example of the differential light
curves is shown in Figure~\ref{fig:curva}(a).
\section{Period analysis}
The period analysis has been performed by means of standard Fourier
analysis and least-squares fitting. In particular, the amplitude
spectra of the differential time series were obtained by means of
Period04 package \citep{lenz}, which considers Fourier as well as
multiple least-squares algorithms. This computer package allows to
fit all the frequencies simultaneously in the magnitude domain
The amplitude spectrum of the differential light curve V650 Tauri -
Comparison is shown in Figure~\ref{fig:curva}(b). As can be seen,
V650 Tauri presents high-amplitude peaks distributed between 12 c/d
and 38 c/d.
The frequencies have been extracted by means of standard
prewhitening method. In order to decide which of the detected peaks
in the amplitude spectrum can be regarded as intrinsic to the star
we follow Breger's criterion given by \citet{breger2}, where it
was shown that the signal-to-noise ratio (in amplitude) should be at
least 4 in order to ensure that the extracted frequency is
significant.
The frequencies, amplitudes and phases are listed in
Table~\ref{tab:frec}. Nine significant frequencies have been
detected in V650 Tauri in this preliminary analysis.
A detailed analysis of these observations will be given in a forthcoming
paper.
\medskip
We would like to thank the staff of the San Pedro M\'artir and Teide
observatories and Xing Long station for their assistence in securing
the observations. This work was partially supported by PAPIIT
IN114309.
|
\section{Inroduction}
The spiking activity of neurons within a local cortical population is typically correlated \cite{1,2,3,4}. As a result, local cortical signals are robust to noise, which is a prerequisite for reliable signal processing in cortex. Under special conditions, coherent activity in a local cortical population is an inevitable consequence of shared presynaptic input \cite{5,6,7,8,9}. Nevertheless, the mechanism for the emergence of correlation, synchronization or even nearly zero-lag synchronization (ZLS) among two or more cortical areas which do not share the same input is one of the main enigmas in neuroscience \cite{7,8,9}.
It has been argued that nonlocal synchronization is a marker of binding activities in different cortical areas into one perceptual entity \cite{8,10,11,12}. This prompted the hypothesis that synchronization may hold key information about higher and complex functionalities of the network. To investigate the synchronization of complex neural circuits we studied the activity modes of networks in which the properties of solitary neurons, population dynamics, delays, connectivity and background noise mimic the inter-columnar connectivity of the neocortex.
\begin{figure*}[t]
\begin{center}
\includegraphics[angle=0,width=0.945\textwidth,totalheight=0.315\textwidth]{fig1.jpg}
\end{center}
\caption{ ZLS and clusters in small neural circuits. (a) Schematic of ZLS of an oriented circuit consisting of four nodes, where node A is stimulated for 5 ms by an external current $I_{ext}=4\mu A/cm^2$. Detailed structure of each node is depicted for node B where filled-circles/empty-circles stand for excitatory/inhibitory neurons and arrows represent reciprocal connections between excitatory and inhibitory neurons. (b) Raster diagram of the firing activity of neurons in (a). (c) Schematic of an oriented circuit consisting of seven nodes which splits into 3-clusters represented by 3 colors. The structure of each node, the distribution of delays and background noise is similar to (a). (d) Raster diagram of the firing activity of neurons in (c). (e) Spike train for (a) with simplified node consisting of only one neuron . (f) Spike train for (c) as for (e). (g) Mixing argument (see text) for (a) where steps are measured in units of $\tau$ and synchronized nodes are composed of an identical set of colors. (h) Mixing argument for (c).
}
\end{figure*}
\section{Neuronal circuit}
We start with a description of the neuronal circuits and define the properties of a neuronal cell, the structure of a node in a network representing one cortical patch, and the connection between nodes. Each neural cell was simulated using the well known Hodgkin Huxley model \cite{13} (see Appendix for details). Each node in the network was comprised of a balanced population of 30 neurons, 80/20 percent of which were excitatory/inhibitory (Fig. 1a). The lawful reciprocal connections within each node were only between pairs of excitatory and inhibitory neurons and were selected at random with probability $p_{in}$. In terms of biological properties it was assumed that distant cortico-cortical connections are (almost) exclusively excitatory whereas local connections are both excitatory and inhibitory \cite{15,16}. In this framework, cortical areas are connected reciprocally across the two hemispheres and within a single hemisphere \cite{16c}, where small functional cortical units (patches) connect to other cortical patches in a pseudo random manner. The number of patches to which a single patch connects varies considerably, where typically, it grows like the square root of the number of cortical neurons \cite{16a} resulting for a mouse in 3 to 6 \cite{16b}, and most likely for humans roughly 150. Hence, we investigated excitatory strongly connected oriented graphs; i.e., if neurons belonging to node A project to neurons belonging to node B, then connections from node B to node A are forbidden; however there is a legal path between any pair of nodes \cite{14}. The connection between neurons belonging to different nodes was excitatory and was selected with probability $p_{out}$. In terms of biological properties, distant cortico-cortical unidirectional connections were exclusively excitatory whereas local connections within one node of the network were both excitatory and inhibitory \cite{15,16}.
The delay between a pair of neurons belonging to the same node was taken from a uniform distribution in the range [1.5,2.5] ms, whereas neurons belonging to different nodes came from a uniform distribution in the range $\lbrack \tau-0.5, \tau+0.5 \rbrack$ ms where $\tau$ was the average time delay including the internal dynamics of a neuron. Results are exemplified below in simulations with $p_{in}$=0.2, $p_{out}$=0.8 and $\tau=20$ ms, unless otherwise indicated. The robustness of the results was tested under the influence of background synaptic noise generated from the synaptic input of a balanced random population of 1000 neurons (see Appendix for details). In the absence of a stimulus an isolated node as well as the entire network has no consistent or periodic firing activity or chaotic activity \cite{ido}.
Figure 1a depicts a neuronal circuit consisting of four nodes and two loops having total delays of $3\tau$ and $4\tau$ with GCD(3,4)=1, where at time t=0 node A is stimulated for 5 ms by an external current $I_{ext}=4\mu A/cm^2$. The raster diagram of the firing activity of the neurons in each node is presented in Fig. 1b. Although the graph does not contain reciprocal connections after a short transient,$\sim200$ ms , the neuronal circuits reach ZLS among all nodes. Figure 1c depicts a neuronal circuit consisting of seven nodes and two loops with total delays of $6\tau$ and $3\tau$ with GCD(6,3)=3, where at time t=0, node B for instance is stimulated for 5 ms by an external current $I_{ext}=4\mu A/cm^2$. The raster diagram of the diluted firing activity of the neurons in each one of the seven nodes is presented in Fig. 1d. Results indicate that after a short transient of less than 200 ms, the neuronal circuit splits into 3-clusters as labeled by three different colors in Figs. 1c and 1d.
There are two sources for the low firing rate in the neuronal circuits \cite{16b,17a,17b}. The first source is the local inhibitory connections that hyperpolarized the membrane potential resulting in a relatively diluted firing pattern in each node. To illustrate this effect the Post-Stimulus Time Histogram (PSTH) of Fig. 1d is presented in Fig. 2, indicating that at any given time only about half of the population fires simultaneously. The second source for low firing is the emergence of m-clusters leading to the periodic activity with a period of m$\tau$ of each node rather than $\tau$ as in the case of ZLS. Hence, the cluster mechanism and low firing rate can appear in the brain activity concurrently. We are well aware that even with the irregularities described here, the statistics of the single neuron firing, as seen here, is far from resembling that of a single cortical neuron in the cortex of a behaving animal. The model described here is but a toy model compared to the full biological reality. With more noise, heterogeneous neural properties and more realistic neuronal model (e.g. adding dendrites), much higher irregularities may appear without disrupting the main results presented here: The behavior of a network of connected nodes is an emergent property of all the inter-nodal connections. Thus, the suitability to describe irregular firing patterns of cerebral cortical area \cite{17a,17b} needs a further investigation and might requires averaging out the regular firing by looking at long recording times.
\begin{figure}[t]
\begin{center}
\includegraphics[angle=0,width=0.32\textwidth,totalheight=0.32\textwidth]{fig2.jpg}
\end{center}
\caption{ PSTH of the spike trains corresponding to Fig. 1b. The emergence of clustering is demonstrated by the synchronized firing for nodes belonging to a cluster ((C,F), (A,D,G) and (B,E)). The synaptic noise in the circuit, as well as the inhibitory connections within each node, result in low firing dynamics, demonstrated by the fact that at any given time only about half of the population of each node (30 neurons) fires simultaneously.}
\end{figure}
\section{GCD-clusters}
In the absence of inhibition and background noise, the firing pattern of each neuron becomes regular and all neurons in a node are synchronized; however, the phase among nodes is unaltered. For the sake of clarity, below we present results for neural circuits where each node was reduced to one excitatory neuron with no background noise, although similar results were obtained in simulations for structured nodes and background noise.
The results of this simplified node characterized by a single neuron for the neural circuits in Figs. 1a and 1c are presented in Figs. 1e (ZLS) and 1f (3-clusters), respectively.
The interplay between the number of clusters and the GCD of the loops that compose a neuronal circuit can best be understood by the self-consistent argument that nodes with identical color are in ZLS and must be driven by the same set of "colors". The trivial solution is always one "color", ZLS; however, the alternative solution consists of exactly GCD "colors", GCD-clusters. An attempt to consistently color nodes serially with a greater number of colors fails, because nodes of the same color have different drives. In the case of $GCD > 1$, GCD-clusters take over the ZLS solution following the mixing argument \cite{18}; the initial condition is a distinct color for each node, time steps are rescaled with $\tau$ and at each time step a node is colored by the union of colors of its driven nodes. The colors of a node at step t indicate the set of nodes/colors at t=0 which are now mixed (integrated) by the node. The mixing argument for the circuit in Fig. 1a is shown in Fig. 1g where after 10 steps all nodes are identical, and colored by 4 colors, indicating a ZLS solution. Similarly the mixing argument for the circuit in Fig. 1c indicates 3-clusters ((A,D,G), (B,E), (C,F)) as depicted in Fig. 1h.
Note that it is possible to consistently color nodes with any common divisor; however, such dynamics requires complex stimuli to more than one node, identical color for nodes in the the initial condition.
\section{Nonlocal mechanism}
A more complex circuit is presented in Fig. 3a consisting of three directed loops with total delays of $6\tau$ , $12\tau$ and $18\tau$ and 25 nodes. The GCD(6,12,18)=6 and 6-clusters (6-colors) were identified in simulations. Small changes in topology can dramatically alter the number of clusters, such that the addition/deletion of one connection can serve as a remote switching mechanism in the circuit \cite{19,20}. Figures 3b, 3c and 2d show that an additional unidirectional connection between two nodes induces a loop with a total delay of $5\tau$ , $4\tau$ and $3\tau$ , respectively. Hence the GCD modifies and switches the 6-cluster solutions to ZLS, 2-clusters and 3-clusters, respectively. One might conclude that in order to generate large loops the number of cortical patches, nodes, needs to increase accordingly, and furthermore that a circuit composed of loops which are only multiples of a given integer (e.g. 6, 12, 18 $.~.~.$) might be far removed from cortical anatomy. Figure 3e indicates that an oriented graph consisting of 6 nodes with diluted connectivity generates loops of sizes $6n\tau$ where n is an integer. Furthermore a shortcut in this condensed representation, as shown by the dashed arrow in Fig. 3e, changes the GCD in a way similar to the expanded representation, Figs 3a-d. We note that a straight chain of neuronal groups with feed forward connections acts like a synfire chain and appropriate connectivity may be found abundantly in the cortex \cite{15}, however the behavior of circuits was not studied so far.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.6\textwidth,height=0.9\textwidth]{fig3.jpg}
\end{center}
\caption{ Non-local mechanism for clustering of complex and condensed circuits. (a) A circuit consisting of 25 simplified nodes, loops with total delays of $6\tau$ , $12\tau$ and $18\tau$ (boundaries of areas A, A+B and C, respectively where $\tau$ is a unit delay between two connected nodes), and a stimulus to one node as in Fig. 1. Nodes split into 6-clusters following the GCD(6,12,18)=6. (b) With an additional unidirectional connection (dashed arrow) and a loop of $5\tau$ , the GCD(5,6,12,18)=1 and the circuit is in ZLS. (c) An additional loop of $4\tau$ as in (b), where GCD(4,6,12,18)=2 and the circuit is in 2 clusters. (d) An additional loop of $3\tau$ as in (b), where GCD(3,6,12,18)=3 and the circuit is in 3 clusters. (e) Schematic condensed representation of 6n loops where n is an integer, emerging in a ring of 6 nodes, each one representing one cortical patch, and with diluted connectivity where the in/out connectivity of each neuron is at least 1. The gray dashed arrow depicts a similar effect as in 2b.}
\end{figure}
\section{Complex external stimuli}
A temporal stimulus to one node of the neural circuit splits the firing pattern of nodes into GCD-clusters and in addition the firing pattern cycle \cite{22} of a node is also equal to the GCD. For a neural circuit with GCD=6 (Figs. 3a and 3e), the firing pattern cycle of a node is exemplified in the first row of Fig. 4a together with its binary representation and the degeneracy of this class of stimuli. Under cyclic permutation symmetry in the 6-clusters arrangement, there are 13 classes of simultaneous stimuli for nodes belonging to one or more clusters as summarized by the first column of Fig. 4a, and as expected, the sum of their degeneracy is $2^6-1=63$. The firing pattern cycle and its binary representation are given by the second and third columns of Fig. 4a.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.60\textwidth,height=0.52275\textwidth]{fig4.jpg}
\end{center}
\caption{ Complex external stimuli and transients. (a) Under permutation symmetry, the 63 different stimuli in the 6-cluster arrangement of the circuit in Fig. 3a are organized into 13 classes (first column), where the degeneracy of each class is given in the last column. The firing pattern cycle of a node together with its binary representation are presented in the second and the third columns, indicating that the number of clusters as well as the firing spike cycle of a node can be any common divisor of the loops. (b) Four node circuit which is in ZLS and the transients for the 15 different stimuli organized in a tree. (c) A square circuit and its 15 different stimuli organized in 5 distinct cyclic flows.
}
\end{figure}
Results indicate that the period of the firing pattern can differ from the GCD (found also for chaotic networks \cite{ido}), for instance the stimulus period in the $4^{th}$ ( $8^{th}$) row of Fig. 4a is 3 (2) and hence the nodes only split into 3-clusters (2-clusters), instead of 6-clusters. In fact the number of clusters can be any common divisor of the loops composing the circuit as a result of stimuli inducing such periodicity, and an example is presented in Fig. 5.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.63\textwidth,height=0.549\textwidth]{fig5.jpg}
\end{center}
\caption{ Simulation results of two connected loops with total delays of $6\tau$ and $12\tau$ consisting of 17 nodes. (a) A drive to one node resulting in GCD(6,12)=6 clusters. (b) A drive to every fourth node in the loop of 12 , represented by $(1,0,0,0,1,0,0,0,1,0,0,0)$, results in GCD(4,6,12)=2 clusters. (c) The drive can be given to any set of nodes with the same colors as in (b) (blue, orange, green) and results in the same clusters. (d) A drive to four consecutive nodes in the loop of 12 represented by $(1,1,1,1,0,0,0,0,0,0,0,0)$, which is characterized by a periodicity of 12. Hence, it results in GCD(6,12)=6 clusters.
}
\end{figure}
The number of clusters as well as the spiking cycle of a given node can identify a class of possible stimuli applied to the circuits. Nevertheless, more detailed information about the stimuli can be deduced rapidly from the transients \cite{24} to synchronization. Fig. 4b shows all the transients to ZLS from the 15 possible stimuli for the circuit in Fig. 1a, whereas Fig. 4c presents the 5 possible cyclic flows for a unidirectional square circuit. It is clear that the length and even knowledge of a partial time ordering of firing nodes in the transient quite clearly identify the stimulus, and the synchronized mode mainly serves as an indicator of the end of the transient.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.6\textwidth,height=0.36\textwidth]{fig6.jpg}
\end{center}
\caption{ ZLS and clusters in heterogeneous circuits. (a) A chain of three neurons with an accumulated $2\tau$ ms delay is equivalent to a chain of two neurons with a $2\tau +4$ ms delay as a result of the 4 ms internal dynamic of the middle neuron. (b) Heterogeneous circuit at ZLS which is equivalent to a homogeneous circuit consisting of $4\tau$ and $5\tau$ connecting loops. (c) Heterogeneous circuit which is equivalent to a homogeneous circuit consisting of $4\tau$ and $6\tau$ connecting loops, 9 nodes and GCD(4,6)=2. The 6 nodes of the heterogeneous circuit are colored according the corresponding nodes in the homogeneous circuit.}
\end{figure}
\section{Heterogenous circuits}
Generalization of the above homogeneous circuit results to heterogeneous circuits is depicted in Fig. 6. Figure 5a indicates that the internal dynamic of a neuron results in an effective delay of about 4 ms. Hence, a chain of three neurons with a total delay of $2\tau$ is equivalent to a chain of two units with one delay of $2\tau +4$ ms and vice versa. Consequently, Fig. 6b presents a heterogeneous circuit which is equivalent to a homogeneous circuit composed of two loops of $4\tau$ and $5\tau$ . Since GCD(4,5)=1, the homogeneous as well as the heterogeneous circuits are in ZLS. Similarly, the delay of $2\tau +4$ ($3\tau +8$) in Fig. 6c can be replaced by homogeneous chains of 3 (4) nodes and the circuit is equivalent to two homogeneous loops of $4\tau$ and $6\tau$. Since GCD(4,6)=2, the original 6 nodes in the heterogeneous circuit, Fig. 6c, are colored according to the corresponding nodes in the homogeneous circuit.
The 4 ms internal dynamic of a neuron was estimated in simulations for large loops. However, for shorter loops, e.g., 6 units, the mapping between homogeneous/heterogeneous networks is found to valid in simulations when the internal dynamics is assumed to be in the range of $\lbrack 2,5 \rbrack$ ms. This range enables a robust synchronization mode of heterogeneous circuits with a wide range of delays, as well as heterogeneous circuits with complex nodes and fluctuating delays as in Fig. 1a.
\section{Concluding remarks}
The activity mode of the entire network cannot simply be described as a "Lego" of small connecting neural circuits with a given activity, since it is governed by a nonlocal quantity, the GCD. These findings challenge the emergence of significant topological motifs and the importance of their role in the functionality of the entire network \cite{26} as well as the impact of statistical properties of complex neural circuits \cite{27}. Rather, they call for a reexamination of sources of correlated activity in cortex where addition/deletion of a connection or more realistically synaptic alternations that can induce transition between decaying and sustained activities can serve as a remote switching mechanism, and indicate that learning induced changes in some connections affects the functionality of the networks more than others. The hypothesis that neural information processing might take place in the transient is suggestive of a much shorter time scale for the inference of a perceptual entity, which might also indicate the emergence of a probabilistic inference consecutive to the accumulation of knowledge during the transient. The notion that complex external stimuli result in several modes of activity also implies richer capabilities of the neural circuit and makes it imperative to analyze complex stimuli that are not triggered simultaneously. Nevertheless, the ways in which neural circuits detect synchronization still remain an enigma.
\section{Appendix}
To simulate neural population dynamics we used Hodgkin-Huxley (HH) type models for action potential generation, in which the membrane potential $V^i$ , of a single neuron $i$ , is described by the following differential equation
\begin{align}
c_m {dV^i \over dt}= & -g_{Na}(m^i)^3h^i(V^i-E_{Na}) \nonumber \\ & -g_k(n^i)^4(V^i-E_k) \nonumber \\ & -g_L(V^i-E_L)+I^i_{syn}+I^i_{ext}+I^i_{noise}
\end{align}
\noindent Here, $c_m=1\mu F/cm^2$ is the membrane capacitance. Maximal conductance and reversal potentials are given by:
\begin{table} [h!]
\begin{center}
\begin{tabular} {|c | c | c |}
\hline
$X $ & $ g_{X} [mS/cm^{2}]$ & $ E_{X} [mV] $ \\ \hline
$N_{a}$ & $120$ & $115$ \\ \hline
$k$ & $36$ & $-12$ \\ \hline
$L$ & $0.3$ & $10.5$ \\ \hline
\end{tabular}
\end{center}
\end{table}
The voltage-gated ion channels $m$, $n$ represent the activation of the sodium and potassium channels and the voltage-gated ion channel, $h$, represent inactivation of the sodium channels. Their dynamics is described by first order kinetic equations for a given neuron $i$:
\begin{equation}
{dY^i \over dt}=\alpha^i_Y(V^i)(1-Y^i)-\beta^i_Y(V^i)Y^i
\end{equation}
\noindent where Y may be substituted by $m$, $n$ and $h$. The experimentally fitted voltage-dependent transition rates are:
\begin{eqnarray}
&\alpha_{m}(V)= \frac{0.1 (V-25)}{1-e^{-0.1(V-25)}} \nonumber\\
&\beta_{m}(V)= 4 e^{-V/18} \nonumber\\
&\alpha_{n}(V)= \frac{0.1 (V-10)}{1-e^{-0.1(V-10)}} \nonumber\\
&\beta_{n}(V)= 0.125 e^{-V/80} \nonumber\\
&\alpha_{h}(V)= 0.07 e^{-V/20} \nonumber\\
&\beta_{h}(V)= \frac{1}{1+e^{-0.1(V-30)}} \nonumber
\end{eqnarray}
Synaptic background noise was simulated by a balance of input from 800 excitatory neurons firing at about 1-3Hz, and 200 inhibitory neurons firing at about 50-100 Hz. Our simulations were adjusted to imitate the behavior of a random biological neural cell. The free parameters were set in a way such that a single cell with no synaptic or external input fired randomly at about 5-7 Hz and the cell activity was noisy around the resting potential with a variance of about 5 mV.
The synaptic transmission between neurons was modeled by a postsynaptic conductance change in the form of an $\alpha$ function
\begin{equation}
\alpha (t) = \frac{e^{-t / \tau_{d}} - e^{-t / \tau_{r}}}{\tau_d - \tau_r}
\end{equation}
\noindent where the decay and rise time of the function are given by $\tau_d=10$ ms and $\tau_r=1$ ms respectively. The synaptic current $I^{i}_{syn}(t)$ takes the form
\begin{equation}
I^{i}_{syn}(t) = -g_{max} \sum_{j} \sum_{t_{j}^{sp}} \alpha (t- t_{j}^{sp} -\tau_{ij}) (V-E_{syn})
\end{equation}
\noindent Here, $\{j\}$ is the group of neurons coupled to neuron i. The internal sum is taken over the train of pre-synaptic spikes occurring at $t^{sp}_{j}$ of a neuron j in the group. The delays arising from the finite conduction velocity of axons are taken into account through the latency time $\tau_{ij}$ in the $\alpha$ function. Excitatory and inhibitory transmissions were differentiated by setting the synaptic reversal potential to be $E_{syn}=60$ mV or $E_{syn}=-20$ mV , respectively. $g_{max}$ describes the maximal synaptic conductance between neurons i and j. We integrated the set of differential equations numerically using Heun's method. The time step of the integration was 0.02ms.
\textbf{Population dynamics:}
Each node in the circuit was considered as one cortical area and comprised a balanced population of 30 excitatory and inhibitory neurons that were coupled sparsely with a fixed probability $p_{in}=0.8$ for every connection between two neurons. Every two neurons that belonged to adjacent nodes were connected with a fixed probability $p_{out}$ and connection strength $g_{max}=0.17 mS/cm^2$. The delays between groups of neurons were taken as a distribution around a mean value. Neurons belonging to the same population had an average delay time of 2 ms and neurons from different populations had delays of 20 ms.
|
\section{Introduction}
In recent years many fields of astrophysics have seen a transition towards increasingly large experiments and surveys. The level of complexity and the costs are rising alongside, requiring careful planning and assessment of the expected performance of the envisaged project at all stages. In forecasts of the statistical constraints on model parameters by future experiments the Fisher matrix \citep{fisher35,tegmark97} has proven to be indispensable \citep[e.g.][]{albrecht06}.
Its ubiquity can largely be attributed to the low computational cost of a Fisher matrix calculation compared to a full mock likelihood analysis, in particular if the data set to be analysed and the number of parameters to be inferred are large. This simplicity comes at the price of a twofold assumption of Gaussianity in the derivation of the Fisher matrix expressions \citep{tegmark97}. First, the data are assumed to be distributed according to a multivariate Gaussian. However, this assumption is shared with the majority of full likelihood analyses to date although precision measurements may require more complicated forms \citep[e.g.][]{bond00,hartlap09}. Second, since the Fisher matrix is defined as the expectation value of the Hessian of the log-likelihood in parameter space, it can only fully represent a Gaussian posterior whose logarithm has constant curvature.
Therefore the confidence levels on model parameters derived from a Fisher matrix are inevitably elliptical. They can describe the posterior distribution close to the point of maximum likelihood and indicate linear degeneracies among parameters via the ellipticity of confidence regions. Fisher matrix analyses fail to identify the shape of the posterior away from its maximum, as well as to detect non-linear dependencies of parameters. However, non-linear model parameter degeneracies are common, and the attempt to minimise or break them can drive the design of experiments. Hence it is desirable to go beyond the assumption of a Gaussian posterior in forecasts for the advanced stages of upcoming precision measurements.
In this work we propose to combine Fisher matrix forecasts with Box-Cox transformations of parameter space to obtain accurate expectations of posterior distributions. \citet{box64} introduced a parametrised set of power transformations with which a wide range of data can be transformed to follow a Gaussian distribution to good approximation. We will apply these transformations to model parameters, instead of data, in order to modify a given posterior into a multivariate Gaussian distribution for which a Fisher matrix analysis is exact. After an inverse Box-Cox transformation the Fisher matrix results will then accurately describe the original posterior. To determine the free parameters of the Box-Cox transformation, the original posterior needs to be sampled, and hence an initial mock likelihood analysis to be run.
We will demonstrate this method with an example from cosmology. Several ambitious surveys are currently planned or designed\footnote{These include e.g. the Large Synoptic Survey Telescope (\texttt{http://www.lsst.org}), the NASA satellite WFIRST (\texttt{http://wfirst.gsfc.nasa.gov}), and the ESA satellite Euclid (\texttt{http://sci.esa.int/euclid}).} that are going to measure the parameters of the cosmological standard model, particularly those of dark matter and dark energy, with high precision. These experiments will investigate several cosmological probes of the large-scale structure of the Universe, the potentially most powerful one being weak gravitational lensing of distant galaxies \citep{albrecht06,peacock06}. Weak lensing features a characteristic non-linear degeneracy between the two best-constrained parameters $\Omega_{\rm m}$ (mean matter density) and $\sigma_8$ (normalisation of matter density fluctuations) as they both govern the overall amplitude of the signal \citep[see e.g.][]{hoekstra06,schrabback09}. Hence a mock weak lensing survey provides an excellent test case, but we emphasise that the method outlined is applicable to any prediction for model parameter constraints.
The paper is organised as follows. Section \ref{sec:theory} details the principles of Box-Cox transformations, our strategies to determine optimal Box-Cox parameters, and the combined Box-Cox-Fisher formalism. In Section \ref{sec:performance} we investigate the performance of the proposed method for a mock weak lensing experiment, comparing different variants in the implementation and quantifying the universality of the Box-Cox-Fisher formalism. We apply this formalism to a test of the degeneracy-breaking capabilities of weak lensing higher-order statistics in Section \ref{sec:application}, before we summarise and conclude on our findings in Section \ref{sec:conclusions}.
\section{Box-Cox transformations of parameter space}
\label{sec:theory}
Power transformations such as the inverse and square-root transformation, or logarithmic transformations are popular choices to render the distribution of data more Gaussian. The Box-Cox transformation unites these cases with a single free parameter per dimension and are hence widely used in various areas of science. Astrophysical applications are rare; one example is the work by \citet{dineen05} who tested cosmic microwave background data for Gaussianity.
For a $N_p$-dimensional variable $\vek{p}$ the Box-Cox transformation in each dimension $\mu=1,\,..\,,N_p$ reads \citep{box64}
\eq{
\label{eq:bctrafo}
\bar{p}_\mu(\lambda_\mu,a_\mu) = \left\{ \begin{array}{ll} \bb{\br{p_\mu + a_\mu}^{\lambda_\mu}-1}/\lambda_\mu & \lambda_\mu \neq 0\\ \ln (p_\mu + a_\mu) & \lambda_\mu = 0 \end{array} \right. \;,
}
where the normalisation has been chosen such that the transformation is continuous in the parameter $\lambda_\mu$ at $\lambda_\mu=0$. We allow for a shift $a_\mu$ as a second free parameter in each dimension. Note that we denote transformed quantities by a bar and drop the dependence on the Box-Cox parameters $\br{\vek{\lambda},\vek{a}}$ unless it needs to be made explicit.
Usually, equation (\ref{eq:bctrafo}) is applied to the elements of a datavector, but we will henceforth understand $p_\mu$ as the parameters of an $N_p$-dimensional parameter space. Then the transform of a given posterior distribution ${\cal P}(\vek{p})$ is given by
\eq{
\label{eq:bcdistribution}
\bar{\cal P}(\bar{\vek{p}}) = {\cal P}(\vek{p})\; J(\vek{p},\bar{\vek{p}})\;,
}
with the Jacobian
\eq{
\label{eq:jacobian}
J(\vek{p},\bar{\vek{p}}) = \prod_{\mu=1}^{N_p} \left| \frac{\dd p_\mu}{\dd \bar{p}_\mu} \right| = \prod_{\mu=1}^{N_p} \br{p_\mu + a_\mu}^{1-\lambda_\mu}\;.
}
The second equality follows directly from equation (\ref{eq:bctrafo}). The first goal is to determine the set of $2 N_p$ parameters, $\br{\vek{\lambda},\vek{a}}$, such that the transformed posterior, $\bar{\cal P}(\bar{\vek{p}})$, is a multivariate Gaussian to good approximation.
\subsection{Optimal Box-Cox parameters}
\label{sec:optimalbc}
Suppose a random sample $\vek{\hat{p}}$ with $n$ elements, i.e. $\bc{\hat{p}_{\mu,1},\,..\,,\hat{p}_{\mu,n}}$ for every $\mu=1,\,..\,,N_p$, is drawn from the posterior ${\cal P}(\vek{p})$, for instance via Monte-Carlo sampling techniques. If the Box-Cox transformed posterior is indeed Gaussian, the distribution is given by
\eqa{
\label{eq:transformedposterior}
{\cal P}(\vek{p}) &=& \bar{\cal P}(\bar{\vek{p}})\; J(\bar{\vek{p}},\vek{p})\\ \nn
&=& \prod_{\mu=1}^{N_p} \br{p_\mu + a_\mu}^{\lambda_\mu-1}\; \frac{1}{\sqrt{(2\pi)^{N_p}\det {\rm Cov}(\bar{\vek{p}})}}\\ \nn
&& \times\; \exp \bc{- \frac{1}{2}\; (\bar{\vek{p}}-\bar{\vek{p}}_{\rm max})^\tau\; {\rm Cov}^{-1}(\bar{\vek{p}})\; (\bar{\vek{p}}-\bar{\vek{p}}_{\rm max})}
}
and has only the Box-Cox parameters, and the mean $\bar{\vek{p}}_{\rm max}$ and covariance
\eq{
\label{eq:defcov}
{\rm Cov}(\bar{\vek{p}}) \equiv \ba{(\bar{\vek{p}}-\bar{\vek{p}}_{\rm max})\;(\bar{\vek{p}}-\bar{\vek{p}}_{\rm max})^\tau}\;
}
of the Gaussian as free parameters. Since $\bar{\cal P}(\bar{\vek{p}})$ is assumed Gaussian, one can employ the standard maximum likelihood estimators for the covariance and mean. The latter simply implies $\bar{\vek{p}}=\bar{\vek{p}}_{\rm max}$, so that the exponential in (\ref{eq:transformedposterior}) is unity. Consequently one obtains the following concentrated log-likelihood for the Box-Cox parameters \citep[for details see][]{box64,velilla93},
\eqa{
\label{eq:bclikelihood}
{\cal L}_{\rm max}(\vek{\lambda},\vek{a}) &=& - \frac{n}{2}\; \ln \det {\rm Cov}\bb{\bar{\vek{\hat{p}}}(\vek{\lambda},\vek{a})}_{\rm ML}\\ \nn
&& + \sum_{\mu=1}^{N_p} \bc{(\lambda_\mu -1) \sum_{i=1}^n \ln (\hat{p}_{\mu,i}+a_\mu)}\;,
}
up to an irrelevant constant. We have added the subscript ML to emphasise that the maximum likelihood estimate for the covariance based on $\vek{\hat{p}}$ is to be used. Maximising this likelihood for a given sample $\vek{\hat{p}}$ should then return Box-Cox parameters $\br{\vek{\lambda},\vek{a}}$ that render $\bar{\cal P}(\bar{\vek{p}})$ as close to Gaussian as possible.
If $N_p$ is small or the likelihood evaluation computationally inexpensive, it may be more convenient and faster to obtain the distribution ${\cal P}(\vek{p})$ directly on a grid instead of using a random sample \citep[see also][]{frommert10}. In this case the transformed posterior can be computed readily via equation (\ref{eq:bcdistribution}) for any combination of Box-Cox parameters. The optimal parameter combination is then found by comparing $\bar{\cal P}(\bar{\vek{p}})$ to a Gaussian distribution with the same mean and covariance, e.g. by minimising the Kullback-Leibler divergence
\eqa{
\label{eq:kullbackleibler}
D_{\rm KL} &=& \int \dd^{N_p} p\; {\cal P}_{\rm ref}(\vek{p}) \ln \frac{{\cal P}_{\rm ref}(\vek{p})}{{\cal P}(\vek{p})}\\ \nn
&\approx& \sum_i {\cal P}_{\rm ref}(\vek{p}_i) \ln \frac{{\cal P}_{\rm ref}(\vek{p}_i)}{{\cal P}(\vek{p}_i)} \prod_{\mu=1}^{N_p} \Delta p_\mu\;.
}
In the second equality we have replaced the integration with a sum over all points of the grid on which the distributions are evaluated, assuming a spacing of the points by $\Delta p_\mu$ in dimension $\mu$. However, we will use $D_{\rm KL}$ to assess the accuracy of the results of our method in Section \ref{sec:performance}, so that we use a different statistic to determine the Box-Cox parameters.
Two one-dimensional distributions can be compared via their quantiles in a QQ-plot. If both distributions are Gaussian, the quantile pairs lie on a straight and hence Pearson's correlation coefficient of the quantiles,
\eq{
\label{eq:rqq}
r_{\rm QQ} = \frac{\ba{\br{Q^{\rm trans}-\langle Q^{\rm trans} \rangle} \bigl(Q^{\rm gauss}-\langle Q^{\rm gauss} \rangle \bigr)}}{\sqrt{\ba{\br{Q^{\rm trans}-\ba{Q^{\rm trans}}}^2}\; \ba{\br{Q^{\rm gauss}-\ba{Q^{\rm gauss}}}^2}}}\;,
}
should attain unity. Here, $Q^{\rm trans}$ denotes the quantiles of the Box-Cox transformed distribution and $Q^{\rm gauss}$ the quantiles of a zero-mean unit-variance Gaussian distribution, the latter readily computed from the cumulative distribution function. In practice we use 30-quantiles to calculate equation (\ref{eq:rqq}). An advantage of $r_{\rm QQ}$ over $D_{\rm KL}$ is that it is independent of the mean and variance of the transformed distribution which therefore do not have to be re-computed for every change in Box-Cox parameters.
Since $r_{\rm QQ}$ can only be applied to one-dimensional distributions, we determine the Box-Cox parameters in every dimension of parameter space from the marginalised posterior in that dimension. When following the approach of using a random sample $\vek{\hat{p}}$ together with equation (\ref{eq:bclikelihood}) to optimise the Box-Cox parameters, we will compare the performance of determining $\br{\vek{\lambda},\vek{a}}$ from the full $N_p$-dimensional posterior and the $N_P$ marginal posteriors.
\subsection{Box-Cox-Fisher formalism}
\label{sec:bcfisher}
Once the optimal Box-Cox parameters are found by either of the methods described in the foregoing section, one can proceed to unite the Box-Cox transformations with the Fisher matrix technique. If the same set of experimental parameters is used for the Box-Cox-Fisher prediction as for the fiducial mock likelihood analysis that the optimal Box-Cox parameters were determined from, one should obtain identical results. Changing the experimental setup in the Box-Cox-Fisher forecasts should then yield similarly accurate results, as long as these parameters do not depart too strongly from those of the mock likelihood analysis such that the shape of the posterior would be modified significantly. The universality with respect to changes in various parameters of the exemplary weak lensing survey will be tested in Section \ref{sec:performance}.
The task is hence to compute the posterior distribution, ${\cal P}(\vek{p})$, of model parameters $\vek{p}$ for a given set of Box-Cox parameters $\br{\vek{\lambda},\vek{a}}$ and a standard Fisher matrix $F^{\rm orig}$, computed for at a fiducial point $\vek{p}_{\rm fid}$ in parameter space. In analogy to equation (\ref{eq:transformedposterior}) the posterior is given by
\eqa{
\label{eq:transformedposteriorfisher}
\nn
{\cal P}(\vek{p}) &=& \!\!\! \sqrt{\frac{\det \bar{F}}{(2\pi)^{N_p}}}\; \exp \bc{- \frac{1}{2}\; (\bar{\vek{p}}-\bar{\vek{p}}_{\rm max})^\tau\; \bar{F}\; (\bar{\vek{p}}-\bar{\vek{p}}_{\rm max})}\\
&& \times\; \prod_{\mu=1}^{N_p} \br{p_\mu + a_\mu}^{\lambda_\mu-1}\;,
}
where we used the transformed Fisher matrix $\bar{F}$ as an estimator for the inverse covariance of the Gaussian of the Box-Cox transformed posterior. The peak position $\bar{\vek{p}}_{\rm max}$ of this Gaussian and $\bar{F}$ are the only unknown quantities in equation (\ref{eq:transformedposteriorfisher}) that have yet to be determined.
In the following we will assume, as in the standard derivation of the Fisher matrix, that the prior is uniform in the region of parameter space where the likelihood deviates significantly from zero. Thus the log-likelihood is given by ${\cal L} = - \ln {\cal P}$, and likewise for the transformed posterior. Then equation (\ref{eq:transformedposteriorfisher}) is equivalent to
\eq{
\label{eq:loglike}
\bar{\cal L} = {\cal L} - \sum_{\mu=1}^{N_p} (1-\lambda_\mu)\; \ln \br{p_\mu + a_\mu}\;.
}
If we designate $\vek{p}_{\rm max}$ as the result of an inverse Box-Cox transformation of $\bar{\vek{p}}_{\rm max}$ and employ the definition of the Fisher matrix, we arrive at the following expression for the transformed Fisher matrix,
\eqa{
\label{eq:bcfisher1}
&& \bar{F}_{\mu\nu} \equiv \ba{ \left. \frac{\partial^2 \bar{\cal L}}{\partial \bar{p}_\mu \partial \bar{p}_\nu} \right|_{\bar{\vek{p}}_{\rm max}}}\\ \nn
&=& \!\!\! \ba{ \left. \frac{\partial^2 \cal L}{\partial p_\mu \partial p_\nu} \right|_{\vek{p}_{\rm max}}} \br{p_{\mu,{\rm max}}+a_\mu}^{1-\lambda_\mu} \br{p_{\nu,{\rm max}}+a_\nu}^{1-\lambda_\nu}\\ \nn
&& +\; \delta_{\mu\nu} \Biggl\{ \lambda_\mu (\lambda_\mu-1) \br{p_{\mu,{\rm max}}+a_\mu}^{-2\lambda_\mu}\\ \nn
&& + (1-\lambda_\mu) \br{p_{\mu,{\rm max}}+a_\mu}^{1-2\lambda_\mu} \ba{ \left. \frac{\partial \cal L}{\partial p_\mu} \right|_{\vek{p}_{\rm max}}} \Biggr\}\;.
}
Here, angular brackets denote expectation values, and $\delta_{\mu\nu}$ is the Kronecker symbol.
At this point we make the simplifying assumption that $\vek{p}_{\rm max} \approx \vek{p}_{\rm fid}$, i.e. that the Box-Cox transformation maps the peak of the original posterior onto the peak of the transformed posterior. As will be demonstrated below, this approximation holds to high accuracy. Alternatively, one could instead Taylor-expand the expectation values of the first and second derivatives of ${\cal L}$ in equation (\ref{eq:bcfisher1}), but this step would necessitate the computation of third-order derivatives of ${\cal L}$ already at the first order of the expansion.
Replacing $\vek{p}_{\rm max}$ by $\vek{p}_{\rm fid}$ in equation (\ref{eq:bcfisher1}), the expectation of the first derivative of the log-likelihood vanishes because it has a maximum at $\vek{p}_{\rm fid}$. Invoking the definition of the standard Fisher matrix for the original distribution, one obtains
\eqa{
\label{eq:bcfisher2}
\bar{F}_{\mu\nu} &\approx& F_{\mu\nu}^{\rm orig}\; \br{p_{\mu,{\rm fid}}+a_\mu}^{1-\lambda_\mu} \br{p_{\nu,{\rm fid}}+a_\nu}^{1-\lambda_\nu}\\ \nn
&& +\; \delta_{\mu\nu}\; \lambda_\mu\; (\lambda_\mu-1) \br{p_{\mu,{\rm fid}}+a_\mu}^{-2\lambda_\mu}\;.
}
We pursue two approaches to determine $\bar{\vek{p}}_{\rm max}$, or equivalently, $\vek{p}_{\rm max}$. Requiring that the transformed posterior peaks at $\bar{\vek{p}}_{\rm max}$ yields the condition
\eqa{
\label{eq:pmaxforward}
\ba{ \left. \frac{\partial \bar{\cal L}}{\partial \bar{p}_\mu} \right|_{\bar{\vek{p}}_{\rm max}}} &=& \ba{ \left. \frac{\partial \cal L}{\partial p_\mu} \right|_{\vek{p}_{\rm max}}} \br{p_{\mu,{\rm max}}+a_\mu}^{1-\lambda_\mu}\\ \nn
&& \;+ (\lambda_\mu-1) \br{p_{\mu,{\rm max}}+a_\mu}^{-\lambda_\mu} = 0\;,
}
which can be numerically solved after Taylor-expanding the expectation value around $\vek{p}_{\rm fid}$,
\eq{
\label{eq:taylordl}
\ba{ \left. \frac{\partial \cal L}{\partial p_\mu} \right|_{\vek{p}_{\rm max}}} \approx \sum_{\nu=1}^{N_p} F_{\mu\nu}^{\rm orig} \br{p_{\nu,{\rm max}}-p_{\nu,{\rm fid}}}\;.
}
Alternatively, one can determine $\vek{p}_{\rm max}$ such that the original distribution ${\cal P}(\vek{p})$ peaks at $\vek{p}_{\rm fid}$, which, using equation (\ref{eq:loglike}), leads to the condition
\eqa{
\label{eq:pmaxbackward}
\ba{ \left. \frac{\partial {\cal L}}{\partial p_\mu} \right|_{\vek{p}_{\rm fid}}} &=& \br{p_{\mu,{\rm fid}}+a_\mu}^{\lambda_\mu-1}\\ \nn
&& \hspace*{-2.7cm} \times\; \sum_{\nu=1}^{N_p} \bar{F}_{\mu\nu} \br{\bar{p}_{\nu,{\rm fid}} - \bar{p}_{\nu,{\rm max}}} - (\lambda_\mu-1) \br{p_{\mu,{\rm fid}}+a_\mu}^{-1} = 0\;.
}
After inserting the approximation given by equation (\ref{eq:bcfisher2}), one obtains an expression that can analytically be solved for $\vek{p}_{\rm max}$. Since both procedures involve approximations, we will compare their performance below in Section \ref{sec:comparison}.
Gaussian priors can be added to the diagonal of $F_{\mu\nu}^{\rm orig}$ in the same way as for the standard Fisher analysis, but if the priors modify the posterior substantially, they also have to be included in the mock likelihood analysis used to find optimal Box-Cox parameters. Note that, when grid or Monte-Carlo sampling this likelihood, one usually defines a maximum range in which the model parameters are allowed to vary. This corresponds to an implicit top-hat prior which cannot be represented in the Fisher matrix formalism. Hence, one has to make sure that the posterior used to determine Box-Cox parameters lies well within the parameter space considered.
\section{Performance}
\label{sec:performance}
To assess the performance of Fisher matrix forecasts combined with Box-Cox transformations, we consider a mock weak lensing survey as outlined in the following. While the modelling is at a level of realism similar to current predictions for planned observational projects, we do not attempt to mimic any particular survey, but rather choose the survey characteristics such that we obtain a posterior distribution of cosmological parameters which serves as a particularly useful benchmark.
Hence, our mock survey will produce a pronounced non-linear degeneracy between the parameters $\Omega_{\rm m}$, the matter density, and $\sigma_8$, the normalisation of matter density fluctuations as an ideal test case. Note that actual future weak lensing surveys will generate much stronger parameter constraints and a reduced $\Omega_{\rm m}-\sigma_8$ degeneracy, so that the Box-Cox-Fisher formalism should perform well in these cases once it does so for the scenario studied in this work.
We will then investigate in detail the implementation outlined in Section \ref{sec:theory}, before answering the question how accurate the Box-Cox-Fisher formalism is when varying the fiducial cosmology, survey parameters, the weak lensing statistic entering the likelihood, and the dimension of the posterior distribution. To be of practical use, the proposed method has to capture the change in the posterior distribution caused by all these variations. Only then can the formalism be employed for efficient forecasting of parameter constraints after a single initial full mock likelihood analysis needed to determine the Box-Cox parameters.
\subsection{Mock weak lensing survey}
\label{sec:survey}
Weak lensing surveys measure the shapes of millions of distant galaxy images which undergo tiny modifications when the light emitted by these galaxies is gravitationally lensed on its way to Earth. Correlating the shapes of pairs of galaxies, one can infer the statistical properties of the matter distribution projected along the line of sight, which in turn depends on the cosmological model. In addition, the weak lensing signal depends on the distances between observer, the structures acting as lenses, and the source galaxy, which provides information about the expansion history of the Universe. For details about gravitational lensing theory we refer the reader to \citet{bartelmann01}; for a recent review on weak lensing measurements see e.g. \citet{munshi08}.
While the majority of weak lensing studies use two-point correlation functions as the observable (see Section \ref{sec:corr}), predictions generally rely on Fourier space measures due to their direct connection to theory and their simple covariance properties. The power spectrum of the dimensionless projected mass density $\kappa$ reads \citep{kaiser92}
\eq{
\label{eq:limber}
C_\kappa(\ell) = \frac{9H_0^4 \Omega_{\rm m}^2}{4 c^4} \int^{\chi_{\rm hor}}_0 \dd \chi\; \frac{g^2(\chi)}{a^2(\chi)}\; P_\delta \br{\frac{\ell}{\chi},\chi}\;,
}
where $\ell$ denotes angular frequency, $H_0$ the Hubble constant, and $a$ the cosmological scale factor. The integral runs over comoving distance $\chi$ up to the horizon distance $\chi_{\rm hor}$. The power spectrum of the three-dimensional matter distribution is given by $P_\delta$, which depends on wavenumber $k=\ell/\chi$ and epoch, specified in terms of comoving distance. The geometrical contributions to equation (\ref{eq:limber}) are collected in the lensing efficiency
\eq{
\label{eq:lenseff}
g(\chi) = \int^{\chi_{\rm hor}}_{\chi} \dd \chi'\, n_{\rm g}\bb{z(\chi')}\, \br{1-\frac{\chi}{\chi'}}\;,
}
where $n_{\rm g}(z)$ denotes the normalised redshift distribution of galaxies in the survey.
We use $C_\kappa(\ell)$ as our weak lensing observable and evaluate it for 100 angular frequency bins, logarithmically spaced between $\ell_{\rm min}=10$ and $\ell_{\rm max}=10^4$. The fiducial cosmology used in our calculations is set to $\Omega_{\rm m}=0.25$, $\sigma_8=0.9$, the baryon density $\Omega_{\rm b}=0.05$, the power-law exponent of the initial matter power spectrum generated by inflation $n_{\rm s}=1.0$, and the Hubble parameter $h=0.7$, where $H_0 = h\, 100\,{\rm km/s/Mpc}$. Moreover the geometry of the Universe is assumed flat by default. To compute $P_\delta$, we employ the transfer function by \citet{eisenstein98} and apply the corrections due to non-linear evolution by \citet{PeacockDodds}.
The projected surface mass density is assumed to be Gaussian distributed, which implies that the covariance is given by \citep{joachimi08}
\eq{
\label{eq:covpower}
{\rm Cov}\bb{C_\kappa(\ell);C_\kappa(\ell')} = \delta_{\ell \ell'}\; \frac{4 \pi}{A_{\rm s}\, \ell\, \Delta \ell} \left( C_\kappa(\ell) + \frac{\sigma_\epsilon^2}{2\bar{n}_{\rm g}} \right)^2,
}
i.e. different angular frequencies are uncorrelated\footnote{Note that the assumption of Gaussianity is simplistic, in particular for high angular frequencies \citep[e.g.][]{kiessling11}, but still widely used for Fisher matrix forecasts (see \citealp{kiessling11b} though).}. Here, $\Delta \ell$ is the width of the angular frequency bin, and $A_{\rm s}=100\,{\rm deg}^2$ the survey size. The random orientations of the intrinsic shapes of source galaxies yield a shape noise contribution to equation (\ref{eq:covpower}), determined by the intrinsic ellipticity dispersion $\sigma_\epsilon=0.35$ and the total number density of galaxies on the sky $\bar{n}_{\rm g}=20\,{\rm arcmin}^{-2}$. We have implemented a redshift distribution of the form
\eq{
\label{eq:nofz}
n_{\rm g}(z) \propto z^2\; \exp \bc{-(z/z_0)^{1.5}}\;,
}
where the characteristic redshift scale $z_0$ is related to the median redshift via $z_0 \approx z_{\rm med}/1.4$. The survey is assumed to have a median redshift $z_{\rm med}=0.9$.
Following widespread practice, we make use of a Gaussian likelihood for $C_\kappa(\ell)$,
\eq{
\label{eq:cslikelihood}
L \propto \exp \bc{ - \frac{1}{2} \sum_{\ell=\ell_{\rm min}}^{\ell_{\rm max}} \frac{\bb{C_\kappa(\ell,\vek{p}_{\rm fid})-C_\kappa(\ell,\vek{p})}^2}{{\rm Cov}\bb{C_\kappa(\ell);C_\kappa(\ell)}}}\;,
}
where the power spectra obtained for the fiducial cosmology $\vek{p}_{\rm fid}$ serve as our mock datavector. We assume flat priors and make sure that the likelihood peaks well inside the region of parameter space considered, so that the posterior is readily obtained from $L$ by renormalisation in parameter space.
\begin{figure*}
\centering
\includegraphics[scale=.42,angle=270]{boxcoxfisher_omsig.ps}
\caption{\textit{Left panel}: $1\sigma$ and $2\sigma$ confidence levels in the $\Omega_{\rm m}-\sigma_8$ plane for the full likelihood analysis (blue dotted lines), the Box-Cox transformed posterior based on marginal distributions (red solid lines), and the Box-Cox-transformed posterior based on the two-dimensional distribution (orange solid lines). For comparison the results for a naive Fisher matrix computation are shown as black lines. \textit{Right panels}: Same as above, but for the marginalised distributions of $\Omega_{\rm m}$ (bottom) and $\sigma_8$ (top).}
\label{fig:parametertrafo}
\end{figure*}
Again assuming Gaussianity, the corresponding Fisher matrix reads
\eq{
\label{eq:standardfisher}
F_{\mu\nu}^{\rm orig} = \sum_{\ell=\ell_{\rm min}}^{\ell_{\rm max}} \frac{\partial C_\kappa(\ell)}{\partial p_\mu}\; {\rm Cov}^{-1}\bb{C_\kappa(\ell);C_\kappa(\ell)}\; \frac{\partial C_\kappa(\ell)}{\partial p_\nu}\;,
}
where both the derivatives and the covariance are evaluated at $\vek{p}_{\rm fid}$. In writing equation (\ref{eq:standardfisher}) we have assumed that the covariance does not depend on cosmology; for the same reason we keep the covariance in equation (\ref{eq:cslikelihood}) fixed at its value for the fiducial set of cosmological parameters.
\subsection{Comparison of implementations}
\label{sec:comparison}
For most of the analysis we will only vary $\Omega_{\rm m}$ and $\sigma_8$ and keep all other cosmological parameters at their fiducial values. We compute the posterior on a grid in the $\Omega_{\rm m}-\sigma_8$ plane according to equation (\ref{eq:cslikelihood}) and also derive the marginal distributions for the two parameters. In Fig.$\,$\ref{fig:parametertrafo} we show confidence levels and marginal distributions for the likelihood analysis as well as for the standard Fisher matrix analysis using equation (\ref{eq:standardfisher}). While the marginal Fisher matrix errors on $\Omega_{\rm m}$ and $\sigma_8$ are still relatively close to the actual results, neither the tails in the marginal distributions, nor the banana-shaped form of the two-dimensional posterior and the extent of the confidence contours along the degeneracy can be reproduced by the standard Fisher matrix.
As a first step in the Box-Cox-Fisher formalism we determine the Box-Cox parameters from the full likelihood, using either the concentrated maximum likelihood from equation (\ref{eq:bclikelihood}) or the QQ-plot correlation coefficient from equation (\ref{eq:rqq}). The latter is restricted to one-dimensional distributions, i.e. in this case the marginal distributions of both $\Omega_{\rm m}$ and $\sigma_8$, whereas $L_{\rm max}$ is calculated for the individual marginal distributions as well as for the two-dimensional posterior. To obtain $L_{\rm max}$, a random sample of size $10^6$ is created from the respective distribution. The optimal values for $(\vek{\lambda},\vek{a})$ for which $L_{\rm max}$ or $r_{\rm QQ}$ attain a maximum are listed in Table \ref{tab:parametertrafo}.
\begin{table*}
\centering
\caption{Kullback-Leibler divergence $D_{\rm KL}$ between the posterior obtained from the full likelihood analysis and the posterior from the Box-Cox transformed Fisher matrices, using different implementations. Shown is $D_{\rm KL}$ for the distribution in the $\Omega_{\rm m}-\sigma_8$ plane in the second column, as well as for the marginalised distributions of $\Omega_{\rm m}$ and $\sigma_8$ in the third and fourth column. In the fifth to eighth column the optimum Box-Cox transformation parameters are listed for every implementation. Box-Cox parameters are either determined from the marginal distributions (1D) or the two-dimensional likelihood (2D). Results when using the different approaches to determining $\vek{p}_{\rm max}$ are also compared. For comparison results are also given for a standard Fisher matrix analysis.}
\begin{tabular}[t]{lrrrrrrr}
\hline\hline
analysis method & $D_{\rm KL}(\Omega_{\rm m})$ & $D_{\rm KL}(\sigma_8)$ & $D_{\rm KL}(\Omega_{\rm m},\sigma_8)$ & $\lambda(\Omega_{\rm m})$ & $a(\Omega_{\rm m})$ & $\lambda(\sigma_8)$ & $a(\sigma_8)$ \\
\hline
standard Fisher & 0.143 & 0.029 & 3.482 & - & - & - & - \\
Box-Cox 1D; $\vek{p}_{\rm max}$ from eq.$\,$(\ref{eq:pmaxbackward}) & 0.008 & 0.008 & 0.022 & -0.74 & 0.03 & 1.54 & 0.28 \\
Box-Cox 1D; $\vek{p}_{\rm max}$ from eq.$\,$(\ref{eq:pmaxforward}) & 0.008 & 0.007 & 0.043 & -0.74 & 0.03 & 1.54 & 0.28 \\
Box-Cox 1D; $\lambda,a$ via $L_{\rm max}$ & 0.010 & 0.008 & 0.040 & -0.09 & -0.08 & 3.73 & 4.00 \\
Box-Cox 2D; $\vek{p}_{\rm max}$ from eq.$\,$(\ref{eq:pmaxbackward}) & 0.013 & 0.013 & 0.017 & -0.03 & -0.03 & 0.87 & -0.46 \\
Box-Cox 2D; $\vek{p}_{\rm max}$ from eq.$\,$(\ref{eq:pmaxforward}) & 0.014 & 0.013 & 0.017 & -0.03 & -0.03 & 0.87 & -0.46 \\
\hline
\end{tabular}
\label{tab:parametertrafo}
\end{table*}
Working on one- or two-dimensional distributions, with $L_{\rm max}$ or $r_{\rm QQ}$ as statistic, results in largely different optimal values for the Box-Cox parameters. To gain further insight, we plot both statistics in the plane spanned by $\lambda$ and $a$ for the marginal distribution of $\sigma_8$ in Fig.$\,$\ref{fig:bcplane_sigma8}, left panel. Both $L_{\rm max}$ or $r_{\rm QQ}$ agree well in the region where they maximise. For a wide range in $(\lambda,a)$-space this maximum lies on a nearly perfect and almost linear degeneracy line.
\begin{figure*}
\centering
\includegraphics[scale=.4,angle=270]{bcplane.ps}
\caption{\textit{Left panel}: Concentrated likelihood $L_{\rm max}$, see equation (\ref{eq:bclikelihood}), and QQ-plot correlation coefficient $r_{\rm QQ}$, see equation (\ref{eq:rqq}), for the marginalised distribution of $\sigma_8$, as a function of Box-Cox transformation parameters $\lambda$ and $a$. Red solid lines correspond to $r_{\rm QQ}$ and indicate a deviation of $10^{-5}$ and $10^{-4}$ from the maximum of 1. The relative deviation of $L_{\rm max}$ from its maximum is shown in grey shading, varying logarithmically between 0.1 (white) and $10^{-5}$ (black). \textit{Right panel}: Skewness and excess kurtosis of the transformed distribution as a function of $\lambda$ and $a$. Levels of constant skewness are shown in red, indicating values of 0.1, 0.01, -0.01, -0.1 from top to bottom. Contours for negative values are dotted. The kurtosis is shown in grey shading, varying linearly between 1 (black) and -0.1 (white). Levels of zero kurtosis are indicated by the black lines.}
\label{fig:bcplane_sigma8}
\end{figure*}
This degeneracy is mirrored in the shape of the Box-Cox transformed distribution, as can be seen in the right panel of Fig.$\,$\ref{fig:bcplane_sigma8}, where we show the skewness and excess kurtosis of the transformed distribution. The degeneracy in maximum $L_{\rm max}$ or $r_{\rm QQ}$ is closely matched by the minimum skewness with values close to zero. The kurtosis also features this degeneracy; however, it does not vanish, but instead obtains a shallow minimum at small negative values along the degeneracy line. Note that the skewness and kurtosis of the original distribution can be read off at $\lambda=1$. In this case contour lines are horizontal.
The mean and variance of the transformed distributions increase along the degeneracy line for larger values of $\lambda$ and $a$, so that the degeneracy can be broken by fixing either of the two lowest-order moments of the transformed distributions. However, since mean and variance are uncritical for our purposes, we leave them as free parameters and simply use the $(\lambda,a)$ combinations on the degeneracy line that our codes produce, the exact values hence determined by numerical effects and the maximisation algorithm used. See e.g. the values for $\lambda$ and $a$ in the third and fourth row of Table \ref{tab:parametertrafo} which lie in the region of maximum $L_{\rm max}$, $r_{\rm QQ}$ and minimum skewness. In the appendix we provide a toy model that illustrates basic properties of Box-Cox transformations including the degeneracy between $\lambda$ and $a$ discussed here.
With optimal values for $\lambda$ and $a$ at hand, we compute the transformed Fisher matrix as given in equation (\ref{eq:bcfisher2}) and subsequently the transformed posterior according to equation (\ref{eq:transformedposteriorfisher}). The resulting confidence contours and marginal distributions, with Box-Cox parameters obtained from the marginal distributions via $r_{\rm QQ}$ (1D) as well as from the full posterior via $L_{\rm max}$ (2D), are also shown in Fig.$\,$\ref{fig:parametertrafo}. Furthermore we provide a quantitative statement on how accurately the Box-Cox transformed posterior matches the actual one by calculating the Kullback-Leibler divergence $D_{\rm KL}$ as given by equation (\ref{eq:kullbackleibler}) between the two distributions in Table \ref{tab:parametertrafo}, again for both the two-dimensional and marginal cases.
Both visual and quantitative inspection demonstrate that the Box-Cox-Fisher formalism excellently reproduces the actual posterior, for all variants of the implementation considered. Compared to the standard Fisher results, the Box-Cox-Fisher formalism improves $D_{\rm KL}$ by a factor of 2 to 4 in the case of the marginal distribution of $\sigma_8$ and by at least an order of magnitude for the marginal distribution of $\Omega_{\rm m}$. The decrease in $D_{\rm KL}$ can mainly be ascribed to the accurate modelling of the non-Gaussian wings of the distributions, but partly also to the shift in the maximum of the marginal distributions away from the fiducial cosmology which the standard Fisher formalism cannot account for.
As the left-hand panel in Fig.$\,$\ref{fig:parametertrafo} suggests, the most blatant discrepancy between the standard and Box-Cox-Fisher analysis happens in the $\Omega_{\rm m}-\sigma_8$ plane, with about two orders of magnitude difference in $D_{\rm KL}$. The overall form of the posterior is represented accurately by the Box-Cox-Fisher contours; only the extent of the $2\sigma$ confidence levels reveals small residual deviations. As expected, if the Box-Cox parameters are derived from the marginal distributions, $D_{\rm KL}$ for the marginal distributions is smaller than for the 2D approach, and vice versa in the case of $D_{\rm KL}$ in the $\Omega_{\rm m}-\sigma_8$ plane.
Note that we have also compared the algorithms given by equations (\ref{eq:pmaxforward}) and (\ref{eq:pmaxbackward}) to calculate $\vek{p}_{\rm max}$ in Table \ref{tab:parametertrafo}. Both perform equally well, but since equation (\ref{eq:pmaxbackward}) can be solved analytically for $\vek{p}_{\rm max}$, we will employ this version henceforth. Moreover we are going to apply the 2D approach, i.e. determining the Box-Cox parameters from the full posterior via $L_{\rm max}$, for the remainder of this paper.
\begin{figure}
\centering
\includegraphics[scale=.39,angle=270]{boxcoxfisher_transformed.ps}
\caption{$1\sigma$ and $2\sigma$ confidence levels in the plane of the Box-Cox transformed parameters. Red contours correspond to the transformed full likelihood, black contours originate from the transformed Fisher matrix given by equation (\ref{eq:bcfisher2}) and centred at $\bar{\vek{p}}_{\rm max}$. The posteriors are Gaussian to good approximation and agree well. The results shown were obtained for the case which is shown in Fig.$\,$\ref{fig:parametertrafo} as orange lines.}
\label{fig:parametertrafo2}
\end{figure}
As seen in Fig.$\,$\ref{fig:bcplane_sigma8}, right panel, the optimal choice of Box-Cox parameters guarantees that the transformed distribution has vanishing skewness and low excess kurtosis, and hence can be assumed to be close to a multivariate Gaussian. Consulting equation (\ref{eq:transformedposteriorfisher}), this transformed distribution should additionally be described well by the transformed Fisher matrix, see equation (\ref{eq:bcfisher2}), with its peak at $\bar{\vek{p}}_{\rm max}$. This is illustrated in Fig.$\,$\ref{fig:parametertrafo2}, and indeed the transformed full posterior is closely matched by the transformed Fisher matrix contours. The slightly more extended confidence contours for the full likelihood might hint at a mildly platykurtic distribution which agrees with the small negative values of excess kurtosis along the degeneracy line in Fig.$\,$\ref{fig:bcplane_sigma8}. The ability of the Box-Cox transformations to change the posterior into a multivariate Gaussian opens up a range of potential applications, as we will discuss further in Section \ref{sec:conclusions}.
\subsection{Varying cosmology and survey parameters}
\label{sec:scaling}
\begin{table*}
\centering
\caption{Kullback-Leibler divergence $D_{\rm KL}$ between the posterior obtained from the full likelihood analysis and the posterior from the Box-Cox transformed Fisher matrices, varying different survey or cosmological parameters as indicated in the first column. Shown is $D_{\rm KL}$ for the distribution in the $\Omega_{\rm m}-\sigma_8$ plane in the second column, as well as the marginalised distributions for $\Omega_{\rm m}$ and $\sigma_8$ in the third and fourth column.}
\begin{tabular}[t]{lrrr}
\hline\hline
parameters changed & $D_{\rm KL}(\Omega_{\rm m})$ & $D_{\rm KL}(\sigma_8)$ & $D_{\rm KL}(\Omega_{\rm m},\sigma_8)$\\
\hline
fiducial parameters & 0.013 & 0.013 & 0.017 \\
$\Omega_{\rm m}: 0.25 \rightarrow 0.225$ & 0.008 & 0.008 & 0.025 \\
$z_{\rm med}: 0.9 \rightarrow 1.0$; $n_{\rm g}: 20 \rightarrow 37.4\,{\rm arcmin}^{-2}$ & 0.005 & 0.005 & 0.019 \\
$\ell_{\rm max}: 10000 \rightarrow 8700$ & 0.015 & 0.015 & 0.019 \\
$A_{\rm s}: 100 \rightarrow 110\,{\rm deg}^2$ & 0.013 & 0.012 & 0.016 \\
\hline
\end{tabular}
\label{tab:parametertrafo2}
\end{table*}
We expect the Box-Cox-Fisher formalism to be particularly useful in an advanced planning stage of an experiment when e.g. the capabilities of breaking model parameter degeneracies come into focus. By then the survey parameters and the analysis strategies should not change radically anymore, but only in relatively small steps and only a few parameters at a time. If that holds true, the general form of the posterior is only moderately modified under these changes, so that one can continue to use the optimal Box-Cox parameters determined from the initial full likelihood analysis.
The Box-Cox-Fisher analysis is repeated for several survey configurations that each differ in one or two parameters from the fiducial survey by about $10\,\%$. These changes are accounted for in the Fisher matrix, but we retain the values of the Box-Cox parameters determined for the fiducial survey. A full likelihood analysis is computed as well for every configuration, but solely for the purpose of assessing the accuracy of the forecast.
We modify the fiducial cosmology by lowering $\Omega_{\rm m}$ by $10\,\%$. A slightly deeper survey is analysed, increasing $z_{\rm med}$ to 1, which also increases the number density of galaxies and consequently reduces the noise contribution. Applying the scaling found by \citet{amara07}, the deeper survey has $\bar{n}_{\rm g}=37.4\,{\rm arcmin}^{-2}$. Moreover we consider the case of discarding the highest angular frequency bins in the analysis, reducing $\ell_{\rm max}$ to 8700. Finally, we increase the survey size by $10\,\%$.
Analogously to the foregoing section, we employ the Kullback-Leibler divergence to compare the Box-Cox-Fisher result with the posterior from the full likelihood analysis. As is evident from Table \ref{tab:parametertrafo2}, $D_{\rm KL}$ for the marginal distributions and the posterior in the $\Omega_{\rm m}-\sigma_8$ plane remains constant to good approximation in all cases.
\subsection{Varying statistic and posterior dimension}
\label{sec:corr}
One of the most likely modifications in mock weak lensing analyses is a change in the statistic used as the observable. We switch to the frequently employed correlation function $\xi_+$ which is related to the power spectrum via \citep{SvWM02}
\eq{
\label{eq:xiplus}
\xi_+(\theta) = \int^\infty_0 \frac{\dd \ell\, \ell}{2 \pi}\; J_0(\ell \theta)\; C_\kappa(\ell)\;,
}
where $J_0$ is the Bessel function of the first kind of order 0. The covariance of the correlation function can directly be determined from equation (\ref{eq:covpower}), as detailed in \citet{joachimi08}. We intend to roughly use the same angular scales as in the power spectrum analysis and thus consider the range $1\,{\rm arcmin}<\theta<5\,{\rm deg}$, divided into 50 logarithmically spaced bins. Note that this range of angular scales does not ensure a similar information content because the angular separation bins are strongly correlated.
\begin{figure*}
\centering
\includegraphics[scale=.42,angle=270]{boxcoxfisher_omsigcorr.ps}
\caption{\textit{Left panel}: $1\sigma$ and $2\sigma$ confidence levels in the $\Omega_{\rm m}-\sigma_8$ plane for the full likelihood analysis (blue dotted lines) and the Box-Cox transformed posterior (red solid lines), using the shear correlation function $\xi_+$ as statistic. The optimal Box-Cox parameters obtained from the power spectrum analysis also yield good results in this case, as indicated by the black solid lines. \textit{Right panels}: Same as above, but for the marginalised distributions of $\Omega_{\rm m}$ (bottom) and $\sigma_8$ (top).}
\label{fig:parametertrafo_corr_2d}
\end{figure*}
Moreover we now use Population Monte-Carlo sampling with CosmoPMC\footnote{\texttt{http://www2.iap.fr/users/kilbinge/CosmoPMC/}} \citep{cappe08,wraith09} to create a random sample of size $10^5$ from the full posterior in order to determine optimal Box-Cox parameters via equation (\ref{eq:bclikelihood}). The results are presented in Fig.$\,$\ref{fig:parametertrafo_corr_2d}, finding again excellent agreement between Box-Cox-Fisher results and full posterior. If the optimal Box-Cox parameters that were obtained for the power spectrum analysis in Section \ref{sec:comparison} are used instead, one arrives at constraints of similar quality. Therefore the Box-Cox-Fisher formalism should also be robust with respect to a change in the weak lensing statistic employed in the Fisher matrix.
As a final test for the practical applicability of the novel forecasting method, we have to verify that it is accurate for a higher-dimensional posterior. Hence we drop the assumption of a spatially flat Universe and vary the density parameter of dark energy, $\Omega_\Lambda$, as well as $n_{\rm s}$ in addition to $\Omega_{\rm m}$ and $\sigma_8$. We perform the analysis as in the foregoing case, using again CosmoPMC to create about $10^6$ random samples of the four-dimensional posterior to determine in total 8 Box-Cox parameters.
\begin{figure*}
\centering
\includegraphics[scale=.38,angle=270]{boxcoxfisher_4d2.ps}
\caption{$1\sigma$ and $2\sigma$ confidence levels for the full likelihood analysis (blue dotted lines) and the Box-Cox transformed posterior (red solid lines) for all two-dimensional marginalised distributions in the four-dimensional parameter space $\bc{\Omega_{\rm m},\sigma_8,\Omega_\Lambda,n_{\rm s}}$. Note that the Fisher matrix as computed by CosmoPMC has been employed in the Box-Cox analysis.}
\label{fig:parametertrafo_corr_4d}
\end{figure*}
The confidence contours of the marginalised posterior distributions for all possible pairs of cosmological parameters are shown in Fig.$\,$\ref{fig:parametertrafo_corr_4d}. The Box-Cox-Fisher formalism yields contours that are able to adopt arbitrary shapes and represent the four-dimensional posterior accurately, including the non-linear degeneracies in the $\Omega_{\rm m}-\sigma_8$ and $\Omega_\Lambda-n_{\rm s}$ planes. The only significant discrepancies between the Fisher-based contours and the confidence levels derived from the Monte-Carlo sample appears in regions where the posterior declines slowly, e.g. for large $\Omega_{\rm m}$ or small $\sigma_8$. The frayed contour lines indicate that these regions are still sparsely sampled by CosmoPMC. This could imply that the Monte-Carlo sample is not suited to allow for a determination of optimal Box-Cox parameters which lead to an accurate posterior shape in these regions. Alternatively, the Box-Cox-Fisher formalism might well be robust enough to produce a precise representation of the posterior also where it is shallow, so that the difference in contour lines would be caused by the insufficient Monte-Carlo sampling in that regime.
As a further example for the reliability of the Box-Cox-Fisher formalism, we initially observed a slight tilt of the Box-Cox-Fisher confidence contours against those from the Monte-Carlo analysis, particularly in the $\Omega_\Lambda-n_{\rm s}$ plane, which could be traced back to a small difference in the correlation functions computed by CosmoPMC and the authors' code. The latter was used to produce the fiducial $\xi_+$ which served as the mock datavector input to CosmoPMC. The small discrepancy in $\xi_+$ leads to a small shift in the maximum likelihood point as determined by CosmoPMC away from the fiducial cosmolgy, as well as slightly different derivatives of $\xi_+$ with respect to cosmological parameters. Consequently, the authors' code and CosmoPMC produce moderately discrepant Fisher matrices. Using the latter in the Box-Cox-Fisher analysis instead results in the excellent agreement shown in Fig.$\,$\ref{fig:parametertrafo_corr_4d}.
\section{An application: Breaking degeneracies in the $\Omega_{\rm m}-\sigma_8$ plane}
\label{sec:application}
\begin{figure*}
\centering
\includegraphics[scale=.42,angle=270]{boxcoxfisher_bispectrum_degeneracy.ps}
\caption{Combined power spectrum and bispectrum constraints on $\Omega_{\rm m}$ and $\sigma_8$. \textit{Left panel}: $1\sigma$ and $2\sigma$ confidence levels for a standard Fisher matrix analysis of two-point weak lensing statistics (red lines), three-point statistics (blue lines), as well as two- and three-point statistics combined (black lines). \textit{Right panel}: Same as above, but for constraints resulting from the Box-Cox-Fisher analysis.}
\label{fig:bispectrum_degeneracy}
\end{figure*}
The Box-Cox-Fisher formalism is applicable to a wide range of problems. For illustrational purposes we provide in the following a toy example which is again built around a mock weak lensing survey. Future experiments which are currently in the planning stages will not be restricted to measuring two-point statistics like $C_\kappa(\ell)$, but make use of higher-order correlations of galaxy shapes. Three-point statistics such as the bispectrum $B_\kappa(\ell_1,\ell_2,\ell_3)$ have been demonstrated to potentially tighten cosmological parameter constraints considerably, e.g. by breaking degeneracies in the $\Omega_{\rm m}-\sigma_8$ plane \citep{berge10}. Since these conclusions rely entirely on standard Fisher analyses, we set out to investigate whether the breaking of $\Omega_{\rm m}-\sigma_8$ degeneracies is affected by the actual, non-elliptical shapes of confidence levels for both two- and three-point weak lensing statistics.
We treat $B_\kappa(\ell_1,\ell_2,\ell_3)$ as our observable three-point statistic and calculate it via \citep[e.g.][]{takada04}
\eqa{
\label{eq:bispectrum}
B_\kappa(\ell_1,\ell_2,\ell_3) &=& \frac{27H_0^6 \Omega_{\rm m}^3}{8 c^6} \int^{\chi_{\rm hor}}_0 \dd \chi\; \frac{g^3(\chi)}{\chi\; a^3(\chi)}\\ \nn
&& \times\; B_\delta \br{\frac{\ell_1}{\chi},\frac{\ell_2}{\chi},\frac{\ell_3}{\chi},\chi}\;,
}
where $B_\delta$ denotes the matter bispectrum. We employ perturbation theory \citep{fry84} to compute $B_\delta$ from the matter power spectrum, applying the corrections due to non-linear structure evolution given in \citet{scoccimarro01}. We employ the bispectrum covariance according to \citet{joachimi09b}, using only the lowest-order term that is given in terms of power spectra. Noting that the bispectrum is only non-zero if its three arguments can form the sides of a triangle (see e.g. \citealp{joachimi09b} for details), we assemble the datavector out of all such combinations, where $\ell_1,\ell_2,\ell_3$ can have 20 logarithmically spaced values between 10 and 1000.
Performing a full mock likelihood analysis for the bispectrum is computationally costly, even if only two cosmological parameters are varied. As a bi-product from another project, we have bispectrum computations on a $20 \times 20$ grid in the $\Omega_{\rm m}-\sigma_8$ plane at our disposal, albeit for a different cosmology than the fiducial survey outlined in Section \ref{sec:survey}. The grid was created for a fiducial cosmology with deviating parameters $\Omega_{\rm b}=0.045$, $h=0.71$, and $n_{\rm s}=0.963$. Furthermore the non-linear correction for the matter power spectrum by \citet{smith03} was used. The bispectra were obtained for a single source redshift, i.e. the redshift distribution in equation (\ref{eq:nofz}) is replaced by a Dirac delta-distribution peaking at $z_{\rm s}=1$. Finally, the angular frequency binning is slightly different, with 18 bins between 10 and 1500.
We make use of the scaling properties of the Box-Cox-Fisher formalism and determine optimal Box-Cox parameters for the bispectrum from a mock likelihood analysis based on the gridded bispectra. The changes in cosmology are not more than $10\,\%$. Besides, they only affect parameters that are kept fixed in this analysis, and that weak lensing is less sensitive to than $\Omega_{\rm m}$ and $\sigma_8$. The single source redshift, $z_{\rm s}$, is similar to the median redshift of the fiducial survey, so that the lensing efficiency should change only mildly, see equation (\ref{eq:lenseff}). Likewise, the different non-linear corrections and angular frequency coverage should not alter the shape of the posterior in the $\Omega_{\rm m}-\sigma_8$ plane significantly.
In the power spectrum analysis we adopt the Box-Cox parameters determined for the fiducial survey. We also use the fiducial survey parameters, except for $\ell_{\rm max}$ which is reduced to 3000. The low angular frequency cut-offs for both two- and three-point statistics are meant to exclude the deeply non-linear clustering regime and thus improve the simplistic approximations for the covariances, particularly for the bispectrum. Additionally, we combine the constraints from two- and three-point statistics by simply adding Fisher matrices or multiplying posteriors, respectively, i.e. we assume that power spectra and bispectra are uncorrelated (which, again, is simplistic but common practice, e.g. \citealp{berge10}).
\begin{table}
\centering
\caption{Marginalised $2\sigma$ constraints on $\Omega_{\rm m}$ and $\sigma_8$ resulting from the standard Fisher and Box-Cox-Fisher analyses of two-point, three-point, and combined two- and three-point weak lensing statistics.}
\begin{tabular}[t]{llrr}
\hline\hline
method & statistics & $\Omega_{\rm m}$ & $\sigma_8$ \\
\hline
& 2pt & $0.25^{+0.11}_{-0.11}$ & $0.90^{+0.23}_{-0.23}$ \\
Fisher & 3pt & $0.25^{+0.06}_{-0.06}$ & $0.90^{+0.09}_{-0.09}$ \\
& 2pt + 3pt & $0.25^{+0.03}_{-0.03}$ & $0.90^{+0.05}_{-0.05}$ \\
\hline
& 2pt & $0.25^{+0.22}_{-0.11}$ & $0.85^{+0.31}_{-0.27}$ \\
Fisher + Box-Cox & 3pt & $0.25^{+0.10}_{-0.07}$ & $0.89^{+0.11}_{-0.12}$ \\
& 2pt + 3pt & $0.25^{+0.05}_{-0.03}$ & $0.90^{+0.06}_{-0.08}$ \\
\hline
\end{tabular}
\label{tab:bispectrum_degeneracy}
\end{table}
In Fig.$\,$\ref{fig:bispectrum_degeneracy} we contrast the parameter constraints in the $\Omega_{\rm m}-\sigma_8$ plane from the standard Fisher matrix and the Box-Cox-Fisher analysis for two-point statistics, three-point statistics, and both data sets combined. The posterior for the bispectrum constraints alone also features the characteristic $\Omega_{\rm m}-\sigma_8$ degeneracy, albeit with a tilted degeneracy line, a property that is also captured by the standard Fisher matrix \citep[see also][]{takada04,berge10}. Since the intersection of the contours is at a sufficiently large angle, the joint constraints in the Box-Cox-Fisher case produce fairly elliptical confidence contours which are of similar size as those resulting from the standard Fisher matrix.
The marginalised constraints on $\Omega_{\rm m}$ and $\sigma_8$ presented in Table \ref{tab:bispectrum_degeneracy} allow for a more quantitative evaluation. Taking into account the accurate shape of the posterior generally increases the $2\sigma$ confidence interval substantially. This increase is stronger the greater the deviation of the posterior from a Gaussian shape, see e.g. the increase by $50\,\%$ for $\Omega_{\rm m}$ in the power spectrum analysis. In the case of the joint two- and three-point constraints, the absolute change in errors is smaller, but still the $2\sigma$ confidence interval grows by about $40\,\%$ ($30\,\%$) for $\sigma_8$ ($\Omega_{\rm m}$).
\section{Conclusions}
\label{sec:conclusions}
In this work we introduced a novel method to compute precise predictions for statistical constraints on model parameters from future experiments. By combining two generic statistical tools -- the Fisher matrix and Box-Cox transformations, we were able to drop the assumption of Gaussianity in parameter space. Applying Box-Cox transformations to model parameters, one arrives at approximately multivariate Gaussian shapes of the posterior. In this transformed space the Fisher matrix can be computed without suffering from the usual limits of the Gaussian assumption. An inverse Box-Cox transformation of the Fisher matrix results then yields realistic posterior distributions in the original parameter space.
We derived the formalism of the combined Fisher and Box-Cox analysis and detailed different approaches to determining the parameters of the Box-Cox transformation from an inital likelihood analysis. Utilising a mock weak lensing survey, we verified the accuracy of the Box-Cox-Fisher formalism and demonstrated that it robustly accounts for changes in various survey parameters and analysis steps. We expect the method to be particularly useful in the advanced planning stages of upcoming experiments and surveys, e.g. to fine-tune the design with repect to the anticipated parameter constraints, or to quantify the breaking of model parameter degeneracies when combining data sets.
A practical implementation of the Box-Cox-Fisher formalism can look as follows:
\begin{enumerate}
\item Obtain information about the full likelihood for a fiducial experiment, for instance from a gridded likelihood in parameter space or via Monte-Carlo sampling.
\item Determine the optimal Box-Cox parameters using the statistics $L_{\rm max}$ or $r_{\rm QQ}$, see equations (\ref{eq:bclikelihood}) and (\ref{eq:rqq}).
\item Calculate the standard Fisher matrix for the exact experimental setup one is interested in.
\item Compute the posterior via equations (\ref{eq:transformedposteriorfisher}), (\ref{eq:bcfisher2}), and (\ref{eq:pmaxbackward}).
\end{enumerate}
The last two steps can be repeated as required for arbitrary values of experimental parameters, as long as these changes do not alter the shape of the posterior too strongly from the initial likelihood analysis. This might for instance happen if several experimental parameters are varied substantially at the same time. Unfortunately, adding new model parameters to the analysis also potentially modifies the posterior significantly, depending on the correlation of this new parameter with the existing ones. Therefore we consider it unlikely that Box-Cox parameters can be determined to sufficent accuracy from low-dimensional sub-spaces of the posterior distribution.
The price to pay for the dramatically more realistic posteriors and confidence regions compare to the standard Fisher analysis is the need for an initial determination of the Box-Cox parameters which requires detailed information about the full posterior distribution. For a realistic number of model parameters the sampling or gridded evaluation of the likelihood is a computationally expensive step. However, complex experiments demand in practice hundreds of forecast calculations, so that switching to a Box-Cox-Fisher prediction after an initial full mock likelihood analysis is still largely beneficial in terms of computational time. Note that the extra calculations required for the new method add only marginally to the time the corresponding standard Fisher matrix computation takes.
We illustrated a potential application of the Box-Cox-Fisher formalism, investigating the effects of precise posterior modelling on the joint constraints by weak lensing two- and three-point statistics in the $\Omega_{\rm m}-\sigma_8$ plane. We find that while the shapes of confidence contours for the individual constraints from power spectra and bispectra change in a pronounced way from the standard Fisher results, the joint posterior is compact and close to the Gaussian form predicted by the standard Fisher matrix, hence confirming in the simple case we considered that three-point statistics can indeed break the $\Omega_{\rm m}-\sigma_8$ degeneracy to a large extent. However, marginal errors on the cosmological parameters increase substantially by up to $50\,\%$ when using the Box-Cox-Fisher analysis instead of standard Fisher matrix forecasts, which certainly needs to be taken into account for predictions of precision measurements.
Generally, the more compact a posterior is, the more it looks Gaussian. Consequently, the local representation around the maximum provided by the Fisher matrix provides a good description of the complete posterior shape in that case. It should be noted that, in order to provide a challenging benchmark to test our method, and to facilitate the covariance calculations, we deliberately designed our exemplary weak lensing survey to yield only weak cosmological constraints. Future weak lensing surveys will perform much better, thereby rendering the Gaussian approximation in parameter space more appropriate for predictions.
Yet, experiments will always be faced with complex posterior distributions. For instance upcoming large-area weak lensing surveys will be used to test modifications of gravity. A popular parametrisation of deviations from General Relativity introduces the gravitational slip and a modification of Newton's constant, where the two parameters are perfectly degenerate and non-linearly related for weak lensing data alone \citep[see e.g.][]{daniel10b}. Comparing the Fisher matrix forecasts for these parameters in \citet{guzik09} with the likelihood analyses in \citet{daniel10b} and \citet{song10}, it is evident that the optimisation of the survey design for modified gravity measurements will need to go beyond the standard Fisher matrix approach.
For future developments of the Box-Cox-Fisher formalism it will prove fruitful to continue the data analysis in the Box-Cox transformed parameter space, and not transform back to the physically motivated model parameters, as done in this work. Then one can fully exploit the Gaussian form of the posterior and apply the whole arsenal of statistical tools that become accurate, or usable in the first place, on Gaussian distributions. One such application, which will be dealt with in a forthcoming publication, is the subsequent decorrelation of the Box-Cox transformed model parameters, which may open up the possibility to define statistically independent variables.
Many steps in statistical data analysis are simplified or improve in accuracy when working with Gaussian distributions, so that one can potentially benefit from Box-Cox transformations in a wide range of problems. For example, \citet{taylor10} have developed an analytical marginalisation technique that works on Gaussian subspaces of the posterior. Using Box-Cox transformations, one can transform sub-spaces of or the complete posterior to a multivariate Gaussian, and moreover assess the non-Gaussianity of a given model parameter to verify whether a transformation is required.
Monte-Carlo Markov Chain (MCMC) methods sample a posterior considerably more efficiently if the distribution is compact and does not feature low-probability tails along degeneracy directions. As an example consider the cosmic microwave background (CMB) likelihood analysis by \citet{tegmark04} who employed the \lq natural\rq\ parametrisation suggested by \citet{kosowsky02}, followed by the diagonalisation of the parameter covariance matrix. Analogously, one could use an intial coarse MCMC sample to determine Box-Cox transformations that render the posterior approximately Gaussian. After an additional decorrelation of parameter space the detailed MCMC analysis could be run with high efficiency on a set of model parameters which are statistically independent and Gaussian distributed to good accuracy. This ansatz is applicable to any kind of likelihood analysis and does not require the existence of a physically motivated set of natural parameters, as in the case of the CMB.
Note furthermore that logarithmic transformations, which constitute a special case of Box-Cox transformations, of the large-scale matter distribution or the weak lensing convergence have recently been shown to enhance the information content of two-point statistics \citep{neyrinck09,seo11}. The potential of the more general Box-Cox transformations in this case is currently under investigation.
\section*{Acknowledgments}
We would like to thank Martin Hendry and Peter Schneider for fruitful discussions. Moreover we acknowledge the help of Martin Kilbinger and Eric Tittley in setting up CosmoPMC. We are grateful to Martin Kilbinger for making CosmoPMC publicly available. BJ acknowledges support by the European DUEL network, project MRTN-CT-2006-036133, and a UK Space Agency Euclid grant.
\bibliographystyle{mn2e}
|
\section{Introduction}
The attempts to clarify the mathematical framework underlying quantum chemistry, solid state physics, quantum field theories (QFT) and connected topics such as -for example, the grounds for the functional approaches, the principles underlying renormalization- as well as attempts to deepen our current understanding of widely used techniques (effective Hamiltonians, adiabatic limits...) are often hindered by very basic questions regarding the underlying mathematical methods. This is particularly obvious when it comes to developing new tools, see e.g. our studies of enhanced algorithms and formulas for the computation of eigenstates of Hamiltonians when the (unperturbed) ground state is degenerate
\cite{BP-09,BMP-10,BDPZ10}, or of the combinatorics of one-particle-irreducibility for interacting systems \cite{BP09}.
The present article focusses on diagrammatics. We argue that Feynman-type diagrammatics belong mathematically to the theory of linear forms on combinatorial Hopf algebras, which allows to generalize the theory to a much wider setting than the classical one.
We cover the usual examples corresponding to vacuum expectations over commutative products of fields where the propagators are represented by edges -this includes for example the various Goldstone-type diagrams and the perturbative expansions parametrized by Feynman diagrams in quantum field theories. However we go beyond and cover the case of expectations over general states, which requires the introduction of generalized diagrams \cite{Djah,BP09}. The same combinatorics happens to provide a pictorial description of cumulants in probability. The Hopf algebraic approach also allows to study expectations of products of free fields (Wightman fields) and derive new Feynman diagram expansions in this setting.
In a second part, we prove general linked cluster theorems, covering all
these cases and making as explicit as possible the link between the
elementary algebra underlying the theorems and the graphical content
(that relies on connectedness in the graph-theoretical sense). Notice
that we avoid deliberately functional methods (see e.g.\cite{Itzykson}):
although very efficient to derive the classical QFT linked cluster
theorem (using a generating function in terms of an external source),
they lack the generality and simplicity of the combinatorial proof.
Moreover, they cannot deal with noncommutative Feynman diagrams
because functional derivatives commute.
In the process, we promote another idea. Namely, linked cluster-type theorems rely ultimately on M\"obius inversion on the partition lattice.
\section*{Acknowledgements} We are particularly grateful to R. Stora. Many letters and documents (among which \cite{Stora08}) he sent us were the initial incentive for the present article -that was conceived to bridge the computations in \cite{BP09} with other approaches to quantum field computations and find some suitable mathematical framework for the (much more advanced) problems these documents suggest. In particular, we aimed at developping a mathematical framework to deal with products of Wightman fields and their average values over general states, one of the problems this article addresses.
\section*{Notation} We use the following convention: since various products will be defined along the article (such as $\ast$ or $\odot$), when we want to emphasize what product is used in an exponential, a logarithm or any other operation, we put the product symbol in exponent, so that $\log^\ast$ emans that we compute the logarithm using the $\ast$ product, $x^{\odot n}$ means that we compute the $n$-th power of $x$ using the $\odot$-product, and so on.
\section{Free combinatorial Hopf algebras}
The objects we will be interested in are combinatorial Hopf algebras in the sense of Joni and Rota \cite{Joni}, that is, bialgebras which coproduct is of combinatorial nature (obtained by ``splitting'' generating symbols according to combinatorial rules encoded by remarkable ``section coefficients'' --in high-energy physics, these coefficients correspond roughly to the symmetry factors of Feynman graphs). More specifically, we will be interested in families of Hopf algebras corresponding physically to bosonic or fermionic systems, to the usual algebraic structure of quantum fields (equipped with a commutative product such as the normal or time-ordered product) and to Wightman fields.
Since our results are more general than what would be required by applications to quantum systems, we state them in full generality and will show later how they specialize to particular physical systems or mathematical problems.
Let $X$ be a fixed ordered set, $X=\{x_1,...,x_n,...\}$. In most applications, $X$ will be infinite and countable, so that the reader may think to $X$ as the set of natural numbers.
The notion of combinatorial Hopf algebra, goes back to \cite{Joni}. The general notion is ill-defined in the litterature (there are many natural candidates, but at the moment no convincing general definition). We choose here a simple and relatively straightforward definition suited for our purposes that reflects some of the natural properties one expects from the notion when the underlying algebra is free commutative or free associative.
\begin{dfn}
We call free commutative (resp. free) \it combinatorial Hopf algebra \rm any connected graded commutative (resp. associative) and cocommutative Hopf algebra $H$ such that:
\begin{itemize}
\item (Freeness) As an algebra, $H$ is the algebra of polynomials (resp. of tensors) over a doubly indexed (finite or countable) set of formal, commuting (resp. noncommuting), variables $\phi_{i}(x_S)$ where $S$ runs over finite subsets of $X$ and $i=1...n_k$, where $k=|S|$ and where the sequence of the $n_k$, $i\in \NN$ is a fixed sequence of integers.
\item (Equivariance) The structure maps are equivariant with respect to maps induced by substitutions in $X$. In other terms, any substitution (that is, any set automorphism) $\sigma$ induces a Hopf algebra automorphism of $H$ which action on the generators is defined by $\sigma(\phi_{i}(x_S)):=\phi_{i}(x_{\sigma(S)})$.
\end{itemize}
\end{dfn}
The $\phi_{i}(x_S)$ are called the \it generators \rm of $H$, the (commutative or associative) monomials in the $\phi_{i}(x_S)$ form a linear basis of $H$ and are therefore called the \it basis elements\rm .
We do not look for the outmost generality, and many of our constructions and definitions can be extended in a fairly straightforward way to more general systems. For example, one might consider free partially commutative Hopf algebras which generators $\phi_{i}(x_S)$ would satisfy partial commutation rules (e.g. all the $\phi_j(x_i)$ would commute for a fixed $i$, but without $\phi_j(x_i)$ and $\phi_k(x_l)$ commuting for $i\not= l$, and so on). The so-constructed Hopf algebras can make sense in various applications and inherit all the properties of free commutative or free combinatorial Hopf algebras that are required for the forthcoming reasonings, we refer e.g. to \cite{PR2002} for details on the subject.
Some further remarks are in order.
Recall first that, by the Leray theorem (see e.g. Prop 4 in \cite{Patras93}), any connected graded commutative Hopf algebra $H$ is free commutative, so that the assumption that $H$ is freely generated as a commutative algebra comes for free when one assumes that $H$ is connected graded commutative.
The key point that makes the Hopf algebra combinatorial (and, as we shall see, suited to Feynman-type graphical reasonings) is that we assume that a set of polynomial generators is fixed and behaves nicely with respect to substitutions in $X$.
The particular case where $n_k=0$ for $k>1$ corresponds to the classical situation where the only propagators showing up in diagrams are the ones that describe the free propagation of a particle.
Let us mention at last that a more pedantical definition of combinatorial Hopf algebras could be given in terms of vector species, following the approach to combinatorial Hopf algebraic structures in \cite{patreu04,PatrasSchocker,AguiarM}.
A very important consequence of the equivariance condition is the following.
\begin{lem} For any subset $S$ of $X$, let us write $H_S$ for the subalgebra of $H$ generated by the $\phi_i(x_T),\ T\subset S$. Then, $H_S$ is a Hopf subalgebra of $H$.
Moreover, any automorphism $\sigma$ of $X$ induces an isomorphism of Hopf algebras from $H_S$ to $H_T$, where $T:=\sigma(S)$.
\end{lem}
The second property is a direct consequence of the first one since $\sigma$ (by definition of the induced map) induces an isomorphism of algebras from $H_S$ to $H_T$.
The proof of the first assertion follows immediately from the equivariance condition. Indeed, notice that an element of $H$ or of $H\otimes H$ is invariant by any substitution that acts as the identity on a finite subset $S$ of $X$ if and only if it is a polynomial in the $\phi_i(x_T),\ T\subset S$ (or, in $H\otimes H$, a sum of tensor products of such polynomials). Now, since,
for $T\subset S$, $\phi_i(x_T)$ is invariant by any substitution that acts as the identity on $S$, the same property holds also true of $\Delta(\phi_i(x_T))$, from which the Lemma follows.
\section{Some remarkable Hopf algebras}
Our favorite examples in view of applications to quantum chemistry, QFT
and solid state physics are simple ones. However the generality chosen
(allowing for example $n_2\not= 0$) is natural to handle nonlocal
interaction terms in the Lagrangians (think for example of the quantum
chemistry approach with Coulomb interaction). Our general approach also paves the way to a unification of QFT techniques (Feynman-type diagrammatics), umbral calculus (duality and linear forms on polynomial algebras) and combinatorics of (possibly ordered) set partitions.
For notational simplicity, we treat only commutative algebras in the usual sense, that is we do not treat explicitly the fermionic case (Grassmann or exterior algebras/ Fermi statistics).
However, as it is well known, there are no difficulties in switching from a bosonic to a fermionic framework -it just requires adding the right signs in the formulas, see \cite{Cassam,BrouderQG}, but handling simultaneously the commutative and anticommutative case would have required introducing consistently signs in all our formulas. For notational simplicity, we decided to stick to the bosonic, commutative, case, and to let the interested reader adapt our results to the anticommutative setting.
We use the langage of particle physics and quantum chemistry (so that the bosonic algebra coincides with the algebra of polynomials).
\begin{dfn} [Bosonic algebra]
We write ${\mathcal B}_k^X$ for the algebra of polynomials over the set of (formal, commuting) variables $$\phi_1(x_1),...,\phi_1(x_n),...;...;\phi_k(x_1),...,\phi_k(x_n),...$$
where $x_i\in X$.
These algebras are naturally equipped with a coproduct $\Delta: \BB\longmapsto\BB\otimes \BB$ that makes them Hopf algebras (that is, $\Delta$ is a map of algebras). The coproduct is defined on the generators $\phi_l(x_i)$ by requiring them to be primitive, that is: $$\Delta (\phi_l(x_i)):=\phi_l(x_i)\otimes 1+1\otimes \phi_l(x_i)$$ and extended multiplicatively to $\PP$ ($\Delta(xy)=\Delta(x)\Delta(y)$).
\end{dfn}
The lower index $k$ should be thought of as the number of quantum fields showing up (in a very broad sense), whereas the $x_i$ should be thought of as points, momenta or more generally dummy integration variables. For later use, we allow $k=\infty$ (to encode the countable set of eigenstates of a given many-body Hamiltonian) and will write simply $\cal B$ for ${\cal B}_k^X$ when no confusion can arise.
\begin{dfn} [Coulomb algebra]
We write ${\mathcal P}_k^X$ for the algebra of polynomials over the set of (formal, commuting) variables $\phi_l(x_i),\ l\leq k, \ \phi(x_{i,j})$, where $i\not= j\in X$.
The coproduct is defined by requiring the generators to be primitive.
\end{dfn}
This algebra is used in non-relativistic many-body physics
and quantum chemistry. The Coulomb interaction desribes
the force between a charge at point $x_i$ and
a charge at point $x_j$. These two points are linked by the
interaction, and a specific variable $\phi(x_{i,j})$ is
used to describe this connection.
\begin{dfn} [Tensor algebra]
We write ${\FF}_k^X$ for the algebra of noncommutative polynomials over the set of variables $$\psi_1(x_1),...,\psi_1(x_n),...;...;\psi_k(x_1),...,\psi_k(x_n),...$$
$x_i\in X$.
The coproduct is defined by requiring the generators to be primitive.
\end{dfn}
Various other free combinatorial Hopf algebras possibly noncommutative but fitting in the general framework of the present article are described (or follow from the results) in \cite{PatrasSchocker,PatrasSchocker2}. Let us quote the Malvenuto-Reuntenauer Hopf algebra and various Hopf algebras of tree-like structures. Although the decorated version of the Connes-Kreimer Hopf algebra of trees in \cite{PatrasSchocker} may have a use for QFT, a more promissing path in that direction is certainly provided by \cite{Gurau}, the belief of which we do share : ``ultimately we think that combinatorics is the right approach to QFT and
that a QFT should be thought of as the generating functional of a certain
weighted species in the sense of \cite{Leroux}''. We restrict however the examples in the present article to well-established domains of QFT and leave further extensions to future work.
\section{Graphication}
Let us start by recalling the construction underlying diagrammatic expansions and show how it can be extended naturally in a noncommutative/nonlocal setting. We call this process ``graphication'', by analogy with the ``arborification'' process underlying tree expansions in analysis and dynamical systems \cite{Ecalle92,Menous07}. In all the article, as a tribute to the physical motivations the ground field is the field of complex numbers (although the results hold over an arbitrary field of characteristic zero).
The reduced coproduct $\overline\Delta : H\longmapsto H\otimes H$ is defined by $\forall x\in H, \overline\Delta(x):=\Delta(x)-x\otimes 1-1\otimes x$. The coproduct and reduced coproduct are coassociative in the sense that $(\Delta\otimes H)\circ \Delta=(H\otimes \Delta)\circ \Delta$ (and similarly for $\overline\Delta$). The iterated coproduct and reduced coproduct maps from $H$ to $H^{\otimes i}$ are therefore well-defined and written $\Delta^{[i]}$, resp. $\overline\Delta^{[i]}$.
We assumed in the definition of combinatorial Hopf algebras that the coproduct is cocommutative: with Sweedler's notation $\Delta (x)=x^{(1)}\otimes x^{(2)}$, this means that $x^{(1)}\otimes x^{(2)}=x^{(2)}\otimes x^{(1)}$. Many of our forthcoming results could be adapted to the noncocommutative setting. However, this hypothesis is particularly usefulf when it comes to graphical encodings.
\begin{dfn}
The graphication map ${\mathcal G}$ is the map from $H$ to $\bigoplus\limits_n(H^{\otimes n})^{sym}\subset \bigoplus\limits_n(H^{\otimes n})$ defined by:
$${\mathcal G}:=\sum\limits_n\frac{\overline\Delta^{[n]}}{n!}.$$
\end{dfn}
Here, the superscript $sym$ in $\bigoplus\limits_n(H^{\otimes n})^{sym}$ means that the elements in the image of ${\mathcal G}$ are sums of symmetric tensor powers of elements of $H$. We will use the canonical isomorphism from covariants $\bigoplus\limits_n(H^{\otimes n})_{sym}:=\bigoplus\limits_nH^{\otimes n}/S_n$ (where $S_n$ stands for the symmetric group of order $n$) to invariants
$$[x_1| ...| x_n]\longmapsto\frac{1}{n!}\sum\limits_{\sigma\in S_n} x_{\sigma(1)}\otimes ...\otimes x_{\sigma(n)}$$ to represent invariants using the bar notation (that is, $y_1\otimes ...\otimes y_n$ in $H^{\otimes n}_{sym}$ is written $[y_1|...|y_n]$). Notice that, by definition (since we deal with covariants), for any permutation $\sigma$, $[y_1|...|y_n]=[y_{\sigma(1)}|...|y_{\sigma(n)}]$.
For example, in the bosonic algebra, abbreviating $\phi_1$ to $\phi$ (a notation we will use without further comments from now on):
$$\GG(\phi(x_1)\phi(x_2)...\phi(x_n))=\sum\limits_{\mathcal I}[\prod_{i\in {I_1}}\phi(x_{i})|...|\prod_{i\in {I_k}}\phi(x_{i})]$$
where ${\mathcal I}$ runs over all partitions $I_1\coprod...\coprod I_k$, $k=1...n$ of $[n]:=\{1,...,n\}$, with $\inf\{i\in I_j\}<\inf\{i\in I_{j+1}\}.$
Or
(isolating some components in the expansion):
$$\GG(\phi(x_1)^4\phi(x_2)^4)=[\phi(x_1)^4\phi(x_2)^4]+ 4[\phi(x_1)|\phi(x_1)^3\phi(x_2)^4]+...$$
$$+ 18[\phi(x_1)^2\phi(x_2)^2|\phi(x_1)^2\phi(x_2)^2]+...$$
$$\GG(\phi_1(x_1)\phi_2(x_1)\phi_3(x_2)^2)=[\phi_1(x_1)\phi_2(x_1)\phi_3(x_2)^2]+...$$
$$+[\phi_1(x_1)|\phi_2(x_1)\phi_3(x_2)^2]+2[\phi_1(x_1)\phi_3(x_2)|\phi_2(x_1)\phi_3(x_1)]+...$$
The same formulas hold in the tensor algebra and in the Coulomb algebras. Notice however that, in the tensor algebra, products are noncommutative, so that the order of the products in the monomials does matter. For example, with self-explanatory notation:
$$[\psi_1(x_1)|\psi_2(x_1)\psi_3(x_2)^2]\not=[\psi_1(x_1)|\psi_3(x_2)\psi_2(x_1)\psi_3(x_2)].$$
We call brackettings the terms such as $[\phi(x_1)^2\phi(x_2)^2|\phi(x_1)^2\phi(x_2)^2]$. The \it length \rm of a bracketting is the number of vertical bars $|$ plus 1. The \it support \rm of a bracketting $\Gamma$ is the set of the $x_i$ showing up in $\Gamma$. For example, $$sup([\phi_3(x_1)\phi_4(x_2)^2|\phi_1(x_1)^2\phi_2(x_8)^2])=\{x_1,x_2,x_8\}.$$
For each $x_i\in sup(\Gamma )$, we write $d_i$ for the total degree of the $\phi_j(x_i)$s in $\Gamma$ (so that for $\Gamma$ as above, $d_1=3,\ d_2=2, \ d_8=2$). For later use, we also introduce the product of two brackettings, which is simply the concatenation product:
$$[u_1|...|u_n]\cdot[v_1|...|v_n]:=[u_1|...|u_n|v_1|...|v_n].$$
These ideas and notation extend in a self-explanatory way to arbitrary combinatorial Hopf algebras. The only change regards the definition of the support and degree: $x_i$ is included in the support of a bracketting whenever there exists a pair $(j,S)$ such that $i\in S$ and $\phi_j(x_S)$ shows up in $\Gamma$ so that, for example,
$$sup([\phi_3(x_1)\phi_4(x_2)^2|\phi_1(x_1)^2\phi_2(x_8)^2|\phi_5(x_{1,5,8,10})])=\{x_1,x_2,x_5,x_8,x_{10}\}.$$
Similarly, the degree $d_8$ of $x_8$ in this bracketting accounts for the $\phi_5(x_{1,5,8,10})$ term and is therefore $d_8=3$.
The coefficients in the right hand side of the equation defining the
graphication will be referred to as symmetry factors. They are closely
related to the structure coefficients for the coproduct (in the basis
provided by monomials in the chosen family of generators, e.g. the
$\phi_i(x_j)$ for the bosonic algebra -these coefficients were called
section coefficients by Joni and Rota \cite{Joni}) but encode also the
symmetries showing up in the coproducts. Whereas, for the bosonic and
Coulomb algebras, symmetry factors are defined without ambiguity, for
other algebras (the tensor algebra or general free combinatorial Hopf
algebras) a given bracketting may appear in the expansion of $\GG(M)$
for various monomials $M$ in the generators (for example, in the tensor
algebra, $[\psi(x_1)|\psi(x_2)]$ appears in the expansion of
$\GG(\psi(x_1)\psi(x_2))$ and of $\GG(\psi(x_2)\psi(x_1))$). In general,
we will therefore write $s_\Gamma^M$ for the symmetry factor of a
bracketting $\Gamma$ in the expansion of $\GG(M)$ and will write simply
$s_\Gamma$ in the particular case of the bosonic and Coulomb algebras (where a unique $M$ exists giving rise to such a factor).
\section{Graphical representation}
Perturbative expansions in particle and solid-state physics are
conveniently represented by various families of diagrams. Feynman
diagrams are the most popular ones, but there are plenty of other
families with construction rules often slightly different from the one
underlying Feynman diagrams. Just to mention one interesting feature,
Feynman diagrams are usually independent of the time-coordinate of the
vertices (this is because of the definition of the so-called Feynman
propagators), whereas other families showing up in solid-state physics take into account causality systematically to construct their diagrams. These ideas are particularly well-explained in \cite{Mattuck}, to which we refer, also for a comprehensive treatment of the zoology of diagrammatic expansions.
Defining a graphical representation associated to a free combinatorial
Hopf algebra depends highly on the structure and particular features of
the algebra. We define various such representations, by increasing order
of complexity, focussing only on the algebras (tensor, bosonic, Coulomb)
we have chosen to investigate in depth. The reader who needs to
construct other taylor-made graphical representations will be able to do
so easily using our recipes (for notational simplicity and to avoid
pointless pedantry, we omit to introduce the most general possible
definitions since the process of designing them is straightforward once the leading principles are understood on some examples).
Let us mention that, when two brackettings $U=[u_1|...|u_n]$ and $V=[v_1|...|v_n]$ have disjoint supports, their product corresponds to the disjoint union of the corresponding graphs (this is the usual product on Hopf algebras of Feynman graphs in QFT, see e.g. \cite{CKI}).
\subsection{Commutative local case: bipartite graphs}
\begin{figure}
\begin{center}
\includegraphics[width=8.0cm]{mobius1.eps}
\caption{Graph of the bracketting
$[\phi_1(x_1)\phi_1(x_2)^2|\phi_2(x_1)^2\phi_1(x_2)^2\phi_2(x_3)|
\phi_3(x_1)\phi_3(x_2)\phi_3(x_3)]$.
\label{fig1}}
\end{center}
\end{figure}
We focus in this section on the bosonic algebra.
Each Feynman bracket (say
$\Gamma=[\phi_1(x_1)\phi_1(x_2)^2|\phi_2(x_1)^2\phi_1(x_2)^2\phi_2(x_3)|
\phi_3(x_1)\phi_3(x_2)\phi_3(x_3)]$, see
fig.~\ref{fig1}) can be represented uniquely by a bipartite (non planar) graph with unoriented colored edges (i.e. by a graph with 2-coloured vertices and colored edges) according to the following rule:
\begin{enumerate}
\item For each $x_i\in sup(\Gamma )$, recall that we write $d_i$ for
the total degree of the $\phi_j(x_i)$s in $\Gamma$. Draw a
$x_i$-labelled black vertex with $d_i$ outgoing colored edges
(the colors being attributed according to the indices $j$ of
the $\phi_j(x_i)$, e.g. a 4-edge black vertex for $x_1$ with one 1-colored edge
(solid line), two 2-colored edges (dotted line) and a 3-colored edge (dashed line).
\item Running recursively from the left to the right of $\Gamma$,
for each term inside brackets and bars (e.g. $\phi_1(x_1)\phi_1(x_2)^2$,
then $\phi_2(x_1)^2\phi_1(x_2)^2\phi_2(x_3)$, then $\phi_3(x_1)\phi_3(x_2)\phi_3(x_3)$),
select randomly according to the colors
and powers showing up in the monomials outgoing edges of the corresponding vertices
(e.g. select one 1-colored edge from the $x_1$ vertex and two 1-colored edges
from the $x_2$ vertex). Connect these edges to a new white vertex
(e.g. a new white vertex with 3 outgoing colored edges).
\end{enumerate}
These edge-colored bipartite graphs will be called from now \it interaction graphs\rm . See \cite{Djah,BP09} for applications.
Some simplifications are possible in many cases of interest. For example, when the $x_i$s are dummy integration variables and can be exchanged freely in any computation (e.g. of scattering amplitudes), the labelling of the black vertices can be omitted.
Another classical situation (encountered with scalar quantum field theories such as $\phi^3$ or $\phi^4$ and in statistical physics) is the one where $k=1$. In that case, there is just one possible color for the edges, so that the coloring of the edges can safely be omitted in the definition of the graphs. In QFT, when several fields coexist (e.g. in QED, where $k=2$), instead of using colors,
practitioners use often different representations for the edges that depend on the fields involved (typically, in QED, edges associated to electrons are plain lines, whereas edges associated to photon propagation are represented by a succession of small waves). Of course, the use of colors or the use of different shapes for the edges are strictly equivalent and a matter of taste and habits.
Another classical simplification occurs when the only brackettings of interest for practical applications are the ones where the only monomials showing up in the brackettings are products of degree 2 of the form $\phi_i(x_j)\phi_i(x_k)$. In the corresponding graphs, the white vertices always have two outgoing edges with the same color, so that these vertices can be erased -what remains is a graph with only black vertices and colored edges (that correspond, physically, to particle types). These are the celebrated Feynman graphs that one encounters in QFT textbooks.
From now on we will identify systematically brackettings and the corresponding interaction graphs.
\subsection{Commutative non local case: tripartite graphs}
By nonlocal case, we mean that some $n_i,\ i>1$ may be different from zero. The canonical example we have in mind is the one of QED in the solid state picture, that is with instantaneous Coulombian interactions (in the particle physics picture the interactions are local and encoded by products of fields in the Lagrangian; this corresponds to the commutative local case).
For simplicity (and in view of the most natural applications), we assume that the only brackettings of interest are those in which the monomials showing up are either monomials in the $\phi_i(x_j)$, either a $\phi_i(x_S), \ |S|>1$ (in other terms, no nontrivial products involving a $\phi_i(x_S)$ should appear in the bracketting).
Each Feynman bracket (say for example $$\Gamma=[\phi_1(x_1)\phi_1(x_2)^2|\phi_2(x_1)^2\phi_1(x_2)^2\phi_2(x_3)|\phi_3(x_{1,2,3})])$$ can be represented uniquely by a tripartite (non planar) graph with two kinds of unoriented edges according to the following rule (see Figure 2):
\begin{figure}
\begin{center}
\includegraphics[width=8.0cm]{mobius2.eps}
\caption{Graph of the bracketting
$[\phi_1(x_1)\phi_1(x_2)^2|\phi_2(x_1)^2\phi_1(x_2)^2\phi_2(x_3)|
\phi_3(x_{1,2,3})]$.
\label{fig2}}
\end{center}
\end{figure}
\begin{enumerate}
\item We distinguish between the two kind of edges in the following way: the first type of edge is drawn as colored lines (the color is an index $i$ with $d_i\not=0$); the second type has no color and is drawn as a sequence of waves (as the photon propagators in QED). We call these edges respectively plain and wavy edges.
\item For each $x_i\in sup(\Gamma )$, draw a $x_i$-labelled black vertex with $d_i$ outgoing edges. Colors are attributed to the plain edges according to the indices $j$ of the $\phi_j(x_i)$, the other edges correspond to $\phi_j(x_S)$s and are wavy edges. For example, we draw a 4-edges black vertex for $x_1$ with one 1-colored edge (represented by a solid line in Fig. 2), two 2-colored edges (dotted lines) and a wavy edge.
\item Running recursively from the left to the right of $\Gamma$, for each term inside brackets and bars (e.g. $\phi_1(x_1)\phi_1(x_2)^2$, then $\phi_2(x_1)^2\phi_1(x_2)^2$, then $\phi_3(x_{1,2,3})$), \begin{itemize}
\item If encountering a monomial involving $\phi_i(x_j)$s, proceed as in the commutative local case:
select randomly according to the colors and powers showing up in the monomials outgoing edges of the corresponding vertices (e.g. select one 1-colored edge from the $x_1$ vertex and two 1-colored edges from the $x_2$ vertex). Connect these edges to a new white vertex (e.g. a new white vertex with 3 outgoing colored edges).
\item If encountering a $\phi_j(x_S)$, select randomly a wavy edge outgoing from the $x_i, \i\in S$ black vertex, connect all these edges to new $j$-labelled grey vertex.
\end{itemize}
\end{enumerate}
These tripartite graphs will be called from now
\emph{nonlocal interaction graphs}. The particular case of solid state physics QED enters in this framework. Moreover, the general case we consider seems new and of interest, since it should allow to treat QED computations in a complex background (e.g. with non trivial vacua). These concrete applications (that originated this work, together with questions and remarks by R. Stora) are left for future work.
Great simplifications occur in many simpler cases of interest. The first
situation encountered in pratice is the one where $n_1=n_2=1$. This
corresponds roughly to the case where one class of particles is present
(say electrons, up to a change from the bosonic to the fermionic
statistics) and where interactions are encoded by nonlocal terms
(Coulomb interaction lines). In that case, there is just one possible color for the edges, so that the coloring of the plain edges can be omitted in the definition of the graphs. Besides, since $n_2=1$, the grey vertices can also be erased, so that one ends up with bipartite graphs with two types of edges (plain and wavy). If one assumes further that the only graphs of interest are those with two outgoing edges, these edges can be also erased. One ends up with the familiar Feynman diagrams with black vertices only and two types of edges corresponding to electron propagators and Coulombian interactions.
\subsection{Noncommutative local case}
We focus in this section on the tensor algebra.
Each Feynman bracket (say $\Gamma=[\psi_1(x_2)\psi_1(x_1)\psi_1(x_2)|\psi_2(x_1)\psi_1(x_2)\psi_2(x_3)\psi_1(x_2)\psi_2(x_1)|
\psi_3(x_3)\psi_3(x_1)\psi_3(x_2)]$) can be represented uniquely by a bipartite (non planar) graph with unoriented colored and locally ordered edges (i.e. by a graph with 2-coloured vertices and colored, locally ordered edges) according to the following rule (the rule will make clear the meaning of ``local order'' that corresponds to an order on the edges reaching a white vertex)
(see fig.~\ref{fig3}):
\begin{figure}
\begin{center}
\includegraphics[width=8.0cm]{mobius3.eps}
\caption{Graph of the bracketting
$[\psi_1(x_2)\psi_1(x_1)\psi_1(x_2)|\psi_2(x_1)\psi_1(x_2)\psi_2(x_3)\psi_1(x_2)\psi_2(x_1)|\psi_3(x_3)\psi_3(x_1)\psi_3(x_2)]$.
\label{fig3}}
\end{center}
\end{figure}
\begin{enumerate}
\item Proceed first as in the commutative case: for each $x_i\in sup(\Gamma )$, recall that we write $d_i$ for the total degree of the $\phi_j(x_i)$s in $\Gamma$. Draw a $x_i$-labelled black vertex with $d_i$ outgoing colored edges (the colors being attributed according to the indices $j$ of the $\phi_j(x_i)$, e.g. a 4-edges black vertex for $x_1$ with one 1-colored edge (solid line), two 2-colored edges (dotted lines) and a 3-colored edge (dashed line).
\item Running recursively from the left to the right of $\Gamma$, for each term inside brackets and bars (e.g. $\psi_1(x_2)\psi_1(x_1)\psi_1(x_2)$, then $\psi_2(x_1)\psi_1(x_2)\psi_2(x_3)\psi_1(x_2)\psi_2(x_1)$...), select randomly according to the colors and powers showing up in the monomials outgoing edges of the corresponding vertices and order them according to the order of the appearance of the colors in the (noncommutative !) monomial (e.g. select a 2-colored edge from the $x_2$ vertex, label it $a$; a 1-colored edge from the $x_1$ vertex, label it $b$; a 2-colored edge from the $x_2$ vertex, label it $c$). Connect these edges to a new white vertex (e.g. a new white vertex with 3 outgoing colored and ordered edges, where the order is defined by the labels).
\end{enumerate}
These edge-colored and locally ordered bipartite graphs will be called from now \it free interaction graphs\rm .
The usual simplifications are possible in many cases of interest, we do not detail them and simply mention that, when the only monomials showing up in brackettings are of the form $\psi_i(x_m)\psi_i(x_l)$, the graphical rule amounts to consider ``classical'' Feynman graphs with labelled vertices and colored and directed propagators.
The noncommutative nonlocal case could be treated similarly by mixing the conventions for the noncommutative local and commutative local cases. The exercice is left to the reader.
\section{Symmetry factors and connectedness}{\label{symfac}}
This section is devoted to a technical but fundamental Lemma that connects the symmetry factors of graphs with the topological notion of connectedness. The lemma is the ground for the proof of linked cluster theorems and is particularly meaningful in the noncommutative case (Wightman fields).
We assume in this section that the Hopf algebra is an arbitrary free or free commutative combinatorial Hopf algebra which generators are primitive elements (this condition is satisfied by all the combinatorial Hopf algebras we have considered so far).
Let us introduce first a further notation. Let $x$ be a basis element (a commutative or noncommutative monomial in the generators) of a combinatorial Hopf algebra which support $S$ decomposes into a disjoint union $T\coprod V$ (the support is defined as for brackettings). We write $x_T$ and $x_V$ and call respectively $T$ and $V$-components of $x$ the two basis elements (possibly equal to zero) defined by $x_T\otimes x_V:= (P_T\otimes P_V)\circ \Delta (x)$, where $P_V$ stands for the projection on $H_V$ orthogonally to all the basis elements that do not belong to $H_V$.
The reader can check that this definition amounts to the following: to get $x_T$, replace, in the expansion of $x$ as a monomial, all the $\phi_i(x_K), K\subset V $ by a 1 and all the $\phi_i(x_K), K\cap V\not=\emptyset, K\cap T\not=\emptyset $ by a zero (and similarly for $x_V$).
\begin{lem}\label{keylemma} Let $\Psi$ be a basis element and $ \Gamma=\Gamma_1\cdot \Gamma_2$ a bracketting or, equivalently, an interaction graph (of any type) such that $sup(\Gamma_1)\cap sup(\Gamma_2)=\emptyset$ and $s_{\Gamma}^\Psi\not=0$ (recall that $\cdot$ stands for the product of brackettings). Topologically, this amounts to assume that $\Gamma$ decomposes as a disjoint union of graphs.
Then:
\begin{itemize}
\item The basis elements $\Psi_{sup(\Gamma_1)}$ and $\Psi_{sup(\Gamma_2)}$ are non zero.
\item Moreover: $s_\Gamma^\Psi=s_{\Gamma_1}^{\Psi_{sup(\Gamma_1)}}s_{\Gamma_2}^{\Psi_{sup(\Gamma_2)}}$
\end{itemize}
In the commutative case, this last identity can be abbreviated to $s_\Gamma=s_{\Gamma_1}s_{\Gamma_2}$.
\end{lem}
Indeed, since the $\phi_i(x_K)$ are primitive, the coproduct of a product, say $a_1...a_n$, of $\phi_i(x_K)$s is the sum of the tensor products $a_{i_1}...a_{i_k}\otimes a_{j_1}...a_{j_{n-k}}$, where $\{i_1,...,i_k\}$ is a (ordered) subset of $[n]$ and $\{j_1,...,j_{n-k}\}$ its (ordered) supplement. The hypothesis $s_{\Gamma}^\Psi\not=0$ ensures that $sup(\psi)= sup(\Gamma_1)\coprod sup(\Gamma_2)$, and that in $\psi$, there is no factor $\phi_i(x_K)$ with $ K\cap V\not=\emptyset$ and $ K\cap T\not=\emptyset $. The first assertion follows.
To prove the second identity, let us first notice that $s_\Gamma^\Psi=s_\Gamma^{\Psi_{sup(\Gamma_1)}\Psi_{sup(\Gamma_2)}}$. Indeed, let us use the same notation as previously and write $\Psi=a_1...a_n$. The coproduct and its iterations are constructed by extracting disjoint subsequences out of the ordered sequence of the $a_i$s. On the other hand, the basis elements showing up in $\Gamma_1$ and $\Gamma_2$ belong to disjoint sets -the relative ordering of the basis elements with support in $sup(\Gamma_1)$ and in $sup(\Gamma_2)$ in the expansion of $\Psi$ do therefore not matter, which proves the identity.
We can now assume without restriction (because of the cocommutativity) that $\Gamma=[S_1,...,S_k,T_1,...,T_l]$ with $\Gamma_1=[S_1,...,S_k]$ $\Gamma_2=[T_1,...,T_l]$.
Besides, since $\Delta$ is an algebra map, we have:
$$\Delta^{[k+l]}(\Psi_{sup(\Gamma_1)}\Psi_{sup(\Gamma_2)})=\Delta^{[k+l]}(\Psi_{sup(\Gamma_1)})\Delta^{[k+l]}(\Psi_{sup(\Gamma_2)}).$$
The multiplicity $\mu_\Gamma:=(k+l)!s_\Gamma^\Psi$ of $\Gamma$ in $\Delta^{[k+l]}(\Psi)$ is therefore obtained by summing the coefficients
of the tensor products $(X_{\sigma(1)},...,X_{\sigma(k+l)})$ in $\Delta^{[k+l]}(\Psi)$, where $\sigma$ runs over $S_{k+l}/Stab((X_1,...,X_{k+l}))$ and $(X_1,...,X_{k+l})=(S_1,...,S_k,T_1,...,T_l)$ (here, $Stab((X_1,...,X_{k+l}))$ stands for the stabilizer of $(X_1,...,X_{k+l})$ in $S_{k+l}$). However, since the coproduct is cocommutative, these coefficients are all equal and one can restrict the computation to the tensor products $(S_{\beta(1)},...,S_{\beta(k)},T_{\alpha(1)},...,T_{\alpha(l)})$, where $\alpha$ and $\beta$ run over permutations in $S_k/Stab((S_1,...,S_k))$ and $S_l/Stab((T_1,...,T_l))$, and multiply the result by the number of $k$-element subsets in $[k+l]=\{1,...,k+l\}$.
We get finally:
$$\mu_\Gamma={{k+l}\choose{k}}\mu_{\Gamma_1}\mu_{\Gamma_2},$$
that is, $s_\Gamma^\Psi=s_{\Gamma_1}^{\Psi_{sup(\Gamma_1)}}s_{\Gamma_2}^{\Psi_{sup(\Gamma_2)}}$ or, in words (and slightly abusively), ``the symmetry factor of an interaction graph is the product of the symmetry factors of its connected components''. Notice that the property holds true also in nonlocal and/or noncommutative cases.
\section{Amplitudes and Feynman rules}\label{machin}
A linear form on a combinatorial Hopf algebra is \it unital \rm if $\rho(1)=1$ and \it infinitesimal \rm if $\rho(1)=0$.
Let us recall, for example, how linear forms on $\mathcal B_k^X$, $X=\{x_1,...,x_n,...\}$, are usually constructed.
Let $T$ be an arbitrary finite sequence of integers.
For any polynomial in $k$ variables, say $P(y_1,...,y_k)=\sum\limits_{1\leq i_1+...+i_k\leq n}p_{i_1,...,i_k}y_1^{i_1}...y_k^{i_k}$, we write $P(T)$ for the polynomial $\prod\limits_{t\in T}P(\phi_1(x_t),...,\phi_k(x_t))\in{\mathcal B}_k^X$.
One can think of $P$ as the interacting part of a Lagrangian.
Natural forms should then be thought of as related physical amplitudes.
For example, for $k=1$, a typical $P$ is $\lambda\frac{y^4}{4!}$, corresponding to the $\phi^4$ theory (see \cite{KS} for details).
The Green functions of this theory are computed via the
formula
\begin{eqnarray*}
G(x_1,\dots,x_n) :=
\frac{\rho\Big(\phi(x_1)...\phi(x_k)e^{i\int \phi^4(x) dx}\Big)}
{\rho\Big(e^{i\int \phi^4(x) dx}\Big)},
\end{eqnarray*}
where $\rho$ denotes the vacuum expectation value
of the time-ordered product of fields
(see section~\ref{QFTsect}).
Let us treat now a radically different example to show the ubiquity of the approach. Let here the role of $P$ be taken by $H_I(t)$, the interacting Hamiltonian of time-dependent perturbation theory (see e.g. our \cite{BP-09,BMP-10}): $H_I(t):=e^{iH_0t}Ve^{-iH_0t}e^{-\epsilon |t|}$.
Let $e_1,...,e_n,...$ be the eigenvectors of $H_0$ with eigenvalues $\lambda_1<\lambda_2\leq ...\leq \lambda_n\leq ...$. We assume for simplicity that the ground state is non degenerate ($\lambda_1\not=\lambda_2$), although the following reasoning holds in full generality due to \cite{BP-09,BMP-10}. The computation of the ground state of the perturbed Hamiltonian $H_0+V$ relies on the computation of the quantities such as:
$Y=<e_1|H_I(t_1)...H(t_p)|e_1>$, where $t\geq t_1>...>t_p> -\infty$.
We set $k=\infty$ and $\phi_{2p}(t_j):=e^{-i\lambda_pt_j-\frac{\epsilon}2|t_j|}<e_p|$, $\phi_{2p+1}(t_j):=e^{+i\lambda_pt_j-\frac{\epsilon}2|t_j|}|e_p>$. Then, $H_I(t)=\sum_{i,j}V_{i,j}\phi_{2j+1}(t)\phi_{2i}(t)$, where $V_{i,j}:=<e_j|V|e_i>$.
The unital form corresponding to the computation of $Y$ is given simply by the form on ${\mathcal F}_\infty^{]-\infty ,t]}$:
$$\rho(\phi_{i_0}(t_0)...\phi_{i_{2k+1}}(t_{2k+1})):=\prod\limits_{0\leq j\leq k}V_{i_{2j+1},i_{2j}}(\phi_{i_{2j}}(t_{2j})|\phi_{i_{2j+1}}(t_{2j+1})),$$
where $(\phi_{i_{2j}}(t_{2j})|\phi_{i_{2j+1}}(t_{2j+1})):=\phi_{i_{2j}}(t_{2j})\phi_{i_{2j+1}}(t_{2j+1})$ if $i_{2j}$ is even and $i_{2j+1}$ odd and zero else.
The value of the form $\rho$ on odd products is $0$. Of course, this example is purely didactical and for such a computation the use of the formalism developped in the present article is largely pointless. It becomes useful when the situation gets more involved. Actually, the simple requirement of taking efficiently into account the divergences arising from the adiabatic expansion may involve advanced combinatorial techniques, see, besides the articles already quoted, our \cite{BDPZ10}.
It is well-known that, in many situations, Green functions such as the ones of the $\phi^4$ theory split into components parametrized by Feynman diagrams. This property also holds for more complex theories and is best explained through Hopf algebraic computations.
Recall first that, since $H$ is a Hopf algebra, the set $ H^\ast$ of
linear forms on $H$ is equipped with an (associative; commutative if $H$
is cocommutative) ``convolution'' product:
$$\forall \rho,\mu \in H^\ast, \rho\ast\mu (x):=\rho(x_{(1)})\mu(x_{(2)}),$$
where we used the Sweedler notation $\Delta (x)=x_{(1)}\otimes x_{(2)}$.
Notice that, if $\rho$ and $\mu$ are infinitesimal forms, $\rho\ast\mu (x):=\rho(x_{\{1\}})\mu(x_{\{2\}})$, where we use the notation $\overline\Delta (x)=x_{\{1\}}\otimes x_{\{2\}}$.
By standard graduation arguments, the convolution logarithm of a unital form $\rho$ is a well-defined infinitesimal form $\tau$ on $\mathcal P$. We extend such a $\tau$ to a linear form (still written $\tau$) on $\PP^{\otimes n}$ (resp. $\PP^{\otimes n}_{sym}$) by: $\tau (x_1\otimes ...\otimes x_n):=\tau(x_1)...\tau(x_n)$ or $\tau[x_1|...|x_n]:=\tau(x_1)...\tau(x_n)$.
Summing up, we get, for $X$ an arbitrary monomial (basis element) in $H$, and since $\tau$ is an infinitesimal form:
\begin{prop}[Feynman diagrams/rules expansion] For an arbitrary unital form on $H$, we have:
$$\rho(X)=\exp^{\ast \tau}(X)=\tau\circ\GG (X),$$
or:
$$\rho(X)=\sum\limits_\Gamma s_\Gamma^X\tau(\Gamma),$$
where $\Gamma$ runs over all the brackettings (or interaction graphs) in the image of $X$ by the graphication map.
\end{prop}
The map $\tau$ acting on the $\Gamma$s is called a Feynman rule. Applying Lemma~\ref{keylemma}, with $H$ as in Sect.~\ref{symfac} we get immediately (with self-explanatory notations):
\begin{lem}\label{syfactrule}
Assume that $\Gamma =\Gamma_1\cdot \Gamma_2$, then:
$$s_\Gamma^X\tau(\Gamma)=s_{\Gamma_1}^{X_1}\tau(\Gamma_1)s_{\Gamma_2}^{X_2}\tau(\Gamma_2).$$
\end{lem}
\section{The combinatorial linked cluster theorem}
The combinatorial linked cluster theorem expands a linear form on a combinatorial Hopf algebra into ``connected'' parts closely related to the topological notion of connectedness. In this section, we show that this expansion is a very general phenomenon related to M\"obius inversion in the partition lattice. Notations and conventions are in Sect.~\ref{symfac}.
We first recall some general facts on the partition lattice and M\"obius inversion that are familiar in combinatorics but probably not well-known by practitioners of Feynman-type diagrammatics and linked cluster theorems (the particular case of M\"obius inversion for the partition lattice we are interested in here seems due to Sch\"utzenberger, we refer to \cite{Comtet} for further details and references on the subject).
For an arbitrary set $S$, partitions $t:=\{T_1,...,T_k\}$ of $S$ (that is: $T_1\coprod ...\coprod T_k=S$, where $\coprod$ stands for the disjoint union) are organized into a poset (partially ordered set, this poset is actually a lattice -two elements have a $max$ and a $min$, this follows easily from the definition of the order). We write $|t|$ for the length of the partition (so that $|t|=k$) and abbreviate the partitions of minimal and maximal length, respectively $\{S\}$ and $\{\{s\},\ s\in S\}$ to $\hat 1$ and $\hat 0$.
The subsets $T_i$ are called the blocks, and the order is defined by
refinement: for any partitions $t$ and $u$, $t\leq u$ if and only if each block of ${t}$ is contained in a block of $u$.
The functions $f(x,y)$ on the partition lattice such that $f(x,y)\not=0$ only if $x\leq y$ form the \it incidence algebra \rm of the lattice.
The (associative) product is defined by convolution: $(f\ast g)(x,y):=\sum\limits_{x\leq z\leq y}f(x,z)g(z,y)$.
The identity of the algebra is Kronecker delta function: $\delta(x,y):=1$ if $x=y$ and $:=0$ else.
The zeta function $\zeta(x,y)$ of the lattice
is defined to be equal to 1 if $x\leq y$ and 0 otherwise.
The
M\"obius function $\mu(x,y)$ is defined to be the inverse of the zeta function for the convolution
product. It can be computed explicitly: for $x\leq t$, where $t=\{T_1,...,T_k\}$, we have:
$$\mu(x,t)=(-1)^{|x|+|t|}(n_1-1)!...(n_k-1)!,$$
where $n_i$ is the number of blocks of $x$ contained in $T_i$.
Using the identities $\mu\ast \zeta(\hat 0,\hat 1)=\zeta\ast\mu (\hat 0,\hat 1)=\delta (\hat 0,\hat 1)=0$, we recover in particular the useful combinatorial formulas:
$$\sum\limits_{0\leq k\leq |S|}\sum\limits_t(-1)^{|t|+|S|}(t_1-1)!...(t_k-1)!=0$$
where $t$ runs over the partitions of length $k$ of $S$ and $t_i$ stands for the number of elements in the $i$-th block $T_i$ of $t$ and, with the same conventions on $t$,
\begin{equation}\label{eqq}
\sum\limits_{0\leq k\leq |S|}\sum\limits_t(-1)^{|t|+1}(|t|-1)!=0
\end{equation}
The key application of these notions is to inclusion/exclusion computations in the partition lattice. Namely, for an arbitrary function $h(x)$ on the lattice, let us set: $h(y)=\sum\limits_{x\leq y}\widetilde{h}(x)$. This formula defines uniquely $\widetilde{h}$ and, in the convolution algebra:
$$(h=\widetilde{h}\ast \zeta )\Leftrightarrow (\widetilde{h}=h\ast \mu )$$
so that $\widetilde{h}(y)=\sum\limits_{x\leq y}h(x)\mu(x,y)$.
Let now $x$ be a basis element of $H$ with support written $S$, where $H$ is as in Sect.~\ref{symfac}. Let $\rho$ be a unital form on $H$. Recall the notation $x_T$ denoting the ``T-component'' of $x$ for an arbitrary $T\subset S$. For an arbitrary set partition $t=\{T_1,...,T_k\}$ of $S$, we extend $\rho$ to a function $\rho^x$ on partitions of $S$ and set:
$$\rho^x(t):=\rho(x_{T_1})...\rho(x_{T_k}).$$
Recall also the decomposition $\rho(x)=\sum_\Gamma s_\Gamma^x\tau(\Gamma)$. The $\Gamma$s are represented by diagrams, among which some are connected. We set $\rho_{conn}(x):=\sum_{\Gamma_c} s_{\Gamma_c}^x\tau(\Gamma_c)$, where the $\Gamma_c$ run over \it connected \rm diagrams.
We can then apply the machinery of inclusion/exclusion to $\rho^x$ and define $\widetilde\rho^x$. The combinatorial linked cluster theorem relates $\rho_{conn}$ and $\widetilde\rho^x$:
\begin{thm}[Combinatorial linked cluster theorem]
We have, for an arbitrary basis element in a free or free commutative combinatorial Hopf algebra which generators are primitive elements:
$$\rho_{conn}(x)=\widetilde\rho^x(\hat 1).$$
\end{thm}
We have indeed:
$$\widetilde\rho^x(\hat 1)=\sum\limits_{t}\rho^x(t)\mu(t,\hat 1)$$
where $t$ runs over the partitions of $S$ and (with the usual notations) $\mu(t,\hat 1)=(-1)^{|t|+1}(|t|-1)!$.
On the other hand, $\rho^x(t)=\rho(x_{T_1})...\rho(x_{T_k})$ and $\rho(x_{T_i})=\sum\limits_{\Gamma_i}s_{\Gamma_i}^{x_{T_i}}\tau(\Gamma_i)$. Lemma~\ref{syfactrule} ensures that $s_{\Gamma_1\cdot ...\cdot\Gamma_k}^x=s_{\Gamma_1}^{x_{T_1}}...s_{\Gamma_k}^{x_{T_k}}$, and we get finally:
$$\widetilde\rho^x(\hat 1)=\sum\limits_{T_1\coprod ...\coprod T_k=S}(-1)^{k+1}(k-1)!\sum\limits_{\Gamma_1,...,\Gamma_k}s_{\Gamma_1\cdot ...\cdot\Gamma_k}^x \tau(\Gamma_1)\cdot ...\cdot \tau(\Gamma_k),$$
where $\Gamma_i$ is a graph showing up in the expansion of $\rho(x_{T_1})$.
Let now $\Psi$ be an arbitrary graph with $s_\Psi^x\not=0$. The graph decomposes uniquely as a union of topologically disjoint graphs $\Psi_1\coprod...\coprod \Psi_n$, where $n$ is the number of connected components of $\Psi$. We write $S_i$ for $sup(\Psi_i)$ and $S_\Psi=\{S_1,...,S_n\}$.
We have to show that the coefficient of $\tau(\Psi)$ in the right hand side of the previous equation is equal to $s_\Gamma^x$ if $n=1$ and to 0 else.
The first property is immediate, since if $\Gamma$ is connected it appears only in the term associated to the trivial partition $\hat 1$ of $S$ and therefore with the coefficient $(-1)^20!s_\Gamma^x=s_\Gamma^x$.
The second property is slightly less immediate, but follows from the
principles of M\"obius inversion together with Lemma~\ref{syfactrule}.
Notice first that $\Gamma$ appears in the right hand side of the
equation in association to all partitions of which $S_\Psi$ is a refinement.
The coefficient of $\Gamma$ is therefore:
$$[\sum\limits_{s_\Psi\leq t\leq \hat 1}(-1)^{|t|+1}(|t|-1)!]s_\Gamma^x$$
which is zero as a consequence of the identity $\zeta\ast\mu (S_\Psi,\hat 1)=\delta(S_\Psi,\hat 1)=0$ in the partition lattice.
\section{The functional linked cluster theorem}
A linear form $\rho$ on a free or free commutative combinatorial Hopf algebra $H$ is called \it symmetric \rm if it is invariant by a bijective relabelling of the variables $x_i$ -so that $\rho(\phi_1(x_2)^8\phi_3(x_5)^2\phi_2(x_9)^3)=\rho(\phi_1(x_4)^8\phi_3(x_2)^2\phi_2(x_1)^3)$, and so on.
When $X$ is ordered, the form is called \it quasi-symmetric \rm if it is invariant by a (strictly) increasing relabelling of the variables, so that e.g. $\rho(\phi_1(x_2)^8\phi_3(x_5)^2\phi_2(x_9)^3)=\rho(\phi_1(x_4)^8\phi_3(x_5)^2\phi_2(x_8)^3)$, but is not necessarily equal to $\rho(\phi_1(x_4)^8\phi_3(x_2)^2\phi_2(x_1)^3)$.
Let us consider now a symmetric or quasi-symmetric unital form $\rho$ on $H$, where $H$ is as in Sect.~\ref{symfac} with an \it infinite \rm ordered index set $X$. For notational simplicity, we will assume that $X=\NN$. That is,
$H$ is an algebra of polynomials (resp. of tensors) over a doubly indexed set of formal, commuting (resp. noncommuting), and primitive variables $\phi_{i}(x_S)$ where $S$ runs over finite subsets of $X$ and $i=1...n_k$, where $k=|S|$ and where the sequence of the $n_k$, $k\in \NN$ is a fixed sequence of integers.
We first generalize the construction of the ``interaction term'' $P$ in Sect.~\ref{machin} as follows. We let ${\mathcal P}= P(T)_{ T\subset X}$ be a family of elements of $H$ such that $P(T)$ is a polynomial (resp. a tensor) in the $\phi_{i}(x_S),\ S\subset T$.
\begin{dfn} We say that ${\mathcal P}$ is \it admissible \rm if and only if
\begin{enumerate}
\item For any order-preserving bijection $\phi$ from $T$ to $R$, $\phi(P(T))=P(R)$.
\item For any $T$ and any partition $U\coprod V=T$ (where $U$ and $V$ inherit the natural order on $T$), we have:
$$\sum_b\mu_b(P_b)_{U}\otimes (P_b)_{V}=P(U)\otimes P(V),$$
where $P(T)=\sum_b\mu_bP_b$ is the unique decomposition of $P(T)$ as a linear combination of basis elements.
\end{enumerate}
\end{dfn}
We let the reader check that the map $P$ constructed in Sect.~\ref{machin} satisfies this requirement.
The composition of $\rho$ with $P$ is a ``scalar species'': the value $\hat\rho (S):=\rho\circ P(S)$ depends only on the number of elements in $S$ (in the quasi-symmetric case, an increasing bijection induces the identity $\hat\rho(S)=\hat\rho(T)$) so that if we set $\hat\rho (|S|):=\frac{\hat\rho (S)}{|S|!}$, the scalar species $\hat\rho$ is entirely characterized by the formal power series
$$\hat\rho (x):=\sum\limits_n\hat\rho(n)x^n.$$
This straightforward remark connects QFT and many-body theory with the various algebraic structures existing on the algebra of formal power series.
Although apparently uselessly pedantical, it is actually useful to understand how these structures connect to the ones existing on scalar species.
There exists a Hopf-like structure on linear combinations of finite sets (see e.g. \cite{PatrasSchocker,PatrasSchocker2} for various developments of these ideas and the related notion of twisted algebras). The coproduct is defined by:
$$\delta(S):=\sum\limits_{U\coprod T=S}U\otimes T$$
where $\coprod$ stands for the disjoint union,
whereas the product is simply induced by the disjoint union of sets (the product of two overlapping sets is not defined).
These maps induce a convolution product written $\odot$ (to distinguish it from the convolution product of forms on $H$) on scalar species: for $\alpha,\beta$ two scalar species we get:
$$\alpha\odot\beta (S):=\sum\limits_{U\coprod T=S}\alpha(U)\beta(T),$$
or: ${\alpha\odot\beta} (x)=\alpha(x)\beta(x)$.
\begin{thm}[General functional linked cluster theorem]
We have, for any unital natural form $\rho$ on $H$ and admissible $\mathcal P$:
$$\log(\hat\rho (x))=\sum\limits_n\sum\limits_{\Gamma_n^c}\frac{s_{\Gamma_n^c}^{n}}{n!}\tau(\Gamma_n^c)x^n,$$
where $\tau:=\log^\ast(\rho)$, $\Gamma_n^c$ runs over the connected Feynman diagrams with vertex set $[n]$ and $s_{\Gamma_n^c}^{n}:=\sum\limits_b\lambda_bs_{\Gamma_n^c}^{b}$, where $P([n])=\sum\limits_b\lambda_b b$ is the decomposition of $P([n])$ as a linear combination of basis elements.
\end{thm}
Proof.
We have:
$$\log(\hat\rho (x))=(\log^\odot\hat\rho)(x)=\sum\limits_n\frac{(\log^{\odot}\hat\rho)([n])}{n!}x^n$$
$$=\sum\limits_n\frac{x^n}{n!}\sum\limits_{I_1\coprod ...\coprod I_k=[n]}\frac{(-1)^{k+1}}{k}\hat\rho (I_1)...\hat\rho (I_k)$$
$$=\sum\limits_n\frac{1}{n!}\sum\limits_k\frac{(-1)^{k+1}}{k}\sum\limits_{I_1\coprod ...\coprod I_k=[n]}\sum\limits_{\Gamma_{I_1},...,\Gamma_{I_k}}s_{\Gamma_{I_1}}^{i_1}...s_{\Gamma_{I_k}}^{i_k}\tau (\Gamma_{I_1})...\tau (\Gamma_{I_k})$$
$$=\sum\limits_n\frac{1}{n!}\sum\limits_k\frac{(-1)^{k+1}}{k}\sum\limits_{I_1\coprod ...\coprod I_k=[n]}\sum\limits_{\Gamma_{I_1},...,\Gamma_{I_k}}s_{\Gamma_{I_1}}^{i_1}...s_{\Gamma_{I_k}}^{i_k}\tau (\Gamma_{I_1}... \Gamma_{I_k})$$
where $\Gamma_{I_i}$ runs over the Feynman brackets in the expansion of $\hat\rho (I_i)$, $i_1:=|I_i|$ and $\Gamma_i\Gamma_j$ denotes the concatenation of two brackets (so that e.g. $[\phi(x_1)|\phi(x_5)^2\phi(x_8)][\phi(x_2)^3|\phi(x_1)]=[\phi(x_1)|\phi(x_5)^2\phi(x_8)|\phi(x_2)^3|\phi(x_1)]$).
Now, let $\Gamma=\Gamma_1\coprod ...\coprod \Gamma_p$ be the (unique) decomposition of a Feynman diagram showing up in the expansion of $\hat\rho ([n])$ into a product of connected (non empty) diagrams. According to Lemma~\ref{keylemma} and since $\mathcal P$ is admissible, for any partition $A_1\coprod ...\coprod A_l$ of $[p]$, we have $s_{\Gamma_{A_1}}^{a_1}...s_{\Gamma_{A_l}}^{a_n}=s_{\Gamma}^n$, where $\Gamma_{A_i}:=\coprod\limits_{j\in A_i}\Gamma_j$ and $a_i$ is the number of vertices of $\Gamma_{A_i}$.
The Theorem amounts then to the following properties: the coefficient of $\tau(\Gamma)$ in $\log(\hat\rho (x))$ is $\frac{s_\Gamma^n}{n!}x^n$ if $p=1$ (that is if the graph is connected) and zero else.
The first property (the connected case) is obvious from the expansion. Let us assume therefore that $p>1$. The property follows once again from the general properties of the partition posets: the equation~(\ref{eqq}) concludes the proof.
\section{Examples}
\subsection{Quantum Field Theory}
\label{QFTsect}
In the quantum theory of the scalar field, the underlying Hopf algebra is the bosonic algebra ${\mathcal B}_k^{X}$, where $k$ is the number of fields and the elements of $X$ stand for dummy position or momentum variables.
A typical example is the $\phi^4$ (scalar) theory with $k=1$ (we write simply $\phi$ for $\phi_1$).
The form $\rho$ computes expectation values of time ordered products of
free fields over the vacuum:
$$\rho(\phi(x_1)...\phi(x_k)):=<0|T(\phi(x_1)...\phi(x_k))|0>.$$
Problems arise when some of the $x_i$s coincide; these problems are the subject of the renormalization theory, we do not address them here.
The physically interesting quantities are the interacting Green functions
\begin{eqnarray}
G(x_1,\dots,x_n) :=
\frac{\rho\Big(\phi(x_1)...\phi(x_k)e^{i\int \phi^4(x) dx}\Big)}
{\rho\Big(e^{i\int \phi^4(x) dx}\Big)}.
\label{Greendef}
\end{eqnarray}
The key point is that $\rho=\exp^*\tau$, where $\tau$ is
zero if its argument has degree different from two and
$\tau\big(\phi(x)\phi(y)\big):=<0|T(\phi(x)\phi(y))|0>$
is the Feynman propagator. The convolution logarithm can
then be written as a sum of Feynman diagrams, where the lines
represent Feynman propagators and the vertices represent
spacetime points $x_i$. It can be checked that the
standard Feynman rules of quantum field theory~\cite{Itzykson} are
exactly recovered by the convolution exponential~\cite{BrouderMN}.
The linked-cluster expansion provides a simple way to
deal with the denominator of eq.~(\ref{Greendef})~\cite{Gross}.
\subsection{Cumulants}
Let $X_1,...,X_n,...$ be a sequence of random variables.
The underlying Hopf algebra for this example is once again the bosonic algebra ${\mathcal B}_1^{\NN}$. The form $\rho$ is defined by:
$$\rho (\phi(x_{i_1})...\phi(x_{i_k})):=E[X_{i_1}...X_{i_k}].$$
This example enters the general commutative local case.
When all the $x_i$s are distinct, $\rho (\phi(x_{i_1})...\phi(x_{i_k}))$ can be expanded as a sum parametrized by Feynman graphs which are disjoint unions of elementary graphs made of distinct black vertices, each joigned to a unique white vertex. The connected Feynman graphs appearing in the expansion correspond to the cumulants $E_c[X_{i_1}...X_{i_k}]$.
The combinatorial linked cluster actually shows that this graphical expansion is equivalent to
the classical identity:
$$E[X_1...X_n]=E_c[X_1...X_n]+\sum\limits_{A_1\cup ...\cup A_k}\prod_{i=1}^kE_c[X_{a_1^i}...X_{a_{j_i}^i}],$$
where $A_1\cup ...\cup A_k$ runs over the proper partitions of $[n]$ and $A_i=\{a_1^i,...,a_{j_i}^i\}$.
When the $X_i$ are copies of a given random variable $X$
and setting $P([n]):=E[X_1...X_n]$ (which is admissible), we recover, using the functional linked cluster theorem:
$$<e^X>_c-1=\log (E(e^X)),$$
with the convention $<1>_c=1$ and $<X^n>_c:=E_c[X_1...X_n]$.
\subsection{Quantum field theory with initial correlation}
In solid state physics and quantum chemistry, the initial
state is generally different from the vacuum. The physically
relevant form becomes (with the same notation as in
the first example)
$$\rho(\phi(x_1)...\phi(x_k)):=<\Phi|T(\phi(x_1)...\phi(x_k))|\Phi>,$$
where $|\Phi>$ is a general state. It is also possible to
consider a mixed state instead of the pure state $|\Phi\rangle$.
Except for very special cases (quasi-free states for bosonic
fields and Slater determinants for fermionic fields)
the convolution logarithm $\tau$ of $\rho$ is then
more complicated than in the first example. In particular,
$\tau$ can be nonzero if its argument have degree different
from two. In quantum optics, expansions in terms of $\tau$ are known
as cluster expansions and they lead to much
better convergence properties~\cite{Kira-08}.
For the fermionic fields, the convolution logarithms $\tau$ are equivalent to
the cumulants of the reduced density matrices, that
are strongly advocated by Kutzelnigg and
Mukherjee~\cite{Kutzelnigg,KutzMukh,Kutzelnigg-00,Kong-10}.
The diagrammatic expansions can then not be done anymore using Feynman diagrams constructed out of Feynman propagators: see e.g. our \cite{BP09} and require the full apparatus of generalized Feynman diagrams for commutative local case.
\subsection{Non-Gaussian measures}
Perturbative expansions in statistical physics for measures of Gaussian type can be performed using the usual Feynman graphs of Sect. \ref{QFTsect}.
This is because the Wick theorem applies.
When dealing with arbitrary functional measures this is not the case any more: higher cumulants (i.e. higher truncated moments or truncated Schwinger functions) have to be taken into account.
Feynman graphs and linked-cluster theorems have been developed by Djah and coll.~\cite{Djah} in this framework.
They were extensively used in several problems of probability
theory~\cite{Djah2,Gottschalk-07,Gottschalk-08}.
These Feynman diagrams are equivalent to those of a quantum field theory with
initial correlation.
\subsection{Free probabilities}
Free probabilities deal with the noncommutative local case and study linear forms on the tensor algebra ${\mathcal F}_1^\NN$. In general, the graphs required to study such forms are free interaction graphs.
In practice, the theory of free probabilities focus often on linear forms with particular properties. This allows for various simplifications and typical properties as far as the corresponding cumulant expansions and their diagrammatic expansions are concerned. In particular, the Speicher's notion of free (or noncrossing) cumulant is obtained from the moment generating function by M\"obius inversion with respect to the lattice of noncrossing partitions (and not with respect to the lattice of partitions), see e.g. \cite{lehner} for further details and references on the subject.
\subsection{Truncated Wightman distributions}
In axiomatic quantum field theory, the
form used in section~\ref{QFTsect} are replaced by
$$W(x_1,...,x_n):=<\Phi|\phi_H(x_1)...\phi_H(x_k)|\Phi>,$$
where the operator product is used instead of the
time-ordered product, the fields are written in the
Heisenberg picture and $|\Phi\rangle$ is the ground
state of the interacting system.
Such functions are called Wightman distributions or
correlation functions.
The main difference with the standard case is that
the Wightman distributions are not symmetric, the
order of the arguments is fixed.
Still, there is a perturbation theory of Wightman functions
that leads to non-commutative Feynman diagrams~\cite{Ostendorf-84}.
Their combinatorics is the same as for the standard case~\cite{BFK}.
Thus, the corresponding convolution logarithm $\tau$
is again zero if its argument is not of degree two,
but now
$\tau\big(\phi(x),\phi(y)\big)
=\langle 0|\phi(x)\phi(y)|0\rangle$ is different
from $\tau\big(\phi(y),\phi(x)\big)$.
However, it is also possible to work at the non-perturbative
level and to define the form
$\rho\big(\phi(x_1)\dots\phi(x_n)\big):=W(x_1,\dots,x_n)$.
In that case, the convolution logarithm $\tau$ is
generally not zero if its argument is not of degree two
and $\tau\big(\phi(x_1)\dots\phi(x_n)\big)$
is now called a truncated Wightman distribution.
The definition of truncated Wightman distributions
was first given by Rudolf Haag in 1958~\cite{Haag58}.
We follow (up to the order of the variables) the definition
of Sandars' paper~\cite{Sanders-10}: ``For $n\ge1$ we let $\mathcal{P}_n$ denote
the set of all partitions of the set $\{1,\dots,n\}$ into pairwise disjoint
subsets, which are ordered from low to high. If $r$ is an ordered set in the
partition $P\in \mathcal{P}_n$ we write $r\in P$ and we denote the elements
of $r$ by $r(1)<\dots <r(|r|)$, where $|r|$ is the number of elements
of $r$. The truncated $n$-point distributions $\omega_n^T$,
$n\ge1$ of a state $\omega$ are defined implicitly in terms
of the $n$-point distributions
\begin{eqnarray*}
\omega\big(\varphi(x_1),\dots,\varphi(x_n)\big)
&=& \sum_{P\in\mathcal{P}_n} \prod_{r\in P}
\omega^T_{|r|}(x_{r(1)},\dots,x_{r(|r|)}).
\end{eqnarray*}
This is exactly the relation between $\rho=\exp^*\tau$ and
$\tau$ for non-commuting variables.
From the physical point of view, the truncation procedure
eliminates the contribution
of the vacuum state as an intermediate state~\cite[p.~271]{Epstein}.
Truncated distributions have many desirable properties.
For instance, they decrease much faster than
Wightman distributions at large space-like separation~\cite{Araki-60}.
\subsection{Nonrelativistic systems with Coulomb interaction}
Let us neglect here the problem of dealing with the Fermi statistics (which amounts essentially to introducing the correct signs in the definition of the Hopf algebra structures, see e.g. \cite{Cassam,BrouderQG}).
Let us consider $n$ electrons in a quantum system
of non-relativistic electrons with a Coulomb interaction
in the external field generated by nuclei.
This is the standard approach of quantum chemistry
and solid-state physics.
We assume that the non-interacting state can be
described by a Slater determinant and that the
particle-hole transformation was used to
deal with occupied states.
The form is defined as in section~\ref{QFTsect}
and the Green functions are now
$$
G(x_1,\dots,x_{n},y_1,\dots,y_n) :=
\frac{\rho\Big(\psi(x_1)\dots\psi(x_n)\psi^\dagger(y_1)
\dots\psi^\dagger(y_n)e^{-i I}\Big)}
{\rho(e^{-i I})},$$
where
$$
I=e^2
\int d t d\mathbf{r} d \mathbf{r'}
\frac{\psi^\dagger(\mathbf{r},t)\psi^\dagger(\mathbf{r}',t)
\psi(\mathbf{r}',t)\psi(\mathbf{r},t)}{8\pi\epsilon_0
|\mathbf{r}-\mathbf{r}'|}.$$
The main difference with quantum field theory is that
the interaction is not local. Still, in the
linked-cluster expansion, we want to consider that
the points $\mathbf{r}$ and $\mathbf{r}'$ in $I$
are connected. In diagrammatic terms,
$\mathbf{r}$ and $\mathbf{r}'$ are connected
by a wavy line. We introduced the tripartite
graphs to deal with this important case.
\bibliographystyle{unsrt}
|
\section{Introduction}
\label{sec:introduction}
The choice of coordinate system often is of great importance when analyzing real-world phenomena and data.
In some cases appropriate basis vectors or basis functions can be chosen from prior knowledge about the system.
If a model does not exist, statistical approaches such as principal component analysis (PCA) or singular value decomposition (SVD) are used.
With an appropriate choice of basis, high-dimensional data often can be represented in a subspace that captures the essential features of the data.
The data analysis process, {\it e.g.},\ classification, clustering or denoising, is then more efficient and reliable.
The standard statistical approaches for dimension reduction are optimal in the sense that specific statistical features of the data are preserved.
Typical drawbacks of this are the high computational complexity, and that all data needs to be re-analyzed when new data is added or that new data is biased by old estimates.
Nonlinear dimension reduction techniques, {\it e.g.},\ autoencoder neural networks \cite{Hinton28072006}, do in some cases perform better than the linear methods in terms of class separability, but this comes at the cost of increased computational complexity.
In this paper we present a dimension reduction technique for tensors that enables incremental encoding and decoding of tensor components.
It does not require that the whole data set is stored and no computationally expensive re-analysis is needed when new data is inserted.
Instead, this method encodes the components in a compactified data structure so that decoding of significant features is possible with high probability.
The range of tensor indices can be extended dynamically, without modifying the component representation (the rank is fixed at construction).
This work was stimulated by a method used in natural language processing that is known as {\it random indexing} \cite{kanerva_2000}, see \cite{sahlgren_2005,kanerva_hyperdimensional_2009} for an introduction.
Random indexing is recently applied in a variety of applications, such as indexing of literature databases \cite{vidya_2010}, event detection in blogs \cite{Jurgens_2009} and graph searching for the semantic web \cite{DamljanovicPC2010}.
The idea of random indexing \cite{kanerva_2000} originates from Pentti Kanervas work on sparse distributed memory \cite{SDM_1988}, and related work on the mathematics of brain-inspired information processing with hyperdimensional symbols, see \cite{kanerva_hyperdimensional_2009} for a recent review.
In terms of basis functions this approach is based on the philosophy that ``randomness is the path of least assumption'', {\it i.e.},\ when little is known beforehand the optimal representation strategy is a random one.
Random indexing of vectors (rank-one tensors) is implemented in the S-Space Package for semantic spaces \cite{Jurgens_2010} and the Semantic Vectors Package \cite{Widdows_2008}.
Here we generalize these approaches to matrices and higher-order tensors in a unified framework:
we outline the equations and properties of {N-way random indexing} (NRI).
The major strengths of this method are that it enables on-line incremental dimension reduction and analysis of complex multi-dimensional data, and that new classes of data can be incorporated without modifying the representation.
The cost of this flexibility is a reduced signal to noise ratio due to the approximate (lossy) nature of the representation scheme, see \sect{simulation}.
Efficient methods for approximate tensor representation are generic and useful tools in the processing of the large amounts of data produced in the modern society.
Higher-order (rank $> 2$) representations are increasingly useful as the processing capacity of computers increases, and allows for more in-depth analyses of complex phenomena.
See \sect{language} for examples.
We restrict this work to spaces with an Euclidean metric, $\delta_{ij}$, so the tensors considered here are practically multi-dimensional arrays.
\subsection{Related work}
\label{sec:relatedwork}
This work was stimulated by Pentti Kanervas work on sparse distributed memory \cite{SDM_1988} and a number of papers related to hyperdimensional computing and neuro-symbolic computing, see \cite{kanerva_hyperdimensional_2009} for a recent review.
In particular it has bearing on, and is a generalization of the random indexing method developed for natural language processing \cite{kanerva_2000,sahlgren_2005}, see also p.~153 in \cite{kanerva_hyperdimensional_2009} where Kanerva mentions the possibility to extend random indexing to two dimensions.
Random projection \cite{Papadimitriou_1998} and random mapping \cite{Kaski_1998} are other methods that utilize high-dimensional random vectors for dimension reduction.
Random projection is used to reduce the dimensionality of a set of points in Euclidean space, while approximately preserving pairwise distances.
This possibility follows from the Johnson--Lindenstrauss lemma \cite{johnson_1984}, which essentially states that a small set of points in high-dimensional space can be mapped into a space of lower dimension such that the distances between the points are approximately preserved.
Random projection has enabled several breakthrough developments, {\it e.g.},\ in the field of algorithms and in elegant alternative proofs.
It is used for combinatorial optimization, machine learning and problems related to information retrieval, see \cite{VempalaBook_2004} for a review.
The method described here is based on random high-dimensional vectors and projection operators also, but the mathematical structure of the algorithm is different from that of the random projection method, and the properties of the technique from an application point of view are different.
Our method enables dimension reduction of tensors of any rank (rank $\leq 3$ is feasible today on a PC).
There are a few well-known algorithms for dimension reduction of tensors, see \cite{TensorReview} for a recent review.
In particular the Tucker decomposition and the parallel factor model are commonly used methods, and there are some extensions of these two algorithms that are commonly used also \cite{TensorReview}.
Tucker decomposition (known also as N-mode PCA, N-mode SVD etc.) is a form of higher-order principal component analysis,
which decomposes a tensor, $c_{ijk\ldots}$, according to the scheme
\begin{equation}
c_{ijk\ldots} = \sum_{\alpha}\sum_{\beta}\sum_{\gamma}\ldots~g_{\alpha\beta\gamma\ldots}~a_{i\alpha}b_{j\beta} c_{k\gamma}\ldots,
\label{eq:tucker}
\end{equation}
where $g_{\alpha\beta\gamma\ldots}$ is the core tensor and $\{a_{i\alpha},~b_{j\beta},~c_{k\gamma}\ldots\}$ are factor matrices.
The factor matrices are usually taken to be orthogonal and can be thought of as the principal components in each dimension of the tensor.
The core tensor represents interactions between the principal components.
Several methods to compute the Tucker decomposition have been developed, see \cite{TensorReview} for a review.
These methods are computationally expensive, because the goal to find an ``optimal'' decomposition with side constraints on the factor matrices is non trivial.
The method presented in this paper is somewhat similar to the Tucker decomposition, because it uses a mathematically equivalent expression for decoding of the representation.
Our method is essentially different from Tucker decomposition in the sense that it is based entirely on random coding, and it thereby avoids the computationally expensive process of calculating the factor matrices.
In our model the factor matrices are randomly generated objects, so-called random indices.
Another benefit is that our method is incremental, but it performs well only with large tensors (high dimensionality is a prerequisite).
In contrast, Tucker decomposition is useful for low-dimensional problems due to the computational complexity of the method.
The parallel factor decomposition is a special case of the Tucker decomposition, which results if the core tensor in \eqn{tucker} is enforced to be superdiagonal.
\section{Indifference property of high-dimensional spaces}
\label{sec:highd}
High-dimensional spaces are different from the two- and three-dimensional spaces that we are naturally trained to imagine.
In two and three dimensions, randomly generated vectors of equal norm are rather similar, {\it e.g.},\ when compared with the dot product.
This is not so at high dimensionality, where nearly all vectors are unsimilar.
Another counter-intuitive property is that the volume of the unit hypersphere relative to that of the hypercube with corresponding width rapidly approaches zero at increasing dimensionality.
The tendency of vectors in high-dimensional binary spaces to be different is well described in \cite{SDM_1988}.
We introduce some key points from that work here and then we will show how to generalize these ideas to ternary space $\{-1,0,1\}^n$, which is important in this work.
Consider the binary space $\{0,1\}^n$ of vectors with length $n$ and equal probability of the states $0$ and $1$.
The distance, $d$, between two binary vectors can be defined as the number of non-zero bits in the bit-wise exclusive or ({\it xor}) of the vectors.
This is equivalent to the square of the Euclidean distance and it corresponds to how many bits that are different in two vectors.
The number of vectors in the space that are at a distance $d$ from a specific vector is given by the binomial coefficient
\begin{equation}
C(n,d) = \binom{n}{d},
\end{equation}
because this is the number of different ways to choose (flip) $d$ bits out of $n$.
The number of vectors at a certain distance from a reference vector therefore follows the binomial distribution with probability $p=1/2$, which has mean $n/2$ and variance $n/4$.
At high values of $n$ the binomial distribution can be approximated with a normal distribution.
If a distribution is approximately normal the proportion within $z$ standard deviations of the mean is erf$(z/\sqrt{2}).$
This implies that the distance distribution is highly concentrated around the mean, because the error function quickly approaches unity for increasing $z$.
For example, $99.7$\% of the distances are within three standard deviations from the mean distance.
Only one billionth $(10^{-9})$ of the distances deviate more than six standard deviations from the mean.
The mean distance is $n/2$ and the standard deviation of distance is $\sqrt{n}/2$.
This implies that the mean distance is $\sqrt{n}$ standard deviations, {\it e.g.},\ $31.6$ standard deviations for $n=1000$.
A striking consequence of this distribution of distances is that practically all vectors in a high-dimensional binary space are located at a distance $\sim n/2$ from any specific vector.
For $1000$-bit vectors the standard deviation is $\sqrt{1000}/2\simeq 15.8$~bits.
Six standard deviations correspond to $95$~bits.
All but one billionth of $1000$-bit vectors are located at $500 \pm 95$~bits from {\it any} specific vector in that space.
The concentration of distances around the mean increases with $n$ and it implies that randomly generated high-dimensional vectors are {\it indifferent} with high probability.
The indifference property is at the core of the ideas discussed in Pentti Kanervas book on sparse distributed memory \cite{SDM_1988}, and related work on the mathematics of brain-inspired information processing with hyperdimensional symbols, see \cite{kanerva_hyperdimensional_2009} for an introduction.
Concepts and symbols represented with arbitrary random high-dimensional patterns are indifferent by chance, but new associations and transformations can be constructed from existing representations with a learning mechanism.
Next we consider indifference (orthogonality) in high-dimensional ternary space.
\section{Properties of high-dimensional ternary space $\{-1,0,1\}^n$}
\label{sec:ternary}
In this work we are interested in ternary vectors, {\it i.e.},\ instead of bits with two possible states, $\{0,1\}$, we consider (balanced) trits with states $\{-1,0,1\}$.
The introduction of a state with negative sign is important, we will return to that in the following section.
This discussion concerns ternary vectors of length $n$ with $k$ positive ($1$) and $k$ negative ($-1$) states, where $k\ll n/2$, {\it i.e.},\ we are interested in sparse ternary vectors with vanishing sum.
The ternary space can be visualized as a subset of an inner product space where orthogonality is defined by a vanishing dot product between two vectors.
With this (the usual) definition of orthogonality, it follows that an $n$-dimensional ternary space has at most $n$ mutually orthogonal vectors.
However, in a high-dimensional space there are many more vectors that are ``nearly orthogonal''.
This is analogous to the high probability of indifference between vectors in high-dimensional binary space, which is described above.
In the following we justify this essential point with explicit results for the approximate orthogonality of sparse vectors in high-dimensional ternary space.
As far as we know the result has not been presented elsewhere.
The total number, $N$, of ternary vectors of length $n$ that has $k$ positive and $k$ negative components is
\begin{equation}
N = \binom{n}{2k}\binom{2k}{k} = \binom{n}{k}\binom{n-k}{k},
\label{eq:nternary}
\end{equation}
because there are $C(n,2k)$ different ways to choose $2k$ non-zero states and
$C(2k,k)$ different ways to distribute the signs to the non-zero states.
The alternative (second) definition above can be interpreted in a similar way.
There are $C(n,k)$ different ways to choose the positive states and $C(n-k,k)$ ways to choose the negative states, or vice versa.
These two definitions are mathematically equivalent.
How many of these $N$ vectors have a dot product that is nearly zero, {\it i.e.},\ how many of them are approximately orthogonal?
Let $d=|\langle\cdot,\cdot\rangle|$ be the absolute value of the dot product between two vectors.
For simplicity we restrict the analysis to $0\leq d\leq k$, because we are interested in approximately orthogonal vectors only.
This restriction does not affect the accuracy of the result.
We assume also that the vectors are sparse so that $n \gg k$.
Imagine a fixed reference vector that is picked at random from the space of $N$ vectors.
This reference vector has $k$ positive states, $k$ negative states and $n-2k$ states that are zero.
The large majority of vectors with $\langle\cdot,\cdot\rangle=\pm d$ with respect to this reference vector will have $d$ states that coincides with the $2k$ non-zero states of the reference vector, and the remaining $2k-d$ non-zero states will be distributed among the $n-2k$ states that are zero in the reference vector.
There are additional vectors with the same value of $d$, because cancellations of type $1+1-1=1$ result from higher-order coincidences.
The relative number of such vectors is, however, insignificant and we therefore neglect them here. This simplification is justified with a numerical calculation, we will return to that below.
The selection of $2k-d$ non-zero states out of $n-2k$ gives a factor of $C(n-2k,2k-d)$.
Then remains the question how many possibilities there are to select those $2k-d$ non-zero states from the $2k$ non-zero states in the reference vector, and how many combinations that arise because of signs.
These questions are not independent, because the number of ways to choose $2k-d$ states from $2k$ states depends on the number of $+1$ states that are chosen, and the relative number of $+1$ states that are chosen will affect also the number of possible permutations.
Accounting for these constraints the number of vectors is
\begin{widetext}
\begin{equation}
N(n,k,d) \simeq
\binom{n-2k}{2k-d}
\sum_{n_+=k-d}^{k} \binom{k}{n_+}\binom{k}{2k-d-n_+}\binom{2k-d}{n_+},
\quad d\leq k,~n\gg k,
\label{eq:ninner}
\end{equation}
\end{widetext}
where $n_+$ denotes the number of positive states that are chosen from the $2k$ non-zero states in the reference vector.
The number of negative states chosen is $n_- = 2k-d-n_+$.
The sum in \eqn{ninner} arises because there are multiple choices for the number of positive states to choose from the reference vector.
At most $k$ positive states can be chosen, {\it i.e.},\ all positive states.
The lower limit of $n_+=k-d$ corresponds to the maximum value for the number of negative states chosen, $n_-=k$.
The first factor in the sum, $C(k,n_+)$, accounts for the number of ways to choose $n_+$ positive states from the $k$ positive states in the reference vector.
Similarly, the second factor accounts for the number of ways to choose $n_-$ negative states from the $k$ negative states in the reference vector.
The last factor accounts for sign permutations when distributing the chosen states to the $2k-d$ non-zero states that are selected by the prefactor.
Since the number of positive and negative signs are fixed, the combinatorial problem solved here should have a hypergeometric character.
This is indeed the case, because the sum in \eqn{ninner} can be replaced with a generalized hypergeometric function,
\begin{widetext}
\begin{eqnarray}
N(n,k,d) &\simeq&
\binom{n-2k}{2k-d}
\binom{2k-d}{k}
\binom{k}{k-d} \nonumber \\
&& \times ~~ {}_3F_2(-d,-k,-k;~1+k-d,1+k-d;~-1),
\quad d\leq k,~n\gg k.
\label{eq:ninnerhyp}
\end{eqnarray}
\end{widetext}
The generalized hypergeometric function, $_3F_2$, is a standard mathematical function that is described, {\it e.g.},\ on-line and in the book \cite{NIST_book}.
If we divide the number of vectors, $N(n,k,d)$, which has a specific value of $d$ with respect to any reference vector, with the total number of vectors in the space, $N$, the result is the relative size of the space as a function of $d$.
The relative size of the space is equivalent to the probability of randomly choosing a vector from the space that has a dot product of $\pm d$ with respect to a reference vector,
\begin{equation}
P(n,k;~\langle\cdot,\cdot\rangle=\pm d) \simeq N^{-1} N(n,k,d), \quad d\leq k,~n\gg k.
\label{eq:pinner0}
\end{equation}
This distribution function is the result that we are looking for, because it describes the probability that randomly chosen vectors from the space are nearly orthogonal.
The numbers $N$ and $N(n,k,d)$ are enormous ($n$ is a high number).
For practical purposes we therefore make a series expansion of factors involving $n$ in the limit $n\rightarrow\infty$.
The result is,
\begin{widetext}
\begin{eqnarray}
P(n,k;~\langle\cdot,\cdot\rangle=\pm d) &\simeq&
\frac{T_1 + T_2}{n^d}
\sum_{i=0}^{d}\frac{(k!)^4}{
\left[(k-d+i)!\right]^2
\left[(k-i)!\right]^2
(d-i)!~i!} \nonumber \\
&=&
\frac{(T_1 + T_2)d!}{n^d}
\binom{k}{d}^2
{}_3F_2(-d,-k,-k;~1+k-d,1+k-d;~-1),
\label{eq:pinner1} \\
T_1 &=& 1-\frac{8k^2+d^2+d-8kd}{2n}, \\
T_2 &=& \frac{1}{n^2}\left[
2(1-2k)^2 k^2
+ \frac{d^4}{8}
+ \left(\frac{5}{12}-2k\right)d^3
+ \left(10k^2-4k+\frac{3}{8}\right)d^2 \right. \nonumber \\
&& \left. +\left(-16k^3+10k^2-2k+\frac{1}{12}\right)d
\right] + {\cal O}(n^{-3}),
\end{eqnarray}
\end{widetext}
where the terms $T_1$ and $T_2$ originate from the series expansion.
The assumptions $d\leq k$ and $n\gg k$ are to be respected in applications of this result.
Numerical results for the dot product between a reference vector and $10^{12}$ randomly chosen ternary vectors are presented in \tab{dotp}.
\begin{table}[h]
\caption{Approximate orthogonality of the high-dimensional space $\{-1,0,1\}^n$.
Tabulated here is the probability, $P$, in \eqn{pinner1} for different values of the vector length, $n$, and number of non-zero components, $2k$.
These probabilities are to be compared with the corresponding probabilities obtained from explicit numerical simulations, $P_{sim}$.
Entries marked with an asterisk demonstrate the effect of neglecting contributions to the inner product arising from higher-order trit combinations (like $\langle\cdot,\cdot\rangle = \ldots+1\times 1\ldots-1\times 1\ldots+1\times 1\ldots=1$) in the analysis leading to \eqn{pinner1}.
The series expansion is marginally applicable in the case $n=10^2$ for low values of $k$, and $n \gg k$ is violated for high $k$.
}
\vspace{1ex}
\centering
\begin{tabular}{| r | r | l r | l r | l r |}
\hline
& & \multicolumn{2}{|c|}{$n=10^2$} & \multicolumn{2}{|c|}{$n=10^3$} & \multicolumn{2}{|c|}{$n=10^4$} \\[0.5ex]
$2k$ & $\langle\cdot,\cdot\rangle$ &
$~P_{sim}$ & $P~~~$ &
$~P_{sim}$ & $P~~~$ &
$~P_{sim}$ & $P~~~$ \\[0.5ex]
\hline
4 & 0 & 8.5e-1 & \quad 8.47e-1& 9.8e-1 & \quad 9.84e-1& $\sim$1.0 & \quad 9.98e-1 \\
& $\pm$1 & 7.3e-2 & 7.29e-2& 7.9e-3 & 7.93e-3& 8.0e-4 & 7.99e-4\\
& $\pm$2 & 2.0e-3 & 1.94e-3& 2.0e-5 & 1.99e-5& 2.0e-7 & 2.00e-7\\
\hline
8 & 0 & 5.5e-1 & *5.17e-1& 9.4e-1 & 9.38e-1& 9.9e-1 & 9.94e-1\\
& $\pm$1 & 1.9e-1 & 1.90e-1& 3.0e-2 & 3.05e-2& 3.2e-3 & 3.20e-3\\
& $\pm$2 & 2.8e-2 & 2.74e-2& 3.9e-4 & 3.86e-4& 4.0e-6 & 3.99e-6\\
& $\pm$3 & 2.0e-3 & 1.96e-3& 2.4e-6 & 2.44e-6& 2.5e-9 & 2.49e-9\\
& $\pm$4 & 7.4e-5 & 7.42e-5& 8.2e-9 & 8.22e-9& $<$$10^{-10}$ & 8.3e-13\\
\hline
12 & 0 & 3.5e-1 & --~~~~& 8.7e-1 & 8.65e-1& 9.9e-1 & 9.86e-1\\
& $\pm$1 & 2.3e-1 & --~~~~& 6.4e-2 & 6.37e-2& 7.1e-3 & 7.10e-3\\
& $\pm$2 & 7.9e-2 & --~~~~& 2.0e-3 & 1.99e-3& 2.2e-5 & 2.17e-5\\
& $\pm$3 & 1.6e-2 & --~~~~& 3.4e-5 & 3.44e-5& 3.7e-8 & 3.69e-8\\
& $\pm$4 & 2.0e-3 & --~~~~& 3.6e-7 & 3.64e-7& $<$$10^{-10}$ & 3.8e-11\\
\hline
16 & 0 & 2.5e-1 & --~~~~& 7.8e-1 & *7.73e-1& 9.7e-1 & 9.75e-1\\
& $\pm$1 & 2.0e-1 & --~~~~& 1.0e-1 & 1.02e-1& 1.3e-2 & 1.25e-2\\
& $\pm$2 & 1.1e-1 & --~~~~& 5.9e-3 & 5.94e-3& 7.1e-5 & 7.09e-5\\
& $\pm$3 & 4.3e-2 & --~~~~& 2.0e-4 & 2.01e-4& 2.3e-7 & 2.34e-7\\
& $\pm$4 & 1.1e-2 & --~~~~& 4.4e-6 & 4.44e-6& 4.9e-10 & 5.0e-10\\
\hline
20 & 0 & 2.0e-1 & --~~~~& 6.9e-1 & *6.72e-1& 9.6e-1 & 9.61e-1\\
& $\pm$1 & 1.8e-1 & --~~~~& 1.4e-1 & 1.39e-1& 1.9e-2 & 1.93e-2\\
& $\pm$2 & 1.2e-1 & --~~~~& 1.3e-2 & 1.31e-2& 1.8e-4 & 1.75e-4\\
& $\pm$3 & 6.5e-2 & --~~~~& 7.4e-4 & 7.36e-4& 9.6e-7 & 9.55e-7\\
& $\pm$4 & 2.7e-2 & --~~~~& 2.8e-5 & 2.78e-5& 3.5e-9 & 3.49e-9\\
\hline
\end{tabular}
\label{tab:dotp}
\end{table}
These numerical results confirm the analytical result.
Observe, however, that the accuracy of the analytical result is poor for low values of $n$ and high values of $k$, as indicated in the table.
This is connected to the assumption that $n\gg k$ in the analysis above. Note the prefactor of $n^{-d}$ in the series expansion, which implies that the probability for high values of $d$ is low.
For historical reasons we consider the example $n=10^4$ and $k=10$, which is typical in Kanervas work \cite{SDM_1988}.
It then follows from \tab{dotp} that $96$\% of the space is orthogonal with respect to a reference vector, and less than $4$\% ($2\times 1.93$) of the space has a dot product of $+1$ or $-1$.
Only $7\times 10^{-9}$ of the space has a dot product with a magnitude higher than or equal to four, which corresponds to $\sim 20$\% non-zero trits in common.
With $25$\% common trits ($d=5$ and $k=10$) the relative size of the space is $2\times 10^{-11}$.
This demonstrates that most of the space is approximately orthogonal to any particular vector in the space.
Analogously, the dot product of pairs of vectors that are randomly chosen from the space follows the same distribution, {\it i.e.},\ equation \eqn{pinner1}.
The probabilities for $n=10^4$ and some different values of $k$ are illustrated in \fig{dotp}.
\begin{figure*}[h]
\centering
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=\linewidth]{dotp2}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=\linewidth]{dotp3}
\end{minipage}
\caption{
Approximate orthogonality of high-dimensional ternary space $\{-1,0,1\}^n$.
The panel on the left-hand side shows the probability \eqn{pinner1} for inner products of sparse ternary vectors of length $n=10^4$ and different numbers of non-zero components, $2k$.
The panel on the right-hand side shows the probability \eqn{pinner1} for $k=10$ and different lengths of the ternary vectors, $n$.
In both cases the horizontal scale is normalized to the maximum value of the inner product, which is $2k$.
Probabilities for absolute values of $\langle\cdot,\cdot\rangle$ higher than $50$\% of the maximum are excluded, because \eqn{pinner1} is valid for $d\leq k$ only.
For $n=10^4$ and $k=4$, {\it i.e.},\ ternary vectors of length ten thousand with four positive and four negative trits, the probability that a randomly generated vector has an inner product of four with respect to a reference vector is about $10^{-12}$.
The probability of an inner product of minus four is about $10^{-12}$ also.
Similarly, for $n=10^4$ and $k=12$ the probability of $50$\% overlap ($\langle\cdot,\cdot\rangle = \pm 12$) is about $10^{-30}$.
Note the prefactor $n^{-|\langle\cdot,\cdot\rangle|}$ in the series expansion of the probability \eqn{pinner1}, which indicates that the probability of high absolute values of the inner product is low.
}
\label{fig:dotp}
\end{figure*}
Note that for a given subset of $N_s$ vectors the distribution of dot products is different, because there are more pairs of vectors than individual vectors in a set.
This is similar to birthday type problems, where the probability that at least two people in a room has the same birthday depends on the number of pairs rather than the number of individuals in the room.
A similar situation arise in the context of hash tables, where collisions are more frequent than suggested by the number of hash values.
The number of pairs in a subset of $N_s$ vectors is $C(N_s,2)=N_s(N_s-1)/2$.
This implies that the number of pairs increases roughly as the square of the number of vectors in the set, and that the number of vectors that can be chosen randomly with a low probability for significant correlations should be reduced with a square root compared to what is suggested by equation \eqn{pinner1}, \tab{dotp} and \fig{dotp}.
Next we present the generalized random indexing algorithm, which is based on this idea of approximate orthogonality.
\section{Random-index coding of tensors}
\label{sec:ricoding}
The concept of random indexing (RI) is introduced in \sect{introduction}.
It was invented for dimension reduction and semantic analysis in the context of natural language processing.
In that context the RI method is used to encode vectors that represent the ``meaning'' of words, so-called semantic vectors, or context vectors.
Here we generalize RI to tensors of arbitrary rank and present the algorithm of N-way random indexing (NRI).
In the following, tensor components are denoted with $c_{ijk\ldots}$ and indices $\{i,~j,~k,\ldots\}$ are used in component space.
Tensor {\it states} are denoted with $s_{\alpha\beta\gamma\ldots}$ and indices $\{\alpha,~\beta,~\gamma,\ldots\}$ are used in state space.
The state tensor has a physical representation that is stored in memory, but it is accessed by encoder and decoder functions only.
Tensor components, $c_{ijk\ldots}$, are related to the states and constitute the input to (output from) the encoder (decoder) function.
The rank of the state tensor is equivalent to that of $c_{ijk\ldots}$, but the state tensor may be of significantly smaller size.
For each index (dimension) of these tensors there is an associated random-index tensor, $r_{{\cal D},i\alpha}$, where ${\cal D}$ is a dimension index.
For vectors ${\cal D}=1$, for matrices ${\cal D} \in \{1,2\}$ and so on.
For given values of ${\cal D}$ and $i$ the state-space components of $r_{{\cal D},i\alpha}$ form a sparse high-dimensional random-index vector, referred to here as {\it index vectors}:
\begin{equation}
r_{{\cal D},i\alpha} = [\ldots~0~0~0~1~0~0~0~\ldots~0~0~0~{-1}~0~0~0\ldots]_{{\cal D},i}.
\end{equation}
Index vectors have a few non-zero components at random positions, {\it i.e.},\ at random values of $\alpha$ (thereby the name ``random index'').
The non-zero components have an absolute value of one and exactly half of them are negative, {\it i.e.},\ this is a sparse balanced ternary vector, see \sect{ternary}
(note that the symbol $k$ in that section is different from the index $k$ defined here).
The number of non-zero components, $\chi_{\cal D}$, is a model parameter, which has a typical value of order ten.
We denote the ranges of state indices, $\{\alpha,~\beta,~\gamma,\ldots\}$, with $[1,n_{\cal D}]$ so that, {\it e.g.},\ $\alpha \in [1,n_1]$ and $\beta \in [1,n_2]$.
Similarly, the ranges of component indices, $\{i,~j,~k,\ldots\}$, are $[1,N_{\cal D}]$.
The length of an index vector is equivalent to the maximum value of the state index, $n_{\cal D}$, in each dimension.
For example, if the state tensor is of size 1000x2000 the index vectors would be of length 1000 (2000) for ${\cal D}=1$ (${\cal D}=2$).
Observe that index vectors can be represented in compact form, {\it e.g.},\ as illustrated in \fig{encoding}, because most components are zero.
An effective way of doing that is to store the indices of non-zero components only.
Signs can be encoded implicitly with the position of the indices, {\it e.g.},\ first half are positive.
The number of non-zero components in an index vector is denoted with $\chi_{\cal D}$, which is an even number.
For each dimension, ${\cal D}$, there are $N_{\cal D}$ index vectors of length $n_{\cal D}$ and each index vector has $\chi_{\cal D}$ non-zero components.
In practical applications index vectors are represented in compact form by at most a few dozen integers, so the storage space required for an NRI representation is essentially determined by the size of the state tensor.
A summary of parameters and definitions is presented in \tab{summary}.
\begin{table*}
\caption{Summary of parameters and definitions.
}
\vspace{1ex}
\centering
\begin{tabular}{l | l}
\hline
Expression & Description \\[0.5ex]
\hline
$c_{ijk\ldots}$ & Tensor components \\
$s_{\alpha\beta\gamma\ldots}$ & State tensor, accessed by encoder/decoder functions only \\
${\cal D}$ & Dimension index ($1 \leq {\cal D} \leq$ rank) \\
$N_{\cal D}$ & Number of index vectors in dimension ${\cal D}$, $\{i,~j,~k,\ldots\} \in [1,N_{\cal D}]$ \\
$n_{\cal D}$ & Length of index vectors in dimension ${\cal D}$, $\{\alpha,~\beta,~\gamma,\ldots\} \in [1,n_{\cal D}]$ \\
$\chi_{\cal D}$ & Number of non-zero components in index vectors of dimension ${\cal D}$ \\
${\cal N}=\prod\chi_{\cal D}$ & Number of states that encode one tensor component \\
$\propto\prod n_{\cal D}$ & Disk/memory space required to store the state tensor \\
$\propto\sum N_{\cal D}\chi_{\cal D}$ & Disk/memory space required to store index vectors \\
\hline
\end{tabular}
\label{tab:summary}
\end{table*}
\subsection{Encoding algorithm}
State components, $s_{\alpha\beta\gamma\ldots}$, are initially set to zero.
This implies that the tensor components $c_{ijk\ldots}$ are zero also (see decoding).
After initialization, the components are updated with addition and subtraction operations, not by assignment.
This is necessary because assignment has unwanted side effects on other components due to the random nature of the representation.
In other words, multiple components are affected when one component is modified and assignment destroys statistics.
A tensor component $c_{ijk\ldots}$ is encoded in the state tensor $s_{\alpha\beta\gamma\ldots}$ using the index vectors.
The addition of a scalar weight $w$ to a tensor component $c_{ijk\ldots}$ corresponds to the operation
\begin{equation}
s_{\alpha\beta\gamma\ldots} \rightarrow s_{\alpha\beta\gamma\ldots}
+ w(r_{1,i\alpha}~r_{2,j\beta}~r_{3,k\gamma}~\ldots),
\label{eq:encoding}
\end{equation}
where the indices $\{i,~j,~k,\ldots\}$ are fixed by the choice of tensor component.
This means that the indices of the tensor component are used to select a particular set of index vectors, forming a subset of approximately orthogonal vectors in state space.
The outer product of index vectors is a tensor with a few ($\cal N$) non-zero components with values $+1$ and $-1$.
It has the same rank and size as the state tensor.
Subtraction of $w$ is defined by the replacement $w \rightarrow -w$ in \eqn{encoding}.
The encoding process is illustrated in \fig{encoding} using compact representation of index vectors.
\begin{figure*}[h]
\centering
\includegraphics[width=9cm]{encoding}
\caption{Encoding of a tensor component with two-way random indexing.
The highlighted rank-two tensor exists in user space, {\it i.e.},\ users can write and read it's component values.
It is a virtual object without physical representation, which is connected to the physical datastructures illustrated in the shaded area via encoder and decoder functions.
In this example a value of $5$ is added to the tensor component $c_{ij}$ of a rank-two tensor of size $N_1 \times N_2$.
The three-step encoding procedure is illustrated in the shaded area.
First the indices $i$ and $j$ of the component are used to select the appropriate index vectors.
The second step is to activate the subset of states that encode the value of $c_{ij}$.
Note that (the sign of) the activated rows and columns correspond to (the sign of) the indices in the index vectors.
The third and last step is to add $5$ to the selected states, while respecting the sign combination of the activations.
A few remarks are in order.
This method performs well for high-dimensional state tensors only, {\it i.e.},\ for high values of all $n_{\cal D}$ (the rank can be low or high).
Index vectors correspond to sparse ternary vectors $\{-1,0,1\}^n$ with vanishing sum, which are represented in compact form by storing the indices of non-zero components only and grouping the indices with respect to their sign.
The three-step process illustrated here is mathematically equivalent to \eqn{encoding}.
}
\label{fig:encoding}
\end{figure*}
\subsection{Decoding algorithm}
The decoding operation is a projection of the state tensor on the index vectors corresponding to the tensor component $c_{ijk\ldots}$
\begin{equation}
c_{ijk\ldots} = {\cal N}^{-1}\sum_{\alpha,\beta,\gamma\ldots}
r_{1,i\alpha}~r_{2,j\beta}~r_{3,k\gamma}~\ldots~s_{\alpha\beta\gamma\ldots},
\label{eq:decoding}
\end{equation}
where $\cal N$ is a normalization factor that is to be defined below.
The encoding procedure \eqn{encoding} is based on a sequence of outer products of index vectors, and the decoding procedure is the corresponding sequence of inner products, which results in a projection of the state tensor on the index vectors.
An essential aspect of this dimension reduction technique is the approximate orthogonality of the high-dimensional index vectors, see \sect{ternary}.
The approximate orthogonality makes it possible to decode many of the significant components of $c_{ijk\ldots}$ from the compressed and distributed representation in the state tensor.
For index vectors of length $n_{\cal D}$, at most $n_{\cal D}$ linearly independent vectors can be constructed (a set of basis vectors).
However, equation \eqn{pinner1} and \fig{dotp} illustrates that for large $n_{\cal D}$ there are many more vectors that are approximately orthogonal.
This makes it possible to encode and decode tensor components in a distributed and approximately orthogonal subspace of the state tensor, which is only partially overlapping the subspaces representing other tensor components.
In practice the result is that tensor components $c_{ijk\ldots}$ with high accumulated weight often can be identified, while tensor components with low accumulated weight disappear in noise.
A practical implementation of the decoding function may therefore return a top-list of components $c_{ijk\ldots}$ with high values, for given values of some of the indices $\{i,~j,~k,\ldots\}$.
These high-value components are to be interpreted as ``likely significant features''.
We return to the details of this approach in the next section, which deals with simulation results.
Note the similarity between the decoding operation \eqn{decoding} and the Tucker decomposition \eqn{tucker}.
The normalization factor in \eqn{decoding} compensates for the sum over states, {\it i.e.},\ the redundancy caused by adding $w$ to multiple states in the encoding procedure.
An index vector in dimension $\cal D$ has $\chi_{\cal D}$ non-zero components.
The normalization factor therefore is
\begin{equation}
{\cal N} = \prod_{\cal D} \chi_{\cal D}.
\label{eq:normalization}
\end{equation}
For example, the rank-two tensor in \fig{encoding} has ${\cal N} = 4\times 2 = 8$, because the index vectors have, respectively, four and two non-zero components.
The quantity $\cal N$ is a measure of the computational complexity of the encoding and decoding processes, because $\cal N$ states are accessed when encoding or decoding the value of one single tensor component.
\begin{figure*}[h]
\centering
\includegraphics[width=9cm]{decoding}
\caption{Decoding of a rank-two tensor component with the random-index method.
The process illustrated here is the inverse of that in \fig{encoding}.
Refer to that figure for further details.
The first step is to select the appropriate index vectors using the indices $i$ and $j$ of the tensor component that is to be decoded.
The index vectors are then used to activate the subset of states that encode the value of $c_{ij}$.
Note that (the sign of) the activated rows and columns correspond to (the sign of) the indices in the index vectors.
The third and last step is to sum the activated states, taking the overall sign into account, and divide the sum with the normalization constant \eqn{normalization}.
Note that the decoded value is similar to that encoded in \fig{encoding}.
The difference ($5.25$ vs $5.00$) is a consequence of the non-zero initial states.
This three-step process is mathematically equivalent to \eqn{decoding}.
}
\label{fig:decoding}
\end{figure*}
\subsection{One-way random indexing}
\label{sec:oneway}
The RI approach used in natural language processing is based on rank-one tensors \cite{kanerva_2000,sahlgren_2005}, {\it i.e.},\ vectors.
We will return to that in \sect{simulation}, but a brief description of this special case is included here for completeness.
In the traditional RI algorithm, each word type that appears in a text corpus is associated with a context vector, and each context ({\it e.g.}, document) is associated with a ternary index vector.
A context vector corresponds to the states of a rank-one tensor, and the index vectors are the ternary index vectors of that tensor.
It would be impractical to construct one state tensor with associated index vectors for each class or item that is to be represented.
Therefore, rank-one NRI tensors can be implemented as a set of rank-one tensors sharing the same index vectors.
For rank~$\geq 1$ the relative size of the index vectors, $(\Sigma N_{\cal D} \chi_{\cal D})/(\Pi n_{\cal D})$, is low and a special solution of that type is not necessary.
\subsection{Summarizing remarks}
The N-way random indexing (NRI) method that is outlined above is a linear dimension reduction technique that is to be used with large-size tensors ($n_{\cal D}>10^3$), because the performance is poor at low dimensionality of the index vectors.
This constraint is related to the probability that randomly generated ternary vectors are approximately orthogonal \eqn{pinner1}, see also \fig{dotp}.
NRI enables computationally efficient incremental dimension reduction of large tensors.
Quantitative details are lost in this process, but qualitative features can be maintained.
This key feature is demonstrated with simulations in the next section, and it follows from a property of the ternary index vectors.
We mentioned in \sect{ternary} that the co-existence of positive and negative states is essential, {\it i.e.},\ that index vectors should be ternary and not binary.
The reason is that, on average, each state is equally likely to be incremented and decremented when random or non-systematic data is encoded.
Random noise does therefore not saturate the states.
It is only via systematic updates of a subset of tensor components that the states gain relatively high amplitudes.
It is only these high-value components that are likely to be decoded correctly with a high weight, while most low-value components will have low weights that are below the noise threshold.
In this perspective NRI-encoded tensor components can be imagined as the momentaneous values of temporal integrals, and only those integrals that have accumulated relatively high values are likely to be correctly decoded and identified.
The simulation results presented in \sect{simulation} illustrates this point.
The incremental feature of this dimension reduction technique is obvious.
A tensor component can be encoded or decoded at any time by updating or reading $\cal N$ states.
No computationally expensive re-calculation of the basis is needed after the addition of data.
Another incremental feature is that the ranges of tensor indices, $N_{\cal D}$, can be increased dynamically by adding more randomly generated index vectors.
This has a low impact on the memory footprint, because index vectors are small entities and the relatively large state tensor is not modified in this process.
No computationally expensive analysis is needed to construct the index vectors, because these vectors are generated randomly.
The order of tensor component updates does not matter, which practically follows from the commutative property of the addition and subtraction operators.
To break this symmetry, {\it e.g.},\ for the purpose of coding time evolution or order, one has to introduce an additional tensor index for that degree of freedom.
Temporal random indexing \cite{Jurgens_2009} is an example, which is used to analyze the time evolution of word semantics to detect novel events in on-line texts.
Note that the decoding operation \eqn{decoding} and the distributed representation of tensor components in the state tensor is mathematically similar to the Tucker decomposition \eqn{tucker}.
Tucker decomposition is a generalization of PCA to tensors of rank~$\geq 3$ (the original work addressed rank-three tensors only) and that approach is therefore computationally expensive, and typically limited to smaller tensors than those addressed by NRI.
The state tensor defined here is analogous to the ``counters'' of Kanervas distributed memory model, which symbolizes synapse weights of neurons receiving signals from address-decoder neurons, see Chapter 4 in \cite{SDM_1988}.
The sparse character of the coding scheme used here is common in biological nervous systems, where sparse codes are essential for enegy-efficient information processing.
These analogues are mentioned here for inspiration only, they are not meant to be interpreted literally.
The distributive character of the encoding operation \eqn{encoding} and the projective character of the decoding operation \eqn{decoding} resembles the structure of neuro-symbolic encoding and decoding operations, such as those in Kanervas ``spatter code'' and Tony Plates ``holographic reduced representation'', see \cite{kanerva_hyperdimensional_2009} for references and an overview.
Next, we present simulation results that demonstrate some key properties of one- and two-way RI.
\section{Numerical study of properties}
\label{sec:simulation}
The randomly generated index vectors used to access the state tensor via outer (distributive) and inner (projective) products are ``nearly orthogonal'' by chance, but not strictly orthogonal, see \fig{dotp}.
Different tensor components are therefore encoded in partially overlapping subspaces of the state tensor, and do therefore interfere with each-other to some degree.
The magnitude of this effect depends on the tensor rank and order of NRI, the length of the index vectors, $n_{\cal D}$, the number of non-zero trits in the index vectors, $\chi_{\cal D}$, the dimension reduction, $\Pi_{\cal D} N_{\cal D}$~:~$\Pi_{\cal D} n_{\cal D}$, and the characteristics of the data that is to be encoded.
In general, tensors with few significant components of comparable magnitude are well represented, while non-sparse tensors with many significant components are poorly represented with NRI.
Components with high values have a negative impact on the reliability of components with significantly lower values.
It is therefore important to control the relative magnitude of significant components.
Observe that the tensor does not have to be sparse to get good performance.
A full tensor can be well represented if it has relatively few high-value components that represent significant features, while the majority of components have low value and are insignificant.
This property is useful if high-dimensional data is accumulated over time and the significant components are unknown beforehand.
In this section we illustrate these aspects with simulations and present an application to natural language processing that further demonstrates the NRI approach.
All simulations are performed with the RITensor software \cite{RITensor}, which includes a C++ template that implements the N-way RI algorithm and some utility functions, {\it e.g.},\ file I/O.
The template has user-defined data types for the state tensor and index vectors, and a mechanism to monitor eventual state saturation.
RITensor includes a Matlab interface also, which is convenient for small-scale experiments.
We consider an example where a dense rank-two tensor, {\it i.e.},\ a matrix, is represented with one-way and two-way RI.
Recall that a rank-two tensor can be represented with one-way RI if each column (or row) is treated as a rank-one tensor, {\it i.e.},\ a vector.
Each column of the matrix represents a class and each row represents a feature that the classes may have.
In principle the matrix elements (tensor components) could represent other items or relations, {\it e.g.},\ the significance of accumulated events in a monitoring task, the weight of edges in a graph, or the weight of word-context relations in natural language processing.
Each component of the tensor is first populated with a random integer drawn from the flat distribution $[0,10]$.
This represents noise and/or insignificant features of the classes.
We then add a relatively low number of significant features to each class, which are represented by high-value components.
The value of these components, $w$, is taken to be 100 or 1,000.
This can be translated into a signal-to-noise ratio (SNR).
The SNR can be expressed in the root-mean-square (RMS) amplitudes of signal and noise, $\text{SNR} = \left(A_{\text{signal}}/A_{\text{noise}}\right)^2$.
We take the number of features to be proportional to the size of the matrix, $N_{\cal D}$, and define the constant of proportionality as $\rho$.
The number of control features in each class then is $\rho N_{\cal D}$, each having weight $w$.
The RMS amplitude of the signal is $\sqrt{\rho N_{\cal D} w^2 / N_{\cal D}} = \sqrt{\rho}w$.
The RMS amplitude of a uniform random distribution on the interval $[0,M]$ follows from the definition and is $\sqrt{M(2M+1)/6}$.
It then follows that the SNR can be expressed in $\rho$, $w$ and $M$ in the following way
\begin{eqnarray}
\text{SNR} &=& 10~\text{log}\left(\frac{A_{\text{signal}}}{A_{\text{noise}}}\right)^2~\text{dB} \nonumber \\
&=& 10~\text{log}\left(\frac{6 \rho w^2}{M(2M+1)}\right)~\text{dB}.
\label{eq:snr}
\end{eqnarray}
For example, the SNR would be $5$~dB if the background is uniformly distributed on the interval $[0,10]$ and there are one percent features, $\rho=0.01$, with weight $w=100$.
The SNR of decoded features is lower, because the RI representation is approximate.
Note that a high SNR does not automatically imply that the data can be accurately represented with NRI, {\it e.g.},\ because a high value of $\rho$ can make the representation inaccurate, see the discussion above.
In the first example we consider a matrix of size $10,000 \times 10,000$ that is encoded with two-way RI in a $5,000\times 5,000$ state, {\it i.e.},\ the dimension reduction is 4:1.
This implies that the index vectors have length $n_{\cal D} = 5,000$.
Unless stated otherwise we will use index vectors with four positive and four negative trits so that $\chi_{\cal D}=8$.
The motivation for this choice is given below.
For each of the $10,000$ classes we add $50$ randomly selected features, which are referred to also as control features.
These features are given a weight of $100$ in one case, and $1,000$ in another case.
The question then is, how many of these fifty control features can be identified?
To answer this question we extract the fifty matrix elements of each class that have the highest decoded weights and compare these with the features that actually were encoded in the matrix.
The result of that simulation is presented in \fig{toplist}, which includes the decoded weights of the first one hundred elements.
\begin{figure*}[h]
\centering
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=\linewidth]{toplist_w100}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=\linewidth]{toplist_w1000}
\end{minipage}
\caption{Average decoded component weights of a rank-two tensor encoded with two-way RI.
In this example a $10,000\times 10,000$ matrix is encoded in a $5,000\times 5,000$ state, {\it i.e.},\ the dimension reduction is 4:1.
Each column of the matrix represents a ``class'' and each row represents a ``feature'' that the classes may have, see the text for an explanation.
The matrix is initialized with uniform random numbers in the range $[0,10]$.
Fifty randomly selected features are then added to each of the ten thousand classes.
In the figure on the (left-) right-hand side these features have weight (100) 1,000, which corresponds to a signal to noise ratio of about (1.5 dB) 21 dB.
The solid curves illustrate the average decoded weight, which has been normalized with the weight of control features (100 and 1,000, respectively).
The horizontal axis is a counter of the first one hundred top-score components returned by the decoding function.
The shaded area represents the standard deviation of the weights obtained for the ten thousand different classes.
The data points with horizontal error bars represent the average and standard deviation of the number of correctly identified features within the first fifty top-score elements.
With a weight of 100 (1,000) of the control features the number of correct features out of 50 is $39 \pm 6.4$ ($46 \pm 1.6$).
}
\label{fig:toplist}
\end{figure*}
The answer to the question posed above is that $39 \pm 6.4$ ($46 \pm 1.6$) of the fifty control features are represented among the top-fifty matrix elements of each class when the weight of the features is $100$ ($1,000$).
These numbers are somewhat higher if the matrix is encoded with one-way RI and the standard deviation of the decoded weights is lower in that case also, we will return to that below.
A number of things can be learned from the result presented in \fig{toplist}:
1) for a fixed number of features and a fixed weight of the features, a higher weight tends to increase the probability that the features are correctly identified via the decoded value of the matrix elements,
2) the standard deviations of the number of correctly identified features and the decoded weights are lower when the weight is higher,
3) it appears that the transition from control features to random noise at an index of 50 is more distinct with the higher weight of $1,000$,
and 4) the decoded weights of the top-fifty elements is distributed around the encoded weight and can deviate from it by at least $\pm 50$ percent.
The transition from features to noise in the top-list of decoded weights is in fact distinct in some cases and then resembles a first-order phase transition in thermodynamic systems.
This point is illustrated in \fig{toplist_fo}, which corresponds to the left-hand side panel of \fig{toplist} with the only modification being that $\rho=0.001$, {\it i.e.},\ there is only $0.1$ percent features in each class.
\begin{figure}[h]
\centering
\includegraphics[width=\figurewidth]{toplist_w100_rho0001}
\caption{
Example of a first-order transition in the average decoded weights of a rank-two tensor encoded with two-way RI.
The case illustrated here is identical to the left-hand side panel of \fig{toplist} with the only difference that the number of control features for each class is ten instead of fifty.
In this case $9.2 \pm 1.6$ features out of ten are identified via the top-ten list of decoded high-value matrix elements.
This is significantly better than the result presented in \fig{toplist} and the transition between features and noise is distinct.
}
\label{fig:toplist_fo}
\end{figure}
A formal analysis of the fidelity of encoded tensor components is beyond the scope of this work.
The interested reader is referred to Kanervas book \cite{SDM_1988} for a similar discussion.
This is an interesting issue for further investigation, because it is related to the question of how to select the length of the list of decoded high-value elements when the number of significant features in the classes is unknown.
Our results provide two guiding principles.
First, the reliability of the high-value elements tend to decrease with their index in the descending list.
Second, when the relative number of features is low and the relative weight of features is high the decoded weights of features is separated from the lower weights of non-features in a distinct transition.
Next we turn to a comparison of one-way and two-way RI, and a study of the effects of varying the dimensionality of index vectors.
The philosophy here is identical to that above, {\it i.e.},\ we consider matrices that encode features of classes.
We vary the relative number of features, $\rho$, from $0.1$ to $10$ percent of the size of the matrix, $N_{\cal D}$.
If $\rho N_{\cal D}$ control features are encoded in a class then a top-list of equally many high-value matrix elements is decoded and the number of control features in that top-list determines the relative number of features retrieved.
This methodology is the same as that used in the analysis leading to the data points with error bars in \fig{toplist}.
The dimension reduction, $\Pi_{\cal D} N_{\cal D}$~:~$\Pi_{\cal D} n_{\cal D}$, is kept fixed at 4:1 in this example also, but the size of the matrix is varied so that the dimensionality of the index vectors varies.
From the analysis of orthogonality in \sect{ternary} one could expect that the relative number of correctly decoded features should increase with dimensionality.
It turns out that this is not the case.
Instead we find that the relative number of correctly decoded features is practically {\em independent} of dimensionality, but that the standard deviation decreases with increasing dimensionality.
In other words, when the dimension reduction ratio is kept fixed and the number of encoded features is proportional to the size of the matrix the effect of increasing the size of the matrix and thereby the dimensionality of index vectors is to reduce the uncertainty in the number of correctly decoded features.
This point is illustrated in \fig{onevstwo}.
\begin{figure*}[h]
\centering
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=7.5cm]{1way-vs-2way}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=7.5cm]{1way-vs-2way_stddev}
\end{minipage}
\caption{
The number of correctly decoded features represented in a rank-two tensor encoded with one-way and two-way RI, see \fig{toplist} and the text for further information.
The horizontal axis represents the relative number of encoded control features, $\rho$, and the vertical axis of the panel on the (right-) left-hand side represents the average (standard deviation) of the relative number of correctly decoded features.
The average is practically independent of the size of the matrix and the dimensionality of the index vectors, provided that the dimension reduction ratio is kept constant, which is the case here.
The standard deviation of the number of correctly decoded features does, however, decrease with increasing dimensionality of the index vectors.
The shaded areas in the panel on the left-hand side illustrate the standard deviations for two different dimensionalities of the index vectors.
In the case of one-way (two-way) RI the higher standard deviation corresponds to a $5,000 \times 5,000$ matrix encoded in $1,250 \times 5,000$ ($2,500 \times 2,500$) states, while the lower standard deviation corresponds to a $10,000 \times 10,000$ matrix encoded in $2,500 \times 10,000$ ($5,000 \times 5,000$) states.
The results presented in the panel on the right-hand side corresponds to these dimensionalities also, and it includes an additional result for a $20,000 \times 20,000$ matrix that is encoded with the same dimension reduction of 4:1.
The higher dimensionality of the index vectors in the $20,000 \times 20,000$ case yields a lower standard deviation than obtained in the other two cases.
}
\label{fig:onevstwo}
\end{figure*}
A prerequisite of this conclusion is that the dimensionality is sufficiently high, {\it i.e.},\ of the order of a thousand components.
It is not valid for low-dimensional index vectors, $n_{\cal D} \lesssim 1000$, because in that domain the average performance decreases.
The details of this performance degradation are presently not understood.
Note that the relative number of correctly decoded features first decreases with an increasing number of encoded features, as expected, but then starts to increase somewhat for $\gtrsim 8$ percent features in the case of two-way RI.
This effect is caused by the increasing probability of obtaining features in the top-list by chance when the relative number of features and the relative length of the top-list increases.
In the case of one-way RI the standard deviation has a maximum around $0.7$--$0.9$ percent features and the nature of this behaviour is not understood.
It may be an effect of chance similar to that described above, but the low value of $\rho$ at the maxima renders such an explanation counter intuitive.
We see no major reason to explore this property in more depth here and therefore leave it as an open issue.
Up to this point the dimension reduction is kept fixed at 4:1.
The next question is how the dimension reduction ratio affects the performance of NRI.
We address this question with an example that is identical to those considered above with the only modification that the dimension reduction, $\Pi_{\cal D} N_{\cal D}$~:~$\Pi_{\cal D} n_{\cal D}$, is varied from 4:1 to 64:1.
The size of the matrix, $N_{\cal D}$, is kept constant when varying the dimension reduction ratio, which implies that the number of features encoded in the classes is constant also.
The result of this simulation is presented in \fig{dimred}.
\begin{figure*}[h]
\centering
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=7.5cm]{dimension_reduction}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=7.5cm]{snr_vs_dimred}
\end{minipage}
\caption{
Effect of the dimension reduction, $\Pi_{\cal D} N_{\cal D}$~:~$\Pi_{\cal D} n_{\cal D}$, on the relative number of correctly decoded features.
The panel on the left-hand side shows the average relative number of correctly decoded features, which is practically independent of matrix size as long as the dimensionality of the index vectors is sufficiently high, see the text for further information.
The panel on the right-hand side shows the signal to noise ratio, defined as the average relative number of correctly decoded features, $\mu$, divided by the corresponding standard deviation, $\sigma$.
Note that this signal to noise ratio refers to the decoded information, while \eqn{snr} refers to the encoded data.
The size of the matrix, $N_{\cal D}$, is taken to be $64,000 \times 64,000$ for both one-way and two-way RI.
At the maximum dimension reduction of 64:1 this corresponds to a state size, $n_{\cal D}$ of $1,000 \times 64,000$ ($8,000 \times 8,000$) for one-way (two-way) RI.
The large size of the matrix is necessary to maintain high dimensionality of the index vectors for one-way RI at high dimension reduction.
The effect of an increasing dimension reduction ratio on the dimensionality of index vectors is lower for two-way RI than for one-way RI, and is increasingly weaker at higher order.
}
\label{fig:dimred}
\end{figure*}
A surprising result of this simulation is that the performance of one-way and two-way RI is comparable at a dimension reduction of 64:1.
The nature of this trend may be related to an interesting scaling phenomena.
Assume that a matrix is square and of size $N \times N$, and that it is encoded with one-way (two-way) RI in a state of size $n \times N$ ($n \times n$).
The dimension reduction ratio, $\xi$, then is $\xi = N/n$ ($\xi = N^2/n^2$) for one-way (two-way) RI.
Solving for the dimensionality of the state matrix, and consequently of the index vectors, we get $n = N/\xi$ for one-way RI and $n = N/\sqrt{\xi}$ for two-way RI.
In general, for any order, $r$, of NRI the dimensionality of the index vectors scales as $n = N \xi^{-1/r}$.
The order of NRI ($r$) therefore effectively reduces the impact of high dimension reduction ratios on the dimensionality of index vectors.
Due to the high computational cost of systematic simulations at high dimension reduction ratios and tensor rank we do not investigate these trends further here, but leave them as an interesting issue for further investigation.
Finally we investigate the dependence of these results on the number of non-zero trits, $\chi_{\cal D}$, in the index vectors.
Up to this point we kept this parameter fixed at $\chi_{\cal D} = 8$, {\it i.e.},\ we used index vectors with four positive and four negative trits.
This is a good compromise, because the performance in the examples considered here increases insignificantly at higher values of $\chi_{\cal D}$, and the computational complexity increases with increasing $\chi_{\cal D}$ since the number of states that needs to be accessed each time a tensor component is encoded or decoded is $\Pi_{\cal D}\chi_{\cal D}$, see \sect{ricoding}.
In \fig{diffk} we illustrate how the average relative number of correctly decoded features varies for different values of $\chi_{\cal D}$ and the relative number of encoded features, $\rho$.
\begin{figure*}[h]
\centering
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=7.5cm]{diffk_one-way}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=7.5cm]{diffk_one-way_std}
\end{minipage}
\vspace{0.3cm} \\
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=7.5cm]{diffk_two-way}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}[b]{7.5cm}
\centering
\includegraphics[width=7.5cm]{diffk_two-way_std}
\end{minipage}
\caption{
The average number of correctly decoded features and the corresponding standard deviation for different numbers of non-zero trits in the index vectors, $\chi_{\cal D}$, and different relative number of encoded features, $\rho \in \{0.1,~0.5,~1,~2,~4,~6,~8,~10\}$ percent.
The matrix considered here has size $5000 \times 5000$ and it is encoded with one-way and two-way RI such that the dimension reduction is 4:1.
Results for $\chi_{\cal D}=8$ are identical to those presented in \fig{onevstwo}.
Refer to that figure and the related text for further information about the $\rho$-dependence of the quantities plotted here.
}
\label{fig:diffk}
\end{figure*}
Illustrated in that figure is also the corresponding standard deviation.
From these results it is clear that there is a significant improvement in performance when increasing $\chi_{\cal D}$ from two to four, and that the improvement thereafter is relatively small.
In two-way RI the number of states associated with one matrix element is $\chi_{1} \times \chi_{2}$ so there is practically a quadratic dependence of the computational complexity on the number of non-zero trits in the index vectors, $\chi_{\cal D}$.
This is the motivation why we settled for $\chi_{\cal D}=8$ in most simulations presented in this section, since it offers practically the same performance as higher values, but at a lower computational cost.
Note the maxima of the standard deviation around $\rho\sim 0.7$ percent in the case of one-way RI, which we noted and commented on also in relation to \fig{onevstwo}.
Here an exception occurs for $\chi_{\cal D}=2$, which does not have such a maxima.
Another surprising feature of these results is the intersection of standard deviations for two-way RI around one percent encoded features.
For low $\rho$ the standard deviation is marginally higher with $\chi_{\cal D}=8$ than $\chi_{\cal D}=4$.
The nature of this detail is not understood.
\section{Application to natural language analysis}
\label{sec:language}
Next we apply this methodology in a statistical study of natural language to further illustrate the NRI approach and some of its properties.
In statistical models of natural language it is common to construct a large word--word or word--document matrix, a so-called {\it co-occurrence} matrix.
This is for example the case in the Hyperspace Analogue to Language (HAL) \cite{lund_1995,lund_1996} and in Latent Semantic Analysis (LSA) \cite{landauer_1997}, which are two pioneering and successful models in this field.
In practical applications the number of words normally exceeds hundreds of thousands.
The number of contexts is necessarily high also, otherwise the statistical information will be insufficient for successful analysis.
Word co-occurrence matrices therefore tend to be large objects, {\it e.g.},\ the relatively simple example considered in the following uses more than 5 billion matrix elements.
Fortunately, co-occurrence matrices are sparse and can therefore be compressed to make the semantic analysis more efficient.
It was demonstrated in \cite{kanerva_2000} that one-way RI can be used to effectively encode co-occurrence matrices for semantic analysis, see also \cite{sahlgren_2005,sahlgren_thesis,kanerva_hyperdimensional_2009}.
The definition of context is model specific, but typically includes a set of words or a document.
In HAL the context is defined by a number of words that immediately surround a given word, while in LSA the context is defined as the document where the word exists.
Linguistically, the former relation can be described as a paradigmatic, {\it i.e.},\ semantic relation, while the latter can be characterized as an associative, {\it i.e.},\ topical relation.
In the traditional RI algorithm, each word type that appears in the data is associated with a context vector, and each context is associated with a ternary index vector.
If the context is defined in terms of the neighbouring words of a given word, which is the strategy that we adopt here, context vectors are created by adding the index vectors of the nearest (typically two or three) preceding and succeeding words every time a word occurs in the data.
If the context is defined as the document where the word exists, context vectors are created by adding the index vectors of all documents where a word occurs, weighted by the frequency of the word in each document.
In either case a context vector is the sum of weighted index vectors of all contexts, defined either by surrounding words or by documents, where that word occurs.
The RI algorithm has traditionally been evaluated using various kinds of vocabulary tests, {\it e.g.},\ the synonymy part of the ``Test of English as a Foreign Language'' (TOEFL)
\cite{kanerva_2000,sahlgren_thesis}.
In the following we revise the synonym identification task presented in \cite{kanerva_2000} with three changes.
First, we want to compare the performance of one-way and two-way RI, so we encode the co-occurrence matrix with both methods.
Second, while Kanerva {\it et al. } used the LSA definition of context, we use a strategy similar to that in HAL and define the context as a window that spans $\pm 2$ words away from the word itself.
This means that for each occurrence of a word there will be four additional word--word correlations encoded in the co-occurrence matrix.
Phrased differently, our co-occurrence matrix is not a word--document matrix, but a word--word matrix where context is defined in terms of word neighbourhood.
This strategy avoids the potential difficulty of defining document (context) boundaries, {\it e.g.},\ in on-line text, and it captures semantic relations between words rather than topical relations.
The length of the context window is a free parameter that affects the quantitative results presented here, but it is not essential for our qualitative discussion.
The third and last difference compared to the study in \cite{kanerva_2000} is that we do not introduce upper and lower limits for the frequencies encoded in the co-occurrence matrix.
High-frequency word-context relations have a negative effect on NRI representations, because any interference (non-orthogonality) between index vectors is amplified by the high weight of the most frequent correlations.
Frequently occurring words like ``the'', ``at'' and ``be'' contribute to extremely high frequencies that render the occurrences of more interesting combinations insignificant, practically by pushing them below the noise threshold.
This effect is stronger for two-way RI than for one-way RI, because in one-way RI the interference affects one dimension only, {\it e.g.},\ contexts interfere but words are kept independent.
In two-way RI both words and contexts interfere and high-frequency word-context combinations needs to be transformed or removed prior to encoding.
The significance of this effect presumably increases with tensor rank, but we have not made simulations to investigate that trend.
Here we include the complete word-context spectrum, including the low- and high-frequency content, and we present results for two different transformations of the spectrum.
In one case we encode the unaltered frequencies directly, and in the other case we take the square root of the frequencies before encoding them.
The effect of the square root is to decrease the significance of high frequencies, and this has a positive effect on the SNR of the RI-encoded co-occurrence matrix.
This choice of transformation was inspired by the qualitative relation $E\sim A^2$ for wave phenomena (where $E$ denotes energy and $A$ amplitude), but it has no obvious connection to that relation and is adopted here without further motivation.
We made some experiments with a logarithmic rescaling of frequencies also, but that seems to be inferior to the square root.
No systematic comparison of transformations was made, and no further efforts were made to optimize the transformation for maximum performance since our focus is on qualitative differences between one-way and two-way RI.
The quantitative results can be further improved with a more careful choice of preprocessing method, {\it e.g.},\ by introducing cuts on the word--word correlation frequencies \cite{karlgren_2001}.
Another example is the preprocessing method used in LSA, where the frequencies are transformed with a logarithm and then divided with the conditional entropy of the context given that the word has occurred \cite{landauer_1997}.
Note, however, that such preprocessing methods invalidate the incremental property of NRI, because the accumulated frequency (weight) must be known before the transformation is made.
This problem can possibly be addressed with a decoding-encoding step when incrementing the matrix elements and/or a more sophisticated pre-processing technique.
We construct the co-occurrence matrix from 37,620 short high-school level articles in the TASA (Touchstone Applied Science Associates, Inc.) corpus
\footnote{The TASA and TOEFL items have been kindly provided by Professor Thomas Landauer, University of Colorado.}.
The text has been morphologically normalized so that each word appears in base form.
It contains about 74,200 word types that are encoded in an co-occurrence matrix with the RITensor software \cite{RITensor}.
For one-way RI we use index vectors of length $1,000$, so that the dimension reduction is $\sim 74,200\times 74,200~:~1,000\times 74,200 \rightarrow 74~:~1$.
In the case of two-way RI we use a state tensor of size $~1,000\times 74,200$, thereby maintaining the same dimension reduction ratio.
We have repeated the two-way calculations with a square state of size $~8,600 \times 8,600$ with similar results.
There are plenty of misspellings (low-frequency words) in the corpus and the most frequent word is ``the'', which occurs nearly 740,000 times.
At second place is ``be'' with just over 420,000 occurrences.
The task consists of eighty TOEFL synonym tests, which contains five words each.
One example of a synonym test considered here is presented in \tab{toefl}.
\begin{table}[h]
\caption{Example of a TOEFL synonym test.
The first word is given and the task is to determine which of the four remaining words that is a synonym of that word.
Illustrated are also the number of occurrences of each word in the TASA (Touchstone Applied Science Associates, Inc.) corpus.}
\vspace{1ex}
\centering
\begin{tabular}{ c | c }
\hline
Word & Number of occurrences \\
\hline
essential (given) & 855 \\
basic & 1920 \\
ordinary & 837 \\
eager & 480 \\
possible & 3348 \\
\hline
\end{tabular}
\label{tab:toefl}
\end{table}
One out of the five words in each test is given and the task is to identify the synonym of that word among the other four words.
There is only one correct synonym in each case, and consequently three incorrect possibilities.
The task to identify the correct synonym is addressed with the NRI-encoded co-occurrence matrix in the following way.
First, the word--word correlation frequencies of the five words in each synonym test are decoded and sorted in descending order (using the function ``find'' in RITensor).
This means that the first (last) word has the highest (lowest) correlation and therefore is most (least) significant.
In practice the non-orthogonality of index vectors cause interference that makes this top-list an approximate one, with decreasing accuracy further down the list.
We therefore select a subset of the high-score word correlations and base the similarity tests on this subset, omitting other word--word correlations with lower decoded weight.
The set of high-score words that correlate with the given word of a synonym test is denoted with $W_0$ and the sets of high-score words that correlate with the four alternative answers are denoted with $W_1$, $W_2$, $W_3$ and $W_4$.
The similarity of two words is measured with the Jaccard index,
\begin{equation}
J(W_0,W_i) = \frac{|W_1\cap W_i|}{|W_1 \cup W_i|},~~~~i\in [1,4].
\label{eq:jaccard}
\end{equation}
This means that two words that share many similar word--word correlations are considered to be similar.
The synonym is identified as the pair of words with the highest value of the Jaccard index out of the four calculated values.
We repeated each test ten times with different sets of index vectors and calculated the average success rate and the standard deviation.
The result is presented in \fig{toefl}.
\begin{figure}[h]
\bigskip
\includegraphics[width=\figurewidth]{toefl_jaccard_avg10runs}
\caption{
Performance of different NRI techniques on TOEFL (Test Of English as a Foreign Language) synonym tests.
Illustrated here is the number of correctly identified synonyms vs. the size of the word--word correlation sets used to calculate the Jaccard index \eqn{jaccard}.
The co-occurrence matrix is encoded with NRI and includes 37,620 short high-school level articles in the TASA (Touchstone Applied Science Associates, Inc.) corpus.
Each TOEFL synonym test consists of five words, out of which two are synonyms.
The task is to identify the synonym of a given word in the set, {\it i.e.},\ there are four alternative answers out of which one is correct.
This implies that an average score of $25$ percent corresponds to random guesses.
These results are based on 80 synonym tests, each comprising five words.
Each test were repeated ten times with new randomly generated index vectors.
Symbols represent average results of the ten simulations, and error bars represent standard deviations.
When performing the synonym similarity test in the traditional way using one-way RI and the cosine of angles between full-length context vectors the result is $47\pm 2.4$ percent correctly identified synonyms, which increases to $51\pm 3.7$ percent when the square root of word frequencies is used.
This is slightly lower than the best results obtained with the Jaccard index test, which are $49\pm 1.7$ and $54\pm 3.9$ percent, respectively.
See the text for further information.
}
\label{fig:toefl}
\end{figure}
From these results it is clear that high-frequency words cause significant interference that has a negative effect on the decoding accuracy.
In particular the two-way result with unmodified word--word correlation frequencies is near the noise threshold of $25$ percent correct synonyms, which is the average score expected from random guesses.
The situation improves significantly when the dominance of high frequencies is reduced with a square-root transformation of the co-occurrence weights before encoding.
On this particular task one-way RI outperforms the two-way method, which is expected because the dimension reduction ratio is identical and the only effect of two-way RI is to introduce additional interference between the words.
One benefit of two-way RI is that more words and contexts can be defined incrementally with a minimum impact on the storage requirements, see \sect{ricoding} for further information.
This property is expected to become increasingly interesting for higher-order tensors, but we leave that issue to a future study since the computational requirements are demanding.
For comparison we calculate also the success rate using the traditional one-way RI approach where word similarity is determined with the cosine of angle between full-length context vectors \cite{kanerva_2000,sahlgren_2005}.
The best result obtained with the Jaccard index comparison is $54\pm 3.9$ percent correct synonyms and the corresponding cosine result is $51\pm 3.7$ percent.
This is an intuitively expected result, because the full-length context vectors include noise that to some extent is excluded in the limited high-frequency lists, and only the significant word--word correlations are meaningful.
It is not clear how to find the optimal length of the top-list, see \fig{toefl}, and we expect that it varies.
One possibility is to further develop the statistical understanding of the top-list, {\it e.g.},\ as illustrated in \fig{toplist} and \fig{toplist_fo}.
This requires an analysis of the fidelity of decoded tensor components, {\it e.g.},\ by generalizing the developments in \cite{SDM_1988}.
\section{Conclusions}
Random indexing (RI) is a useful incremental dimension reduction technique that has been successfully applied for a decade in vector space models of natural language.
It has recently been applied in a number of other contexts and the method is generalized here to matrices and higher-order tensors, which means that the method can be applied to new classes of complex problems.
One example is natural language processing, where the input data can have both high rank and (very) high dimensionality, and where typically only a fraction of the actual feature space is informative and useful.
There are a number of obvious ways to formulate and utilize high-rank input data for natural language processing applications, {\it e.g.},\
by augmenting standard co-occurrence matrices with temporal information ({\it cf.}\ Temporal RI \cite{Jurgens_2009})
or linguistic relations \cite{baroni:lenci,cruys},
or by incorporating structural information in distributed representations \cite{clark:pulman,yeung:tsang}.
However, there have been few attempts at extending traditional matrix-based natural language processing methods to tensors.
One reason for this is the considerable computational costs involved in working with tensors, and this is something that N-way RI (NRI) is likely to facilitate.
The explicit mathematical formulation of NRI that is presented here serves also as a starting point for further theoretical developments, {\it e.g.},\ concerning storage capacity and decoding accuracy.
In particular, the analytical result for the approximate orthogonality of ternary vectors, see \fig{dotp}, is useful for that purpose and has not been published before as far as we know.
Representations of information with NRI, either in the traditional way as vectors or in the form of higher-order tensors, is most accurate when the encoded features are sparse and of comparable magnitude.
The presence of a noisy or otherwise non-significant background of information that makes the representation non-sparse does not pose a problem as long as the magnitude of these components is significantly lower than that of the interesting features.
An amplitude ratio of one order of magnitude can be sufficient if the density of significant features is low.
We note that a low density of features can result in a first-order transition from significant decoded features to noise, and that this transition becomes continuous when the density of features increases or the difference of magnitudes between features and noise is low.
This property is interesting since it can allow for an accurate discrimination between signal and noise in some problems.
The possibility to represent non-sparse tensors in a compact representation that can be incrementally and efficiently encoded is one major advantage of NRI over other methods.
Other appealing features are the possibility to extend the size of the tensor at the insignificant cost of generating new sparse random vectors and that decoding is computationally efficient.
The NRI method depends on high dimensionality of the index vectors and is therefore suitable for complex problems only.
We note that there is an interesting scaling relation between the dimension reduction ratio, $\xi$, the rank, $r$, and size, $N$, of the tensor, and the dimensionality of index vectors, $n$: $n = N \xi^{-1/r}$.
This suggests that high-rank tensors may suffer less from high dimension reduction ratios than vectors and matrices, but that is a matter of further investigation.
In this work we simulate two-way and one-way RI, {\it i.e.},\ representations of matrices and sets of vectors, respectively.
The computationally more demanding task to study the behaviour of higher-order RI is an issue for future work and applications.
Note, however, that the prototype software \cite{RITensor} supports NRI of tensors with any rank.
Our simulation results show that the performance of NRI depends significantly on these parameters: the density of features and their relative magnitude, the dimension reduction ratio and the dimensionality of the index vectors.
The comparison between one-way RI of a set of vectors and two-way RI of matrices shows that the signal to noise ratio of decoded features is higher for one-way RI, when all comparable parameters are equivalent.
This is to be expected, because the two-way algorithm introduces additional interference, practically between the vectors of the one-way method.
The benefit of two-way RI is that the size of the matrix can be extended in both directions by generating additional index vectors.
It would be interesting to know how this trend continues at higher rank of the tensor.
A systematic study of that aspect requires a distributed computing approach and has so far not been attempted by us.
An unexpected result is that the average performance of NRI does not depend significantly on the dimensionality of the index vectors, provided that the dimensionality exceeds a critical threshold of about $n=10^3$.
Below the threshold the average number of correctly decoded features decreases notably with decreasing dimensionality, but above the threshold the average is practically independent of dimensionality.
The standard deviation of the number of correctly decoded features is, however, significantly dependent on the dimensionality.
When increasing the dimensionality of index vectors the standard deviation decreases.
We find that the number of non-zero trits in the index vectors, $\chi_{\cal D}$ ($=2k$), has an effect on the performance but that the effect is smaller than expected from it's impact on the approximate orthogonality of index vectors.
The performance increases notably when increasing $\chi_{\cal D}$ from two to four, but there is apparently no practical reason to go beyond $\chi_{\cal D}=8$.
This indicates that there is a tradeoff between the indifference of index vectors and the magnitude of interference between different tensor components.
Further theoretical work is needed to clarify these findings.
\bibliographystyle{abbrvurl.bst}
|
\section{\@startsection {section}{1}{\z@}%
{-3.5ex \@plus -1ex \@minus -.2ex}%
{2.3ex \@plus.2ex}%
{\noindent\normalsize\bfseries\uppercase}}
\makeatother
\makeatletter
\renewcommand\subsection{\@startsection{subsection}{2}{\z@}%
{-3.25ex\@plus -1ex \@minus -.2ex}%
{1.5ex \@plus .2ex}%
{\noindent\normalsize\bfseries}}
\makeatother
\makeatletter
\renewcommand\subsubsection{\@startsection{subsubsection}{2}{\z@}%
{-3.25ex\@plus -1ex \@minus -.2ex}%
{1.5ex \@plus .2ex}%
{\noindent\normalsize\itshape}}
\makeatother
\let\oldfigure=\figure
\let\endoldfigure=\endfigure
\renewenvironment{figure}{
\begin{oldfigure}
\begin{center}
}{
\end{center}
\end{oldfigure}}
\newenvironment{figure2}{
\begin{figure*}
\begin{center}
}{
\end{center}
\end{figure*}}
\let\oldtable=\table
\let\endoldtable=\endtable
\renewenvironment{table}{
\begin{oldtable}
\begin{center}
\begin{footnotesize}
}{
\end{footnotesize}
\end{center}
\end{oldtable}}
\newenvironment{table2}{
\begin{table*}
\begin{center}
\begin{footnotesize}
}{
\end{footnotesize}
\end{center}
\end{table*}}
\citestyle{aa}
\let\oldthebibliography=\thebibliography
\let\endoldthebibliography=\endthebibliography
\renewenvironment{thebibliography}[1]{
\renewcommand\bibsection{\section*{References}}
\begin{oldthebibliography}{#1}
\setlength{\itemsep}{0pt}
\begin{small}
}{
\end{small}
\end{oldthebibliography}}
\newcommand{Paper~I\xspace}{Paper~I\xspace}
\newcommand{\ensuremath{\mathrm{d}}}{\ensuremath{\mathrm{d}}}
\newcommand{\U}[1]{\ensuremath{\mathrm{~#1}}}
\newcommand{\U{erg}}{\U{erg}}
\newcommand{\U{erg~s}}{\U{erg~s}}
\newcommand{\U{yr}}{\U{yr}}
\newcommand{\U{Myr}}{\U{Myr}}
\newcommand{\U{Gyr}}{\U{Gyr}}
\newcommand{\U{pc}}{\U{pc}}
\newcommand{\U{kpc}}{\U{kpc}}
\newcommand{\U{Mpc}}{\U{Mpc}}
\newcommand{\U{M}_{\odot}}{\U{M}_{\odot}}
\newcommand{\U{km\ s^{-1}}}{\U{km\ s^{-1}}}
\newcommand{H{\sc i} }{H{\sc i} }
\newcommand{H{\sc ii} }{H{\sc ii} }
\newcommand{\reft}[1]{Table~\ref{tab:#1}}
\newcommand{\reff}[1]{Figure~\ref{fig:#1}}
\newcommand{\tn}[1]{[\emph{T.N.}: #1]}
\newcommand{\todo}[1]{\par\noindent$\bullet$ \fbox{\parbox[t]{0.96\columnwidth}{To do: \texttt{#1}}}\par}
\newcommand{\fb}[1]{{\bf{#1}}}
\newcommand{\mc}[1]{\mathcal{#1}}
\newcommand{\e}[1]{\mathrm{e}^{#1}}
\begin{document}
\thispagestyle{empty}
\onecolumn
\noindent{\LARGE\bf On the dynamical evolution of globular clusters\\ II- The isolated cluster}
\vspace{0.7cm}
\noindent{\Large Michel~H\'enon$^{\star}$, translated by Florent~Renaud$^{1,2}$}
\vspace{0.2cm}
{\footnotesize \noindent \it
$^{\star}$ Institut d'Astrophysique, Paris (now at the Observatoire de Nice)\\
$^1$ Observatoire Astronomique, Universit\'e de Strasbourg, 11 rue de l'Universit\'e, F-67000 Strasbourg, France\\
$^2$ Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge, CB3 0HA, UK\\
\phantom{$^2$} \emph{<EMAIL>}}\\
\noindent{\footnotesize Originaly received on October 23, 1964; translated on October 7, 2010}
\vspace{0.5cm}
\begin{flushright}
\begin{minipage}{12cm}
\setlength{\parindent}{15pt}
This paper is an English translation of Michel H\'enon's article, \emph{Sur l'\'evolution dynamique des amas globulaires. II- L'amas isol\'e} originally published in French in the Annales d'Astrophysique, Vol. 28, p.62 (1965).
A translation of the first paper of this series (Paper~I\xspace: \emph{Sur l'\'evolution dynamique des amas globulaires}, H\'enon 1961, Annales d'Astrophysique, Vol. 24, p.369) is also available.
Conventions and notations are as in the original version, for consistency. The English version is written so that it is as faithful to the French text as possible. The translator added some notes [\emph{T.N.}] for the sake of clarity, when required. French, English and Russian abstracts written by M.~H\'enon are available in the original version of the paper. The English part is reproduced below.
FR thanks Michel H\'enon for the enthusiasm and kindness he expressed when he was asked for permission to translate his work, as well as Douglas Heggie and Mark Gieles for their careful proofreading.
\end{minipage}
\end{flushright}
\vspace{0.5cm}
\begin{multicols}{2}
\section*{Original abstract}
The homologic \tn{homologous} model which was defined and studied in an earlier paper \citep{Henon1961} is computed for the case of an isolated cluster. This model is compared to von Hoerner's numerical results \citep{vonHoerner1963}. The agreement is quite good for the density low [\emph{sic} law]; the distribution of energies and the rate of escape. The broad features of the evolution are: tendency towards the homologic model, building up of an infinite central density, formation of close binaries, expansion of the cluster as a whole. An unexplained discrepancy remains in the rate of evolution.
\section{Introduction}
In a previous work \citep[hereafter Paper~I\xspace]{Henon1961}, we have obtained a system of fundamental equations that allows one to calculate the dynamical evolution of a star cluster from a given initial state. We have then calculated and studied in detail a particular solution of this system: the ``homologous model'', which has the property of remaining self-similar in time, the evolution being reduced to the scaling of the physical quantities.
The cluster was supposed to be plunged into an inhomogeneous field of forces, due to the Galaxy. The effect of such a field (often called ``tidal effect'') is to limit the cluster radius to a specific value $r_e$, linked with the mass of the cluster through the relation (Paper~I\xspace, Equation 3.8):
\begin{equation}
r_e \propto \mc{M}_e^{1/3}.
\end{equation}
Every star further than the distance $r_e$ definitely escapes from the cluster. This way, a loss of stars is continuously happening, and the cluster mass decreases with time. From Equation~1, the radius decreases too.
The present paper focusses on the different case of an \emph{isolated cluster}, i.e. when the external forces are zero or negligible. Then, the radius $r_e$ is infinite, and the stars cannot escape anymore\footnote{A more detailed study of this phenomenon is given in \citet{Henon1960}; we find that the escape rate is not exactly zero, but negligible in practice. Similar results have been obtained by \citet{Woolley1962}.} (see Paper~I\xspace, Chapter~III). Hereafter, the mass $\mc{M}_e$ of the cluster is constant over time. All the equations given in Paper~I\xspace are still valid in this case, as soon as $\lambda = \infty$, and one can solve them by use of to numerical computations (see Paper~I\xspace, Chapter~V).
\end{multicols}\twocolumn\noindent
\section{Results}
\subsection{Structure}
Once again, we find that only one unique homologous model exists. The four fundamental functions $\fb{F}(\fb{E})$, $\fb{D}(\fb{U})$, $\fb{R}(\fb{U})$, $\fb{Q}(\fb{E})$ are given in \reft{1} and plotted in \reff{1}. As in Paper~I\xspace, all the variables relative to the homologous model are in bold font; $\fb{E}$ is the energy, $\fb{U}$ is the potential, $\fb{F}$ is the distribution function, $\fb{D}$ the spatial density, $\fb{R}$ the distance to the centre; $\fb{Q}$ is a functional form without particular physical meaning (one can show, in fact, that $\fb{Q}$ is the \emph{adiabatic invariant} associated to the total energy $\fb{E}$, i.e. the volume of phase-space enclosed within the surface $\fb{E} = \mathrm{cst}$; see \citealt{LeviCivita1934}). \reff{1} is quite similar to the analogous figure in Paper~I\xspace, the main difference being that here, $\fb{R}$ goes to infinity when $\fb{U}$ goes to zero, because the cluster has an infinite radius.
\begin{table2}
\caption{The isolated homologous model}
\label{tab:1}
\begin{tabular}{r@{}lr@{}lr@{}lr@{}lr@{}lr@{}l}
\multicolumn{2}{c}{$\fb{E}$ or $\fb{U}$} & \multicolumn{2}{c}{$\fb{F}$} & \multicolumn{2}{c}{$\fb{D}$} & \multicolumn{2}{c}{$\fb{R}$} & \multicolumn{2}{c}{$\fb{Q}$} & \multicolumn{2}{c}{$\fb{D}_p$} \\
\hline
\hline
-5& & 131&.81 & 134&.66 & 0&.1126 & 0&.0002927 & 39&.86\\
-4&.8 & 106&.43 & 106&.45 & 0&.1256 & & \phantom{.000}4025 & 34&.72\\
-4&.6 & 85&.80 & 83&.83 & 0&.1402 & & \phantom{.000}5554 & 30&.13\\
-4&.4 & 69&.05 & 65&.71 & 0&.1566 & & \phantom{.000}7680 & 26&.05\\
-4&.2 & 55&.46 & 51&.26 & 0&.1753 & 0&.001064 & 22&.41\\
-4& & 44&.44 & 39&.75 & 0&.1964 & & \phantom{.00}1479 & 19&.18\\
-3&.8 & 35&.52 & 30&.63 & 0&.2204 & & \phantom{.00}2060 & 16&.33\\
-3&.6 & 28&.30 & 23&.43 & 0&.2478 & & \phantom{.00}2879 & 13&.80\\
-3&.4 & 22&.47 & 17&.76 & 0&.2793 & & \phantom{.00}4039 & 11&.58\\
-3&.2 & 17&.77 & 13&.34 & 0&.3153 & & \phantom{.00}5691 & 9&.629\\
-3& & 13&.98 & 9&.898 & 0&.3571 & & \phantom{.00}8058 & 7&.928\\
-2&.8 & 10&.94 & 7&.247 & 0&.4058 & 0&.01148 & 6&.451\\
-2&.6 & 8&.495 & 5&.220 & 0&.4629 & & \phantom{.0}1647 & 5&.176\\
-2&.4 & 6&.544 & 3&.689 & 0&.5304 & & \phantom{.0}2382 & 4&.085\\
-2&.2 & 4&.989 & 2&.546 & 0&.6110 & & \phantom{.0}3479 & 3&.160\\
-2& & 3&.754 & 1&.707 & 0&.7085 & & \phantom{.0}5141 & 2&.384\\
-1&.8 & 2&.777 & 1&.104 & 0&.8283 & & \phantom{.0}7709 & 1&.744\\
-1&.6 & 2&.010 & 0&.6818 & 0&.9782 & 0&.1177 & 1&.226\\
-1&.4 & 1&.413 & 0&.3958 & 1&.171 & & \phantom{.}1837 & 0&.8170\\
-1&.2 & 0&.9516 & 0&.2112 & 1&.425 & & \phantom{.}2955 & 0&.5060\\
\\
-1& & 0&.6018 & 0&.09978 & 1&.778 & & \phantom{.}4948 & 0&.2816\\
-0&.9 & 0&.4623 & 0&.06429 & 2&.011 & & \phantom{.}6533 & \\
-0&.8 & 0&.3437 & 0&.03904 & 2&.300 & & \phantom{.}8779 & 0&.1324\\
-0&.7 & 0&.2444 & 0&.02191 & 2&.667 & 1&.206 & \\
-0&.6 & 0&.1631 & 0&.01103 & 3&.151 & 1&.706 & 0&.0462\\
-0&.5 & 0&.09898 & 0&.004740 & 3&.822 & 2&.510 & 0&.0227\\
-0&.4 & 0&.05150 & 0&.001584 & 4&.817 & 3&.909 & 0&.00928\\
-0&.3 & 0&.02015 & 0&.000350 & 6&.457 & 6&.66 & 0&.00262\\
-0&.2 & 0&.00371 & 0&.000033 & 9&.710 & 13&.4: & 0&.000346\\
-0&.1 & 0&.00011 & 0&.000000 & 19&.41 & 41&.1: & 0&.000005\\
0& & 0& & 0& & &$\infty$ && $\infty$ & 0\\
\hline
\end{tabular}
\end{table2}
\begin{figure}
\includegraphics[width=\columnwidth]{f1}
\caption{Isolated homologous model: the four fundamental functions $\fb{F}(\fb{E})$, the distribution function; $\fb{D}(\fb{U})$, the spatial density; $\fb{R}(\fb{U})$, the distance to the centre; $\fb{Q}(\fb{E})$, the adiabatic invariant.}
\label{fig:1}
\end{figure}
Below $\fb{E} = -5$ or $\fb{U} = -5$, one can use the asymptotic forms of (4.34) and (4.35) from Paper~I\xspace. Furthermore, for $\fb{E}\to 0$, or $\fb{U}\to 0$, the asymptotic forms are:
\begin{eqnarray}
\fb{F} &=& a_0\ \e{-\sigma'\sqrt{-\fb{E}}},\\
\fb{D} &=& 2 \sqrt{\pi} a_0\ \sigma^{-3/2} (-\fb{U})^{9/4}\ \e{-\sigma'\sqrt{-\fb{U}}},\nonumber\\
\fb{R} &=& -\fb{M}_e / \fb{U},\nonumber\\
\fb{Q} &=& \frac{\pi\ \fb{M}_e^3}{12\sqrt{2}}(-\fb{E})^{-3/2},\nonumber
\end{eqnarray}
with:
\[ a_0 = 6.391, \qquad \sigma = 3.544. \]
The values of constants found in the equations are:
\begin{eqnarray}
K & = & -1.363,\\
c & = & -1.097,\nonumber\\
b & = & +\frac{2}{3}c = -0.7313,\nonumber\\
\fb{M}_e & = & 1.946 \qquad \textrm{(total mass)},\nonumber\\
\fb{L}_e & = & 1.217 \qquad \textrm{(total kinetic energy)}.\nonumber
\end{eqnarray}
The spatial density $\fb{D}$ is plotted in \reff{2} as a function of the radius, and compared to those of the non-isolated model (the two curves are normalized to get the same asymptotical form for $\fb{R}\to 0$). Toward the outskirts, the models are naturally very different, as one yields a finite radius while the other is infinite. However, close to the centre, both models overlap. Therefore, the presence of an external field only modifies the outer layers of the cluster, as one could have expected. To be clearer, we can state that the two models are almost identical within the sphere enclosing half of the total mass (the radius of this sphere, the so-called median (\tn{half-mass}) radius, is $\fb{R} = 0.7563$ for the isolated cluster). In particular, we note that the isolated cluster, like the non-isolated cluster, yields a density that goes to infinity as $\fb{R}^{-2}$ when $\fb{R}$ goes to zero.
\begin{figure}
\includegraphics[width=\columnwidth]{f2}
\caption{spatial density of the isolated homologous model (on top and right) and of the non-isolated model (on bottom and left), as a function of the radius.}
\label{fig:2}
\end{figure}
The last column of \reft{1} gives the projected density $\fb{D}_p$ as a function of $\fb{R}$. It goes to infinity as $\fb{R}^{-1}$ when $\fb{R}$ goes to zero. For $\fb{R} \to \infty$, its asymptotic form is:
\begin{equation}
\fb{D}_p = 4\pi a_0\ \sigma^{-2}\ \fb{M}_e^{5/2}\ \fb{R}^{-3/2} \e{-\sigma \sqrt{\fb{R}/\fb{M}_e}}.
\end{equation}
\subsection{Evolution}
The equations (5.14) of Paper~I\xspace, which give the two parameters $\beta$ and $\gamma$ of the homology as functions of time, become:
\begin{eqnarray}
\beta & = & \beta_0 \left(1 - \frac{T}{T_1}\right)^{-2/3},\\
\gamma & = & \gamma_0 \left(1 - \frac{T}{T_1}\right)^{-1}.\nonumber
\end{eqnarray}
From this, one can derive the variations of the physical quantities as functions of time. In particular, the radius of the cluster evolves as
\begin{equation}
r = r_0 \left(1 - \frac{t}{t_1}\right)^{2/3}.
\end{equation}
Here, $r$ represents any quantity which characterizes the size of the cluster, e.g. the median radius defined above. $t_1$ is negative (see Paper I, 5.13). The variation of $r$ is plotted on \reff{3}. We see that the isolated cluster \emph{expands} with time (while the non-isolated cluster contracts).
\begin{figure}
\includegraphics[width=\columnwidth]{f3}
\caption{Isolated homologous model: evolution of the radius.}
\label{fig:3}
\end{figure}
This result differs from the usual idea according to which internal encounters make the cluster contract. This can be easily explained. The classical reasoning is as follows: the stars which escape from the cluster take with them practically zero energy, so that the total energy of the cluster remains constant; according to the virial theorem, the potential energy also remains constant; as the total number of stars decreases, the distances between them must decrease too. But here, on the one hand there are no escapers because the cluster is isolated. On the other hand, the fact that the parameter $K$ is non-zero and negative (as shown in Paper~I\xspace, Chapter~IV) implies the existence of a \emph{negative flux of energy} toward the centre of the cluster. The phenomenon occurs in the same way as for non-isolated clusters. This negative energy accumulates in the centre, thanks to the formation of increasingly close binaries or multiple stars. At the same time, the energy in the rest of the cluster decreases (in absolute value); inducing an increase of the distance between the stars, i.e. a dilation of the cluster.
The curve on \reff{3} continues to infinity in the future; it is therefore impossible to define a remaining lifetime for the cluster, as opposed to the non-isolated case. In theory, the evolution of the cluster goes on indefinitely. In practice however, while the cluster expands continuously, it will eventually feel the gravitational effect of another object; we are back to the non-isolated case studied in Paper~I\xspace.
The curve in \reff{3} is limited toward small $t$'s. Thus, one can obtain an upper limit equal to $|t_1|$ for the age of the isolated cluster. This is strictly true only when one assumes that the cluster has always followed the homologous model, but is probably true in an approximate fashion in other cases. This limit can be calculated for a real cluster of which one knows the mass and size, thanks to (2.24), (5.13) and (8.18) of Paper~I\xspace; one obtains:
\begin{equation}
|t_1| = \frac{1}{G^{1/2}\ m\ |c|\ \ln{(n)}} \left(\frac{\mc{M}_e}{\fb{M}_e}\right)^{1/2} \left(\frac{r}{\fb{R}}\right)^{3/2}.
\end{equation}
\section{Comparisons with the numerical experiments of von Hoerner}
There is no such thing as a real system with which the present model could be usefully compared. The open clusters and the globular clusters can surely not be considered as isolated; their observed radius is close to the limit set by the galactic field, in general. The elliptical galaxies could, in many cases, be considered as isolated systems to a good approximation; but their internal relaxation time is extremely long, so that their structure is probably not set by \tn{stellar} encounter effects. Finally, for galaxy clusters, the distribution function cannot be precisely determined, in particular in the external regions.
However, the present model can be compared with the two ``experimental clusters'' of \citet[hereafter v.H.]{vonHoerner1963}. These are imaginary clusters made of 25 stars, whose evolution has been numerically computed for a time much longer than the relaxation time. No other forces than mutual attraction are taken into account; therefore, these clusters lie in the isolated case. Furthermore, the masses of the stars are taken to be equal; as in the homologous model.
\subsection{Density law}
The evolution of the spatial distribution of the clusters is plotted in v.H., Figure~5. For the comparison, we consider the final state, the most relaxed, i.e. $t=5.4$ in the first (\tn{experimental cluster}) case and $t=2.9$ in the second case (corresponding to the mean of the fifth and fourth last columns of Table~2 of v.H., respectively). In order to compare these distributions with those of the homologous model, it is necessary to scale $r$ and $\rho$ correctly. These scalings derive from the median radius and the total mass. The two clusters have median radii of 0.840 and 1.006 respectively (from v.H., Table~10) and a total mass of unity; the homologous model has a median radius of 0.7563 and a mass of:
\begin{equation}
\int_0^{\infty} 4\pi \fb{D}\ \fb{R}^2\, \ensuremath{\mathrm{d}}\fb{R} = 4\pi \fb{M}_e = 24.45 \qquad \textrm{(Paper~I\xspace, 2.45)}.
\end{equation}
One finds:
\begin{eqnarray}
\fb{R}/r & = & 0.900 \quad \textrm{or} \quad 0.752;\\
\fb{D}/\rho & = & 33.5 \quad \textrm{or} \quad 57.5.\nonumber
\end{eqnarray}
\reff{4} shows the result. The agreement \tn{between the theoretical and numerical relations} is quite good, especially when considering that there is no free parameter in this comparison, and when noticing that the initial state of the clusters, also visible on \reff{4} was very different from that of the homologous model.
\begin{figure}
\includegraphics[width=\columnwidth]{f4}
\caption{Comparison of the density law: homologous model (solid line) and numerical experiments (\tn{dotted and} dashed lines).}
\label{fig:4}
\end{figure}
\subsection{Energy distribution}
von Hoerner also gives the energy distribution of the stars (v.H., Figure~9 and last column of Table~14). The energy is normalized to its mean value $\overline{\fb{E}}$, and transformed into another quantity $\varepsilon$, for graphical purposes:
\begin{equation}
\varepsilon = \textrm{arcsinh}\left(\frac{\fb{E}}{\overline{\fb{E}}}\right);
\end{equation}
and $f(\varepsilon)\ \ensuremath{\mathrm{d}}\varepsilon$ is the fraction of the stars whose energy is within $\varepsilon$ and $\varepsilon + \ensuremath{\mathrm{d}}\varepsilon$.
To obtain the corresponding curve in the case of the homologous model, one must first calculate the mean energy of a star. It is easy to demonstrate that, for a virialised cluster, this mean energy is minus three times the mean kinetic energy (see v.H., Eq.~23), i.e.:
\begin{equation}
\overline{\fb{E}} = -3 \frac{\fb{L}_e}{\fb{M}_e} = -1.876.
\end{equation}
Furthermore, the fraction of the stars having an energy between $\fb{E}$ and $\fb{E} + \ensuremath{\mathrm{d}}\fb{E}$ is (Paper~I\xspace, 2.42):
\[\frac{\fb{F} \fb{Q}'\ \ensuremath{\mathrm{d}}\fb{E}}{\fb{M}_e}.\]
When using the new variable $\varepsilon$, this expression must be equal to $f(\varepsilon)\ \ensuremath{\mathrm{d}}\varepsilon$. Thus, using (10), one gets:
\begin{equation}
f(\varepsilon) = \frac{\fb{F} \fb{Q}'\ \sqrt{\fb{E}^2+\overline{\fb{E}}^2}}{\fb{M}_e}.
\end{equation}
The distribution obtained is plotted in \reff{5}, and is compared with the results of von Hoerner. Note that, here again, there is no free-parameter. The agreement is excellent from $\varepsilon = 0$ to $\varepsilon = 2.5$, i.e. for the energies ranging from 0 to about 6 times the mean energy. This interval contains almost all the stars (see v.H., Figure~8). In particular, the ``shoulder'' of the distribution next to $\varepsilon = 0$ is very well reproduced. For $\fb{E}$ [\emph{sic}, $\varepsilon$] larger than 2.5, the results differ; the stars with highly negative energies are in excess \tn{in the numerical experiments of von Hoerner} compared to the homologous model. This probably comes from the formation of very close binaries, detected by von Hoerner (his Table~5), and predicted by theory as described above.
\begin{figure}
\includegraphics[width=\columnwidth]{f5}
\caption{Comparison of the energy distribution: homologous model (solid line) and numerical experiments (\tn{dotted and} dashed lines).}
\label{fig:5}
\end{figure}
Let's digress about this topic. One could be surprised that the theory leads to the disagreement illustrated in \reff{5}, although it predicted the existence of binaries. In fact, the homologous model presented here, as well as all the other theoretical models presented up to now, to our knowledge, is fundamentally unable to account for binaries. We assume, indeed, that the state of the cluster is fully described by the distribution function (Paper~I\xspace, 2.1):
\[\varphi(x,y,z,v_x,v_y,v_z,m,t) \quad \textrm{or, in short:}\quad \varphi(\tau,t),\]
which gives the probability to have a star in a small volume of phase-space, at a given time. Doing this, we implicitly assume that the stars are \emph{independent} from each other, meaning that the probability of finding a star at a given point of phase-space does not depend on the presence or absence of stars in other points of phase-space. However, this is only a hypothesis and it is not necessarily true. It seems that this fundamental point has not received all the care it deserves; it is even ignored in the classical stellar dynamics studies.
But this hypothesis contradicts the existence of binaries in the cluster. Indeed, if pairs exist, the presence of a star at a point of (ordinary) space leads to a higher probability to find another star in the neighbourhood. In phase-space, the probability is also modified. To get a correct mathematical description of this effect, we must introduce the \emph{double distribution function}:
\[\varphi_2(\tau_1, \tau_2, t),\]
which gives the probability that a star is at the point $\tau_1$ \emph{and} another star is at the point $\tau_2$, at the time $t$.
The same way, it would be necessary to introduce a triple distribution function $\varphi_3$ to account for triple stars, and so on. These multiple distribution functions are common in plasma physics \citep[see e.g.][Chapter~6]{Delcroix1963} and it would be interesting to introduce them in stellar dynamics too. This would perhaps allow us to calculate exactly, e.g., the formation rate and the properties of binaries in a cluster. We could also expect from this a more precise theory about the effect of encounters.
\subsection{Evolution}
In Paper~I\xspace, Chapter~VII, we have shown that the homologous model is the final state toward which a cluster tends, whatever its initial form was. This is clearly confirmed by the results of von Hoerner (v.H., Figures~5 and 6): his clusters, which initially have a constant density inside a sphere, progressively concentrate in the centre, and form a very extended halo at the same time; the density law tends toward the theoretical curve of the homologous model (\reff{4}).
We have also shown that in a cluster which has initially a finite central density, this density rapidly increases and tends toward infinity. This phenomenon is clearly visible in the Figure~4 of von Hoener, where the ``density radius'' $\mathrm{R}_\rho$ tends to zero. It is easy to verify (v.H., Eq.~10) that this radius can become zero only if the central density is infinite. One can also directly see the rapid growth of the central density (v.H., Table~9).
Finally, we have seen above that the isolated homologous model expands with time (\reff{3}). This also appears on the right-hand side of the Figure~6 of von Hoerner: once the clusters have reached the homologous state, the evolution is only a dilatation of the entire cluster, which is seen on the figure as a translation of all the points to the top. This is especially visible in the second case.
Now, let's make a quantitative comparison. We first have to adapt the timescales. We have (Paper~I\xspace, Eq.~8.9):
\begin{equation}
\frac{\ensuremath{\mathrm{d}} t}{\ensuremath{\mathrm{d}}\fb{T}} = G^{-1/2}\ m^{-1}\ [\ln{(n)}]^{-1} \left(\frac{\mc{M}_e}{\fb{M}_e}\right)^{1/2} \left(\frac{r}{\fb{R}}\right)^{3/2};
\end{equation}
$\fb{T}$ being the time variable of the homologous model, while $t$ is the real time. We have $\mc{M}_e = mn$; and when using the definition of the unit of time given by von Hoerner (v.H., p.54), we get:
\begin{equation}
\frac{\ensuremath{\mathrm{d}} t}{\ensuremath{\mathrm{d}}\fb{T}} = n^{1/2}\ [\ln{(n)}]^{-1}\ \fb{M}_e^{-1/2} \left(\frac{r}{\fb{R}}\right)^{3/2}.
\end{equation}
The $r/\fb{R}$ ratio does not vary much with time (v.H., Table~10); we can assume it remains constant and adopt the values given by (9). With $n=25$, we get:
\begin{equation}
\frac{\ensuremath{\mathrm{d}} t}{\ensuremath{\mathrm{d}}\fb{T}} = 1.303 \quad \textrm{or} \quad 1.707.
\end{equation}
From Paper~I\xspace, (2.32) and (5.13), we find out:
\begin{equation}
t_1 = -1.188 \quad \textrm{or}\quad -1.556,
\end{equation}
and (6) gives the theoretical rate of change of the radius, at $t=0$:
\begin{equation}
\frac{\ensuremath{\mathrm{d}}\log{(r)}}{\ensuremath{\mathrm{d}} t} = 0.244 \quad \textrm{or}\quad 0.186.
\end{equation}
The observed rate of change for the median radius is (v.H., Table~10; we used the mean value, over the entire evolution):
\begin{equation}
\frac{\ensuremath{\mathrm{d}}\log{(r_\mathrm{med})}}{\ensuremath{\mathrm{d}} t} = 0.026 \quad \textrm{or}\quad 0.031.
\end{equation}
Here, we find a clear disagreement: the two clusters evolve much slower than predicted by theory, by a factor of 10 or 6, approximatively.
We can also calculate the time required to form an infinite central density. Back to the calculation made in Paper~I\xspace, Chapter~VII, and assuming that the parameter $d$ takes the same value here, we find that a isolated cluster, starting from a weakly concentrated state, must get an infinite central density after a time:
\begin{equation}
\fb{T}_2= 0.3052,
\end{equation}
i.e.
\begin{equation}
t_2 = 0.473 \quad \textrm{or}\quad 0.620,
\end{equation}
from Paper~I\xspace, (5.12), (2.32) and (15) above.
But the Figure~4 of von Hoerner shows that the infinite central density appears after about $t = 6$ for the first cluster, and $t = 4$ for the second. Here again, the observed evolution rate is smaller than the theoretical one, by a factor 13 or 6.
The reason for such a disagreement has not been found.
\subsection{Escape rate}
As mentioned above, the escape rate of an isolated cluster is much smaller than that of a non-isolated cluster, but is not exactly zero. It reads \citep[Eq.~15]{Henon1960}:
\begin{equation}
\frac{\ensuremath{\mathrm{d}} n}{\ensuremath{\mathrm{d}} t} = -0.00868 \sqrt{\frac{Gmn}{r_0^3}},
\end{equation}
where $r_0$ is the radius enclosing, in projection, the half of the total mass.
The clusters from von Hoerner have initially the shape of an homogeneous sphere of radius unity; in this case, we have $r_0 = 0.6083$. Furthermore, $n=25$. Taking into account the choice for the unit of time (\tn{v.H.,} p. 54), we obtain:
\begin{equation}
\frac{\ensuremath{\mathrm{d}} n}{\ensuremath{\mathrm{d}} t} = -0.0915.
\end{equation}
The total duration of the two examples calculated with $n=25$ is (v.H., Table~8):
\[\Delta t = 6.23 + 3.86 = 10.09;\]
so that the number of escapers should be $\Delta n = -0.92$. In fact, we measure: $\Delta n = -1$ (v.H., Table~6).
Four other examples have been calculated with $n=16$. Their duration is not given by von Hoerner. However, the total number of relaxation times covered is 88, instead of 64 for the examples with $n=25$. The number of escapers should be larger in proportion, i.e. $\Delta n = -1.27$. The number measured is: $\Delta n = -1$. We note that here, the agreement between the theory and the numerical experiments is very good.
\section{Conclusion}
We have compared the results obtained for the same problem: structure and evolution of an isolated cluster, thanks to two very different methods, almost opposite: the first one (the homologous model) describes the cluster with a continuous distribution function, whose evolution is ruled by a Fokker-Planck type equation; the second one (numerical experiments) represents the cluster as a discrete system made of material points, whose evolution is given by the ordinary laws of mechanics.
The results are generally in very good agreement. This constitutes a solid argument in favor of the validity of both theoretical methods.
Some hypotheses are also confirmed. In the case of the homologous model, we had to assume an isotropic distribution for the velocities everywhere in space, in order to reduce the complexity of the equations. No restriction of this kind exists in the numerical experiments, where the entire evolution of the system of points is computed exactly. The agreement of the results, particularly for the density law (\reff{4}), indicates that this hypothesis about isotropy does not strongly affect the structure of the cluster. We can reasonably suppose that the same applies to the non-isolated clusters, studied in Paper~I\xspace.
Furthermore, in the case of the numerical experiments, one could ask whether a system of only 25 points can properly represent a real cluster, made of a much larger number of stars. But the homologous model, with its continuous distribution function, can be considered as the limiting case of a system made of an infinite number of points. The similar results obtained in both cases are thus, very probably, also valid for all the intermediate cases, i.e. from $n=25$ to $n=\infty$; this range includes all the real stellar systems, which are thus ``surrounded'' by the two methods. We can also claim that, for the questions we address, 25 is a good approximation of infinity.
Therefore, it seems recommended to develop further the method of the numerical experiments, which is still in its early stages. Addressing simultaneously the questions of stellar dynamics with theory and numerical experiments would, certainly, allow us to make more solid and faster progresses than in the past.
|
\section{Introduction}
\subsection{Upper limits}
In general, an upper limit is a probabilistic statement bounding one
of several unknown parameters determining the observed data at hand.
While it would be hard to derive general properties applicable in any
possible data analysis context, we will for the illustration purpose
consider a simple case here: a sinusoidal signal in white Gaussian
noise. This example exhibits many similarities with commonly
encountered real-world problems, including the use of Fourier methods,
nuisance parameters, trials factors, partly analytical and numerical
analysis, etc., and we believe is general enough to yield valuable
insights.
\subsection{The frequentist case}
The frequentist detection approach is based on some \textsl{detection
statistic}~$d$, which for given data is then used to derive a
significance statement along the lines of \textsl{``If the data were
only noise (null hypothesis~$H_0$), a detection statistic value~$\geq
d_0$ would have been observed with probability~$p$.''} ($\prob(d\geq
d_0\,|\,H_0)=p$). The probability~$p$ here is the p-value, and a low
p-value is associated with a great significance. In the case of a
non-detection, the statement then may be reversed to an upper limit
statement \textsl{``Had the signal amplitude been $\geq A^\star$, a
larger detection statistic value ($\geq d_0$) would have been observed
with at least 90\% probability''} ($\prob(d\geq d_0\,|\, A\geq
A^\star)\geq 90\%$), where $A^\star$ is the 90\% confidence upper
limit (e.g.\ \cite{AbbottEtAl2004c,BradyCreightonWiseman2004}).
\subsection{The Bayesian case}
In the Bayesian framework, detection and parameter estimation are more
separate problems; for detection purposes one would need to derive the
\textsl{marginal likelihood}, or \textsl{Bayes factor}, which (in
conjunction with the prior probabilities for the ``signal'' and
``noise~only'' hypotheses $H_1$ and $H_0$) allows one to derive the
probability for the presence of a signal. The detection statement
would then be \textsl{``(Given the observed data~$y$,) the probability
for the presence of a signal is~$p$.''} ($\prob(H_1|y)=p$). The upper
limit statement on the other hand is a matter of parameter estimation;
given the joint posterior distribution of all unknowns in the model,
one would need to marginalize to get the posterior distribution of the
parameter of interest alone. The upper limit statement would then be
``\textsl{(Given the observed data and the presence of a signal,) the
amplitude is $\leq A^\star$ with 90\% probability.''}
($\prob(A\leq A^\star\,|\, y,H_1)=90\%$)
\cite{BDA,AbbottEtAl2010a}.
\section{The data model}
We assume the data~$y$ to be a time series given by a parameterized
signal~$s$ and additive noise~$n$:
\begin{equation}
y(t_i) \; = \; s(t_i) + n(t_i),
\end{equation}
where $i=1,\ldots,N$ and $t_i=i\Delta_t$.
The (sinusoidal) signal is given by
\begin{equation}
s(t) \;=\; A\,\sin(2\pi f t + \phi),
\end{equation}
where $A\geq0$ is the amplitude, $0\leq\phi<2\pi$ is the phase, and
$f \in \{\frac{j_1}{N\Delta_t},\ldots,\frac{j_k}{N\Delta_t}\}$ is
the frequency, where $1\leq j_1,\ldots,j_k \leq \frac{N}{2}-1$ defines
the range of possible (Fourier) frequencies.
The number $k$ of frequency bins may be varied and constitutes the
so-called ``trials factor'' here. The noise~$n$ is assumed to be
white and Gaussian with variance~$\sigma^2$.
\section{Frequentist approach}\label{sec:FreqentistApproach}
If there were no unknown parameters in the signal model, then,
following from the Neyman-Pearson lemma, the optimal detection
statistic would be given by the \textsl{likelihood ratio} of the two
hypotheses. In the case that the hypotheses include unknowns
(composite hypotheses) as in our case, this is commonly treated using
the \textsl{generalized likelihood ratio} framework, that is, by
considering the ratio of \textsl{maximized} likelihoods, where
maximization is done over the unknown parameters \cite{MGB}.
In our case, we have a 3-dimensional parameter space under the signal
model. The conditional likelihood for a given frequency may be
maximized analytically over phase and amplitude. The \emph{profile
likelihood} (maximized conditional likelihood for given frequency, as
a function of frequency) is eventually proportional to the time
series' periodogram. The generalized likelihood ratio detection
statistic then is given as the periodogram maximized over the
frequency range of interest:
\begin{equation}\label{eqn:DetStatDefinition}
d^2 \;:=\; \max_j \, {\textstyle \frac{2}{N\sigma^2}} \big|\tilde{y}_j\big|^2
\end{equation}
where $\tilde{y}_j$ is the (complex valued) $j$th element of the
discretely Fourier transformed time series~$y$. The
``$\frac{2}{N\sigma^2} \big|\tilde{y}_j\big|^2$'' term (the
periodogram) maximized over in~(\ref{eqn:DetStatDefinition}) is in
fact also the \textsl{matched filter} for a sinusoidal signal
\cite{WainsteinZubakov}, and the maximum~$d^2$ is commonly referred to
as the ``loudest event'' \cite{BradyCreightonWiseman2004}.
The detection statistic's distribution may be derived analytically
under both hypotheses~$H_0$ and $H_1$, as this is a particular case of
an \textsl{extreme value statistic} \cite{MGB}. Under the null
hypothesis, $d^2$ is the maxi\-mum of $k$ independently
$\chi_2^2$-distributed random variables; the cumulative distribution
function (CDF) of~$d^2$ is given by
\begin{equation}\label{eqn:DetStatDistnH0}
F_{d^2;H_0}(x) \;=\; \prob(d^2\leq x \,|\, H_0) \;=\; \bigl(F_{\chi_2^2}(x)\bigr)^k
\end{equation}
where $F_{\chi_2^2}$ is the CDF of a $\chi_2^2$ distribution, and $k$
again is the number of independent frequency bins, or ``trials''.
This is essentially the ``background distribution'' of~$d^2$.
Under the signal hypothesis~$H_1$, $d^2$~is the maximum of $(k\!-\!1)$
independently $\chi_2^2$-distributed random variables \emph{and} one
noncentral-$\chi_2^2(\lambda)$-distributed variable with noncentrality
parameter~$\lambda=\frac{N}{2\sigma^2}A^2$. The corresponding CDF
under $H_1$ then is
\begin{equation}\label{eqn:DetStatDistnH1}
F_{d^2;H_1}(x) \;=\; \bigl(F_{\chi_2^2}(x)\bigr)^{(k-1)} \times F_{\chi_{2,\lambda}^2}(x)
\end{equation}
where $F_{\chi_{2,\lambda}^2}$ is the CDF of a noncentral $\chi_2^2$
distribution with parameter~$\lambda$.
For some observed detection statistic value~$d^2_0$, the (detection)
significance is determined by the p\mbox{-}value $\prob(d^2 \geq d^2_0
\,|\,H_0) = \int_{d^2_0}^\infty p(d^2|H_0)\,\differential d^2$. The
90\% loudest-event upper limit is given by the smallest amplitude
value~$A^\star$ for which $\int_{d^2_0}^\infty
p(d^2\,|\,A,H_1)\,\differential d^2 \geq 90\%$, so that $\prob(d^2\geq
d^2_0 \,|\, A \geq A^\star,H_1)\geq 90\%$.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{roever_final_fig1}
\caption{The integrals to be computed for a frequentist and a
Bayesian 90\%~upper limit are very different. The Bayesian
integral is computed along the vertical amplitude axis,
conditioning on the observed detection statistic
value~$d^2=d^2_0$. The frequentist integral goes along the
horizontal axis of possible realisations of~$d^2$ for any
given amplitude. (Example values here: $N=100$,
$\Delta_t=1$, $\sigma^2=1$, $k=49$, $d^2_0=11$.)}
\label{fig:integration}
\end{figure}
\section{Bayesian approach}
We assume uniform prior distributions on phase, frequency, and
amplitude. Given the (3-dimensional) likelihood function
\cite{Bretthorst}, one can then derive joint and marginal posterior
distributions $\prob(A,\phi,f \,|\, y)$ and $\prob(A|y)$. However,
Monte Carlo simulations show that --- in this particular model --- the
amplitude's marginal posterior distribution is virtually unaffected by
whether one considers the complete data~$y$, or only the ``loudest
event''~$d^2$. The essential information about the signal amplitude
is contained in that loudest event, and the marginal amplitude
posterior is dominated by the conditional distribution of the loudest
frequency bin. We find that the main difference between the two kinds
of limits in this model is \textsl{not} due to maximization vs.\
integration of the posterior; in the following we will therefore
consider only the simpler, directly comparable, and more illustrative
case of a Bayesian loudest event limit based on $\prob(A|d^2)$ instead
of $\prob(A|y)$.
Our relevant observable now is the ``loudest event''~$d^2$.
The likelihood function $\prob(d^2|A)$ was defined
through~(\ref{eqn:DetStatDistnH1}) in the previous section. The 90\%
upper limit on the amplitude is given by the amplitude~$A^\star$ for
which $\int_0^{A^\star}p(A\,|\,d^2,H_1) \, \differential A = 90\%$,
so that $\prob(A < A^\star\,|\,d^2,H_1)=90\%$.
\section{Comparison}
The likelihood function here is a function of two parameters: the
observable~$d^2$ and the amplitude parameter~$A$\@. Since the
amplitude prior is assumed uniform, the posterior distribution is
simply proportional to the likelihood, which allows for a nice
comparison of both approaches. \Fref{fig:integration} illustrates the
integrations performed for both the frequentist and the Bayesian upper
limits for some particular realisation~$d^2=d^2_0$.
Since the data~$y$ are reduced to a single observable~$d^2$, there
also is a one-to-one mapping from $d^2$ to the upper limit~$A^\star$.
\Fref{fig:mapping} shows both resulting upper limits as a function of
the ``loudest event''~$d^2$.
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{roever_final_fig2}
\caption{The mapping from observable~$d^2$ to the upper limit on amplitude.
The bottom panel shows the ``background'' distribution of $d^2$ under $H_0$.
(Example values here: $N=100$, $\Delta_t=1$, $\sigma^2=1$, $k=49$.)}
\label{fig:mapping}
\end{figure}
An important feature to note is that the frequentist limit will be
\textsl{zero} for certain values of~$d^2$. The point at (and below)
which this happens is the lower 10\% quantile of the
``background'' distribution
of~$d^2$ under~$H_0$~(\ref{eqn:DetStatDistnH0}) --- at this point the
probability of observing a larger $d^2$ value is (by definition) 90\%
for zero-amplitude signals already, which makes zero the 90\% upper
limit. Note that this implies that if $H_0$ in fact is true, 10\% of
all 90\% upper limits will be zero. Note also that this is consistent
with the intended 90\% \textsl{coverage} of frequentist confidence
bounds --- if the upper limit is supposed to fall above and below the
true amplitude value with 90\% and 10\% probabilities respectively,
then 10\% of the upper limits \textsl{must} be zero under~$H_0$.
Having the distribution of the detection statistic
(equations~(\ref{eqn:DetStatDistnH0}), (\ref{eqn:DetStatDistnH1})) and
the mapping from $d^2$ to upper limit (\Fref{fig:mapping}) allows us
to derive the distribution of upper limits for given parameters.
\Figure[b]~\ref{fig:limitdistns} illustrates the behaviour of the
resulting upper limits for different values of amplitude~$A$ and
trials factor~$k$. The left panel shows that for large amplitudes the
two limits behave roughly the same, as one could already see from
\Fref{fig:mapping}, while for low amplitudes the posterior upper limit
will level off and will not rule out amplitude values below a certain
noise level. The frequentist limit's distribution on the other hand
reaches all the way down to zero, and in particular the 90\%~limit's
10\%~quantile follows a straight line of slope~1 and intercept~0 ---
the frequentist 90\%~limit is (by construction) essentially a
statistic that has its 10\%~quantile at the true amplitude value.
The right panel of \Fref{fig:limitdistns} shows the differing
behaviour of both limits as a function of the trials factor~$k$ when
the true amplitude is zero. The frequentist limit's 10\% quantile
remains at zero (the true value), while the posterior limit is bounded
away from zero but otherwise tends to yield tighter constraints on the
amplitude, especially for large~$k$.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{roever_final_fig3}
\caption{The distribution of upper limits as a function of amplitude
(left panel) and trials factor (for zero amplitude; right
panel). Note that the frequentist 90\% limit is
essentially a statistic that is designed to have its
10\%~quantile at the true amplitude value.}
\label{fig:limitdistns}
\end{figure}
\section{Conclusions}
The most obvious technical difference between frequentist vs.\
Bayesian upper limits is in maximization vs.\ integration over
parameter space. This, however, is not --- at least in the example
discussed here --- the primary origin of discrepancies between the
two. When founding \textsl{both} limits on maximization (i.e., the
``loudest event''), the behaviour of the Bayesian limit is affected
very little; so the crucial information about the signal amplitude is
in fact contained in the loudest event. Both kinds of upper limits
behave very similarly for ``loud'' signals, i.e., a large
signal-to-noise ratio (SNR), but their differences become apparent in
the interesting case of (near\hbox{-}) zero amplitude signals. While
the Bayesian upper limit expresses what amplitude values may be ruled
out with 90\% certainty based on the data (and model assumptions), the
frequentist upper confidence limit is defined solely through its
``coverage'' property. The frequentist 90\%~limit needs to end up
above and below the true amplitude value with 90\% and 10\%
probability respectively, which simply means that the frequentist
limit may be any random variable that has its 10\% quantile at the
true amplitude. This in particular implies that for a true amplitude
of $A=0$ the limit has a 10\% chance of being zero as well, and it
makes the frequentist limit very hard to actually interpret, not only
if it actually happens to turn out as zero. When considering the
effect of the trials factor (or look-elsewhere effect) in the low-SNR
regime where both limits behave differently, the posterior-based limit
will usually yield tighter constraints especially for large trials
factors, but it will never be zero.
The Bayesian upper limit based on the amplitude's posterior
distribution will of course change with changing prior
assumptions. For simplicity, we assumed an (improper) uniform
amplitude prior here, but this should actually be a conservative
choice in some sense, for a realistic prior in the continuous
gravitational-wave context would in general be much more concentrated
towards low amplitude values (something like the --- also improper ---
prior with density $p(A)\propto \frac{1}{A^4}$).
Another question is how exactly one would do the actual computations
for a Bayesian upper limit in practice --- the frequentist upper
limits are usually not computed via direct analytical or numerical
integration of the likelihood, but the integral (see
\Fref{fig:integration}) is determined in a nonparametric fashion via
Monte Carlo integration and bootstrapping of the data. While the
frequentist limit requires finding the amplitude~$A^\star$ at which
the integral ($\prob(d^2>d^2_0\,|\,A=A^\star)$) yields the desired
confidence level, an analogous procedure to derive the Bayesian upper
limit would probably require Monte Carlo sampling of $\prob(d^2|A)$
across the range of all amplitudes~$A$ in order to then do the
integral in the orthogonal direction.
Further complications arise especially for the frequentist limit when
the signal model gets more complex. The general procedure required
for the Bayesian upper limit is rather obvious --- determine the
marginal posterior distribution of amplitude~$\prob(A|y)$, then
determine the 90\%~quantile. The frequentist procedure on the other
hand may run into major problems. For example, if there are multiple
parameters affecting the signal's SNR, a ``loudest event'' might be
hard to define, or to translate into a constraint on the amplitude. As
there may not be a simple one-to-one connection between SNR and
amplitude parameter as in the present case, the ``loudest event'' may
not be the only relevant figure to constrain the signal amplitude.
The consideration of nuisance parameters is generally tricky in a frequentist
framework and may effectively suggest the use of a Bayesian procedure instead
\cite{Searle2008}.
Computation also becomes more complicated if the frequency parameter
is not restricted to (``independent'') Fourier frequencies. Note that
the reasoning behind the generalized likelihood ratio approach (see
Sec.~\ref{sec:FreqentistApproach}) leading to the ``loudest event''
concept was very much an ad-hoc construction in the first place.
Another notable related concept is that of a \textsl{power constrained
upper limit}. In search experiments, these may be based on the
\textsl{sensitivity} of the search procedure. In case the search
yielded no detection, one can state the signal amplitude that would
have been detected with 90\%~probability; this number may then also be
used as a lower bound on the frequentist limit (\textsl{``don't rule
out what you wouldn't be able to detect''}). However, this kind of
statement requires the specification of another, additional parameter: the
corresponding false alarm rate defining the threshold of what is
considered a ``detection'', and as such is inseparably connected to
the detection procedure (see also \Fref{fig:sensitivity}).
In particle physics a different approach is commonly taken; here the
sensitivity is usually specified as the expected upper limit for many
repetitions of the experiment in the absence of a signal. This figure
would correspond to the solid lines at zero amplitude in
Fig.~\ref{fig:limitdistns}.
An important point to note is that both these sensitivity statements
do not depend on the observed data.
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{roever_final_fig4}
\caption{Illustration of the determination of a
90\%~\textsl{detection sensitivity} threshold.
Such a statement would be independent of the observed data,
and it requires the specification of an additional
parameter: the corresponding false alarm rate defining the
threshold of what is considered a ``detection''. (Here:
$N=100$, $\Delta_t=1$, $\sigma^2=1$, $k=49$.)}
\label{fig:sensitivity}
\end{figure}
\bibliographystyle{unsrt}
|
\section{Level statistics of elliptic billiard}
\begin{figure}[htp]
\begin{center}
\includegraphics[width=2.1in]{Images/Ps.EB.labelled.eps}
\includegraphics[width=2.1in]{Images/variance.EB.e3000.d5.labelled.eps}
\includegraphics[width=2.1in]{Images/rigidityWidth.EB.labelled.eps}
\includegraphics[width=2.1in]{Images/rigidity.EB.labelled.eps}
\end{center}
\caption{\label{fig:EB} EB's $P(s)$ (a), $\Sigma$ (b), $\Delta_3(\epsilon, E)$ saturation with $E$ (c), long-range oscillations of $\Delta_3^\infty(\epsilon)$ with energy interval width (d), calculated by averaging over an ensemble of 1350 samples of $\sigma$. (a) Black dots: numerical result. Red line: Poisson distribution. Inset: shape of quarter EB. (b) Black line: numerical result. Green line: theoretical result calculated by averaging over variance calculated from 200 shortest POs. (c) Black line: numerical result. Red line: saturated value of $\Delta_3(\epsilon,E)$. (d) Black line: saturated $\Delta_3^\infty(\epsilon)$. Red line: the scaling relation in Eq. \ref{eq:scaling_relation}.}
\end{figure}
The boundary of elliptic billiard (EB) is defined as $x^2/a^2 + y^2/b^2 = 1$ with aspect ratio $\sigma = b/a$ and $ab = 1$ \cite{ayant87.I, ayant87.II, nakamura88, waalkens97}. There are four symmetry classes of eigenvalues and eigenfunctions according to the reflection symmetry of eigenfunctions against the $x$ and $y$ axes. We study the level statistics of the odd-odd class, which implies a quarter EB with Dirichlet boundary condition. The eigenvalues are computed by direct diagonalization of the Hamiltonian \cite{ayant87.I, ayant87.II, nakamura88}.
There are two kinds of averaging used for analysis of statistics in quantum chaos: ensemble averaging, e.g., over realization of disorder, and spectral averaging \cite{taoma.level.repulsion}. For integrable systems, only spectral averaging had been done before Refs. \cite{wickramasinghe05, wickramasinghe08}, in which the system parameter is sampled from algebraic numbers around the central value. A better way of sampling is sampling from a normal distribution \cite{taoma.modified.Kepler}. In this article, for EB $\sigma$ is sampled from a normal distribution centered around $1/2$ and for circular billiard (CB) around 1.
The nearest neighbor spacing distribution $P(s)$ with the level spacing $s$, level number variance $\Sigma(\epsilon, E)$ with energy $\epsilon$ and energy interval width $E$, and spectral rigidity $\Delta_3(\epsilon, E)$ are shown in Fig. \ref{fig:EB}. $P(s)$ generally follows Poisson distribution, but displays deviation at small $s$. This is a general phenomenon showing level repulsion of integrable systems \cite{taoma.level.repulsion}.
The oscillations of $\Sigma(\epsilon, E)$ are explained by the semiclassical equation
\begin{equation}
\Sigma(\epsilon, E)
= \sum_{j} \frac{8A_{j}^2}{\hbar^{N-1}T_{j}^2}\sin^2 \left(\frac{ET_{j}}{2\hbar}\right),
\end{equation}
where $j$ counts periodic orbits (PO), $\hbar$ is the Planck constant, $A_{j}$, $T_{j}$ amplitude and period of PO respectively, $2N$ the dimension of phase space \cite{wickramasinghe08}. Like rectangular billiard (RB), the match between numerical and theoretical results is achieved only after ensemble averaging in theoretical calculation \cite{taoma.modified.Kepler}. Spectral rigidity saturates. Like RB, $\Delta_3$ of EB generally follows the squared-energy scaling relation \cite{taoma.modified.Kepler}:
\begin{equation}\label{eq:scaling_relation}
\Delta_3^\infty(\epsilon) \propto \sqrt\epsilon ,
\end{equation}
but displays long-range oscillations.
All the statistics support that EB is a generic integrable system.
\subsection{spectral rigidity and global variance}
\begin{figure}[htp]
\begin{center}
\includegraphics[width=2.5in]{Images/globalVariance.RB.eps} \\ \includegraphics[width=2.5in]{Images/globalVariance.CB.eps} \\ \includegraphics[width=2.5in]{Images/globalVariance.EB.eps}
\end{center}
\caption{\label{fig:globalVariance} Comparison between spectral rigidity and global variance. Red line: $\Delta_3^\infty(\epsilon)$. Blue line: $\Sigma_g(\epsilon)$ of RB, quarter EB, and quarter CB. The ensemble size of RB calculating $\Sigma_g$ is $10^5$. The oscillation of $\Sigma_g(\epsilon)$ is unlikely to disappear in the limit of infinite ensemble size.}
\end{figure}
The long-range oscillations of spectral rigidity show some global oscillations of level density. To reveal this global oscillations, we investigate a special case of level number variance, the global variance defined as
\begin{equation}\label{eq:global_variance}
\Sigma_g( \epsilon )
\equiv \langle [\delta\mathscr{N}(\epsilon) ]^2 \rangle
\equiv \langle [\mathscr{N}(\epsilon) - \langle\mathscr{N}(\epsilon)\rangle ]^2 \rangle,
\end{equation}
where $\mathscr{N}(\epsilon)$ is the spectral staircase. Global variance is also a special case of the correlation function of spectral staircase defined as $\langle \delta\mathscr{N}(\epsilon_1) \delta\mathscr{N}(\epsilon_2) \rangle$ \cite{serota.correlation.function}.
It is proved that the the rigidity and global variance are equal: \cite{serota.correlation.function, wickramasinghe08}
\begin{equation}\label{eq:global_variance_rigidity_relation}
\Sigma_g(\epsilon) = \sum_{j}\frac{2A_{j}^2}{\hbar^{N-1}T_{j}^2} = \Delta_3 .
\end{equation}
The above equation is partially correct as $\Sigma_g$ fluctuates around $\Delta_3$ for RB, CB and EB as demonstrated in Fig. \ref{fig:globalVariance}. For RB, $\Sigma_g$ and $\Delta_3$ follow the squared energy scaling relation. For CB or EB, the long-range oscillations of $\Sigma_g$ and $\Delta_3$ are ``synchronized''. The scales of oscillations of $\Sigma(\epsilon,E)$ and $\Delta_3^\infty(\epsilon)$ are different. The former is $E \sim \sqrt{\epsilon}$, and the latter is longer than this.
We think the long-range oscillations of $\Delta_3$ and $\Sigma_g$ originate from collective effect of type-R POs of EB or CB and exist in any billiard with whispering gallery modes. The whispering gallery POs have approximately equal orbit length or multiples of that and hence they coherently cause global fluctuation of level density.
\section{Periodic orbits of elliptic billiard}
\begin{figure}[htp]
\begin{center}
\includegraphics[width=1.5in]{Images/orbitsR31.eps}
\includegraphics[width=1.5in]{Images/orbitsR41.eps}
\includegraphics[width=1.5in]{Images/orbitsR52.eps}
\includegraphics[width=1.5in]{Images/orbitsO4.eps}
\caption{\label{fig:orbitsR345}
$R_{3,1}$, $R_{4,1}$, $R_{5,2}$, and $O_4$ families of POs.
Red line: envelope ellipse or hyperbola. }
\end{center}
\end{figure}
Semiclassical spectral properties are closely related to classical POs.
We have the following theorems concerning classical POs of EB, demonstrated in Fig. \ref{fig:orbitsR345}.
First, the numerical work of hundreds of families of POs verifies that all the POs except two on axes are continuous. Inside the area covered by a family of POs, there are four directions to find POs in the family. This makes the factor $c$ in Eq. \ref{eq:amplitude_PO_CB} below same for every continuous families of POs.
Second, the envelope of a family of POs is an ellipse or a hyperbola.
Third, the envelope ellipses, hyperbolas and EB are confocal. In Fig. \ref{fig:orbitsR345}, the foci of the envelope ellipse for $R_{3,1}$ are at $x = \pm\sqrt{1.228268^2 - 0.09297^2} = \pm \sqrt{3/2}$, which is same as EB; the foci of the envelope hyperbola of $O_4$ are at $x = \pm\sqrt{1.1547^2 + 0.408248^2} = \pm\sqrt{3/2}$.
Fourth, all the POs of a family with the same envelope curve have the same orbit length, for example all the $R_{3,1}$ POs have orbit length 6.0322.
Fifth, for type-R POs, every orbit line is tangent with the envelope ellipse and for type-O either the orbit line or its extended line is tangent with the envelope hyperbola.
\section{Fourier analysis of spectrum of rectangular, circular billiards}
To reveal the underlying POs of semiclassical spectrum, we perform Fourier analysis on the spectrum. In this section and the section below, we do not perform ensemble averaging on the Fourier analysis. The positions of peaks of Fourier analysis are identified with orbit lengths. Numerical work of EB confirms that all the peaks have the same half-width. Hence the height of a peak is proportional to amplitude of a peak and proportional to the amplitude of a PO.
\begin{equation}\label{eq:height_peak_vs_amplitude_PO}
\textit{height of peak} \propto
\textit{amplitude of peak} \propto
\textit{amplitude of PO} .
\end{equation}
\subsection{Fourier analysis of spectrum}
The Fourier analysis of spectrum has been successfully applied to analyze Sinai billiard, hydrogen atom in a strong magnetic field \cite{stockmann00}.
The trace formula expresses the fluctuation of level density as a summation over (families of) POs:
\begin{equation}\label{eq:trace_formula}
\rho(k) = \sum_{j} A_{j}( L_{j} ) e^{ikL_{j}+\alpha},
\end{equation}
where $L_{j}$ is orbit length of a family of POs and $\alpha$ the Maslov index. We do not consider $\alpha$ in this article. If we extend $A_{j}( L_{j} )$ to $A( l )$: the amplitude of POs over orbit length by extending $L_{j}$ to the whole line of $l$, we have
\begin{equation}
\rho(k)
= \int_{-\infty}^\infty A( l ) e^{ikl} dl.
\end{equation}
The reverse of a PO has orbit length $-l$.
The inverse Fourier analysis gives the amplitude
\begin{equation}\label{eq:inverse_trace_formula_smoothed}
A( l )
= \frac{1}{2\pi} \int_{-\infty}^\infty \rho(k) e^{-ikl} dk.
\end{equation}
The discreteness of spectrum makes Fourier analysis into a summation over eigen-momentum $k_i$, where $i$ counts eigenvalues:
\begin{equation}\label{eq:inverse_trace_formula_discrete}
A( l )
= \frac{1}{2\pi} \sum_{i} e^{-ik_i l},
\end{equation}
and
\begin{equation}
A_{j}( L_{j} )
= \frac{1}{2\pi} \sum_{i} e^{-ik_i L_{j}}.
\end{equation}
For simplicity, we ignore the subscript $j$ below.
\subsection{rectangular billiard}
\begin{figure}[htp]
\begin{center}
\includegraphics[width=3.0in]{Images/Fourier.rectangle.Golden.eps}
\caption{\label{fig:Fourier_rectangle}
Black line: Fourier analysis of RB with two sides as 1 and $(\sqrt{5}+1)/2$. Red square: semiclassical theory. The $L=2$ peak is normalized to 1 in both numerical and theoretical calculations. }
\end{center}
\end{figure}
A direct quantum mechanical calculation proves that for RB the amplitude of a family of POs is given by
\begin{equation}\label{eq:amplitude_PO_RB}
A(L) \propto \frac{c}{\sqrt{L}},
\end{equation}
where the momentum space factor $c = 1/2$ for POs parallel to boundary, and 1 otherwise \cite{wickramasinghe05, taoma.modified.Kepler} as for the former there are two directions to find POs from the same family while for the latter there are four directions. Although Eq. \ref{eq:amplitude_PO_RB} is simple, it gives correct position and height of almost every peak shown in Fig. \ref{fig:Fourier_rectangle}. The rare large discrepancy for some POs is caused by interference of close-by POs.
\subsection{circular billiard}
\begin{figure}[htp]
\begin{center}
\includegraphics[width=3.0in]{Images/Fourier.circular.eps}
\caption{\label{fig:Fourier_circular} Black line: Fourier analysis of CB. Red square: semiclassical theory. The $L=4$ peak is normalized to 1 in both numerical and theoretical calculations. }
\end{center}
\end{figure}
For CB, different POs cover different areas in position space. To account for this difference, we add the factor $S$, the area covered by a family of POs in position space (in the case of EB, the area covered by orbit lines in Fig. \ref{fig:orbitsR345}) and the amplitude of POs reads
\begin{equation}\label{eq:amplitude_PO_CB}
A(L) \propto \frac{c S}{\sqrt{L}},
\end{equation}
where the momentum space factor $c = 1/2$ for POs along diameters, and 1 otherwise as for the former there are two directions to find POs from the same family while for the latter there are four directions. For a family of POs denoted by $(n, m)$ with $m\leq n/2$, where in a period, the particle collides $n$ times with the boundary and makes $m$ revolutions,
\begin{eqnarray}
&S = \pi[ 1 - \cos^2(\frac{m\pi}{n})^2 ] \\
&L = 2 n \sin(\frac{m\pi}{n}) .
\end{eqnarray}
Eq. \ref{eq:amplitude_PO_CB} is verified in Fig. \ref{fig:Fourier_circular} except for interfering close-by orbits.
\subsection{summary}
The position and height of peaks of RB and CB are correctly given by the amplitude formulae of POs in Eqs. \ref{eq:amplitude_PO_RB} and \ref{eq:amplitude_PO_CB}. From our knowledge, the factor $S$ in Eq. \ref{eq:amplitude_PO_CB} has never been pointed out before. Its existence in EB is also confirmed below. An argument to justify Eqs. \ref{eq:amplitude_PO_RB} and \ref{eq:amplitude_PO_CB} is as follows.
The amplitude of a family of POs is the linear superposition of all the individual POs in the family. For integrable systems, we assume that the amplitude of each PO is only decided by its orbit length and given by $1/\sqrt{L}$, which is derived theoretically in the case of RB or more generally from Berry-Tabor formula \footnote{The factor $1/T$ in $A^2\propto 1/T$ with period $T$ in the appendix of \cite{wickramasinghe08} gives the same result.}. In this assumption, orbit stability plays no role. All the POs of a family contribute equally to the amplitude. The factor $cS$ is used to quantify how many POs are in a family. $c$ gives the number of POs in the momentum space and $S$ in the position space. For RB, every family of POs covers the whole billiard and $S$ plays no role therein.
Now we have a trace formula for integrable systems from Eq. \ref{eq:trace_formula}:
\begin{equation}\label{eq:trace_formula_generic_integrable_system}
\delta\rho(k) = \sum_{j} \frac{c S}{\sqrt{L}} e^{ikL + \alpha} .
\end{equation}
Given its similarity with Gutzwiller trace formula \cite{gutzwiller90}, this formula gives some new information compared with Berry-Tabor trace formula \cite{berry1977calculating.bound.spectrum}. It is worthwhile to understand its relation with the latter.
\section{Fourier analysis of spectrum of elliptic billiard}
\begin{figure}[htp]
\begin{center}
\includegraphics[width=3.3in]{Images/Fourier.Theory.EB.eps}
\end{center}
\caption{\label{fig:Fourier.Theory.EB} Red arrows: positions of unstable and stable isolated POs. The Fourier analysis contains eigenvalues of all the four symmetry classes of EB with $\sigma = 1/2$ but the almost degenerate eigenvalues are removed by only considering one of them. The heights of peaks are normalized by defining the peak for $R_{3,1}$ equal to 1 in both numerical result and theory.}
\end{figure}
Comparison of Fourier analysis of EB with the semiclassical theory in Eq. \ref{eq:amplitude_PO_CB} is shown in Fig. \ref{fig:Fourier.Theory.EB}. The positions of peaks match well with very small discrepancy. At $L = 2.83$, the short peak is due to the unstable PO with $L = 4b$, indicating that the unstable isolated PO is unimportant to the level fluctuation. At $L = 5.66$, the stable isolated PO with $L = 4a$ contributes another short peak. The isolated POs contribute little to level fluctuation, which was already demonstrated in the modified Kepler problem containing one stable isolated PO \cite{taoma.modified.Kepler}. The unimportance of isolated POs can be explained by our argument in the last section as its position space factor $S = 0$. At $L = 6.32$, the continuous $R_{4,1}$-POs contribute significantly. Around $L = 6.6$, there are many peaks overlapping together. These are the contribution from $R_{n,1}$ with $n\gg 1$. These POs has the orbit length around the perimeter equal to 6.85.
For heights of peaks, good match between numerical and theoretical results is achieved for type-R POs except for $R_{5,2}$ but not for type-O POs. Two kinds of discrepancies exist. First, for $O_4$, $O_{8}$, $O_{10}$, $O_{14}$, $O_{16}$, the discrepancy is a factor around 1. Second, for $O_6$ the discrepancy is very large. For the equally spaced $O_4, 2O_4, 3O_4, 4O_4$ (repetition of $O_4$) and $O_6, 2O_6, 3O_6$ (repetition of $O_6$), their relative magnitudes are correctly given by the factor $1/\sqrt{L}$.
\section{Conclusions}
Although EB is a generic integrable system, it contains several new characteristics. The second order statistics including spectral rigidity and global variance display long-range oscillations, not noticed in other systems before. We think such oscillations exist for any billiard with whispering gallery modes. The global variance seems to be a statistical instrument worth further studying. The Fourier analysis of spectrum gives the position and amplitude of POs. For RB and CB, the numerical results match our proposed theory of amplitude of POs. For EB, only the type-R POs fully match our theory. Further work is needed on type-O POs. Future work will address the orbital magnetic response of EB. Type-R and type-O POs enclose different magnetic flux and hence they should show different orbital magnetic response. In summary, EB is a non-trivial integrable system that will further enrich our understanding of quantum chaos.
|
\section{Introduction}
Throughout this paper $G=(V,E)$ is a finite, simple, and connected
graph of order $n(G)$. The distance between two vertices $u$ and
$v$, denoted by $d_G(u,v)$, is the length of a shortest path
between $u$ and $v$ in $G$. We write it simply $d(u,v)$ when no
confusion can arise. Also, the diameter of $G$,
$max_{\{u,v\}\subseteq V(G)}d(u,v)$, is denoted by $diam(G)$. The
vertices of a connected graph can be represented by different
ways, for example, the vectors which theirs components are the
distances between the vertex and the vertices in a given subset of
vertices.
For an ordered set $W=\{w_1,w_2,\ldots,w_k\}\subseteq V(G)$ and a
vertex $v$ of $G$, the $k$-vector
$$r(v|W):=(d(v,w_1),d(v,w_2),\ldots,d(v,w_k))$$
is called the (metric) {\it representation} of $v$ with respect
to $W$. The set $W$ is called a {\it resolving set} (locating set)
for $G$ if distinct vertices have different representations. In
this case we say the set $W$ {\it resolves} $G$. Elements in a
resolving set are called {\it landmarks}. A resolving set $W$ for
$G$ with minimum cardinality is called a {\it basis} of $G$, and
its cardinality is the {\it metric dimension} of $G$, denoted by
$\beta(G)$.
The concept of (metric) representation is introduced by
Slater~\cite{Slater1975} (see~\cite{Harary}). For more results in
this concept see~\cite{cameron,trees,cartesian
product,bounds,sur1,Discrepancies}. \\
It is obvious that for every graph $G$ of order $n$, $1\leq \beta(G)\leq
n-1$. Yushmanov~\cite{Yushmanov} improved this bound to $\beta(G)\leq
n-diam(G)$.
Khuller et al.~\cite{landmarks} and Chartrand et al.~\cite{Ollerman}
independently proved that $\beta(G)=1$ if and only
if $G$ is a path. Also, all graphs with metric dimension two are
characterized by Sudhakara and Hemanth Kumar~\cite{dim=2}.
Chartrand et al.~\cite{Ollerman} proved that the
only graph of order $n$, $n\geq 2$, with metric dimension $n-1$
is the complete graph $K_n$. They also provided a
characterization of graphs of order $n$ and
metric dimension $n-2$. In~\cite{idea} the problem of
characterization of all graphs of order $n$ with metric dimension
$n-3$ is proposed. In this paper, we answer to this question and
provide a characterization of graphs with metric dimension $n-3$.
First in next section, we present some definitions and known
results which are necessary to prove our main theorem.
\section{Preliminaries}
In this section, we present some definitions and known results
which are necessary to prove our main theorem. The notations
$u\sim v$ and $u\nsim v$ denote the adjacency and none-adjacency
between vertices $u$ and $v$, respectively. An edge with end
vertices $u$ and $v$ is denoted by $uv$. A path of order $n$,
$P_n$, and a cycle of order $n$, $C_n$, are denoted by
$(v_1,v_2,\ldots, v_n)$ and $(v_1,v_2,\ldots,v_n,v_1)$,
respectively.
\par We say an ordered set
$W$ {\it resolves} a set $T$ of vertices in $G$, if the
representations of vertices in $T$ are distinct with respect to
$W$. When $W=\{x\}$, we say that the vertex $x$ resolves $T$. To
see that whether a given set $W$ is a resolving set for $G$, it is
sufficient to look at the representations of vertices in
$V(G)\backslash W$, because
$w\in W$ is the only vertex of $G$ for which $d(w,w)=0$.\\
In~\cite{Ollerman} all graphs of order $n$ with metric dimension
$n-2$ are characterized as follows.
\begin{lem} {\rm\cite{Ollerman}} \label{n-2}
Let $G$ be a graph of order $n\geq 4$. Then $\beta(G)=n-2$ if and
only if $G=K_{s,t}~(s,t\geq 1), G=K_s\vee\overline K_t~(s\geq 1,
t\geq 2)$, or $G=K_s\vee (K_t\cup K_1)~ (s,t\geq 1)$.
\end{lem}
For each vertex $v\in V(G)$, let $\Gamma_i(v):=\{u\in
V(G)~|~d(u,v)=i\}$. Two distinct vertices $u,v$ are {\it twins} if
$\Gamma_1(v)\backslash\{u\}=\Gamma_1(u)\backslash\{v\}$. It is
called that $u\equiv v$ if and only if $u=v$ or $u,v$ are twins.
In~\cite{extermal}, it is proved that the relation $\equiv$ is an
equivalent relation. The equivalence class of the vertex $v$ is
denoted by $v^*$. The {\it twin graph} of $G$, denoted by $G^*$,
is the graph with vertex set $V(G^*):=\{v^*~|~v\in V(G)\}$, where
$u^*v^*\in E(G^*)$ if and only if $uv\in E(G)$. It is easy to
prove that $u,v$ are adjacent in $G$ if and only if all vertices
of $u^*$ are adjacent to all vertices of $v^*$, hence the
definition of $G^*$ is well defined. For each subset $S\subseteq
V(G)$, let $S^*$ denote the set $\{v^*\in V(G^*)~|~v^*\subseteq
S\}$. Also, by $\Gamma_i^*(v^*)$, we mean the set $\{u^*\in
V(G^*)~|~d_{G^*}(u^*,v^*)=i\}$. For each $v\in v^*$, it is
immediate that $\Gamma^*_i(v^*)=\Gamma^*_i(v)$ if $i\neq 0$.
Furthermore, we define \begin{description} \item$R_1(v):=\{x\in
\Gamma_1(v)~|~\exists y\in \Gamma_2(v):x\sim y\}$, \item
$R_2(v):=\Gamma_1(v)\backslash R_1(v)$,\item $R^*_1(v^*):=\{x^*\in
\Gamma^*_1(v^*)~|~\exists y^*\in \Gamma^*_2(v^*):x^*\sim
y^*\}$,\item $R^*_2(v^*):=\Gamma^*_1(v^*)\backslash R^*_1(v^*)$.
\end{description}
\par We say a set $S$ of vertices is {\it homogeneous}
if the induced subgraph by $S$ in $G$, $G[S]$, is a complete or
an empty subgraph of $G$. In this terminology, it is proved
in~\cite{extermal} that each vertex of $G^*$ is a homogeneous
subset of $V(G)$.
\begin{obs} {\rm\cite{extermal}}\label{twins}
If the vertices $u,v$ are twins in a graph $G$ and $S$ resolves
$G$, then $u$ or $v$ is in $S$. Moreover, if $u\in S$ and $v\notin
S$, Then $(S\setminus\{u\})\cup \{v\}$ also resolves $G$.
\end{obs}
\begin{pro} {\rm\cite{extermal}}\label{diam G^*}
Let $G\neq K_1$ be a graph. Then $diam(G^*)\leq diam(G).$
Moreover, if $u,v\in V(G)$ are not twin vertices of $G$, then
$d_{G^*}(u^*,v^*)=d_G(u,v)$.
\end{pro}
As in ~\cite{extermal}, we say that $v^*\in V(G^*)$ is of type:
\begin{description}
\item $\bullet$ (1) if $|v^*|=1$, \item$\bullet$ (K) if
$G[v^*]\cong K_r$ and $r\geq 2$, \item $\bullet$ (N) if
$G[v^*]\cong \overline{K_r}$ and $r\geq 2$.
\end{description}
A vertex of $G^*$ is of type (1K) if it is of type (1) or (K). A
vertex is of type (1N) if it is of type (1) or (N). A vertex is of
type (KN) if it is of type (K) or (N). We denote by $\alpha(G^*)$
the number of vertices of $G^*$ of type (K) or (N). It is obvious
that $G$ is uniquely determined by $G^*$, and the type and
cardinality of each vertex of $G^*$.
\par Hernando et al.~\cite{extermal} characterized all
graphs of order $n$, diameter $d$ and metric dimension $n-d$ by
the following theorem.
\begin{lem} {\rm\cite{extermal}}\label{n-d}
Let $G$ be a graph of order $n$ and diameter $d\geq 3$. Let $G^*$
be the twin graph of $G$. Then $\beta(G)=n-d$ if and only if $G^*$
is one of the following graphs:
\begin{description}
\item 1. $G^*\cong P_{d+1}$ and one of the following cases holds\\
(a) $\alpha(G^*)\leq 1$;\\ (b) $\alpha(G^*)=2$, the two vertices
of $G^*$ not of type (1) are adjacent, and if one is a leaf of
type (K), then the other is also of type (K);\\
(c) $\alpha(G^*)=2$, the two vertices of $G^*$ not of type (1) are
at distance $2$ and both are of type (N); or\\
(d) $\alpha(G^*)=3$ and there is a vertex of type (KN) adjacent to
two vertices of type (N). \item 2. $G^*\cong P_{d+1,k}$ {\rm(}the
path $(u^*_0,u^*_1,\ldots,u^*_d\rm)$ with one extra vertex
adjacent to $u^*_{k-1}${\rm)} for some integer $k\in [3,d-1]$, the
degree-$3$ vertex $u^*_{k-1}$ of $G^*$ is of any type, each
neighbor of $u^*_{k-1}$ is of type (1N), and every other vertex is
of type (1). \item 3. $G^*\cong P^\prime_{d+1,k}$ {\rm(}the path
$(u^*_0,u^*_1,\ldots,u^*_d)$ with one extra vertex adjacent to
$u^*_{k-1}$ and $u^*_k${\rm)} for some integer $k\in [2,d-1]$, the
three vertices in the cycle are of type (1K), and every other
vertex is of type (1).
\end{description}
\end{lem}
To prove our main theorem, we need the following propositions.
\begin{pro}\label{sugraph}
Let $H$ be a subgraph of a graph $G$. If $\beta(H)=n(H)-t$ and
$d_H(u,v)=d_G(u,v)$ for all pairs of vertices in $H$, then
$\beta(G)\leq n(G)-t$.
\end{pro}
\begin{proof} {Let $W$ be a basis of $H$. Since $d_H(u,v)=d_G(u,v)$ for
all pairs of vertices in $H$, $W$ as a subset of $V(G)$ resolves
all vertices of $H$. Hence, $T=W\cup (V(G)\backslash V(H))$ is a
resolving set for $G$ where, $|T|=n(G)-t$. Therefore, the metric
dimension of $G$ is at most $n(G)-t$. }\end{proof}
\begin{cor}\label{sugraph d=2}
If $H$ is an induced subgraph of a graph $G$, where $diam(H)=2$,
and $\beta(H)=n(H)-t$ for some positive integer $t$, then
$\beta(G)\leq n(G)-t$.
\end{cor}
\begin{cor}\label{sugraphH<K<G}
Let $H$ be an induced subgraph of a graph $G$, and $G$ be an
induced subgraph of a graph $R$, where $diam(H)=diam(G)=2$,
$\beta(R)=n(R)-t$, and $\beta(H)=n(H)-t$ for some positive integer
$t$. Then $\beta(G)=n(G)-t$.
\end{cor}
\begin{proof}{
Let $\beta(G)=n(G)-s$, for some positive integer $s$. By
Corollary~\ref{sugraph d=2}, we have $n(G)-s=\beta(G)\leq n(G)-t$
and $n(R)-t=\beta(R)\leq n(R)-s$. Therefore, $s=t$. }\end{proof}
\begin{pro}\label{dim G^*<n-t}
Let $G$ be a graph and $G^*$ be the twin graph of $G$. If
$\beta(G^*)=n(G^*)-t$ for some positive integer $t$, then
$\beta(G)\leq n(G)-t$.
\end{pro}
\begin{proof}{ Let $S^*$ be a basis of $G^*$ and $T^*=V(G^*)\backslash S^*$.
We choose a vertex $v$ from each $v^*\in T^*$ and let
$T=\{v~|~v^*\in T^*\}$. Since $S^*$ is a basis of $G^*$, for each
pair of vertices $u^*,v^*\in T^*$, there exist a vertex $x^*\in
S^*$ such that $d_{G^*}(x^*,u^*)\neq d_{G^*}(x^*,v^*)$. Note that
neither $u$ nor $v$ is twin with $x$, for each $x\in x^*$.
Therefore, by Proposition~\ref{diam G^*}, we have
$d_{G^*}(x^*,u^*)=d_G(x,u)$ and $d_{G^*}(x^*,v^*)=d_G(x,v)$.
Hence, $d_G(x,u)\neq d_G(x,v)$, which implies that the set
$S=\bigcup_{v^*\in S^*}v^*$ resolves $T$. Therefore,
$V(G)\setminus T$ is a resolving set for $G$ of cardinality
$n(G)-t$, thus $\beta(G)\leq n(G)-t$. }\end{proof}
\begin{pro}\label{B(G)>n-n*}
Let $G$ be a graph and $G^*$ be the twin graph of $G$. Then
$\beta(G)\geq n(G)-n(G^*)$.
\end{pro}
\begin{proof}{
If $S$ resolves $G$, then Observation~\ref{twins} shows that $S$
contains at least $|v^*|-1$ vertices from each $v^*\in V(G^*)$.
Hence, $\beta(G)\geq n(G)-n(G^*)$. }\end{proof}
\section{\label{sec:sts}Main Results}
Let $G$ be a connected graph of order $n$ and metric dimension
$n-3$. Since $n-3=\beta(G)\leq n-diam(G)$, it follows that
$diam(G)\leq 3$. If $diam(G)=1$, then $G\cong K_n$, contrary to
$\beta(G)=n-3$. If $diam(G)=3$, then in Theorem~\ref{n-d}, let
$d=3$, which obtains a characterization of graphs $G$ with
$\beta(G)=n-3$, where $diam(G)=3$, (Note that in this case the
interval $[3,2]$ is empty, hence, Case 2 dose not occur).
Therefore, it is enough to consider the case $diam(G)=2$. The
following theorem is our main result, which is a characterization
of all graphs with metric dimension $n-3$ and diameter $2$.
\begin{theorem} \label{d=2}
Let $G$ be a graph of order $n$ and diameter $2$ and $G^*$ be the
twin graph of $G$. Then $\beta(G)=n-3$ if and only if $G^*$
satisfies in one of the following structures:
\begin{description}
\item $G_1$: $G^*\cong K_3$ and has at most one vertex of type
(1K);
\item $G_2$: $G^*\cong P_{3}$ and one of the following cases holds:\\
(a) The degree-$2$ vertex is of type (N) and one of the leaves is of type (K)
and the other is of any type;\\
(b) One of the leaves is of type (K), the other is of type (KN)
and the degree-$2$ vertex is of any type; \item $G_3$: $G^*$ is
the paw {\rm(}a triangle with a pendant edge{\rm)}, and the
degree-$3$ vertex is of any type, one of the degree-$2$ vertices
is of type (N), the other is of type (1K), the leaf is of type
(1N). Moreover, a degree-$2$ vertex of type (K) yields the leaf
and the degree-$3$ vertex are not of type (N); \item $G_4$:
$G^*\cong C_5$, and each vertex is of type (1); \item $G_5$: $G^*$
is $C_5$ with a chord, and the adjacent degree-$2$ vertices are of
type (1) and the other vertices are of type (1K); \item $G_6$:
$G^*$ is a $C_5$ with two adjacent chords. The degree-$4$ vertex
is of any type, the others are of type (1K). Furthermore, two
non-adjacent vertices are not of type (K), and two adjacent
vertices are not of different types (K) and (N); \item $G_7$:
$G^*$ is a kite with a pendant edge adjacent to a degree-$3$
vertex, and the leaf is of type (1), the degree-$4$ and degree-$3$
vertices are of type (1K), one of the degree-$2$ vertices is of
type (K) and the other is of type (1). \item $G_8$: $G^*$ is the
kite, and one of the degree-$2$ vertices is of type (K) the other
is of type (1), one of the degree-$3$ vertices is of type (N), and
the other is of type (1K);\item $G_9$: $G^*\cong C_4$, and two
adjacent vertices are of type (K) and others are of type (1);\item
$G_{10}$: $G^*\cong C_4\vee K_1$, and two degree-$3$ adjacent
vertices are of type (K), degree-$4$ vertex is of type (1K), and
others are of type (1).
\end{description}
In Figure~\ref{G^*and dim=n-3} the scheme of the above $10$
structures are shown.
\end{theorem}
\begin{figure}[ht]\hspace{1.5cm}
\vspace*{5cm}\special{em:graph fig1.bmp} \vspace*{1.5cm}
\vspace{1.3cm}
\caption{\label{G^*and
dim=n-3}{\scriptsize The twin graphs of graphs with diameter $2$
and metric dimension $n-3$.} }
\end{figure}
\subsection{\label{necessity} Proof of Necessity} Throughout this
section, $G$ is a graph of order $n$, diameter $2$, metric
dimension $n-3$, and $G^*$ is the twin graph of $G$. Note that,
Proposition~\ref{diam G^*} implies that $diam(G^*)\leq 2$. Through
a sequence of lemmas and propositions, we show that $G^*$ has one
of the structures $G_1$ to $G_{10}$.
\begin{pro}\label{G^*=K_n pro}
If $diam(G^*)=1$, then $G^*$ satisfies in structure $G_1$.
\end{pro}
\begin{proof}{
Let $u^*,v^*$ be two vertices of $G^*$ of type (1K). Since $G^*$
is a complete graph, every pair of vertices $u\in u^*$ and $v\in
v^*$ of $G$ are twins, thus $u^*=v^*$. Hence, $G^*$ has at most
one vertex of type (1K). If vertices $v^*_1,v^*_2,v^*_3$, and
$v^*_4$ of $G^*$ (except possibly $v^*_1$) are of type (N), then
we choose an arbitrary vertex $v_i\in v^*_i$, for each $i$, $1\leq
i\leq 4$, and $u_i\in v^*_i\backslash\{v_i\}$, for each $i$,
$2\leq i\leq 4$. Let $T=\{u_2,u_3,u_4\}$. It follows
that$$r(v_1|T)=(1,1,1),\quad r(v_2|T)=(2,1,1),\quad
r(v_3|T)=(1,2,1),\quad r(v_3|T)=(1,1,2).$$ Hence, the set
$V(G)\backslash\{v_1,v_2,v_3,v_4\}$ is a resolving set for $G$ and
$\beta(G)\leq n-4$. This contradicts our assumption,
$\beta(G)=n-3$. Therefore, $G^*$ has at most three vertices.
Assume $G^*\cong K_t$ for some integer $t\in [1,3]$. If $t=1$,
then $G\cong K_n$ or $G \cong\overline K_n$. If $t=2$, then
$G\cong K_{r,s}$ or $G\cong K_r \vee \overline K_s$. Since
$\beta(G)=n-3$, the above cases are impossible. Consequently
$G^*\cong K_3$, which is the desired conclusion. }\end{proof} The
remainder of this section will be devoted to the case
$diam(G^*)=2$. It is clear that in this case, there exists a
vertex $v\in V(G)$ such that $\Gamma^*_2(v^*)\neq \emptyset$ and
$\Gamma_2(v)\neq\emptyset$.
\begin{lemma}\label{G^*=K_3}
If \- $\Gamma_2^*(v^*)\neq\emptyset$ and $v\in v^*$, then
$\bigcup_{u^*\in \Gamma_i^*(v^*)}u^* \subseteq \Gamma_i(v)$,
where $i\in\{1,2\}$, $\bigcup_{u^*\in R_2^*(v)}u^*
\linebreak\subseteq R_2(v)$, and $\bigcup_{u^*\in
R_1^*(v)}u^*=R_1(v)$. Moreover, if $v^*$ is of type (1K), then
$\bigcup_{u^*\in \Gamma_2^*(v^*)}u^*=\Gamma_2(v)$,
$R_1^*(v^*)=R_1^*(v)$, and $R_2^*(v^*)=R_2^*(v)$.
\end{lemma}
\begin{proof}{ It is clear that a vertex in
$\Gamma_1(v)$ is not twin with any vertex of $\Gamma_2(v)$.
Therefore, all twins of vertices in $\Gamma_1(v)$ are in the set
$\Gamma_1(v)\cup \{v\}$ and all twins of vertices in $\Gamma_2(v)$
are in $\Gamma_2(v)\cup \{v\}$. This gives $\Gamma_i(v)\backslash
v^*=\bigcup_{u^*\in \Gamma_i^*(v^*)}u^*$, for $i\in\{1,2\}$.
Hence, $\bigcup_{u^*\in \Gamma_i^*(v^*)}u^*\subseteq \Gamma_i(v)$,
when $i\in\{1,2\}$. Note that all twins of vertices in $R_1(v)$
are in $R_1(v)$, because each member of $R_1(v)$ is adjacent to
$v$ and has at least one neighbor in $\Gamma_2(v)$ while the
vertices in $R_2(v)\cup\{v\}$ have not such neighbors. Thus
$\bigcup_{u^*\in R_1^*(v)}u^*=R_1(v)$. By the same reason, all
twins of vertices in $R_2(v)$ are in $R_2(v)\cup\{v\}$.
Consequently, $\bigcup_{u^*\in R_2^*(v)}u^*=R_2(v)\backslash
v^*\subseteq R_2(v)$.
\par Now let $v^*$ is of type (1K). Therefore, $v$ can only be
twin with the vertices of $R_2(v)$. Hence, $\bigcup_{u^*\in
\Gamma_2^*(v^*)}u^*=\Gamma_2(v)$. It is clear that
$R^*_1(v^*)\subseteq R^*_1(v)$. If there exist a vertex $u^*\in
R^*_1(v)\backslash R^*_1(v^*)$, then $u^*\in R^*_2(v^*)$, because
$\Gamma^*_1(v^*)=R^*_1(v^*)\cup R^*_2(v^*)$. Since $v^*\subseteq
R_2(v)\cup \{v\}$, all neighbors of each vertex $u\in u^*$ are in
the set $\Gamma_1(v)\cup\{v\}$, this contradicts the fact that
$u\in R_1(v)$. Therefore, $R_1^*(v^*)=R_1^*(v)$ and consequently,
$R_2^*(v^*)=R_2^*(v)$.
}\end{proof}
\begin{lemma}\label{one homogeneous}
For each $v$, where $\Gamma_2(v)\neq\emptyset$, at least one of
the sets $\Gamma_1(v)$ and $\Gamma_2(v)$ is homogeneous.
\end{lemma}
\begin{proof}{
Let $\Gamma_2(v)\neq\emptyset$. On the contrary, suppose that both
$\Gamma_1(v)$ and $\Gamma_2(v)$ are not homogeneous. Therefore,
there exist vertices $v_1, v_2$, and $v_3$ in $\Gamma_1(v)$, and
vertices $u_1, u_2$, and $u_3$ in $\Gamma_2(v)$ such that $v_1\sim
v_2, v_2\nsim v_3$ and $u_1\sim u_2$, $u_2\nsim u_3$. If
$W^\prime=\{v,v_2,u_2\}$, then the representations of
$v_1,v_3,u_1$ and $u_3$ with respect to $W^\prime$ are as follows
$$r(v_1|W^\prime)=(1,1,*),\quad r(v_3|W^\prime)=(1,2,*),\quad
r(u_1|W^\prime)=(2,*,1),\quad r(u_3|W^\prime)=(2,*,2),$$where $*$
is $1$ or $2$. These four representations are distinct, hence,
$V(G)\backslash\{v_1,v_3,u_1,u_3\}$ is a resolving set for $G$.
Thus $\beta(G)\leq n-4$, which is a contradiction. Therefore, at
least one of the sets $\Gamma_1(v)$ or $\Gamma_2(v)$ is
homogeneous. }\end{proof} By Lemma~\ref{one homogeneous}, to
complete the proof of necessity, we need to consider the following
two cases.
\vspace{4mm} \\{\bf Case 1.} There exist a vertex $v\in V(G)$ such
that $\Gamma^*_2(v^*)\neq\emptyset$ and $\Gamma_1(v)$ is
homogeneous. \vspace{4mm}\\ By the assumption of Case 1, the
following results are obtained.
\begin{fact}\label{R_1^*<2}
$|R^*_1(v^*)|\leq 2$.
\end{fact}
\begin{proof}{ Since every vertex of $R^*_1(v^*)$ has a neighbor in
$\Gamma^*_2(v^*)$ and $\Gamma_1(v)$ is homogeneous, for distinct
vertices $x^*,y^*\in R^*_1(v^*)$, the sets $\Gamma^*_1(x^*)\bigcap
\Gamma^*_2(v^*)$ and $\Gamma^*_1(y^*)\bigcap \Gamma^*_2(v^*)$ are
distinct nonempty sets. Therefore, the set $\Gamma^*_2(v^*)$
resolves the vertices of $R^*_1(v^*)$ in $G^*$. Moreover, since
every vertex of $R^*_1(v^*)$ has a neighbor in $\Gamma^*_2(v^*)$
and $v^*$ has not such a neighbor, if we compute the
representations of the vertices in $R^*_1(v^*)\cup\{v^*\}$ with
respect to $\Gamma^*_2(v^*)$, then the representation of each
vertex in $R^*_1(v^*)$ has a coordinate 1 while all coordinates of
$r(v^*|\Gamma^*_2(v^*))$ are 2. Therefore, $\Gamma^*_2(v^*)$
resolves the set $R^*_1(v^*)\cup\{v^*\}$ consequently,
$V(G^*)\backslash(R^*_1(v^*)\cup\{v^*\})$ is a resolving set for
$G^*$. Thus $\beta(G^*)\leq n(G^*)-|R^*_1(v^*)\cup\{v^*\}|$. On
the other hand, by Propositions~\ref{dim G^*<n-t}, we have
$\beta(G^*)\geq n(G^*)-3$. Therefore,
$|R^*_1(v^*)\cup\{v^*\}|\leq 3$, which yields $|R^*_1(v^*)|\leq
2$. }
\end{proof}
\begin{lemma}\label{R_2=empty}
If $\Gamma_2(v)$ is not homogeneous, then $R_2(v)$ and
$R^*_2(v^*)$ are empty sets.
\end{lemma}
\begin{proof}{
Note that $|R^*_1(v^*)|\geq 1$, otherwise
$\Gamma^*_2(v^*)=\emptyset$. Since $\Gamma_2(v)$ is not
homogeneous, there exist vertices $i,j,$ and $k$ in $\Gamma_2$
such that $i\sim j$ and $j\nsim k$. Now if
$R_2(v)\neq\emptyset$, then let $r_1\in R_1(v)\cap \Gamma_1(j)$,
$r_2\in R_2(v)$ and $W_0=\{v,j\}$. Thus, $r(i|W_0)=(2,1)$,
$r(k|W_0)=(2,2)$, $r(r_1|W_0)=(1,1)$ and $r(r_2|W_0)=(1,2)$, and
so $\beta(G)\leq n-4$. This contradiction implies
$R_2(v)=\emptyset$.
\par If $R^*_2(v^*)\neq\emptyset$, then $\bigcup_{u^*\in R^*_2(v^*)}u^*\nsubseteq
R_2(v)$ and hence, by Lemma~\ref{G^*=K_3}, $v^*$ is of type (N).
Therefore, there exist two adjacent vertices $a,b\in
\bigcup_{u^*\in\Gamma^*_2(v^*)}u^*$, otherwise $\Gamma_2(v)$ is
homogeneous. Since $diam(G^*)=2$, there exist a vertex $r_1\in
\bigcup_{u^*\in R^*_1(v^*)}u^*$ such that $r_1\sim a$. Now let
$v_1,v_2\in v^*,~r_2\in\bigcup_{u^*\in R^*_2(v^*)}u^*$, and
$W=\{v_1,a\}$. Thus, we have $$r(v_2|W)=(2,2),\quad
r(r_1|W)=(1,1),\quad r(r_2|W)=(1,2),\quad r(b|W)=(2,1).$$
Therefore, $V(G)\backslash\{v_2,r_1,r_2,b\}$ is a resolving set
for $G$, with cardinality $n-4$, which contradicts our assumption
$\beta(G)=n-3$. Consequently $R^*_2(v^*)=\emptyset$.}
\end{proof}
\begin{fact}\label{Gamma^*_2<3}
$|\Gamma^*_2(v^*)|\leq 3$.
\end{fact}
\begin{proof}{ If $\Gamma_2(v)$ is homogeneous, then since every vertex
of $\Gamma^*_2(v^*)$ has a neighbor in $R^*_1(v^*)$ and
$\Gamma^*_2(v^*)$ is homogeneous, for distinct vertices
$x^*,y^*\in\Gamma^*_2(v^*)$, the sets $\Gamma^*_1(x^*)\bigcap
R^*_1(v^*)$ and $\Gamma^*_1(y^*)\bigcap R^*_1(v^*)$ are distinct
nonempty sets. Therefore, for each pair of vertices
$x^*,y^*\in\Gamma^*_2(v^*)$ there exist a vertex $r^*_1\in
R^*_1(v^*)$ such that $r^*_1$ resolves $x^*$ and $y^*$ in $G^*$.
Hence, $R^*_1(v^*)$ resolves all vertices of the set
$\Gamma^*_2(v^*)$. This implies that
$V(G^*)\backslash\Gamma^*_2(v^*)$ is a resolving set for $G^*$,
which yields $\beta(G^*)\leq n(G^*)-|\Gamma^*_2(v^*)|$. On the
other hand, by Propositions~\ref{dim G^*<n-t}, we have
$\beta(G^*)\geq n(G^*)-3$. Thus, $|\Gamma^*_2(v^*)|\leq 3$.
\par Now let $\Gamma_2(v)$ is not
homogeneous. By Fact~\ref{R_1^*<2}, $|R_1^*(v^*)|\leq 2$. Thus, we
consider two cases, $|R_1^*(v^*)|=1$ and $|R_1^*(v^*)|=2$. In the
case $|R_1^*(v^*)|=1$, let $R^*_1(v^*)=\{r^*_1\}$, $r_1\in r^*_1$,
and for each $l\in\Gamma_2(v)$, $N_1(l)=\Gamma_1(l)\cap
\Gamma_2(v)$, and $N_2(l)=\Gamma_2(l)\cap \Gamma_2(v)$. Since
$\Gamma_2(v)$ is not homogeneous, there exists a vertex
$x\in\Gamma_2(v)$ such that both $N_1(x)$ and $N_2(x)$ are
nonempty sets. Let $a\in N_1(x)$ and $y\in N_2(x)$. Note that
$\Gamma^*_2(v^*)=\{x^*\}\cup N^*_1(x)\cup N^*_2(x)$, and $x$
resolves $a$ and $y$. Hence, if $N_1(x)$ is not homogeneous, then
there exist vertices $i,j,k\in N_1(x)$ such that $i\sim j$ and
$j\nsim k$. Thus, $W^\prime=\{v,x,j\}$ resolves the
$\{i,k,y,r_1\}$, this contradiction yields $N_1(x)$ is
homogeneous. By a similar reason $N_2(x)$ is also homogeneous.
Note that all different neighbors of vertices in $N^*_1(x)$ are
in the set $N^*_2(x)$, because $N_1(x)$ is homogeneous and its
vertices share their neighbors in $\Gamma_1(v)\cup\{x\}$.
Similarly every different neighbor of vertices in $N^*_2(x)$ are
in $N^*_1(x)$, hence, $N^*_1(x)$ and $N^*_2(x)$ resolves each
other. Now let $W_1=N^*_2(x)\cup\{x^*\}$. If each vertex of
$N^*_1(x)$ has a none-neighbor vertex in $N^*_2(x)$, then the
representation of each vertex of $N^*_1(x)$ with respect to
$N^*_2(x)$ has a coordinate $2$, all coordinates of
$r(r^*_1|N^*_2(x))$ is $1$, and $r(v^*|N^*_2(x))$ is entirely
$2$. Consequently, $W_1$ resolves $N^*_1(x)\cup\{r^*_1,v^*\}$.
Thus, $\beta(G)=n-3$ implies that $|N^*_1(x)|\leq 1$. Moreover,
if there exist a vertex $a^*\in N^*_1(x)$ such that $a^*$ is
adjacent to all vertices of $N^*_2(x)$, then $N^*_1(x)$ has at
most two vertices. Otherwise, there are two distinct vertices
$b^*,c^*\in N^*_1(x)$ such that they are different from $a^*$,
and $r(b^*|N^*_2(x))$ and $r(c^*|N^*_2(x))$ are not entirely $1$,
and so $W_1$ resolves four vertices $a^*,~b^*,~c^*$, and $v^*$,
contrary to $\beta(G)=n-3$. Hence, $|N^*_1(x)|\leq 2$.
Furthermore, $|N^*_1(x)|= 2$ yields that there exist a vertex
$a^*\in N^*_1(x)$ such that $a^*$ is adjacent to all vertices of
$N^*_2(x)$. On a similar way $|N^*_2(x)|\leq 2$, moreover,
$|N^*_2(x)|= 2$ only if there exist a vertex $y^*\in N^*_2(x)$
such that $y^*$ is none-adjacent to all vertices of $N^*_1(x)$.
Thus, at most one of the sets $N^*_1(x)$ and $N^*_2(x)$ can have
two vertices, because it is impossible that there exist a pair of
vertices $a^*\in N^*_1(x),~y^*\in N^*_2(x)$ such that $a^*$ is
adjacent to all vertices of $N^*_2(x)$ and $y^*$ is none-adjacent
to all vertices of $N^*_1(x)$. Consequently
$|\Gamma^*_2(v^*)|\leq 4$. We claim that $|\Gamma^*_2(v^*)|=4$ is
impossible.
\par If $|\Gamma^*_2(v^*)|=4$,
then one of the two blow cases can be happened. \vspace{4mm}\\
1. $|N^*_1(x)|=2$ and $|N^*_2(x)|=1$. Let $N^*_1(x)=\{a^*,b^*\}$,
$N^*_2(x)=\{y^*\}$, $y^*\sim a^*$, $v\in v^*,a\in a^*,b\in
b^*,x\in x^*$ and $y\in y^*$. If $a^*\sim b^*$, then $x^*$ and
$b^*$ are twins, see Figure~\ref{3}(a). Since $x^*\sim b^*$, one
of them is of type (N). Note that $b^*$ is not of type (N),
because $N^*_1(x)$ is homogeneous and $a^*\sim b^*$. Hence, $b^*$
is of type (1K) and $x^*$ is of type (N). Thus, the set
$V(G)\backslash\{r_1,y,a,x\}$ is a resolving set for G of size
$n-4$, which is impossible. Therefore, $a^*\nsim b^*$, this gives
the set $V(G)\backslash\{v,r_1,x,a\}$ is a resolving set for $G$.
Which is a contradiction.\vspace{4mm}\\
2. $|N^*_1(x)|=1$ and $|N^*_2(x)|=2$. Let $N^*_1(x)=\{a^*\}$,
$N^*_2(x)=\{y^*,z^*\}$, $z^*\sim a^*$, $v\in v^*,a\in a^*,x\in
x^*,y\in y^*,z\in z^*$ and $S=\{x,y\}$. Thus, $r(z|S)=(2,*)$,
$r(a|S)=(1,2)$, $r(r_1|S)=(1,1)$ and $r(v|S)=(2,2)$, where $*=1$
or $2$. Note that $\beta(G)=n-3$ yields $*=2$, see
Figure~\ref{3}(b). Therefore, $x^*$ and $z^*$ are twins. Since
they are none-adjacent vertices, one of them is of type (K).
Clearly $z^*$ is not of type (K), otherwise, since $N^*_2(x)$ is
homogeneous, we have $*=1$, which is impossible. Consequently,
$x^*$ is of type (K), this implies that the set
$(x^*\backslash\{x\})\cup\{z,y\}$
resolves $\{v,r_1,a,x\}$. Thus, $\beta(G)\leq n-4$. \vspace{4mm}\\
These contradictions yield, $|\Gamma^*_2(v^*)|\leq 3$, when
$|R^*_1(v^*)|=1$.
\par To complete
the proof we need only to consider the case $|R^*_1(v^*)|=2$. In
this case, since all different neighbors of vertices in $R_1(v)$
are in $\Gamma_2(v)$, $|R^*_1(v^*)|=2$ implies that there exist a
vertex $a\in\Gamma_2(v)$ such that $\Gamma_2(a)\cap
R_1(v)\neq\emptyset$. Let $T=\Gamma_2(v)\backslash\{a\}$. Since
$\Gamma_2(v)$ is not homogeneous, it has at least three vertices,
hence $|T|\geq 2$. If $T$ is not homogeneous, then there exists a
vertex $i\in T$ such that $i$ resolves two vertices of $T$.
Moreover, we know that $a$ resolves two vertices of $R_1(v)$.
Hence, $\{v,i,a\}$ resolves at least four vertices of $G$. Thus,
we obtain a resolving set for $G$ of size $n-4$, which is
impossible. Therefore, $T$ is homogeneous.
\par If $a$ is adjacent to a vertex in $T$, then $a$
is adjacent to all vertices of $T$, otherwise $\{a,v\}$ resolves
four vertices. If a vertex $t\in T$ is not adjacent to some
vertices of $R_1(v)$, then similar to above, it can be seen that
$\Gamma_2(v)\backslash\{t\}$ is homogeneous and $t$ is adjacent or
none-adjacent to all vertices of $\Gamma_2(v)\backslash\{t\}$.
This implies that $\Gamma_2(v)$ is homogeneous, which is a
contradiction with the assumption. Hence, every vertex of $T$ is
adjacent to all vertices of $R_1(v)$. Therefore, all vertices of
$T$ are twins and form a vertex $b^*$ in $G^*$. Thus,
$\Gamma^*_2(v^*)$ consists of two vertices $a^*$ and $b^*$, where
$a^*$ is of type (1) and $b^*$ is of type (KN). }
\end{proof}
\vspace{-.7cm}\begin{figure}[ht]\hspace{1cm}
\vspace*{5cm}\special{em:graph fig2.bmp} \vspace*{1.5cm}
\vspace{-1.3cm}
\caption{\label{3} \footnotesize
Different cases of $N^*_1(x)$ and $N^*_2(x)$.}
\end{figure}
\begin{pro}\label{case2R^*=1} If $|R^*_1(v^*)|=1$, then $G^*$ satisfies one of the
structures $G_2$, $G_3$, or $G_7$.
\end{pro}
\begin{proof}{
Let $R^*_1(v^*)=\{r^*_1\}$. First let $\Gamma_2(v)$ is
homogeneous. Therefore, every vertex of $\Gamma^*_2(v^*)$ is
adjacent to $r^*_1$ and all vertices of $\Gamma^*_2(v^*)$ are
twins. Consequently, since $\Gamma_2(v)$ is homogeneous, all
vertices of $\bigcup_{u^*\in\Gamma^*_2(v^*)}u^*$ are twins. This
gives $|\Gamma^*_2(v^*)|=1$. Now, If $R^*_2(v^*)=\emptyset$, then
$G^*\cong P_3$ and $\alpha(G^*)\geq 2$, otherwise $\beta(G)=n-2$.
It is easy to check that in both cases $\alpha(G^*)=2$ and $3$, at
least one of the leaves is of type (K), otherwise $\beta(G)=n-2$.
Let $G^*=(x^*_1,x^*_2,x^*_3)$ and $x^*_1$ be of type (K). If
$x^*_2$ is of type (1K) and $x^*_3$ is of type (1), then
$\beta(G)=n-2$. This contradiction implies that, if $x^*_3$ is of
type (1), then $x^*_2$ is of type (N). That implies $G^*$
satisfies the structure $G_2$.
\par If $R^*_2(v^*)\neq\emptyset$,
then all vertices of $R^*_2(v^*)$ have a neighbor in $R^*_1(v^*)$,
otherwise $diam(G^*)\geq 3$. Since $\Gamma_1(v)$ is homogeneous,
every vertex of $R^*_2(v^*)$ is adjacent to every vertex of
$R^*_1(v^*)$. Hence, all vertices of $R^*_2(v^*)$ are twins. This
implies that $\Gamma_1(v)$ is a clique, all vertices of
$\bigcup_{u^*\in R^*_2(v^*)}u^*$ are twins and so
$|R^*_2(v^*)|=1$. In this case $G^*$ is the paw and one of the
degree-$2$ vertices is $v^*$ and the other belongs to
$\Gamma^*_1(v^*)$. Therefore, the structure of $G^*$ is as
Figure~\ref{(3),(4)}(a). Hence, $x^*$ and $v^*$ are adjacent
twins. Therefore, one of them is of type (N), otherwise $x^*=v^*$,
which contradicts $R^*_2(v^*)\neq\emptyset$. Since $\Gamma_1(v)$
is a clique, $x^*,y^*$ are of type (1K). Thus, $v^*$ is of type
(N). If $z^*$ is of type (K), then we choose arbitrary fixed
vertices $x\in x^*,v_1,v_2\in v^*,y\in y^*$ and $z_1,z_2\in z^*$.
Hence, $$r(v_1|\{v_2,z_2\})=(2,2),\, r(x|\{v_2,z_2\})=(1,2),\,
r(y|\{v_2,z_2\})=(1,1),\, r(z_1|\{v_2,z_2\})=(2,1).$$ Thus, the
set $V(G)\backslash\{x,y,z_1,v_1\}$ is a resolving set for $G$,
and $\beta(G)\leq n-4$. Consequently, $z^*$ is of type (1N).
Similarly if $x^*$ is of type (K) and $z^*$ is of type (N), then
$V(G)\backslash\{x,y,z_1,v_1\}$ is a resolving set for $G$ that is
a contradiction. Hence, $G^*$ satisfies the structure $G_3$.
\par Now let $\Gamma_2(v)$ is not homogeneous.
By the same notation as the proof of Fact~\ref{Gamma^*_2<3}, we
can see that the vertices of $N_1(x)$ can only be twins with each
other and $x$, also the vertices of $N_2(x)$ can only be twins
with each other, $x$, and $v$. By Lemma~\ref{R_2=empty}, we have
$R^*_2(v^*)=\emptyset$. Hence, if $|\Gamma^*_2(v^*)|=1$, then
$G^*\cong P_3$, all vertices of $N_2(x)$ are twins with $v$, and
all vertices of $N_1(x)$ are twins with $x$. Thus, $x^*$ is of
type (K), $v^*$ is of type (N), and $r^*_1$ is of any type. In
this case $G^*$ satisfies the structure $G_2$.
\par When $|\Gamma^*_2(v^*)|=2$, there exist three cases.\vspace{4mm}\\
1. The vertex $x$ and all vertices of $N_1(x)$ are twins and there
exist some vertices in $N_2(x)$ which are not twins with $v$ and
they are twins by themselves. Thus, the vertices of
$N_2(x)\backslash v^*$ form exactly one vertex, $y^*$ in $G^*$.
Hence, $x^*$ is of type (K) and $x^*\nsim y^*$. Therefore,
$y^*$ and $v^*$ are twins. Since $v^*\nsim y^*$, one of them is
of type (K). Note that, if $v^*$ is of type (K), then there exist
some vertex $u\in V(G)$ which is adjacent to all vertices of
$\Gamma_1(v)\cup\{v\}$ and is not adjacent to any vertex of
$\Gamma_2(v)$. Hence, $u\in R_2(v)$, which is impossible, because
by Lemma~\ref{R_2=empty}, $R_2(v)=\emptyset$. Thus, $y^*$ is of
type (K). Let $r_1\in r^*_1$, $y\in y^*$, $x\in x^*$ and $v\in
v^*$. Then $(x^*\backslash\{x\})\cup(y^*\backslash\{y\})$ resolves
$\{v,r_1,x,y\}$, this contradicts our assumption $\beta(G)=n-3$.
Consequently, this case does not happen.
\vspace{4mm}\\
2. There exist some vertices of $N_2(x)$ which are twins with $x$
and the rest are twins with $v$, and all vertices of $N_1(x)$ are
twins. Therefore, the vertices of $N_1(x)$ create a vertex
$a^*$ in $G^*$, $a^*\sim x^*$, and $x^*$ is of type (N). Hence,
$G^*$ is the paw, with the leaf $v^*$, the degree-$3$ vertex
$r^*_1$, and degree-$2$ vertices $a^*$ and $x^*$. If $a^*$ is of
type (N), then $V(G)\backslash\{v,r_1,x,a\}$ is a resolving set
for $G$, where $r_1\in r^*_1,\- v\in v^*,\-x\in x^*$ and $a\in
a^*$. This contradiction yields $a^*$ is of type (1K). Because
$R_2(v)=\emptyset$, $v^*$ is not of type (K). Therefore, $v^*$ is
of type (1N). By a similar method as before, we deduce that, if
$a^*$ is of type (K), then $v^*$ and $r^*_1$ can not be of type
(N). Thus, $G^*$ satisfies in structure $G_3$.
\vspace{4mm}\\
3. All vertices of $N_2(x)$ are twins with $v$, and there exist
some vertices in $N_1(x)$ which are not twins with $x$. Hence, the
vertices of $N_1(x)\backslash x^*$ form a unique vertex $a^*$ in
$G^*$, $a^*\sim x^*$, and consequently $G^*$ is the paw. Note that
$v^*$ is the leaf and its type is (N), the vertex $r^*_1$ has
degree $3$, and $x^*,\-a^*$ are degree-$2$ vertices. Since $x^*$
and $a^*$ are adjacent twins, one of them is of type (N). Also,
since all vertices of $N^*_2(x)$ are twins with $v$, $x^*$ is of
type (1K), and so $a^*$ is of type (N). Therefore, $G^*$ has the
structure $G_3$.
\par Finally, if $|\Gamma^*_2(v^*)|=3$, then we have three
following cases.\vspace{4mm}\\
1. Every vertex of $N_2(x)$ is twin with $v$ or $x$, and
$N^*_1(x)$ has two vertices $a^*$ and $b^*$. In this case $a^*$
and $b^*$ are twins, hence $a^*=b^*$, because $N_1(x)$ is
homogeneous. Thus, $|\Gamma^*_2(v^*)|\leq 2$. Therefore this is
not
the case.\vspace{4mm}\\
2. Every vertex of $N_1(x)$ is twin with $x$, and $N^*_2(x)$ has
two vertices $y^*$ and $z^*$. Hence, $y^*$ and $z^*$ are twins.
Since $N_2(x)$ is homogeneous, $y^*=z^*$, which is a contradiction
with $|\Gamma^*_2(v^*)|=3$.\vspace{4mm}\\
3. There exist some vertices in $N_1(x)$ which are not twins with
$x$, also there exist some vertices in $N_2(x)$ which are neither
twins with $x$ nor $v$. In this case, each one of $N^*_1(x)$ and
$N^*_2(x)$ has exactly one vertex $a^*$ and $y^*$, respectively.
If $a^*\nsim y^*$, then $v^*$ and $y^*$ are none-adjacent twins.
Hence one of them is of type (K). Since $R_2(v)=\emptyset$, $v^*$
is of type (1N), and so $y^*$ is of type (K). Let $v\in v^*$,
$y\in y^*$, $a\in a^*$, and $x\in x^*$. Then
$V(G)\backslash\{y,x,r_1,v\}$ is a resolving set for $G$. This
contradiction yields $a^*\sim y^*$, and in consequence, $G^*$ is a
kite with a pendant edge, adjacent to a degree-$3$ vertex. Thus,
$y^*$ and $x^*$ are none-adjacent twins, hence one of them is of
type (K). By symmetry, we can assume $x^*$ is of type (K). Since
$R_2(v)=\emptyset$, $v^*$ is of type (1N). As we see before,
$v^*,\-a^*$, and $r_1^*$ are not of type (N) and $y^*$ is not of
type (KN). Therefore, $G^*$ satisfies the structure $G_7$ and the
proof is completed.}\end{proof}
\vspace{-.7cm}\begin{figure}[ht]\hspace{3cm}
\vspace*{5cm}\special{em:graph fig3.bmp} \vspace*{1.5cm}
\vspace{-2.3cm}
\caption{\label{(3),(4)}
{\footnotesize$|\Gamma^*_2(v^*)|=1$ and $|\Gamma^*_2(v^*)|=2$.}}
\end{figure}
Heretofore we have considered the case
$|R^*_1(v^*)|=1$. Now we investigate the case $|R^*_1(v^*)|=2$.
\begin{pro}\label{case2R^*=2} If $|R^*_1(v^*)|=2$
and $\Gamma_2(v)$ is not homogeneous, then $G^*$ satisfies the
structure $G_7$.
\end{pro}
\begin{proof}{ By the same notation as the proof of
Fact~\ref{Gamma^*_2<3}, let $\Gamma^*_2(v^*)=\{a^*,b^*\}$, where
$a^*$ is of type (1) and $b^*$ is of type (KN). Thus, if $a^*\sim
b^*$, then $b^*$ is of type (N), however $a^*\nsim b^*$ implies
that $b^*$ is of type (K), because $\Gamma_2(v)$ is not
homogeneous. Let $R_1^*(v^*)=\{x^*,y^*\}$. Since $R_1(v)$ has a
none-adjacent vertex to $a$, the vertex $a^*$ has exactly one
neighbor in $R^*_1(v^*)$. By the proof of Fact~\ref{Gamma^*_2<3},
$b^*\sim x^*$ and $b^*\sim y^*$, thus $G^*$ is the $4$-cycle,
$C_4=(v^*,x^*,b^*,y^*,v^*)$ with the pendant edge $x^*a^*$ and
possibly extra edges $a^*b^*$ and $x^*y^*$, see
Figure~\ref{(3),(4)}(b). Because $diam(G^*)=2$, at least one of
the edges $a^*b^*$ and $x^*y^*$ exists. If $a^*\sim b^*$, then
$b^*$ is of type (N). Let $v\in v^*$, $a\in a^*$, $b\in b^*$,
$x\in x^*$, and $y\in y^*$. Consequently, the set $a^*\cup
(b^*\backslash\{b\})$ resolves $\{v,x,b,y\}$, since $b^*$ is of
type (N). This contradiction yields $a^*\nsim b^*$, and so
$x^*\sim y^*$, $a^*$ is of type (1), and $b^*$ is of type (K).
Note that $x^*$ and $y^*$ are not of type (N), otherwise
$\Gamma_1(v)$ is not homogeneous. Also, we can see easily that
$v^*$ is not of type (KN). In this case $G^*$ satisfies the
structure $G_7$.}
\end{proof}
Now, we need only to consider the case $|R^*_1(v^*)|=2$ when
$\Gamma_2(v)$ is homogeneous. In this case, if
$|\Gamma^*_2(v^*)|=1$, then all vertices of $R_1(v)$ are twins and
consequently, $|R^*_1(v)|=1$, which contradicts $|R^*_1(v^*)|=2$.
Therefore, $|\Gamma^*_2(v^*)|\geq 2$.
\begin{lemma}\label{Claim5}
If $\Gamma_2(v)$ is homogeneous and $|R^*_1(v^*)|=2$, then
$R^*_2(v^*)=\emptyset$.
\end{lemma}
\begin{proof} { By Fact~\ref{Gamma^*_2<3} and above argument, we
have $2\leq |\Gamma^*_2(v^*)|\leq 3$. Suppose on the contrary that
$R^*_2(v^*)\neq\emptyset$. Since $diam(G^*)=2$ and $\Gamma_1(v)$
is homogeneous, every vertex of $R^*_2(v^*)$ is adjacent to every
vertex of $R^*_1(v^*)$. In this way all vertices of $R^*_2(v^*)$
are twins, therefore all vertices of $\bigcup_{u^*\in
R^*_2(v^*)}u^*$ are twins, and so $|R^*_2(v^*)|=1$. Let
$R^*_2(v^*)=\{r^*_2\}$, $R^*_1(v^*)=\{x^*,y^*\}$, and
$\{a^*,b^*\}\subseteq\Gamma^*_2(v^*)$. Therefore, $r^*_2$ is
adjacent to the vertices $v^*,x^*$ and $y^*$. Note that
$\Gamma_1(v)$ and $\Gamma^*_1(v^*)$ are cliques, because
$\Gamma_1(v)$ is homogeneous and $r^*_2$ is adjacent to $x^*$ and
$y^*$. Thus, $r^*_2,x^*$ and $y^*$ are of type (1K). Since all
neighbors of $r^*_2$ and $v^*$ are shared, $r^*_2$ and $v^*$ are
adjacent twins. For $r^*_2\neq v^*$, one of them is of type (N).
Because $r^*_2$ is of type (1K), $v^*$ is of type (N).
We choose arbitrary fixed vertices $v_1,v_2\in v^*,r_2\in
r^*_2,x\in x^*,y\in y^*,a\in a^*$ and $b\in b^*$ and set
$T=\{v_1,a,b\}$. Since $a^*\neq b^*$ and $\Gamma_2(v)$ is
homogeneous, one of the vertices $a^*$ and $b^*$has only one
neighbor in $R^*_1(v^*)$ and the other have one or two. Without
loss of generality, we can assume that $x^*$ is the only neighbor
of $a^*$ in $R^*_1(v^*)$, see Figure~\ref{2}(b). Thus,
$$r(v_2|T)=(2,2,2),\quad r(x|T)=(1,1,*),\quad
r(r_2|T)=(1,2,2),\quad r(y|T)=(1,2,1),$$ where $*$ is $1$ or $2$.
However, the four above representations are distinct. This yields
the contradiction $\beta(G)\leq n-4$. Therefore,
$R^*_2(v^*)=\emptyset$. }\end{proof}
\begin{lemma}\label{claim3}
If $\Gamma_2(v)$ is homogeneous and $|\Gamma^*_2(v^*)|=3$, then
there is exactly one vertex $a^*\in\Gamma^*_2(v^*)$ such that
$a^*$ is adjacent to all vertices of $R^*_1(v^*)$.
\end{lemma}
\begin{proof} { Let $|\Gamma^*_2(v^*)|=3$. We have seen that $R^*_1(v^*)$
resolves the set $\Gamma^*_2(v^*)$. Suppose on the contrary that
there is not any vertex of $\Gamma^*_2(v^*)$ adjacent to all
vertices of $R^*_1(v^*)$. Hence, at least one coordinate of the
representation of each vertex in $\Gamma^*_2(v^*)$, with respect
to $R^*_1(v^*)$ is 2. While every coordinate of
$r(v^*|R^*_1(v^*))$ is 1. Therefore, $R^*_1(v^*)$ is a resolving
set for $G^*[R^*_1(v^*)\cup\Gamma^*_2(v^*)\cup\{v^*\}]$ with
cardinality $n(G^*)-4$, since $|\Gamma^*_2(v^*)\cup \{v^*\}|=4$.
It follows that $\beta(G^*)\leq n(G^*)-4$, and
Proposition~\ref{dim G^*<n-t} implies that $\beta(G)\leq n-4$,
which is a contradiction. Therefore, there exists a vertex
$a^*\in\Gamma^*_2(v^*)$ adjacent to all vertices of $R^*_1(v^*)$.
If there exists another vertex $b^*\in\Gamma^*_2(v^*)$ adjacent to
all of $R^*_1(v^*)$, then $a^*$ and $b^*$ are twins, since
$\Gamma^*_2(v^*)$ is homogeneous. This implies that $a^*=b^*$
while $|\Gamma^*_2(v^*)|=3$. Therefore, such a vertex in
$\Gamma^*_2(v^*)$ is unique. }
\end{proof}
\begin{lemma}\label{claim4}
If $\Gamma_2(v)$ is homogeneous and $|R^*_1(v^*)|=2$, then
$|\Gamma^*_2(v^*)|\leq 2$.
\end{lemma}
\begin{proof} {On the contrary, suppose
$|\Gamma^*_2(v^*)|=3$. By Lemma~\ref{claim3}, there exists
exactly one vertex $a^*\in\Gamma^*_2(v^*)$ such that $a^*$ is
adjacent to all vertices of $R^*_1(v^*)$. Let
$R^*_1(v^*)=\{x^*,y^*\}$ and $\Gamma^*_2(v^*)=\{a^*,b^*,c^*\}$.
Each one of vertices $b^*$ and $c^*$ has at least one neighbor in
$R^*_1(v^*)$ and by Lemma~\ref{claim3}, they have exactly one
neighbor in $R^*_1(v^*)$. If their neighbors in $R^*_1(v^*)$ are
same, then they are twins, since $\Gamma^*_2(v^*)$ is homogeneous.
This implies that every pair of vertices $b\in b^*$ and $c\in c^*$
are twins (because $\Gamma_2(v)$ is homogeneous) consequently,
$b^*=c^*$, which contradicts $|\Gamma^*_2(v^*)|=3$. Thus, one of
them, say $b^*$, is (only) adjacent to $x^*$ and the other $c^*$,
is (only) adjacent to $y^*$, see Figure~\ref{2}(a) (dotted edges
may be exist or not).
Now$$r(v^*|\{b^*,c^*\})=(2,2), \, r(x^*|\{b^*,c^*\})=(1,2), \,
r(y^*|\{b^*,c^*\})=(2,1), \, r(a^*|\{b^*,c^*\})=(*,*),$$where $*$
is $1$ or $2$. If $*=1$, then $r(a^*|\{b^*,c^*\})=(1,1)$, and so
$V(G^*)\backslash\{a^*,v^*,x^*,y^*\}$ is a resolving set of size
$n(G^*)-4$ for $G^*$, this contradiction yields $*=2$. Hence
$\Gamma_2(v)$ and $\Gamma^*_2(v^*)$ are independent sets, since
$\Gamma_2(v)$ is homogeneous.
\par Since $R^*_2(v^*)=\emptyset$, if $v^*$ is of type (1N),
then $v^*$ and $a^*$ are twins and every pair of vertices $v\in
v^*$ and $a\in a^*$ are twins (because both $a^*$ and $v^*$ are of
type (1N)), and so $v^*=a^*$, that is a contradiction. Therefore,
$v^*$ is of type (K). For arbitrary fixed vertices $v_1,v_2\in
v^*,x\in x^*,y\in y^*,a\in a^*,b\in b^*$ and $c\in c^*$ and
$T=\{v_1,a,c\}$, we have
$$r(v_2|T)=(1,2,2),\quad r(x|T)=(1,1,2),\quad r(y|T)=(1,1,1),\quad
r(b|T)=(2,2,2).$$ Hence, $V(G)\backslash\{v_2,x,y,b\}$ is a
resolving set for $G$, with cardinality $n-4$. This contradiction
implies that $|\Gamma^*_2(v^*)|\leq 2$.}
\end{proof}
\vspace{-.7cm}\begin{figure}[ht]\hspace{3cm}
\vspace*{5cm}\special{em:graph fig4.bmp} \vspace*{1.5cm}
\vspace{-2.3cm}
\caption{\label{2} \footnotesize$|\Gamma^*_2(v^*)|=3$ and
$|\Gamma^*_1(v^*)|=3$.}
\end{figure}
On account of the above results, we need only assume that
$|\Gamma^*_2(v^*)|=|R^*_1(v^*)|=2$ and $|R^*_2(v^*)|\leq~1$.
\begin{pro}
If $|R^*_1(v^*)|=2$ and $\Gamma_2(v)$ is homogeneous, then $G^*$
satisfies one of the structures $G_4$ to $G_7$.
\end{pro}
\begin{proof}{
Let $R^*_1(v^*)=\{x^*,y^*\}$,
$\Gamma^*_2(v^*)=\{a^*,b^*\}$. Then $G^*$ is as described in
Figure~\ref{4}. If $a^*\nsim b^*$, then $x^*\sim y^*$ and $x^*\sim
b^*$, otherwise $diam(G^*)=3$, a contradiction. Let $G_0$ be the
path $(a^*,x^*,v^*,y^*,b^*)$. Thus, $G^*$ must be one of the
following five graphs:
\begin{description}
\item $H^*_1:=G_0+a^*b^*$, \item $H^*_2:=G_0+a^*b^*+x^*b^*$, \item
$H^*_3:=G_0+a^*b^*+x^*y^*$, \item
$H^*_4:=G_0+a^*b^*+x^*b^*+x^*y^*$, \item
$H^*_5:=G_0+x^*b^*+x^*y^*$.
\end{description}
We fix the vertices $v\in v^*,x\in x^*,y\in y^*,a\in a^*$ and
$b\in b^*$ in each $H^*_i$, $1\leq i\leq 5$. Note that $H^*_1\cong
C_5$. If $G^*\cong H^*_1$, then all vertices of $H^*_1$ are of
(1), otherwise (with a simple computation) $\beta(G)\leq n-4$. In
this case $G^*$ satisfies structure $G_4$.
\par When $G^*\cong
H^*_2$, $x^*$ and $y^*$ are not of type (K), because $\Gamma_1(v)$
is homogeneous and $x^*\nsim y^*$. Similarly $a^*$ and $b^*$ are
not of type (N). If $x^*$ or $y^*$ is of type (N), then
$V(G)\backslash\{v,x,y,b\}$ is a resolving set for $G$, with
cardinality $n-4$. Also $v^*$ of type (N) or (K) yields
$V(G)\backslash\{v,x,y,b\}$ or $V(G)\backslash\{v,x,a,b\}$ is a
resolving set for $G$ of size $n-4$, respectively. These
contradictions show that $G^*$ satisfies the structure $G_5$ if
$G^*\cong H^*_2$.
\par Let $G^*\cong H^*_3$. Since $\Gamma_1(v)$
and $\Gamma_2(v)$ are homogeneous, $x^*\sim y^*$ and $a^*\sim b^*$
imply that $x^*,y^*,a^*,b^*$ are not of type (N). If $a^*$ or
$b^*$ is of type (K), then $V(G)\backslash\{v,x,y,a\}$ or
$V(G)\backslash\{x,y,v,b\}$ is an $(n-4)$-vertex resolving set for
$G$, respectively. Also $v^*$ of type (N) yields the resolving
set, $V(G)\backslash\{x,y,v,b\}$ for $G$. These contradictions
imply that $G^*$ has the structure $G_5$.
\par If $G^*\cong H^*_4$
and one of the vertices $v^*$ or $y^*$ is of type (N), then the
set $V(G)\backslash\{v,x,y,b\}$ or $V(G)\backslash\{x,y,a,b\}$ is
a resolving set for $G$ of cardinality $n-4$. Thus $v^*$ and $y^*$
are of type (1K). By symmetry, the vertices $a^*$ and $b^*$ are of
type (1K), too. If non-adjacent vertices $v^*$ and $b^*$ are of
type (K), then the set $V(G)\backslash\{v,x,y,b\}$ is a resolving
set of size $n-4$. Similarly none-adjacent vertices $a^*$ and
$y^*$ are not of type (K) simultaneously. Also, if none-adjacent
vertices $a^*$ and $v^*$ are of type (K), then
$V(G)\backslash\{v,x,y,a\}$ resolves $G$, which is impossible.
Therefore, none-adjacent vertices are not of the same type (K).
Moreover, if $x^*$ is of type (N), and $y^*$ or $v^*$ is of type
(K), then $V(G)\backslash\{v,x,y,a\}$ is a resolving set for $G$
of cardinality $n-4$. By the same reason, if $x^*$ is of type (N),
then $a^*$ and $b^*$ are not of type (K). Thus, $G^*$ satisfies
the structure $G_6$.
\par Finally, assume that $G^*\cong H^*_5$. Since $v^*\neq b^*$
and these vertices are none-adjacent twins in $H^*_5$, at least
one of them is of type (K). Hence, $v^*$ is of type (K) and $b^*$
is of type (1N), because $\Gamma_2(v)$ is homogeneous and
$a^*\nsim b^*$. If $b^*$ is of type (N), then
$V(G)\backslash\{v,x,y,b\}$ resolves $G$, a contradiction. It
follows that $b^*$ is of type (1). By the similar way as before,
one can see that $a^*$ is of type (1), and both $x^*$ and $y^*$
are of type (1K), and thus $G^*$ satisfies the structure $G_7$.}
\end{proof}
\begin{figure}[ht]\hspace{4cm}
\vspace*{5cm}\special{em:graph fig5.bmp} \vspace*{1.5cm}
\vspace{-2.8cm}
\caption{\label{4} \footnotesize
$|\Gamma^*_1(v^*)|=|\Gamma^*_2(v^*)|=2$ and both $\Gamma^*_1(v^*)$
and $\Gamma^*_2(v^*)$ are homogeneous.}
\end{figure}
\noindent{\bf Case 2.} For each vertex $v\in V(G)$ with
$\Gamma^*_2(v^*)\neq\emptyset$, $\Gamma_1(v)$ is not homogeneous.
\vspace{4mm}\\
We choose a fix vertex $v\in V(G)$ with the property
$\Gamma^*_2(v^*)\neq\emptyset$. Lemma~\ref{one homogeneous}
concludes that in this case, $\Gamma_2(v)$ is homogeneous. For
each vertex $x\in\Gamma_1(v)$, let
$M_1(x):=\Gamma_1(v)\cap\Gamma_1(x)$ and
$M_2(x):=\Gamma_1(v)\cap\Gamma_2(x)$. Since
$M_2(x)\subseteq\Gamma_2(x)$ and $\Gamma_2(x)$ is homogeneous,
$M_2(x)$ is also homogeneous. If $M_1(x)$ is not homogeneous, then
there exist vertices $i$, $j$, and $k$ in $M_1(x)$ such that
$i\sim j$ and $k\nsim j$. Thus, for each pair of vertices $y\in
M_2(x)$ and $c\in\Gamma_2(v)$, we have $r(i|\{v,x,j\})=(1,1,1)$,
$r(k|\{v,x,j\})=(1,1,2)$, $r(y|\{v,x,j\})=(1,2,*)$,
$r(c|\{v,x,j\})=(2,*_1,*_2)$, where $*$, $*_1$ and $*_2$ are $1$
or $2$. However, these representations are distinct, which is a
contradiction. Therefore, $M_1(x)$ is homogeneous.
\begin{pro}\label{claim 3.1} If there exist a vertex
$x\in R_2(v)$ with $M_2(x)\neq\emptyset$, then $G^*$ satisfies the
structure $G_6$.
\end{pro}
\begin{proof} { Since $x\in R_2(v)$, we have
$\Gamma_1(x)=M_1(x)\cup\{v\}$. Note that $v$ is adjacent to all
vertices of $M_1(x)$. Since $M_1(x)$ is homogeneous and
$\Gamma_1(x)$ is not homogeneous, we conclude $M_1(x)$ is an
independent set and contains at least two vertices. Now let $m_1$
and $m_2$ be two arbitrary vertices in $M_1(x)$. Thus, $m_1$
resolves $m_2$ and $v$, hence $m_1$ can not resolve any pair of
vertices in $\Gamma_2(x)$, otherwise the set $\{x,m_1\}$ resolves
at least four vertices. Therefore, $m_1$ is adjacent to all
vertices of $\Gamma_2(x)$ or none-adjacent to all of them. Since
$m_1$ is an arbitrary vertex of $M_1(x)$, all vertices of
$\Gamma_2(x)$ have same neighbors in $M_1(x)$. Note that
$\Gamma_2(x)=M_2(x)\cup\Gamma_2(v)$, because $x\in R_2(v)$. Also
all vertices of $M_2(x)$ are adjacent to $v$, and all vertices of
$\Gamma_2(v)$ are not adjacent to $v$. Thus, every pair of
vertices in $M_2(x)$ and also every pair of vertices of
$\Gamma_2(v)$ are twins. Let $t^*=M_2(x)$ and $s^*=\Gamma_2(v)$ be
the corresponding vertices in $G^*$. Moreover, the vertices of
$M_1(x)$ that are adjacent to all of $\Gamma_2(x)$ are twins and
form a vertex $y^*$ in $G^*$, also the remaining vertices of
$M_1(x)$ are twins with each other and create a vertex $z^*$ in
$G^*$. Therefore, $G^*$ has at most six vertices $v^*$, $x^*$,
$y^*$, $z^*$, $t^*$, and $s^*$, where $x^*$ is adjacent to $v^*$,
$z^*$, and $y^*$. Also, $v^*$ and $y^*$ are adjacent to all
vertices except $s^*$ and $z^*$, respectively. There is no other
edge in $G^*$ except possibly $s^*t^*$, see Figure~\ref{5}(a).
\par If all
of these six vertices exist, then $d(z^*,s^*)=3$, which
contradicts $diam(G^*)=2$. Since $s^*=\Gamma_2(v)$, the vertex
$z^*$ does not exist. It is clear that $y^*$ is of type (N),
because $M_1(x)$ is an independent set of size at least two. Let
$v\in v^*$, $x\in x^*$, $y_1,y_2\in y^*$, $s\in s^*$, and $t\in
t^*$. If $s^*\nsim t^*$, then $v^*\cup x^*\cup t^*\subseteq
\Gamma_2(s)$. But the set $v^*\cup x^*\cup t^*$ is not
homogeneous, and so $\Gamma_2(s)$ is not homogeneous, this
contradiction yields $s^*\sim t^*$. Thus, since $s^*\cup
t^*\subseteq \Gamma_2(x)$, the adjacent vertices $s^*$ and $t^*$
are of type (1K). Moreover, $x^*\cup v^*\subseteq \Gamma_2(s)$
and $x^*\sim v^*$ yields $x^*$ and $v^*$ are of type (1K). But as
we see before, when $y^*$ is of type (N) the other vertices can
not be of type (K), hence they must be of type (1), also, two
none-adjacent vertices are not of the same type (K). Therefore,
$G^*$ satisfies the structure $G_6$. }
\end{proof}
Now let for each vertex $u\in R_2(v)$, one of the sets $M_1(u)$ or
$M_2(u)$ is empty. Note that every vertex of $R_2(v)$ has a
neighbor in $R_1(v)$, otherwise $diam(G)\geq 3$. Hence,
$M_1(u)\neq\emptyset$. Consequently, $M_2(u)=\emptyset$, for each
$u\in R_2(v)$. Since $diam(G)=2$,
$M_1(u)=\Gamma_1(v)\backslash\{u\}$. Therefore, every vertex of
$R_2(v)$ is adjacent to all vertices of $R_1(v)$, $R_2(v)$ is a
clique, and $|R^*_2(v^*)|\leq 1$. We consider the cases $R_1(v)$
is homogeneous or not homogeneous, separately.
\begin{figure}[ht]\hspace{4cm}
\vspace*{5cm}\special{em:graph fig6.bmp} \vspace*{1.5cm}
\vspace{-3.2cm}
\caption{\label{5} \footnotesize
$|\Gamma^*_1(v^*)|=4$ and $|\Gamma^*_1(v^*)|=2$ in Case 2.}
\end{figure}
\begin{pro}\label{claim32}
If for each vertex $u\in R_2(v)$, the set $M_2(u)=\emptyset$ and $R_1(v)$ is homogeneous,
then $G^*$ satisfies the structure $G_2$.
\end{pro}
\begin{proof} { Let $H=G[\{v\}\cup R_1(v)\cup\Gamma_2(v)]$. It is
clear that $H$ is an induced subgraph of $G$ with diameter two. By
Corollary~\ref{sugraph d=2}, $n(H)-3\leq\beta(H)\leq n(H)-2$.
First suppose $\beta(H)=n(H)-2$. Theorem~\ref{n-2} yields $H$ is
one of the graphs $K_{r,s}$, $K_r\vee \overline K_s$, or
$K_r\vee(K_1\cup K_s)$. If $H=K_{r,s}$, then $R_1(v)$ and
$\Gamma_2(v)$ are independent sets. Now, for each $t\in
\Gamma_2(v)$ the set $\Gamma_1(t)$ is a nonempty independent set
in $G$, which is a contradiction with the assumption that
$\Gamma_1(t)$ is not homogeneous, for such a $t$. Consequently,
$H\neq K_{r,s}$. Note that $R_2(v)$ is a clique and all vertices
of $R_2(v)$ are adjacent to all vertices of $R_1(v)$, since for
each vertex $u\in R_2(v)$ the set $M_2(u)$ is empty. Therefore,
$R_1(v)$ is not a clique, otherwise $\Gamma_1(v)$ is homogeneous.
On the other way, if $H=K_r\vee \overline K_s$ or
$H=K_r\vee(K_1\cup K_s)$, then $R_1(v)$ is a clique. This
contradiction yields $\beta(H)=n(H)-3$.
\par Since $R_1(v)$ and $\Gamma_2(v)$ are homogeneous, the graph $H$
with its vertex $v$ satisfies the conditions of Case 1. Thus,
$H^*$ has one of the six structures $G_2$ to $G_7$. If
$R_2(v)=\emptyset$, then $\Gamma_1(v)=R_1(v)$ is homogeneous,
which is a contradiction. Therefore, $R_2(v)\neq\emptyset$.
Hence, we have $R_2(v)$ is a clique and all its vertices are
adjacent to all vertices of $R_1(v)$, consequently, $R^*_1(v^*)$
is not a clique, otherwise $\Gamma_1(v)$ is homogeneous. Note
that, in the structures $G_3$, $G_6$, and $G_7$, $R^*_1(v^*)$ is
a clique, therefore these structures do not occur.
\par In graphs with
structure $G_5$, when both neighbors of $v^*$ have degree $3$,
$R^*_1(v^*)$ is a clique. This implies that, if $H^*$ has the
structure $G_5$, then $v^*$ is adjacent to a degree-$2$ vertex and
a degree-$3$ vertex. Here, if $t^*\in \Gamma^*_2(v^*)$ is the
degree-$2$ vertex and $t\in t^*$, then $\Gamma_1(t)$ is a clique,
contradicts the assumption that $\Gamma_1(v)$ is not homogeneous.
Therefore, $H^*$ has not the structure $G_5$.
\par When $H^*$ has the structure $G_4$, $H^*\cong C_5$. If $a^*$ is a neighbor of $v^*$, then
$R_2(v)\cup a^*$ is a resolving set for $G$ with cardinality
$n-4$. These contradictions imply that $H^*$ can only have the
structure $G_2$.
\par Let $H^*$ satisfies the structure $G_2$.
If $H^*$ has not the condition (a) of the structure $G_2$, then
the only vertex of $R^*_1(v^*)$ is of type (1K), therefore,
$R_1(v)$ is a clique or $|R_1(v)|=1$, hence $\Gamma_1(v)$ is a
clique, which is impossible. Thus, $H^*$ has the condition (a) of
the structure $G_2$, consequently, the degree-$2$ vertex of $H^*$
is of type (N), hence, $R_1(v)$ is an independent set of size at
least $2$. If $v^*$ as a vertex of $H^*$ is of type (N), then by
condition (a), the other leaf of $H^*$ is of type (K) and
consequently, $V(G)\backslash\{v,r_1,r_2,b\}$ resolves $G$, where
$v\in v^*,~r_1\in R_1(v),~r_2\in R_2(v)$, and $b\in \Gamma_2(v)$.
Therefore, $v^*$ as a vertex of $H^*$ is of type (1K). Hence,
since all vertices of $R_2(v)$ are adjacent to all vertices of
$\Gamma_1(v)\cup\{v\}$, $v$ is twin with all vertices of $R_2(v)$.
Consequently, $v^*$ as a vertex of $G^*$ is of type (1K) and $G^*$
satisfies the structure $G_2$. }
\end{proof}
We investigate the case $R_1(v)$ is not homogeneous for two
possibility, $|\Gamma^*_2(v^*)|\geq 2$ and $|\Gamma^*_2(v^*)|=~1$,
separately.
\begin{pro}\label{claim3-3}
If for each $u\in R_2(v)$ the set $M_2(u)=\emptyset$, $R_1(v)$ is
not homogeneous, and $|\Gamma^*_2(v^*)|\geq 2$, then $G^*$
satisfies the structure $G_6$.
\end{pro}
\begin{proof}{ Since $\Gamma_2(v)$ is homogeneous,
$V(G^*)\backslash\Gamma^*_2(v^*)$ is a resolving set for $G^*$.
Hence, Proposition~\ref{dim G^*<n-t} implies that
$\Gamma^*_2(v^*)$ has at most three vertices. Note that, there
exist a vertex $x\in R_1(v)$ such that both sets $M_1(x)\cap
R_1(v)$ and $M_2(x)\cap R_1(v)$ are nonempty sets, because
$R_1(v)$ is not homogeneous. We have $M_1(x)\cap R_1(v)$ and
$M_2(x)\cap R_1(v)$ are homogeneous, because $M_1(x)$ and $M_2(x)$
are homogeneous. On the other hand, all vertices of $R_2(v)$ are
adjacent to all vertices of $R_1(v)$, since for each vertex $u\in
R_2(v)$ the set $M_2(u)$ is empty. Also, $(M^*_2(x)\cap
R^*_1(v^*))\cup\Gamma^*_2(v^*)$ resolves $M^*_1(x)\cap
R^*_1(v^*)$, because all different neighbors of the set
$M^*_1(x)\cap R^*_1(v^*)$ are in the set $(M^*_2(x)\cap
R^*_1(v^*))\cup\Gamma^*_2(v^*)$. Moreover, in the representations
of all vertices of $M^*_1(x)\cap R^*_1(v^*)$ with respect to
$\Gamma^*_2(v^*)$, at least one of the coordinates is $1$. While
all coordinates of $r(v^*|\Gamma^*_2(v^*)$ are $2$. Hence,
$(M^*_2(x)\cap R^*_1(v^*))\cup\Gamma^*_2(v^*)$ resolves
$\{v^*\}\cup(M^*_1(x)\cap R^*_1(v^*))$. Now Proposition~\ref{dim
G^*<n-t} gives $|\{v^*\}\cup(M^*_1(x)\cap R^*_1(v^*))|\leq 3$, and
consequently $|M^*_1(x)\cap R^*_1(v^*)|\leq 2$. Similarly
$|M^*_2(x)\cap R^*_1(v^*)|\leq 2$.
\par Since all neighbors of $\Gamma^*_2(v^*)$ are
in $R^*_1(v^*)$ and $\Gamma_2(v)$ is homogeneous,
$|\Gamma^*_2(v^*)|\geq 2$ implies that there exist vertices
$z^*\in R^*_1(v^*)$ and $t^*\in\Gamma^*_2(v^*)$ such that
$z^*\nsim t^*$. Let $z\in z^*$. Then $z\nsim t$, for each $t\in
t^*$. Since $z$ has a neighbor $t^\prime\in\Gamma_2(v)$, $z$ is
adjacent to all vertices of $R_1(v)\backslash\{z\}$ or not
adjacent to all these vertices, otherwise the set $\{v,z\}$
resolves four vertices of $G$. Moreover, if
$R_1(v)\backslash\{z\}$ is not homogeneous, then there exist three
vertices $i,j,k\in R_1(v)\backslash\{z\}$ such that $j$ resolves
$\{i,k\}$, and so $\{v,z,j\}$ resolves $\{i,k,t,t^\prime\}$, which
is impossible. Thus, $R_1(v)\backslash\{z\}$ is homogeneous.
Therefore, $G[R_1(v)]\cong K_1\vee\overline{K_l}$ or $K_1\cup K_l$
for some positive integer $l\geq 2$, because $R_1(v)$ is not
homogeneous. It follows that all vertices of
$R_1(v)\backslash\{z\}$ have a neighbor and a none-neighbor vertex
in $R_1(v)$. Hence, each vertex of $R_1(v)\backslash\{z\}$ is
adjacent or none-adjacent to all vertices of $\Gamma_2(v)$, since
$\beta(G)=n-3$. But by definition of $R_1(v)$, each vertex of
$R_1(v)$ has a neighbor in $\Gamma_2(v)$. From this, each vertex
of $R_1(v)\backslash\{z\}$ is adjacent to all vertices of
$\Gamma_2(v)$. Thus, all vertices of $R_1(v)\backslash\{z\}$ are
twins, and consequently they form a vertex $y^*$ of type (KN) in
$G^*$, and $z^*$ is a vertex of type (1) in $G^*$. Therefore,
$|R^*_1(v^*)|=2$ and $y^*$ is adjacent to all vertices of
$\Gamma^*_2(v^*)$.
\par If $|\Gamma^*_2(v^*)|=3$, then $z^*$ can
adjacent to one or two vertices of $\Gamma^*_2(v^*)$, however two
vertices of $\Gamma^*_2(v^*)$ coincide, because $\Gamma_2(v)$ is
homogeneous. Hence, $|\Gamma^*_2(v^*)|=2$. Let
$\Gamma^*_2(v^*)=\{r^*,s^*\}$. Then $G^*[V(G^*)\backslash
R^*_2(v^*)]$ is as illustrated in Figure~\ref{5}(b). If both edges
$y^*z^*$ and $r^*s^*$ do not exist, then $d(r^*,z^*)=3$, which
contradicts $diam(G^*)=2$, therefore one of them exists. Let $y\in
y^*,\-r\in r^*,\-s\in s^*$. If $y^*\sim z^*$ and $r^*\nsim s^*$,
then $y^*$ is of type (N), since $\Gamma_1(v)$ is not homogeneous.
Also, $r^*$ is of type (K), otherwise $\Gamma_1(r)=y^*$ is an
independent set, which is impossible. This this leads to the
$(n-4)$-vertex resolving set, $V(G)\backslash\{v,y,z,r\}$, for
$G$. This contradiction shows that $r^*\sim s^*$ in $G^*$.
\par If
$y^*\nsim z^*$, then $y^*$ is of type (K), since $\Gamma_1(v)$ is
not homogeneous. Moreover, $s^*$ and $r^*$ are of type (1K), since
$\Gamma_2(v)$ is homogeneous. However $\Gamma_1(r)$ is a clique,
this contradiction shows that both edges $r^*s^*$ and $y^*z^*$
exist in $G^*$. Therefore, $y^*$ is of type (N), since
$\Gamma_1(v)$ is not homogeneous. Furthermore, $r^*$ and $s^*$ are
of type (1K), because $\Gamma_2(v)$ is homogeneous. Also, since
$\Gamma_2(v)$ is a clique, $v^*$ is of type (1K). As we see
before, when $y^*$ is of type (N), the other vertices of $G^*$ are
not of type (K), hence other vertices are of type (1). Moreover,
two none-adjacent vertices are not of the same type (K).
Consequently, $G^*$ satisfies the structure $G_6$.
}\end{proof}
\begin{pro}\label{claim3-4}
If for each $u\in R_2(v)$ the set $M_2(u)=\emptyset$, $R_1(v)$ is
not homogeneous, and $|\Gamma^*_2(v^*)|=1$, then $G^*$ satisfies
one of the structures $G_8$ to $G_{10}$.
\end{pro}
\begin{proof}{ Since $R_1(v)$ is not homogeneous, there exist a vertex $x\in
R_1(v)$ such that $M_2(x)\neq\emptyset$. If there is no edge
between $R^*_1(v^*)\cap M^*_1(x)$ and $R^*_1(v^*)\cap M^*_2(x)$ or
all vertices of $R^*_1(v^*)\cap M^*_1(x)$ are adjacent to all
vertices of $R^*_1(v^*)\cap M^*_2(x)$, then $|R^*_1(x)|\leq 3$,
because all distinct neighbors of the vertices in these two sets
are in each other. In the same manner as the proof of
Proposition~\ref{claim3-3}, we can see that each one of the sets
$R^*_1(v^*)\cap M^*_1(x)$ and $R^*_1(v^*)\cap M^*_2(x)$ has at
most two vertices. Since $R_2(v)$ is a clique and every vertex of
it is adjacent to all vertices of $R_1(v)$, if
$R_2(v)\neq\emptyset$, then all vertices of $R_2(v)$ are twins
with $v$, and so $v^*$ is of type (K). Also, if
$R_2(v)=\emptyset$, then $v^*$ and the only vertex of
$\Gamma^*_2(v^*)$, say $w^*$, are twins. Since $w^*\nsim v^*$, one
of them is of type (K). Note that $R_2(v)=\emptyset$ shows $v^*$
is not of type (K), consequently, $w^*$ is of type (K). In this
case, by symmetry of $v^*$ and $w^*$ (see Figure~\ref{6}(a)), we
can assume that $v^*$ is of type (K), and so all vertices of
$R_2(v)$ are twins with $v$, and in consequence, without loss of
generality we can assume that $R^*_2(v^*)=\emptyset$.
\par Now let both $R^*_1(v^*)\cap M^*_1(x)$
and $R^*_1(v^*)\cap M^*_2(x)$ have two vertices $\{a^*,b^*\}$ and
$\{y^*,z^*\}$, respectively, see Figure~\ref{6}(a). Since all
distinct neighbors of $z^*$ and $y^*$ are in $\{a^*,b^*\}$,
$\{a^*,b^*\}$ resolves $\{y^*,z^*\}$. Suppose that $v_1,v_2\in
v^*,~y\in y^*,~z\in z^*$, and $w\in w^*$. Then $a^*\cup b^*$
resolves $\{y,z\}$. Also, each coordinate of $r(y|x^*)$ and
$r(z|x^*)$ is $2$. While $r(v_1|x^*)=r(w|x^*)$ is entirely $1$,
$r(v_1|v^*\backslash\{v_1\})$ is completely $1$, and all
coordinates of $r(w|v^*\backslash\{v_1\})$ are $2$. Therefore,
$V(G)\backslash \{y,z,v_1,w\}$ is a resolving set for $G$, which
is impossible. Hence, at least one of the sets $R^*_1(v^*)\cap
M^*_1(x)$ or $R^*_1(v^*)\cap M^*_2(x)$ has one vertex.
\par If $R^*_1(v^*)\cap
M^*_2(x)$ has one vertex and $R^*_1(v^*)\cap M^*_1(x)$ has two
vertices, then $a^*\neq b^*$ and $y^*=z^*$. Thus, the only
distinct neighbor of $a^*$ and $b^*$ is $y^*$, and so $y^*$ is
adjacent to exactly one of them, say $a^*$. Now
$r(v_1|\{v_2,x,a\})=(1,1,1)$, $r(y|\{v_2,x,a\})=(1,2,1)$,
$r(b|\{v_2,x,a\})=(1,1,*)$, and $r(w|\{v_2,x,a\})=(2,1,1)$, where
$*=1$ or $2$. If $*=2$, then $\beta(G)\leq n-4$, therefore $*=1$.
It follows that $x^*$ and $b^*$ are twins and they are adjacent,
hence one of them is of type (N). Since $R^*_1(v^*)\cap M^*_1(x)$
is homogeneous, $a^*\sim b^*$ gives $b^*$ is of type (1K), and so
$x^*$ is of type (N). In this case, $V(G)\backslash\{x,a,y,w\}$ is
a resolving set for $G$, a contradiction.
\par If $R^*_1(v^*)\cap
M^*_1(x)$ has one vertex and $R^*_1(v^*)\cap M^*_2(x)$ has two
vertices, then $a^*=b^*$ and $y^*\neq z^*$. Hence, $a^*$ is
adjacent to exactly one of the vertices $z^*$ and $y^*$, say
$y^*$. Thus, $V(G)\backslash \{v_1,y,z,w\}$ is a resolving set for
$G$. This contradiction yields $|R^*_1(v^*)|\leq 3$. Since
$R_1(v)$ is not homogeneous, $|R^*_1(v^*)|\geq 2$. First, let
$|R^*_1(v^*)|=3$ and $R^*_1(v^*)=\{x^*,a^*,y^*\}$. Then $G^*$ is
as Figure~\ref{6}(b). If $a^*\nsim y^*$, then $x^*$ and $a^*$ are
adjacent twins, consequently, one of them, say $x^*$, is of type
(N). This gives the resolving set, $V(G)\backslash \{v_1,x,y,w\}$
for $G$, therefore $a^*\sim y^*$, and so $x^*$ and $y^*$ are
twins, hence, one of them, say $x^*$, is of type (K). If $y^*$ is
of type (KN), then $V(G)\backslash\{a,x,y,w\}$ is a resolving set
for $G$, thus $y^*$ is of type (1). By the same argument $w^*$ is
of type (1). Note that $a^*$ is of type (1K), otherwise
$V(G)\backslash\{v_1,y,a,w\}$ is a resolving set for $G$. Thus
$G^*$ satisfies the structure of $G_{10}$.
\par When
$|R^*_1(v^*)|=2$, two cases can be happened.\vspace{4mm}\\
1. All vertices of $R_1(v)\cap M_1(x)$ are twins with $x$. In this
case $G^*\cong C_4$ and $x^*=a^*$. Since $R_1(v)\cap M_1(x)$ is
not empty, $x^*$ is of type (K). If $y^*$ is of type (KN), then
$V(G)\backslash\{v_1,x,y,w\}$ is a resolving set for $G$, hence
$y^*$ is of type (1). By the same reason $w^*$ is of type (1),
consequently $G^*$ satisfies the structure
$G_9$.\vspace{4mm}\\
2. All vertices of $R_1(v)\cap M_2(x)$ are twins with $x$. In this
case $G^*$ is the kite and $x^*=y^*$. Since $R_1(v)\cap M_2(x)$ is
not empty, $x^*$ is of type (N). If $a^*$ is of type (N), then
$V(G)\backslash\{v_1,x,a,w\}$ is a resolving set for $G$, hence
$a^*$ is of type (1K). If $w^*$ is of type (KN), then
$V(G)\backslash\{v_1,x,a,w\}$ is a resolving set for $G$. Thus,
$G^*$ satisfies the structure $G_8$, which proves the proposition.
}
\end{proof}
Now, the proof of necessity is completed.
\begin{figure}[ht]\hspace{3.7cm}
\vspace*{5cm}\special{em:graph fig7.bmp} \vspace*{1.5cm}
\vspace{-3.3cm}
\caption{\label{6}\footnotesize
$|\Gamma^*_2(v^*)|=1$ and $R_1(v)$ is not homogeneous.}
\end{figure}
\subsection{\label{Sufficiency} Proof of Sufficiency}
In this section we prove that, if $G$ is a graph of order $n$ and
$diam(G)=2$ such that $G^*$ has one of the structures $G_1$ to
$G_{10}$ in Theorem~\ref{d=2}, then $\beta(G)=n-3$. In the
following we consider each structure $G_1$ to $G_{10}$, as shown
in Figure~\ref{G^*and dim=n-3}, separately. In each case, we
assume that $i\in i^*,~j\in j^*,~c\in c^*,~p\in p^*,~q\in q^*$,
and $u\in u^*$. \vspace{4mm}\\
{$\bf G_1$}. Since $G^*$ has three
vertices, by Proposition~\ref{B(G)>n-n*}, $\beta(G)\geq n-3$. On
the other hand, by Theorem~\ref{n-2}, $\beta(G)\neq n-2$, also $G$
is not a complete graph. Therefore, $\beta(G)\leq n-3$. Hence
$\beta(G)=n-3$. \vspace{4mm}\\
{$\bf G_2$}. Similar to above, we
deduce $\beta(G)=n-3$. \vspace{4mm}\\
{$\bf G_3$}. Let $H,~H_1,~H_2$, and $H_3$ be four graphs such that
their twin graphs are as $G_3$ in Figure~\ref{G^*and dim=n-3}.
Also assume that in $H^*$ the vertex $j^*$ is of type (N) and the
other vertices are of type (1), in $H^*_1$ both $i^*$ and $u^*$
are of type (K), $j^*$ is of type (N), and $p^*$ is of type (1),
in $H^*_2$ both $j^*$ and $p^*$ are of type (N), $i^*$ is of type
(K), and $u^*$ is of type (1), in $H^*_3$, $u^*$ is of type (1)
and other vertices are of type (N).
\par Thus, $H$ is an
induced subgraph of $G$ and $G$ is an induced subgraph of $H_t$,
for some $t$, $1\leq t\leq 3$. Now we get the metric dimension of
$H$ and $H_t$'s, for $1\leq t\leq 3$. Since in $H^*$, the vertex
$j^*$ is of type (N), each resolving set for $H$ contains at least
$|j^*|-1$ vertices of $j^*$. Moreover,
$r(u|j^*\backslash\{j\})=r(i|j^*\backslash\{j\})$, hence,
$j^*\backslash\{j\}$ is not a resolving set for $H$. It is easy to
check that $(j^*\backslash\{j\})\cup\{u\}$ is a resolving set, and
so a basis of $H$. Thus, $\beta(H)=n(H)-3$. Since in $H^*_1$ the
vertices $i^*,~j^*$, and $u^*$ are not of type (1), each resolving
set for $H_1$ contains at least $|i^*|-1,\-|j^*|-1$, and $|u^*|-1$
vertices of $i^*,\-j^*$, and $u^*$, respectively. On the other
hand $r(u|i^*\cup j^* \cup u^*\backslash\{i,j,u\})=r(i|i^*\cup j^*
\cup u^*\backslash\{i,j,u\})$, therefore $i^*\cup j^* \cup
u^*\backslash\{i,j,u\}$ does not resolve $H_1$. It is easy to see
that $i^*\cup j^* \cup u^*\backslash\{j,u\}$ resolves $H_1$,
hence, it is a basis of $H_1$, and consequently
$\beta(H_1)=n(H_1)-3$. By a same argument, we can see
$\beta(H_t)=n(H_t)-3$, $2\leq t\leq 3$. Now, since
$diam(H)=diam(G)=2$, by Corollary~\ref{sugraphH<K<G}, we have
$\beta(G)=n-3$. \vspace{4mm}\\
{$\bf G_4$}. It is clear that
$\beta(C_5)=2$. \vspace{4mm}\\
{$\bf G_5$}. Let $H$ and $R$ be two graphs such that their twin
graphs are as $G_5$, all vertices of $H^*$ are of type (1) and in
$R^*$ both vertices $p^*$ and $q^*$ are of type (1) and other
vertices are of type (K). Therefore, $H$ is an induced subgraph of
$G$ and $G$ is an induced subgraph of $R$. It is clear that
$\beta(H)=2=n(H)-3$. Each resolving set for $R$ contains at least
$|i^*|-1,~|j^*|-1$, and $|u^*|-1$ vertices from $i^*,\-j^*$, and
$u^*$, respectively. Let $W=i^*\cup j^* \cup
u^*\backslash\{i,j,u\}$. Then $r(i|W)=r(j|W)=r(u|W)$, hence, $W$
is not a resolving set for $R$. Also adding one of the vertices
$i,\-j$, and $u$ to $W$, can not provide a resolving set for $R$,
because $\{i,j,u\}$ is a clique in $R$. Since $diam(R)=2$, neither
$p$ nor $q$ can not resolve more than two vertices from the set
$\{i,j,u\}$. Thus, $\beta(R)\geq |W|+2=n(R)-3$. Since $R$ is not
complete graph and none of the graphs in Theorem~\ref{n-2},
$\beta(R)\leq n(R)-3$. Hence, $\beta(R)=n(R)-3$. Since
$diam(H)=diam(G)=2$,
Corollary~\ref{sugraphH<K<G} yields $\beta(G)=n-3$. \vspace{4mm}\\
{$\bf G_6$}. Let $H,~H_1,~H_2$, and $H_3$ be four graphs with twin
graphs as $G_6$ in Figure~\ref{G^*and dim=n-3}. Moreover, assume
that, all vertices of $H^*$ are of type (1), in $H^*_1$ the vertex
$i^*$ is of type (N) and the other vertices are of type (1), in
$H^*_2$ both $p^*$ and $q^*$ are of type (1) and the other
vertices are of type (K), in $H^*_3$ both $p^*$ and $u^*$ are of
type (1) and the other vertices are of type (K). Hence, $H$ is an
induced subgraph of $G$ and $G$ is an induced subgraph of $H_t$,
for some $t$, $1\leq t\leq 3$. It is clear that
$\beta(H)=2=n(H)-3$. Each resolving set for $H_1$ contains at
least $|i^*|-1$ vertices from $i^*$. But $r(u|i^*\backslash
\{i\})=r(j|i^*\backslash \{i\})=r(p|i^*\backslash
\{i\})=r(q|i^*\backslash \{i\})$ and there is no vertex of set
$\{u,j,p,q\}$ such that adding it to $i^*\backslash\{i\}$ provides
a resolving set for $H_1$, hence $\beta(H_1)\geq
|i^*|+1=n(H_1)-3$. Since $H_1$ is not complete graph,
Theorem~\ref{n-2} gives $\beta(H_1)\leq n(H_1)-3$, and so
$\beta(H_1)=n(H_1)-3$. Each resolving set for $H_2$ contains at
least $|i^*|-1,\-|j^*|-1$, and $|u^*|-1$ vertices from
$i^*,\-j^*$, and $u^*$, respectively. Assume $W=i^*\cup j^*\cup
u^*\backslash\{i,j,u\}$. It follows that $r(i|W)=r(j|W)=r(u|W)$,
hence $W$ does not resolve $H_2$. It is easy to check that to
provide a resolving set for $H_2$ we need to add at least two
vertices from $V(H_2)-W$ to $W$. Thus, $\beta(H_2)\geq
|W|+2=n(H_2)-3$. On the other hand, Theorem~\ref{n-2} implies
$\beta(H_2)\leq n(H_2)-3$, and consequently $\beta(H_2)=n(H_2)-3$.
By the same way, $\beta(H_3)=n(H_3)-3$. Since $diam(H)=diam(G)=2$,
Corollary~\ref{sugraphH<K<G} implies $\beta(G)=n-3$.
\vspace{4mm}\\
{$\bf G_7$}. Let $H$ and $R$ be two graphs such that their twin
graphs are as $G_7$ in Figure~\ref{G^*and dim=n-3}. Moreover,
assume that in $H^*$ the vertex $u^*$ is of type (K) and the other
vertices are of type (1), and in $R^*$ both vertices $p^*$ and
$q^*$ are of type (1) and other vertices are of type (K).
Therefore, $H$ is an induced subgraph of $G$ and $G$ is an induced
subgraph of $R$. It is clear that each resolving set for $H$
contains at least $|u^*|-1$ vertices of $u^*$. Moreover,
$r(u|u^*\backslash\{u\})=r(i|u^*\backslash\{u\})=r(j|u^*\backslash\{u\})$
and to provide a resolving set for $H$, we must add at least two
vertices from the set $\{u,i,j,p,q\}$ to $u^*\backslash\{u\}$,
hence $\beta(H)\geq |u^*|+1=n(H)-3$. Also by Theorem~\ref{n-2},
$\beta(H)\leq n(H)-3$, thus $\beta(H)=n(H)-3$. Since $i^*,\-j^*$,
and $u^*$ are not of type (1) in $R^*$, each resolving set for $R$
contains at least $|i^*|-1,~|j^*|-1$, and $|u^*|-1$ vertices of
$i^*,\-j^*$, and $u^*$, respectively. For $W=i^*\cup j^* \cup
u^*\backslash\{i,j,u\}$, we have $r(i|W)=r(j|W)=r(u|W)$, hence $W$
is not a resolving set for $R$. To provide a resolving set for
$R$, we need to add at least two vertices from the set
$\{u,i,j,p,q\}$ to $W$, and consequently, $\beta(R)\geq
|W|+2=n(R)-3$. Theorem~\ref{n-2} shows that $\beta(R)\leq n(R)-3$,
hence $\beta(R)=n(R)-3$. Since $diam(H)=diam(G)=2$,
Corollary~\ref{sugraphH<K<G} yields $\beta(G)=n-3$. \vspace{4mm}\\
{$\bf G_8$}. Let $H$ and $R$ be two graphs such that their twin
graphs are as $G_8$ in Figure~\ref{G^*and dim=n-3}, where in $H^*$
both $j^*$ and $p^*$ are of type (1), $u^*$ is of type (K) and
$i^*$ is of type (N), and in $R^*$ both vertices $j^*$ and $u^*$
are of type (K), $i^*$ is of type (N) and $p^*$ is of type (1).
Then $G$ is one of the graphs $H$ or $R$. Theorem~\ref{n-2} shows
that $\beta(H)\leq n(H)-3$ and $\beta(R)\leq n(R)-3$. Note that
each resolving set for $H$ contains at least $|u^*|-1$ and
$|i^*|-1$ vertices from $u^*$ and $i^*$, respectively. If
$S=(i^*\cup u^*)\backslash\{i,u\}$, then $r(u|S)=r(j|S)$.
Therefore, $\beta(H)\geq |S|+1=n(H)-3$, and so $\beta(H)=n(H)-3$.
It is clear that every resolving set for $R$ contains at least
$|i^*|-1,~|j^*|-1$, and $|u^*|-1$ vertices of $i^*,~j^*$, and
$u^*$, respectively. Let $W=i^*\cup j^* \cup
u^*\backslash\{i,j,u\}$. Then $r(j|W)=r(u|W)$, hence $W$ is not a
resolving set for $R$, and so $\beta(R)\geq |W|+1=n(R)-3$. It
follows that $\beta(R)=n(R)-3$. Consequently, $\beta(G)=n(G)-3$.
\vspace{4mm}\\
{$\bf G_9$}. Let $G^*$ be as $G_9$ in
Figure~\ref{G^*and dim=n-3}, where $i^*$ and $u^*$ are of type (K)
in $G^*$, and the other vertices are of type (1). Each resolving
set for $G$ contains at least $|i^*|-1$ and $|u^*|-1$ vertices of
$i^*$ and $u^*$, respectively. For $W=i^*\cup
u^*\backslash\{i,u\}$, we have $r(i|W)=r(u|W)$, hence $W$ is not a
resolving set for $G$, and so $\beta(G)\geq |W|+1=n(G)-3$.
Theorem~\ref{n-2} implies $\beta(G)\leq n(G)-3$. Consequently,
$\beta(G)=n(G)-3$. \vspace{4mm}\\
{$\bf G_{10}$}. Let $H$ and $R$ be two graphs such that their twin
graphs are as $G_{10}$ in Figure~\ref{G^*and dim=n-3},
furthermore, in $H^*$ both $u^*$ and $c^*$ are of type (K) and
other vertices are of type (1), and in $R^*$ both vertices $i^*$
and $p^*$ are of type (1) and other vertices are of type (K). Then
$G$ is one of the graphs $H$ or $R$. Theorem~\ref{n-2} shows that
$\beta(H)\leq n(H)-3$ and $\beta(R)\leq n(R)-3$. Note that each
resolving set for $H$ contains at least $|u^*|-1$ and $|c^*|-1$
vertices from $u^*$ and $c^*$, respectively. If $S=c^*\cup
u^*\backslash\{c,u\}$, then $r(u|S)=r(j|S)=r(c|S)$, therefore $S$
does not resolve $H$. To provide a resolving set for $H$, we need
to add at least two vertices from the set $\{u,i,j,c,p\}$ to $S$,
and so $\beta(H)\geq |S|+2=n(H)-3$, hence $\beta(H)=n(H)-3$. It is
clear that every resolving set for $R$ contains at least
$|u^*|-1,\-|j^*|-1$, and $|c^*|-1$ vertices from $u^*,\-j^*$, and
$c^*$, respectively. Let $W=u^*\cup j^* \cup
c^*\backslash\{u,j,c\}$. Hence, $r(u|W)=r(j|W)=r(c|W)$, therefore
$W$ is not a resolving set for $R$. Clearly, to provide a
resolving set for $R$, we must add at least two vertices from the
set $\{u,i,j,c,p\}$ to $W$, hence $\beta(R)\geq |W|+2=n(R)-3$,
thus $\beta(R)=n(R)-3$. Consequently, $\beta(G)=n(G)-3$.\\ This
completes the proof of the sufficiency of
Theorem~\ref{d=2}.\hfill{$\rule{2mm}{2mm}$}
|
\section{Introduction}
The modern research academic is judged as never before: a large variety of
metrics are now employed to determine the worth and merit of researchers,
particularly when it comes to hiring. Anecdotally, one of the chief metrics used
is the Hirsch index ($h$-index; Hirsch 2005).
The $h$-index is formally defined as follows:
``A scientist has index $h$ if $h$ of
his or her $N_p$ papers have at least
$h$ citations each and the other
($N_p - h$) papers have $\leq h$ citations each'' (Hirsch 2005).
Its modest
simplicity is probably a prime factor in its rapid pick-up
by major publishers (Ball 2005; Anon 2005).
Moreover, this index is particularly useful as it has superior predictive power
(in terms of productivity) for the
future of researchers compared to the total number of career citations, career
publications and mean citations per paper (Hirsch 2007).
Although other metrics and analyses exist (cf.\ Pearce 2004; Kurtz et al.\ 2005;
Egghe 2006;
Kosmulski 2006;
Jin 2006;
Blustin 2007;
Jin et al.\ 2007;
Bornmann, Mutz \& Daniel 2008;
Wu 2010; Zyczkowski 2010), the $h$-index remains as the most prominent
of its class in the field.
Recently, Conti et al.\ (2011) presented work on the Astronomer's H-R Diagram
(number of Google search results versus citations and $h$-index) for members
of the American Astronomical Society (AAS). Contained within that presentation
are a number interesting concepts: a top-ten list of $h$-index of AAS members
(spanning the range $94<h<118$) and the $h$ indices of all AAS members.
This work is motivated by the Conti et al.\ (2011) presentation and seeks
to determine the typical range of $h$-index in Australian Astronomy which
may be of use for future employers and employees in the community.
The format of this work is as follows.
In Section~2, we give an overview of the dataset that we use: the membership
of the Astronomical Society of Australia.
In Section~3 we determine percentiles of the $h$-index distribution
for a variety of ASA membership categories, including students.
To attempt to normalize relative to opportunity, we
re-evaluate the $h$-index distribution as a function
time elapsed since Ph.D.\ award date in Section~4. Our conclusions are
presented in Section~5.
\section{Data}
To determine the $h$-index of Australian astronomers, we make use of the
Astronomical Society of Australia (ASA) membership list. The membership
list is a fair representation of the Australian astronomical community: the majority
of professional astronomers are members.
Membership of the ASA comes in several difference categories which we
use a single letter to abbreviate and detail in
Table~\ref{tab:cats}. The advantage of the ASA membership list is that
we can differentiate different grades of members (i.e.\ amateurs from
professional astronomers who actively publish) to better probe
the $h$-index in these sub-categories.
\begin{table*}
\begin{center}
\caption{Categories of ASA membership, showing the number of members (N) in each
category and the fraction of each that were excluded (X) from the subsequent analysis
due to name confusion (see Section~2).
We ignore the extra categories of
Associate Society member (i.e.\ persons who are members of other learned societies that
are likely to not possess Ph.D.'s in astronomy) and corporate members.
It is important to note that individual researchers can belong
to multiple categories; e.g.\ retired, overseas fellows.
}\label{tab:cats}
\begin{tabular}{llll}
\hline Category & N & X & Notes \\
\hline
M & 262 & 0.10 & Member: Full professional member with a Ph.D.\ in astronomy or related discipline. \\
F & 81 & 0.10 & Fellow: Senior members of the community with potentially decades of experience. \\
S & 164 & 0.09 & Student: Post-graduate students studying toward Ph.D.'s in astronomy. \\
H & 15 & 0.00 & Honourary: Elected by the ASA council for distinguished contributions. \\
R & 54 & 0.11 & Retired.\\
O & 56 & 0.10 & Overseas.\\
A & 14 & 0.00 & Associate Members: \\
& & & Educators, communicators and amateur astronomers lacking a Ph.D.\\
\hline
\end{tabular}
\medskip\\
\end{center}
\end{table*}
For each ASA member, we then implement a search in NASA's
Astrophysics Data System (ADS) to return a list of all refereed publications.
We then sort this list according to citations to determine the $h$-index for
each ASA member. We note for prosperity that these searches were implemented
on 24th--25th January 2011 and were correct on a best-efforts basis as of said
date range.
A big issue in this methodology is attempting to tie down each individual to
unique entries in ADS.
Although the present author is blessed with a very rare surname,
others in the community are not. For more common surnames, we use the
first name and the middle initials to help determine the $h$-index
of specific researchers, including attempting common substitutions
for first names; e.g., `Bill' for
`William', etc.
However, for the very common surnames (e.g.\ Smith),
this is not always possible. Therefore, the subsequent analysis in this
work does not include any names for which we could not adequately
differentiate a single individual in the literature in a reasonable amount
of time.
This affects $\sim$10\% of the membership list and is labelled
X in Table~\ref{tab:cats}.
We caution that
the subsequent analysis should therefore be regarded as incomplete:
the inclusion of these names could increase or decrease the relative
rankings of individuals within the ASA membership.
We also note that we make no attempt to exclude self-citations
in our analysis (e.g.\ Pimbblet 2011). Finally, it may be
the case that some of the categories may not be up-to-date due
to (e.g.) student members gaining their Ph.D.'s and either not
upgrading to full membership status immediately or the list itself not
being updated immediately.
\begin{figure}
\begin{center}
\includegraphics[scale=1, angle=0, width=3.4in]{ASAhist.ps}
\caption{Histogram of $h$-index for all ASA members, excluding
those for whom name confusion could not be resolved.}\label{fig:hist}
\end{center}
\end{figure}
\section{$h$-index by ASA Membership Category}
This simplest point of departure for the $h$-index analysis is to
pull out the top ten: those people who could rightly be called academic
giants in their own right in the community.
To do this, we simply rank all professional
members who are not based
overseas (i.e.\ categories M+F+S+H+R; Table~\ref{tab:cats}).
This top ten list is presented in Table~\ref{tab:topten}.
Although we refrain from commenting on individuals in this
list, it is instructive to compare it to Conti et al.'s (2011)
list. The $h$-index values for the top ten AAS members is much
higher than the Australian ($94 < h < 118$ versus $53 < h < 77$).
Examination of Conti et al.'s (2011) figures suggests that membership
of very large observational programmes such as the Sloan Digital Sky
Survey (SDSS; e.g.\ Abazajian et al.\ 2009) can boost researcher's
$h$-index above mean values. It is certainly the case that the
Australian top ten is dominated by non-SDSS professionals and
we therefore suggest that most of the difference seen between
the two samples could be due to this effect. Indeed, seven of
the AAS's top ten (Conti et al.\ 2011)
are contained in the author list of Abazajian et al.\
(2009). However, we do note that the Australian top ten does
contain a number of members of other consortia (not as large or
extensive as SDSS) such
as the 2dF Galaxy Redshift Survey (e.g.\ Colless et al.\ 2001).
Moreover, the majority of the listed researchers
in Table~\ref{tab:topten} also feature
in Thomson Reuters' ISI's highly cited list (isihighlycited.com)
for space science.
\begin{table}[h]
\begin{center}
\caption{Top 10 $h$-index for Australian Astronomy, excluding
overseas professionals.}\label{tab:topten}
\begin{tabular}{lll}
\hline Rank & Name & $h$ \\
\hline
=1 & Ken FREEMAN & 77 \\
=1 & Jeremy MOULD & 77 \\
3 & Karl GLAZEBROOK & 71 \\
4 & Dick MANCHESTER & 68 \\
5 & Michael DOPITA & 64 \\
6 & Warrick COUCH & 61 \\
7 & Matthew COLLESS & 56 \\
8 & Brian SCHMIDT & 54 \\
=9 & Mike BESSELL & 53 \\
=9 & Joss BLAND-HAWTHORN & 53 \\
\hline
\end{tabular}
\medskip\\
\end{center}
\end{table}
But what of the rest of the community? In Figure~\ref{fig:hist} we
display a histogram of $h$-index for all ASA members. This graph
is dominated by those members having a zero $h$-index or slightly
above, much as the AAS community is (Conti et al.\ 2011).
The vast majority of these members are student members, many of whom
are likely to not have published. Even if they have published, the
duration of the Ph.D.\ may mean that sufficient time has not elapsed
to gain large numbers of citations and that only the very exceptional
papers produced by students garner large number of citations immediately.
Clearly students in present-day
collaborations such as WiggleZ (Drinkwater et al.\ 2010)
will benefit from this effect much in the same way as SDSS members
receive a boost.
To analyse the content of Figure~\ref{fig:hist} in a more in-depth manner,
we now create sub-samples of the ASA membership according to grade
and determine various percentiles of the $h$-index distribution.
These percentiles are presented in Table~\ref{tab:cents}. We do
not present results for
individual categories H, O, R and A due to low numbers.
This can be seen in the relatively tiny difference between
the percentiles quoted for M+F+R-O versus M+F-R-O samples
in Table~\ref{tab:cents}.
We start by discussing the student membership result.
At the upper echelons, students appear to have an $h$-index
comparable of junior professionals. But a careful analysis
of the membership list reveals that this is exactly what these
students are: junior professionals who should be in the M
category.
We argue that anything above the 90th percentile
for the S category should be regarded with suspicion.
Naturally, the fellows occupy much higher $h$-index values
than the regular members do. The effect of adding or removing the
retirees from the M+F sample is modest: the most noticeable effect
is at the upper echelons of the scale.
However, the major problem of this analysis is that it does not
specify the $h$-index relative to opportunity. To remedy this,
we now try to divide up the ASA membership according to years
since the award of a Ph.D.
\begin{table*}
\begin{center}
\caption{Percentiles of $h$-index distribution by ASA membership grade.
N gives the total number of members for each row.
}\label{tab:cents}
\begin{tabular}{rlllllllll}
\hline
Category & N & \multicolumn{7}{c}{Percentile} \\
& & 25 & 50 & 75 & 90 & 95 & 97.5 & 99 \\
\hline
M & 235 & 6.0 & 12.0 & 21.0 & 32.5 & 43.0 & 47.8 & 52.4 \\
F & 73 & 21.0 & 30.0 & 39.8 & 50.0 & 59.2 & 69.8 & 75.4 \\
S & 149 & 0.0 & 0.0 & 2.0 & 4.0 & 5.0 & 6.7 & 9.5 \\
M+F-R-O & 227 & 7.0 & 15.0 & 28.0 & 41.0 & 45.0 & 53.0 & 63.7 \\
M+F+R-O & 266 & 6.0 & 15.0 & 28.0 & 40.0 & 45.4 & 53.3 & 68.8 \\
M+F+S-R-O & 374 & 1.0 & 5.0 & 19.0 & 33.6 & 42.3 & 46.7 & 59.7 \\
\hline
\end{tabular}
\medskip\\
\end{center}
\end{table*}
\section{$h$-index by Years Since Ph.D. Award}
Even the award year of a Ph.D.\ must be regarded with healthy
suspicion as a metric for performance relative to opportunity.
This is especially true for early-career researchers who may complete their
Ph.D.\ whilst undertaking their first post-doctoral placement
and for the many researchers who have had significant time
away from the profession; the present author included.
To determine the award date of the Ph.D., we use the results
from ADS where available. If the Ph.D.\ is not listed in
ADS, then we use the date of the second
first-author refereed publication by the
member as a compromise proxy for this date, given the
distribution of the S sample in Table~\ref{tab:cents}. This date
was determined for all ASA members in the M+F+R-O category.
Where no date could be determined by either method, the
member was simply removed from the list. This may
have the effect of meaning that the percentiles for
this sample are upper limits as we have missed
doctoral researchers who have few first author publications.
We present the percentiles of this distribution in Table~\ref{tab:centsy}.
The results show a fairly steady progression as one increases
in seniority from Ph.D.\ award date without any
obvious discrepancies, as may be expected.
However, one comment to be made is that there seems many
less young professionals in the samples than there perhaps
should be (given the numbers in more senior years).
This tentatively suggests that new recruits to Australian
Astronomy may not be joining the ASA immediately.
\begin{table*}
\begin{center}
\caption{Percentiles of $h$-index distribution by years since Ph.D.\ award
(explicitly: 2011 minus the award date).
Only members in M+R+F-O categories are included.
N gives the total number of members for each row.
}\label{tab:centsy}
\begin{tabular}{rlllllllll}
\hline
Years since & N & \multicolumn{7}{c}{Percentile} \\
Ph.D.\ award & & 25 & 50 & 75 & 90 & 95 & 97.5 & 99 \\
\hline
0--5 & 32 & 5.0 & 6.0 & 9.0 & 10.0 & 13.2 & 15.8 & 16.3 \\
6--10 & 42 & 8.5 & 14.0 & 19.0 & 22.0 & 25.0 & 27.8 & 29.7 \\
11--20 & 74 & 12.5 & 19.0 & 28.5 & 39.6 & 43.4 & 53.1 & 58.4 \\
$>$20 & 94 & 12.0 & 25.0 & 38.0 & 47.2 & 55.1 & 62.9 & 77.0 \\
\hline
\end{tabular}
\medskip\\
\end{center}
\end{table*}
Further, not all areas and sub-disciplines
of science and astronomy may be equal. Those researchers involved
in (e.g.) instrumentation may have a very different $h$-index
distribution to those researching observational cosmology (particularly
those in larger-sized consortia).
\section{Conclusions}
This work has presented an analysis of the $h$-index distributions
for present members of the ASA. As well as deriving a top ten
(Table~\ref{tab:topten}), we
have presented the percentiles for various sub-samples of the
ASA's membership, including student statistics (Table~\ref{tab:cents}).
We have also attempted to analyse the distribution relative to
opportunity by detailing the percentiles by time elapsed since
Ph.D.\ award date (Table~\ref{tab:centsy}).
Clearly the $h$-index is a crude estimator of the value
of a researcher and should not be used in isolation to other
metrics, even if it is a good predictor of future
productivity (Hirsch 2007).
It will be instructive to re-visit this analysis
in future years or decades to determine how the field
has changed.
We terminate this work with a $caveat\ emptor$: there are known deficiencies
in this analysis such as numerous
missing persons (who are not ASA members)
whose statistics may alter the results presented.
We have tried to be up-front with various caveats throughout this work,
but there may yet be unknown unknowns present as well.
Further, there may exist transcription errors that went
un-detected during the data assembly stage. However, as far as possible,
we believe the numbers quoted in this work are accurate.
\section*{Acknowledgments}
KAP thanks Bryan Gaensler for tweeting about Conti et al.'s presentation
from the 2011 AAS meeting, and Michael J.\ Morgan for
many discussions on how to interpret the $h$-index
in the context of Australian Astronomy
which inspired this present work. I also thank the
anonymous referee for a positive review of the manuscript that
has improved its content.
This research has made use of NASA's Astrophysics Data System Bibliographic Services.
|
\section{Introduction}
The special electronic structure of graphene shows up in its electronic properties. \cite{dassarma, eerev} Most attention has been paid to the electronic conductivity \cite{castroneto09} which, due to the strong energy dependence of the density of states, can be tuned significantly with a gate voltage. For dirty graphene, this gate dependence is furthermore modified by elastic scattering \cite{hwang07} due to potential inhomogenities forming charge puddles. At high voltages, also inelastic scattering due to optical phonons appears \cite{meric08}, but the low-energy inelastic scattering due to electron-electron (e-e) \cite{mirlin, mueller} or electron- acoustic phonon scattering (e-ph) \cite{tse, Janne} does not directly influence the conductivity because the mean free paths for them are typically larger than the elastic mean free path. \cite{dassarma, castroneto09}
In this paper we study the effect of low-energy inelastic scattering in graphene by using Josephson critical current measurements to determine heat transport in the system. The idea is to apply a heater voltage and to measure the increased temperature via nearby thermometers. The measurement is performed at sub-kelvin temperatures and low voltages, thereby providing access to low-energy inelastic scattering processes and allowing to disregard scattering from optical phonons. To perform the measurement we use three superconducting electrodes fabricated on graphene. Two of them (thermometer electrodes, C and R in Fig.~\ref{fig:set}) lie close to each other, so that we can measure a finite supercurrent through them, and the third one (heater electrode, L in Fig.~\ref{fig:set}) is used for heating the system. The supercurrent is sensitive to the electron temperature or, more accurately, to the electron distribution function on the graphene region between the superconductors \cite{du08} and therefore acts as an electron thermometer.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{./set_cdr.eps}
\caption{(color online). (a) Optical image of the studied graphene flake with three Ti/Al electrodes marked L, C and R. The contour of the flake is highlighted by a a dashed line. (b) Differential resistance of the long (L-C) and short (C-R) sections as a function of gate voltage at $T=4$~K.}
\label{fig:set}
\end{figure}
We make two measurements: First we measure the supercurrent as a function of temperature, and thus calibrate the thermometer and find parameters for a microscopic theory describing the supercurrent in the junction. Second, we apply a voltage to the heater electrode, supplying heat into the electron system, and measure again the supercurrent in the presence of the heater voltage. The magnitude of the supercurrent in the presence of the heater voltage is sensitive to the strength of inelastic relaxation inside graphene, allowing us to measure it.
The Joule heat generated in the presence of a bias voltage is dissipated either to the electrodes or to phonons.\cite{tse, Janne, kubakaddi09} However, since the electron-acoustic phonon coupling is so weak in graphene, most of the heat escapes into the electrodes, even though this process is blocked at low energies by the superconducting gap $\Delta$. The escape is possible as a result of processes transferring excitations from low energies to above the gap, in particular multiple Andreev reflections \cite{MAR-S} and e-e scattering. The former tends to broaden the electron distribution below the gap and thus to increase the effective temperature at these energies, while the latter drives the distribution towards a quasiequilibrium (Fermi function) form having a lower effective temperature. Our thermometer is most sensitive to the distribution at energies below $\Delta$, and therefore it is a sensitive probe of the e-e scattering strength in graphene. But to extract the magnitude of the e-e scattering we have to abandon the simple effective temperature description used for example in Ref.~\onlinecite{roddaroup10} and rather solve a full kinetic Boltzmann equation, with e-e scattering included explicitly with a collision integral.
This paper is organized as follows: In Sec.~\ref{sec:exps}, we describe the experimental setup used to carry out the measurements. In Sec.~\ref{sec:thry}, we formulate our theoretical model and consider the different sources of inelastic relaxation. In Sec.~\ref{sec:res} we combine the theoretical and experimental results to provide an estimate for the strength of relaxation in the system. Finally, we discuss the implications of these results in Sec.~\ref{sec:dis}.
\section{Experiments}\label{sec:exps}
An optical image of the studied monolayer graphene sample is shown in Fig.~\ref{fig:set}(a). The $2.8$ $\mu$m long and $4.0$ $\mu$m wide graphene area in between leads L and R is partially interrupted by a $1.0$ $\mu$m wide lead C. This latter lead is $1.7$ $\mu$m and $0.4$~$\mu$m far from leads L and R, respectively, with all the distances measured between leads' internal edges. The graphene flake has been exfoliated with a semiconductor wafer dicing tape and deposited on top of a 250~nm thick SiO$_2$ layer. The oxide isolates the graphene flake from a highly {\it p}-doped Si substrate used as a back gate in our experiments. Three Ti/Al (10~nm/50~nm) metallic contacts were patterned by using standard electron-beam lithography techniques, and evaporated in ultra-high vacuum. A $10^{-10}$~mbar vacuum during the metal evaporation guarantees highly transparent contacts, which are needed to observe proximity-induced supercurrents. The sample was measured at low temperatures, down to 80~mK, in a dry dilution cryostat BF-SD250 from Bluefors. The sample contacts are electrically connected to room temperature electronics via one-meter long thermocoaxes, low-pass RC filters (cut-off frequency of 1~kHz) and one-meter twisted pairs, protecting the sample from the room-temperature electrical noise. Below the critical temperature of $\sim600$~mK, the Ti/Al leads become superconducting, resulting in the formation of three different superconductor-graphene-superconductor (SGS) junctions.
Figure~\ref{fig:set}(b) shows the gate voltage ($V_g$) dependence of the differential resistance ($R_d$) of the graphene sections L-C, C-R and of the whole flake (L-R), at 4~K. The entire section L-R presents a peak in $R_d$ at a gate voltage $V_{\rm CNP}=-11$~V. This resistance peak is associated with a minimum charge carrier density and takes place at the charge neutrality point (CNP). The negative value of $V_{\rm CNP}$ indicates that the flake is {\it n}-doped in the absence of gate voltage. The change in the resistance is smaller in the {\it p}-doped region than in the {\it n}-doped one. This asymmetry comes from the {\it n}-doping by the Ti/Al leads \cite{lee, muellert, Janne2} and the resulting formation of {\it p-n} junctions when bulk of the graphene is {\it p}-doped by the gate for $V_g<V_{\rm CNP}$. For $V_g>V_{\rm CNP}$, we estimate from the Drude model a mobility of $3500$~cm$^2$V$^{-1}$s$^{-1}$ and a mean-free-path of $70$~nm at $V_g=30$~V. The electrical transport through the studied sections is therefore diffusive. The resistance $R_N$ of the short section (C-R) is continuously decreasing with $V_g$ and no resistance peak is observed within the investigated gate voltage range. The short section is thus always {\it n}-doped. This is because the contacts affect the sample on a scale of micrometers so that the average doping in the short section is stronger.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{./IVs_80mK_diff_gate.eps}
\caption{(color online). Current-voltage characteristic of the short section C-R at $V_g=-30$~V obtained by successively sweeping up (blue curve) and down (red curve) the bias current at 80~mK. The SGS junction jumps from the zero-voltage state into the resistive state at the critical current $I_c$ and comes back into its initial state at the retrapping current $I_r$. Inset: $IV$ curve of the short section C-R at $V_g=30$~V and T=80~mK.}
\label{fig:IVs_diff_gate}
\end{figure}
We now focus on the current-voltage ($IV$) characteristics of the long and short sections measured at 80~mK for different gate voltages. The $IV$ curves generally contain a supercurrent branch characterized by a zero-voltage state when the bias current is kept below the critical current $I_c$ (see Fig.~\ref{fig:IVs_diff_gate}). At $I_c$, the SGS junction jumps into the resistive state. The initial superconducting state is recovered when the current is biased below the so-called retrapping current $I_r$. For the short sample, at 80~mK, $I_r$ always differ from $I_c$ leading to a hysteretic $IV$ curve. The critical current $I_c$ depends on the gate voltage and decreases when the normal resistance increases as seen in Fig.~\ref{fig:set}. When the gate voltage is tuned from $-30$ to $30$~V, the critical current increases from $47$ to $120$~nA and the normal resistance is lowered from 708 to 337~$\Omega$, respectively.
\begin{comment}
During this gate voltage change the characteristic $I_cR_N$ product rises from 30~$\mu V$ to $40$~$\mu V$. For the long sample, $I_cR_N$ is around 11~$\mu$V at the CNP and 14~$\mu$V at $V_g=30$~V.
The value of the $I_cR_N$ product characterizes the regime of the diffusive SGS junction. In the short regime, the coherence length $\xi$ of the Andreev pair in the graphene flake is by definition much longer than the junction length $L$, and $I_cR_N=2.1\Delta/e$. The superconducting gap $\Delta$ has been extrated from conductance measurements and is equal to $100$~$\mu$eV. In our experiments, $I_cR_N$ stays always below 2.1$\Delta/e$. We deduce that $\xi$ is on the order of $L$, and that our SGS junctions are in the long regime.
\end{comment}
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{./IV_long_sections_cdr.eps}
\caption{(color online). $IV$ curves of sections L-R and L-C at $V_g=-3.4$~V and $T=80$~mK. Three voltage transitions are highlighted by the arrows.}
\label{fig:IV_long_section}
\end{figure}
Fig.~\ref{fig:IV_long_section} shows the current-voltage characteristics of the sections L-C and L-R at $V_g=-3.4$~V. Both $IV$ curves present a supercurrent branch with the same critical current of 16~nA, and identical differential resistance values (650~$\Omega$) at sufficiently low bias. This is understood by the presence of a supercurrent through the short section (C-R) keeping leads C and R at the same potential. The transition of the short sample into the resistive state is identified by a second voltage jump. This jump takes place at a critical current of $156$~nA for the section L-C and $104$~nA for the section L-R. The differential resistances increase to 685 and 810~$\Omega$ for sections L-C and L-R, respectively.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{./IVs_curves_diff_temperatures_cdr.eps}
\caption{(color online). Temperature dependence of the critical (solid line) and retrapping (dashed line) currents at $V_g=-3.4$~V. Inset: Current-voltage characteristic of the short section C-R at $V_g=-3.4$~V for different temperatures. The current is swept up from -120 to 120~nA.}
\label{fig:IVs_diff_temp}
\end{figure}
The $IV$ characteristics of our SGS junctions are strongly changing with temperature. Fig.~\ref{fig:IVs_diff_temp} shows the temperature dependence of the $IV$ curve of the short section (C-R) at $V_g=-3.4$~V. The critical current decreases with increasing the temperature. The retrapping current remains almost constant until 0.3 K and then goes down. The hysteresis of the $IV$ curve is reduced as the temperature goes up and disappears at $0.45$~K. Similar results are found at different gate voltages, and also in the long section L-C. The temperature dependence of the critical current is used in Sec.~\ref{sec:res} to extract the charge carrier mean free path. From the shape of this dependence we can already tell that we are in the long junction regime, where the length of the junction $L$ is longer than the superconducting coherence length $\xi$. Most importantly, the strong temperature dependence of the critical current allows us to use SGS junctions as electronic thermometers.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{./Ic_vs_Vlc_cdr.eps}
\caption{(color online). Critical (circle points) and retrapping (cross points) currents as a function of the voltage $V_{\rm L-C}$ across the section L-C. The triangle points corresponds to the deduced electronic temperature.}
\label{fig:Ic_vs_Vlc}
\end{figure}
Keeping the bath temperature at 80~mK, the electronic temperature can be changed by injecting a dissipative current in between leads L and C. As shown in Fig.~\ref{fig:Ic_vs_Vlc} for $V_g=-3.4$~V, the critical current decreases in the short section (C-R) as the voltage $V_{\rm L-C}$ across the long section L-C goes up. The retrapping current remains initially almost constant at low voltages and decreases at around $0.15$~mV. By assuming that the electronic distribution follows a Fermi distribution, we can directly relate the critical current to the electonic temperature. We find that the temperature amounts to around $0.52$~K at $V_{\rm L-C}=0.35$~mV. However, as we show in Sec.~\ref{sec:thry}, the electronic distribution function may differ from a Fermi distribution in which case the electronic temperature is not properly defined. Consequently, the electronic temperature directly deduced from the measurement of $I_c$ is only an effective temperature (see Eq.~\eqref{eq:Teff} below).
\section{Theoretical model}\label{sec:thry}
We compose a model where a graphene flake is divided into two parts by superconducting electrodes with energy gaps $\Delta$ and treat the system as effectively one-dimensional with the essential dimension aligned along the $x$-direction. This setup is illustrated in Fig.~\ref{fig:graph} where the two distinct regions are numbered as 1 and 2.
\subsection{Distribution function and supercurrent}\label{sec:distrib}
Physical observables, such as the supercurrent, can now be determined from the electron distribution function $f(\epsilon)$ which is a function of both energy $\epsilon$ and position $x$, although below the latter is not explicitly written down. The distribution function satisfies the time-independent diffusion equation (disregarding the proximity effect)
\begin{equation}\label{eq:diff}
-D\frac{\partial^2 f(\epsilon)}{\partial x^2}=I(f(\epsilon)).
\end{equation}
Here $D$ is the diffusion constant, related to the Fermi velocity $v_F=10^6$ m/s and the transport relaxation time $\tau$ for elastic scattering by
\begin{equation}\label{eq:D}
D=v_F^2\tau/2,
\end{equation}
and $I(f)$ is the collision integral for the inelastic processes in the flake (most importantly e-e and e-ph scattering).
\begin{figure}[htb]
\centering
\includegraphics[width=0.9\columnwidth]{./graphene.eps}
\caption{(color online). Sketch of the model with graphene flake divided into two regions: the heater (region 1) and thermometer (region 2). The relative sizes of the regions are determined from critical-current measurements as explained in the text.}
\label{fig:graph}
\end{figure}
At the graphene-superconductor interfaces we have two kinds of boundary conditions depending on whether we are at the end points $x=0$ and $x=L$ or at the boundary of regions 1 and 2, at $x=x_c$. At the end points \cite{PauliTero}
\begin{subequations}\label{eq:bcend}
\begin{equation}\label{eq:eq}
f(\epsilon)=f_{L/R}(\epsilon), \quad |\epsilon|>\Delta
\end{equation}
\begin{equation}\label{eq:balance}
f(\mu+\epsilon)=1-f(\mu-\epsilon),\quad |\epsilon|<\Delta
\end{equation}
\begin{equation}\label{eq:heatblock}
\partial_x f(\mu+\epsilon)=\partial_x f(\mu-\epsilon),\quad |\epsilon|<\Delta.
\end{equation}
\end{subequations}
Here, $f_{L/R}(\epsilon)$ is the equilibrium Fermi function at the corresponding potential, i.e.,\ $f_L(\epsilon)=f_0(\epsilon-eV)$ and $f_R(\epsilon)=f_0(\epsilon)$ as $f_0(\epsilon)\equiv \left[\exp{\left[(\epsilon-\mu)/k_BT_{\rm bath}\right]}+1\right]^{-1}$ at base temperature $T_{\rm bath}$ and chemical potential $\mu$ with respect to the Dirac point. Below the gap, these boundary conditions conserve the balance of positive and negative charge excitations \eqref{eq:balance} and ensure that there is no energy current entering the superconductor \eqref{eq:heatblock}, due to Andreev reflection. Above the gap, we simply have continuity of the particle distribution across the contact since the interface is assumed transparent \eqref{eq:eq}. At the dividing superconductor at $x=x_c$, we require
\begin{subequations}\label{eq:bcmid}
\begin{equation}\label{eq:bcmida}
\left.f(\epsilon)\right|_{x=x_{c-}}=\left.f(\epsilon)\right|_{x=x_{c+}}
\end{equation}
\begin{equation}
f(\epsilon)=f_0(\epsilon), \quad |\epsilon|>\Delta
\end{equation}
\begin{equation}
f(\mu+\epsilon)=1-f(\mu-\epsilon),\quad |\epsilon|<\Delta
\end{equation}
\begin{equation}\label{eq:bccont}
\begin{split}
\partial_x\left.f(\mu+\epsilon)\right|_{x=x_{c-}}-\partial_x\left.f(\mu-\epsilon)\right|_{x=x_{c-}}=&\\
\partial_x\left.f(\mu+\epsilon)\right|_{x=x_{c+}}-\partial_x\left.f(\mu-\epsilon)\right|_{x=x_{c+}},&\quad |\epsilon|<\Delta.
\end{split}
\end{equation}
\end{subequations}
In addition to the requirement of continuity across $x=x_c$, Eq.~\eqref{eq:bcmida}, we have the same condititions as above apart from Eq.~\eqref{eq:bccont} which requires that also the energy current is conserved across $x=x_c$. Physically, we assume that the superconductor on top of the graphene flake acts as an electrode for the high-energy electrons and does not noticeably affect the low-energy ones. We also assume above and in the following that $\Delta$ is constant: independent of position or heating in the system.
The degree of inelastic scattering determines the state of the system which, in the presence of superconductors, can become a quite complicated nonequilibrium state when e-e relaxation is incomplete. \cite{MAR-S} On the other hand, when e-e relaxation is complete, two separate alternatives are possible: The system can either be in the equilibrium state $f_0(\epsilon)$ with bath temperature $T_{\rm bath}$ or in a so-called quasiequilibrium state $f_0(\epsilon)$ with electron temperature $T_e>T_{\rm bath}$ depending on whether the region in question is heated in one way or another. In our configuration, regions 1 and 2 are separated at $x=x_c$ by a grounding superconductor effective at $\epsilon>\Delta$. Therefore, for the ideal case of complete e-e relaxation, region 2 remains in equilibrium at temperature $T_{\rm bath}$ and region 1 in quasiequilibrium with $T_e$ determined by the heating voltage $V$. However, when e-e relaxation is incomplete, the presence of the heat link between the regions makes it a priori possible that the system in region 2 is in any of these states: equilibrium, quasiequilibrium or nonequilibrium. Irrespective of the state of the system, we can illustrate the heat distribution in the system by defining a (local) effective electron temperature
\begin{equation}\label{eq:Teff}
k_BT_e=\frac{\sqrt{6}}{\pi}\sqrt{\int_{-\infty}^\infty\epsilon\left[f(\epsilon)-1+\theta(\epsilon)\right]d\epsilon}
\end{equation}
where $\theta(\epsilon)$ is the Heaviside step function. This equates the energy in the system to the thermal energy.
For the quantitative results, we solve the distribution function $f(\epsilon)$ from the diffusion equation Eq.~\eqref{eq:diff} and when $V=0$, we use the equilibrium Fermi function $f(\epsilon)=f_0(\epsilon)$ with a given bath temperature $T_{\rm bath}$. In both cases, the supercurrent through graphene is \cite{Belzig}
\begin{equation}\label{eq:IS}
I_S(\phi)=\frac{C}{eR_N}\int_0^\infty d\epsilon j_S(\epsilon,\phi)[1-2f(\epsilon)],
\end{equation}
where we average $f(\epsilon)$ over the $x$-coordinate. Here, $C\lesssim1$ is a prefactor of the order unity describing the imperfections in the measurement. The spectral supercurrent $j_S(\epsilon,\phi)$ of a diffusive SNS system is defined in Ref.~\onlinecite{Belzig} and we determine it numerically at the phase $\phi=0.6\pi$, which gives a fair approximation for the low-temperature critical current $I_c\equiv \max I_S(\phi)\approx I_S(0.6\pi)$.
We note that for graphene the diffusion constant $D$ can depend strongly on the type of disorder and on doping. For screened Coulomb impurities,\cite{adam,Janne2} $D\propto\tau\propto |\mu|$. Below, $D$ is essentially fit to the experiments, so that we do not need to specify the nature of the scatterers. The use of a semiclassical diffusion approach has some limitations for the description of graphene, however. The experimentally found finite minimal conductivity at the CNP cannot be explained without quantum-mechanical effects and a more detailed description of the impurities and the nonuniform doping effects they may introduce (charge puddles). These are not an issue if the graphene is strongly {\it n} or {\it p} doped. In our case, for $V_g>V_{\rm CNP}$ the doping is of {\it n} type everywhere, but as mentioned above, for $V_g<V_{\rm CNP}$ {\it p-n} junctions are expected to emerge close to the contacts.\cite{lee,muellert,Janne2} For these reasons, and because of the assumed one-dimensionality, our description should be viewed only as an effective model, in particular when $V_g<V_{\rm CNP}$.
\subsection{Inelastic interactions: electron-phonon}\label{sec:e-ph}
We start by estimating the strength of e-ph contribution in the inelastic collision integral. For this, we deem it sufficient to pay attention only to acoustic phonons for the range of temperatures and voltages relevant to our experiments ($T,eV/k_B\sim 1$~K $\cong 0.1$~meV). The collision integral for acoustic phonons in graphene has been derived in Ref.~\onlinecite{Janne} but here we only need the e-ph power discussed in Refs.~\onlinecite{Janne, giazotto06, kubakaddi09, bistritzer09}. We assume the distribution $f(\epsilon)$ is sufficiently well defined by the effective electron temperature $T_e$ of Eq.~\eqref{eq:Teff}. Then, in the limit $(c/v_F)|\mu|\gg k_BT_e$, where $c$ is the sound velocity, the e-ph power is
\begin{equation}\label{eq:Peph}
P_{\rm e-ph}=\Sigma(\mu) A(T_e^4-T_{\rm bath}^4),
\end{equation}
with a $\mu$-dependent interaction constant $\Sigma(\mu)$ and the area of the flake $A$. This power law is applicable up to very close to the Dirac point since the condition $\mu=\hbar v_Fk_F=\hbar v_F\sqrt{\pi\epsilon_r\epsilon_0\delta V_g/ed}\gg k_BT_ev_F/c$ translates to
\begin{equation}
\delta V_g\gg\left(\frac{k_BT_e}{\hbar c}\right)^2\frac{ed}{\pi\epsilon_r\epsilon_0}
\end{equation}
for the distance from the Dirac point in terms of gate voltage $\delta V_g\equiv |V_g-V_{\rm CNP}|$. For $T_e\sim 1$~K, thickness of the gate oxide $d=250$~nm, relative permittivity $\epsilon_r=4$ and $c=0.02v_F=2\times 10^4$~m/s, this becomes $\delta V_g\gg 15$~mV. In our experimental data the minimal $\delta V_g$ is about $1.5$ V, and so the condition is not violated. This estimate neglects the effect of charge puddles close to the CNP, but we expect Eq.~\eqref{eq:Peph} to hold whenever the use of our semiclassical approach is justified.
The total power injected into the system is $P_{\rm in}=V^2/R_N$ and if electron-phonon coupling is absent, all of this escapes into the leads. If, furthermore, e-e interactions are neglected, the electron distribution function is comprised of discrete steps due to multiple Andreev reflections \cite{MAR-S} so that highest local effective temperature of the system is roughly $k_BT_e^{\rm max}\approx(\Delta+eV)/4$. \cite{myfoot} As a result, energy escape takes place also at $eV\ll\Delta$. In the presence of e-e interactions, $T_e$ is reduced from this value so that for strong interactions, $k_BT_e^{\rm max}\approx eV/4$ as in a diffusive conductor without superconductivity (Notice that, at high voltages, a hot spot is formed in the middle of the normal-conducting region 1, so that $T_e^{\rm max}$ becomes high while the average $T_e$ remains much lower). For $T_e\gg T_{\rm bath}$
\begin{equation}
\frac{P_{\rm e-ph}}{P_{\rm in}}=\frac{\Sigma AT_e^4R_N}{V^2},
\end{equation}
and we see that the relative importance of e-ph interaction increases in two cases: first with increasing $eV$ and second with $eV$ decreasing below $\Delta$ when e-e interaction is absent. We evaluate the both possibilities to obtain a range for $eV$ where $P_{\rm e-ph}$ is significant.
The value for the e-ph interaction constant has not been measured but we obtain a theoretical upper estimate $\Sigma=k_B^4\mathcal{D}^2\pi^2|\mu|/(60\hbar^5\rho v_F^3c^3)<6.9\times 10^{-3}$~W/K$^4$m$^2$ for graphene mass density $\rho=0.76$~mg/m$^2$, deformation potential constant $\mathcal{D}=10$~eV (see, for example, Ref.~\onlinecite{borysenko}) and chemical potential $\mu=0.22$~eV, corresponding to $V_g=+30$~V. This is roughly an order of magnitude smaller than e-ph interaction in metals, \cite{giazotto06} when the reduced dimensionality of graphene is taken into account using thickness $\sim 1$~\AA. Setting $A\approx 2.8$~$\mu$m $\times 4.0$~$\mu$m, $\Delta/k_B\approx 1.12$~K, and $R_N\approx 1$~k$\Omega$, yields $\frac{P_{\rm e-ph}}{P_{\rm in}} < 0.013\times (k_BT_e/\Delta)^4/(eV/\Delta)^2$. We then have the result:
\begin{equation}
\frac{P_{\rm e-ph}}{P_{\rm in}}<5\times 10^{-5} \left(\frac{eV}{\Delta}\right)^2\left(1+\frac{\Delta}{eV}\right)^4,
\end{equation}
assuming $k_BT_e=(\Delta+eV)/4$. The ratio is above 1\% when $eV<0.083\Delta$ or $eV>12\Delta$, while our measurements are focused on the range $eV=0.5\ldots 5\Delta$. We emphasize that in this estimate we used the hot-spot temperature which is much larger than the average $T_e$ at high voltages. On the other hand, at low voltages, we assumed a total absence of e-e interaction resulting in a high $T_e$ of the order $\Delta$. Generally, $T_e$ can be expected to be even smaller than estimated above and we conclude that for our experiment e-ph coupling can be neglected.
\begin{comment}
At very small voltages, any finite $P_{\rm e-ph}$ becomes dominant as long as e-e interaction is incomplete so that $T_e-T_{\rm bath}$ remains finite.
In the general case we calculate $T_e$ as function of the applied bias voltage $V$ numerically, with the e-e interaction included as described below. Then we compare the magnitude of the expected the e-ph power to $P_{in}$ with
However, when any form of inelastic relaxation is present, $T_e$ is determined by interplay of the participating processes and it is generally impossible to estimate it without a full solution to the problem. We compare the magnitude of the e-ph effect to $P_{\rm in}$ using the values for $T_e$ we have determined numerically for different $V$ based on the form of e-e interaction in Sec.~\ref{sec:e-e}. For $T_e\gg T_{\rm bath}$, when e-ph interaction is strong,
\begin{equation}
\frac{P_{\rm e-ph}}{P_{\rm in}}=\frac{\Sigma AT_e^4R_N}{V^2}.
\end{equation}
At very small voltages, any finite $P_{\rm e-ph}$ becomes dominant as long as e-e interaction is incomplete so that $T_e-T_{\rm bath}$ remains finite.
The value for the e-ph interaction constant has not been measured but we obtain a theoretical upper estimate $\Sigma=k_B^4\mathcal{D}^2\pi^2|\mu|/(60\hbar^5\rho v_F^3c^3)<6.9\times 10^{-3}$~W/K$^4$m$^2$ for graphene mass density $\rho=0.76$~mg/$m^2$, deformation potential constant $\mathcal{D}=10$~eV (see, for example, Ref.~\onlinecite{borysenko}) and chemical potential $\mu=0.22$~eV, corresponding to $V_g=+30$~V. This is roughly an order of magnitude smaller than e-ph interaction in metals, \cite{giazotto06} when the reduced dimensionality of graphene is taken into account using thickness $\sim 1$~\AA. The numerically determined $k_BT_e$ is $\sim 0.2\Delta$ at $eV=0.5\Delta$ and, despite the strong spatial dependence, always below $k_BT_e^{\rm max}\sim 1.2\Delta$ at $eV=5\Delta$, almost independent of the gate voltage (Notice that, at high voltages, a hot spot is formed in the middle of the normal-conducting region 1, so that $T_e^{\rm max}$ becomes high while the average $T_e$ remains much lower). Setting $A\approx 2.8$~$\mu$m $\times 4.0$~$\mu$m, $\Delta/k_B\approx 1.12$~K, and $R_N\approx 1$~k$\Omega$, we have the upper estimates $\frac{P_{\rm e-ph}}{P_{\rm in}}<9\times 10^{-5}$ at $eV=0.5\Delta$ and $\frac{P_{\rm e-ph}}{P_{\rm in}}<1.1\times 10^{-3}$ at $eV=5\Delta$. We conclude that for our experiment e-ph coupling can be neglected.
\end{comment}
\subsection{Inelastic interactions: electron-electron}\label{sec:e-e}
Coulomb interactions in graphene have been discussed, for example, in Refs.~\onlinecite{mirlin, mueller, tse2}. However, most of the existing results are for clean, charge-neutral graphene, where the golden-rule collision integrals are furthermore plaqued by divergences.\cite{mueller} A full theory of e-e interactions for diffusive graphene that would be valid at both the Dirac point and at finite doping is currently lacking. In particular it is not known if a well-defined quasiequilibrium state ever exists in diffusive graphene biased far from equilibrium, although it is often a convenient assumption.\cite{Janne2} Since the interband relaxation due to e-e collisions is expected to be weak,\cite{mirlin,foster09,balev} electrons and holes (or electrons in the conduction and valence bands) may in any case have to be treated separately.\cite{balev}
As explained above, our semiclassical approach restricts our calculation in principle to the strongly-doped regime, where only
one charge carrier is dominant. In this case the system may be expected to behave somewhat similarly to other disordered
two-dimensional conductors. Thus, a reasonable starting point for an effective description is the Altshuler-Aronov theory\cite{AA}
for diffusive normal metals. The collision integral for a well-screened diffusive wire in two dimensions is \cite{AA, RamSmth}
\begin{equation}\label{eq:Iee}
\begin{split}
I_{e-e}(f(\epsilon))=
\kappa_{e-e}&\int_{-\infty}^\infty d(\hbar\omega)\;
\int_{-\infty}^\infty d\epsilon'(\hbar\omega)^{-1}\\
&\times\left[I^{\rm in}(\omega,\epsilon,\epsilon')-I^{\rm out}(\omega,\epsilon,\epsilon')\right]
\end{split}
\end{equation}
with
$I^{\rm in}(\omega,\epsilon,\epsilon')=[1-f(\epsilon)][1-f(\epsilon')]f(\epsilon-\hbar\omega)f(\epsilon'+\hbar\omega)$
and $I^{\rm out}(\omega,\epsilon,\epsilon')=f(\epsilon)f(\epsilon')[1-f(\epsilon-\hbar\omega)][1-f(\epsilon'+\hbar\omega)]$.
The prefactor is given by $\kappa_{e-e}=1/(8|\mu|\tau)$ and while the distribution functions are assumed position-dependent, we disregard any such dependence in $\kappa_{e-e}$ for simplicity.
\begin{comment}
We can also estimate the e-e relaxation rate, defined as a functional derivative
\begin{equation}\label{eq:funcd}
\frac{1}{\tau_{e-e}(\epsilon)}=\left.\frac{\delta I_{e-e}(f(\epsilon))}{\delta f(0)}\right|_{f(\epsilon)=f_0(\epsilon)}.
\end{equation}
\begin{equation}\label{eq:funcd}
\frac{1}{\tau_{e-e}(\epsilon)}=\left.\frac{\partial I_{e-e}(f(\epsilon))}{\partial h}\right|_{h=0}(\eta(\epsilon))^{-1},
\end{equation}
when $f(\epsilon)$ is expressed as a sum $f(\epsilon)=f_0(\epsilon)+h\eta(\epsilon)$.
\end{comment}
We can also estimate the e-e relaxation rate, which reads \cite{RamSmth2}
\begin{equation}\label{eq:tee}
\frac{1}{\tau_{e-e}}=\frac{k_BT_e}{2m_eD}\ln\frac{T_1}{T_e}
\end{equation}
in energy-averaged form valid at low-energies $\sim k_BT_e$. The result is obtained from the collision integral, where an energy cut-off is required to remove a logarithmic divergence at small $\omega$, characteristic to two-dimensional systems.\cite{abraham} The cut-off energy is $\hbar\omega_0=4k_BT_e^3/T_1^2$ and this gives $T_1=m_e^2D^3e^4N_F^2/\hbar^2k_B\varepsilon_0^2$, also for Eq.~\eqref{eq:tee}. Here, $m_e$ is the electron mass, $\varepsilon_0$ the vacuum permittivity and $N_F$ the density of states at the Fermi level. For graphene, the latter becomes \cite{Grev}
\begin{equation}\label{eq:NF}
N_F=\frac{2}{\pi}\frac{|\mu|}{(\hbar v_F)^2},
\end{equation}
including spin degeneracy.
\subsection{Final model in dimensionless form}
With the e-e interaction as the sole contributor to the inelastic relaxation, we now cast the diffusion equation Eq.~\eqref{eq:diff} in a dimensionless form directly applicable for numerical solution:
\begin{equation}\label{eq:diffnodim}
\partial_{\tilde x}^2f=-K_{e-e}\tilde I_{e-e},
\end{equation}
where $\tilde x=x/L$ with $L$ denoting the total length of the graphene flake. The dimensionless parameter $K_{e-e}$ describing the strength of e-e interaction in two dimensions is derived from $\kappa_{e-e}$ given above and expressed in terms of experimentally relevant parameters. When all energies are normalized by $\Delta$,
\begin{equation}\label{eq:Kee}
K_{e-e}=\frac{L^2}{D}\Delta\kappa_{e-e}=\frac 14\frac{R_N}{R_Q}\frac{\Delta}{E_{\rm Th}}\frac{w}{L},
\end{equation}
where we have used the formula $\sigma=e^2N_FD=L/R_Nw$ for conductivity and Eqs.~\eqref{eq:D} and \eqref{eq:NF} for $D$ and $N_F$.
Here, $R_N$ is the normal-state resistance of the graphene flake, $R_Q\equiv h/e^2$ the quantum of resistance, $E_{\rm Th}=\hbar D/L^2$ the Thouless energy, $w$ the width of the flake and $L$ its length. Note that $K_{e-e}$ now depends on the dimensions of the sample and we define it, together with the other extensive materials parameters, here for the whole flake. Since $E_{\rm Th}\propto D$ and $R_N\propto D^{-1}$, the parameter $K_{e-e}$ is predicted to scale as $K_{e-e}\propto D^{-2}\propto R_N^2$.
We note again that while $K_{e-e}$ does not depend explicitly on $\mu$, such dependence is in principle present through the diffusion constant $D(\mu)$ (and hence $R_N$ and $E_{\rm Th}$). Below, we use only the single parameter $K_{e-e}$ to characterize the e-e interaction and extract it from the experiments.
\section{Results}\label{sec:res}
\subsection{Electron-electron strength}
In order to access the e-e scattering strength $K_{e-e}$, we make two measurements. In the first one, the bath temperature is varied. In the second, an injection voltage at the left superconductor is used to heat up the graphene flake. The critical current and the retrapping current in region 2 are then measured as functions of temperature and voltage, respectively. In the first measurement, the system is in thermal equilibrium, whereas in the second one, it can be in a nonequilibrium state. Therefore, we use the first measurement to determine $E_{\rm Th}$ in regions 1 and 2, and the second measurement to compare the experiments with our theoretical model.
\begin{table}
\caption{Graphene properties determined for different gate voltages. The results for $E_{\rm Th}$ and $C$ (Eq.~\eqref{eq:IS}) are based on critical-current measurements for which we have data for all gate voltages apart from $V_g=-30$~V and $V_g=-3.4$~V in region 1.}
\label{table:ETh}
\begin{center}
\begin{tabular}{ | c | c || c | c | c || c | c | c |}
\hline
& & \multicolumn{3}{|l||}{Region 1} & \multicolumn{3}{| l|}{Region 2} \\
$V_g$ & $\mu$(meV) & $R_N$ & $E_{\rm Th}/\Delta$ & $C$ & $R_N$ & $E_{\rm Th}/\Delta$ & $C$ \\ \hline
$-30~V$ & 150 & $760~\Omega$ & N/A & N/A & $708~\Omega$ & $1/3.5$ & 0.33 \\
$-9.5~V$ & 42 & $1100~\Omega$ & $1/6.4$ & 0.11 & $547~\Omega$ & $1/2.35$ & 0.42 \\
$-3.4~V$ & 96 & $650~\Omega$ & N/A & N/A & $467~\Omega$ & $1/2.5$ & 0.39 \\
$+30~V$ & 220 & $330~\Omega$ & $1/7.7$ & 0.18 & $337~\Omega$ & $1/2.7$ & 0.36 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{./EThfit_long.eps}
\caption{(color online). Critical current in region 1 as a function of bath temperature. Solid lines are the fits to the theory, using $E_{\rm Th}$ and $C$ as the fitting parameters.}
\label{fig:EThfit_long}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{./EThfit_short.eps}
\caption{(color online). Critical current in region 2 as a function of bath temperature. Solid lines are the fits to the theory, using $E_{\rm Th}$ and $C$ as the fitting parameters.}
\label{fig:EThfit_short}
\end{figure}
In thermal equilibrium, we calculate the theoretical value for $I_c$ simply by using an equilibrium function $f_0$ at temperature $T_{\rm bath}$ in Eq~\eqref{eq:IS}. Here and below, $\Delta=100\;\mu$eV $\cong 1.1$~K so that the critical temperature $T_c\sim$ 0.6~K. Since $j_S(\epsilon,\phi)$ in Eq.~\eqref{eq:IS} is dependent on $E_{\rm Th}$, we may fit the calculated value of $I_c$ to the measured one at different $V_g$ (Figs.~\ref{fig:EThfit_long} and~\ref{fig:EThfit_short}), to obtain the values in Table~\ref{table:ETh}. These values for $E_{\rm Th}$ are subsequently used to determine the effective lengths $L_1$ and $L_2$ of regions 1 and 2, respectively. We assume that the diffusion constant $D$ is constant throughout the graphene flake so that
\begin{equation}
\frac{L_1}{L_2}=\sqrt{\frac{E_{\rm Th}({\rm region\;2})}{E_{\rm Th}({\rm region\;1)}}}\approx \frac 53.
\end{equation}
This approximate value is used in the simulation and assuming $L=L_1+L_2$, we have $L_1=1.75$~$\mu$m and $L_2=1.05$~$\mu$m. For consistency, we may also check the elastic scattering length $l=2L^2E_{\rm Th}/\hbar v_F$ resulting from the values we obtained for $E_{\rm Th}$ and depending on $V_g$, we have $l=90\ldots 140$~nm. At $V_g=30$~V, the value $l=120$~nm is relatively close to the experimentally determined $l=70$~nm and the difference can be due to inaccuracy in determining the effective length.
\begin{table}
\caption{Parameter values determined for the graphene flake. Resistance is calculated as the sum of $R_N$ in region 1 and 2, whereas Thouless energy is obtained assuming that the diffusion constant determined from the $I_c$-measurements in region 2 is homogeneous throughout the whole flake. The theoretical value for $K_{e-e}$ is calculated from Eq.~\eqref{eq:Kee}.}
\label{table:Kee}
\begin{center}
\begin{tabular}{ | c || c | c | c | c | c |}
\hline
\multicolumn{6}{|l|}{The whole flake}\\
$V_g$ & $R_N$ & $E_{\rm Th}/\Delta$ & $D$ & $K_{e-e}^{\rm fit}$ & $K_{e-e}^{\rm theory}$\\ \hline
$-30~V$ & $1468~\Omega$ & $1/24.9$ & 0.046~m$^2$/s & 30.8 & 0.51 \\
$-9.5~V$ & $1647~\Omega$ & $1/16.6$ & 0.069~m$^2$/s & 51.9 & 0.38 \\
$-3.4~V$ & $1117~\Omega$ & $1/17.8$ & 0.065~m$^2$/s & 26.6 & 0.28 \\
$+30~V$ & $667~\Omega$ & $1/19.5$ & 0.060~m$^2$/s & 7.6 & 0.18 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{./IcVtogether.eps}
\caption{(color online). Critical current in region 2 as a function of heater voltage. Solid lines are the fits to the theory, using $K_{e-e}$ as the fitting parameter at a single point $eV<\Delta$ (The exact value depends on the gate voltage).}
\label{fig:IcV}
\end{figure}
In the second experiment we use the heater voltage $V$ at $T_{\rm bath}=80$~mK and, as a result, $I_c$ in region 2 is dependent on the strength of the e-e interaction $K_{e-e}$. We proceed by fitting the theoretical and experimental results at a single small voltage $eV<\Delta$ with $K_{e-e}$ as the fit parameter. We then use the fitted $K_{e-e}$ to compute the whole $I_c(V)$-curve. The results for different $V_g$ are given in Fig.~\ref{fig:IcV}. The corresponding $K_{e-e}$ are listed in Table~\ref{table:Kee}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{./Kee_vs_RN.eps}
\caption{(color online). Values for the e-e interaction strength $K_{e-e}$ determined from the experimental data and shown with respect to the normal state resistance of the graphene flake $R_N$.}
\label{fig:KvsR}
\end{figure}
From Eq.~\eqref{eq:Kee} we expect the e-e strength to scale as $K_{e-e}\sim R_N^2$. The values of $K_{e-e}$ extracted from the experiments are shown in Fig.~\ref{fig:KvsR} as a function of $R_N$ from Table~\ref{table:Kee}, assuming additionally that $R_N\rightarrow 0$, $K_{e-e}\rightarrow 0$. The data supports the quadratic behavior.
\subsection{State of the system: Quasi- or nonequilibrium?}
We determine the state of the system and assess how sensitive it is to any changes in $K_{e-e}$ by looking at the electron distribution functions $f(\epsilon)$. In our theoretical model based on metallic e-e interaction, gate voltage has no explicit effect on the results and any gate dependence is due only to such dependence in the materials parameters such as $R_N$. We therefore use the parameter values from the case where $V_g=-30$~V as a representative example. In addition, we set $eV=2.5\Delta$ so that any pecularities due to the nonequilibrium state should be visible both in region 1 and 2. A true nonequilibrium distribution with $K_{e-e}=0$ is given in Fig.~\ref{fig:f0} and we notice that, in the absence of relaxation, $f(\epsilon)$ remains constant in region 2. This happens because the electrostatic potential at the middle superconductor is fixed. The numerically determined $f(\epsilon)$ for $K_{e-e}=30.8$ (Table~\ref{table:Kee}) is shown in Fig.~\ref{fig:f1}. We see that while the electrons clearly have a nonequilibrium distribution near the superconductors, where $f(\epsilon)$ in any case is strongly affected by the boundary conditions of Eqs.~\eqref{eq:bcend} and~\eqref{eq:bcmid}, in the middle of both regions 1 and 2 $f(\epsilon)$ is smoothed very close to a thermal distribution.
The question now regarding the state of the system is: is $f(\epsilon)$ in region 2 effectively a thermal (quasi)equilibrium distribution? If so, the e-e interaction is weak enough to let some of the energy injected into region 1 leak into region 2, but still so strong that it forces the system in region 2 into a thermal state. To answer this, we calculate $I_c$ as a function of heater voltage for several values of $K_{e-e}$ in Fig.~\ref{fig:KtestIc}. First, the expected value of supercurrent is clearly dependent on the value of $K_{e-e}$ in this range so the measurement can be used to determine $K_{e-e}$ with satisfactory precision. Second, as there is no plateau of constant $I_c$ at small voltages, the system is only in complete equilibrium state at $V=0$. Third, at voltages $eV\gtrsim\Delta$ there is a clear difference between the critical current obtained using actual nonequilibrium distributions and their quasiequilibrium counterparts, i.e.,\ equilibrium functions $f_0(\epsilon)$ with $T_e$ determined from $f(\epsilon)$. This implies that not only electron heating but also the formation of nonequilibrium state affects the observable supercurrent at $eV\gtrsim\Delta$. For voltages smaller than this, a quasiequilibrium description for the system is apt.
\begin{figure}[t!]
\centering
\includegraphics[width=0.95\columnwidth]{./Vg300inV20Keex0oravankepponen.eps}
\caption{(color online). Electron distribution functions $f(\epsilon)$ for a representative case ($V_g=-30$~V, $eV=2.5\Delta$) at different positions of the graphene flake. Here, e-e interactions are neglected, i.e.,\ $K_{e-e}=0$. Dashed line is the quasiequilibrium function $f_0(\epsilon)$ at the effective temperature $T_e$ corresponding to the nonequilibrium function $f(\epsilon)$.}
\label{fig:f0}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=0.95\columnwidth]{./Vg300inV20porkkana.eps}
\caption{(color online). Electron distribution functions $f(\epsilon)$ for a representative case ($V_g=-30$~V, $eV=2.5\Delta$) at different positions of the graphene flake. Here, the strength of the e-e interaction corresponds to the value of Table~\ref{table:Kee} which we obtained by fitting to the experimental data: $K_{e-e}=30.8$. Dashed line is the quasiequilibrium function $f_0(\epsilon)$ at the effective temperature $T_e$ corresponding to the nonequilibrium function $f(\epsilon)$. The spectral supercurrent $j_S(\epsilon)$ (Eq.~\eqref{eq:IS}) is shown as a black dotted line with arbitrary units and centered so that $j_S=0$ when $|\epsilon|<E_{\rm Th}=\Delta/3.5$.}
\label{fig:f1}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{./Vg300_Ic_eri_Kee.eps}
\caption{(color online). Calculated critical current as a function of voltage for different values of the e-e strength $K_{e-e}$. The solid lines are the results from the nonequilibrium distributions $f(\epsilon)$, whereas the dashed lines demonstrate how $I_c$ differs if quasiequilibrium functions with effective electron temperatures are used instead. Blue crosses correspond to the measured values.}
\label{fig:KtestIc}
\end{figure}
\subsection{Relaxation time}
We finally estimate the e-e relaxation time in our sample. From Eq.~\eqref{eq:tee}, we find $\tau_{e-e}=2\ldots4$~ns at $T_e=T_{\rm bath}=80$~mK and $\tau_{e-e}=400\dots 700$~ps at $T_e=0.5$~K, depending on the gate voltage.
The relaxation rate is expected to be proportional to the electron-electron scatterin strength $K_{e-e}$. Our measurements therefore suggest e-e relaxation times which are smaller than the values above by a factor $K_{\rm e-e}^{\rm fit}/K_{\rm e-e}^{\rm theory}$. Then, at $T_e=T_{\rm bath}$, $\tau_{e-e}=30\ldots 70$~ps and $T_e=0.5$~K, $\tau_{e-e}=5\dots 13$~ps. We are not aware of any comparable results in the low-energy regime. The e-e scattering times have been estimated for graphene in Ref.~\onlinecite{tse2}, but there the graphene is ballistic, the energy scale much larger and the obtained scattering times, consequently, (even) much smaller.
\section{Discussion}\label{sec:dis}
We have measured the strength of e-e interaction in graphene at four different gate voltages. From Eq.~\eqref{eq:Kee} we find the expected values for the parameter $K_{e-e}$ describing the interaction strength as given in Table~\ref{table:Kee}. The measured values are roughly 40-140 times larger than expected from the Altshuler-Aronov theory with the largest differences closer to the Dirac point. Discrepancies between theory and experiments are reported also in metallic wires (Ag), \cite{Huard} but there the differences are sample-specific and only up to a factor of 20, with the measured value larger there as well. Even though the strength of the e-e interaction is stronger than expected, the system is still not thermalized and the incomplete e-e relaxation can be seen at heater voltages $V$ well below the superconducting gap by measuring the critical current. The result at low voltages is seen as an energy leak from the heater junction and increasing temperature in the thermometer region. For $eV>\Delta$, also the electrons in the thermometer are driven into a nonequilibrium state where a thermal description with an effective temperature $T_e$ is no longer enough. In addition to finding the magnitude of the e-e scattering strength, we find that the scattering strength exhibits a significant gate dependence, presumably due to changes in charge density as the gate voltage is varied.
Finally, we note that we have also estimated the interaction strength between electrons and acoustic phonos in our setup with an aim to measure it. However, since the e-e interaction is relatively strong, we expect that $T_e$ is at most of the order of the heater voltage except when $eV\ll\Delta$. Consequently, the expected electron-phonon power at $eV\lesssim\Delta$ becomes even lower than predicted by our estimate in Sec.~\ref{sec:e-ph}. At the other end of the scale, $eV>\Delta$, the effective temperature needs to be increased yet further for a high ratio $P_{\rm e-ph}/P_{\rm in}\sim T_e^4/V^2$. This results in notable heating also in the thermometer region, making a critical current measurement such as the one used here very difficult, unless a strong thermal isolation is established between the heater and the thermometer regions.
\begin{acknowledgments}
The authors thank Pauli Virtanen and Matti Laakso for helpful discussions. JV gives thanks to Finnish Foundation for Technology Promotion for their support. JKV and TTH acknowledge the support from the Academy of Finland and TTH further acknowledges the funding European Research Council (Grant No. 240362-Heattronics)
\end{acknowledgments}
|
\section{Introduction}
In 1930 Hecke (\cite{hecke1930}) proved an interesting result concerning the representation of $\PSL_2(\bbF_q)$ on the space of holomorphic differentials of $X(q)$. Analyzing the character table of this group we see two `special' irreducible representations (here denoted $\pi_+$ and $\pi_-$). If we call $m_+$ and $m_-$ the multiplicity of $\pi_+$ and $\pi_-$ in that representation, then Hecke showed that $m_+ - m_- = h(-q)$. In this paper we present a translation and explanation of the ideas contained in \cite{hecke1930}.
We attempt to use the same notation as Hecle whenever it is defined, otherwise we use standard notation. A notable exception is the choice of notation for the irreducible representations of $\PSL_2(\bbF_q)$.
We assume familiarity with representation theory, Riemann surfaces and some knowledge of modular forms at an introductory level (for instance, the first chapters of \cite{dia&shu2005}). When needed, we refer to results in other topics as well.
\section{Fixing notations}
In these notes, $q$ will denote a prime number (which we assume to satisfy $q \equiv 3 \pmod{4}$ and $q > 3$).
\[
\Gamma(q) = \left\{ \begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \SL_2(\bbZ) \ \bigg| \ \begin{pmatrix} a & b \\ c & d \end{pmatrix} \equiv \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix} \pmod q \right\}
\]
\[
\Gamma_1(q) = \left\{ \begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \SL_2(\bbZ) \ \bigg| \ \begin{pmatrix} a & b \\ c & d \end{pmatrix} \equiv \begin{pmatrix} 1 & * \\ 0 & 1 \end{pmatrix} \pmod q \right\}
\]
\[
M_2(\Gamma(q)) := \textrm{space of modular forms of weight $2$ with respect to } \Gamma(q)
\]
(as defined in \cite{dia&shu2005}; in particular, our modular forms are holomorphic on $\mathcal{H}$ and at the cusps)
\vspace{.5cm}
We have a well-known (right) action of $\SL_2(\bbZ)$ on $M_2(\Gamma(q))$: if $\gamma = \begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \SL_2(\bbZ)$ and $\varphi \in M_2(\Gamma(q))$, then
\[
(\varphi[\gamma]_2) (z) := (cz+d)^{-2} \varphi(\gamma z)
\]
where $\gamma z := \frac{az+b}{cz+d}$.
Since the elements of $M_2(\Gamma(q))$ are invariant under the action of $\Gamma(q)$ and $-I$, the original action induces an action of $\PSL_2(\bbF_q) = \dfrac{\SL_2(\bbZ)}{\{\pm I \} \cdot \Gamma(q)}$.
\vspace{.5cm}
Let $\zeta := \exp\left(\frac{2 \pi i}{q}\right)$ and $P = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} \in PSL_2(\bbF_q)$.
We define the $\bbC$-vector space $V := \left\{ f \in M_2(\Gamma(q)) \ | \ f[P]_2 = \zeta f \right\}$ and denote $z := \dim_{\bbC} V$.
We also define the following modular curves (which are Riemann surfaces): $X(q) = \Gamma(q) \backslash \mathcal{H}^*$ and $X_1(q) = \Gamma_1(q) \backslash \mathcal{H}^*$ (where $\mathcal{H}^*$ is the union of the upper half-plane and the cusps).
Finally, throughout this paper, if $\pi$ is a representation of a group $G$ on a vector space $V$, $\pi(g)$ denotes the element in $GL(V)$ or its trace depending on the context (sometimes we write $\tr(\pi(g))$ so that it is clear what we mean).
\section{First Remarks}
Note that the action of $\PSL_2(\bbF_q)$ on $M_2(\Gamma(q))$ induces a representation (since it is a right action, the representation is given by $\gamma \cdot \varphi = \varphi[\gamma^{-1}]_2$).
Recall that $M_2(\Gamma(q)) = E_2(\Gamma(q)) \oplus S_2(\Gamma(q))$, where $E_2(\Gamma(q))$ is the Eisenstein space and $S_2(\Gamma(q))$ is the space of cusp forms. This decomposition is also a decomposition of $M_2(\Gamma(q))$ as a representation of $\PSL_2(\bbF_q)$. Moreover, $S_2(\Gamma(q))$ is isomorphic to the space of holomorphic differentials of $X(q)$.
$\PSL_2(\bbF_q)$ can be naturally identified with a subgroup of $\Aut(X(q))$ and its action on $S_2(\Gamma(q))$ defined in the previous section is (via these identifications) the action of a subgroup of $\Aut(X(q))$ on the space of holomorphic differentials of $X(q)$.
If $\pi$ denotes the representation from last paragraph, then it is known that $\pi(\gamma) + \overline{\pi(\gamma)} = 2 - t$ where $t$ is the number of fixed points of $\gamma \in \PSL_2(\bbF_q) \subseteq \Aut(X(q))$ (this is called the Lefschetz Fixed Point Formula; cf. \cite{farkas&kra1992}). So, if $\pi(\gamma) \in \bbR$, we can obtain $\pi(\gamma)$ by computing $(2-t)/2$.
Looking at the character table of $\PSL_2(\bbF_q)$ we see that all the irreducible representations have real traces except in the case $q \equiv 3 \pmod{4}$ (where there are two irreducible representations which do not have real traces).
Thus, from now on we assume $q \equiv 3 \pmod{4}$ and $q > 3$ (the case $q = 3$ is simple and can be treated individually). Hecke actually studied all the cases in \cite{hecke1930} (for the interested reader: Hecke defines $\varepsilon = (-1)^{(q-1)/2}$ and deals with both cases simultaneously).
\section{Character Table of $PSL_2(\bbF_q)$}
In this section we recall the character table of $\PSL_2(\bbF_q)$ (following the presentation given in \cite{casselmana}).
Representations of the conjugacy classes are:
\[
\begin{array}{cccc}
\pm \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix} & & & \\
\pm \begin{pmatrix} t & 0 \\ 0 & 1/t \end{pmatrix} & \sim & \pm \begin{pmatrix} 1/t & 0 \\ 0 & t \end{pmatrix} & (t \neq \pm 1) \\
\pm \begin{pmatrix} a & -b \\ b & a \end{pmatrix} & \sim & \pm \begin{pmatrix} a & b \\ -b & a \end{pmatrix} & (a^2 + b^2 = 1) \\
\pm \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} & & & \\
\pm \begin{pmatrix} 1 & -1 \\ 0 & 1 \end{pmatrix} & & &
\end{array}
\]
If we let $E = \bbF_q(\sqrt{-1})$, then $N^1_{E / \bbF_q}$ embeds in $\SL_2(\bbF_q)$ via $a + b \sqrt{-1} \mapsto \begin{pmatrix} a & -b \\ b & a \end{pmatrix}$. Since $N^1_{E / \bbF_q} / \{ \pm 1 \}$ has a unique element of order $2$ (namely $\sqrt{-1}$), it also has a unique character of order $2$ (call it $\rho_0$).
The irreducible representations of $\PSL_2(\bbF_q)$ are the following:
\begin{itemize}
\item the trivial representation of dimension $1$
\item the Steinberg representation
\item representations $\pi_{\chi}$ parametrized by the characters $\chi$ of the group $\bbF_q^{\times} / \{ \pm 1 \}$
\item representations $\pi_{\rho}$ parametrized by the characters $\rho \neq \rho_0$ of the group $N^1_{E / \bbF_q} / \{ \pm 1 \}$
\item representations $\pi_+$ and $\pi_-$, corresponding to the character $\rho_0$
\end{itemize}
Finally, we give the character table:
\begin{table}[H]
$
\begin{array}{| c | c | c | c | c | c | c| }
\hline & id & St & \pi_{\chi} & \pi_{\rho} & \pi_+ & \pi_- \\ \hline
I = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix} & 1 & q & q+1 & q-1 & (q-1)/2 & (q-1)/2 \\ \hline
\begin{pmatrix} t & 0 \\ 0 & 1/t \end{pmatrix} & 1 & 1 & \chi(t) + \chi(1/t) & 0 & 0 & 0 \\ \hline
\begin{pmatrix} a & -b \\ b & a \end{pmatrix} & 1 & -1 & 0 & -\rho(\epsilon) - \rho(1 / \epsilon) & -\rho_0(\epsilon) & -\rho_0(\epsilon) \\ \hline
P = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} & 1 & 0 & 1 & -1 & \overline{\mathfrak{G}} & \mathfrak{G} \\ \hline
P^{-1} = \begin{pmatrix} 1 & -1 \\ 0 & 1 \end{pmatrix} & 1 & 0 & 1 & -1 & \mathfrak{G} & \overline{\mathfrak{G}} \\ \hline
\end{array}
$
\caption{Irreducible representations of $\PSL_2(\bbF_q)$}
\label{tbl:irreps}
\end{table}
\vspace{.5cm}
where $\varepsilon = a + b \sqrt{-1} \in E$,
\[
\mathfrak{G} = \mysum_{\left( \frac{x}{q} \right) = 1} \exp(2 \pi i x / q) = \mysum_{\left( \frac{x}{q} \right) = 1} \zeta^x
\]
and $\overline{\mathfrak{G}}$ is its complex conjugate.
\section{Computing $z$ using representation theory}
In this section we will compute $z$ viewing $M_2(\Gamma(q))$ as a representation of $\PSL_2(\bbF_q)$.
\subsection{General remarks on representations of $\PSL_2(\bbF_q)$}
For each representation $\pi$ and each $n \in \{ 0, \dotsc, q-1 \}$, let $p_{\pi}(n)$ denote the multiplicity of $\zeta^n$ viewed as an eigenvalue of $\pi(P^{-1})$. (Sometimes we shall simply write $p$ instead of $p_{\pi}$). With this notation, we have the following:
\[
f_\pi := \tr(\pi(I)) = p(0) + p(1) + \dotsb + p(q-1)
\]
\[
\tr(\pi(P)) = p(0) 1 + p(1) \zeta + \dotsb + p(q-1) \zeta^{q-1}
\]
(recall that $\pi(\gamma)$ can be diagonalized for any $\gamma \in \PSL_2(\bbF_q)$).
So, given $f_\pi$ and $\tr(\pi(P))$, we can determine $p(0), p(1), \dotsb, p(q-1)$.
We can therefore compute $p(n)$ for each of the irreducible representations (cf. table \ref{tbl:irreps}) obtaining the following:
\vspace*{.1cm}
\begin{center}
\begin{tabular}{r c l}
$id$ & & $p(0) = 1$ and $p(n) = 0$ for all $n>0$ \vspace*{.05cm} \\
$St$ & & $p(n) = 1$ for all $n$ \vspace*{.05cm} \\
$\pi_\chi$ & & $p(0) = 2, p(1) = \dotsb = p(q-1) = 1$ \vspace*{.1cm} \\
$\pi_\rho$ & & $p(0) = 0, p(1) = \dotsb = p(q-1) = 1$ \\
$\pi_+$ & & $p(n) = \left\{ \begin{array}{c c l} 1 & \textrm{, } & \left( \frac{n}{q} \right) = 1 \\ 0 & \textrm{, } & \textrm{ otherwise } \end{array} \right.$ \\
$\pi_-$ & & $p(n) = \left\{ \begin{array}{c c l} 1 & \textrm{, } & \left( \frac{n}{q} \right) = -1 \\ 0 & \textrm{, } & \textrm{ otherwise } \end{array} \right.$
\end{tabular}
\end{center}
As an example of these computations, let us determine $p_{\pi_+}(n)$ for all $n$. We have the equations
\[
\frac{q-1}{2} = p(0) + p(1) + \dotsb + p(q-1)
\]
\[
\begin{array}{c c l}
\mysum\limits_{\left( \frac{x}{q} \right) = 1} \zeta^x & = & p(0)1 + p(1) \zeta + \dotsb + p(q-1) \zeta^{q-1} \\
& = & [p(0) - p(q-1)]1 + [p(1) - p(q-1)] \zeta + \dotsb + [p(q-2) + p(q-1)]\zeta^{q-2}
\end{array}
\]
Hence, for $0 \leq n \leq q-2$
\[
\begin{array}{ccl}
p(n) - p(q-1) & = & \left\{ \begin{array}{ccl}
1 & , & \left( \frac{n}{q} \right) = 1 \\
0 & , & \textrm{ otherwise}
\end{array} \right.
\end{array}
\]
Now, since there are exactly $(q-1)/2$ squares in $\bbF_q^{\times}$ the first equation reads
\[
\frac{q-1}{2} = (q-1) p(q-1) + \frac{q-1}{2}
\]
and so
\[
p(q-1) = 0
\]
This yields
\[
\begin{array}{ccl}
p(n) & = & \left\{ \begin{array}{ccl}
1 & , & \left( \frac{n}{q} \right) = 1 \\
0 & , & \textrm{ otherwise}
\end{array} \right.
\end{array}
\]
So, if we have a representation $W$ of $\PSL_2(\bbF_q)$ which decomposes as
\[
W = \alpha \cdot St \ \oplus \ \beta_+ \cdot \pi_+ \ \oplus \ \beta_- \cdot \pi_- \ \oplus \ \mysum_{\chi} \gamma_{\chi} \cdot \pi_{\chi} \ \oplus \ \mysum_{\rho} \delta_{\rho} \cdot \pi_{\rho}
\]
then
\[
\dim_{\bbC} \left\{ w \in W \ | \ P^{-1}w = \zeta w \right\} = \alpha + \beta_+ + \mysum \gamma_{\chi} + \mysum \delta_{\rho}
\]
\subsection{Applying the general remarks to our case}
\label{ssec:applying_gen_rmks}
Now we apply the previous subsection to the representation on $M_2(\Gamma(q))$ (recall that if $\gamma \in \PSL_2(\bbF_q)$ and $\varphi \in M_2(\Gamma(q))$, then $\gamma \varphi = \varphi[\gamma^{-1}]_2$).
As we saw earlier, $M_2(\Gamma(q)) = E_2(\Gamma(q)) \ \oplus \ S_2(\Gamma(q))$. So, we may study $E_2(\Gamma(q))$ and $S_2(\Gamma(q))$ separately.
It is known that, as a representation,
\[
E_2(\Gamma(q)) = St \ \oplus \ 2 \mysum_{\chi} \pi_{\chi}
\]
Let
\begin{equation}
\label{eqn:decomp_of_s2}
S_2(\Gamma(p)) = x \cdot St \ \oplus \ y_+ \cdot \pi_+ \ \oplus \ y_- \cdot \pi_- \ \oplus \ \mysum_{\chi} u_{\chi} \cdot \pi_{\chi} \ \oplus \ \mysum_{\rho} v_{\rho} \cdot \pi_{\rho}
\end{equation}
be the decomposition of the space of cusp forms (notice $id$ does not occur because of the well known fact that $\dim M_2(\SL_2(\bbZ)) = 0$).
Then, since there are $\frac{q + 1}{4} - 1$ irreducible representations of the form $\pi_{\chi}$, the previous subsection gives us
\[
z = x + 1 + y_+ + \mysum_{\chi} u_{\chi} + 2 \left( \frac{q + 1}{4} - 1 \right) + \mysum_{\rho}v_{\rho}
\]
Our goal for the rest of this section is to simplify this expression. For this, we will define some notations and use them to help us (these are actually introduced in \cite{hecke1928}).
\vspace{.5cm}
\begin{df} Given the decomposition (\ref{eqn:decomp_of_s2}), we define
\[
\begin{array}{ccc}
U = \mysum\limits_{\chi} u_{\chi} & , & V = \mysum\limits_{\rho} v_{\rho} \\
Y = y_- + y_+ & , & S = Y + 2U + 2V
\end{array}
\]
\end{df}
\vspace{.5cm}
Using this notation, we obtain
\begin{equation}
\label{eqn:formula_z}
z = \frac{y_+ - y_-}{2} + \frac{q-1}{2} + \frac{Y}{2} + U + V + x
\end{equation}
\vspace{.5cm}
\begin{df} If $H \leq \PSL_2(\bbF_q)$, we define
\[
Z(H) := \dim_{\bbC} \{ f \in S_2(\Gamma(q)) \ | \ f[\gamma]_2 = f, \forall \gamma \in H \}
\]
\end{df}
\vspace{.5cm}
Notice that $Z(H)$ is just the multiplicity of the identity representation of $H$ in $S_2(\Gamma(q))$. Thus, by representation theory,
\begin{equation}
\label{eqn:formula_Z(H)}
Z(H) = \frac{1}{|H|} \mysum_{h \in H} \left( x St(h) + y_+ \pi_+(h) + y_- \pi_-(h) + \mysum_{\chi} u_{\chi} \pi_{\chi}(h) + \mysum_{\rho} v_{\rho} \pi_{\rho}(h) \right)
\end{equation}
Notice $H_1 := \left\{ \begin{pmatrix} a & -b \\ b & a \end{pmatrix} \in \PSL_2(\bbF_q) \ \bigg| \ a, b \in \bbF_q \right\}$ is a subgroup of order $(q+1)/2$. Also, $H_2 := \left\{ \begin{pmatrix} t & 0 \\ 0 & 1/t \end{pmatrix} \in \PSL_2(\bbF_q) \ \bigg| \ t \in \bbF_q^\times \right\}$ is a subgroup of order $(q-1)/2$ (because $\bbF_q^\times$ is cyclic of order $q-1$ and $I = -I$ in $\PSL_2(\bbF_q)$). This motivates our next definition:
\begin{df} $Z(\frac{q+1}{2}) := Z(H_1)$ \ and \ $Z(\frac{q-1}{2}) := Z(H_2)$.
\end{df}
Using equation (\ref{eqn:formula_Z(H)}), we obtain the following:
\begin{equation}
\label{eqn:specific_formula_Z(H)}
\begin{array}{c}
Z(\frac{q-1}{2}) = 3x + Y + 2U + 2V \\ \\
Z(\frac{q+1}{2}) = x + Y + 2U + 2V
\end{array}
\end{equation}
Hence,
\begin{equation}
\label{eqn:specific_formula_x_S}
\begin{array}{c}
x = \frac{1}{2} Z(\frac{q-1}{2}) - \frac{1}{2} Z(\frac{q+1}{2}) \\ \\
S = \frac{3}{2} Z(\frac{q+1}{2}) - \frac{1}{2} Z(\frac{q-1}{2})
\end{array}
\end{equation}
So,
\begin{equation}
\frac{Y}{2} + U + V + x = \frac{S}{2} + x = \frac{1}{4} \left( Z\left(\frac{q+1}{2}\right) + Z\left(\frac{q-1}{2}\right) \right)
\end{equation}
Note $Z(H)$ is $\dim S_2(\Gamma)$ for a certain congruence subgroup $\Gamma$ and, so, is equal to the genus of the corresponding modular curve. Hence one can compute (Hecke computes this in \cite{hecke1928})
\begin{equation}
Z\left(\frac{q+1}{2}\right) + Z\left(\frac{q-1}{2}\right) = \frac{q^2-1}{6} - (q-1)
\end{equation}
So,
\begin{equation}
\label{eqn:z_repn}
z = \frac{y_+ - y_-}{2} + \frac{q-1}{4} + \frac{q^2-1}{24}
\end{equation}
\section{Computing $z$ using Riemann-Roch}
\subsection{Some general definitions and results}
Before we actually compute $z$ we will introduce some definitions and present some general facts about automorphic factors and its relation with modular forms. For the most part we will follow \cite{rankin77}.
Throughout this subsection, $\Gamma \subseteq \SL_2(\bbZ)$ denotes a congruence subgroup, $\overline{\Gamma}$ its image in $\PSL_2(\bbZ)$ and if $T = \begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \Gamma$ and $z \in \mathcal{H}$, then $T:z = cz + d$. Moreover, $P = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} \in \SL_2(\bbZ)$.
\begin{df}
A map $\nu : \Gamma \times \mathcal{H} \rightarrow \bbC$ is called an \emph{automorphic factor of weight $k$} if
\begin{enumerate}
\item $z \mapsto \nu(T,z)$ is holomorphic for all $T \in \Gamma$
\item $|\nu(T,z)| = |T:z|^k$ for all $T \in \Gamma$ and $z \in \mathcal{H}$
\item $\nu(ST, z) = \nu(S,Tz) \nu(T,z)$ for all $S,T \in \Gamma$ and $z \in \mathcal{H}$
\item if $-I \in \Gamma$, then $\nu(-T,z) = \nu(T,z)$ for all $T \in \Gamma$ and $z \in \mathcal{H}$
\end{enumerate}
\end{df}
Let us denote $\mu(T,z) := (T:z)^k$. One can prove that $\mu(T,z) = v(T) \mu(T,z)$, where $|v(T)| = 1$. The map $v$ is called the \emph{multiplier system} associated with $\nu$. Note that it determines $\nu$.
\begin{df}
If $L \in \SL_2(\bbZ)$, then we define the \emph{parameter of the cusp} $L \infty$ (with respect to $\Gamma$ and $\nu$) to be the only $\kappa_L \in [0,1)$ such that $v(L P^{n_L} L) = \exp(2 \pi i \kappa_L)$ where $n_L$ is the width of the cusp $L \infty$ with respect to $\Gamma$. (One proves that this depends only on the orbit of $L \infty$ via the action of $\Gamma$).
\end{df}
\begin{df}
An \emph{unrestricted modular form} of weight $k$ for the group $\Gamma$ with respect to the automorphic factor $\nu$ (or multiplier system $v$) is a holomorphic function $\varphi: \mathcal{H} \rightarrow \bbC$ such that $f(Tz) = \nu(T,z) f(z)$ for all $T \in \Gamma$ and $z \in \mathcal{H}$. The space of all such functions is denoted $M'_k(\Gamma, v)$. (Note: the definition found in \cite{rankin77} is a little more general; moreover, the notation used is slightly different).
\end{df}
\begin{fact}
Let $\varphi \in M'_k(\Gamma,v)$, $L \in \SL_2(\bbZ)$ and $n_L$ be the width of the cusp $L \infty$ with respect to $\Gamma$. Define $\varphi[L]_k (z) := \mu(L,z)^{-1} \varphi(Lz)$ and $\varphi[L]_k^*(z) := \exp(-2 \pi i \kappa_L z / n_L) \varphi[L]_k (z)$. Then
\begin{enumerate}
\item $\varphi[L]_k (z) = (z + n_L) = \exp(2 \pi i \kappa_L) \varphi[L]_k(z)$
\item $\varphi[L]_k^*(z + n_L) = \varphi[L]_k^*(z)$
\end{enumerate}
\end{fact}
\begin{df}
A function $\varphi \in M'_k(\Gamma, v)$ is called a \emph{modular form} of weight $k$ for the group $\Gamma$ with respect to the automorphic factor $\nu$ (or multiplier system $v$) if $\varphi$ is holomorphic at the cusps.
\end{df}
\begin{rmk}
A function $\varphi \in M'_k(\Gamma, v)$ is said to be \emph{holormophic at the cusps} if for all $L \in \SL_2(\bbZ)$, $\varphi[L]_k^*(z) = \mysum\limits_{m = N_L}^{\infty} a_m t^m$ for some $N_L \in \bbZ_{\geq 0}$ where $t = \exp(2 \pi i z / n_L)$.
\end{rmk}
\begin{df}
If $\varphi \in M_k(\Gamma, v) \backslash \{ 0 \}$ and $z \in \mathcal{H}^*$, then we define the \emph{order of $\varphi$ at $z$} by
\[
\ord(\varphi, z, \Gamma) = \left\{
\begin{array}{lcl}
\frac{\ord(\varphi, z)}{\overline{\Gamma}_{z}} & , & \textrm{ if } z \in \mathcal{H} \\
\kappa_L + N_L & , & \textrm{ if } z = L \infty
\end{array}
\right.
\]
(where $\ord(\varphi, z)$ is just the order of $\varphi$ at $z$ as a holomorphic function on $\mathcal{H}$, $\overline{\Gamma}_{z}$ is the stabilizer of $z$ in $\overline{\Gamma}$ and $N_L$ is as in the previous remark such that $a_{N_L} \neq 0$).
\end{df}
\begin{fact} Let $\varphi \in M_k(\Gamma, v)$. Then
\[
\varphi[L]_k(z) = \exp(2 \pi i \kappa_L z / n_L) \mysum\limits_{m = N_L}^{\infty} a_m(L) \exp(2 \pi i z / n_L)
\]
(This equality justifies the definition of the order of $\varphi$ at a cusp)
\end{fact}
\begin{df}
Given $\varphi \in M_k(\Gamma, v) \backslash \{ 0 \}$, we define
\[
\ord(\varphi, \Gamma) = \mysum_{z \in \Gamma \backslash \mathcal{H}^*} \ord(\varphi, z, \Gamma)
\]
\end{df}
\begin{thm}(\textit{theorem 4.1.4 in \cite{rankin77}})
\label{thm:valence_formula}
If $\varphi \in M_k(\Gamma, v) \backslash \{ 0 \}$ then $\ord(\varphi, \Gamma) = \dfrac{\mu k}{12}$, where $\mu = \left[ \PSL_2(\bbZ) : \overline{\Gamma} \right]$.
\end{thm}
This theorem is sometimes called `Valence Formula'.
\subsection{Computing $z$}
\label{ssec:computing_z}
We are now going to use Riemann-Roch to compute $z$. (throughout this section, we view $P = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix}$ in $\SL_2(\bbZ)$ or in $\PSL_2(\bbF_q)$ interchangeably depending on the context).
First, using that $\Gamma_1(q) = \left< \Gamma(q), P \right>$ we notice that $V = M_2(\Gamma_1(q), v)$ where $v$ is the multiplier system associated with the automorphic factor $\nu$ given by
\[
\begin{array}{ccccc}
\nu & : & \Gamma_1(q) \times \mathcal{H} & \longrightarrow & \bbC \\
& & (\gamma,z) & \longmapsto & (cz+d)^2 \zeta^b
\end{array}
\]
where $\gamma = \begin{pmatrix} a & b \\ c & d \end{pmatrix}$. So $v : \Gamma_1(q) \rightarrow \bbC$ is given by $v(\gamma) = \zeta^b$.
Now fix $\varphi \in V \backslash \{ 0 \}$ and define $\widetilde{V} := \left\{ \frac{\varphi_1}{\varphi} \ \big| \ \varphi_1 \in V \right\} \subseteq \bbC(X_1(q)) = $ set of rational functions on $X_1(q)$. Note that $\widetilde{V} \cong V$ (as $\bbC$-vector spaces) and, hence, $z = \dim_{\bbC} \widetilde{V}$.
The idea is to write $\widetilde{V} = L(D) = \{ f \in \bbC(X_1(q)) \mid div(f) \geq -D \}$ for a suitable divisor $D$. We then need to find the degree of $D$ in order to apply Riemann-Roch.
$D$ is basically the ``divisor'' of $\varphi$ but first we have to understand what we mean by ``divisor'' of $\varphi$. Notice the order of $\varphi$ as defined in the previous section might not be an integer for elliptic points and cusps. So we have to deal with those points.
First note that since $q \geq 3$ then $\Gamma_1(q)$ has no elliptic elements. Hence we only need to deal with cusps. We need to find all cusps with non-zero parameter ($\kappa_L \neq 0$).
Let $\dfrac{r}{s}$ be a cusp of $\Gamma_1(q)$ (and assume $\gcd(r,s) = 1$) and $L = \begin{pmatrix} r & x \\ s & y \end{pmatrix} \in \SL_2(\bbZ)$ and assume $n = n_L$ is the width of the cusp $r/s = L \infty$. Then
\begin{equation}
\label{eqn:stab_cusp}
L P^n L^{-1} = \begin{pmatrix} 1 - nrs & nr^2 \\ -ns^2 & 1 + nrs \end{pmatrix}
\end{equation}
Note that if $n \equiv 0 \pmod{q}$, then $\kappa_L = 0$. If $n \not\equiv 0 \pmod{q}$, then $s \equiv 0 \pmod{q}$ (otherwise $n \neq n_L$). So the cusps $r/s$ that are problematic are the ones with $s \equiv 0 \pmod{q}$. One checks that these are represented by
\begin{equation}
\label{eqn:bad_cusps}
\dfrac{r}{q} \ , \ r = 1, 2, \dotsc, \frac{q-1}{2}
\end{equation}
(and these are not $\Gamma_1(q)$-equivalent).
Hence, by (\ref{eqn:stab_cusp}), we see that $n_L = 1$ for all the cusps in (\ref{eqn:bad_cusps}). Moreover, for those cusps, $v(L P^{n_L} L^{-1}) = \exp(2 \pi i r^2/q) = \zeta^{r^2}$ and, thus, $\kappa_L = \left\{ \frac{r^2}{q} \right\} = $ the fractional part of $r^2/q$.
The conclusion is that $\widetilde{V} = L(D)$ where $D$ is a divisor of degree $m = \mu / 6 - \mysum\limits_{r=1}^{(q-1)/2} \left\{ \frac{r^2}{q} \right\}$ where $\mu = \left[ \PSL_2(\bbZ) : \overline{\Gamma_1(q)} \right] = (q^2 - 1)/2$ (by theorem \ref{thm:valence_formula}).
\begin{claim}
Let $g$ be the genus of $X_1(q)$ (which is known to be $g = (q-5)(q-7)/24$), then $m \geq 2g - 2$.
\end{claim}
\begin{proof}
$m - 2g + 2 = (q^2-1)/12 - \mysum\limits_{r=1}^{(q-1)/2} \left\{ \frac{r^2}{q} \right\} - (q-5)(q-7)/12 + 2 = (12q-36)/12 - \mysum\limits_{r=1}^{(q-1)/2} \left\{ \frac{r^2}{q} \right\} + 2 = q - 1 - \mysum\limits_{r=1}^{(q-1)/2} \left\{ \frac{r^2}{q} \right\} > 0$.
\end{proof}
So Riemann-Roch gives us $z = \dim_{\bbC} \widetilde{V} = m - g + 1$.
Denote $\chi(n) = \left( \frac{n}{q} \right)$ the Legendre symbol. Since $\left\{ \frac{r^2}{q} \right\} = \frac{r^2 \mod{q}}{q}$ we obtain that
\[
\mysum\limits_{r=1}^{(q-1)/2} \left\{ \frac{r^2}{q} \right\} = \frac{1}{2} \mysum\limits_{n=1}^{q-1} \frac{n}{q}(1 + \chi(n)) = \frac{q-1}{4} + \frac{1}{2q} \mysum_{n=1}^{q-1} n \chi(n)
\]
Hence, we can compute
\[
z = m - g + 1 = \dotsb = \frac{q^2 + 6q - 7}{24} - \frac{1}{2q} \mysum_{n=1}^{q-1} n \chi(n)
\]
Now, using Dirichlet's class number formula, we obtain
\begin{equation}
\label{eqn:z_RR}
z = \frac{q^2 + 6q - 7}{24} - \frac{1}{2}h(-q)
\end{equation}
where $h(-q) = $ class number of $\bbQ(\sqrt{-q}) / \bbQ$.
\section{Conclusion and final remark}
Using equation (\ref{eqn:z_repn}) from section \ref{ssec:applying_gen_rmks} and equation (\ref{eqn:z_RR}) from section \ref{ssec:computing_z} we finally obtain what we wanted:
\[
m_+ - m_- = h(-q)
\]
As a final remark, we note that a different proof of this result of Hecke can be found in \cite{casselmana}.
|
\section{Introduction}
The market risk premium is one of the main factors that drives the return of
any given portfolio of assets. Hence it is a key quantity for hedge funds, pension
funds, and numerous other investors. The risk premium can make investments
grow smoothly or jump up and down widely, often in an unpredictable manner.
In spite of its importance in asset allocation, however, the risk premium is
notoriously difficult to estimate from observed price processes of various risky
assets (see, e.g., Rogers 2001). Is it possible then to estimate the risk premium
from current prices of financial derivatives?
If $\{S_t\}$ denotes the price process of a risky asset and $h(s)$ is the payout
function of a European contingent claim expiring at $T$, then the price of this
derivative is given by the expectation of the cash flow $h(S_T)$, suitably
discounted, in the risk-neutral measure. Because asset price processes in
the risk-neutral measure are independent of the market risk premium, one
might be tempted to conclude therefore that derivative prices are likewise
independent of the risk premium. Indeed, in the case of the
Black-Scholes-Merton model where all relevant parameters are constant in
time, the risk premium parameter essentially drops out of various derivative
pricing formulae. Notwithstanding this example, it is worth bearing in mind that
the choice of the pricing measure does depend on the choice of the risk
premium. Thus, derivative prices in general will depend implicitly on the risk
premium, often in a nonlinear way. It follows that calibration of the market
risk premium from option prices is feasible within a given modelling framework
(Brody \textit{et al}. 2011).
The main purpose of the present paper is to address the question whether it is
possible, at least in principle, to determine the risk premium unambiguously, if
the totality of arbitrage-free market prices for various derivatives were available.
We shall find that the market risk premium consists of two components in an
additive manner (for models based on Brownian filtrations): The first of
the two, which we might call a `systematic' component, depends explicitly on
the term structure of the market, while the second, which we might call an
`idiosyncratic' component, is independent of the term structure of the market,
and thus can be identified as \textit{pure noise}. We show that the systematic
component can in principle be determined from current market data, whereas
the idiosyncratic noise component is strictly `hidden' and thus cannot be
inferred from derivative prices. Therefore, the risk premium can be backed
out from market data only up to an indeterminable additive noise.
Although the noise component cannot be inferred directly, it nevertheless has
an impact on the dynamics of asset prices under the physical measure, even
though it does not reflect the `true' state of affairs. Hence a spontaneous creation
of superfluous noise can move the price of an asset in an essentially arbitrary
direction. In particular, because the risk premium, and hence its noise component,
is a vectorial quantity, the direction of the noise vector can at times lie close to
the directions of volatility vectors of the share prices of a particular industrial sector,
leading to the creation of a `bubble' for that sector by pushing up those share
prices. When a more reliable information concerning the state of that sector is
unveiled, the direction of the risk premium vector is likely to change so as to
generate a negative component in the excess rate of return. This can be
exacerbated by an increase in the magnitude of asset volatilities due to
information revelation,
thus leading to a `burst'. Such a scenario need not be confined to a particular
financial sector; the existence of the so-called `equity premium puzzle' over a
specified period can likewise be attributed to the prevailing noise that points in
the general direction of the equity market volatility, but not in the direction of the
bond market volatility.
Needless to say, our formulation does not explain the cause of the creation
of anomalous price movements such as a financial bubble or an equity premium;
nor does it address the predictability of these events. In fact, according to our
characterisation, bubbles can at best be identified retrospectively, after their
bursts. Nevertheless, we are able to describe the mechanism by which such
anomalous price movements are generated in a simple and intuitive manner. In
particular, since our characterisation of a bubble is different from those more
commonly used in the literature, we are able to circumvent the analysis based on
subtle distinctions between local and true martingales. We also provide a heuristic
argument why the hidden noise might have the tendency of creating equity premium.
\section{Pricing kernel}
For definiteness, we shall be adopting the pricing kernel approach (see, e.g.,
Cochrane 2005, Bj\"ork 2009). We model the financial market on a probability
space $(\Omega,\mathcal{F}, \mathbb{P})$ with filtration
$\{{\mathcal F}_t\}_{t\geq0}$. Here
$\mathbb{P}$ denotes the `physical' probability measure, and
$\{{\mathcal F}_t\}$ is assumed to be generated by a multi-dimensional
Brownian motion. Expectation under
${\mathbb P}$ is denoted $\mathbb{E}[-]$, and for the conditional expectation
with respect to ${\mathcal F}_t$ we write ${\mathbb E}_t[-]$. Two other probability
measures enter the ensuring discussion; these are the risk-neutral measure
${\mathbb Q}$ and an auxiliary measure ${\mathbb R}$ to be described below.
Expectations in these measures will be written $\mathbb{E}^{\mathbb Q}[-]$ and
$\mathbb{E}^{\mathbb R}[-]$, respectively.
We assume that the market is free of arbitrage opportunities, and that there is
an established pricing kernel, but market completeness is not assumed. These
assumptions imply the existence of a unique preferred risk-neutral measure
${\mathbb Q}$. The pricing kernel, denoted here by $\{\pi_t\}_{t\geq0}$, is a
positive supermartingale with the property that if $S_T$ is the price at time $T$
of an asset that pays no dividend, then the price at time $t$ of the asset is given
by
\begin{eqnarray}
S_t = \frac{1}{\pi_t}\, {\mathbb E}_t [\pi_T S_T] .
\label{eq:1}
\end{eqnarray}
In particular, if $S_T=1$, then (\ref{eq:1}) gives the pricing formula for the discount
bond: $P_{tT} = {\mathbb E}_t [\pi_T]/\pi_t$.
We shall proceed by discussing some properties of the pricing kernel that are
relevant to our analysis here. In addition to being a positive supermartingale, the
pricing kernel fulfils the condition that ${\mathbb E}[\pi_t]\to0$ as $t\to\infty$. A
positive supermartingale possessing this property is known as a \textit{potential}.
It follows that every pricing kernel can be represented as a potential, and conversely
every potential constitutes an admissible pricing kernel. The Doob-Meyer
decomposition then shows that $\{\pi_t\}$ can be represented uniquely in the form:
\begin{eqnarray}
\pi_t = {\mathbb E}_t[A_\infty] - A_t,
\end{eqnarray}
where $\{A_t\}_{t\geq0}$ is an increasing adapted process such that $A_\infty$ is
finite. Note that ${\mathbb E}_t[A_\infty]$ is a uniformly integrable martingale. We
may define $\{A_t\}$ according to
\begin{eqnarray}
A_t = \int_0^t a_s \mbox{$\rm d$} s
\end{eqnarray}
for some adapted nonnegative process $\{a_t\}$. Hence it suffices to choose the
process $\{a_t\}$ to model the pricing kernel, and this leads to the potential
approach of Rogers (1997) to model term structure dynamics. A substitution shows
that
\begin{eqnarray}
\pi_t = \int_t^{\infty} {\mathbb E}_t [a_u] \mbox{$\rm d$} u.
\label{eq:4}
\end{eqnarray}
The representation (\ref{eq:4}) resembles that of Flesaker and Hughston (1996,1997),
if we make the following identification. First, writing $\rho_0(T) = -\partial_T P_{0T}$,
where $P_{0T}$ is the initial discount function, we see that the processes
$\{M_t(u)\}_{t\geq0,u\geq t}$ defined by
\begin{eqnarray}
M_{t}(u) = \frac{{\mathbb E}_t[a_u]}{\rho_0(u)}
\end{eqnarray}
is a one-parameter family of positive martingales, i.e. for each fixed $u\geq t$,
$\{M_t(u)\}$ is a martingale. This follows on account of the martingale property
of the conditional expectation ${\mathbb E}_t[a_u]$. In terms of these positive
martingales, the pricing kernel can be expressed in the Flesaker-Hughston
form:
\begin{eqnarray}
\pi_t = \int_t^{\infty} \rho_0(u) M_t(u) \mbox{$\rm d$} u.
\label{eq:6}
\end{eqnarray}
From the martingale representation theorem we deduce that the dynamical equations
satisfied by the positive martingale family $\{M_t(u)\}$ take the form:
\begin{eqnarray}
\mbox{$\rm d$} M_t(u) = M_t(u) v_t(u) \mbox{$\rm d$} \xi_t,
\label{eq:7}
\end{eqnarray}
where $\{v_t(u)\}$ is a family of adapted (in general vectorial) processes and
$\{\xi_t\}$ is a standard multi-dimensional
Brownian motion under the ${\mathbb P}$ measure. We observe therefore that
modelling the pricing kernel is equivalent to modelling the one-parameter family
of volatility processes $\{v_t(u)\}$. On account of (\ref{eq:6}) and (\ref{eq:7}) we
deduce, by an application of Ito's lemma, that
\begin{eqnarray}
\frac{\mbox{$\rm d$} \pi_t}{\pi_t} = -r_t \mbox{$\rm d$} t - \lambda_t \mbox{$\rm d$} \xi_t,
\label{eq:dpi}
\end{eqnarray}
where
\begin{eqnarray}
r_t = \frac{\rho_0(t) M_t(t)}{\int_t^{\infty} \rho_0(u) M_t(u) \mbox{$\rm d$} u}
\end{eqnarray}
is the short rate, and
\begin{eqnarray}
\lambda_t = -\frac{\int_t^\infty \rho_0(u) v_t(u) M_t(u)\mbox{$\rm d$} u}
{\int_t^{\infty} \rho_0(u) M_t(u) \mbox{$\rm d$} u}
\label{eq:10}
\end{eqnarray}
is the market risk premium. The fact that the drift
of $\{\pi_t\}$ can be identified with the short rate can be seen by applying the
martingale condition (\ref{eq:1}) on the money market account $\{B_t\}$
satisfying $\mbox{$\rm d$} B_t = r_t B_t \mbox{$\rm d$} t$. That is, the drift of $\{\pi_t B_t\}$ vanishes if
and only if the drift of $\{\pi_t\}$ is $\{-r_t\}$. Similarly, let us write $\{\mu_t\}$ for
the drift of a risky asset $\{S_t\}$ that pays no dividend, and $\{-\lambda_t\}$ for
the volatility of $\{\pi_t\}$. Then the martingale condition on $\{\pi_t S_t\}$ implies
that $\mu_t=r_t+\lambda_t\sigma_t$, which shows that $\{\lambda_t\}$ indeed
expresses the excess rate of return above the risk-free rate in unit of volatility.
An advantage of working with the pricing kernel is that once a model is chosen
for the volatility processes $\{v_t(u)\}$ of the martingale family, we are able not
only to price a wide range of derivatives via the pricing formula
${\mathbb E}[\pi_T H_T]$, where $H_T$ is the payout of a derivative, but also
to obtain a model for the interest rate term structure. Furthermore, a model for
$\{v_t(u)\}$, which can be calibrated by use of market data for derivative prices,
implies a process for the risk premium $\{\lambda_t\}$ according to the
prescription (\ref{eq:10}), and this in turn can be used for asset allocation purposes.
This is the sense in
which derivative prices can be used to calibrate the risk premium, within any
modelling framework (cf. Brody \textit{et al}. 2011). The issues that we would
like to address here are: (a) the ambiguity associated with the determination of
the risk premium from market data; and (b) the identification of the origin of this
ambiguity. For these purposes, it is useful to examine the probabilistic
characterisation of the pricing kernel, within the term structure density approach
of Brody and Hughston (2001).
\section{Probabilistic representation of the pricing kernel}
To proceed, we shall make the following observation that the positivity of
nominal interest and the requirement that a bond with infinite maturity must
have vanishing value imply that $\rho_0(T) = -\partial_T P_{0T}$ defines a
probability density function on the positive half-line (Brody and Hughston 2001).
More generally, the positivity of the martingale family $\{M_t(u)\}$ implies that
$\{\rho_t(u)\}$ defined by
\begin{eqnarray}
\rho_t(u) = \frac{\rho_0(u) M_t(u)}{\int_0^{\infty} \rho_0(u) M_t(u) \mbox{$\rm d$} u}
\label{eq:12}
\end{eqnarray}
is a measure-valued process, i.e. $\rho_t(u)\geq 0$ for all $t$ and all $u$; and
\begin{eqnarray}
\int_0^\infty \rho_t(u) \mbox{$\rm d$} u = 1
\end{eqnarray}
for all $t\geq0$.
The measure-valued process thus introduced suggests the existence of a random
variable $X$ whose conditional density under some probability measure is given by
(\ref{eq:12}). Furthermore, an application of Ito's lemma on (\ref{eq:12}) shows that
\begin{eqnarray}
\frac{\mbox{$\rm d$} \rho_t(u)}{\rho_t(u)} = \left( v_t(u) - {\hat v}_t\right)
\left( \mbox{$\rm d$}\xi_t - {\hat v}_t\mbox{$\rm d$} t\right),
\label{eq:14}
\end{eqnarray}
where
\begin{eqnarray}
{\hat v}_t = \int_0^\infty v_t(u) \rho_t(u) \mbox{$\rm d$} u
\end{eqnarray}
can be thought of as the conditional expectation of $v_t(X)$.
Indeed, the dynamical equation (\ref{eq:14}) takes the form of a Kushner equation,
thus implies the existence of the following auxiliary filtering problem. For simplicity
of exposition, let us for now assume that the one-parameter family of volatility
processes $\{v_t(u)\}$ is deterministic. We introduce a probability space
$(\Omega,{\mathcal F},{\mathbb R})$, upon which $X$ is a positive random variable
with density $\rho_0(u)$. The meaning of the measure ${\mathbb R}$ will be
examined at a later point. On this probability space, consider the following
information process
\begin{eqnarray}
\xi_t = \int_0^t v_s(X) \mbox{$\rm d$} s + \beta_t ,
\label{eq:16}
\end{eqnarray}
where $\{\beta_t\}$ is an ${\mathbb R}$-Brownian motion, independent of $X$. The
task of the `observer' is thus to determine the best estimate of $X$ given the data
$\{\xi_s\}_{0\leq s\leq t}$. Assuming that the criteria for optimality is an estimator that
minimises the quadratic error, standard results in filtering theory (cf. Wonham 1965,
Liptser and Shiryaev 2001) show that the best estimate for $X$ is the expectation
of $X$ with respect to the \textit{a posteriori} density:
\begin{eqnarray}
\frac{\mbox{$\rm d$}}{\mbox{$\rm d$} u} {\mathbb R}(X<u|{\mathcal F}_t) = \frac{\rho_0(u)
\exp\left(\int_0^t v_s(u){\rm d}\xi_s - \frac{1}{2}\int_0^t v_s^2(u){\rm d}s\right)}
{\int_0^{\infty} \rho_0(u) \exp\left(\int_0^t v_s(u){\rm d}\xi_s - \frac{1}{2}\int_0^t
v_s^2(u){\rm d}s\right) \mbox{$\rm d$} u},
\label{eq:17}
\end{eqnarray}
where ${\mathcal F}_t=\sigma(\{\xi_s\}_{0\leq s\leq t})$.
Notice that the right side of (\ref{eq:17}) is in fact identical to the right side of
(\ref{eq:12}), provided that the measure change between ${\mathbb P}$ and
${\mathbb R}$ are suitably defined (recall that $\{\xi_t\}$ in the ${\mathbb P}$
measure is a Brownian motion, while in the ${\mathbb R}$ measure it is a
drifted Brownian motion of (\ref{eq:16})).
The above-specified filtering problem leads to the following probabilistic interpretation
for the pricing kernel. Writing
\begin{eqnarray}
N_t = \int_0^{\infty} \rho_0(u) \exp\left(\int_0^t v_s(u){\rm d}\xi_s - \mbox{$\textstyle \frac{1}{2}$}
\int_0^t v_s^2(u){\rm d}s\right) \mbox{$\rm d$} u
\label{eq:18}
\end{eqnarray}
for the normalisation of the conditional density $\{\rho_t(u)\}$, we see that the
pricing kernel is given by the `unnormalised' conditional probability that $X>t$:
\begin{eqnarray}
\pi_t = N_t \, {\mathbb R}_t(X>t),
\label{eq:19}
\end{eqnarray}
where for simplicity we have written ${\mathbb R}_t(-)={\mathbb R}(-|{\mathcal F}_t)$
for the conditional probability. Further, the price of a discount bond admits a
probabilistic representation in the ${\mathbb R}$ measure:
\begin{eqnarray}
P_{tT} = \frac{{\mathbb R}_t(X>T)}{{\mathbb R}_t(X>t)}.
\label{eq:20}
\end{eqnarray}
This formula shows that the price process of a discount bond is given by the
ratio of the ${\mathbb R}$-conditional probability that the positive random variable
$X$ taking values greater than $T$ and that of $X$ taking values greater than $t$.
By use of the Bayes formula, we deduce that (\ref{eq:20}) can alternatively be
expressed in the form
\begin{eqnarray}
P_{tT} = {\mathbb R}_t(X>T|X>t) ,
\label{eq:21}
\end{eqnarray}
since the set $\{X>t\}$ contains the set $\{X>T\}$. This is essentially the
representation obtained by Brody and Friedman (2009) for the discount bond
using the information-based approach to interest rate modelling.
It is worth remarking that the random variable $X$ has the dimension of time.
In Brody and Friedman (2009), $X$ was interpreted as the arrival time of liquidity
crisis, in the narrow sense of a cash demand. Hence, under this interpretation,
(\ref{eq:21}) shows that the bond price at $t$ is the probability that the timing of
the occurrence of a cash demand is beyond $T$, given that it has not yet
occurred at $t$, and given the noisy information (\ref{eq:16}) concerning the value
of $X$, in a suitably chosen measure ${\mathbb R}$.
\section{Back to the market measure}
The normalisation $\{N_t\}$ can be used to effect a measure change ${\mathbb R}
\to {\mathbb P}$. To see this, note first that the process $\{W_t\}$ defined by
\begin{eqnarray}
W_t = \xi_t - \int_0^t {\mathbb E}_s^{\mathbb R}[v_s(X)] \mbox{$\rm d$} s
\label{eq:22}
\end{eqnarray}
is an ${\mathbb R}$-Brownian motion with respect to the filtration $\{{\mathcal F}_t\}$
generated by the information process (\ref{eq:16}). In fact, this is just the innovations
representation for the filtering problem posed above. Thus, the Brownian property
can be verified by checking that $\{W_t\}$ satisfies the martingale condition:
\begin{eqnarray}
{\mathbb E}_t^{\mathbb R}\left[W_T\right] &=& {\mathbb E}_t^{\mathbb R}\left[
\int_0^T v_s(X) \mbox{$\rm d$} s + \beta_T - \int_0^T {\mathbb E}_s^{\mathbb R}[v_s(X)] \mbox{$\rm d$} s
\right] \nonumber \\ &=&
{\mathbb E}_t^{\mathbb R}\left[ \int_0^t v_s(X) \mbox{$\rm d$} s + \beta_t - \int_0^t
{\mathbb E}_s^{\mathbb R}[v_s(X)]
\mbox{$\rm d$} s \right] \nonumber \\ && + {\mathbb E}_t^{\mathbb R}\left[ \int_t^T v_s(X) \mbox{$\rm d$} s
+ (\beta_T-\beta_t) - \int_t^T {\mathbb E}_s^{\mathbb R}[v_s(X)] \mbox{$\rm d$} s \right]
\nonumber \\ &=& W_t,
\end{eqnarray}
where we have made use of the martingale property
${\mathbb E}_t^{\mathbb R}[{\mathbb E}_s^{\mathbb R}[v_s(X)]]=
{\mathbb E}_t^{\mathbb R}[v_s(X)]$ of the conditional expectation for $s>t$,
and the tower property of
conditional expectation to deduce ${\mathbb E}_t^{\mathbb R}[\beta_T] =
{\mathbb E}_t^{\mathbb R}[\beta_t]$. Along with $(\mbox{$\rm d$} W_t)^2=\mbox{$\rm d$} t$, L\'evy's
characterisation shows that $\{W_t\}$ is an ${\mathbb R}$-Brownian motion.
On the other hand, an application of Ito's lemma on (\ref{eq:18}) gives
\begin{eqnarray}
\frac{\mbox{$\rm d$} N_t}{N_t} = {\hat v}_{t} \mbox{$\rm d$} \xi_t,
\end{eqnarray}
from which it follows that
\begin{eqnarray}
N_t = \exp \left( \int_0^t {\hat v}_{s} \mbox{$\rm d$} \xi_s - \mbox{$\textstyle \frac{1}{2}$} \int_0^t {\hat v}_{s}^2
\mbox{$\rm d$} s \right)
\label{eq:25}
\end{eqnarray}
is a positive martingale satisfying $N_0=1$. Hence $\{N_t\}$ can be used as the
likelihood process to change the probability measure. Specifically, for any
${\mathcal F}_t$-measurable random variable $Z_t$ we have
\begin{eqnarray}
{\mathbb E}_s^{\mathbb R}\left[Z_t\right] = \frac{1}{N_s}{\mathbb E}_t^{\mathbb P}
\left[N_tZ_t\right] \quad {\rm and} \quad {\mathbb E}_s^{\mathbb P}\left[Z_t\right] =
N_s {\mathbb E}_t^{\mathbb R}\left[\frac{1}{N_t}Z_t\right].
\end{eqnarray}
In particular, (\ref{eq:22}) and (\ref{eq:25}) shows that $\{\xi_t\}$ is a Brownian
motion under the ${\mathbb P}$ measure. In addition, we find that the random
variable $X$ has the same probability law under ${\mathbb P}$ as under
${\mathbb R}$, and that $X$ and $\xi_t$ are ${\mathbb P}$-independent. These
properties can be verified by showing
\begin{eqnarray}
{\mathbb E}^{\mathbb P}[\mbox{$\rm e$}^{{\rm i}(x\xi_t+yX)}]=
{\mathbb E}^{\mathbb P}[\mbox{$\rm e$}^{{\rm i}x\xi_t}]{\mathbb E}^{\mathbb P}[\mbox{$\rm e$}^{{\rm i}yX}]
\end{eqnarray}
for all real $x,y$, and calculating the right side explicitly.
We remark that the conditional probability ${\mathbb R}_t(X>t)$ appearing in
(\ref{eq:19}) can be interpreted as representing the pricing kernel in the
${\mathbb R}$ measure. Specifically, writing $\Pi_t$ for ${\mathbb R}_t(X>t)$,
we deduce from (\ref{eq:17}) that
\begin{eqnarray}
\Pi_t = \frac{\int_t^\infty \rho_0(u) \exp\left(
\int_0^t v_s(u){\rm d}\xi_s - \frac{1}{2}\int_0^t v_s^2(u){\rm d}s\right)\mbox{$\rm d$} u}
{\int_0^{\infty} \rho_0(u) \exp\left(\int_0^t v_s(u){\rm d}\xi_s - \frac{1}{2}\int_0^t
v_s^2(u){\rm d}s\right) \mbox{$\rm d$} u}.
\label{eq:27}
\end{eqnarray}
A short calculation making use of (\ref{eq:22}) shows that the
${\mathbb R}$-pricing kernel (\ref{eq:27}) can be expressed manifestly in the
Flesaker-Hughston representation:
\begin{eqnarray}
\Pi_t = \int_t^\infty \rho_0(x) G_t(x) \mbox{$\rm d$} x,
\end{eqnarray}
where $\{G_t(x)\}$ is a one-parameter family of positive $\mathbb{R}$-martingales:
\begin{eqnarray}
G_t(x) = \exp \left( \int_0^t \tilde{v}_s(x) \mbox{$\rm d$} W_t -\mbox{$\textstyle \frac{1}{2}$} \int_0^t \tilde{v}_s(x)^2
\mbox{$\rm d$} t \right),
\label{eq:29}
\end{eqnarray}
and where $\tilde{v}_t(x) = v_t(x) - {\mathbb E}_t^{\mathbb R}[v_t(X)]$. The
dynamical equation satisfied by the $\mathbb{R}$-pricing kernel therefore reads
\begin{eqnarray}
\frac{\mbox{$\rm d$} \Pi_t}{\Pi_t} = -r_t \mbox{$\rm d$} t - ({\hat v}_{t} +\lambda_t) \mbox{$\rm d$} W_t,
\end{eqnarray}
where ${\hat v}_{t} = {\mathbb E}_t^{\mathbb R}[v_t(X)]$ and
$\lambda_{t} = -{\mathbb E}_t^{\mathbb R}[v_t(X)|X>t]$.
\section{Indeterminacy of the risk premium}
Returning to the ${\mathbb P}$-measure, we recall that once a parametric model
for the martingale volatility $\{v_t(x)\}$ is chosen, then prices of derivatives will
in general depend on this model choice. Hence $\{v_t(x)\}$ can be calibrated from
derivative prices. The initial term structure density $\rho_0(u)$, on the other hand,
can be calibrated from the initial yield curve. By substituting these ingredients
in (\ref{eq:10}) we thus obtain a market implied risk premium, subject to the model
choice. Of course, any tractable model is unlikely to fit all derivative prices. One
can nevertheless ask whether it is possible to fix $\{v_t(x)\}$ in a hypothetical
situation where one has access to the totality of liquidly-traded derivative prices
and an unlimited computational resource, i.e. whether it is possible in principle to
fix $\{v_t(x)\}$ unambiguously. Perhaps not surprisingly, the answer is negative.
To see this, suppose that the volatility of the Flesaker-Hughston martingale family
is decomposed in the form
\begin{eqnarray}
v_t(u) = \phi_t(u) - \alpha_t ,
\label{eq:31}
\end{eqnarray}
where the vector process $\{\alpha_t\}$ is independent of $X$, and has no
parametric dependence on $u$. The minus sign here is purely a matter of
convention. Then writing
\begin{eqnarray}
L_t = \exp\left(-\int_0^t \alpha_s \mbox{$\rm d$} \xi_s - \mbox{$\textstyle \frac{1}{2}$} \int_0^t \alpha^2_s \mbox{$\rm d$} s\right),
\end{eqnarray}
we find that the pricing kernel takes the form
\begin{eqnarray}
\pi_t = L_t \int_t^\infty \rho_0(u)\, \mbox{$\rm e$}^{ \int_0^t \phi_s(u) {\rm d} \xi_s -
\frac{1}{2} \int_0^t \phi^2_s(u) {\rm d} s + \int_0^t \phi_s(u)\alpha_s{\rm d} s} \mbox{$\rm d$} u.
\end{eqnarray}
It follows that the price at time $t$ of a contingent claim, with payout $H_T=
h(S_T)$ at $T>t$, is given by
\begin{eqnarray}
H_t &=& {\mathbb E}_t\left[ \frac{\pi_T}{\pi_t}\, H_T \right] \nonumber \\ &=&
{\mathbb E}_t\left[ \frac{L_T}{L_t} \frac{\int_T^\infty \rho_0(u)\, \mbox{$\rm e$}^{ \int_0^T
\phi_s(u) {\rm d} \xi_s - \frac{1}{2} \int_0^T \phi^2_s(u) {\rm d} s + \int_0^T
\phi_s(u)\alpha_s{\rm d} s} \mbox{$\rm d$} u}{\int_t^\infty \rho_0(u)\, \mbox{$\rm e$}^{ \int_0^t \phi_s(u)
{\rm d} \xi_s - \frac{1}{2} \int_0^t \phi^2_s(u) {\rm d} s + \int_0^t \phi_s(u)
\alpha_s{\rm d} s} {\rm d} u} \, H_T \right] \nonumber \\ &=&
{\mathbb E}_t^\alpha\left[ \frac{\int_T^\infty \rho_0(u)\, \mbox{$\rm e$}^{ \int_0^T
\phi_s(u) {\rm d} \xi_s - \frac{1}{2} \int_0^T \phi^2_s(u) {\rm d} s + \int_0^T
\phi_s(u)\alpha_s{\rm d} s} \mbox{$\rm d$} u}{\int_t^\infty \rho_0(u)\, \mbox{$\rm e$}^{ \int_0^t \phi_s(u)
{\rm d} \xi_s - \frac{1}{2} \int_0^t \phi^2_s(u) {\rm d} s + \int_0^t \phi_s(u)
\alpha_s{\rm d} s} {\rm d} u} \, H_T \right],
\label{eq:34}
\end{eqnarray}
where we have used $\{L_t\}$ as a density martingale to change the measure.
Evidently, under the new measure ${\mathbb P}^\alpha$, the process
$\{\xi_t^\alpha\}$ defined by
\begin{eqnarray}
\xi_t^\alpha = \xi_t + \int_0^t \alpha_s \mbox{$\rm d$} s
\label{eq:35}
\end{eqnarray}
is a standard Brownian motion. Substituting (\ref{eq:35}) in (\ref{eq:34}) we deduce
that
\begin{eqnarray}
H_t =
{\mathbb E}_t^\alpha\left[ \frac{\int_T^\infty \rho_0(u)\, \mbox{$\rm e$}^{ \int_0^T
\phi_s(u) {\rm d} \xi_s^\alpha - \frac{1}{2} \int_0^T \phi^2_s(u) {\rm d} s }
\mbox{$\rm d$} u}{\int_t^\infty \rho_0(u)\, \mbox{$\rm e$}^{ \int_0^t \phi_s(u) {\rm d} \xi_s^\alpha -
\frac{1}{2} \int_0^t \phi^2_s(u) {\rm d} s} {\rm d} u} \, H_T \right],
\label{eq:36}
\end{eqnarray}
which is identical to the pricing formula under the ${\mathbb P}$ measure
had $\{\alpha_t\}$ been identically zero in
the first place, on account of the following observation. The price of the underlying
asset at time $T$ can be expressed in the form
\begin{eqnarray}
S_T = S_0 \exp\left( \int_0^T \left(r_s+\lambda_s\sigma_s-\mbox{$\textstyle \frac{1}{2}$}\sigma_s^2\right)
\mbox{$\rm d$} s + \int_0^T \sigma_s \mbox{$\rm d$} \xi_s \right),
\label{eq:37}
\end{eqnarray}
where $\{\sigma_t\}$ is the volatility of $\{S_t\}$. Now if the volatility of the
martingale family $\{M_t(u)\}$ takes the form (\ref{eq:31}), then the risk premium
can be expressed as
\begin{eqnarray}
\lambda_t = \lambda_t^\alpha + \alpha_t,
\label{eq:38}
\end{eqnarray}
where $\{\lambda_t^\alpha\}$ is the risk premium in the $\mathbb{P}^\alpha$
measure:
\begin{eqnarray}
\lambda_t^\alpha = -\frac{\int_t^\infty \rho_0(u) \phi_t(u) \mbox{$\rm e$}^{
\int_0^t \phi_s(u) {\rm d} \xi_s^\alpha - \frac{1}{2} \int_0^t \phi^2_s(u)
{\rm d}s} \mbox{$\rm d$} u}{\int_t^{\infty} \rho_0(u) \mbox{$\rm e$}^{ \int_0^t \phi_s(u)
{\rm d}\xi_s^\alpha - \frac{1}{2} \int_0^t \phi^2_s(u) {\rm d} s} \mbox{$\rm d$} u} .
\label{eq:39}
\end{eqnarray}
Substituting (\ref{eq:35}) and (\ref{eq:38}) in (\ref{eq:37}), we obtain
\begin{eqnarray}
S_T = S_0 \exp\left( \int_0^T \left(r_s+\lambda_s^\alpha\sigma_s
-\mbox{$\textstyle \frac{1}{2}$}\sigma_s^2\right) \mbox{$\rm d$} s + \int_0^T \sigma_s \mbox{$\rm d$} \xi_s^\alpha \right).
\label{eq:40}
\end{eqnarray}
We thus find that the probability law of the random variable $S_T$, and hence
$H_T$, under the ${\mathbb P}$-measure with $\alpha_t=0$, is the same as that
under the ${\mathbb P}^{\alpha}$-measure with $\alpha_t\neq0$. It follows that
any \textit{addition} of terms in the martingale volatility $\{v_t(u)\}$ that is
independent of the parameter $u$ does not affect current price levels.
The above result shows that the risk premium vector $\{\lambda_t\}$ can be
determined from market prices of derivatives only up to an additive vectorial
term $\{\alpha_t\}$. This freedom, however, is not arbitrary; it can only
arise from a constant (i.e. independent of the parameter $u$) addition to the
volatility of the martingale family in the form of (\ref{eq:31}).
\section{Information-based interpretation}
The ambiguity in the determination of the risk premium can be interpreted from the
viewpoint of information-based asset pricing theory of Brody \textit{et al}. (2007).
In the information-based pricing framework one models the market filtration
directly in the form of an information process concerning market factors relevant
to the cash flows of a given asset. Our objective here, which extends the previous
work of Brody and Friedman (2009), is to analyse the model (\ref{eq:16}) for the
information process that determines the pricing kernel.
The interpretation of the information process (\ref{eq:16}) is as follows. Market
participants are concerned with the realised value of the random variable $X$,
which, in a certain sense can be interpreted as the timing of a serious liquidity
crisis. In reality, market participants observe price processes, or equivalently
the underlying Brownian motion family $\{\xi_t\}$. As indicated above, under the
physical ${\mathbb P}$-measure the random variables $X$ and $\xi_t$ are
independent. However, market participants `perceive' information with certain
risk adjustments characterised by the density martingale $\{N_t\}$ of (\ref{eq:25}).
In this risk-adjusted measure, the path $\{\xi_t\}$ represents the aggregate of
noisy information for the value of $X$ in the form of (\ref{eq:16}). The
`signal' concerning the value of $X$, in particular, is revealed to the market
through the structure function $\{v_t(u)\}$, which in turn determines the
volatility structure of the pricing kernel, and hence the risk premium.
Suppose that the structure function $\{v_t(u)\}$ takes the form (\ref{eq:31}),
where $\{\alpha_t\}$ is independent of $X$. Then because (\ref{eq:16})
represents the information process for the random variable $X$, the constant
$\{\alpha_t\}$ combines with the `noise' term $\{\beta_t\}$. In other words,
the choice of $\{\alpha_t\}$ is entirely equivalent to the choice of noise; the
Brownian noise is replaced by a drifted Brownian noise. This change of noise
composition does not affect current asset prices, and therefore is not
directly detectable from market data, even though asset-price drifts are
modified, in general in an unidentifiable manner. Note that the point of view
that the indeterminacy of the asset price drifts is caused by noise has been
put forward heuristically by Black (1986); our observation thus formalises this
argument more precisely.
It is worth remarking briefly the observation made in Brody and Friedman (2009)
concerning the form of the structure function $\{v_t(u)\}$ in the absence of the
noise drift $\{\alpha_t\}$. Since small values of
$X$ imply imminent liquidity crisis, in an ideal market the signal-to-noise ratio
of the information process (\ref{eq:16}) should be large for small values of $X$,
as compared to large values of $X$. In other words, under normal market
conditions we expect the signal magnitude $|v_t(u)|$ be decreasing in $u$ for
every $t$. Conversely, if $|v_t(u)|$ is increasing in $u$, then the excess rate of
return above the short rate for discount bonds, i.e. the inner product of the risk
premium and the discount bond volatility, is negative, yielding negative excess
rate of return due to the inverted form of the structure function $\{v_t(u)\}$.
\section{Anomalous price behaviour}
The fact that current asset prices are unaffected by changes in the structure of
the noise term does not imply that $\{\alpha_t\}$ can be ignored altogether. Indeed,
(\ref{eq:38}) shows that the existence of such a component does shift the
risk premium. Since the drift of an asset with volatility $\{\sigma_t\}$ is given in
the ${\mathbb P}$-measure by $r_t+\lambda_t\sigma_t$, the noise-induced
drift $\alpha_t\sigma_t$ can generate various anomalous price dynamics under
the physical ${\mathbb P}$-measure.
As an example, let us consider the case of an anomalous price growth, or a
bubble. In the large vector space of asset volatilities, it is inevitable that
volatility vectors form clusters consisting of different sectors or industries. This
is because, by definition, a given sector of companies share analogous
risk exposures. Now if an anomalous noise component $\{\alpha_t\}$ at some
point in time emerges to point in the direction of one of these volatility clusters,
then this can cause a sharp rise in the share prices of that sector. Since
the noise vector $\{\alpha_t\}$ carries no real economic information, this can
be identified as a bubble, where prices of a set of assets grow sharply, and
independently of the `true' state of affairs, without seriously affecting price
processes of other assets. Similarly, at a later time, the magnitude of
$\{\alpha_t\}$ can diminish. In particular, more reliable information concerning
the true state of affairs may be revealed, which in turn leads to an increase
in the magnitudes of volatilities on the one hand, while on the other hand the
risk premium vector can point in a direction such that the inner product
$\lambda_t\sigma_t$ takes a large negative value; thus leading to a bubble
`burst'.
In the finance and economics literature, there exists a substantial work on the
study of various aspects of financial bubbles (see, e.g., Camerer 1989 for an
early review). It is important to note that our characterisation of a bubble is
motivated by an information-based perspective. One commonly used definition
of a bubble, on the other hand, is given by the difference between the current
price and the expected discounted future cash flows in the risk-neutral
measure (cf. Tirole 1985, Heston \textit{et al}. 2007). Under this definition,
discounted asset prices in the risk-neutral measure can be modelled by use
of strict local martingales (Cox and Hobson 2005, Jarrow \textit{et al}. 2007,
2010), within the arbitrage-free pricing framework.
While this formulation of a bubble leads to the unravelling of many interesting
mathematical subtleties underlying fundamental theorems of asset pricing, from
an information-theoretic viewpoint the plausibility of such a definition for a bubble
seems questionable. In particular, a mathematical definition of a financial bubble
that involves no reference to the ${\mathbb P}$ measure seems restrictive; a
bubble, after all, is a phenomenon seen under the ${\mathbb P}$ measure.
The pricing kernel approach, on the other hand, is based on a stronger assumption
that if $\{S_t\}$ represents the price process of a liquidly traded asset, then
$\{\pi_tS_t\}$ must be a true ${\mathbb P}$-martingale. As such, the discounted
${\mathbb Q}$-expectation of future asset price necessarily agrees with the
current value, or else there are arbitrage opportunities.
The conventional definition of a financial bubble in terms of the inequality
$S_t>\pi_t^{-1}{\mathbb E}_t[\pi_T S_T]$ is sometimes justified heuristically by
the fact that some traders, when they are under the impression that there is a
bubble and thus traded prices are above the `fundamental' values, will
nevertheless participate in the apparent bubble with the view that they can
withdraw from their positions before the crunch (see, e.g., Camerer 1989 and
references cited therein). This example and other similar ones are often used in
support of the argument that some traders are willing to purchase stocks even
when they
know that the price level is above its fundamental value. The shortcomings in
such an argument are that (a) the role of market filtration is not adequately taken
into account; and that (b) the fact that such a stock purchase is equivalent to the
purchase of an American option is overlooked. A more plausible characterisation
of a bubble participation seems to be as follows. Given the information
$\{{\mathcal F}_t\}$, a trader estimates that there is a bubble that will continue
to grow for a while. Hence, subject to the filtration, the best estimate of the
future cash flow for this trader, with a suitable risk-adjustment, is given by
$\sup_{\tau}\pi_t^{-1}{\mathbb E}_t[\pi_\tau S_\tau]$, where $\tau$ is a stopping time
when the stock is sold. If this expectation agrees to the current price level, then
a transaction occurs. Conversely, it seems implausible that a transaction takes
place if the best estimate by a rational trader of a discounted cash flow is lower
than the current price level.
The view we put forward here is that a bubble in an asset ought to be identified
with an anomaly in the rate of return of that asset, and not with an anomaly in the
price level itself.
Here, a precise definition of an `anomaly' in the drift is essentially what we have
described above, namely, the existence of an additive term in the volatility of
the martingale family $\{M_t(u)\}$ that is constant in the parameter $u$. Based
on this definition, it is admissible that price processes behave in a manner that
does not always reflect what one might perceive as the true state of affairs, had
one possessed better information concerning the true worth of the assets. Put
the matter differently, decisions concerning transactions that ultimately lead to
price dynamics are made in accordance with the unfolding of information. Since
this information is necessarily noisy, the best filters chosen by market participants
will inevitably deviate from true values of assets being priced. If the noise
structure changes, then it is only reasonable that the dynamical aspects of these
deviations will likewise change. In particular, the increment of the innovations
representation---that characterises the arrival of `real' information over the
interval $[t,t+\mbox{$\rm d$} t]$---is
given by
\begin{eqnarray}
\mbox{$\rm d$} W_t = \mbox{$\rm d$} \xi_t - {\hat \phi}_t \mbox{$\rm d$} t + \alpha_t \mbox{$\rm d$} t,
\label{eq:41}
\end{eqnarray}
where ${\hat\phi}_t={\mathbb E}_t^{\mathbb R}[\phi_t(X)]$,
and this illustrates in which way the existence of a nonzero noise drift
$\{\alpha_t\}$ affects the dynamics.
Our characterisation of anomalous price dynamics is not confined to the
consideration of financial bubbles. Again, in the large vector space of asset
volatilities, it seems plausible that equity market volatilities and fixed-income
volatilities generally lie on distinct subspaces. If the noise vector $\{\alpha_t\}$
has a tendency to lie in the direction of equity-volatility subspace, then this
naturally leads to an excess growth in the equity market, explaining the
phenomena of the so-called equity premium puzzle, where over time the
rate of return associated with the equity market considerably exceeds that
of the bond market (see, e.g., Kocherlakota 1996 for a review).
\section{Relation to the risk-neutral measure}
We have established the relation between the auxiliary probability measure
${\mathbb R}$ and the physical measure ${\mathbb P}$. The relation between
the latter and the risk-neutral measure ${\mathbb Q}$, on the other, involves the
risk premium process $\{\lambda_t\}$. To recapitulate these two relations, we
have
\begin{eqnarray}
\mbox{$\rm d$} W_t = \mbox{$\rm d$} \xi_t - {\hat v}_t \mbox{$\rm d$} t \quad {\rm and} \quad \mbox{$\rm d$} W_t^* = \mbox{$\rm d$} \xi_t
+ \lambda_t \mbox{$\rm d$} t,
\label{eq:42}
\end{eqnarray}
where
\begin{eqnarray}
{\hat v}_t = \frac{\int_0^\infty \rho_0(u) v_t(u) M_t(u)\mbox{$\rm d$} u}
{\int_0^{\infty} \rho_0(u) M_t(u) \mbox{$\rm d$} u} \quad {\rm and} \quad
\lambda_t = -\frac{\int_t^\infty \rho_0(u) v_t(u) M_t(u)\mbox{$\rm d$} u}
{\int_t^{\infty} \rho_0(u) M_t(u) \mbox{$\rm d$} u} ,
\label{eq:43}
\end{eqnarray}
and where we let $\{W_t^*\}$ denote the ${\mathbb Q}$-Brownian motion. By
combining the two relations in (\ref{eq:42}) we deduce at once that the
measure-change density martingale is given by
\begin{eqnarray}
\left.\frac{\mbox{$\rm d$}{\mathbb Q}}{\mbox{$\rm d$} \mathbb{R}}\right|_{{\mathcal F}_t} = \exp\left(-
\int_0^t ({\hat v}_s+\lambda_s)\mbox{$\rm d$} W_s -\mbox{$\textstyle \frac{1}{2}$} \int_0^t ({\hat v}_s+\lambda_s)^2
\mbox{$\rm d$} s \right),
\end{eqnarray}
which determines the general relation between ${\mathbb Q}$ and ${\mathbb R}$.
As indicated above, a closer inspection on (\ref{eq:43}), however, shows that
\begin{eqnarray}
{\hat v}_t = {\mathbb E}_t^{\mathbb R}[v_t(X)] \quad {\rm and} \quad
\lambda_t = -{\mathbb E}_t^{\mathbb R}[v_t(X)|X>t] .
\label{eq:45}
\end{eqnarray}
In other words, under the restriction $X>t$, we have, conditionally, ${\hat v}_t+
\lambda_t=0$. Therefore, the auxiliary measure ${\mathbb R}$, whose existence
is ensured by the lack of arbitrage and the existence of pricing kernel, can be
interpreted as an extension of the risk-neutral measure. Conversely, by
restricting to the event $X>t$, we can think of ${\mathbb R}$ indeed as the
risk-adjusted measure.
\section{Stochastic volatility}
So far we have analysed the case for which $\{v_t(x)\}$ is a deterministic function
of time. The fact that the volatility structure of the martingale family $\{M_t(x)\}$ is
deterministic, however, does not imply deterministic volatilities for asset prices.
On the contrary, even for an elementary discount bond, the associated volatility
process is highly stochastic. Hence when we speak about a `stochastic volatility'
we have in mind the volatility for the martingale family $\{M_t(x)\}$, whereas the
stochasticity for asset prices is presumed.
From the viewpoint of practical implementation, it probably suffices to restrict
attention to deterministic volatility structures, since deterministic volatilities for
$\{M_t(x)\}$ give rise to a range of sophisticated stochastic volatility models for
asset prices. Indeed, it is shown in Brody \textit{et al}. (2011) that even in the
very restricted case of a single factor model with the time-independent volatility
$v_t(x)=\mbox{$\rm e$}^{-\sigma x}$ that depends only on one model parameter $\sigma$, it is
possible to calibrate caplet prices across different maturities reasonably accurately.
It is nevertheless of interest to enquire whether the auxiliary information process
exists in the more general context of stochastic volatilities. For this purpose, let
us begin by considering the case where $\{v_t(x)\}$ admits the decomposition
(\ref{eq:31}) and where $\{\phi_t(x)\}$ is deterministic and $\{\alpha_t\}$ is chosen
such that the noise term $n_t\equiv\int_0^t\alpha_s\mbox{$\rm d$} s+\beta_t$ is an
$\{{\mathcal F}_t^\beta\}$-measurable Gaussian process. Then an application of
the martingale representation theorem shows that $\{n_t\}$ admits a
decomposition of the form
\begin{eqnarray}
n_t = \int_0^t b_s \mbox{$\rm d$} s + \int_0^t \gamma_s \mbox{$\rm d$} \beta_s,
\end{eqnarray}
where $\{b_s\}$ and $\{\gamma_s\}$ are deterministic. A short calculation then
shows that an auxiliary information process
\begin{eqnarray}
\xi_t = \int_0^t \phi_s(X) \mbox{$\rm d$} s + n_t
\end{eqnarray}
in the ${\mathbb R}$ measure indeed exists, with the property that the scaled
information process $\int_0^t\gamma_s^{-1}\mbox{$\rm d$}\xi_s$ determines the market
Brownian motion and that $\{b_t\}$ plays the role similar to that of a
deterministic $\{\alpha_t\}$ in the previous analysis, and hence is not determinable
form current market prices.
The foregoing example shows how one can model the random rise and fall of
anomalous price dynamics. More generally, the structure function $\{v_t(x)\}$ can
depend in a general way on the history of the information process up to time $t$.
In this case, we obtain a generic stochastic volatility model for the martingale
family. Provided that the structure function is sufficiently well behaved so that
relevant stochastic integrals exist, the auxiliary information process can be seen
to exist in the ${\mathbb R}$ measure. To illustrate this, consider an elementary
`toy model' for which information process takes the form of an Ornstein-Uhlenbeck
process:
\begin{eqnarray}
\xi_t = \mbox{$\rm e$}^{\sigma X t}\int_0^t \mbox{$\rm e$}^{-\sigma Xs} \mbox{$\rm d$} \beta_s,
\end{eqnarray}
where $\sigma$ is a parameter, and $X$ and $\{\beta_t\}$ are independent. Such
an information process corresponds to a stochastic volatility model for which the
volatility process is given by a linear function of the ${\mathbb P}$-Brownian
motion: $v_t(x)=\sigma x \xi_t$.
\section{Discussion}
The main results of the paper are as follows: We have derived the existence of an
auxiliary filtering problem underlying arbitrage-free modelling of the pricing kernel;
the solution of which determines the volatility structure of the
positive martingale family $\{M_t(u)\}$ appearing in the Flesaker-Hughston
representation for the pricing kernel. We have demonstrated that the structure of
the ambient information process fully characterises the risk premium process
$\{\lambda_t\}$. We have shown, under the Brownian-filtration setup, that
$\{\lambda_t\}$ admits a canonical decomposition into two terms in an additive
manner; the systematic term that can be calibrated from current market data for
derivative prices, and the idiosyncratic term that cannot be estimated (unless, of
course, one can estimate drift processes of risky assets), and thus can be
identified as noise.
It is worth emphasising that these results hold irrespective of our choice of
interpretation. Nevertheless, our characterisation of anomalous price dynamics
seems sufficiently compelling, for, such phenomena are ultimately observed
under the physical measure ${\mathbb P}$. One might ask what causes the
evolution of the noise drift $\{\alpha_t\}$. This is an interesting
econometric question that, however, goes beyond the scope of the present
investigation. It
suffices to remark that the random variable $X$ that constitutes the signal
component of the ambient information process has units of time, and thus is
ultimately linked to the term structure of financial markets. One possible
explanation of the excess equity premium therefore is that fixed-income market
intrinsically embodies more information concerning the term structure as
compared to the equity market, and this imbalance is manifested in the form of
an additional drift in the noise component pointing generally towards the
direction of equity volatility vectors.
\vskip 10pt
\noindent The authors thank Mark Davis, Robyn Friedman, Matheus Grasselli,
Lane Hughston, Andrea Macrina, and Bernhard Meister for comments and
stimulating discussions.
|
\section{Results}
\subsection{A Quantitative Model for $c_{i}(r)$}
For each scientist $i$, we find that $c_{i}(r)$ can be approximated by
a scaling regime for small $r$ values, followed by a truncated scaling regime for large $r$ values.
Recently a novel distribution, the discrete generalized beta distribution (DGBD)
\begin{equation}
c_{i}(r) \equiv A_{i} r^{-\beta_{i}} (N_{i}+1-r)^{\gamma_{i}}
\label{Cr}
\end{equation}
has been proposed as a model for rank profiles in the social and natural sciences that
exhibit such truncated scaling behavior \cite{DGBfunc, RankOrder}.
The parameters $A_{i}$, $\beta_{i}$, $\gamma_{i}$ and $N_{i}$ are each defined for a given $c_{i}(r)$ corresponding to an
individual scientists $i$, however we suppress the index $i$ in some equations to keep the notation concise. We estimate the
two scaling parameters $\beta_{i}$ and $\gamma_{i}$ using {\it Mathematica} software to perform a multiple linear regression of $\ln c_{i}(r) = \ln A_{i} - \beta_{i} \ln r + \gamma_{i} \ln (N_{i}+1-r)$ in the base functions $\ln r$ and $\ln (N_{i}+1-r)$. In our fitting procedure we replace $N$ with
$r_{1}$, the largest value of $r$ for which $c(r) \geq 1$ (we find that $r_{1} / N_{i} \approx 0.84 \pm 0.01$ for careers in datasets A and B). Figs. \ref{Ztop5} and \ref{ZAveTop100} demonstrate the utility of the DGBD to represent
$c_{i}(r)$, for both large and small $r$. The regression correlation coefficient $R_{i}>0.97$ for all $\ln
c_{i}(r)$ profiles analyzed.
The DGBD proposed in \cite{DGBfunc} is an improvement over the Zipf law (also called the generalized power-law or Lotka-law \cite{EggheLotka}) model and the stretched exponential
model \cite{H} since it reproduces the varying curvature in $c_{i}(r)$ for both small and large $r$.
Typically, an exponential cutoff is imposed in the power-law model, and justified
as a finite-size effect. The DGBD does not require this assumption, but rather, introduces a second scaling exponent
$\gamma_{i}$ which controls the curvature in $c_{i}(r)$ for large $r$ values. The DGBD has been
successfully used to model numerous rank-ordering profiles analyzed in \cite{DGBfunc, RankOrder} which arise in the
natural and socio-economic sciences. The relative values of the $\beta_{i}$ and
$\gamma_{i}$ exponents are thought to capture two distinct mechanisms that contribute to the evolution of $c_{i}(r)$
\cite{DGBfunc, RankOrder}. Due to the data limitations in this study, we are not able to study the dynamics in $c_{i}(r)$ through time. Each $c_{i}(r)$ is a ``snapshot'' in time, and so we can only conjecture on the evolution of $c_{i}(r)$ throughout the career. Nevertheless, we believe that there is likely a positive feedback effect between the ``heavy-weight'' papers and ``newborn'' papers, whereby the reputation of the ``heavy-weight'' papers can increase the exposure and impact the perceived significance of ``newborn'' papers during their infant phase.
Moreover, the 2-regime power-law behavior of $c_{i}(r)$ suggests that the reinforcement dynamics can be quantified by the scale-free parameters $\beta$ and $\gamma$.
The $\beta_{i}$ value determines the relative change in the $c_{i}(r)$
values for the high-rank papers, and thus it can be used to further distinguish
the careers of two scientists with the same $h$-index.
In particular, smaller $\beta$ values characterize flat profiles with relatively low contrast between the high and low-rank regions of any given profile, while larger $\beta$ values indicate a sharper separation between the two regions.
In Fig. \ref{ZAveTop100}(a) we plot $c_{i}(r)$ for each scientist from dataset [A] as well as
the average of the 100 individual curves $\overline{c}(r) \equiv\frac{1}{100}\sum_{i=1}^{100} c_{i}(r)$ (see Figs. \ref{ZAveR100} and \ref{ZAveAsst100}
for analogous plots for datasets [B] and [C]).
We find robust power-law scaling
\begin{equation}
\overline{c}(r) \sim r^{-\beta} \ \ \ [\beta \approx 0.92 \pm 0.01]
\end{equation}
for $10^{0} \leq r \leq 10^{2}$.
Interestingly, this $\beta$ value is similar to the scaling exponents calculated for other rank-size (Zipf) distributions
in the social and economic sciences, e.g. word frequency \cite{ZipfLawWord}
and city size \cite{DGBfunc, RankOrder,ZipfLawCity}.
\begin{figure}
\centering{\includegraphics[width=0.45\textwidth]{Fig2.pdf}}
\caption{\label{ZAveTop100} {\bf Data collapse of each $c_{i}(r)$ along a universal curve}. A comparison of 100 rank-citation profiles $c_{i}(r)$ demonstrates the statistical regularity in career publication output. Each scientist produces a cascade of papers of varying impact between the $c_{i}(1)$ pillar paper down to the least-known paper $c_{i}(N_{i})$. {\bf (a)} Zipf rank-citation profiles $c_{i}(r)$ for 100 scientists listed in dataset [A].
For reference, we plot the average $\overline{c}(r)$ of these 100 curves and find $\overline{c}(r) \sim r^{-\beta}$ with $\beta =
0.92 \pm 0.01$. The solid green line is a least-squares fit to $\overline{c}(r)$ over the range $1 \leq r
\leq 100$.
We also plot the $H_{2}(r)$ and $H_{80}(r)$ lines for reference.
{\bf (b)} We re-scale the curves in panel (a), plotting $c_{i}(r') \equiv c_{i}(r)/ A(r_{1}+1-r)^{\gamma}$, where we use
the best-fit $\gamma_{i}$ and $A_{i}$ parameter values for each individual $c_{i}(r)$ profile. Using the rescaled rank value $r' \equiv
r^{\beta_{i}}$, we
show excellent data collapse onto the expected curve $c(r') = 1/r'$. (see Figs. \ref{ZAveR100} and \ref{ZAveAsst100}
for analogous plots for dataset [B] and [C] scientists). Green data points correspond to the average $c(r')$ value with 1$\sigma$ error bars calculated using all 100 $c_{i}(r')$ curves separated into logarithmically spaced bins.
}
\end{figure}
We calculate each $\beta_{i}$ value using a multilinear least-squares regression of $\ln c_{i}(r)$ for $ 1 \leq r \leq
r_{1}$ using the DGBD model defined in Eq.~[\ref{Cr}].
To properly weight the data points for better regression fit over the entire
range, we use only $20$ values of $c_{i}(r)$ data points that are equally spaced on
the logarithmic scale in the range $r \in [1,r_{1}]$. We elaborate the details of this fitting technique in the methods section.
We plot five empirical $c_{i}(r)$ along with their corresponding best-fit DGBD functions in Fig. \ref{Ztop5} to demonstrate the goodness of fit for the
entire range of $r$.
In order to demonstrate the common functional form of the DGBD model, we
collapse each $c_{i}(r)$ along a universal
scaling function $c(r') = 1/r'$, by using the rescaled rank values $r' \equiv r^{\beta_{i}}$ defined for each curve.
In Figs. \ref{ZAveTop100}(b), \ref{ZAveR100}(b) and \ref{ZAveAsst100}(b), we plot the quantity $c_{i}(r') \equiv c_{i}(r)/ A(r_{1}+1-r)^{\gamma}$,
using the best-fit $\gamma_{i}$ and $A_{i}$ parameter values for each individual
$c_{i}(r)$ profile. While the curves in Fig. \ref{ZAveTop100}(a) are jumbled and distributed over a large range of $c(r)$ values,
the rescaled $c_{i}(r)$ curves in Fig. \ref{ZAveTop100}(b) all lie approximately along the predicted curve $c(r') = 1/r'$.
\subsection{Using $c_{i}(r)$ to quantify career production and impact}
A main advantage of the $h$-index is the simplicity in which it is
calculated, e.g. {\it ISI Web of Knowledge} \cite{WofK} readily
provides this quantity online for distinct authors.
Another strength of the $h$-index is its stable growth with respect to changes in $c_{i}(r)$ due to time and information-dependent factors \cite{Hstability}. Indeed, the $h$-index is
a ``fixed-point" of the citation profile. This time stability is evident in the observed growth rates of $h$ for scientists.
Average growth rates, calculated here as $h/L$, where $L$ is the duration in years between a given author's first and
most recent paper,
typically lie in the range of one to three units per year (this annual growth rate corresponds to
the quantity $m$ introduced by Hirsch \cite{H}). Annual growth rates $h/L \approx 3$ correspond to
exceptional scientists (for the histogram of $P(h/L)$ see the Fig. \ref{H1byCLPDF} and for $h/L$ values see the SI text (Tables S1-S4)).
As a result, $h/L$ is a good predictor for future achievement along with $h$ \cite{H2}.
It is truly remarkable how a single number, $h_{i}$, correlates with other measures of impact. Understandably, being just a single number, the $h$-index cannot
fully account for other factors, such as variations in citation standards and coauthorship patterns across discipline \cite{hindexResearchers,hindexFields, ProConH}, nor can $h_{i}$
incorporate the full information contained in the entire $c_{i}(r)$ profile.
As a result, it is widely appreciated that the $h$-index can underrate the value of the best-cited papers, since once a paper transitions into the region $ r \leq h_{i}$,
its citation record is discounted, until other less-cited papers with
$r>h_{i}$ eventually overcome the rank ``barrier" $r = h_{i}$. Moreover, as noted in \cite{H}, the papers for which $r > h_{i}$ do not contribute any additional credit.
\begin{figure}
\centering{\includegraphics[width=0.49\textwidth]{Fig3.pdf}}
\caption{\label{CrstarH}
{\bf Limitations to the use of the $h$-index alone.}
The $h$-index can be insufficient in comprehensively representing $c_{i}(r)$.
{\bf (a)} The $h$-index does not contain any information about $c_{i}(r)$ for $r < h_{i}$, and can shield a scientist's most successful accomplishments which are the basis for much of a scientist's reputation.
This is evident in the cases where $c(r_{i}^{*}) \gg h_{i}$, in which case the $h$-index cannot account for the stellar impact of the papers. {\bf (b)}
For a given $h_{i}$ value, prolific careers are characterized by a large $\beta_{i}$ value, as it is harder to maintain large $\beta_{i}$ values for large $h_{i}$. As a result, the
$\beta_{i}$ vs $h_{i}$ parameter space can be used to identify anomalous careers and to better compare two scientists with similar $h_{i}$ indices. We find that a third career metric $C_{i}$, the total number of citations to the papers of author $i$, can be calculated with high accuracy by the scaling relation $C_{i} \sim h_{i}^{1+\beta_{i}}$, which we illustrate in Fig. \ref{CbetaH1}(b). }
\end{figure}
Instead of choosing an arbitrary $h_{p}$ as an productivity-impact indicator,
we use the analytic properties
of the DGBD to calculate a crossover value $r_{i}^{*}$.
In the methods section, we derive an exact expression for $r_{i}^{*}$
which highlights the distinguished papers of a given author.
To calculate $r_{i}^{*}$, we use the logarithmic derivative $\chi(r) \equiv d\ln c(r) / dr$
to quantify the relative change in $c_{i}(r)$ with increasing $r$.
We defined papers as ``distinguished'' if they satisfy the inequality $c_{i}(r)/c_{i}(r+1) > \exp (\overline{\chi})$, where $\overline{\chi}$ is the average value of $\chi(r)$ over the entire range of $r$ values.
This inequality selects the peak papers which are significantly more cited than their neighbors.
The peak region $r \in [1,r_{i}^{*}]$ corresponds to a ``knee" in $c_{i}(r)$ when plotted on log-linear axes. The dependence of $\overline{\chi}$ and $r_{i}^{*}$ on the three DGBD parameters $\beta_{i}$, $\gamma_{i}$ and $N_{i}$ are provided in the methods section.
The advantage of $r_{i}^{*}$ is that
this characteristic
rank value is a comprehensive representation of the stellar papers in the high-rank scaling regime since it depends on the DGBD parameter values $\beta_{i}$, $\gamma_{i}$ and $N_{i}$, and thus
probes the entire citation profile.
Fig. \ref{CrstarH} shows a scatter plot of the ``$c$-star'' $c_{i}^{*} \equiv c_{i}(r_{i}^{*})$ and $h_{i}$
values calculated for each scientist and demonstrates that there is a
non-trivial relation between these two
single-value indices. It also shows that for scientists within a small range of $c^{*}$ there is a large variation in the corresponding $h$ values, in some cases straddling across
all three sets of scientists. Also, there are several $c_{i}^{*}$ values which significantly
deviate from the trend in Fig. \ref{CrstarH}, which is plotted on log-log axes.
These results reflect the fact that
the $h$-index cannot completely incorporate the entire $c_{i}(r)$ profile.
We plot the histogram of $c_{i}^{*}$ and $r_{i}^{*}$ values in Figs. \ref{CstarPDF} and \ref{rstarPDF}, respectively.
To further contrast the values of $c_{i}^{*}$ and the $h$-index, we propose the
``peak indicator" ratio $\Lambda_{i} \equiv c_{i}^{*} / h_{i}$, which corrects specifically
for the $h$-index penalty on the stellar papers in the peak region of $c_{i}(r)$.
Thus, all papers in the peak region of $c_{i}(r)$ satisfy the condition $c_{i}(r)\geq h_{i} \Lambda_{i}$.
In an extreme example, R. P. Feynman has a peak value $\Lambda \approx 36$, indicating that his best papers are
monumental pillars with respect to his other papers which contribute to his $h$-index. Fig. \ref{PeakPDF} shows the histogram of $\Lambda_{i}$ values, with typical values for
dataset [A] scientists $\langle \Lambda \rangle \approx 3.4 \pm 3.9$, and for dataset [B] scientists $\langle \Lambda \rangle \approx 2.2 \pm 1.1$.
This indicator can only be used to compare scientists with similar $h$ values, since a small $h_{i}$ can result in a large $\Lambda_{i}$.
An alternative ``single number'' indicator is $C_{i}$, an author's total number of citations
\begin{equation}
C_{i}=\sum_{r=1}^{N} c_{i}(r) \ ,
\label{Csum}
\end{equation}
which incorporates the entire $c_{i}(r)$ profile. However, it has been shown that $\sqrt{C_{i}}$ correlates well with $h_{i}$ \cite{Hredner}, a result which we will demonstrate in Eq. [\ref{Cbh}] to follow directly from a $c_{i}(r)$ with $\beta_{i} \approx 1$.
\begin{figure}
\centering{\includegraphics[width=0.49\textwidth]{Fig4.pdf}}
\caption{\label{CbetaH1}
{\bf Aggregate publication impact $C$.}
The total number of citations $C_{i}\sim h_{i}^{1+\beta_{i}}$ is also comprehensive productivity-impact measure.
For most best-fit DGBD model curves, the $C_{i}$ value is preserved with high precision.
This shows that the difference between a given $c_{i}(r)$ and the corresponding best-fit DGBD model function are negligible on the macroscopic scale.
{\bf (a)} The exact aggregate number of citations $C_{i}$, calculated from $c_{i}(r)$ using Eq. [\ref{Csum}], can be analytically approximated by $C_{\beta, h}$ using Eq. [\ref{Cbh}] which depends only on the scientist's $\beta_{i}$ and $h_{i}$ values.
{\bf (b)}
We justify the use of the DGBD model defined in Eq. [\ref{Cr}] for the approximation of $c_{i}(r)$ by comparing the aggregate citations $C_{i}$ with the expected aggregate citations $C_{m} = \sum_{r=1}^{r_{1}} c_{m}(r)$ calculated from the best-fit DGBD model $c_{m}(r)$. Including the extra scaling-parameter, as in the DGBD model, improves the agreement between the theoretical and empirical $C_{i}$ values in (a) and (b). We plot the line $y=x$ (dashed-green line) for visual reference. }
\end{figure}
We test the aggregate properties of $c_{i}(r)$ by
calculating
the aggregate number of citations $C_{\beta, h}$ for a given profile,
\begin{equation}
C_{\beta, h} \equiv \sum_{r=1}^{N} A r^{-\beta} \approx h^{1+\beta} \sum_{r=1}^{N'} r^{-\beta} = h^{1+\beta}H_{N',
\beta} \sim h^{1+\beta}
\label{Cbh}
\end{equation}
where $H_{N', \beta}$ is the {\it generalized harmonic number} and is of order $O(1)$ for $\beta \approx 1$.
We neglect the $\gamma_{i}$ scaling regime since the low-rank papers do not significantly contribute to an author's $C_{i}$ tally.
We approximate the coefficient $A$ in Eq.~[\ref{Cbh}] using the definition $c(h)\equiv h$, which implies that
$A/h^{\beta} \approx h$.
We use the value $N' \equiv 3h$, so that $C_{\beta, h}$ can be approximated by only the two parameters $h_{i}$ and $\beta_{i}$ for
any given author.
We justify this choice of $N'$ by examining the rescaled $c_{i}(r/h)$, which we consider to be
negligible beyond rank $r = 3 h_{i}$ for most scientists.
In Fig.~\ref{CbetaH1}(a), we plot for each scientist the predicted $C_{\beta, h}$ value versus the empirical $C_{i}$ value,
and we find excellent agreement with our theoretical prediction $C_{i}\sim h_{i}^{1+\beta_{i}}$ given by Eq. [\ref{Cbh}]. In Fig.~\ref{CbetaH1}(b), we plot
for each scientist the
total number of citations $C_{m} = \sum_{r=1}^{r_{1}} c_{m}(r)$ using the best-fit DGBD model $c_{m}(r) \equiv c_{i}(r; \beta_{i}, \gamma_{i}, A_{i}, r_{1})$ to approximate $c_{i}(r)$. The excellent agreement demonstrates that the fluctuations in the residual difference $c_{m}(r) - c_{i}(r)$ cancel out on the aggregate level. Furthermore, a comparison of the quality of agreement between the theoretical $C_{i}$ values and the empirical $C_{i}$ values in Fig.~\ref{CbetaH1}(a) and (b) shows the importance of the additional $\gamma_{i}$ scaling regime in the DGBD model.
\section{Discussion}
We use the DGBD model to provide an
analytic description of $c_{i}(r)$ over the entire range of $r$, and provide a deeper quantitative understanding of
scientific impact arising from an author's career publication works. The DGBD model exhibits scaling
behavior for both large and small $r$, where the
scaling for small $r$ is quantified by the exponent $\beta_{i}$, which for many scientists analyzed, can be approximated
using only
two values of the generalized $h$-index $h_{p}$ (see SI text).
In particular, we show that for a given $h$-value, a larger $\beta_{i}$ value corresponds to a more prolific publication career, since $C_{i}\sim h_{i}^{1+\beta_{i}}$.
Many studies analyze only the high rank values of
generic Zipf ranking profiles $c(r)$, e.g. computing the scaling regime for $r<r_{c}$ below some some rank cutoff $r_{c}$.
However, these
studies cannot quantitatively relate the large observations to the small observations within the system of interest.
To account for this shortcoming, our method for calculating the crossover
values $r_{i}^{*}\equiv \overline{r}_{-}$, $r_{\mathsf x }$, and $\overline{r}_{+}$, which we elaborate in the methods section, can be used in general to quantitatively distinguish relatively large
observations and relatively small observations within the entire set of observations.
Moreover, the DGBD model has been shown to have wide application in
quantifying the Zipf rank profiles in various phenomena \cite{RankOrder}.
To measure the upward mobility of a scientist's career, in the SI text we address the question: given that a scientist has index $h$, what
is her/his most likely $h$-index value $\Delta t$ years in the future?
In consideration of the bulk of $c_{i}(r)$, and following from the regularity of $c_{i}(r)$ for $r\approx h$,
we propose a model-free gap-index
$G(\Delta h)$ as both an estimate and a target for future achievement which can be used in the review of career
advancement. The gap index $G(\Delta h)$, defined as a proxy for the total number of
citations a scientist needs to reach a target value $h+\Delta h$, can detect the potential for fast $h$-index growth by quantifying $c_{i}(r)$
around $h$. This estimator differs from other estimators for the time-dependent $h$-index \cite{DynamicH, StochasticModel, SimulatingH} in that $G(\Delta h)$ is model independent.
Even though the productivity of scientists can vary substantially \cite{Scientists, ShockleyProductivity, ProductDiff,
huber98, PetersonPNAS}, and despite the complexity of success in academia, we find remarkable statistical regularity in the
functional form of $c_{i}(r)$ for the scientists analyzed here from the physics community. Recent work in
\cite{Scientists, UnivCite, Rad2} calculates the citation distributions of papers from various disciplines and shows that proper normalization of impact measures can
allow for comparison across time and discipline.
Hence, it is likely that the publication careers of productive scientists in many disciplines obey the statistical
regularities observed here for the set of 300 physicists. Towards developing a model for career evolution, it is still unclear how
the relative strengths of two contributing factors (i) the extrinsic cumulative advantage effect \cite{Matthew1, Matthew2, Scientists} versus (ii) the intrinsic role
of the "sacred spark" in combination with intellectual genius \cite{ProductDiff} manifest in the parameters of the DGBD model.
With little calculation, the $\beta_{i}$ metric developed here, used in conjunction with the $h_{i}$, can better answer
the question, ``How popular are your papers?'' \cite{howpopular}.
Since the cumulative impact and productivity of individual scientists are also found to obey
statistical laws \cite{Scientists, BB2}, it is possible that the competitive nature of scientific advancement can be
quantified and utilized in order to monitor career progress.
Interestingly, there is strong evidence for a governing mechanism of career progress based on cumulative advantage \cite{Scientists,cumadvprocess, BB2}
coupled with the
the inherent talent of an individual, which results in statistical regularities in the career achievements of scientists as well as professional athletes \cite{BB0,BB2,BB3}.
Hence, whenever data are available \cite{CompSocScience, SocialDynamicsRev}, finding statistical regularities
emerging from human endeavors is a first step towards better understand the dynamics of human productivity.
\section{Methods}
\subsection{Selection of scientists and data collection}
We use disambiguated ``distinct author'' data from {\it ISI Web of Knowledge}.
This online database is host to comprehensive data that is well-suited for developing testable models for
scientific impact \cite{DiffusionRanking,Scientists,Rad2} and career progress \cite{BB2}.
In order to approximately control for discipline-specific publication and citation factors, we analyze 300 scientists from the field of physics.
We aggregate all authors who published in {\it Physical Review Letters} (PRL) over the 50-year period 1958-2008 into a common dataset. From this dataset, we rank the scientists using the citations shares metric defined in
\cite{Scientists}. This citation shares metric divides equally the total number of citations a paper receives among the
$n$ coauthors, and also normalizes the total number of citations by a time-dependent factor to account for citation
variations across time and discipline.
Hence, for each scientist in the PRL database, we calculate a cumulative number of citation shares received
from only their PRL publications. This tally serves as a proxy for his/her scientific impact in all journals.
The top 100 scientists according to this citation shares metric comprise dataset [A].
As a control, we also choose 100
other dataset [B] scientists, approximately randomly, from our ranked PRL list. The selection criteria for the control
dataset [B] group are that an author must have published between 10 and 50 papers in PRL. This likely ensures that the
total publication history, in all journals, be on the order of 100 articles for each author selected.
We compare the tenured scientists in datasets A and B with 100 relatively young assistant professors in dataset [C].
To select dataset [C] scientists, we chose two assistant professors from the top 50 U.S. physics and astronomy departments (ranked
according to the magazine {\it U.S. News}).
For privacy reasons, we provide in the SI tables only the abbreviated initials for each scientist's name (last name initial, first and middle name initial, e.g. L, FM). Upon request we can provide full names.
We downloaded datasets A and B from ISI Web of Science in Jan. 2010 and dataset C from ISI
ISI Web of Science in Oct. 2010.
We used the ``Distinct Author Sets" function provided by ISI in order
to increase the likelihood that only papers published by each given author are analyzed. On a case by case basis, we performed further
author disambiguation for each author.
\subsection{Statistical significance tests for the $c(r)$ DGBD model}
We test the statistical significance of the DGBD model fit using the $\chi^{2}$ test
between the 3-parameter best-fit DGBD $c_{m}(r)$ and the empirical $c_{i}(r)$.
We calculate the $p$-value for the $\chi^{2}$ distribution with $r_{1}-3$ degrees of freedom and find, for each data set, the number $N_{> p_{c}}$ of $c_{i}(r)$ with $p$-value $> p_{c}$: $N_{> p_{c}}$ =
4 [A], 19 [B], 22 [C] for $p_{c}=0.05$, and 8 [A], 22 [B], 37 [C] for $p_{c} = 0.01$.
The significant number of $c_{i}(r)$ which
do not pass the $\chi^{2}$ test for $p_{c} = 0.05$, results from the fact that the DGBD is a scaling function over several orders of magnitude in both $r$ and $c_{i}(r)$ values, and so the residual differences $[c_{i}(r) - c_{m}(r)]$ are not expected to be normally distributed since there is no characteristic scale for scaling functions such as the DGBD. Nevertheless, the fact that so many $c_{i}(r)$ do pass the $\chi^{2}$ test at such a high significance level, provides evidence for the quality-of-fit of the DGBD model. For comparison, none of the $c_{i}(r)$ pass the $\chi^{2}$ test using the power-law model at the $p_{c} = 0.05$ significance level. In the next section, we will also compare the macroscopic agreement in the total number of citations for each scientist and the total number of citations predicted by the DGBD model for each scientist, and find excellent agreement.
s
\begin{figure*}
\centering{\includegraphics[width=0.6\textwidth]{Fig6.pdf}}
\caption{\label{demo} {\bf Characteristic properties of the DGBD.}
We graphically illustrate the derivation of the characteristic $c_{i}(r)$ crossover values that locate the two tail regimes of $c_{i}(r)$, in particular, the
distinguished ``peak" paper regime corresponding to paper ranks $r \leq r^{*}$ (shaded region).
The crossover between two scaling regimes suggests a
complex reinforcement relation between the impact of a scientist's most famous papers and the impact of his/her other papers.
{\bf(a)} The $c_{i}(r)$ plotted on log-log axes with $N=278$, $\beta=0.83$ and $\gamma=0.67$, corresponding to the average values of the Dataset [A] scientists.The hatched magenta curve is the $H_{1}(z)$ line on the log-linear
scale with corresponding $h$-index value $h = 104$. The $r^{*}$ value for $c_{i}(r)$ is not visibly obvious.
{\bf(b)}
We plot on log-linear axes the centered citation profile $c_{i}(z)$ (solid black curve) given by the symmetric rank transformation $z = r -z_{0}$ in Eq. [\ref{cz}]. This representation better highlights the peak paper regime, but fails to highlight the power-law $\beta$ scaling.
{\bf(c)} We plot the corresponding logarithmic derivative $\chi(z)$ of $c(z)$ (solid black curve),
which represents the relative change in $c(z)$. The dashed red line corresponds to $-\overline{\chi}$, where $\overline{\chi}
$ is the average value of $\chi(z)$ given by Eq. [\ref{avechi}]. The values of $\overline{z}_{\pm}$, indicated by the
solid vertical green lines, are defined as the intersection of $\overline{\chi}$ with $\chi(z)$ given by Eq. [\ref{zpm}].
The regime $z<\overline{z}_{-}$ corresponds to the best papers of a given author. The hatched blue line corresponds to
$z_{\mathsf x }^{-}$ which marks the crossover between the $\beta$ and $\gamma$ scaling regimes. }
\end{figure*}
\subsection{Derivation of the characteristic DGBD $r$ values }
Here we use the analytic properties of the DGBD defined in Eq. [\ref{Cr}] to calculate
the special $r$ values from the parameters $\beta$, $\gamma$ and $N$ which locate the two tail regimes of $c(z)$, and in particular, the distinguished paper regime.
The scaling features of the DGBD do not
readily convey any characteristic scales which distinguish the two scaling regimes.
Instead, we use the properties of $\ln c_{i}(r)$ to characterize the crossover
between the high-rank and the low-rank regimes of $c_{i}(r)$.
We begin by considering $c_{i}(r)$ under the centered rank transformation $z = r- z_{0}$, where $z_{0}=(N+1)/2$, then
\begin{equation}
c(z) = A \frac{(z_{0}-z)^{\gamma}}{(z_{0}+z)^{\beta}} \ ,
\label{cz}
\end{equation}
in the domain $z \in [ -(z_{0}-1), (z_{0}-1)]$.
The logarithmic derivative of $c(z)$ expresses the relative change in $c(z)$,
\begin{eqnarray}
\chi(z) &\equiv& \frac{d \ln c(z)}{dz} = \frac{d c(z)/dz}{c(z)} \nonumber \\
&=& - \Big( \frac{\gamma}{z_{0}-z} + \frac{\beta}{z_{0}+z}\Big)= - m \Big( \frac{1 + \theta x}{1-x^{2}}\Big) \ ,
\end{eqnarray}
where $x=z/z_{0}$, $\theta = \frac{\gamma- \beta}{\gamma + \beta}$, and $m= \Big(\frac{\gamma + \beta}{z_{0}}\Big)$.
The extreme values of $\frac{d \ln c(z)}{dz}$ for $z_{0} \gg 1$ are given by
\begin{eqnarray}
\frac{d \ln c(z)}{dz}\Big|_{z = -(z_{0}-1)} \approx - \beta \\
\frac{d \ln c(z)}{dz}\Big|_{z = z_{0}-1} \approx - \gamma
\end{eqnarray}
and the average value $\overline{\chi}$ is calculated by,
\begin{eqnarray}
\overline{\chi} &\equiv& \langle \frac{d \ln c(z)}{dz} \rangle \nonumber \\
&=& \frac{-m}{(1-1/z_{0})-(1/z_{0}-1)}\int_{-(1-1/z_{0})}^{(1-1/z_{0})} dx \frac{(1+\theta x)}{1-x^{2}} \nonumber\\
&=& \frac{- m}{2} \ln N
\end{eqnarray}
The function $\chi(z)$ takes on the value of $\overline{\chi}$ twice at the values $\overline{z}_{\pm} = z_{0} \overline{x}_{\pm}$ corresponding to the solutions to the quadratic equation,\\
\begin{equation}
\overline{\chi} = - m \Big( \frac{1 + \theta x}{1-x^{2}}\Big) \ ,
\label{avechi}
\end{equation}
which has the solution
\begin{eqnarray}
\overline{x}_{\pm} &=& -\frac{\theta}{\ln N}\pm\frac{\sqrt{(\ln N)^{2}-2\ln N + \theta^{2}}}{\ln N} \nonumber \\
& \approx & -\frac{\theta}{\ln N}\pm \sqrt{1-2/\ln N}
\label{zpm}
\end{eqnarray}
for $\theta^{2}/\ln^{2}N \ll 1$.
Converting back to rank, then
\begin{eqnarray}
\overline{r}_{\pm}
\approx \big(\frac{N}{2}\big)\big(1-\frac{\theta}{\ln N}\pm \sqrt{1-2/\ln N}\big) \ ,
\label{rminus}
\end{eqnarray}
and so the value $r^{*}\equiv \overline{r}_{-}$ is the special rank value which distinguishes the
set of excellent papers of each given author. The $c$-star value $c_{i}(r^{*})$ is thus a characteristic value arising from the special analytic properties of $c_{i}(r)$. This method for determining the
crossover value $r^{*}$ can be applied to any general Zipf profile which can be modeled by the DGBD.
Furthermore, the crossover $z_{\mathsf x }$ between the $\beta$ scaling regime and the $\gamma$ scaling regime is calculated from the inflection points of $\ln c(z)$,
\begin{eqnarray}
0 = \frac{d^{2} \ln c(z)}{dz^{2}}\vert_{z=z_{\mathsf x }}
= \frac{-\gamma}{(z_{0}-z_{\mathsf x })^{2}} + \frac{\beta}{(z_{0}+z_{\mathsf x })^{2}}
\end{eqnarray}
which has 2 solutions $z_{\mathsf x }^{\pm} = z_{0} \Big( \frac{1\pm \zeta}{1\mp \zeta}\Big)$ , where $\zeta \equiv \sqrt{\gamma/\beta}$. Only $\vert z_{\mathsf x }^{-} \vert <z_{0}$ is a physical solution. Transforming back to rank values, we find $r_{\mathsf x } = z_{0} + z_{\mathsf x }^{-} = z_{0} \frac{2}{ 1 + \zeta} = \frac{N+1}{ 1 + \zeta}$. We illustrate these special $z$ values in Fig. \ref{demo}.
\section{Acknowledgments}
We thank J. E. Hirsch and J. Tenenbaum for helpful suggestions.
|
\section{INTRODUCTION}
Manycore and accelerator-based programming models have become promising techniques in high-performance computing. Especially, novel uses of graphic processing units (GPU) with the parallel computing architecture CUDA \citep{NVIDIA2010} have revealed the potential of general-purpose GPU (GPGPU) computing. Moreover, nowadays the developments of GPU applications have moved beyond the single GPU stage, and both performance and parallel efficiency of the applications must be optimized. In astrophysical simulations, considerable performance speed-ups in multi-GPU systems have been demonstrated in a broad range of applications, for example, the direct $N$-body simulations \citep[e.g.,][]{Schive2008,Gaburov2009,Spurzem2011}, Barnes-Hut tree algorithm \citep{Hamada2009}, and reionization simulations \citep{AT2010}.
\citet{Schive2010a} present a parallel GPU-accelerated adaptive-mesh-refinement (AMR) code named \emph{GAMER} (GPU-accelerated Adaptive-MEsh-Refinement), which is dedicated to high-performance and high-resolution astrophysical simulations. The AMR implementation is based on constructing a hierarchy of grid patches with an oct-tree data structure, and the relaxing total variation diminishing scheme \citep[RTVD;][]{JX1995} with directional splitting is adopted in the hydrodynamic solver. A hybrid CPU/GPU model is adopted in the code, in which both hydrodynamic and gravity solvers are implemented into GPU and the AMR data structure is manipulated by CPU. An order of magnitude performance speed-up is demonstrated on the \emph{Laohu} GPU cluster at the High Performance Computing Center at National Astronomical Observatories of China (NAOC), using up to 128 Tesla C1060 GPUs \citep{Spurzem2011}.
Directionally unsplit algorithms have the advantage of maintaining spatial symmetry and have been extended to magnetohydrodynamic (MHD) simulations, where the divergence-free constraint is preserved \citep[e.g.,][]{Stone2008}. However, due to the requirement of 3D stencils, it was unclear whether these schemes could be implemented into GPU with high performance. Accordingly, in this work, we describe the extensions to the original GAMER code, including the implementation of several directionally unsplit Riemann-solver-based hydrodynamic schemes, and the hybrid MPI/OpenMP/GPU parallelization. Benchmarks using up to 32 Tesla C2050 GPUs on the \emph{Dirac} GPU cluster at the National Energy Research Scientific Computing Center at Lawrence Berkeley National Laboratory (NERSC/LBNL) are reported.
The structure of this paper is as follows. In Section 2, we introduce the numerical algorithms of the directionally unsplit hydrodynamic schemes implemented in the code. In Section 3, we describe the GPU implementation and the performance comparison between CPU and GPU hydrodynamic solvers. Several optimizations, especially the hybrid MPI/OpenMP/GPU parallelization, are described in Section 4. In Section 5, we present the overall performance speed-ups in both uniform-mesh simulations and AMR simulations, using up to 32 Fermi GPUs. Finally, we summarize the results in Section 6.
\section{NUMERICAL ALGORITHMS}
In this work, three directionally unsplit hydrodynamic algorithms are implemented into GAMER, including the MUSCL-Hancock method \citep[MHM; see][for an introduction]{Toro2009}, a variant of the MUSCL-Hancock method \citep[hereafter referred to as the VL scheme;][]{Falle1991}, and the corner-transport-upwind scheme \citep[CTU;][]{Colella1990}. The three-dimensional CTU scheme adopted in the code is a simplified version proposed by \citet{GS2008}, which requires only six solutions to the Riemann problem per cell per time-step. Also note that the VL and CTU schemes have also been implemented in the widely-adopted code \emph{Athena} \citep{Stone2008,SG2009}, and hence enables a direct comparison between GAMER and Athena in terms of both accuracy and performance.
Since the CTU scheme is more complicated than MHM and VL, in the following we highlight the main procedures in CTU for updating solutions by one time step $\triangle t$, and emphasize the major differences between CTU and the other two schemes. For a more comprehensive description of the CTU implementation, please see \citet{Stone2008}.
\renewcommand{\labelenumi}{(\arabic{enumi})}
\begin{enumerate}
\item Evaluate the left and right interface values for all cell interfaces in all three spatial directions by the 1D spatial data reconstruction. For CTU, an intermediate step advanced by the characteristics is also included to evaluate the interface values at $\triangle t/2$.
\item Evaluate the fluxes across all cell interfaces by solving the Riemann problem.
\item Correct the half-step cell interface values obtained in step (1) by computing the transverse flux gradients, and solve the Riemann problem with the corrected data to obtain the new fluxes across all cell interfaces. This step is only necessary in CTU.
\item Update solutions by $\triangle t$ by the conservative integration.
\item Store the fluxes across the boundaries of all grid patches. This step is only necessary in AMR simulations.
\end{enumerate}
The characteristic tracing step and the evaluation of the transverse flux gradients are not required in either MHM or VL. In MHM, the half-step solutions are obtained by first computing the fluxes from the reconstructed interface values in step (1) without Riemann solvers and then advancing the interface values by $\triangle t/2$ by the flux differences in all three directions. In VL, the half-step solutions are obtained by computing the first-order fluxes with Riemann solver prior to step (1).
For the data reconstruction in step (1), GAMER supports both piecewise linear method (PLM) and piecewise parabolic method (PPM). In general, the latter is less diffusive and can resolve the density cusps in self-gravity systems with higher spatial resolution. Several slope limiters are implemented in both methods, including the generalized minmod limiter, van Leer-type limiter, and van Albada-type limiter \citep[see][for an introduction]{Toro2009}. We also support the hybrid limiter adopted in Athena, which combines the generalized minmod and van Leer-type limiters. The spatial interpolation can be applied to either primitive variables or characteristic variables. Performing interpolation on characteristic variables can reduce the non-physical oscillations in some cases, for example, in the shock tube problems. However, in cosmological simulations where low-density voids will always form, this method is found to be less robust and can yield negative density more easily.
The code supports several Riemann solvers, including the exact solver based on \citet{Toro2009}, HLLE solver \citep{Einfeldt1991}, HLLC solver \citep[see][for an introduction]{Toro2009}, and Roe's solver \citep{Roe1981}. The exact solver is much more time-consuming compared with other solvers and is implemented primarily for obtaining reference solutions. The HLLC and Roe's solvers give comparable accuracy, while the HLLE solver is more diffusive in circumstances requiring the information of contact waves. However, \citet{Einfeldt1991} showed that the HLLE solver is positively conservative; that is, the density and pressure are positive-definite. He also showed that in certain initial data the linearization approximation adopted in Roe's solver will fail and lead to negative density or pressure in the intermediate region of the Riemann-problem solution. If this failure is detected during simulations, we follow the same strategy adopted in Athena, in which either HLLE, HLLC, or exact solver is used to calculate the fluxes at the cell interfaces where the Roe's solver fails.
\section{GPU COMPUTING}
In this section, we describe the strategy to map the directionally unsplit hydrodynamic schemes described in Section 2 into the CUDA parallel computing architecture. We also present the performance comparison between CPU and GPU solvers.
\subsection{CUDA Implementation}
In GAMER, the fluid data are always decomposed into grid patches, each of which consists of $8^3$ cells. In addition, due to the oct-tree data structure, we can always group the eight nearby patches into a single \emph{patch group} (which contains $16^3$ cells) before updating data, so that the additional workload to prepare the ghost-cell data can be reduced in comparison with a single patch. In CUDA implementation, each patch group will be computed by one thread block. We do not require to store all simulation data in the GPU global memory. Instead, we only need to ensure that the workload in each GPU is high enough to fully exploit its computing power. Accordingly, the number of patch groups sent into GPU at a time depends on hardware specifications rather than simulation scale, and the maximum simulation scale will not be limited by the relatively small memory in GPUs (4 GB in Tesla C1060 and 3 GB in Tesla C2050). Furthermore, since all simulation data still reside in the CPU memory, it reveals the potential of combining GPU computing with the out-of-core technique, by which the large storage space of multiple hard disks can be utilized as the additional virtual memory to significantly increase the total amount of memory available \citep{Schive2010b}.
The number of threads per thread block is a free parameter and the optimal value is also hardware-dependent. Typically, we use 128 threads per thread block in Tesla C1060 and 512 threads per thread block in Tesla C2050, in which cases each thread will compute the solutions of multiple cells. For a typical 3D loop in C language, we have
\begin{lstlisting}
// array format: Array3D[NZ][NY][NX]
for (unsigned int k=k_start; k<k_start+k_size; k++) {
for (unsigned int j=j_start; j<j_start+j_size; j++) {
for (unsigned int i=i_start; i<i_start+i_size; i++) {
Array3D[k][j][i] = ... ;
}}}
\end{lstlisting}
, which is converted to the following GPU kernel:
\begin{lstlisting}
// array format: Array1D[NZ*NY*NX]
unsigned int count = threadIdx.x;
unsigned int blocksize = blockDim.x;
unsigned int i, j, k, index;
while (count<i_size*j_size*k_size) {
i = i_start + coun
j = j_start + coun
k = k_start + count/(i_size*j_size);
index = (k*NY + j)*NX + i;
Array1D[index] = ... ;
count += blocksize;
}
\end{lstlisting}
Note that additional workload, mainly associated with the expensive integer division and modulo operations, is introduced into the GPU kernel. However, it will not have a large performance impact as long as the number of arithmetic operations performed on the 1D array is large enough. Furthermore, since this kind of conversion poses no constraints on the number of threads per thread block, it makes the code much more flexible and also makes the performance tuning in different generations of GPUs more easily.
In GPU computing, one of the most important keys to achieving high performance is to efficiently utilize the small shared memory (16 KB per multiprocessor in Tesla C1060 and 48 KB per multiprocessor in Tesla C2050) in order to hide the latency of accessing data from the global memory. For the directionally split schemes, taking advantage of the shared memory is more straightforward and efficient since the integration only requires 1D stencils. Therefore, the data can be decomposed into many small 1D segments to fit into the shared memory. Moreover, the number of ghost cells in both ends of the 1D data segment can be much smaller than the total number of cells in one segment, and hence the computational overhead resulted from this 1D data decomposition is small. The shared memory has been fully utilized in the directionally split hydrodynamic schemes and the GPU Poisson solver in GAMER \citep{Schive2010a}.
In comparison, for the directionally unsplit hydrodynamic schemes described in Section 2, some calculations are essentially 3D operations which require 3D stencils (e.g., the steps (3) and (4) in CTU), while some are still 1D operations (e.g., the step (1) in CTU). For MHM, the half-step prediction also requires a 3D stencil since the solutions are updated by taking account of the flux differences in all three directions in a single step. Utilizing the shared memory in these 3D operations is more tricky, and the most straightforward solution is to decompose the data into small 3D tiles to fit into the shared memory. However, it is arguable whether this method can significantly improve the overall performance, especially when we consider the relatively high surface/volume ratio in each small 3D tile and the high arithmetic intensity in the Riemann-solver-based schemes. In some preliminary tests we find that no substantial performance improvement is obtained by using the shared memory. Therefore, in the current implementation, we only use the global memory for the directionally unsplit hydrodynamic schemes. Despite that, considerable performance speed-up is still achieved and will be detailed in the next subsection. This approach makes the code more flexible to be run in GPUs with smaller shared memory (e.g., Tesla C1060), and it also makes it relatively straightforward to convert CPU hydrodynamic solvers to GPU solvers. Implementation of the shared memory in the directionally unsplit schemes will be investigated in the future. Also note that temporary global memory arrays are allocated to store the intermediate results during the integration (e.g., the fluxes and the left and right interface values). These arrays do not need to be transferred between host (CPU) memory and device (GPU) memory.
The integration procedure is carefully organized in order to minimize the performance impact of the high latency and relatively low bandwidth of the global memory access. For example, in the half-step prediction in MHM, the six interface values of one cell are all evaluated \emph{before} the thread proceeds to the next cell, so that the fluxes can be stored in the registers and the interface values can be updated without reloading data from the global memory. Another example is that the fluxes across the patch boundaries are stored immediately after solving the Riemann problem at these boundaries. By doing so, we do not need to perform additional memory copies in the global memory after all fluxes are evaluated, despite the fact that the latter method is more straightforward to implement.
In Fermi GPUs, the following tips are found to be able to fine-tune the performance. The compilation flag \emph{-Xptxas -dlcm=ca} is applied to enable both L1 and L2 caches. The flags \emph{-prec-div=false -ftz=true} are also used since they are found to improve the performance by 17\% without sacrificing the overall accuracy in most cases. During the initialization of CUDA devices, the cache configurations of the kernels of the directionally unsplit hydrodynamic solvers are set to \emph{cudaFuncCachePreferL1} via the function \emph{cudaFuncSetCacheConfig} in order to have a larger L1 cache. In a few cases, we find that the \emph{\underline{\ \ }forceinline\underline{\ \ }} function qualifier can enhance the performance, especially for computationally expensive functions (e.g., the Roe's Riemann solver). Additionally, it is observed that using structures instead of arrays to store the temporary 5-element fluid data delivers higher performance and also reduces the local memory usage in pre-Fermi GPUs.
Finally, we notice that many key components in different integration schemes are identical, for example, the data reconstruction, Riemann solver, full-step update, and storing fluxes. Therefore, in GAMER these functions are developed with flexible arguments so that they can be efficiently shared among different schemes. It makes the code more maintainable and extensible.
\subsection{GPU Performance}
All the tests in this work are performed on Tesla C2050 GPUs and Intel Xeon E5530 CPUs, except for the tests aiming to compare the performances between different hardware configurations. The CPU codes are compiled with \emph{gcc v4.4.2} and the GPU codes are compiled with \emph{nvcc v3.2}. The \emph{-O3} optimization flag is adopted in both CPU and GPU tests. Single precision variables are adopted in all performance experiments. Additionally, in order to have a fair comparison, the timing measurements of GPU always include the time to copy data between host and device memory.
Figure \ref{fig:Performance_GPUSolver} shows the performance comparisons between one GPU and one CPU core as a function of the number of cells sent into GPU at a time. The top left panel shows the speed-ups of different hydrodynamic schemes. For comparison, we also include the results of two directionally split schemes: the RTVD scheme and the weighted average flux scheme \citep[WAF; see][for an introduction]{Toro2009}. Note that in the directionally split schemes we utilize the fast shared memory, while in the unsplit schemes we only use the global memory. Nevertheless, the latter still achieve substantially higher performance speed-ups due to their higher arithmetic intensity. Maximum speed-ups of 114, 122, and 111 are demonstrated in MHM, VL, and CTU, respectively. Also note that both VL and CTU require six Riemann solvers per cell per time-step, while MHM only requires three. Therefore, the MHM scheme is measured to be about 1.5 times faster than the other two schemes.
The \emph{CUDA stream} can be used to overlap the memory copy between host and device memory with the kernel execution, and, in general, using more CUDA streams can improve the efficiency of overlapping and hence deliver higher performance. The performances of the CTU scheme with different numbers of CUDA streams are given in the bottom left panel in Figure \ref{fig:Performance_GPUSolver}. The performance is improved by 26\% by using four CUDA streams as compared to the result using only one CUDA stream. It is also found that using more than four CUDA streams does not further improve the overall performance. Therefore, throughout the tests in this work, four CUDA streams are always adopted unless the number is explicitly specified. The method to implement the CUDA stream in GAMER is described in Section 4.1.
The top right panel in Figure \ref{fig:Performance_GPUSolver} compares the performances of the CTU scheme with different data reconstruction methods (PLM and PPM) and different Riemann solvers (HLLE, HLLC, and Roe's solvers). The highest speed-ups achieved in different methods range between 96 and 111, which do not differ significantly. In general, the PPM data reconstruction gives higher speed-ups than the PLM data reconstruction, and the Roe's solver delivers higher speed-ups than the HLLE and HLLC solvers. In addition, we find that the optimal performance is achieved when we have one to two patch groups per multiprocessor, in which case the CTU scheme with the PPM data reconstruction only requires about 110 MB to 220 MB memory in Tesla C2050. Having too many patch groups per multiprocessor can deteriorate the performance.
In order to have more comprehensive performance measurements, we also compare the timing results conducted in three different GPU systems: the \emph{Dirac} GPU cluster at NERSC/LBNL, the \emph{Laohu} GPU cluster at NAOC, and the GPU cluster at the National Center for High-performance
Computing of Taiwan (hereafter referred to as NCHC). The Dirac system contains 48 nodes connected by QDR InfiniBand. Each node is equipped with two Intel Xeon E5530 CPUs and one NVIDIA Tesla C2050 GPU. The Laohu system has 85 nodes, each with two Intel Xeon E5520 CPUs and two NVIDIA Tesla C1060 GPUs. The experimental GPU system in NCHC has two Intel Xeon X5670 CPUs and four NVIDIA Tesla M2070 GPUs in each node. The measured speed-ups in different systems are presented in the bottom right panel in Figure \ref{fig:Performance_GPUSolver}. The maximum number of cell updates per second achieved by one Fermi GPU is $3.0\times 10^{7}$, which is about 1.72 times faster than the performance achieved by on Tesla C1060 GPU.
Note that the timing experiments presented above only measure the performance of the hydrodynamic solver. In other words, they do not include the elapsed time of all other operations in AMR simulations. Therefore, the values in Figure \ref{fig:Performance_GPUSolver} can be considered as the optimal overall speed-ups we can achieve in any AMR simulation using GAMER. Also note that so far the performance comparisons are all based on one GPU versus one CPU core. The overall performance speed-up with multiple GPUs and CPU cores is presented in Section 5.
\section{OPTIMIZATION}
In this section, we describe several performance optimizations applied \emph{outside} the GPU kernel. Since GAMER adopts a hybrid CPU/GPU implementation, these optimizations have been demonstrated to be extremely important for achieving higher overall performance.
\subsection{Asynchronous Memory Copy}
As already mentioned in the previous section, the memory copy between host and device memory can be performed concurrently with the kernel execution by managing the CUDA stream. Specifically, the kernel launch and memory copy associated with different CUDA streams can be performed in parallel, while the operations associated with the same CUDA stream will be performed sequentially. The following code shows an example using $ns$ CUDA streams and assigning $np$ patch groups to each stream.
\begin{lstlisting}
// input array format: float HostArray_In [ns*np][SizePerPatchGroup_In]
// output array format: float HostArray_Out[ns*np][SizePerPatchGroup_Out]
const unsigned int MemSize_In = np*SizePerPatchGroup_In *sizeof(float);
const unsigned int MemSize_Out = np*SizePerPatchGroup_Out*sizeof(float);
for (unsigned int s=0; s<ns; s++) {
cudaMemcpyAsync( DeviceArray_In + s*np, HostArray_In + s*np, MemSize_In,
cudaMemcpyHostToDevice, Stream[s] );
Kernel <<< np, 512, 0, Stream[s] >>> ( DeviceArray_In + s*np, DeviceArray_Out + s*np, ... );
cudaMemcpyAsync( HostArray_Out + s*np, DeviceArray_Out + s*np, MemSize_Out,
cudaMemcpyDeviceToHost, Stream[s] );
}
\end{lstlisting}
\subsection{Concurrent Execution between CPU and GPU}
In GAMER, before invoking the GPU solver, we need to prepare the input array in CPU, which stores the interior and ghost-cell data of the patch groups to be calculated. This step is referred as the \emph{preparation step}. In addition, after receiving the updated data from GPU, we need to copy the data of different patches to their corresponding arrays in CPU. This step is referred as the \emph{closing step}. Experiments show that the preparation step can be very expensive, especially in AMR simulations where spatial and temporal interpolations are required to obtain the ghost-cell data if the targeted patches are adjacent to the coarse-fine boundaries. This step is the performance bottleneck in the previous version of GAMER and can take up to 3 times longer than the GPU computation.
To alleviate this issue, we notice that in principle the preparation and closing steps and the GPU solver can be performed concurrently, provided that they are targeting different patches. Therefore, by taking advantage of the fact that the GPU solver is an asynchronous function, we can overlap the executions of the preparation and closing steps in CPU with the GPU computation. The implementation and efficiency of this optimization are described in more detail in \citet{Schive2010a}.
\subsection{Hybrid MPI/OpenMP/GPU Parallelization}
The optimization of making CPU and GPU work in parallel greatly improves the overall performance, especially in the cases where the computation times of CPU and GPU are comparable (e.g., when relatively low-end GPUs are adopted). However, in both Dirac and Laohu systems, it does not completely eliminate the performance bottleneck described above due to the fact that the preparation step in CPU can take significantly longer than the GPU hydrodynamic solver. As a result, the overall performance speed-up is still much lower than the optimal value given in Section 3.2. To solve this issue, in this work we further implement the OpenMP parallelization in CPU computation.
In most GPU applications, one CPU core can take charge of one GPU, and the multi-GPU parallelization is achieved using MPI \citep[e.g.,][]{Schive2008,Schive2010a}. Accordingly, in the GPU clusters with more CPU cores than GPUs, some CPU cores will be idle during simulations. It is not an issue for the kind of GPU applications where the overall speed-up is dominated by the GPU performance. However, for complicated programs such as GAMER, the overall performance is determined by the performance of the hybrid CPU/GPU computing, and therefore it is crucial to fully exploit the computing power of both the multi-core CPUs and GPUs. To this end, we have implemented the hybrid MPI/OpenMP/GPU parallelization in GAMER. Each MPI process is responsible for one GPU, and the CPU computation allocated to each MPI process is further parallelized with OpenMP. For example, in a GPU cluster with $N_{node}$ nodes and each node is equipped with $N_{core}$ CPU cores and $N_{GPU}$ GPUs, we can run the simulation with $N_{node}\times N_{GPU}$ MPI processes and launch $N_{core}/N_{GPU}$ OpenMP threads in each process.
In GAMER, the AMR implementation is realized by constructing a hierarchy of grid patches, and different patches can be evaluated in parallel. Accordingly, the MPI implementation is based on the rectangular domain decomposition, and the OpenMP implementation is based on the patch-level parallelization. OpenMP is applied to most CPU computations, including both the preparation and closing steps, correcting the coarse-grid data, calculating the evolution time-step, and the grid refinement. Since different patches contain the same number of cells, the computation workload associated with each patch is approximately the same. Therefore, in most cases, OpenMP can be implemented straightforwardly and high parallel efficiency can be achieved.
\section{OVERALL PERFORMANCE}
In this section, we present the overall performance comparisons between CPUs and GPUs in both uniform-mesh simulations and AMR simulations. In each case, we compare the achieved speed-ups with different optimization levels described in Section 4, and show results for both single-GPU and multi-GPU. The CTU scheme with the PPM data reconstruction and Roe's Riemann solver is adopted in all tests presented in this section. Performance is measured on the Dirac system.
\subsection{Uniform-mesh Performance}
Figure \ref{fig:Performance_GAMER_vs_Athena} shows the overall performance speed-up in the uniform-mesh 3D blast wave test as a function of the spatial resolution ($N$). First, we verify that the performance of GAMER without GPU acceleration and OpenMP is as fast as Athena. The performance difference is less than 6\% when $N\ge 32^3$. This result is very important as it makes the performance comparison between CPU and GPU in GAMER more convincing. The CPU-only performance scales linearly with the number of OpenMP threads when there are at least one patch groups per thread. For the performance with GPU acceleration, results with different optimization levels are shown together for comparison. The unoptimized code still achieves a speed-up of 35, and factors of 39, 63, and 101 speed-ups are further demonstrated when different optimizations described in Section 4 are implemented successively. This is a very encouraging result since the speed-up achieved by the fully optimized code closely approaches the optimal value given in Section 3.2. The maximum number of cell updates per second is $2.9\times 10^{7}$, which is 25 times faster than the CPU-only performance using four cores. We also compare the accuracies of physical results obtained separately by GAMER and Athena, and the relative differences are on the order of the machine precision in both single-precision and double-precision experiments.
Another important feature in GAMER is that it is very memory-efficient as compared with Athena; this is expected due to the patch-based decomposition employed in the code. For the directionally unsplit hydrodynamic schemes described in Section 2, usually we need to store the left and right interface values and fluxes at \emph{all} cell interfaces, which can be very memory-consuming. However, in GAMER we only need to store these data for the \emph{patches being computed in parallel}, the number of which is generally much smaller than the total number of patches. Experiments show that for the adiabatic hydrodynamic simulations with $N=400^3$, Athena consumes about 13 GB of memory, while GAMER only consumes roughly 3 GB of memory.
Figure \ref{fig:Scalability_NoAMR} shows the parallel scalability of GAMER in uniform-mesh simulations. The result is excellent for weak scaling, in which we let each GPU compute $512^3$ cells. The 32-GPU run is 31.6 times faster than the single-GPU performance, which corresponds to a parallel efficiency of 98.8\%. Strong scaling is much more challenging. By fixing the simulation resolution to $1024\times 1024\times 512$, the 32-GPU run still achieves a factor of 27.2 speed-up over the single-GPU performance, giving a parallel efficiency of 85.0\%.
\subsection{AMR Performance}
To demonstrate the performance in the AMR simulations with sufficiently complicated grid structure, we follow the similar timing experiments performed in \citet{Schive2010a} and measure the performance in purely baryonic cosmological simulations. The root-level resolution is $256^3$ and up to four refinement levels are used, giving $4096^3$ effective resolution. Note that the self-gravity is included here, and a GPU Poisson solver based on the successive overrelaxation method \citep[SOR; see][for an introduction]{Press2007} is adopted for the refinement levels. The GPU SOR solver alone is measured to be 84 times faster than the CPU counterpart on the Dirac system.
Figure \ref{fig:Performance_AMR_SingleGPU} shows the overall performance speed-up at redshift $z=2$ using a single GPU. The fully optimized GPU code demonstrates speed-ups of 84.0 and 21.6 as compared with the CPU-only single-core and quad-core performances, respectively. The maximum speed-up is slightly lower than the value obtained in the uniform-mesh simulations, which is mainly due to the relatively lower speed-up ratio achieved by the GPU Poisson solver. We emphasize that the fully optimized performance (inverted triangles) is 2.3 times faster than the partially optimized performance without OpenMP (triangles), which demonstrates the importance of adopting the hybrid MPI/OpenMP parallelization. We also find that using more than four OpenMP threads does not further improve the overall performance, which is expected since the overall performance will be limited by the GPU performance when more than four OpenMP threads are used and the concurrent execution between CPU and GPU is enabled. This fact reveals the plausibility of installing more than one high-end GPUs in each multi-core computing node, such as the hardware configuration adopted on the Laohu system.
Figure \ref{fig:Performance_AMR_MultiGPU} shows the overall performance speed-up using multiple GPUs as a function of the number of MPI ranks. For example, the 8-GPU performance is compared to the CPU-only performance using 8 cores, each of which resides on a different computing node. In the 32-GPU test, the maximum speed-ups are 71.4 and 18.3 as compared with the CPU-only single-core and quad-core performances, respectively. The performance decrement results from the increasing MPI communication time, which takes about 11\% of the total execution time in the 32-GPU test. This issue was found to be of minor importance in the simulations with higher spatial resolution (and hence lower surface/volume ratio). For example, for the 128-GPU benchmark on the Laohu GPU cluster at NAOC, the MPI communication time takes less than 2\% of the total execution time \citep{Spurzem2011}. Also note that this issue can potentially be largely alleviated by overlapping communication with computation. Finally, we point out that, in the current implementation, the load is unbalanced among different GPUs in the AMR simulations due to the rectangular domain decomposition. This issue will be addressed elsewhere.
\section{SUMMARY}
We have introduced the directionally unsplit hydrodynamic schemes newly implemented in GAMER, including the MUSCL-Hancock method, a variant of the MUSCL-Hancock method, and the corner-transport-upwind scheme. In each scheme, we support different data reconstruction methods (PLM and PPM), different Riemann solvers (HLLE, HLLC, and Roe's solvers), and also different slope limiters. All schemes have been implemented in GPU using NVIDIA CUDA, and up to two orders of magnitude performance speed-up has been demonstrated as compared to the performance using a single CPU core.
Several optimizations have been implemented in the code, including the asynchronous memory copy, the concurrent execution between CPU and GPU, and the hybrid MPI/OpenMP/GPU parallelization, by which we can fully exploit the computing power in a heterogeneous CPU/GPU system. OpenMP has been shown to be able to eliminate the performance bottleneck in the previous version of GAMER, and hence considerably improve the overall performance.
We have presented the overall performances of both uniform-mesh simulations and AMR simulations, measured on the Dirac cluster at NERSC/LBNL. In uniform-mesh tests, single GPU achieves a performance of $2.9\times 10^{7}$ cell updates per second, which is 101 times faster than the performance using a single CPU core and 25 times faster than the quad-core performance. We also directly compare GAMER with the well-known code Athena in adiabatic hydrodynamic tests, in which two orders of magnitude performance speed-up is also demonstrated. Weak scaling with 98.8\% parallel efficiency and strong scaling with 85.0\% parallel efficiency are achieved in the 32-GPU experiments. In AMR tests, 32-GPU run achieves speed-ups of 71.4 and 18.3 as compared to the performances using 32 and 128 CPU cores, respectively.
In Athena, both the VL and CTU schemes have been extended to MHD simulations \citep{GS2008,Stone2008,SG2009}. Following their work, we have performed some preliminary tests on a MHD solver with the CTU scheme and constrained transport technique, and achieved a factor of 60 speed-up when compared with a single CPU core. Details of the implementation of MHD in GAMER will be reported in a future communication.
\section{ACKNOWLEDGEMENTS}
A substantial part of simulations presented in this work were performed on the Dirac GPU cluster at the National Energy Research Scientific Computing Center at Lawrence Berkeley National Laboratory (NERSC/LBNL). We would like to thank Hemant Shukla, John Shalf, and Horst Simon in the International Center for Computational Science (ICCS) for providing the access to this system. The special supercomputer Laohu at the High Performance Computing Center at National Astronomical Observatories of China, funded by Ministry of Finance under the grant ZDYZ2008-2, has also been used. We want to thank Rainer Spurzem, Peter Berczik, and Gao Wei for helping conduct simulations on this system. Simulations were also performed on the National Center for High-Performance Computing of Taiwan. Finally, we are grateful to Evghenii Gaburov for insightful suggestions and sharing the source code. This work is supported in part by the National Science Council of Taiwan under the grant NSC97-2628-M-002-008-MY3.
.
|
\section{Introduction}
While data on large scale structures point towards a Universe dominated by dark matter and dark energy, e.g. \citet{komatsu11}, the nature of these is still a deep mystery \citep[e.g.,][]{frieman08,wiltshire08,bertone10,kroupa10}. In this context, it is good to keep in mind that this conclusion essentially relies on the
assumption that gravity is correctly described by Einstein's General Relativity in the extreme weak-field limit, a regime where the need for dark matter itself prevents the theory from being tested. Until this double dark mystery is solved, it is thus worth investigating alternative paradigms and their implications.
For instance, Modified Newtonian dynamics \citep[MOND]{milgrom83} naturally explains various {\it spiral} galaxy scaling relations \citep{tully77, mcgaugh00, mcgaugh04}. The existence of a very tight baryonic Tully-Fisher relation for disk galaxies \citep{mcgaugh05, trachternach09} is for instance one of the remarkable predictions of MOND. The corresponding relation for early-type galaxies is much more
difficult to investigate because they are pressure-supported systems, and the equivalent circular velocity curves determined from the velocity dispersion profiles suffer from the well-known degeneracy with anisotropy. However, some studies circumvented this problem: for instance, \citet{krona00} used data on 21 elliptical galaxies to construct non-parametric models from which circular velocity curves, radial profiles of mass-to-light ratio, and anisotropy profiles as well as high-order moments could be computed. This led \citet[hereafter G01]{gerhard01} to publish benchmark scaling relations for ellipticals. It was e.g. shown for the first time that circular velocity curves tend to become flat at much larger accelerations than in spiral galaxies. This would seem to contradict the MOND prescription, for which flat circular velocities typically occur well below the acceleration threshold $a_0 \sim 10^{-8}$~cm~s$^{-2}$, but not at accelerations of the order of a few times $a_0$ as in ellipticals. Also \citet[hereafter T09]{thomas09} published scaling relations for dark matter halos of 18 Coma galaxies, using similar prescriptions
as G01. We remark that G01 employed spherical models while the models of T09 are axisymmetric.
Not many studies have considered the predictions of MOND in elliptical galaxies. \citet{milgrom84} showed that pressure-supported isothermal systems have finite mass in MOND with the density at large radii falling approximately as $r^{-4}$. It was also shown that there exists a mass-velocity dispersion
relation of the form $(M/10^{11}M_\odot) \approx (\sigma_r/100\,\,
{\rm kms}^{-1})^4$ which is similar to the observed Faber-Jackson relation \citep{sanders2000, sanders2010}, and that, in order to match the fundamental plane, MOND models must deviate from being strictly isothermal and isotropic: a radial orbit anisotropy in the outer regions is needed \citep{sanders2000,cardone11}. \citet{tiret07} and \citet{angus08} also analyzed the distribution of velocity dispersion of PNe on scales of 20 kpc, and of satellites on very large scales of the order of 400~kpc around red isolated ellipticals, showing that MOND allowed to fit both scales successfully.
Hereafter, we make general remarks on the properties of spherical galaxies within MOND, and their scaling relations. We first point out a remarkable property of elliptical galaxies exhibiting a flattening of their circular velocity curve at small radii: such a flattening in the intermediate gravity regime is actually generated by a baryonic density distribution following a Jaffe profile in these parts of the galaxies. We then further show that the observational scaling relations for the dark halos of the elliptical galaxy sample by G01 are strikingly similar to the theoretical ``phantom'' halos of MOND (i.e. the halo that would produce in Newtonian gravity the same additional gravity as MOND), with one remarkable exception: MOND predicts that the product of the central density with the core radius should be constant, as recently observed for spiral galaxies \citep{donato09, gentile09}.
\section{Flat circular velocity curves and the Jaffe profile}
Although it has been argued that some ellipticals do not need any dark matter or enhancement of gravity \citep{aaron03}, there are many counter-examples \citep{magorrian, richtler04, schuberth06, kumar07}. Such a recent example is the elliptical galaxy NGC 2974 where the presence of an HI disk allowed a more or less direct measurement of circular velocities \citep{weijmans08}. There is also evidence that elliptical galaxies exhibit flat circular velocity curves , but that, contrary to spiral galaxies, this happens in the {\it inner regions} where $g>a_0$ (e.g., G01, Weijmans et al. 2008). Such a flattening of circular velocities is {\it a priori} not expected in the strong to intermediate gravity regime in MOND, and poses the question of how to analytically interpret it.
In the intermediate gravity regime, the transition from Newtonian to MONDian dynamics is described by the $\mu$-function of MOND. Many concordant studies have recently shown that, in spiral galaxies, the ``simple'' transition of \citet{fb05} is a good representation of the data \citep*[for an extensive discussion]{gentile11}. In a spherical system, with this simple transition, the enclosed (baryonic) mass $M_M(r)$ needed to produce the same gravitational potential in MOND as the (baryonic+dark) mass $M_N(r)$ in Newtonian gravity is:
\begin{equation}
M_M (r) = M_N(r) - \left( \frac{1}{M_N(r)} + \frac{G}{r^2a_0} \right)^{-1}.
\end{equation}
In a region where the circular velocity is constant $v_c=V$ (even if $g>a_0$), one can write $M_N(r)= V^2r/G$, and thus after some algebra
\begin{equation}
M_M(r) = \frac{V^4}{a_0 G} \cdot \frac{r}{r+V^2/a_0}.
\end{equation}
Remarkably, this enclosed mass profile corresponds precisely to a Jaffe profile \citep{jaffe83} with scale-radius $r_j=V^2/a_0$ (meaning that the acceleration is $a_0$ at $r_j$),
and with total mass
$M_{\rm tot}=V^4/(a_0 G)$.
Indeed, as the enclosed mass $M_M(r) = M_M(r_0) + 4 \pi \int_{r_0}^r{\rho(R) R^2 dR}$, this enclosed mass profile corresponds locally to the density profile:
\begin{equation}
\rho(r) = \frac{M_{\rm tot}}{4 \pi} \cdot \frac{r_j}{r^2~ (r+r_j)^2},
\end{equation}
with the characteristic surface density (see also Milgrom 1984) $M_{\rm tot}/r_j^2=a_0/G$. This profile is of course not valid for the very inner parts of an elliptical galaxy, where $V$ is not constant. Let us also note that (i) it was already known that a Jaffe profile produces a flat circular velocity curve at $r \ll r_j$ in Newtonian gravity, which MOND generalizes to radii $r \sim r_j$; (ii) $M_{\rm tot}$ does not necessarily have to be the real total mass of the galaxy, as the Jaffe profile fit to the density distribution could have a cut-off in the outer parts. In that case, the constant circular velocity $V$ would actually fall slightly above the prediction from the baryonic Tully-Fisher relation of spiral galaxies.
Interestingly, this is precisely what is observed for the G01 sample
of ellipticals.
The fact that elliptical galaxies can exhibit (equivalent) circular velocity curves that are flat in the intermediate gravity regime is thus analytically understood in MOND by the fact that the outer regions of ellipticals can be approximated by a Jaffe profile with a large scale-radius, i.e. in regions well within the intermediate gravity regime rather than in the deep-MOND regime. These flat circular velocity curves would have been impossible with exponential density profiles (as encountered in
spiral galaxies), meaning that the fact that circular velocity curves become flat quicker in ellipticals does not come as a surprise in the context of MOND.
This finding looks like an interesting possibility to devise new tests of MOND based on photometry. However,
in reality it might be difficult: not many spherical galaxies with a precisely measured density profile are
dynamically investigated out to large radii, and have enough tracers to measure the higher order moments and constrain the anisotropy. Moreover, light might not trace the baryonic mass precisely. As an example, the circular velocity in NGC 2974, which can be traced by an HI disk, becomes constant at around 5 kpc and
has the value 300~${\rm kms}^{-1}$, which would correspond to a Jaffe scale radius of 23 kpc.
Unfortunately, NGC 2974 is neither spherical nor does its photometry reach
large radii so that it does not serve well as a test object. In any case, Weijmans et al.~(2008, their Fig.~20) showed that the reverse procedure (going from the density to the circular velocity curve) leads to a very good fit.
\section{Dark matter scaling relations for phantom halos of ellipticals}
\begin{table*}[t]
\caption{The table shows for baryonic Hernquist profiles with mass and effective radius described by the first two columns, the corresponding parameters of the MONDian ``phantom'' halo
represented by a logarithmic potential fitted from the center to two effective radii of the baryonic profile. The columns are the baryonic mass, the luminous effective radius, the core radius $r_0$ and the asymptotic velocity $v_0$ of the
log-halo, its surface density, central density and phase space density, the latter as defined by G01.The last column column gives the acceleration in units of
$a_0$ at a radius of 2 $R_{eff}$ for each Hernquist model. The predictions for MONDian halos in the previous columns are valid
only for galaxies embedded in an external field smaller than this
value.}
\begin{center}
\begin{tabular}{cccccccc}
baryonic mass $[M_\odot]$ &$ R_{eff}$ [kpc] & $r_0$ [kpc] & $v_0$[${\rm kms}^{-1}$] & $S$$[M_\odot/pc^2]$ & $\rho_0$$ [M_\odot/pc^3]$ & $f_{ps}$& acc.[$a_0$]\\
\hline
$10^{12}$ & 14.1 & 8.83 & 244 & 374 & 0.04 & 8.82$\times$ $10^{-9}$ & 1.53\\
$8\times10^{11}$ & 11.8 & 7.60 & 228 & 379 & 0.05 & $1.19\times 10^{-8}$ & 1.55 \\
$5\times 10^{11}$ & 8.06 & 5.20 & 193 & 393 & 0.076 & $ 3.02\times 10^{-8}$ & 1.91 \\
$2\times10^{11}$ & 3.84 & 2.87 & 146 & 411 & 0.143 & $1.3\times 10^{-7}$ & 2.96 \\
$10^{11}$ & 1.47 & 1.24 & 99 & 438 & 0.35 & $1.02\times 10^{-6}$ & 8.44\\
$5 \times10^{10}$ & 1.25 & 1.04 & 90 & 431 & 0.41 &$1.6\times 10^{-6}$ & 6.08 \\
\hline
\end{tabular}
\end{center}
\label{surfdens}
\end{table*}
We now apply the reverse procedure, and check whether the phantom halos predicted by the simple transition of MOND (Famaey \& Binney 2005) comply with the observational scaling relations of dark halos of ellipticals. As stated above, Jaffe profiles are not good descriptions of the very inner parts of ellipticals. We hereafter rather choose Hernquist profiles \citep{hernquist90} to represent the baryonic content of ellipticals: these are realistic enough and allow for an exhaustive exploration of their properties without varying too many free parameters. Such a Hernquist-model is described by its total mass $M$ and scale-radius $r_{H}$.
The profile of the Newtonian circular velocity curve then reads
\begin{equation}
v_N(r) = \sqrt{\frac{G M r}{(r+r_H)^2}}
\end{equation}
where
the scale-radius $r_H$ of the Hernquist-model is related to the effective (half-light) radius by $ R_{eff} = 1.815 \, r_H$.
Adopting the simple transition formula between the Newtonian and the MONDian regime, one
finds for the MOND circular velocity
\begin{equation}
v_{M} = \sqrt{v_N^2(r)/2 + \sqrt{v_N^4(r)/4 + v_N^2(r) a_0 r}},
\end{equation}
and the MONDian phantom halo has the circular velocity
\begin{equation}
v_{\rm phantom} = \sqrt{\sqrt{v_N^4(r)/4 + v_N^2(r) a_0 r}-v_N^2(r)/2}
\end{equation}
To enable the comparison with the scaling relations of G01, where the dark matter halos are adopted as logarithmic halos, we fit $v_{\rm phantom}(r)$ to the circular velocity $v_{\rm log}(r)$ of a logarithmic halo with asymptotic circular velocity $v_0$ and core radius $r_0$:
\begin{equation}
v_{\rm log}(r) = v_0 r/\sqrt{r_0^2+r^2}.
\end{equation}
The fits are performed within the inner two effective radii \footnote{Let us note that these fits are not particularly good: the circular velocity curve $v_c(r) = v_0 r/(r_0+r)$ would have provided better
fits, but the core radius of the corresponding halo would then be systematically smaller with respect to G01}. The fitted central dark matter density is then given by
\begin{equation}
\rho_0 = 3 (v_0/r_0)^2/(4 \pi G).
\end{equation}
The characteristic central phase space density is defined (see G01) as
\begin{equation}
f_{ps} = 3^{3/2} \rho_0/v_0^3.
\end{equation}
The characteristic surface density within $r_0$ is then also defined as (see also Donato et al. 2009):
\begin{equation}
S = \rho_0 \cdot r_0.
\end{equation}
Table \ref{surfdens} lists these fitted parameters for six baryonic Hernquist masses over a large mass range. The combinations of the masses and effective radii in Table \ref{surfdens} follow equation (5) of G01 (in accordance with the fundamental plane), where we transformed their luminosities into masses by using $M/L_B$=8 for all galaxy baryonic masses.
\begin{figure}[]
\begin{center}
\includegraphics[width=0.4\textwidth]{gerhard_thomas.pdf}
\caption{The figure shows the surface density, central density, and the central phase space density logarithmically (see the text for
more explanations) of the phantom dark halos (circles) for different baryonic Hernquist masses from Table \ref{surfdens} together with the relations given by G01 (dotted lines) and T09 (dashed lines). Note that these parameters observationally exhibit a very large scatter around the mean relations. Note also that the relations of T09 are not given explicitly in their paper but have been constructed from their Table 3 omitting
galaxies with young stellar cores. }
\label{fig:gerhard}
\end{center}
\end{figure}
Fig.\ref{fig:gerhard} shows the values of the fitted dark halo parameters derived from applying MOND to the baryonic Hernquist-profiles, together with the observational scaling relations given by G01 (dotted lines) and T09 (dashed lines). The upper panel shows the characteristic phase space density, the middle panel the central volume density, and the lower panel the characteristic surface density.
Let us note that the plotted relations are indicative only, since the data (Fig.18 of G01 and Figs. 1 and 4 of T09) show a very large scatter even when logarithmically displayed. However, within this observational uncertainty, it is remarkable that some features are perfectly reproduced, particularly the slopes of the phase space density and of the central volume density as a function of baryonic mass (given the observational scatter, the almost perfect reproduction of the central volume density of G01 might of course be {\it partly} coincidental).
As first emphasized by G01, the phase-space density values are at a given mass higher than in spirals, which means that under the $\Lambda$ Cold Dark Matter paradigm, dark halos of ellipticals cannot be the result of collisionless mergers of present-day spirals, but must have been assembled at a very early time, when the cosmological density was higher. In MOND this is of course not necessarily the case, as the phase-space argument does not apply to phantom halos.
One also notes a remarkable exception to the scaling relations: the fitted characteristic dark matter surface density $S$ is fully independent from the Hernquist parameters, and it is systematically lower than in G01 and T09. We emphasize that this constancy is not related to the special relation of mass and effective radius. Varying $R_{eff}$ by a factor of two at a given mass does not change the constant surface density significantly. This prediction of MOND thus brings the value closer to the (also constant) value of $S$ observed in spiral galaxies, ${\rm log}S=2.1$ \citep{donato09}. Let us note that MOND also predicts the observed constant value of $S$ in spirals, which is somewhat lower because (i) spirals are a bit deeper into the MOND regime (\citealt{milgrom09}) and (ii) their flattened baryonic profiles lead to a somewhat higher Newtonian gravity at a given mass, and in turn a somewhat lower MOND contribution to the phantom halo.
On the first glance one might interpret this constancy and the other scaling relations as a clear signature of MOND in
ellipticals: however, CDM may also predict that the surface density within the scale radius
of NFW halos weakly depends on dark matter total mass \citep{boyarsky10}.
For spiral galaxies, this is of little interest as it is known that cuspy profiles often do not fit rotation curves \citep{fb05, deblok10, gentile05}, the mystery then being how to erase the cusp by the feedback from the baryons while keeping the product $ \rho_0 r_0$ constant. In elliptical galaxies, the situation is less clear as NFW profiles often do fit the data equally well as cored profiles \citep{schuberth10}. We thus fitted NFW profiles to the same MONDian phantom halos
and found a perfect agreement. The question remains whether these NFW-halos are ''cosmological'' or in other words, fulfill the relation between virial mass and concentration predicted by cosmological simulations. Fig.\ref{fig:M_C} displays for our Hernquist masses the resulting concentrations of the NFW-halos (open circles) corresponding to the MONDian phantom halos, while the triangles show the concentration values expected from the equation (9) of \citet{maccio08}, using 200 times the critical density as the mean density within the virial radius (standard cosmology: h=0.7, $\Omega_m = 0.3$, $\Omega_\Lambda=0.7$).
One concludes that for high masses the MONDian phantom halos are not distinguishable from cosmological NFW halos, given also that the simulations predict considerable scatter. For smaller masses the difference between MONDian phantom halos and NFW
cosmological halos is larger.
\begin{figure}[]
\begin{center}
\includegraphics[width=0.4\textwidth]{M_C.pdf}
\caption{The figure shows for our 6 Hernquist masses the concentration parameters of the associated NFW-halos, if the MONDian phantom halos
are fitted by NFW profiles (open circles). The triangles are the concentration parameters expected from the relation quoted by \citet{maccio08}}
\label{fig:M_C}
\end{center}
\end{figure}
\section{External field effect}
Due to the non-linearity of MOND and its associated breaking of the
Strong Equivalence Principle, a MONDian stellar system embedded in an
external gravitational field (EF) stronger than its own internal field
behaves in a quasi-Newtonian way, with an effectively higher
gravitational constant
\citep{milgrom83,famaey07}. Most of the sample galaxies are located in clusters or groups where the EF might
have an influence. \citet{wu10} for instance showed how the EF can lead to the
lopsidedness of an originally axisymmetric non-isolated galaxy.
While it is beyond the scope of this research note to evaluate in detail the EF in the present sample, a very rough estimation is presented in Fig. \ref{EF}, which plots for the
Virgo and the Coma cluster the accelerations based on the extrapolations of the mass models cited in the figure caption (these extrapolations are only meant to give an order of magnitude
estimate, but should not be taken as rigorous models). This can be compared with the internal accelerations at 2 $R_{eff}$ for
the Hernquist models in Tab.\ref{surfdens}. Indicated are the projected distances of galaxies in the Virgo and
Coma region.
The positions of the Virgo galaxies correspond to the middle points of their NGC numbers, while the Coma
galaxies are plotted as small open circles.
One concludes that the EF should have no influence in the two samples at the galactocentric distances which we consider.
\begin{figure}[]
\begin{center}
\includegraphics[width=0.4\textwidth]{EF_thomas.pdf}
\caption{This plot estimates the external field acting on Virgo (dashed) and Coma (dotted) galaxies. Abscissa is the projected distance in Mpc from M87 and
NGC 4874, respectively. Ordinate is the
acceleration in units of $a_0$. Adopted distances for Virgo and Coma are 15 Mpc and 100 Mpc, respectively. The values for Virgo are generated by an extrapolation of the mass model for M87 of \citet{mclaughlin99}. Some G01 galaxies are indicated by their NGC numbers.
The values for Coma are generated by using the NFW dark halo from \citet{lokas03}. Small open circles are the 18 T09 galaxies whose
projected distances are taken from \citet{godwin83}.
The comparison with Table \ref{surfdens} shows that the EF is small.}
\label{EF}
\end{center}
\end{figure}
\section{Conclusion}
Here we showed that (i) in MOND, galaxies exhibit a flattening of their circular velocity curve at high gravities ($g>a_0$) if they are described by a Jaffe profile with characteristic surface density $a_0/G$ in the region where the circular velocity is constant (since this is not possible for exponential profiles, it is remarkable that such flattenings of circular velocity curves at high accelerations are only observed in elliptical galaxies); (ii) the phantom halos of ellipticals predicted by MOND (i.e., the dark halos that would produce in Newtonian gravity the same additional gravity as MOND) can be fitted by logarithmic halos which perfectly reproduce the observed scaling relations of ellipticals for phase-space densities and central volume densities $\rho_0$; (iii) these halos have a constant characteristic surface density $\rho_0r_0$; (iv) contrary to spirals (for which there are more data in the very central parts), the phantom halos of ellipticals can as well be fitted by cuspy NFW halos, the concentration of which is in accordance with the theoretical predictions of $\Lambda$CDM for the highest masses, but in slight disagreement for baryonic masses smaller than $10^{11} M_\odot$: a modern, large, sample of elliptical galaxies, which are dynamically well investigated out to large radii and cover a large range of masses, will thus be required to get discriminating power. But in any case, and whatever the true physical reason for it, it is remarkable that a recipe (MOND) known to fit rotation curves of spiral galaxies with remarkable accuracy also apparently predicts the observed distribution of ``dark matter'' in
elliptical galaxies.
\begin{acknowledgements}
We thank an anonymous referee for a thoughtful report.
TR acknowledges financial support from the Chilean Center for Astrophysics,
FONDAP Nr. 15010003, from FONDECYT project Nr. 1100620, and
from the BASAL Centro de Astrofisica y Tecnologias
Afines (CATA) PFB-06/2007. BF acknowledges the support of the Humboldt foundation. GG is a postdoctoral researcher of the FWO-Vlaanderen (Belgium).
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
\label{intr}
In hard X-ray (HXR) emission of many flares quasi-periodic variations were observed with time-intervals between pulses, $P \sim 10-60$\,s [see \inlinecite{lip78}; see also review of \inlinecite{n+m09} and references therein].
In our previous paper (\opencite{paper1}, Paper I) we attempted to investigate relationship between HXR loop-top (LT) sources and the quasi-periodic variations. Main difficulty was that sequences of pulses are usually short, so that it is difficult to carry out comprehensive analysis of their quasi-periodicity. Therefore we used time-interval, $\Delta{t}$, between the strongest pulses as a simple estimate of the quasi-period $P$.
In the present paper we have selected four flares which have longer sequences of HXR pulses, so that it was possible to carry out detailed analysis of their quasi-periodicity (Section \ref{qper}). Section \ref{16jan} contains detailed analysis of 16 January 1994 flare. In Section \ref{det} we estimate values of electron density and magnetic field strength inside several HXR loop-top sources. Section \ref{sum} contains discussion and conclusions.
\section{Observations and their analysis}\label{obs}
We used HXR observations recorded by {\sl Yohkoh} Hard X-ray Telescope, HXT, \cite{kos91} [light-curves and images] and {\sl Compton Gamma Ray Observatory} Burst and Transient Source Experiment, BATSE, \cite{fis92} [light-curves].
\subsection{Analysis of quasi-periodicity of HXR pulses}\label{qper}
We have selected four flares which have longer sequences of HXR pulses, so that detailed analysis of their quasi-periodicity was possible. Main difficulty in the analysis was the fact that during impulsive phase the pulses occur simultaneously with quick increase of total HXR intensity (see Figure \ref{93may14pgm}a). Therefore we have applied a method of normalization which is commonly used by radioastronomers (see \opencite{fle08}). A normalized time series, $S(t)$, is:
\begin{equation}
S(t) = \frac{F(t) - \hat{F}(t)}{\hat{F}(t)},
\end{equation}
where $F(t)$ is the measured HXR flux and $\hat{F}(t)$ is a running average of $F(t)$. The red line in Figure \ref{93may14pgm}a shows $\hat{F}(t)$ calculated with averaging time $\delta{t} = 30$\,s. The normalized time series, $S(t)$ is shown in Figure \ref{93may14pgm}b. Quasi-periodicity of the pulses is clearly shown in this Figure and it is confirmed by power spectrum seen in Figure \ref{93may14pgm}c (Fourier transform of $S(t)$). The power spectrum has been calculated for time-interval 22:01-22:05:20 UT (Impulsive Phase, IP).
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{p2revf01.eps}}
\hfill \caption{Analysis of the HXR light-curve (25-50 keV, {\sl Compton Gamma Ray Observatory}/BATSE observations) for a flare of 14 May 1993. (a) HXR light-curve. The red line shows running average calculated with averaging time ${\delta}t = 30$ s. (b) Normalized light-curve, $S(t)$ [see Equation (1)]. Vertical dashed lines show the time-interval which was used to calculate power spectrum. (c) Power spectrum calculated for the normalized light-curve, $S(t)$. $P$ is the period corresponding to the peak in power spectrum.} \label{93may14pgm}
\end{figure}
We see in Figures \ref{93may14pgm}a and \ref{93may14pgm}b that before the impulsive phase, between 21:58:40 and 22:00:20 UT, three increasing pulses occurred. Mean time-interval between the pulses is about 31\,s which is close to the quasi-period, $P$, seen in the power spectrum (28\,s). This suggests that these pulses also belong to the quasi-periodic sequence seen during impulsive phase.
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{p2revf02.eps}}
\hfill \caption{23-33 keV light-curve recorded by {\sl Yohkoh}/ Hard X-ray Telescope for a flare of 16 January 1994. The red line connects peaks of HXR pulses to visualize their non-linear increase and saturation. The noise is due to statistical fluctuations of counting rate. The dashed vertical line shows the beginning of impulsive phase.} \label{94jan16lc}
\end{figure}
Figure \ref{94jan16lc} shows a HXR light-curve of 16 January 1994 flare. Like in Figure \ref{93may14pgm}a, during impulsive phase we see a sequence of increasing pulses. The red line connects tops of the pulses to show their quick non-linear increase and saturation. Before the impulsive phase (BIP) there is a sequence of weak pulses. Normalization and power-spectrum analysis is shown in Figure \ref{94jan16pgm}. The power spectrum in Figure \ref{94jan16pgm}c has been calculated for time interval A, i.e. BIP plus IP. Figure \ref{94jan16pgm}d shows power spectrum calculated for the decay phase (time-interval B), where the pulses were weak and their profiles were disturbed, but the quasi-period had been retained.
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{p2revf03.eps}}
\hfill \caption{The same as in Figure 1, for a flare of 16 January 1994. the power spectra have been calculated for time intervals A and B and they are shown in panels (c) and (d).} \label{94jan16pgm}
\end{figure}
Figure \ref{98aug18lc} shows HXR light-curve of 18 August 1998 flare. Its characteristic feature is that quasi-period during increase of HXR emission is shorter than during decrease of the emission. Therefore analysis of the light-curve has been done separately for time-intervals A and B (see Figures \ref{98aug18pgma} and \ref{98aug18pgmb}). In our model of oscillating magnetic traps changes of quasi-period are explained as being due to changes of length of oscillating traps (see Section 3).
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{p2revf04.eps}}
\hfill \caption{33-53 keV {\sl Yohkoh}/HXT light-curve for a flare of 18 August 1998. Power spectrum analysis has been carried out for time intervals A and B, separately, and results are presented in Figures 5 and 6.} \label{98aug18lc}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{p2revf05.eps}}
\hfill \caption{Power spectrum analysis for time-interval A of the flare of 18 August 1998.} \label{98aug18pgma}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{p2revf06.eps}}
\hfill \caption{Power spectrum analysis for time-interval B of the flare of 18 August 1998.} \label{98aug18pgmb}
\end{figure}
Figure \ref{93mar12lc} shows HXR light-curves of 12 March 1993 flare. General behavior is similar to that seen in Figure 2: There are low-amplitude oscillations before impulsive phase and quick increase of HXR emission after the onset of impulsive phase. Specific feature are short-period oscillations after HXR maximum. Therefore analysis of the light-curve has been done separately for time-intervals A and B (see Figures \ref{93mar12pgma} and \ref{93mar12pgmb}). Change of the quasi-period is clearly seen.
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{p2revf07.eps}}
\hfill \caption{HXR light-curves for a flare of 12 March 1993 recorded by {\sl Yohkoh}/HXT and {\sl CGRO}/BATSE. Very good agreement of the fluctuations in both curves is seen. Power spectrum analysis has been carried out for time-intervals A and B which are marked by vertical lines.} \label{93mar12lc}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{p2revf08.eps}}
\hfill \caption{Power spectrum analysis for time-interval A of the flare of 12 March 1993.} \label{93mar12pgma}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{p2revf09.eps}}
\hfill \caption{Power spectrum analysis for time-interval B of the flare of 12 March 1993.} \label{93mar12pgmb}
\end{figure}
\subsection{Analysis of 16 January 1994 flare} \label{16jan}
We have chosen the flare of 16 January 1994 for further analysis because loop-top and footpoint sources can be easily recognized in its HXR images (Figure \ref{mozaic}). We wanted to investigate what happened in the loop-top source during transition from low-amplitude oscillations to quick increase of HXR emission seen in Figures \ref{94jan16lc} and \ref{94jan16pgm}. Toward this end we have reconstructed {\sl Yohkoh}/HXR images in energy channels L (14-23 keV), M1 (23-33 keV), and M2 (33-53 keV) for the time of Before Impulsive Phase (BIP) oscillations (upper row in Figure \ref{mozaic}), about the beginning of impulsive phase (middle row) and for maximum of IP (lower row).
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{sp_p2_f10.eps}}
\hfill \caption{HXR images of the flare of 16 January 1994 in three
energy channels: 14-23, 23-33, and 33-53 keV (vertical columns).
Upper row: images recorded before impulsive phase, middle row:
recorded near the beginning of impulsive phase, lower row: recorded
during impulsive phase. See text for discussion.} \label{mozaic}
\end{figure}
We see that before impulsive phase HXR footpoints are strong which means that accelerated electrons easily escape from the LT source. But at the beginning of impulsive phase the footpoints are weaker than the loop-top source (F $<$ LT) in 14-23 and 23-33 keV emission which indicates that most of accelerated electrons deposit their energy within the LT source.
Next we have determined mean temperature, $T$, emission measure,
$EM$, and mean electron number density, $N$, in the HXR LT source,
from its {\sl Yohkoh} soft-X-ray (SXR) images, using filter-ratio
method (from Be119 and Al12 images). This analysis has been done in
the following way: We have integrated SXR fluxes from area $A = 4$
{\sl Yohkoh} pixels [i.e. 4.9 $\times$ 4.9 (arcsec)$^2$] at the
center of HXR LT source. Next we determined diameter, $d$, of the
HXR source according to isocontour $0.5I_{max}$, where $I_{max}$ is
the maximum intensity within the source. We assumed that extension
of the HXR source along the line of sight is also $d$ and calculated
the volume of the emitting plasma as $V = 1.15 A d$ (here 1.15 is a
correction factor which takes into account that actual size of the
HXR source is somewhat larger than that determined from isocontour
$0.5I_{max}$). Mean electron number density was calculated as $N =
\sqrt{EM/V}$ and obtained time-variation of the temperature, $T(t)$,
and density, $N(t)$, are shown in Figures \ref{temp} and \ref{nel}.
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{sp_p2_f11.eps}}
\hfill \caption{Time-variation of the mean temperature in HXR
loop-top source of 16 January 1994 flare.} \label{temp}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{sp_p2_f12.eps}}
\hfill \caption{Time-variation of the mean electron number density
in HXR loop-top source of 16 January 1994 flare.} \label{nel}
\end{figure}
Figure \ref{temp} shows that increase of energy release in the LT source began already about 23:10:40 UT. Sharp peaks of temperature about 23:12:30 and 23:16 UT (Figure \ref{temp}) are correlated with peaks in HXRs (Figure \ref{94jan16pgm}b) which confirms that significant part of the energy of accelerated electrons is deposited within the loop-top source (the temperature peaks are somewhat, $\sim$16 s, delayed relative to the HXR peaks which is due to accumulation of plasma heating by accelerated electrons). Significant random fluctuations of $N(t)$ in Figure \ref{nel} are mostly due to random errors in estimates of the source volume from individual HXR images. Systematic increase of density with time seen in Figure \ref{nel} is due to chromospheric evaporation flow.
X-ray images for a flare of 30 October 1992 are shown in Figure
\ref{92oct30img}. We see similar behavior like in Figure
\ref{mozaic}: Footpoint sources were stronger than LT source before impulsive phase and the LT source was stronger during impulsive phase.
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{sp_p2_f13.eps}}
\hfill \caption{HXR images of the flare of 30 October 1992 in two
energy channels: 14-23 and 23-33 keV (vertical columns). Upper row:
images recorded before impulsive phase, lower row: images recorded
during impulsive phase.} \label{92oct30img}
\end{figure}
\subsection{Determination of magnetic field strength and electron
densities for flares which were investigated in Paper I}
\label{det}
In Paper I we used $N = 1.0 \times 10^{10}$\, cm$^{-3}$ as typical
value of the electron number density in HXR LT sources during
impulsive phase. This value has been obtained by \inlinecite{k+l08}
from {\sl RHESSI} soft-X-ray images of the sources. In Paper I we
used this value of $N$ to estimate magnetic field strength in
oscillating magnetic traps:
\begin{equation}
B_2 = v_2 \sqrt{4 {\pi} {\rho}}
\end{equation}
where $v_2$ is the wave speed estimated from analysis of HXR
oscillations and ${\rho} \approx N m_H$, $m_H$ is the mass of
proton. Obtained values of $B_2$ are given in Table \ref{tab1}.
For eight flares of those investigated in Paper I it was possible to
determine mean electron density, $N$, for HXR LT source from SXR
images using the method described in Section \ref{16jan}. The values
are given in Table \ref{tab1}. We see that most values are
significantly higher than that assumed in Paper I ($N = 1.0 \times
10^{10}$ cm$^{-3}$). This indicates that the value of $N$ used in
Paper I, and therefore also magnetic field strength, $B_2$, were
underestimated. We have calculated corrected values, $B_2^{\ast}$,
of the magnetic field strength using Equation (2) with the new
values of $N$ (see Table \ref{tab1}).
Let us note that values of electron number density, $N$, derived
from {\sl Yohkoh}/ SXT images are reliable, since:
\begin{enumerate}
\item They have been confirmed by independent method (see
\opencite{b+j07}),
\item Determination of $N$ from {\sl Yohkoh}/Al12 images does not
depend significantly on temperature estimates, since instrumental
response function for the Al12 images is nearly constant in wide
range of temperatures (7-40 MK -- see curve {\sl f} in Figure 9 of
\opencite{tsu91}).
\end{enumerate}
\begin{table}
\caption{Values of mean electron density and magnetic field
strength} \label{tab1}
\begin{tabular}{ccccc}
\hline Date & UT & $B_2$ [G]$^{\rm a}$ & $N$ [$10^{10}$ cm$^{-3}$]$^{\rm b}$ & $B_2^{\ast}$ [G]$^{\rm c}$ \\
\hline 28 Jun 92 & 14:00 & 36 & 1.6 & 45 \\
8 May 98 & 01:59 & 40 & 2.3 & 62 \\
31 Oct 91 & 09:11 & 26 & 6.0 & 64 \\
14 May 93 & 22:05 & 30 & 12 & 104 \\
7 Jun 93 & 14:22 & 33 & 15 & 127 \\
16 Jan 94 & 23:17 & 35 & 8.0 & 98 \\
23 Jun 00 & 14:26 & 28 & 15 & 61 \\
25 Nov 00 & 18:39 & 21 & 9 & 63 \\
\hline
\end{tabular}
\begin{list}{}{}
\item[$^{\rm a}$] magnetic field strength estimated in Paper I
(assuming $N = 1 \times 10^{10}$ cm$^{-3}$)
\item[$^{\rm b}$] mean electron number density in HXR LT source at
HXR maximum, estimated in the present paper
\item[$^{\rm c}$] magnetic field strength calculated with the
new value of $N$
\end{list}
\end{table}
\section{Discussion and summary} \label{sum}
In Paper I we have proposed a model of oscillating magnetic traps to explain quasi-periodic oscillations seen in HXRs. In this model we have assumed that:
\begin{enumerate}
\item Magnetic structure at the top of flaring loop is triangular (``cusp-like''; Figure \ref{scheme}).
\item Electrons are efficiently accelerated during compression of magnetic traps (\opencite{s+k97},\hspace*{1mm} \opencite{k+k04},\hspace*{1mm} \opencite{b+s05}).
\end{enumerate}
The cusp-like structure is now a standard model of solar flare magnetic field -- see \inlinecite{asc04}. At the top of this structure (near P in Figure \ref{scheme}) magnetic reconnection occurs. Reconnected magnetic field and plasma flow into the confined volume BPC with high velocity. Hence, it seems to be obvious that, this should excite magnetosonic oscillations within the BPC volume.
Magnetic fields PB and PC reconnect at P (Figure \ref{scheme}a). This generates a sequence of magnetic traps which move downward, the traps overtake each other, they collide and undergo compression (Figure \ref{scheme}b). During the compression particles are accelerated within the traps, magnetic pressure, gas pressure, and the pressure of accelerated particles increase, so that the compression is stopped, the traps can expand and undergo magnetosonic oscillations.
\begin{figure}
\centerline{\includegraphics[width=1\textwidth]{jj+mtf2.eps}}
\caption{{\bf a} Schematic diagram showing a cusp-like magnetic
configuration. Magnetic fields PB and CP reconnect at P. The thick
horizontal line is the chromosphere. {\bf b} Detailed picture of the
BPC cusp-like magnetic structure.} \label{scheme}
\end{figure}
During the compression parameters of the traps undergo strong changes. At the beginning of compression the trap ratio, $\chi = B_{max}/B_{min}$, is high ($B_{max}$ is the magnetic field strength at magnetic mirrors and $B_{min}$ is the strength at the middle of the trap). The trap ratio decreases during compression and it reaches the lowest value, ${\chi}_{min}$, at the end of compression. Then the electrons reach highest energies and they can most easily escape from the trap. Plasma density determines the fraction of accelerated electrons which lose their energy within the trap, i.e. before they escape from the loop-top source. This fraction depends on the energy, since Coulomb losses depend on energy (see changes of the footpoint/loop-top intensity ratio along rows and along columns in Figure \ref{mozaic}).
Distribution of the oscillations within the BPC volume is determined by the time-profile, $v(t)$, of the velocity of reconnection flow which excites the oscillations. Its central part ($dv/dt \approx 0$) determines a ``main trap'' which contains many electrons and which is responsible for strong HXR pulses. Wings of the velocity profile ($dv/dt \ne 0$) are responsible for excitation of oscillations of many other (``secondary'') traps whose oscillations are shifted in phase. Superposition of many HXR pulses coming from these secondary traps gives a ``quasi-smooth'' emission (seen as the emission below the pulses in Figures \ref{93may14pgm}-\ref{93mar12pgmb}). We see that this quasi-smooth emission is stronger than pulses in the investigated flares which indicates that most of the volume BPC was filled with the ``secondary'' traps (their volume was greater than the volume of main trap).
In many flares HXR pulses are not clearly seen during quick increase of HXR emission, i.e. during beginning of impulsive phase (see example in Figure \ref{93mar12pgmb}). According to our model of oscillating magnetic traps, in such cases the time-profile, $v(t)$ of reconnection flow is quasi-rectangular, i.e. its maximum is flat. Therefore, it excites many oscillating traps of similar power, which have different length (see Figure \ref{scheme}b) and their oscillations are shifted in phase. This gives strong quasi-smooth HXR emission. Clear quasi-periodic pulses are seen only near maximum of HXR emission (see Figure \ref{93mar12pgmb}).
When we observe a long sequence of quasi-periodic oscillations (like in Figures \ref{93may14pgm}-\ref{93mar12pgmb}), this indicates that:
\begin{enumerate}
\item The oscillations have been excited by a short pulse of reconnection flow (otherwise the oscillations would be more chaotic).
\item The oscillations are self-maintained, i.e. some feedback mechanism operates which causes that the oscillations do not decay, but they increase during impulsive phase.
\end{enumerate}
\inlinecite{k+l08} investigated HXR LT sources from {\sl RHESSI} observations and they have found that the energy contained in non-thermal electrons is usually higher than thermal energy. This means that also the pressure of non-thermal electrons is higher than gas pressure, $p_{NT} > p$. This suggests that the pressure, $p_{NT}$, of non-thermal electrons is an important factor in the feedback mechanism which maintains the oscillations of magnetic traps. During compression of a magnetic trap the pressure $p_{NT}$ steeply increases and this causes increase of the amplitude of next expansion of the trap. This, in turn, causes increase of restoring force (i.e. the tension of bent magnetic field lines) and therefore the pulse of the pressure $p_{NT}$ will be stronger during the next compression. Hence, there is a feedback between the intensity of pulses of the pressure $p_{NT}$ and the amplitude of the magnetic trap oscillation and this feedback is responsible for quick increase of the oscillations during impulsive phase.
It may be interesting to note that this mechanism of maintaining the oscillations is analogous to the mechanism of instability which is responsible for stellar pulsations (``non-adiabatic pulsations''): in both cases the point is that additional energy is accumulated during compression of the gas (it is the energy of helium ionization in the case of Cepheids and it is energy of non-thermal electrons in our case).
Main results of the present paper are the following:
\begin{enumerate}
\item It has been confirmed that quasi-periodic oscillations (QPO) occur in HXR emission of solar flares.
\item We have found that low-amplitude QPO occur before impulsive phase of some flares.
\item We have found that quasi-period of the oscillations can change in some flares. We interpret this as being due to changes of the length of oscillating magnetic traps.
\item During impulsive phase most of the energy of accelerated (non-thermal) electrons is deposited within the HXR loop-top sources (Section \ref{16jan}). [For weak HXR flares this is seen at lower energies (14-23 keV), but not at 23-33 keV.]
\item We argue that the basic properties of the HXR oscillations can be explained in terms of a simple model of oscillating magnetic traps (see Paper I). This model allows us also to explain large number of electrons which are accelerated during impulsive phase: Observations show that the amplitude of the oscillations quickly increases and therefore the traps are filled with increasing amount of plasma coming from chromospheric evaporation. This is main source of electrons which undergo acceleration.
\item We suggest that a feedback between the pressure of accelerated electrons and the amplitude of the following expansion of magnetic trap is the mechanism which causes the quick increase of the amplitude of oscillations.
\item We have also determined improved values of electron number density and magnetic field strength for HXR loop-top sources of several flares which were investigated in Paper I. Obtained values fall within the limits of $N \approx (2 -15) \times 10^{10}$ cm$^{-3}$, $B \approx (45 - 130)$ gauss.
\end{enumerate}
Main advantages of the model of oscillating magnetic traps are the following:
\begin{enumerate}
\item Acceleration of electrons occurs in a large volume (BPC in Figure \ref{scheme}).
\item During development of impulsive phase plasma is delivered into BPC volume by chromospheric evaporation flow whose density is higher than the density of surrounding corona. This explains why number of accelerated electrons is very high.
\end{enumerate}
\begin{acks}
The {\sl Yohkoh} satellite is a project of the Institute of Space and Astronautical Science of Japan. The {\sl Compton Gamma Ray Observatory} is a project of NASA. We would like to thank anonymous referee for valuable remarks which helped us to improve this paper. This work was supported by Polish Ministry of Science and High Education grant No. N\,N203\,1937 33.
\end{acks}
|
\section*{ }{\bf Abstract:}
We study the mini--superspace quantization of spatially homogeneous (Bianchi)
cosmological universes sourced by a Dirac spinor field. The quantization of the homogeneous
spinor leads to a finite-dimensional fermionic Hilbert space and thereby to a multi-component
Wheeler-DeWitt equation whose main features are: (i) the presence of spin-dependent
Morse-type potentials, and (ii) the appearance of a q-number squared-mass term, which
is of order ${\cal O}(\hbar^2)$, and which is affected by ordering ambiguities. We give the
exact quantum solution of the Bianchi type-II system (which contains both scattering states
and bound states), and discuss the main qualitative
features of the quantum dynamics of the (classically chaotic)
Bianchi type-IX system. We compare the exact
quantum dynamics of fermionic cosmological billiards to previous works that described
the spinor field as being either classical or Grassmann-valued.
\section{Introduction}
The main aim of this work is to clarify the physical structure of the coupled quantum Einstein-Dirac
system within a minisuperspace cosmological setting, and, in particular, its dynamics
in the vicinity of a cosmological
singularity. Let us recall that the seminal work of Belinsky, Khalatnikov and Lifshitz (BKL) \cite{BKL} (see also Misner \cite{Mi}) has brought
into light the chaotically `` oscillatory" behaviour, near the singularity, of the diagonal components of the metric both
in homogeneous Bianchi IX cosmological models, and in the ``general (classical) solution'' of Einstein equations.
Recently, this chaotic BKL behaviour acquired a new significance through the discovery of its unexpected link with infinite-dimensional
Kac-Moody algebras \cite{DaHe2,Damour:2001sa,DHN} (for reviews, see \cite{DHN2,Henneaux:2007ej}).
The presence of such a chaotic behaviour depends both on the spacetime dimension, $D$,
and on the matter content of the considered cosmological model. For instance, pure gravity (i.e. the vacuum Einstein equations) has a
chaotic, BKL-like behaviour in spacetime dimension $ D \leq 10$, but a monotonic, Kasner-like behaviour in spacetime dimensions $ D \geq 11$ \cite{Demaret:1986ys}.
The effect, near a singularity, of a general {\it bosonic} matter content (scalar field and $p-$forms), in any spacetime dimension $D$,
has been studied in Ref. \cite{DHN2}. The main result is that
almost all the bosonic degrees of freedom ``freeze" near the singularity (i.e. admit limits, at each spatial point, as $ t \to 0$), except for the
diagonal part of the spatial metric, parametrized as $g_{a a} = \exp (- 2 \beta_a)$
(together with any scalar field $\varphi = \beta_0$, if present). The dynamics of the $\beta_a$'s can be represented as that of a
"ball" (of position $ \beta_a$) moving on a Lorentzian (or hyperbolic) billiard. The latter billiard motion
can then either be chaotic, or non-chaotic (i.e. ultimately monotonic), depending on the geometry of the `` billiard table",
which is a polyhedron domain in hyperbolic space.
The quantum mini--superspace versions of some of these bosonic cosmological
billiards have been studied in several works \cite{Misner:1969ae,Ryan72,RyanShepley75,Montani:2007vu}
While the effect of {\it bosonic} matter on the dynamics near a singularity is well understood, the effect of {\it fermionic} matter
is less well understood. This difference in understanding has both technical and physical roots. From the technical point of view,
the Hamiltonian description of spinor fields, coupled to gravity, is rather subtle and complex. The Hamiltonian description
of a Dirac (spin $\frac{1}{2}$) field, coupled to gravity (notably in Bianchi spacetimes), has been clarified in several papers,
see Refs. \cite{NelsonTeitelboim,IsNe,HenPRD,HenIHP,ObRy,Jantzen:1982je,Christodoulakis:1983qb,D'Eath:1986mc}. Here, we shall not consider the case of the gravity coupling to
a Rarita-Schwinger (spin $\frac{3}{2}$) field, i.e. the case of ``supersymmetric (quantum) cosmology", though we view the Einstein-Dirac system as a toy model for the
Einstein-Rarita-Schwinger case. See, e.g., \cite{D'Eath:1996at} for a entry in the literature on
supersymmetric quantum cosmology.
On the other hand, the physical meaning of having spinorial sources in cosmology has remained somewhat obscure. Various (conflicting)
approaches to this issue have been assumed in the literature. We shall distinguish three different approaches to the treatment of
spinorial fields, say $\psi$: (i) bilinears\footnote{Some authors even replace the $\psi$ variables themselves by real numbers.}
in $\psi$ (in the source terms, $T^{\mu \nu} $, for gravity) are replaced by real numbers (or ``c-numbers");
(ii) the $\psi$ variables are treated as Grassmann-valued (or ``G-numbers");
or (iii) the $\psi$'s are treated as quantum operators (or ``q-numbers").
When treating the $\psi$'s as G-numbers, they do not affect the BKL chaos (if the latter is present in absence of fermions).
Indeed, $T^{\mu \nu} \propto \psi^2$ only modifies the ``soul" $g_2 + g_4 + \cdots$ of the metric $g$, i.e. the part of the Grassmann expansion
$ g = g_0 + g_2 + g_4 \cdots$ of $g$ which contains an even (non zero) number of Grassmann generators. By contrast, the ``body" $g_0$ of $g$ (i.e. the
part of $g$ which does not contain any Grassmann generator) is unaffected by the $\psi$'s, and is entirely determined by the bosonic
sector of the theory. This approach has been recently used \cite{DaHi} to discuss how the chaotic behaviour of the body $g_0$ of the metric induces a
corresponding chaotic dynamics in the (Grassmann-level-one) fermions $ \psi= \psi_1 + \psi_3 + \cdots$.
By contrast, if one replaces the $ \sim \psi^2$ source terms by c-numbers (that react back on the body $g_0$ of the metric),
they can drastically modify the asymptotic behaviour near the singularity. Indeed, it has been argued by Belinsky and Khalatnikov \cite{BK}
that the presence of a Dirac field would, if so treated, ultimately destroy the BKL chaos. This result has been confirmed,
and streamlined, by the Hamiltonian treatment of \cite{dBHP} which showed that the $ \sim \psi^2$ source terms modify the billiard
dynamics of the (logarithmic) cosmological scale factors $\beta_a$ by adding a (positive) ``squared-mass term" $\mu^2$ to the effective
(Lorentzian-signature) Hamiltonian describing the dynamics of $\beta_a$. Indeed, such a squared-mass term (in Lorentzian $\beta$ space)
slows down the motion of the $\beta$ particle and ultimately prevent collisions on the cushions of the billiard table.
In the present work we shall treat the fermions as q-numbers and face the problem of discussing the meaning of the back reaction of fermions
within a quantum framework. This is a notoriously difficult problem, but we shall be able to get answers by restricting ourselves
to the mini-superspace framework of homogeneous cosmological spacetimes.
In that framework, the quantization of the spinor $\psi$ is equivalent to
considering that the wavefunction $\Phi$ of the universe is a multi-component object,
say $\Phi_\sigma$, where the discrete index $\sigma$ labels the finite number of independent
states allowed by the anticommutation relations of $\psi$. In particular, as we shall see in
detail below, when $\psi$ is
a Majorana spinor the discrete index $\sigma$ takes 4 values and can be identified with the
spinor index of an auxiliary $SO(4)$ Clifford algebra defined by the quantum operators
$\Gamma_\alpha= 2 g^{\frac14} \psi_\alpha$, $\alpha=1,2,3,4$. The appearance
in the wavefunction of such
a discrete spinor-type label when considering zero-mode (spatially independent)
Fermionic operators is well known from the Ramond string case \cite{Ramond:1971gb}
, and has also been
used in some studies of quantum supersymmetric minisuperspace cosmologies,
see, e.g., \cite{Obregon:1998hb}.
Most of our discussion will allow for a general (class A) Bianchi model,
but we shall discuss the quantum dynamics in details only for two cases:
the Bianchi IX (or `mixmaster') model, and the (non-chaotic) Bianchi II model.
We shall also compare the quantum solution to the various ways of discussing its classical analogs.
Our paper is organized as follows. Section 2 discusses the classical Lagrangian formulation
(with a specific fixing of the vielbein in terms of the metric) of an homogeneous spinor
field coupled to a Bianchi metric. This is followed in Sec. 3 by the corresponding classical
Hamiltonian formulation. The quantization of this system, and the discussion of the quantum
dynamics of the Bianchi-IX and Bianchi-II Einstein-Dirac systems is presented in Sec. 4. The comparison of the q-number and G-number approaches
is discussed in Sec. 5, while our main conclusions are presented in Sec. 6. Finally,
three appendices present complementary material, namely: Appendix A : alternative
Hamiltonian approach; Appendix B: alternative fixing of the dreibein; and Appendix C:
classical dynamics of the Bianchi-II case.
\section{Classical Lagrangian formulation of a homogeneous spinor field coupled to a Bianchi cosmological metric}\label{sec2}
\setcounter{equation}{0}
\subsection{Choice of approach}\label{ssec2.1}
The coupling of a spinor field to gravity poses special problems in view of the need to use a vielbein, say $h_{\ \mu}^{\hat\alpha}$, in addition to the metric $g_{\mu\nu}$, to describe the spinor degrees of freedom. The metric and vielbein components are related by
\begin{equation}
\label{eq2.1}
g_{\mu\nu} = \eta_{\hat\alpha\hat\beta} \, h_{\ \mu}^{\hat\alpha} \, h_{\ \nu}^{\hat\beta} \, .
\end{equation}
Here, and in the following, we shall use hatted indices to denote ``flat indices'' referring to a local orthonormal frame where $\eta_{\hat\alpha\hat\beta} = {\rm diag} (-1,+1,+1,+1)$. [We restrict ourselves to a four-dimensional spacetime, and use greek indices to denote spacetime indices, and latin indices to denote spatial ones.]
\smallskip
There are basically two different approaches to the description of the coupled gravity-spinor (or Einstein-Dirac) dynamics. Either: (i) the gravity degrees of freedom are only described by means of the metric components $g_{\mu\nu}$, the vielbein ones $h_{\ \mu}^{\hat\alpha}$ being (locally) determined in terms of $g_{\mu\nu}$ by some suitable gauge-fixing of the local Lorentz symmetry $SO(3,1)$, or (ii) one does not break the local Lorentz symmetry, and describes the gravity degrees of freedom by means of the redundant vielbein variables $h_{\ \mu}^{\hat\alpha}$. The approach (i) is technically simpler, but depends on the choice of a specific gauge-fixing of the local $SO(3,1)$ symmetry. The approach (ii) does not depend (until one discusses explicit solutions) on the choice of a gauge-fixing of $SO(3,1)$, but involves more constraints than the first approach, namely the constraints linked to the local $SO(3,1)$ gauge symmetry. We shall use the approach (i) in the text, and discuss the approach (ii) in an Appendix. Note that the approach
(i) has been advocated in Ref. \cite{ObRy}, but that its implementation in that reference
differs from the one we shall use (and does not take advantage of the useful automorphic
potentialities of the matrix $S$ discussed below).
\smallskip
Having decided on the approach (i), we need to choose a specific way of fixing the local $SO(3,1)$ gauge symmetry, {\it i.e.} of determining a specific vielbein $h_{\ \mu}^{\hat\alpha}$, when given the metric components $g_{\mu\nu}$ (in some coordinate system, or, more generally, some
non-holonomic frame). Here, the (assumed) symmetry properties of Bianchi models come to our help. Let us recall that the geometry of homogeneous spacetimes admits the special form
\begin{eqnarray}
\label{eq2.2}
ds^2 &= &g_{\mu\nu} (x) \, dx^{\mu} \, dx^{\nu} \nonumber \\
&= &-N^2 (t) \, dt^2 + g_{ab} (t) (N^a(t) \, dt + \tau^a (x)) (N^b (t) \, dt + \tau^b (x))
\end{eqnarray}
where the one-forms $\tau^a (x) = \tau^a_i (x) \, dx^i$ only depend on the spatial coordinates used in the $t = {\rm const.}$ slices. In the case (considered here) of Bianchi geometries, {\it i.e.} such that the spatial slices admit a simply transitive Lie group $G$ preserving the geometry, the one-forms $\tau^a$ can be chosen to be invariant under the group $G$, {\it i.e.} $\pounds_{\xi_b} \tau^a = 0$, where $\pounds$ denotes a Lie derivative and $\xi_b = \xi_b^i (x) \, \partial / \partial x^i$, with $b = 1,2,3$, a basis of three infinitesimal generators of $G$. With a suitable choice of $\tau^a$, the structure constants $C^c \, _{ab}$ of (the Lie algebra of) $G$, which enter the Lie brackets of the Killing vectors, $[\xi_a , \xi_b] = + \, C^c \, _{ab} \ \xi_c$, also enter the Cartan differential of the forms $\tau^a$, namely
\begin{equation}
\label{eq2.3}
d\tau^a = + \, \frac{1}{2} \, C^a \, _{bc} \ \tau^b \wedge \tau^c \, .
\end{equation}
Note that the constants $C^a \, _{bc}$ enter with the {\it same} sign in $[\xi_b , \xi_c]$ and in $d \tau^a$, because there are two flips of sign when going from the $\xi_a$'s to the $\tau^a$'s: one flip between the bracket of $[\xi , \xi]$ and the $d$ of the co-frame dual to the $\xi$'s, and a second flip between the structure constants of the left action of $G$ (defined, say, by the $\xi$'s) and those of its right action (which commutes with the $\xi$'s, and corresponds to the invariant (co)-frame linked to the $\tau$'s).
\smallskip
Apart from the $\tau^a$'s (which depend on the spatial coordinates only), the other objects entering the homogeneous metric (\ref{eq2.2}) only depend on the (coordinate) time, $t$. In addition, by using some time-dependent change of coordinates, one can ensure that the components of the shift vector $N^a(t)$ vanish for all times. We shall generally assume that this is the case, but we have left them in Eq.~(\ref{eq2.2}) as a reminder that, in the Hamiltonian formalism, the $N^a$'s enter the action as Lagrange multipliers of the three diffeomorphism constraints ${\mathcal H}_a \approx 0$. Similarly, the ``lapse'' $N(t)$ enters the action as a Lagrange multiplier for the Hamiltonian constraint ${\mathcal H} \approx 0$. Finally, the dynamical variables of Bianchi geometries are the six
functions of time $g_{ab} (t)$.
\subsection{Fixing the local Lorentz gauge symmetry}\label{ssec2.2}
In view of the special structure (\ref{eq2.2}) of the geometry it is natural to choose a vielbein co-frame $\theta^{\hat\alpha} = h_{\ \mu}^{\hat\alpha} \, dx^{\mu}$ of the form
\begin{eqnarray}
\label{eq2.4}
\theta^{\hat 0} &= &N(t) \, dt \, , \\
\label{eq2.5}
\theta^{\hat a} &= &\sum_b e^{-\beta_a(t)} \, S^{\hat a} \, _b \ (t) (\tau^b + N^b(t) \, dt) \, ,
\end{eqnarray}
where the matrix $S^{\hat a} \, _b$ satisfies\footnote{We generally use Einstein's summation convention, except when there are ambiguities, as in Eqs.~(\ref{eq2.5}) or (\ref{eq2.6}).}
\begin{equation}
\label{eq2.6}
\sum_a e^{-2\beta_a} \, S^{\hat a} \, _b \ S^{\hat a} \, _c = g_{bc} \, .
\end{equation}
In other words, the matrix $S^{\hat a} \, _b$, or rather its inverse $S^b \, _{\hat a}$, such that
\begin{equation}
\label{eq2.7}
S^{\hat a} \, _c \ S^c \, _{\hat b} = \delta_{\hat b}^{\hat a} \, , \quad S^a \, _{\hat c} \ S^{\hat c} \, _b = \delta_b^a \, ,
\end{equation}
transforms the quadratic form $g_{bc}$ into a diagonal form
\begin{equation}
\label{eq2.8}
g_{ab} \, S^a \, _{\hat c} \ S^b \, _{\hat d} = [{\rm diag} (e^{- 2\beta_1} , e^{-2\beta_2} , e^{-2\beta_3})]_{\hat c \hat d} \, .
\end{equation}
The idea behind the representation (\ref{eq2.6}) is to encode the six independent components of $g_{ab}$ into two sets of three variables: (1) three ``diagonal'' degrees of freedom $\beta_1 , \beta_2 , \beta_3$; and (2) three ``off diagonal'' ones, parametrizing the ``diagonalizing'' matrix $S$. For such a decomposition to be uniquely defined, one needs to restrict the structure of the ($3$ by $3$) matrix $S$ by six conditions. This can be achieved in several different ways. For instance, one could require the lower-than-diagonal elements of the matrix $S^{\hat a} \, _b$ to vanish, and its diagonal elements to be equal to 1; this would correspond to the so-called Iwasawa decomposition of $h^{\hat a} \, _b$ (and, correspondingly, of $g_{cd} = \delta_{\hat a \hat b} \ h^{\hat a} \, _c \ h^{\hat b} \, _d)$, which is unique, and which was found to be useful in recent work on the hidden presence of Kac-Moody symmetries in gravity theories \cite{DHN2}. We shall discuss in Appendix B the use of this decomposition in the study of the dynamics of type~II Bianchi cosmologies. However, in the case of the most generic (class~A) Bianchi models, namely type~IX and type~VIII, the Iwasawa decomposition is rather inconvenient.
As was emphasized by R.T.~Jantzen \cite{Jantzen:1982je,Jantzen79,Jantzen:2001me}
it is quite advantageous to use decompositions (\ref{eq2.5}), (\ref{eq2.6}) with a matrix $S^{\hat a} \, _b$ restricted to belonging to the {\it automorphism group}, say ${\mathcal A}$, of the Lie algebra ${\frak G}$ of $G$. Explicitly, this means that ${\mathcal A}$ is the group of linear transformations which leave invariant the structure constants $C^a \, _{bc}$.
The reason why the choice of a ``diagonalizing'' matrix $S$ belonging to the automorphism group ${\mathcal A}$ is advantageous is that (as we shall see explicitly below) the potential terms in the Hamiltonian can be expressed in terms of the $\beta$'s and of the components, say $\overline C^{\overline a} \, _{\overline b \overline c}$, of the structure constants w.r.t. the intermediate, co-frame $\overline\tau^{\overline a} = S^{\hat a} \, _b \ \tau^b$. These components are given by
\begin{equation}
\label{eq2.9}
\overline C^{\overline a} \, _{\overline b \overline c} = S^{\hat a} \, _{a'} \ S^{b'} \, _{\hat b} \ S^{c'} \, _{\hat c} \ C^{a'} \, _{b'c'} \, .
\end{equation}
For a general choice of $S^{\hat a} \, _b$ the components $\overline C^{\overline a} \, _{\overline b \overline c}$ would depend on the off-diagonal variables entering $S^{\hat a} \, _b$. For instance, when using an Iwasawa decomposition of $g_{ab}$ (with an upper diagonal matrix $S$), the $\overline C^{\overline a} \, _{\overline b \overline c}$ explicitly depend on the off-diagonal variables $\nu_{12} , \nu_{23} , \nu_{13}$ entering Eq.~(\ref{Iwasrep}), so that the potential terms in the Hamiltonian also depend on these off-diagonal variables. By contrast, by definition of the automorphism group ${\mathcal A}$ (as fixing ${\frak G}$ and therefore the $C$'s) when $S \in {\mathcal A}$ the components $\overline C^{\overline a} \, _{\overline b \overline c}$ are simply equal to the original $C^a \, _{bc}$, and thereby do not introduce any dependence on the off-diagonal metric variables.
\smallskip
When $G$ is a simple group (and when its Dynkin diagram has no symmetries), the automorphism group ${\mathcal A}$ of ${\frak G}$ is the adjoint group of $G$. In the case of Bianchi type~IX (where $G = SU(2)$), this means that ${\mathcal A}$ is $SO(3)$ (which is the
quotient $SU(2)/Z_2$). In that case, $S^{\hat a} \, _b$ is an orthogonal matrix, and the decomposition (\ref{eq2.6}) is the so-called ``Gauss decomposition'', corresponding to the diagonalization of the quadratic form $g_{ab}$ w.r.t. a given Euclidean metric $\delta_{ab}$. [Such
a Gauss decomposition was advocated by M. Ryan \cite{Ryan72,RyanShepley75}.]
In the case of Bianchi type~VIII, the automorphism group is $SO(1,2)$, which means that one should use an ``hyperbolic'' generalization of the Gauss decomposition of $g_{ab}$ w.r.t. a given Lorentzian metric $\eta_{ab} = {\rm diag} (-1 , -1 , +1)$. The special metrics, $\delta_{ab}$ or $\eta_{ab}$, that enter here are simply (modulo a suitable normalization) the Cartan-Killing metrics $k_{ab}$ associated to the Lie algebra ${\frak G}$, say
\begin{equation}
\label{eq2.10}
k_{ab} = - \, \frac{1}{2} \, C^c \, _{ad} \ C^d \, _{bc} \, .
\end{equation}
For instance, in the usual basis for type~IX we have $C^a \, _{bc} = \varepsilon_{abc}$, so that $k_{ab} = + \, \delta_{ab}$.
\smallskip
Summarizing: In the Bianchi type~IX case, we parametrize the six metric degrees of freedom contained in $g_{ab} (t)$ by means of the three diagonal variables $\beta_1 (t) , \beta_2 (t) , \beta_3 (t)$ and the three Euler angles $\theta_1 (t) , \theta_2 (t) , \theta_3 (t)$ parametrizing the {\it orthogonal} metric $S^{\hat a} \, _b \, (\theta_1 , \theta_2 , \theta_3)$ entering the Gauss decomposition (\ref{eq2.6}) of the quadratic form $g_{ab}$ (``transformation to principal axes''). On the other hand, in the Bianchi type~VIII case, the parametrization of $g_{ab} (t)$ by three diagonal variables $\beta_1 (t) , \beta_2 (t) , \beta_3 (t)$ and three ``diagonalizing angles'' $\theta_1 (t) , \theta_2 (t) , \theta_3 (t)$ should be done by an hyperbolic $(S \in SO (1,2))$ generalization of the Gauss decomposition, {\it i.e.} a transformation of $g_{ab}$ to principal axes w.r.t. a Lorentzian metric $\eta_{ab} = {\rm diag} (-1 , -1 , +1)$. [Such a transformation is always possible. It can be built from the eigenvectors of the non positive-definite quadratic form $\eta_{ab}$ w.r.t. the positive-definite one $g_{ab}$ .]
\smallskip
Then, in terms of such a parametrization, $g_{ab} \leftrightarrow (\beta_1 , \beta_2 , \beta_3 \, , \ S^{\hat a} \, _b \ (\theta_1 , \theta_2 , \theta_3))$ of the metric, we gauge-fix the local Lorentz symmetry by defining the specific vielbein $\theta^{\hat\alpha} (\beta_a , \theta_a) = h_{\mu}^{\hat\alpha} (\beta_a , \theta_a) \, dx^{\mu}$ by means of Eqs.~(\ref{eq2.4}), (\ref{eq2.5}).
\subsection{Lagrangian formulation of the Einstein-Dirac system}\label{ssec2.3}
Having uniquely determined (for types~IX and VIII) a vielbein $\theta^{\hat\alpha} \, _{\mu} \ dx^{\mu}$ in terms of the usual metric degrees of freedom $g_{ab} (t)$, $N^a (t)$, $N(t)$, we can now consider the general Einstein(-Hilbert)-Dirac Lagrangian density,
\begin{equation}
\label{eq2.11}
L = L_{EH} + L_D \, ,
\end{equation}
where\footnote{We use units such that $16 \pi \, G = 1 = c$.}
\begin{eqnarray}
\label{eq2.12}
L_{EH} &= &\sqrt{- \, ^4g} \ ^4R \, , \\
\label{eq2.13}
L_D &= &\sqrt{- \, ^4g} \ \left( \overline\Psi \, \gamma^{\hat\alpha} \, \nabla_{\hat\alpha} \, \Psi - m \, \overline\Psi \, \Psi \right) \, ,
\end{eqnarray}
as a functional of the metric $g_{\mu\nu}$, and of the spinor field $\Psi$ (w.r.t. the gauge-fixed vielbein $h_{\mu}^{\hat\alpha} (g_{\nu\lambda})$). We use gamma matrices adapted to our mostly plus signature $-+++$, namely
\begin{equation}
\label{eq2.14}
\gamma^{\hat\alpha} \, \gamma^{\hat\beta} + \gamma^{\hat\beta} \, \gamma^{\hat\alpha} = 2 \, \eta^{\hat\alpha\hat\beta} \, {\rm 1\mkern-4mu I} \, ,
\end{equation}
with an anti-hermitian $\gamma^{\hat 0}$ (satisfying $(\gamma^{\hat 0})^2 = -{\rm 1\mkern-4mu I}$), and three hermitian $\gamma^{\hat 1} , \gamma^{\hat 2} , \gamma^{\hat 3}$. We then choose
\begin{equation}
\label{eq2.15}
\beta := i \, \gamma_{\hat 0} = - \, i \, \gamma^{\hat 0} \, ,
\end{equation}
with the involutive $\beta$ ($\beta^2 = + {\rm 1\mkern-4mu I}$) being used to define the usual Dirac conjugate (as defined in the mostly minus signature)
\begin{equation}
\label{eq2.16}
\overline\Psi := \Psi^{\dagger} \, \beta \, .
\end{equation}
The frame covariant derivative of the spinor entering the Dirac action Eq.~(\ref{eq2.13}) is $\nabla_{\hat\alpha} \, \Psi \equiv h^{\mu} \, _{\hat\alpha} \ \nabla_{\mu} \, \Psi$, where $h^{\mu} \, _{\hat\alpha}$ is the vielbein frame dual to the vielbein co-frame $h^{\hat\alpha} \, _{\mu}$ ({\it i.e.} $h^{\hat\alpha} \, _{\mu} \ h^{\mu} \, _{\hat\beta} = \delta^{\hat\alpha} \, _{\hat\beta}$), and where the world-index covariant derivative $\nabla_{\mu} \, \Psi$ is given by
\begin{equation}
\label{eq2.17}
\nabla_{\mu} \, \Psi = \partial_{\mu} \, \Psi + \frac{1}{4} \, \omega_{\hat\alpha\hat\beta \, \mu} \, \gamma^{\hat\alpha\hat\beta} \, \Psi \, ,
\end{equation}
where (denoting $h_{\hat\alpha \mu}\equiv \eta_{\hat\alpha \hat\beta} h^{\hat\beta}\,_{\mu}$)
$$
\omega_{\hat\alpha\hat\beta \, \mu} := h_{\hat\alpha \nu} \, \nabla_{\mu} \, h^{\nu}\,_{\hat\beta}
$$
are the connection components of the vielbein (with last index taken as world index), and
where\footnote{Everywhere we use brackets $[\cdots]$ around indices to denote antisymmetrization with weight one.}
\begin{equation}
\label{eq2.18}
\gamma^{\hat\alpha\hat\beta} := \frac{1}{2} \, (\gamma^{\hat\alpha} \, \gamma^{\hat\beta} - \gamma^{\hat\beta} \, \gamma^{\hat\alpha}) \equiv \gamma ^{[\hat\alpha} \gamma^{\hat\beta]} \, .
\end{equation}
We shall only consider here class~A Bianchi models, {\it i.e.} models satisfying $C^a \, _{ac} = 0$, which is equivalent to saying that the dualization of the structure constants w.r.t. the antisymmetric lower indices, $n^{ad} := \frac{1}{2} \, \varepsilon^{bcd} \, C^a \, _{bc}$, yields a symmetric
tensor density. In other words
\begin{equation}
\label{eq2.19}
C^a \, _{bc} = \varepsilon_{bcd} \, n^{ad} \, ,
\end{equation}
where $\varepsilon_{abc} = \varepsilon_{[abc]}$ (with $\varepsilon_{123} = + \, 1$), and $n^{ab} = n^{ba}$. For type~IX, one has $n^{ab} = \delta^{ab}$ in the usual basis, while, for type~VIII $n^{ab} = {\rm diag} (+1,+1,-1)$. [Note that the Cartan-Killing metric $k_{ab}$, Eq.~(\ref{eq2.10}), associated to the $C$'s is quadratic in $n^{ab}$. In type~IX $k_{ab} = \delta_{ab} = {\rm diag} (+1,+1,+1)$ is numerically equal to $n^{ab}$, while in type~VIII $k_{ab} = {\rm diag} (-1,-1,+1)$ is of signature $--+$, independently of whether one chooses a basis where $n^{ab} = {\rm diag} (+1,+1,-1)$ or $n^{ab} = {\rm diag} (-1,-1,+1)$.]
\smallskip
It is well-known that the dynamics of class~A Bianchi models derives from a Lagrangian which is obtained simply by substituting in the general action (\ref{eq2.11}) the symmetry-reduced form of the metric, {\it i.e.} Eq.~(\ref{eq2.2}). This is also true when the metric is coupled to a homogeneous spinor. Here, we shall define the spatial homogeneity of a spinor $\Psi$ simply as meaning that the components of the spinor w.r.t. any frame $h_{\hat\alpha} := h_{\hat\alpha}^{\mu} \, \partial_{\mu}$ which is invariant under the homogeneity group $G$, {\it i.e.} $\pounds_{\xi_a} \, h_{\hat\alpha} = 0$, depend only on time\footnote{In some cases, some spatial variation of $\Psi$, of the type $\pounds_{\xi_a} \, \Psi = i \, \lambda_a \, \Psi$, with real quantities $\lambda_a$ subject to the integrability constraint $\lambda_a \, C^a \, _{bc} = 0$, is compatible with the homogeneity of the geometry \cite{HenIHP}. However, such a generalization is allowed neither in the case (we shall focus on) of a Majorana spinor, nor in the case of simple Lie algebras, such as type~IX or type~VIII, which are of most physical interest.}.
\smallskip
We know on general grounds \cite{ADM} that the lapse, $N$, and the shift vector, $N^a$, will enter the final Hamiltonian action $S = \int (pdq - H(q,p) \, dt)$ as Lagrange multipliers of, respectively, the Hamiltonian constraint, ${\mathcal H}$, and the diffeomorphism (or momentum) constraints, ${\mathcal H}_a$. Namely, the Hamiltonian is of the general form
\begin{equation}
\label{eq2.20}
H = \int (N {\mathcal H} + N^a \, {\mathcal H}_a) \, \mu
\end{equation}
where $\mu \equiv \tau^1 \wedge \tau^2 \wedge \tau^3$ is the spatial volume density [in the co-frame $(dt,\tau^1 , \tau^2 , \tau^3)$], and where
\begin{eqnarray}
\label{eq2.21}
{\mathcal H} &= &\sqrt g \, (2 \ ^4G_0^0 - T_0^0) = \sqrt g \, (^4R_0^0 - \, ^4R_a^a - T_0^0) \, , \\
\label{eq2.22}
{\mathcal H}_a &= &\sqrt g \, (2 \ ^4G_a^0 - T_a^0) = \sqrt g \, (2 \ ^4R_a^0 - T_a^0) \, .
\end{eqnarray}
Here, $g$ denotes the determinant of the spatial metric $g_{ab}$ (w.r.t. the spatial-coframe $\tau^a$, see Eq.~(\ref{eq2.2})), $^4G_{\beta}^{\alpha} \equiv \, ^4R_{\beta}^{\alpha} - \frac{1}{2} \, ^4R \, \delta_{\beta}^{\alpha}$ the spacetime Einstein tensor, and $T_{\beta}^{\alpha}$ the matter stress-energy tensor. The factor 2 multiplying $^4G_{\alpha}^0$ in the equations above represents $(8\pi G)^{-1}$ in the units we use where $16\pi G = 1$.
\smallskip
Knowing in advance the structure (\ref{eq2.20}), one can simplify the computation of the Hamiltonian by working in the special quasi-Gaussian gauge where
\begin{equation} \label{gauge}
N = \sqrt g \, , \ {\rm and} \ N^a = 0 \,.
\end{equation}
[This gauge was found useful in many previous cosmological studies, see
e.g. \cite{BKL,Ryan72,DHN2}.]
In addition, we shall assume (for notational simplicity) that we consider the dynamics of a (comoving) piece of a homogeneous universe which has a unit (comoving) volume $1 = \int \mu$. This allows one to identify the total Lagrangian with the Lagrangian density:
$$
S = \int \mu \int dt \, L(q,\dot q) = \int dt \, L (q,\dot q) \, .
$$
\subsection{Gravity part of the Lagrangian}\label{ssec2.4}
The gravity part of the Lagrangian, which generically reads (modulo a total divergence)
\begin{equation}
\label{eq2.23}
L_{EH} = N \, \sqrt g \, [g^{ac} \, g^{bd} \, K_{ab} \, K_{cd} - (g^{ab} \, K_{ab})^2 + R(g)] \, ,
\end{equation}
in terms of the spatial scalar curvature $R(g) \equiv \, ^3R (g)$, and of the second fundamental form\footnote{Note that Ref. \cite{ADM} defines $K_{ab}$ with the opposite sign.},
\begin{equation}
\label{eq2.24}
K_{ab} := \frac{1}{2N} \, (\partial_t \, g_{ab} - D_a \, N_b - D_b \, N_a)
\end{equation}
($D$ denoting the $3$-dimensional covariant derivative), reads, when working in the gauge (\ref{gauge})
\begin{equation}
\label{eq2.25}
L_{EH} = T_g (g,\dot g) - V_g (g) \, .
\end{equation}
Here $T_g$ denotes the ``kinetic-energy'' part of the gravity Lagrangian $(\dot g \equiv \partial_t \, g)$,
\begin{equation}
\label{eq2.26}
T_g (g,\dot g) = \frac{1}{4} \, g^{ac} \, g^{bd} \, \dot g_{ab} \, \dot g_{cd} - \frac{1}{4} \, (g^{ab} \, \dot g_{ab})^2 \, ,
\end{equation}
while $V_g$ denotes its ``potential'' part,
\begin{equation}
\label{eq2.27}
V_g (g) = - \, g \, R(g) \, .
\end{equation}
Using the decomposition (\ref{eq2.6}) of $g_{ab}$ into the three diagonal variables $\beta_1$, $\beta_2$, $\beta_3$, and the three angles $\theta_1 , \theta_2 , \theta_3$ parametrizing the $SO(3)$ [respectively $SO(1,2)$] matrix $S^{\hat a} \, _b$ in the type~IX (resp. type~VIII) case, we can express the gravity part (\ref{eq2.25}) of the Lagrangian in terms of $\beta_a , \theta_a$ and $\dot\beta_a , \dot\theta_a$.
\smallskip
The spatial scalar curvature of a general homogeneous metric $g_{ab} (t) \, \tau^a (x)$ $\tau^b (x)$ is expressible in terms of the structure constants of Eq.~(\ref{eq2.3}), namely
(see, e.g., \cite{SpindelGBS})
\begin{equation}
\label{eq2.28}
R(g) = - \frac{1}{4} \, C^a \, _{bc} \ C^{\underline a} \, _{\underline b \underline c} - \frac{1}{2} \, C^a \, _{bc} \ C^b \, _{a \underline c} - C^a \, _{ac} \ C^a \, _{a \underline c}
\end{equation}
where it is understood (for notational transparence) that a summation over, say, $a$ and $\underline a$ denotes an appropriate contraction by means of $g_{ab}$ or its inverse $g^{ab}$ ({\it e.g.} $A_c \, B_{\underline c} \equiv g^{cc'} \, A_c \, B_{c'}$). In the class~A case the last term in Eq.~(\ref{eq2.28}) vanishes, while the other ones can be expressed in terms of the (symmetric) dual, $n^{ad}$ of $C^a \, _{bc}$ (see Eq.~(\ref{eq2.19}). This leads to the following simple expression for the ``gravity potential'' $V_g$, Eq.~(\ref{eq2.27}),
\begin{equation}
\label{eq2.29}
V_g (g) = n^{ab} \, n^{\underline a \underline b} - \frac{1}{2} \, (n^{a \underline a})^2 \equiv g_{aa'} \, g_{bb'} \, n^{ab} \, n^{a'b'} - \frac{1}{2} \, (g_{ab} \, n^{ab})^2 \, .
\end{equation}
Inserting the decomposition (\ref{eq2.6}) into this result, then yields the expression of $V_g$ in terms of $\beta_a$ and the matrix $S$. As announced above, the fact that $S^{\hat a} \, _b$ was chosen to belong to the automorphism group ${\mathcal A}$ (leaving the structure constants $C^a \, _{bc}$ invariant) implies that the right-hand-side (r.h.s.) of Eq.~(\ref{eq2.28}) is independent of $S$, and only depends on the diagonal variables $\beta_a$. This is also true for the potential $V_g = - \, g \, R(g)$ if $S$ belongs to the ``special'' subgroup of ${\mathcal A}$ having $\det S = 1$. In the cases we consider here (types IX and VIII) this is automatically the case as ${\mathcal A} = SO(3)$ or $SO(1,2)$. Finally, we conclude that $V_g (g)$ is given by the same expression that it would have if $g_{ab}$ had been assumed to be diagonal, namely
\begin{eqnarray}
\label{eq2.30}
V_g (g) = V_g (\beta) &\equiv &n_1^2 e^{-4\beta_1} + n_2^2 e^{-4\beta_2} + n_3^2 e^{-4 \beta_3} - \frac{1}{2} \, (n_1 e^{-2 \beta_1} + n_2 e^{-2 \beta_2} + n_3 e^{-2 \beta_3})^2 \nonumber \\
&= &\frac{1}{2} \, [n_1^2 e^{-4 \beta_1} + n_2^2 e^{-4\beta_2} + n_3^2 e^{-4 \beta_3}] \nonumber \\
&&- \, [n_1n_2e^{-2(\beta_1 + \beta_2)} + n_2n_3e^{-2(\beta_2 + \beta_3)} +n_3n_1 e^{-2(\beta_3 + \beta_1)}] \, ,
\end{eqnarray}
where $n_a$ denote the diagonal components of $n^{ab} ={\rm diag} (n_1,n_2,n_3)$.
Turning now to the kinetic part $T_g$, Eq.~(\ref{eq2.26}), we need to evaluate it in terms of the $\beta - S$ parametrization (\ref{eq2.6}) of $g_{ab}$. To be explicit we should, at this stage, choose a specific convention for the definition of the three (possibly generalized) Euler angles $\theta_1 , \theta_2 ,\theta_3$ parametrizing the (pseudo-)orthogonal matrix $S$. It is, however, better to introduce a notation for the ``angular velocity'', say $w$, of the matrix $S$, before specifying its expression in terms of $\dot\theta_1 , \dot\theta_2 , \dot\theta_3$ and the $\theta$'s. We define $w^{\hat a} \, _{\hat b}$ by writing
\begin{equation}
\label{eq2.31}
\dot S^{\hat a} \, _c = \sum_{\hat b} w^{\hat a} \, _{\hat b} \ S^{\hat b} \ _c \, ,
\end{equation}
or $\dot S = w \, S$, if we consider $S \equiv S^{\hat a} \, _b$ and $w \equiv w^{\hat a} \, _{\hat b}$ as matrices. In this matrix notation, the decomposition (\ref{eq2.6}) reads $g = S^T \, e^{-2\beta} \, S$, where $g$ denotes here the matrix $g_{ab}$, $\beta$ the diagonal matrix $\beta_a \, \delta_{ab}$, and the superscript $T$ the transposition of a matrix. Differentiating the matrix $g$ then yields
\begin{equation}
\label{eq2.32}
\dot g = S^T (-2 \dot\beta e^{-2\beta} + e^{-2\beta} \, w + w^T \, e^{-2\beta}) \, S \, .
\end{equation}
At this stage, the calculation depends on whether the matrix $S$ is orthogonal ($SO(3)$; type~IX) or pseudo-orthogonal ($SO(1,2)$; type~VIII). We shall henceforth focus on the type~IX case, giving only some indications of the differences that arise in the type~VIII case. In the type~IX case we have
(when using the usual basis where $n^{ab}=\delta^{ab}$) $S^T S = SS^T = {\rm 1\mkern-4mu I}$ so that the ``matrix angular velocity'' $w$ defined by Eq.~(\ref{eq2.31}) is antisymmetric in the usual sense: $w^T = -w$. [In the type~VIII case $w$ would be antisymmetric in the Lorentzian sense, {\it i.e.} after considering $w_{\hat a \hat b} : = \eta_{\hat a \hat a'} \, w^{\hat a'} \, _{\hat b}$ where $\eta_{\hat a \hat b} = {\rm diag} (-1,-1,+1)$.] Inserting this knowledge in Eq.~(\ref{eq2.32}) then yields an explicit expression for $\dot g_{ab}$ of the form (denoting $w^{\hat a \hat b} \equiv \delta^{\hat b} \, _{\hat c} \ w^{\hat a} \, _{\hat c} = w^{\hat a} \, _{\hat b}$)
\begin{equation}
\label{eq2.33}
\dot g_{cd} = S^{\hat a} \, _c \ S^{\hat b} \, _d \ \overline k_{\overline a \overline b} \, ,
\end{equation}
\begin{equation}
\label{eq2.34}
\overline k_{\overline a \overline b} \equiv - \, 2 \, \dot\beta_a \, e^{-2\beta_a} + (e^{-2\beta_a} - e^{-2\beta_b}) \, w^{\hat a \hat b} \, .
\end{equation}
Here, we can think of $\overline k_{\overline a \overline b}$ as the components of the covariant tensor $k_{ab} := \dot g_{ab}$ w.r.t. to the ``rotated'' co-frame
\begin{equation}
\label{eq2.35}
\overline\tau^{\overline a} := S^{\hat a} \, _b \ \tau^b \, .
\end{equation}
The co-frame (\ref{eq2.35}) is intermediate between the basic co-frame $\tau^a$ (w.r.t. which $ds^2 = -N^2 \, dt^2 + g_{ab} \, \tau^a \, \tau^b$), and the orthonormal frame $\theta^{\hat\alpha}$ (w.r.t. which $ds^2 = \eta_{\hat\alpha\hat\beta} \, \theta^{\hat\alpha} \, \theta^{\hat\beta}$). Indeed, in the co-frame $(dt , \overline\tau^{\overline a})$ we have a non-Minkowskian, but {\it diagonal} form of the spacetime metric, namely: $ds^2 = -N^2 \, dt^2 + \underset{a}{\sum} \ e^{-2\beta_a} (\overline\tau^a)^2$, with $N = \sqrt{\det g_{ab}} = e^{-(\beta_1 + \beta_2 + \beta_3)}$.
\smallskip
As the gravitational kinetic-energy term (\ref{eq2.26}) is manifestly invariant under any linear change of basis of the co-frame $\tau^a$, it can be rewritten as
\begin{equation}
\label{eq2.36}
T_g = \frac{1}{4} \, \overline g^{\overline a \overline c} \ \overline g^{\overline b \overline d} \ \overline k_{\overline a \overline b} \ \overline k_{\overline c \overline d} - \frac{1}{4} \, (\overline g^{\overline a \overline b} \ \overline k_{\overline a \overline b})^2 \, ,
\end{equation}
where $\overline g^{\overline a \overline b}$ are the components of the contravariant metric w.r.t. the intermediate frame $\overline\tau^{\overline a}$, namely: $ \overline g^{\overline a \overline b} = e^{+2\beta_a} \, \delta_{ab}$. This yields the explicit result
\begin{equation}
\label{eq2.37}
T_g = T_{\beta} (\dot\beta) + T_w (w,\beta) \, ,
\end{equation}
where
\begin{equation}
\label{eq2.38}
T_{\beta} (\dot\beta) = \sum_a \dot\beta_a^2 - \left( \sum_a \dot\beta_a \right)^2 = -2 \, (\dot\beta_1 \, \dot\beta_2 + \dot\beta_2 \, \dot\beta_3 + \dot\beta_3 \, \dot\beta_1) \, ,
\end{equation}
and (in the type IX case)
\begin{eqnarray}
\label{eq2.39}
T_w^{\rm IX} (w,\beta) &= &2 \sinh^2 (\beta_1 - \beta_2) (w^{\hat 1 \hat 2})^2 + 2 \sinh^2 (\beta_2 - \beta_3)(w^{\hat 2 \hat 3})^2 \nonumber \\
&+ &2 \sinh^2 (\beta_3 - \beta_1)(w^{\hat 3 \hat 1})^2 \, .
\end{eqnarray}
Summarizing so far: the gravity part of the Lagrangian has the form
\begin{equation}
\label{eq2.40}
L_{EH} (g,\dot g) = T_{\beta} (\dot\beta) + T_w (w,\beta) - V_g (\beta) \, ,
\end{equation}
with a $\beta$-kinetic energy given by (\ref{eq2.38}), a rotational kinetic energy linked to the dynamics of the Euler angles entering $S(\theta_a)$ given by (\ref{eq2.39}) and a potential energy given by (\ref{eq2.30}) (with $n^{ab}=\delta^{ab}$). Note that the matrix $S$, {\it i.e.} the Euler angles $\theta_1 , \theta_2 , \theta_3$, do not explicitly enter the result (\ref{eq2.40}). As explained above, they do not enter the potential term $V_g$ because $S$ was chosen to belong to the automorphism group of the Lie algebra ${\mathfrak G}$. However, they do implicitly enter the rotational kinetic energy $T_w$, as the rotational angular velocity $w^{\hat a \hat b}$ do depend both on $\theta_a$ and $\dot\theta_a$. For instance, if we define the Euler angles as in the standard references \cite{Landau-Lifchitz,Goldstein}, namely using the $z-x-z$ convention, {\it i.e.} a matrix
\begin{equation}
\label{eq2.41}
S = \begin{pmatrix}
\cos\psi &\sin\psi &0 \\ -\sin \psi &\cos \psi &0 \\ 0 &0 &1
\end{pmatrix} \begin{pmatrix}
1 &0 &0 \\ 0 &\cos \theta &\sin\theta \\ 0 &-\sin\theta &\cos\theta
\end{pmatrix} \begin{pmatrix}
\cos\varphi &\sin\varphi &0 \\ -\sin \varphi &\cos \varphi &0 \\ 0 &0 &1
\end{pmatrix} ,
\end{equation}
the angular velocities $w^{\hat a \hat b}$ (with $w = \dot S \, S^{-1}$) are given by
\begin{eqnarray}
\label{eq2.42}
w^{\hat 1 \hat 2} &= &\dot\varphi \cos\theta + \dot\psi \, , \nonumber \\
w^{\hat 2 \hat 3} &= &\dot\varphi \sin\theta \sin\psi + \dot\theta \cos\psi \, , \nonumber \\
w^{\hat 3 \hat 1} &= &\dot\varphi \sin\theta \cos\psi - \dot\theta \sin\psi \, .
\end{eqnarray}
As several authors (see, e.g., \cite{Ryan72}) have previously remarked, the rotational part $T_w$ of the gravity Lagrangian is analogous to the kinetic energy of a rotating body (``asymmetric top''), namely $T_w = \frac{1}{2} (I_1 \, \Omega_1^2 + I_2 \, \Omega_2^2 + I_3 \, \Omega_3^2)$, where $\Omega_1 \equiv w^{\hat 2 \hat 3}$, $\Omega_2 \equiv w^{\hat 3 \hat 1}$, $\Omega_3 \equiv w^{\hat 1 \hat 2}$ are the body-frame components of the angular velocity and $I_3 = 4 \sinh^2 (\beta_1 - \beta_2)$, {\it etc.} the (body-frame) moments of inertia. Note, however, that, contrary to a usual rigid solid, the body-frame moments of inertia are time-dependent. Indeed, they depend on the $\beta$'s, which have a (coupled) dynamics determined by their kinetic energy $T_{\beta}$, their potential energy $V_{\beta}$, and their couplings to the other variables [and notably the Euler angles themselves, through the term $T_w (\dot\theta_a , \theta_a , \beta_a)$].
\subsection{Spinor part of the Lagrangian}\label{ssec2.5}
Let us now derive the explicit form of the spinor part, $L_D$, Eq.~(\ref{eq2.13}), of the Lagrangian. To do this, we need the explicit expression of the connection coefficients, say $\omega_{\hat\alpha\hat\beta\hat\gamma} \equiv \omega_{\hat\alpha\hat\beta\mu} \, h^{\mu}_{\ \hat\gamma} \,$, of our specifically chosen vielbein $\theta^{\hat\alpha} =h^{\hat\alpha} \,_{\mu} dx^{\mu}$, defined in Eq.~(\ref{eq2.5}) in terms of the metric degrees of freedom $\beta_a$, $S (\theta_a)$ (parametrizing $g_{ab}$ {\it via} Eq.~(\ref{eq2.6})). [In this
subsection we take as above a vanishing shift vector $N^a=0$, but
allow for an arbitrary lapse $N(t)$.] If we consider the ``structure constants'', say ${\mathcal C}^{\hat\alpha} \, _{\hat\beta\hat\gamma}$, of the orthonormal co-frame $\theta^{\hat\alpha}$, defined as
\begin{equation}
\label{eq2.43}
d \theta^{\hat\alpha} = \frac{1}{2} \, {\mathcal C}^{\hat\alpha} \, _{\hat\beta \hat\gamma} \ \theta^{\hat\beta} \wedge \theta^{\hat\gamma} \, ,
\end{equation}
the (frame) connection coefficients $\omega_{\hat\alpha\hat\beta\hat\gamma}$ can be expressed as
\begin{equation}
\label{eq2.44}
\omega_{\hat\alpha\hat\beta\hat\gamma} = \frac{1}{2} \, ({\mathcal C}_{\hat\alpha\hat\beta\hat\gamma} + {\mathcal C}_{\hat\beta\hat\gamma\hat\alpha} - {\mathcal C}_{\hat\gamma\hat\alpha\hat\beta}) \, ,
\end{equation}
where we denoted ${\mathcal C}_{\hat\alpha\hat\beta\hat\gamma} := \eta_{\hat\alpha\hat\sigma} \, {\mathcal C}^{\hat\sigma} \, _{\hat\beta\hat\gamma}$.
\smallskip
As $\theta^{\hat 0} = N(t) \, dt$ only involves $t$, we have $d \theta^{\hat 0} = 0$, so that
\begin{equation}
\label{eq2.45}
{\mathcal C}^{\hat 0} \, _{\hat\beta\hat\gamma} = 0 \qquad (\mbox{and} \quad {\mathcal C}_{\hat 0 \hat\beta \hat\gamma} = 0) \, .
\end{equation}
On the other hand $\theta^{\hat a}$ involves both $t$ and the spatial coordinates (that are implicit in $\tau^a (x)$). Using Eq.~(\ref{eq2.3}), one finds that the Cartan differential of $\theta^{\hat a}$ reads (no summation on $a$, but summation on $b,c,d$))
\begin{equation}
\label{eq2.46}
d \theta^{\hat a} = \partial_t (e^{-\beta_a} \, S^{\hat a} \, _{b}) \, dt \wedge \tau^b + \frac{1}{2} \, e^{-\beta_a} \, S^{\hat a} \, _d \ C^d \, _{bc} \ \tau^b \wedge \tau^c \, .
\end{equation}
Rewriting the r.h.s. in terms of $\theta^{\hat 0} = N dt$ and $\theta^{\hat a} = e^{-\beta_a} \, S^{\hat a} \, _b \ \tau^b$ yields the structure constants ${\mathcal C}^{\hat a} \, _{\hat 0 \hat b}$ and ${\mathcal C}^{\hat a} \, _{\hat b \hat c}$, namely (no summation on $a,b,c$)
\begin{equation}
\label{eq2.47}
{\mathcal C}^{\hat a} \, _{\hat 0 \hat b} \equiv - {\mathcal C}^{\hat a} \, _{\hat b \hat 0} = - \frac{1}{N} \, \dot\beta_a \, \delta_{ab} + e^{-\beta_a + \beta_b} \, \frac{w^{\hat a}\,_{ \hat b}}{N}
\end{equation}
\begin{equation}
\label{eq2.48}
{\mathcal C}^{\hat a} \, _{\hat b \hat c} = e^{-\beta_a + \beta_b + \beta_c} \, \overline C^{\overline a} \, _{\overline b \overline c} \, ,
\end{equation}
where the $\overline C$ coefficients are the structure constants w.r.t. the intermediate frame $\overline\tau^{\overline a} \equiv S^{\hat a} \, _b \ \tau^b$, as defined in Eq.~(\ref{eq2.9}) above. Again, the choice of the matrix $S$ as belonging to the automorphism group of the Lie algebra ${\mathfrak G}$ implies that the $\overline C$ coefficients are simply equal to the original structure constants $C^a \, _{bc}$.
\smallskip
Inserting the results (\ref{eq2.45}), (\ref{eq2.47}), (\ref{eq2.48}) in Eq.~(\ref{eq2.44})
(and considering the special case of type IX, i.e. $w^T =-w$)
yields the explicit expressions of the connection coefficients, namely
\begin{equation}
\label{eq2.49}
\omega_{\hat 0 \hat b \hat c} = - \frac{1}{2} \, \left[ {\mathcal C}_{\hat b \hat 0 \hat c} + {\mathcal C}_{\hat c \hat 0 \hat b} \right] = \frac{1}{N} \, \dot\beta_b \, \delta_{bc} + \sinh (\beta_b - \beta_c) \, \frac{w^{\hat b \hat c}}{N}
\end{equation}
\begin{equation}
\label{eq2.50}
\omega_{\hat a \hat b \hat 0} = \frac{1}{2} \, \left[ - {\mathcal C}_{\hat a \hat 0 \hat b} + {\mathcal C}_{\hat b \hat 0 \hat a} \right] = -\cosh (\beta_a - \beta_b) \, \frac{w^{\hat a \hat b}}{N} \, ,
\end{equation}
\begin{equation}
\label{eq2.51}
\omega_{\hat a \hat b \hat c} = \frac{1}{2} \, \left( e^{-\beta_a + \beta_b + \beta_c} \, C^a \, _{bc} + e^{-\beta_b + \beta_c + \beta_a} \, C^b \, _{ca} - e^{-\beta_c + \beta_a + \beta_b} \, C^c \, _{ab} \right) \, .
\end{equation}
In the last expressions, one has simply $C^a \, _{bc} = \varepsilon_{abc}$ when using the usual basis $\tau_a$ for type~IX, so that
\begin{equation}
\label{eq2.52}
\omega_{\hat a \hat b \hat c} = \frac{1}{2} \, e^{\beta_1 + \beta_2 + \beta_3} \, (e^{-2\beta_a} + e^{-2\beta_b} - e^{-2\beta_c}) \, \varepsilon_{abc} \, .
\end{equation}
When inserting these results into the Dirac Lagrangian (\ref{eq2.13}), {\it i.e.}
\begin{equation}
\label{eq2.53}
L_D = N \, \sqrt g \left( \overline\Psi \, \frac{\gamma^{\hat 0}}{N} \, \partial_t \, \Psi + \overline\Psi \, \gamma^{\hat\gamma} \, \omega_{\hat\alpha\hat\beta\hat\gamma} \, \frac{\gamma^{\hat\alpha\hat\beta}}{4} \, \Psi - m \, \overline \Psi \Psi \right) \, ,
\end{equation}
there arise several types of terms: (a) a term involving $\partial_t \, \Psi$; (b) some terms involving $\partial_t \, \beta$; (c) terms involving the rotational velocities $w \sim \dot\theta$; (d) terms involving the purely spatial components $\omega_{\hat a \hat b \hat c}$ of the connection; and (e) the term involving the mass $m$ of the spinor field. Let us first note that the lapse cancels out in all the terms involving one time derivative ($\partial_t \, \Psi$, $\partial_t \, \beta$ or $w$), while it contributes a factor $N^{+1}$ in the ``potential'' terms (d) and (e). Let us first focus on the terms (a) and (b), {\it i.e.} those involving either $\partial_t \, \Psi$ or $\partial_t \, \beta$. They are easily found to be
\begin{eqnarray}
\label{eq2.54}
L_D &= &\sqrt g \, \overline\Psi \, \gamma^{\hat 0} \left( \dot\Psi - \frac{1}{2} \, (\dot\beta_1 + \dot\beta_2 + \dot\beta_3) \, \Psi \right) + \ldots \nonumber \\
&= &g^{\frac{1}{2}} \, \overline\Psi \, \gamma^{\hat 0} \left( \dot\Psi + \frac{\partial_t \, g^{1/4}}{g^{1/4}} \, \Psi \right) + \ldots \, ,
\end{eqnarray}
where we introduced the logarithmic derivative of $g^{\frac{1}{4}} = e^{-\frac{1}{2} (\beta_1 + \beta_2 + \beta_3)}$. This shows (as had been used in many previous works), that the replacement of the original spinor variable $\Psi$ by the rescaled spinor
\begin{equation}
\label{eq2.55}
\chi : = g^{\frac{1}{4}} \, \Psi
\end{equation}
disposes of the coupling to the $\dot\beta_a$'s. This rescaling has also the effect of absorbing the prefactor $\sqrt g$ in Eq.~(\ref{eq2.53}) into the various spinor bilinears: $\sqrt g \, \overline\Psi (\ldots) \Psi \equiv \overline\chi (\ldots) \chi$. Note, however, that the lapse prefactor $N$ remains in factor of the bilinears of types (d) and (e), {\it i.e.} those involving no time derivatives.
\smallskip
Finally, after using the rescaling (\ref{eq2.55}) (and remembering the convention (\ref{eq2.15})), we end up with a spinor part of the Lagrangian of the form
\begin{eqnarray}
\label{eq2.56}
L_D &= &i \, \chi^{\dagger} \, \dot\chi - \cosh (\beta_1 - \beta_2) \, w^{\hat 1 \hat 2} \, \Sigma^{\hat 1 \hat 2} - \cosh (\beta_2 - \beta_3) \, w^{\hat 2 \hat 3} \, \Sigma^{\hat 2 \hat 3} \nonumber \\
&&- \, \cosh (\beta_3 - \beta_1) \, w^{\hat 3 \hat 1} \, \Sigma^{\hat 3 \hat 1} - V_{s \, {\rm grav}} - V_{s \, {\rm mass}} \, ,
\end{eqnarray}
where we have introduced the short-hand notation
\begin{equation}
\label{sigma}
\Sigma^{\hat a \hat b} := \frac{1}{2} \, \overline\chi \, \gamma^{\hat 0} \, \gamma^{\hat a \hat b} \, \chi = \frac{i}{2} \, \chi^{\dagger} \, \gamma^{\hat a \hat b} \, \chi
\end{equation}
for the spinor bilinears that couple to the rotational velocities $w^{\hat a \hat b}$ (we used (\ref{eq2.15}) in the last equation). We are including a factor $\frac{1}{2}$ in the definition (\ref{sigma}) so that the hermitian operator $\Sigma^{\hat a \hat b}$ (in the quantum theory) measures the (second quantized) spin of the spinor field $\chi$ (with eigenvalues $\pm \, \frac{1}{2}$ or $0$; see below).
\smallskip
In addition to the (body-frame) ``spin-angular-velocity'' coupling terms $\propto w^{\hat a \hat b} \, \Sigma^{\hat a \hat b}$ in the spinor Lagrangian, there are also (velocity-independent) ``spinor potential terms''\footnote{The classical analogs of these terms were discussed in Ref. \cite{ObRy}, where
$\chi$ was treated as a c-number.}, that are naturally divided into two separate contributions: (i) the spinor potentials $V_{s \, {\rm grav}}$ coming from the coupling to the spatial connection coefficients $\omega_{\hat a \hat b \hat c}$; and (ii) the spinor potential $V_{s \, {\rm mass}}$ coming from the Dirac mass term $m \, \overline\Psi \Psi$. The original expression (from (\ref{eq2.53})) of the gravitational-spinor potential is
\begin{equation}
\label{eq2.57}
V_{s \, {\rm grav}} = - \frac{N}{4} \, \overline\chi \, \gamma^{\hat c} \, \omega_{\hat a \hat b \hat c} \, \gamma^{\hat a \hat b} \, \chi \, .
\end{equation}
Using the gamma identity (where $\gamma^{\hat a \hat b \hat c} \equiv \gamma^{[\hat a}\gamma^{\hat b} \gamma^{\hat c]}$)
\begin{equation}
\label{eq2.58}
\gamma^{\hat c} \, \gamma^{\hat a \hat b} = \gamma^{\hat c \hat a \hat b} + \eta^{\hat c \hat a} \, \gamma^{\hat b} - \eta^{\hat c \hat b} \, \gamma^{\hat a} \, ,
\end{equation}
and the fact that all the traces of $\omega_{\hat a \hat b \hat c}$ vanish (because of the vanishing of $C^a \, _{ab}$), the potential (\ref{eq2.57}) can be rewritten, using (\ref{eq2.44}), as
\begin{equation}
\label{eq2.59}
V_{s \, {\rm grav}} = - \sum_{a,b,c} \, \frac{N}{8} \, \overline\chi \ {\mathcal C}^{\hat a} \, _{\hat b \hat c} \ \gamma^{\hat a \hat b \hat c} \, \chi \, .
\end{equation}
Using $N = \sqrt g = e^{-(\beta_1 + \beta_2 + \beta_3)}$ and (\ref{eq2.48}), this yields,
in any (class A) Bianchi type (with $n^{ab} = {\rm diag} (n_1,n_2,n_3)$)
\begin{equation}
\label{eq2.60bis}
V_{s \, {\rm grav}} = - \frac{1}{4} \, n_1 e^{-2\beta_1} \, \overline\chi \, \gamma^{\hat 1 \hat 2 \hat 3} \, \chi - \frac{1}{4} \, n_2 e^{-2\beta_2} \, \overline\chi \, \gamma^{\hat 2 \hat 3 \hat 1} \, \chi - \frac{1}{4} \, n_3 e^{-2\beta_3} \, \overline\chi \, \gamma^{\hat 3 \hat 1 \hat 2} \, \chi,
\end{equation}
where, for instance, the first term comes from the $C^1 \, _{23}$ structure constant, and is associated with the corresponding $(1;2,3)$ gravitational wall in $\beta$-space, namely $w^g_{1;23} = \beta_1 - \beta_2 - \beta_3 + \underset{a}{\sum} \, \beta_a = 2 \, \beta_1$ (when considering, as we do here, the $3$-dimensional case).
We see (in agreement with Refs.~\cite{dBHP,DaHi}) that the corresponding spinor coupling involves
\begin{equation}
\label{eq2.61}
\overline\chi \, \gamma^{\hat 1 \hat 2 \hat 3} \, \chi = i \, \chi^{\dagger} \, \gamma_{\hat 0} \, \gamma^{\hat 1 \hat 2 \hat 3} \, \chi \, .
\end{equation}
Another peculiar feature of $3$ dimensions is that all the different gravitational walls, $w^g_{1;23}$, $w^g_{2;31}$ and $w^g_{3;12}$, involve the same spinor bilinear $\overline\chi \, \gamma^{\hat 1 \hat 2 \hat 3} \, \chi = \overline\chi \, \gamma^{\hat 2 \hat 3 \hat 1} = \overline\chi \, \gamma^{\hat 3 \hat 1 \hat 2}$. This implies that, for instance in the type~IX case, the spinorial-gravitational coupling explicitly reads
\begin{equation}
\label{eq2.60}
V_{s \, {\rm grav}}^{\rm IX } = - \frac{1}{4} \, ( e^{-2\beta_1} \, + e^{-2\beta_2} \, +e^{-2\beta_3} ) \
\overline\chi \, \gamma^{\hat 1 \hat 2 \hat 3} \, \chi ,
\end{equation}
Finally, the last term in the spinor Lagrangian (\ref{eq2.56}) reads
\begin{equation}
\label{eq2.62}
V_{s \, {\rm mass}} = + \, m \, N \, \sqrt g \ \overline\Psi \, \Psi = m \, N \, \overline\chi \, \chi = m \, e^{-(\beta_1 + \beta_2 + \beta_3)} \, \overline\chi \, \chi \, ,
\end{equation}
with
\begin{equation}
\label{eq2.63}
\overline\chi \, \chi = i \, \chi^{\dagger} \, \gamma_{\hat 0} \, \chi \, .
\end{equation}
Contrary to what happened for $V_{s \, {\rm grav}}$, this term does not correspond to one of the gravitational walls entering $V_g (\beta)$. It would correspond to the wall form
$w_{\Lambda} (\beta)$ associated to a cosmological constant $\Lambda$. Indeed, in the $N = \sqrt g$ gauge, $\Lambda$ generates a term $\propto \Lambda \, N \, \sqrt g = \Lambda \, e^{-2w_{\Lambda} (\beta)}$ with $w_{\Lambda} (\beta) = \underset{a}{\sum} \, \beta_a$. As usual (see, {\it e.g.}, Ref.~\cite{dBHP}) one expects the corresponding ``spinor wall'' to have a halved exponent, {\it i.e.} $\propto e^{-w_{\Lambda} (\beta)}$, which is the case of the spinor mass term (\ref{eq2.62}).
\section{Classical Hamiltonian formulation of the Einstein-Dirac Bianchi system.}
Having obtained a gauge-fixed Lagrangian formulation of the Einstein-Dirac system, let us
now show how to pass to a Hamiltonian formalism.
\subsection{From the Lagrangian to the Hamiltonian in presence of derivative couplings}\label{ssec2.6}
A (well known) peculiarity of the Einstein-Dirac Lagrangian is the presence of {\it derivative couplings} between gravity and the spinor. In our homogeneous-cosmology context (and after the rescaling (\ref{eq2.55}) of the spinor) these derivative couplings are the terms $\propto w^{\hat a \hat b} \, \Sigma^{\hat a \hat b}$ in (\ref{eq2.56}), where $w^{\hat a \hat b}$ are linear in the time derivatives of the Euler angles (see Eq.~(\ref{eq2.42})). A well-known instance of such derivative couplings is given by the coupling of a charged particle to a magnetic field say (suppressing indices)
\begin{equation}
\label{eq2.64}
L(q,\dot q) = \frac{1}{2} \, m \, \dot q^2 + e \, \dot q \, A - e \, V \, .
\end{equation}
Let us recall the effect of the derivative coupling $ e \, \dot q \, A$, when going from the Lagrangian to the Hamiltonian formalism. First, it modifies the relation between the (canonical) momentum and the velocity, namely
\begin{equation}
\label{eq2.65}
p = \frac{\partial L}{\partial \dot q} = m \, \dot q + e \, A \, .
\end{equation}
Second, $e \, A$ cancels out when computing the energy (because ``a magnetic force does no work'')
\begin{equation}
\label{eq2.66}
E(q,\dot q) := p \, \dot q - L = (m \, \dot q + e \, A) \, \dot q - \frac{1}{2} \, m \, \dot q^2 - e \, A \, \dot q + e \, V = \frac{1}{2} \, m \, \dot q^2 + e \, V \, .
\end{equation}
However, third, the $e \, A$ coupling reappears in the Hamiltonian because one must replace $\dot q$ by its expression in terms of $p$:
\begin{equation}
\label{eq2.67}
H(q,p) = [E(q,\dot q)]_{\dot q(p)} = \left[ \frac{1}{2} \, m \, \dot q^2 + e \, V \right]_{\dot q (p)} = \frac{(p-e \, A)^2}{2m} + e \, V \, .
\end{equation}
This mechanism is easily seen to hold for any derivative coupling which is linear in time derivatives. [In a more general case the mass $m$ in (\ref{eq2.64}) becomes some $q$-dependent quadratic form, and the $m^{-1}$ factor in $H(q,p)$ becomes the inverse quadratic form.] The crucial end result is that the Hamiltonian is numerically equal to the sum of the original kinetic energy terms (without the velocity coupling term $e \, \dot q \, A$) and of the potential energy, but with the replacement of the concerned velocities by their expressions in terms of their ($e \, A$-shifted) canonical momenta.
\subsection{Hamiltonian formulation of the Einstein-Dirac system}\label{ssec2.7}
We can now write down the explicit Hamiltonian of the Einstein-Dirac system, for homogeneous configurations, when using the gauge-fixed vielbein (\ref{eq2.4}), (\ref{eq2.5}) (with zero shift vector). More precisely, the Hamiltonian action density has the form
\begin{eqnarray}
\label{eq2.68}
L_{\rm Ham} &= &\sum_a \, \pi_{\beta_a} \, \dot\beta_a + \sum_a p_{\theta_a} \, \dot\theta_a + i \, \chi^{\dagger} \, \dot\chi \nonumber \\
&&- \, \widetilde N \, \widetilde{\mathcal H} (\beta , \pi_{\beta} , \theta , p_{\theta} , \chi^{\dagger} , \chi) - N^a \, {\mathcal H}_a (\beta , \pi_{\beta} , \theta , p_{\theta} , \chi^{\dagger} , \chi) \, ,
\end{eqnarray}
where we have introduced the canonical momenta $\pi_{\beta_a} , p_{\theta_a}$, respectively conjugated to the three $\beta$'s, and to the three Euler angles $\theta_a$, and where $\widetilde N$ denotes the rescaled lapse $\widetilde N := N/\sqrt g$. We shall work in a gauge where $\widetilde N = 1$ and $N^a = 0$, which makes the Hamiltonian in the Hamiltonian action, $L_{\rm Ham} = p \, \dot q - H$, simply equal to $\widetilde{\mathcal H}$. Note that $\widetilde{\mathcal H} \equiv \sqrt g \, {\mathcal H}$ where ${\mathcal H}$ is the usual Arnowitt-Deser-Misner Hamiltonian density entering Eq.~(\ref{eq2.21}) (${\mathcal H}$ is a spatial density of weight $+1$, while $\widetilde{\mathcal H}$ has weight $+2$). Note also that we did not introduce any notation for the conjugate momentum of $\chi$ as the Dirac action is first order, {\it i.e.} already in $p \, \dot q - H$ form (with $p = i \, \chi^{\dagger}$ when $q=\chi$). Using the results above (notably in the last Section) the value of $\widetilde{\mathcal H}$ is the sum of kinetic-energy and potential-energy terms:
\begin{eqnarray}
\label{eq2.69}
\widetilde{\mathcal H} &= &T_{\beta} (\pi_{\beta}) + T_w (\beta , \pi_w , \chi , \chi^{\dagger}) + V_g (\beta) \nonumber \\
&&+ \, V_{s \, {\rm grav}} (\beta , \chi , \chi^{\dagger}) + V_{s \, {\rm mass}} (\beta , \chi , \chi^{\dagger}) \, .
\end{eqnarray}
The various potential terms are the same as written above: $V_g (\beta)$ is given by Eq.~(\ref{eq2.30}), $V_{s \, {\rm grav}} (\beta , \chi , \chi^{\dagger})$ by Eq.~(\ref{eq2.60bis}), and $V_{s \, {\rm mass}} (\beta , \chi , \chi^{\dagger})$ by Eq.~(\ref{eq2.62}). [One should use $\overline\chi := i \, \chi^{\dagger} \, \gamma_{\hat 0}$ to replace $\overline \chi$ in terms of $\chi^{\dagger}$.]
\smallskip
The $\beta$-kinetic-term\footnote{When discussing the $\beta$-kinetic-term in the Lagrangian ($L= T-V$), we do not take out a factor $\widetilde N$,
as we do in discussing the corresponding term in the Hamiltonian.} is originally given (in any dimension, and for any time gauge, i.e. any value of
$\widetilde N := N/\sqrt g$) by
\begin{equation}
\label{eq2.70}
T_{\beta} (\dot\beta) = \frac{1}{\widetilde N} \, G_{ab} \, \dot\beta_a \, \dot\beta_b \equiv \frac{1}{\widetilde N} \,\left[ \sum_a \dot\beta_a^2 - \left( \sum_a \dot\beta_a \right)^2 \right]\, ,
\end{equation}
which defines the $\beta$-space metric $G_{ab}$. [Though we put the $a$ index on $\beta$
as a subscript, one should think of it as a contravariant index $\beta^a$.]
After the rescaling (\ref{eq2.55}) of the spinor field, $\dot\beta$ appears only in the $\beta$ kinetic term, so that the conjugate momenta $\pi_{\beta_a}$ to the $\beta_a$'s are given by
\begin{equation}
\label{eq2.71}
\pi_{\beta_a} = \frac{2}{\widetilde N} \, G_{ab} \, \dot\beta_b \, .
\end{equation}
This leads to the following value for the $\beta$ kinetic energy expressed in terms of $\pi_{\beta}$
\begin{equation}
\label{eq2.72}
T_{\beta} (\pi_{\beta}) = \frac{1}{4} \, G^{ab} \, \pi_{\beta_a} \, \pi_{\beta_b} \, ,
\end{equation}
where $G^{ab}$ is the inverse of $G_{ab}$, {\it i.e.} (in our $3+1$ dimensional case)
\begin{equation}
\label{eq2.73}
G^{ ab} \, \pi_{\beta_a} \, \pi_{\beta_b} = \sum_a \pi_{\beta_a}^2 - \frac{1}{2} \left( \sum_a \pi_{\beta_a}\right)^2 \, .
\end{equation}
Note that, if we were in a space dimension $d \ne 3$, the coefficient of the second term on the r.h.s. of (\ref{eq2.73}) would be $\frac{1}{d-1}$ instead of $\frac{1}{2}$.
\smallskip
It remains to discuss the rotational kinetic term $T_w$ linked to the angular motion parametrized by the three Euler angles $\theta_a$. To simplify
this discussion, we shall henceforth come back to using the time gauge (\ref{gauge}), as we did in Section 2.
[The formulas above for the $\beta$ kinetic terms
were written in a general time gauge as a reminder of the various occurrences of $N$ and $\sqrt g$ in the
kinetic part of the action.]
As the rotational velocities $w^{\hat a \hat b} \sim \dot\theta_c$ are linearly coupled to $\chi$ {\it via} the spinor bilinears $\Sigma^{\hat a \hat b}$ (see Eq.~(\ref{eq2.56})), we must apply the result of the previous Section to this term. To ease the writing of the Hamiltonian version of the rotational kinetic term $T_w$ it is convenient to introduce the following momentum-like variables
\begin{equation}
\label{eq2.74}
\pi_{w^{ab}} := \frac{\partial L}{\partial \, w^{ab}}
\end{equation}
{\it i.e.} explicitly (in the type IX case)
\begin{equation}
\label{eq2.75}
\pi_{w^{ab}} = 4 \sinh^2 (\beta_a - \beta_b) \, w^{ab} - \cosh (\beta_a - \beta_b) \, \Sigma^{ab} \, ,
\end{equation}
as deduced from Eqs.~(\ref{eq2.39}) and (\ref{eq2.56}). Here, and henceforth, we have simplified the notation by dropping the carets over the (``flat'') indices $a,b$ on $w^{ab}$ and $\Sigma^{ab}$. [Note that the partial derivatives in Eq.~(\ref{eq2.74}) are done w.r.t. the three {\it restricted} independent components $w^{12}$, $w^{23}$, $w^{31}$, of $w$; the other ones being defined in terms of these by, {\it e.g.}, $w^{21} := -w^{12}$, {\it etc.}]
\smallskip
Solving Eq.~(\ref{eq2.75}) for the $w$'s as functions of the $\pi_w$'s, then yields the following Hamiltonian version $(\sim (p-eA)^2 / (2m))$ of the rotational kinetic term
\begin{eqnarray}
\label{eq2.76}
T_w (\beta , \pi_w , \chi , \chi^{\dagger}) &= &\frac{1}{8 \sinh^2 (\beta_1 - \beta_2)} \, [\pi_{w^{12}} + \cosh (\beta_1 - \beta_2) \, \Sigma^{12}]^2 \nonumber \\
&+ &\frac{1}{8 \sinh^2 (\beta_2 - \beta_3)} \, [\pi_{w^{23}} + \cosh (\beta_2 - \beta_3) \, \Sigma^{23}]^2 \nonumber \\
&+ &\frac{1}{8 \sinh^2 (\beta_3 - \beta_1)} \, [\pi_{w^{31}} + \cosh (\beta_3 - \beta_1) \, \Sigma^{31}]^2 \, .
\end{eqnarray}
If one wanted to express the Hamiltonian ${\mathcal H}$ completely in terms of canonically conjugated variables, one should replace the non-canonical, momentum-like variables $\pi_w$ by the linear combinations of the Euler-angle conjugate momenta $p_{\theta_a}$ (see Eq.~(\ref{eq2.68})) that they represent. As $\pi_w \, w = w \, \partial L / \partial \, w = \dot\theta \, \partial L / \partial \, \dot\theta$ $= p_{\theta} \, \dot\theta$, the transpose of the matrix $A$ appearing on the r.h.s. of (\ref{eq2.42}) yields the link $p_{\theta} = A^T \pi_w$, namely
\begin{eqnarray}
\label{eq2.77}
p_{\varphi} &= &\cos \theta \, \pi_{w^{12}} + \sin \theta \sin \psi \, \pi_{w^{23}} + \sin \theta \cos \psi \, \pi_{w^{31}} \, , \nonumber \\
p_{\theta} &= &\cos \psi \, \pi_{w^{23}} - \sin \psi \, \pi_{w^{31}} \, , \nonumber \\
p_{\psi} &= &\pi_{w^{12}} \, ,
\end{eqnarray}
whose inverse reads
\begin{eqnarray}
\label{eq2.78}
\pi_{w^{12}} &= &p_{\psi} \, , \nonumber \\
\pi_{w^{23}} &= &\sin \psi \, \frac{p_{\varphi} - \cos \theta \, p_{\psi}}{\sin \theta} + \cos \psi \, p_{\theta} \, , \nonumber \\
\pi_{w^{31}} &= &\cos \psi \, \frac{p_{\varphi} - \cos \theta \, p_{\psi}}{\sin \theta} - \sin \psi \, p_{\theta} \, .
\end{eqnarray}
The canonical Poisson brackets $\{ \theta_a , \theta_b \} = 0 = \{p_{\theta_a} , p_{\theta_b} \}$ and $\{ \theta_a , p_{\theta_b} \} = \delta_{ab}$ are easily found to imply the following brackets for the non-canonical variables $\pi_w$:
\begin{eqnarray}
\label{eq2.79}
\{ \pi_{w^{12}} , \pi_{w^{23}} \} &= &- \, \pi_{w^{31}} \, , \nonumber \\
\{ \pi_{w^{23}} , \pi_{w^{31}} \} &= &- \, \pi_{w^{12}} \, , \nonumber \\
\{ \pi_{w^{31}} , \pi_{w^{12}} \} &= &- \, \pi_{w^{23}} \, .
\end{eqnarray}
Note that these are the {\it opposite-sign} brackets, compared to those of a usual angular momentum vector: $\{ L_x , L_y \}$ $= + \, L_z$, {\it etc.} Indeed, in the analogy of (\ref{eq2.76}) (without the spin term) with the Hamiltonian of an asymmetric top, {\it i.e.}
\begin{equation}
\label{eq2.80}
H_{\rm top} = \frac{L_1^2}{2 I_1} + \frac{L_2^2}{2 I_2} + \frac{L_3^2}{2 I_3} \, ,
\end{equation}
the quantities $L_1 , L_2 , L_3$ (which are analogous to $\pi_{w^{23}} , \pi_{w^{31}} , \pi_{w^{12}}$) represent the (non-conserved) body-frame components of the angular momentum, which have {\it opposite-type} brackets, namely $\{ L_1 , L_2 \} = - L_3$, {\it etc.} This change of sign is linked to the fact that the body-frame $L_1 , L_2 , L_3$ are related to the (normal-bracket, and conserved) space-frame $L_x , L_y , L_z$ by the time-dependent matrix $S$ effecting the rotation between the space-frame and the body-frame.
\smallskip
For most purposes, the expression (\ref{eq2.76}) of the rotational energy in terms of the $\pi_w$'s, together with the knowledge of the Poisson brackets (\ref{eq2.79}) between the $\pi_w$'s, suffices to discuss the Hamiltonian dynamics of the system.
\smallskip
Summarizing: the total Hamiltonian of the gravity-spinor system is given by Eq.~(\ref{eq2.69}), together with Eq.~(\ref{eq2.72}) $[T_{\beta} (\pi_{\beta})]$, Eq.~(\ref{eq2.76}) $[T_w (\beta , \pi_w , \chi)]$, Eq.~(\ref{eq2.30}) $[V_g (\beta)]$, Eq.~(\ref{eq2.60}) $[V_{s \, {\rm grav}} (\beta , \chi)]$ and Eq.~(\ref{eq2.62}) $[V_{s \, {\rm mass}} (\beta , \chi)]$. The Poisson brackets between the various variables is given by canonical pairs appearing in the $p \dot q$ terms in Eq.~(\ref{eq2.68}). [See also (\ref{eq2.79}).] When considering ``classical'' spinor variables, one should use an odd, Grassmann-valued $\chi$, and define a graded (anti-) bracket under which $\chi$ is conjugated to $i \, \chi^{\dagger}$. The quantum case will be discussed in detail below.
\subsection{Diffeomorphism constraints}\label{ssec2.8}
For simplifying the calculations, we have assumed above that we were working in a gauge where the shift vector $N^a$ vanishes. However, as recalled in Eq.~(\ref{eq2.68}), when we relax this assumption we know that $N^a$ will simply enter as a Lagrange multiplier of the diffeomorphism constraint ${\mathcal H}_a \approx 0$, where (see Eq.~(\ref{eq2.22})
\begin{equation}
\label{eq2.81}
{\mathcal H}_a = \sqrt g \, (2 \ ^4R_a^0 - T_a^0) = -2 \, D^b \, \pi_{ab}^{\rm kin} - \sqrt g \, T_a^0 \, .
\end{equation}
Here, the tensorial objects refer to the spatial metric $g_{ab}$ ({\it e.g.} $D_c \, g_{ab} \equiv 0$), and all indices are projected on the co-frame $\tau^a$, and its dual. In addition, $\pi_{\rm kin}^{ab}$ denotes the kinematical part of the conjugate momentum $\pi^{ab}$ of the metric $g_{ab}$. In the electromagnetic model discussed above, this would be the $m \, \dot q$ part, without the $e \, A$ shift in Eq.~(\ref{eq2.65}). In our case, $\pi_{\rm kin}^{ab}$ denotes the part of $\pi^{ab}$ proportional to the second fundamental form $K$, {\it i.e.} (in covariant form) $\pi_{ab}^{\rm kin} =\sqrt{g} ( K_{ab} - g_{ab} \, K^c \, _c)$, when using the $+ \, \dot g_{ab}$ convention (\ref{eq2.24}) for the definition of $K_{ab}$. Finally, the matter term in (\ref{eq2.81}) comes, in our case, from the stress-energy tensor of the spinor field. From Ref.~\cite{HenIHP}, the explicit expression of the r.h.s. of (\ref{eq2.81}) for a Bianchi spacetime reads
\begin{equation}
\label{eq2.82}
{\mathcal H}_a = 2 \, \pi_{\rm kin}^{bd} \, g_{bc} \, C^c \, _{ad} + \frac{1}{2} \, \Sigma^d \, _c \ C^c \, _{ad} \, ,
\end{equation}
where $\Sigma^d \, _c := h^d \, _{\hat a} \ h_{\hat b c} \, \Sigma^{\hat a \hat b}$.
\smallskip
Inserting the decomposition $h^{\hat a} \, _b = e^{-\beta_a} \, S^{\hat a} \, _b$ of the dreibein (with $g_{cd} = \delta_{\hat a \hat b} \, h^{\hat a} \, _c \ h^{\hat b} \, _d$), and transforming all tensor indices to the intermediate frame $\overline\tau^{\overline a} = S^{\hat a} \, _b \ \tau^b$, except the indices on $\Sigma^{\hat a \hat b}$ which must remain ``flat'' ({\it i.e.} referring to $\theta^{\hat a} = e^{-\beta_a} \, \overline\tau^{\overline a}$), one finds that (\ref{eq2.82}) becomes
\begin{equation}
\label{eq2.83}
\overline{\mathcal H}_{\overline m} := S^a \, _{\hat m} \ {\mathcal H}_a = \sum_{\overline p , \overline q} \left( 2 \, \overline\pi_{\rm kin}^{\overline p \overline q} \, e^{-2 \beta_p} + \frac{1}{2} \, e^{\beta_q - \beta_p} \, \Sigma^{\hat q \hat p} \right) \overline C^{\overline p} \, _{\overline m \overline q} \, .
\end{equation}
Note that, because of the covectorial nature of ${\mathcal H}_a$, the definition of $\overline{\mathcal H}_{\overline m}$ involves the inverse of the matrix $S^{\hat a} \, _b$. The automorphism property of the matrix $S$ guarantees that the $\overline\tau$-frame structure constants $\overline C$ entering (\ref{eq2.83}) are numerically equal to the original, $\tau$-frame, structure constants $C$. In addition, the rewriting of the $\pi^{ab} \, \dot g_{ab}$ term in the Hamiltonian action as $\underset{a}{\sum} \, \pi_{\beta_a} \, \dot\beta_a + \pi_{w^{12}} \, w^{12} + \pi_{w^{23}} \, w^{23} + \pi_{w^{31}} \, w^{31}$, shows that the components $\overline\pi_{\rm kin}^{\overline 1 \, \overline 2} = \overline\pi_{\rm kin}^{\overline 2 \, \overline 1}$ are linked to $\pi_{w^{12}}^{\rm kin}$ {\it via}
\begin{equation}
\label{eq2.84}
\pi_{w^{12}}^{\rm kin} = 2 \, (e^{-2 \beta_1} - e^{-2 \beta_2}) \, \overline\pi_{\rm kin}^{\overline 1 \, \overline 2} \, ,
\end{equation}
and similarly for the cyclic permutations $23$ and $31$. This implies that Eq.~(\ref{eq2.83}) reads, say for $\overline m = \overline 3$,
\begin{equation}
\label{eq2.85}
\overline{\mathcal H}_{\overline 3} = - \, \pi_{w^{12}}^{\rm kin} + \cosh (\beta_1 - \beta_2) \, \Sigma^{\hat 1 \hat 2} \, .
\end{equation}
From Eq.~(\ref{eq2.75}), we see that the last, spin contribution in (\ref{eq2.85}) has the correct magnitude and sign to transform the kinematic contribution $\pi_{w^{12}}^{\rm kin} \equiv 4 \sinh^2 (\beta_1 - \beta_2) \, w^{12}$ into the full momentum-like variable $\pi_{w^{12}}$. Finally, this shows
(in agreement with \cite{Jantzen:1982je}) that $\overline{\mathcal H}_{\overline 3} = - \, \pi_{w^{12}}$, {\it i.e.}, more explicitly
\begin{equation}
\label{eq2.86}
S^a \, _{\hat m} \ {\mathcal H}_a = - \, \pi_{w^{\hat p \hat q}} \, ,
\end{equation}
where $\hat m \, \hat p \, \hat q$ is a cyclic permutation of $123$. In other words, in the analogy where $\pi_{w^{23}} = L_1$, {\it etc.} are the {\it body-frame} angular momenta, we have that ${\mathcal H}_1 = - \, L_x$, {\it etc.} are (minus), the {\it space-frame} angular momenta. As in the asymmetric top situation, the ${\mathcal H}_a$, contrary to the $\pi_w$, are conserved because of the first-class bracket $\{ {\mathcal H}_a , {\mathcal H} \} = 0$. Let us also note that the diffeomorphism constraints satisfy the Poisson-bracket algebra
\begin{equation}
\label{eq2.87}
\{ {\mathcal H}_a , {\mathcal H}_b \} = \, - \, C^c \, _{ab} \ {\mathcal H}_c \, .
\end{equation}
[Here, the minus sign comes from the minus sign in the relation ${\mathcal H}_1 = - \, L_x$.]
The simple link (\ref{eq2.86}) between the diffeomorphism constraints and the (body-frame) angular momenta $\pi_w$ is not an accident but derives from basic symmetry properties. Indeed, while the three quantities ${\mathcal H}_a$ generate (in phase space) adjoint transformations of the homogeneity group $G$ \cite{HenIHP}, the three angular momenta $\pi_w$ generate, by definition, the group of the matrices $S(\theta_1 , \theta_2 , \theta_3)$ used in the parametrization (\ref{eq2.5}), (\ref{eq2.6}). However, for simple groups, such as type~IX and type~VIII, the latter group coincides with the adjoint representation of $G$.
\subsection{Explicit forms of some (class A) Bianchi Hamiltonians}\label{ssec2.9}
The full Hamiltonian action (\ref{eq2.68}) implies two types of constraints: (i) the diffeomorphism constraints ${\mathcal H}_a \approx 0$ linked to the arbitrariness of the shift vector $N^a$, and (ii) the Hamiltonian constraint $\widetilde{\mathcal H} \approx 0$ linked to the arbitrariness of the rescaled lapse $\widetilde N = N/\sqrt g$. When considering general (class A) Bianchi models, the expression and the number of effective diffeomorphism constraints strongly depend on the structure constraints of the homogeneity group. Let us summarize the results for the various Bianchi types.
\smallskip
In the Bianchi type~I case, {\it i.e.} when $C^a \, _{bc} = 0$, we see on the general expression (\ref{eq2.82}) that ${\mathcal H}_a \equiv 0$, {\it i.e.} that there are no diffeomorphism constraints. In that case, one can still conventionally decide to use the decomposition (\ref{eq2.6}) with a matrix $S (\theta_1 , \theta_2 , \theta_3) \in SO(3)$. One then ends up with an Hamiltonian action of the form (\ref{eq2.68}), except for the last term $N^a \, {\mathcal H}_a$ which is absent. This type~I action describes the coupled dynamics of the variables $\beta_1 , \beta_2 , \beta_3$, $\theta_1 , \theta_2 , \theta_3$, and $\chi$. The explicit form of the Hamiltonian is given (when $\widetilde N = N/\sqrt g$) by the r.h.s. of (\ref{eq2.69}), where, however, several terms vanish because of the vanishing of the $C$'s. Specifically, the two potential terms in (\ref{eq2.69}) linked to gravitational walls, namely $V_g (\beta)$ and $V_{s \, {\rm grav}} (\beta , \chi)$, are identically zero in type~I. Finally, the type~I Hamiltonian has the form
\begin{equation}
\label{eq2.88}
\widetilde{\mathcal H}^I = T_{\beta} (\pi_{\beta}) + T_w (\beta , \pi_w , \chi , \chi^{\dagger}) + V_{s \, {\rm mass}} (\beta , \chi , \chi^{\dagger})
\end{equation}
where $T_{\beta} (\pi_{\beta})$ is (universally) given by Eq.~(\ref{eq2.72}), $V_{s \, {\rm mass}} (\beta , \chi)$ by Eq.~(\ref{eq2.62}), and where $T_w (\beta , \pi_w , \chi)$ is given by the full expression (\ref{eq2.76}). Note that this Hamiltonian (which must be submitted to the single constraint $\widetilde{\mathcal H} \approx 0$) describes the coupled dynamics of the variables $\beta_1 , \beta_2 ,\beta_3$; $\varphi , \theta , \psi$, and $\chi$. It is a generalized asymmetric top dynamics (for the rotational degrees of freedom $\varphi , \theta , \psi$), where the moments of inertia $I_3 (\beta) = 4 \sinh^2 (\beta_1 - \beta_2)$, {\it etc.} have their own (coupled) dynamics, and which includes a coupling between the (bosonic) rotor and spinor degrees of freedom. As our main physical focus here concerns type~IX, we shall not discuss in detail the type~I dynamics. Let us only note that Ref.~\cite{BK} has explicitly solved the equations of motion of type~I dynamics, when using a different way of fixing the rotational state of the dreibein, and treating $\chi$ as a classical quantity.
\smallskip
The Bianchi type~II case cannot be directly described by the expressions derived above from the $SO(3)$ parametrization (\ref{eq2.6}) because (contrary to the type~I and type~IX cases) its structure constants are not invariant under the full $SO(3)$ group. Refs.~\cite{Jantzen:1982je,Jantzen79,Jantzen:2001me} has indicated other useful choices of three-dimensional matrix groups leaving invariant the $C$'s. To give a different perspective on the type~II case, and relate it to recent work on cosmological billiards, we treat the type~II case in an Appendix by using an Iwasawa decomposition of the metric, {\it i.e.} an upper triangular matrix $S$ in Eqs.~(\ref{eq2.5}) and (\ref{eq2.6}).
\smallskip
Leaving aside a discussion of the other {\it degenerate} class-A Bianchi types ({\it i.e.} VI$_0$ and VII$_0$), let us come back to the main focus of this work, namely the non-degenerate types~IX and VIII. In both cases, the link (\ref{eq2.86}) between the three diffeomorphism constraints, and the three body-frame angular momenta $\pi_w$, holds. [As mentioned above, in type~VIII the rotational angles and momenta refer to $SO(1,2)$ with metric ${\rm diag} (+1,-1,-1)$, rather than to $SO(3)$ with ${\rm diag} (+1,+1,+1)$.] We can therefore greatly simplify the dynamics by reducing the Hamiltonian by imposing the (first class) constraints ${\mathcal H}_a = 0$, {\it i.e.} $\pi_w = 0$, thereby eliminating both the $p_{\theta} \, \dot\theta$ terms in Eq.~(\ref{eq2.68}) and the $N^a \, {\mathcal H}_a$ contribution. This leads to a reduced Hamiltonian action of the simpler form
\begin{equation}
\label{eq2.89}
L_{\rm Ham}^{\rm VIII,IX} = \sum_a \pi_{\beta_a} \, \dot\beta_a + i \, \chi^{\dagger} \, \dot\chi - \widetilde N \, \widetilde{\mathcal H}^{\rm VIII,IX} (\beta , \pi_{\beta} , \chi^{\dagger} , \chi) \, ,
\end{equation}
where ($Z = {\rm VIII}, {\rm IX}$ being a label for the Bianchi type)
\begin{equation}
\label{eq2.90}
\widetilde{\mathcal H}^Z := T_{\beta} (\pi_{\beta}) + T_w^{(0)} (\beta , \chi) + V_g^Z (\beta) + V_{s \, {\rm grav}}^Z (\beta , \chi) + V_{s \, {\rm mass}} (\beta , \chi) \, .
\end{equation}
Here: $T_{\beta} (\pi_{\beta})$ and $V_{s \, {\rm mass}} (\beta , \chi)$ have the universal structure given above; $V_g^Z (\beta)$ is given for any Bianchi type (including degenerate ones) by Eq.~(\ref{eq2.29}); $V_{s \, {\rm grav}}^Z$ is given for any Bianchi type by Eq.~(\ref{eq2.59}), with ${\mathcal C}^{\hat a} \, _{\hat b \hat c}$ given by Eq.~(\ref{eq2.48}) (with $\overline C = C$ when $S$ leaves the $C$'s invariant, as is the case for $Z = {\rm VIII}$ and IX); and, finally, the rotational energy term takes, in type IX, the simplified form obtained by replacing $\pi_w \to 0$ in (\ref{eq2.76})
\begin{eqnarray}
\label{eq2.91}
T_w^{(0){\rm IX}} (\beta , \chi) &= &\frac{1}{8} \, \coth^2 (\beta_1 - \beta_2) (\Sigma^{12})^2 + \frac{1}{8} \coth^2 (\beta_2 - \beta_3)(\Sigma^{23})^2 \nonumber \\
&&+ \, \frac{1}{8} \coth^2 (\beta_3 - \beta_1) (\Sigma^{31})^2 \, .
\end{eqnarray}
The type VIII case is obtained by particularizing to the case $ (n_1, n_2, n_3) = (+1,+1,-1)$
the result for a general $n^{ab} = {\rm diag} (n_1, n_2, n_3)$, which reads
\begin{equation}
\label{eq2.92}
T_w^{(0)} (\beta , \chi) = \frac{1}{8} \, \left( \frac{ n_1 e^{\beta_2-\beta_1} + n_2 e^{\beta_1-\beta_2} }{ n_1 e^{\beta_2-\beta_1} - n_2 e^{\beta_1-\beta_2} } \right)^2 (\Sigma^{12})^2 + {\rm cyclic} \, .
\end{equation}
The Hamiltonian (\ref{eq2.90})--(\ref{eq2.92}) describes the coupled dynamics of the diagonal degrees of freedom of the metric coupled to a spinor. In the next Section we shall discuss its quantization.
\section{Quantum formulation of the coupled spinor-Bianchi-IX system}\label{sec3}
\setcounter{equation}{0}
\subsection{Dependence on the Euler angles}\label{ssec3.1} We shall denote by $\Phi$ the abstract quantum state of the system. Following Dirac, we interpret the various classical constraints, ${\mathcal H} \approx 0$, ${\mathcal H}_a \approx 0$ as constraints on $\Phi$ of the type
\begin{equation}
\label{eq3.1}
\widehat{\mathcal H} \, \Phi = 0 \, ,
\end{equation}
\begin{equation}
\label{eq3.2}
\widehat{\mathcal H}_a \, \Phi = 0 \, ,
\end{equation}
where $\widehat{\mathcal H}$ and $\widehat{\mathcal H}_a$ are suitably defined operatorial versions of ${\mathcal H}$ and ${\mathcal H}_a$. We shall work in the gravity configuration space $\beta_1 , \beta_2 , \beta_3$; $\theta_1 , \theta_2 , \theta_3$, together with a suitable description of the spinor degrees of freedom, labelled, say, by $\sigma$ (see below). This leads to a wave function of the universe described by $\Phi (\beta , \theta , \sigma)$. Let us start by remarking that the classical link (\ref{eq2.86}) naturally suggests an ordering such that Eq.~(\ref{eq3.2}) becomes
\begin{equation}
\label{eq3.3}
\widehat\pi_{w^{ab}} \, \Phi (\beta , \theta , \sigma) = 0 \, .
\end{equation}
The classical angular momenta $\pi_w$ are linear combinations of the canonical angular momenta $p_{\theta_a}$, see Eq.~(\ref{eq2.78}). The canonical quantization $\widehat p_{\theta_a} = -i \, \partial / \partial \, \theta_a$ (with $\hbar = 1$), together with the natural ordering of the $\widehat\pi_w$ as differential operators on the $SO(3)$ space\footnote{Note that the $\theta$'s parametrize $SO(3)$ matrices, rather than the $SU(2)$ group $G$ which constitutes the cosmological space. The ``factor $2$'' difference between $SO(3)$ and $SU(2)$ might have been important, if we had needed to consider situations where $\widehat\pi_w \, \Phi \ne 0$.} naturally leads to concluding that the quantum constraint (\ref{eq3.3}) means that $\Phi (\theta_1 , \theta_2 , \theta_3)$ does not depend on the Euler angles: $\partial_{\theta_a} \Phi = 0$. More directly, we can say that Eq.~(\ref{eq3.3}) means that $\Phi$ is a zero-angular-momentum state on $SO(3)$; {\it i.e.} an ``$s$-wave''. As $\Phi$ does not depend on the $\theta$'s, we can simply ignore them in the following. Note that this is consistent with having started the quantization procedure at the level of the final, reduced type~IX Hamiltonian, namely Eq.~(\ref{eq2.90}), which does not contain the $\theta$'s. From this point of view, the quantum description is cleaner than the classical one. Indeed, in the classical description the Euler angles satisfy first-order differential equations obtained by setting to zero the l.h.s. of Eq.~(\ref{eq2.75}). This yields a non-trivial dynamics for the $\theta$'s sourced by the spinor bilinears $\propto \Sigma^{ab} (t)$, and influenced by the time-dependence of the $\beta$'s. In the quantum description, the $\theta$'s do not appear at all, which is nicely consistent with the fact that they have the character of gauge parameters.
\subsection{Quantum description of the spinor degrees of freedom}\label{ssec3.2}
After the rescaling (\ref{eq2.55}) of the spinor, the spinor kinetic term is simply
\begin{equation}
\label{eq3.4}
L_s^{\rm kin} = i \, \chi^{\dagger} \dot\chi \equiv i \sum_{\alpha = 1}^4 \chi_{\alpha}^{\dagger} (t) \, \dot \chi_{\alpha} (t)
\end{equation}
where we made explicit the Dirac-spinor index $\alpha = 1,2,3,4$, which was kept implicit up to now. In the case where $\chi$ is a generic {\it Dirac spinor}, with $4$ independent {\it complex} components, the quantization of (\ref{eq3.4}) is done by the standard anticommutator result (with $\hbar = 1$)
\begin{equation}
\label{eq3.5}
[\chi_{\alpha} , \chi_{\beta}]_+ = 0 = [\chi_{\alpha}^{\dagger} , \chi_{\beta}^{\dagger}]_+ \, ; \quad [\chi_{\alpha} , \chi_{\beta}^{\dagger}]_+ = \delta_{\alpha\beta} \, ,
\end{equation}
where $[A,B]_+ \equiv AB + BA$ denotes an anticommutator, and where all the $\chi$'s are taken at the same instant $t$. In other words, each one of the four $\chi_{\alpha}$'s can be viewed (at each instant $t$) as an independent (complex) fermionic destruction operator, with $\chi_{\alpha}^{\dagger}$ being the corresponding fermionic creation operator. To each pair $(\chi_{\alpha} , \chi_{\alpha}^{\dagger})$ then corresponds a two-dimensional (complex) Hilbert space, so that to the $4$ mutually anticommuting pairs $(\chi_{\alpha} , \chi_{\alpha}^{\dagger})$, $\alpha = 1,\ldots , 4$ correspond a fermionic Hilbert space of dimension $2^4 = 16$.
\smallskip
We can, however, simplify the problem by demanding, from the start, that $\chi$ be a {\it Majorana spinor}, {\it i.e.} be restricted to contain only $4$ independent {\it real} components. More precisely, let us use the following explicit Majorana representation of the $4$ Dirac gamma matrices $\gamma^{\widehat\alpha}$ entering the spinor action:
$$
\gamma^{\widehat 0} = \begin{pmatrix} 0 &i \, \sigma_2 \\ i \, \sigma_2 &0 \end{pmatrix} \, , \ \gamma^{\widehat 1} = \begin{pmatrix}\sigma_1 &0 \\ 0 &\sigma_1 \end{pmatrix} \, , \ \gamma^{\widehat 2} = \begin{pmatrix}-\sigma_3 &0 \\ 0 &-\sigma_3 \end{pmatrix} \, ,
$$
\begin{equation}
\label{eq3.6}
\gamma^{\widehat 3} = \begin{pmatrix} 0 &-i \, \sigma_2 \\ i \, \sigma_2 &0 \end{pmatrix} \, ,
\end{equation}
where $\sigma_1 = \begin{pmatrix} 0 &1 \\ 1 & 0 \end{pmatrix}$, $\sigma_2 = \begin{pmatrix} 0 &-i \\ i &0 \end{pmatrix}$, $\sigma_3 = \begin{pmatrix} 1 &0 \\ 0 &-1 \end{pmatrix}$ are the standard Pauli matrices. The four gamma matrices (\ref{eq3.6}) are all real, with the 3 spatial $\gamma$'s being symmetric (and therefore hermitian), while $\gamma^{\hat 0}$ is anti-symmetric ({\it i.e.} anti-hermitian).
\smallskip
When using a real representation such as (\ref{eq3.6}), the reality condition for a Majorana spinor is simply $\chi^* = \chi$, {\it i.e.} the condition that each component $\chi_{\alpha}$ be real ({\it i.e.} hermitian, as an operator). In that case, the quantization of the standard spinor kinetic term (\ref{eq3.4}), {\it i.e.} $i \, \underset{\alpha}{\sum} \, \chi_{\alpha} \, \dot\chi_{\alpha}$, involves an extra factor $\frac{1}{2}$, namely
\begin{equation}
\label{eq3.7}
[\chi_{\alpha} , \chi_{\beta}]_+ \equiv \chi_{\alpha} \, \chi_{\beta} + \chi_{\beta} \, \chi_{\alpha} = \frac{1}{2} \, \delta_{\alpha\beta} \, .
\end{equation}
The need for the factor $\frac{1}{2}$ in the quantization condition (\ref{eq3.7}) can be viewed in several ways. The most direct way is that the normalization of the basic field commutators should be chosen so that the universal Heisenberg-representation equation of motion for operators (with $\hbar = 1$)
\begin{equation}
\label{eq3.8}
i \ \frac{d}{dt} \, Q = [Q,H]
\end{equation}
(where $[A,B] \equiv AB-BA$ denotes a commutator) hold true, and be consistent with the usual Euler-Lagrange equations of motion. As the Euler-Lagrange variation of the kinetic term $\delta \int dt \, i \, \chi \, \dot\chi$ (for a real, anticommuting $\chi$) involves a factor 2 (after integrating by parts), one needs the factor $\frac{1}{2}$ in (\ref{eq3.7}).
\smallskip
An indirect check consists of decomposing a general, complex Dirac spinor, with kinetic term (\ref{eq3.4}), into real ({\it i.e.} Majorana) and imaginary parts, say $\chi = \chi_R + i \, \chi_I$, $\chi^{\dagger} = \chi_R - i \, \chi_I$, where $\chi_R$ and $\chi_I$ are both Majorana. It is then easily checked that the standard anticommutators (\ref{eq3.5}) imply that $\chi_R$ and $\chi_I$ both satisfy (\ref{eq3.7}), and mutually anti-commute. In addition, it is easily seen that if we insert the decomposition $\chi = \chi_R + i \, \chi_I$ in the full Dirac action (\ref{eq2.56}), it simply decomposes as the sum of two {\it decoupled} Majorana actions, one involving $\chi_R$ and one involving $\chi_I$. [This is true because all the spinor bilinears, including the mass term $\propto \chi^{\dagger} \, \gamma_{\widehat 0} \, \chi$, involve an antisymmetric Clifford-algebra matrix, such as $\gamma^{\widehat 1} \, \gamma^{\widehat 2}$, $\gamma_{\widehat 0} \, \gamma^{\widehat 1 \, \widehat 2 \, \widehat 3}$ or $\gamma_{\widehat 0}$, sandwiched between $\chi^{\dagger}$ and $\chi$.] This shows that considering a complex Dirac spinor source is equivalent to coupling gravity to the sum of two independent Majorana spinor sources. At the Hilbert-space level, the Dirac fermionic state space is just the tensor product of two independent Majorana fermionic state spaces. For simplicity, we shall consider in the following the simpler, single Majorana case, {\it i.e.} $\chi^* = \chi$. Note, in particular, that in that case the fermionic Hilbert space is simply of dimension $2^2 = 4$. Indeed, this fermionic Majorana space is the (irreducible) representation space of the algebra (\ref{eq3.7}), which is simply a {\it Clifford algebra} on the $4$-dimensional Euclidean space. Actually, if we introduce the notation
\begin{equation}
\label{eq3.9}
\Gamma_{\alpha} := 2 \, \chi_{\alpha} \, , \qquad \alpha = 1,\ldots , 4
\end{equation}
we see that the four operators (in Hilbert space) $\Gamma_{\alpha}$, $\alpha = 1,\ldots , 4$, satisfy the standard $O(4)$ Clifford algebra relations
\begin{equation}
\label{eq3.10}
\Gamma_{\alpha} \, \Gamma_{\beta} + \Gamma_{\beta} \, \Gamma_{\alpha} = 2 \, \delta_{\alpha\beta} \, .
\end{equation}
In particular, each $\Gamma_{\alpha}$ has a unit square. In addition, note that the product
\begin{equation}
\label{eq3.11}
\Gamma_5 := \Gamma_1 \, \Gamma_2 \, \Gamma_3 \, \Gamma_4 \equiv 16 \, \chi_1 \, \chi_2 \, \chi_3 \, \chi_4 \, ,
\end{equation}
defines an operator which anticommutes with each $\chi_{\alpha}$, and which has also a unit square: $\Gamma_5^2 = +1$. [This contrasts with the $O(3,1)$ case where $\gamma_5 \equiv \gamma_{\widehat 0} \, \gamma_{\widehat 1} \, \gamma_{\widehat 2} \, \gamma_{\widehat 3}$ squares to {\it minus} $1$, because of the time-like character of $\gamma_{\widehat 0}$.]
\smallskip
Summarizing: when $\chi$ is taken to be a Majorana spinor, the second-quantiza\-tion of $\chi$, Eq.~(\ref{eq3.7}), leads to a $4$-dimensional space of fermionic degrees of freedom. In other words, the wavefunction of the universe, $\Phi (\beta_a , \theta_a , \sigma)$, is labelled, besides the $6$ continuous bosonic labels $(\beta_1 , \beta_2 , \beta_3 ; \theta_1 , \theta_2 , \theta_3)$ of the gravitational minisuperspace, by a discrete index $\sigma$, which takes $4$ values, and which is algebraically equivalent to the spinor index of the Dirac representation of the $O(4)$ Clifford algebra (\ref{eq3.10}). When $\chi$ is taken to be a complex, Dirac spinor, the second quantization of $\chi$, Eq.~(\ref{eq3.5}), leads to a $16$-dimensional space of fermionic degrees of freedom, which is the tensor product of two independent Majorana-spinor spaces. Note that if one works in a Heisenberg picture, where the $\chi (t)$'s evolve in time according to the equations of motion (\ref{eq3.8}), the corresponding Clifford algebra generators (\ref{eq3.9}) must also be viewed as time-evolving objects, rather than fixed numerical gamma matrices.
\subsection{Explicit form of the Einstein-Majorana Bianchi-IX quantum Hamiltonian}\label{ssec3.3}
After having taken into account the diffeomorphism constraints (\ref{eq3.2}) in the form (\ref{eq3.3}), so that the wave function $\Phi (\beta , \sigma)$ depends only on the three diagonal metric variables $\beta_1 ,\beta_2 ,\beta_3$ and the discrete spin-space label $\sigma$, it is natural, in the gauge $\widetilde N \equiv N / \sqrt g = 1$, to quantize the gravitational degrees of freedom by replacing the classical $\beta$-momenta $\pi_{\beta}$ [which enter the Hamiltonian (\ref{eq2.90}) only through $T_{\beta} (\pi_{\beta})$, Eq.~(\ref{eq2.72})] by the differential operators $\widehat\pi_{\beta_a} := -i \, \partial / \partial \, \beta_a$. This leads to interpreting the Hamiltonian constraint as the following Klein-Gordon-like Wheeler-DeWitt equation
\begin{equation}
\label{eq3.12}
\widehat{\widetilde{\mathcal H}} { \ }^{\!\!\rm IX} \, \Phi (\beta , \sigma) = 0 \, .
\end{equation}
with
\begin{equation}
\label{eq3.13}
\widehat{\widetilde{\mathcal H}} { \ }^{\!\!\rm IX} := T_{\beta} (\widehat{\pi}_{\beta}) + \widehat{T}_w^{(0)} (\beta , \chi) + V_g^{\rm IX} (\beta) + \widehat V_{s \, {\rm grav}}^{\rm IX} (\beta , \chi) + \widehat V_{s \, {\rm mass}} (\beta , \chi) + c \, ,
\end{equation}
\begin{equation}
\label{eq3.14}
T_{\beta} (\widehat{\pi}_{\beta}) = - \frac{1}{4} \, G^{ab} \, \partial_{\beta_a} \, \partial_{\beta_b} = - \frac{1}{4} \, (\partial_{\beta_1}^2 + \partial_{\beta_2}^2 + \partial_{\beta_3}^2) + \frac{1}{8} \, (\partial_{\beta_1} + \partial_{\beta_2} + \partial_{\beta_3})^2 \, ,
\end{equation}
where $V_g^{\rm IX} (\beta)$ is the usual type-IX potential (\ref{eq2.30}), and where the quantum versions of all the $\beta - \chi$ coupling terms $\widehat T_x^{(0)} (\beta , \chi)$, $\widehat V_{s \, {\rm grav}}^{\rm IX} (\beta , \chi)$, $\widehat V_{s \, {\rm mass}} (\beta , \chi)$, are simply obtained from their classical expressions by interpreting the various spinor bilinears they contain as operators in the fermionic space defined, say in the Majorana case, by Eq.~(\ref{eq3.7}). This yields
\begin{eqnarray}
\label{eq3.15}
\widehat{T}_w^{(0)} (\beta , \chi) &= &\frac{1}{8} \coth^2 (\beta_1 - \beta_2) \, (\widehat\Sigma^{12})^2 + \frac{1}{8} \coth^2 (\beta_2 - \beta_3) \, (\widehat\Sigma^{23})^2 \nonumber \\
&&+ \, \frac{1}{8} \coth^2 (\beta_3 -\beta_1) \, (\widehat\Sigma^{31})^2 \, ,
\end{eqnarray}
where
\begin{equation}
\label{eq3.16}
\widehat\Sigma^{12} := \frac{i}{2} \, \chi^{\dagger} \, \gamma^{\widehat 1 \, \widehat 2} \, \chi \, , \ \widehat\Sigma^{23} := \frac{i}{2} \, \chi^{\dagger} \gamma^{\widehat 2 \, \widehat 3} \, \chi \, , \ \widehat\Sigma^{31} := \frac{i}{2} \, \chi^{\dagger} \gamma^{\widehat 3 \, \widehat 1} \, \chi \, ,
\end{equation}
are operators acting in the $4$-dimensional fermionic space. Similarly, denoting by ``$+$ cyclic'' the addition of two other cyclically permuted terms,
\begin{eqnarray}
\label{eq3.17}
V_{s \, {\rm grav}}^{\rm IX} (\beta , \chi) &= &- \frac{1}{4} \, e^{-2\beta_1} (i \, \chi^{\dagger} \, \gamma_{\widehat 0} \, \gamma^{\widehat 1 \, \widehat 2 \, \widehat 3} \, \chi) + \mbox{cyclic} \nonumber \\
&= &- \frac{1}{4} \, (e^{-2\beta_1} + e^{-2\beta_2} + e^{-2\beta_3}) (i \, \chi^{\dagger} \, \gamma_{\widehat 0} \,\gamma^{\widehat 1 \, \widehat 2 \, \widehat 3} \, \chi) \, ,
\end{eqnarray}
and,
\begin{equation}
\label{eq3.18}
\widehat V_{s \, {\rm mass}} (\beta , \chi) = m \, e^{-(\beta_1 + \beta_2 + \beta_3)} \, (i \, \chi^{\dagger} \, \gamma_{\widehat 0} \, \chi) \, .
\end{equation}
Finally, we have allowed for the addition of an ``ordering constant'' $c$ in Eq.~(\ref{eq3.13}). There are several motivations for allowing for such an additional constant. First, if we were working in a different gauge, {\it e.g.} $N=1$, {\it i.e.} $\widetilde N = g^{-1/2} = e^{(\beta_1 + \beta_2 + \beta_3)}$ instead of $\widetilde N = 1$, we would have had to quantize ${\mathcal H} = \frac{1}{4} \, e^{(\beta_1 + \beta_2 + \beta_3)}$ $G^{ab} \, \pi_{\beta_a} \, \pi_{\beta_b}$ instead of $\frac 14 G^{ab} \, \pi_{\beta_a} \, \pi_{\beta_b}$. More generally the quantization
of ${\mathcal H} = \frac{1}{4} \, e^{2 \alpha(\beta)}$ $G^{ab} \, \pi_{\beta_a} \, \pi_{\beta_b}$,
where $\alpha(\beta)= \alpha_a \beta_a$ is some linear form in the $\beta$'s, leads (after
reabsorbing $e^{\alpha(\beta)}$ in the definition of the wavefunction)
to an ordering ambiguity which is equivalent to adding a constant
$c \propto G^{ab} \alpha_a \alpha_b$ in Eq.~(\ref{eq3.13}). Second, the contribution (\ref{eq3.15}) being {\it quartic} in the non-commuting $\chi$'s has also ordering ambiguities (though it is natural to define it as we did, {\it i.e.} by taking the squares of the spinor bilinears (\ref{eq3.16}), which have no ordering ambiguities because $\gamma^{\widehat 1 \, \widehat 2}$, {\it etc.}, are antisymmetric matrices). Finally, there is a basic ambiguity in the gravity-spinor Lagrangian due to the possible use of first-order (Palatini) versus second-order formulations. It has been shown long ago \cite{Weyl29} that this leads to an ambiguity in the action density proportional to
\begin{eqnarray}
\label{eq3.19}
\Delta {\mathcal L} &= & \frac{1}{3!} \, \sqrt{- \, ^4 g} \ (\overline\psi \, \gamma^{\widehat\alpha \, \widehat\beta \, \widehat\gamma} \, \psi) (\overline\psi \, \gamma_{\widehat\alpha \, \widehat\beta \, \widehat\gamma} \, \psi) \nonumber \\
&= &\widetilde N \, [- (i \, \chi^{\dagger} \, \gamma^{\widehat 1 \, \widehat 2} \, \chi)^2 - (i \, \chi^{\dagger} \, \gamma^{\widehat 2 \, \widehat 3} \, \chi)^2 - (i \, \chi^{\dagger} \, \gamma^{\widehat 3 \, \widehat 1} \, \chi)^2 \nonumber \\
&&+ \ (i \, \chi^{\dagger} \, \gamma_{\widehat 0} \, \gamma^{\widehat 1 \, \widehat 2 \, \widehat 3} \, \chi)^2] \, .
\end{eqnarray}
We will come back below to the latter ($\beta$-independent when $\widetilde N = 1$, but $\chi$-dependent) ambiguity. Finally, note that one could also consider adding to the quantum
Dirac action other Lorentz-invariant contributions quadratic in spinor bilinears,
$\Delta {\mathcal L} \sim (\overline\psi \, \gamma^{A} \, \psi)^2$,
because no symmetry prevents the appearance of such quartic contributions.
\subsection{Properties of the spinor bilinears}\label{ssec3.4}
For the time being, let us note that, in the representation (\ref{eq3.6}), one can compute explicit expressions for the Majorana-spinor bilinears entering the Hamiltonian. Let the $4$ components of the Majorana spinor $\chi$ in the representation (\ref{eq3.6}) be denoted as $\chi_1 , \chi_2 , \chi_3 , \chi_4$, and let $\chi^T$ denote the transpose of the column vector $\chi_{\alpha}$, {\it i.e.} the horizontal row $\chi^T = (\chi_1 , \chi_2 , \chi_3 , \chi_4)$. We find
\begin{eqnarray}
\label{eq3.20}
\chi^T \, \gamma^1 \, \gamma^2 \, \chi &= &\chi_1 \, \chi_2 - \chi_2 \, \chi_1 + \chi_3 \, \chi_4 - \chi_4 \, \chi_3 \, , \nonumber \\
\chi^T \, \gamma^2 \, \gamma^3 \, \chi &= &\chi_2 \, \chi_3 - \chi_3 \, \chi_2 + \chi_1 \, \chi_4 - \chi_4 \, \chi_1 \, , \nonumber \\
\chi^T \, \gamma^3 \, \gamma^1 \, \chi &= &\chi_3 \, \chi_1 - \chi_1 \, \chi_3 + \chi_2 \, \chi_4 - \chi_4 \, \chi_2 \, ,
\end{eqnarray}
for the bilinears entering $\widehat T_w^{(0)}$ (without the $i/2$ prefactor), while
\begin{equation}
\label{eq3.21}
\chi^T \, \gamma_0 \, \gamma_1 \, \gamma_2 \, \gamma_3 \, \chi = \chi_1 \, \chi_2 - \chi_2 \, \chi_1 - \chi_3 \, \chi_4 + \chi_4 \, \chi_3 \, ,
\end{equation}
and
\begin{equation}
\label{eq3.22}
\chi^T \, \gamma_0 \, \chi = \chi_2 \, \chi_3 - \chi_3 \, \chi_2 - \chi_1 \, \chi_4 + \chi_4 \, \chi_1 \, ,
\end{equation}
are the bilinears entering $\widehat V_{s \, {\rm grav}}$ and $\widehat{V}_{s \, {\rm mass}}$.
\smallskip
In view of (\ref{eq3.20}) it would seem that $\widehat T_w^{(0)}$, which contains the squares of these bilinears, is a hopelessly complicated expression. However, things simplify very much if we use the Clifford-algebra properties of the $\chi$'s. More precisely, using the notation (\ref{eq3.9}), and simple properties such as $\Gamma_1^2 = 1$, $(\Gamma_1 \, \Gamma_2)^2 = -1$, and $\Gamma_1 \, \Gamma_2 \, \Gamma_5 = - \Gamma_3 \, \Gamma_4$ (with the definition (\ref{eq3.11}) of $\Gamma_5$), one finds the following simplified expressions for the spinor bilinears
\begin{eqnarray}
\label{eq3.23}
\chi^T \gamma^{12} \, \chi &= &\Gamma_1 \, \Gamma_2 \, \frac{1-\Gamma_5}{2} \, , \nonumber \\
\chi^T \gamma^{23} \, \chi &= &\Gamma_2 \, \Gamma_3 \, \frac{1-\Gamma_5}{2} \, , \nonumber \\
\chi^T \gamma^{31} \, \chi &= &\Gamma_3 \, \Gamma_1 \, \frac{1-\Gamma_5}{2} \, , \nonumber \\
\chi^T \gamma_{0123} \, \chi &= &\Gamma_1 \, \Gamma_2 \, \frac{1+\Gamma_5}{2} \, , \nonumber \\
\chi^T \gamma_0 \, \chi &= &\Gamma_2 \, \Gamma_3 \, \frac{1+\Gamma_5}{2} \, .
\end{eqnarray}
Note that these expressions contain the projectors
\begin{equation}
\label{eq3.24}
P_{\pm} := \frac{1 \pm \Gamma_5}{2} \, ,
\end{equation}
which satisfy the usual properties of Dirac's helicity projectors $(1 \pm i \, \gamma_5)/2$ [with $\gamma_5 = \gamma_0 \, \gamma_1 \, \gamma_2 \, \gamma_3$ so that $(i \, \gamma_5)^2 = +1$], namely: $P_+ + P_- = 1$, $P_+^2 = P_+$, $P_-^2 = P_-$, $P_+ \, P_- = 0$.
\smallskip
Let us also note that the above bilinears satisfy simple commutation relations among themselves. First, as $\Gamma_{ab} := \Gamma_{[a} \Gamma_{b]}$ ({\it i.e.} $\Gamma_{11} = 0$, $\Gamma_{12} = \Gamma_1 \, \Gamma_2$, {\it etc.}) commutes with $\Gamma_5$, and $P_+ \, P_- = 0$, we see that the first three bilinears in (\ref{eq3.23}) commute with the last two. Second, one easily checks that (together with its cyclic analogs)
\begin{equation}
\label{eq3.25}
\left[ \chi^T \, \frac{\gamma^{12}}{2} \, \chi \, , \ \chi^T \, \frac{\gamma^{23}}{2} \, \chi \right] = \chi^T \, \frac{\gamma^{13}}{2} \, \chi \, ,
\end{equation}
{\it i.e.} adding the factor $i$ transforming the anti-hermitian $\frac{1}{2} \, \chi^T \gamma^{ab} \, \chi$ into the hermitian operators $\widehat{\Sigma}^{ab}$ of Eq.~(\ref{eq3.16}),
\begin{equation}
\label{eq3.26}
\left[ \widehat\Sigma^{12} , \widehat\Sigma^{23} \right] = -i \, \widehat\Sigma^{31} \, ,
\end{equation}
together with its cyclic kins. In other words, the second-quantized operators $\widehat\Sigma^{12}$, $\widehat\Sigma^{23}$, $\widehat\Sigma^{31}$ satisfy the (reverse-sign) commutation law of body-frame angular momenta (the classical Poisson bracket $\{ L_1 , L_2 \} = -L_3$ mentioned above becoming $[L_1 , L_2] = -i \, L_3$ at the quantum level).
\smallskip
Let us note that, more generally, one can deduce directly from the basic anticommutation relations (\ref{eq3.7}) that the commutators of two bilinears $\chi^T A \chi \equiv \chi_{\alpha} \, A_{\alpha\beta} \, \chi_{\beta}$ and $\chi^T B \, \chi \equiv \chi_{\alpha} \, B_{\alpha\beta} \, \chi_{\beta}$ (where $A$ and $B$ are some, say antisymmetric, $\chi$-independent matrices) is equal to
\begin{equation}
\label{3.27}
[\chi^T A \, \chi \, , \ \chi^T B \, \chi] = \chi^T \, [A,B] \, \chi \, ,
\end{equation}
where $[A,B]_{\alpha\beta} = A_{\alpha\sigma} \, B_{\sigma\beta} - B_{\alpha\sigma} \, A_{\sigma\beta}$ is the usual matrix commutator.
\smallskip
In addition the squares of the bilinears (\ref{eq3.23}) simplify. Notably, $(\chi^T \gamma^{12} \, \chi)^2 = (\Gamma_1 \, \Gamma_2)^2 \, P_-^2 = -P_-$, {\it i.e.}
\begin{equation}
\label{eq3.28}
\left( \widehat\Sigma^{12} \right)^2 = \left( \widehat\Sigma^{23} \right)^2 = \left( \widehat\Sigma^{31} \right)^2 = \frac{1}{4} \, P_- \, .
\end{equation}
\subsection{``Squared-mass'' in the Wheeler-DeWitt equation}\label{ssec3.5}
The main result of Ref.~\cite{BK} was the finding that a classical spinor source ({\it i.e.} treating the spinor bilinears $\chi^T A \, \chi$ as $c$-numbers) modifies the usual quadratic relation $\underset{a}{\sum} \, p_a^2 - \left( \underset{a}{\sum} \, p_a \right)^2 = 0$ satisfied by Kasner exponents (in the Bianchi type~I case, {\it i.e.} far from any potential wall) into the relation\footnote{We normalize the type~I spatial metric so that $\sqrt g = T$, where $T$ is the proper time.}
\begin{equation}
\label{eq3.29}
\sum_a p_a^2 - \left( \sum_a p_a \right)^2 = - \frac{1}{2} \left[ (\Sigma^{12})^2 + (\Sigma^{23})^2 + (\Sigma^{31})^2 \right] \, ,
\end{equation}
where the $\Sigma^{ab}$ are the spinor bilinears defined above (treated as $c$-numbers). In the language of cosmological billiards (and using the time gauge $N = \sqrt g = T$, i.e. a coordinate time $t=- \ln T$), the result (\ref{eq3.29}) means that the Lorentzian squared velocity in $\beta$-space $\dot\beta^2 := G_{ab} \, \dot\beta_a \, \dot\beta_b = \underset{a}{\sum} \, p_a^2 - \left( \underset{a}{\sum} \, p_a \right)^2$ is negative, $\dot\beta^2 < 0$, {\it i.e. time-like}, rather than light-like ($\dot\beta^2 = 0$) as usual. Belinsky and Khalatnikov \cite{BK} argued that this implied that a classical spinor would (like a scalar field does) ultimately quench the chaotic, oscillatory regime near the singularity, and ultimately turn it into a monotonic Kasner-like, power-law behaviour (because the $\beta$-particle can ultimately move on a time-like geodesic which no longer hits any gravitation wall).
\smallskip
The result (\ref{eq3.29}) was understood via an Hamiltonian approach, and within a more general Kac-Moody coset model approach, by de~Buyl, Henneaux and Paulot \cite{dBHP}. These authors found that, far from the walls associated to a general coset model, the mass-shell condition for the $\beta$-momenta $\pi_{\beta_a} \propto G_{ab} \, \dot\beta_b$ was modified, by the coupling to a classical spinor, from the usual light-like condition $\pi_{\beta}^2 = 0$ (where $\pi_{\beta}^2 := G^{ab} \, \pi_{\beta_a} \, \pi_{\beta_b}$) to a massive-shell condition
\begin{equation}
\label{eq3.30}
\pi_{\beta}^2 + \mu^2 = 0 \, ,
\end{equation}
where the general expression for the ($\beta$-space) squared-mass $\mu^2$ was found to be
\begin{equation}
\label{eq3.31}
\mu_{\rm coset}^2 = \frac{1}{2} \sum_{\alpha} (i \, \chi^{\dagger} J_{\alpha}^s \, \chi)^2 \, .
\end{equation}
Here, $\chi$ is the coset Dirac field [which corresponds to the rescaled Einstein-Dirac field (\ref{eq2.55})], and $\alpha$ labels all the (positive) Kac-Moody roots (including their degeneracy) that enter the bosonic coset model $G/K$. Each root $\alpha$ is a linear form in the $\beta$'s, and corresponds (in Iwasawa gauge) to a bosonic wall potential proportional to $\exp (- \, 2 \, \alpha(\beta))$. The object $J_{\alpha}^s$ in (\ref{eq3.31}) is an anti-hermitian generator which represents (in the spinor representation $s$) the ``rotation'' $E_{\alpha} - E_{-\alpha}$ of the maximally compact subgroup $K$ of $G$ associated to the root $\alpha$. The extra factor $i$ in (\ref{eq3.31}) turns $J_{\alpha}^s$ into a hermitian operator, so that $i \, \chi^{\dagger} J_{\alpha}^s \, \chi$ is real and $\mu^2$ is given (for a Kac-Moody coset $G/K$) by an infinite sum of positive quantities.
\smallskip
In the language of the general Kac-Moody result (\ref{eq3.30}), (\ref{eq3.31}) (which uses a normalization where $\pi_{\beta_a} = G_{ab} \, \dot\beta_b$, without the factor $2$ entering our Eq.~(\ref{eq2.71}) above) the original Belinsky-Khalatnikov result (\ref{eq3.29}) corresponds to the squared-mass
\begin{equation}
\label{eq3.32}
\mu_{\rm BK}^2 = \frac{1}{2} \left[ (\Sigma^{12})^2 + (\Sigma^{23})^2 + (\Sigma^{31})^2 \right] \, .
\end{equation}
This squared-mass is of the general form (\ref{eq3.31}) (see below) but includes only three roots, namely the roots corresponding to the so-called ``symmetry walls''\footnote{When considering
potential wall forms $w_A(\beta)$ from a coset perspective we denote them as $\alpha_A(\beta)$.},
$\alpha_{12} (\beta) = \beta^2 - \beta^1$, $\alpha_{23} (\beta) = \beta^3 - \beta^2$ and $\alpha_{13} (\beta) = \beta^3 - \beta^1$. One might wonder whether the presence of only those three symmetry roots, and, in particular, the absence of any contribution coming from the gravitational-wall roots $\alpha^g_{1;23}(\beta)=2 \beta_1$, etc.
(which are crucial to generate the BKL chaos) is due to the approximate nature of the treatment of Ref.~\cite{BK}, based on a Bianchi type-I model, {\it i.e.} a model which contains all the symmetry walls, but no gravitational walls. Separately from this issue, we wish here to use our results above to clarify the effect of considering, as one should, the spin source as being quantum, rather than classical, as was assumed both in \cite{BK} and in \cite{dBHP}.
\smallskip
One can define the squared-mass term in the Wheeler-DeWitt equation as the term remaining with $\widehat\pi_{\beta}^2$ in a BKL-type limit $\beta \to + \infty$ in which one stays far away from all exponential walls. We have already mentioned that $\widehat V_{s \, {\rm grav}}^{\rm IX}$, Eq.~(\ref{eq3.17}), contains (half) gravitational walls, while $\widehat V_{s \, {\rm mass}}$, Eq.~(\ref{eq3.18}), contains a (half) cosmological-constant-related exponential wall. All these terms, as well as the usual bosonic type-IX potential $V^{\rm IX} (\beta)$, Eq.~(\ref{eq2.30}), will tend to zero in this far-wall BKL limit. It remains to consider the contribution $\widehat T_w^{(0)}$, Eq.~(\ref{eq3.15}). Using the hyperbolic-trigonometry identity
$ \coth^2 x = 1 + 1/\sinh^2 (x)$,
we see that we can split $\widehat T_w^{(0)}$ as
\begin{equation}
\label{eq3.34}
\widehat T_w^{(0)} = \frac{1}{4} \, \widehat{\mu}_{\Sigma}^2 + \widehat{V}^{\rm centrif} (\beta , \chi) \, ,
\end{equation}
where
\begin{equation}
\label{eq3.35}
\widehat{\mu}_{\Sigma}^2 = \frac{1}{2} \left[ (\widehat\Sigma^{12})^2 + (\widehat\Sigma^{23})^2 + (\widehat\Sigma^{31})^2 \right] \, ,
\end{equation}
and
\begin{equation}
\label{eq3.36}
\widehat{V}^{\rm centrif} (\beta , \chi) = \frac{1}{8} \, \frac{(\widehat\Sigma^{12})^2}{\sinh^2 (\beta_1 - \beta_2)} + {\rm cyclic} \, .
\end{equation}
One recognizes in (\ref{eq3.36}) the sum of three $\sinh^{-2}$-type centrifugal-wall potentials, which is the form taken by symmetry walls when one uses a Gauss decomposition of the off-diagonal components of the metric rather than an Iwasawa decomposition
(see \cite{Ryan72,RyanShepley75}).
Far from the symmetry (or centrifugal) walls, {\it i.e.} when $\vert \beta^2 - \beta^1 \vert \gg 1$, {\it etc.}, $\widehat{V}^{\rm centrif}$ becomes a sum of exponentially small terms $\propto \exp (-2 \, \vert \beta^2 - \beta^1 \vert) + {\rm cyclic}$. This leaves as only terms contributing to the dynamics in the far-wall limit
\begin{equation}
\label{eq3.37}
\widehat{\widetilde{\mathcal H}} { \ }^{\!\!\rm IX} = \frac{1}{4} \, \left[ \widehat\pi_{\beta}^2 + \widehat\mu_{\Sigma}^2 + 4c \right] + \mbox{(exponentially small terms)},
\end{equation}
where $\widehat\pi_{\beta}^2 = - G^{ab} \, \partial_{\beta_a} \, \partial_{\beta_b}$ is the Klein-Gordon operator associated to the $\beta$-space Lorentzian metric.
\smallskip
This result shows that, in Bianchi type-IX (and also, as one easily sees, in type VIII), the coupling of gravity to a quantum spinor generates a $q$-number squared mass in the Klein-Gordon-like Wheeler-DeWitt equation given by
$$
\widehat\mu^2 = \widehat\mu_{\Sigma}^2 + 4c \, ,
$$
where $\widehat\mu_{\Sigma}^2$ is defined in Eq.~(\ref{eq3.35}), and where $c$ is the ordering ambiguity that we allowed for in Eq.~(\ref{eq3.13}).
\smallskip
Note, first, that if we take a formal classical limit of $\widehat\mu_{\Sigma}^2$, Eq.~(\ref{eq3.35}), we recover the Belinsky-Khalatnikov result $\mu_{\rm BK}^2$, Eq.~(\ref{eq3.32}), without any extra contribution coming from the gravitational walls (which were explicitly taken into account in our calculation, contrary to the type-I computation of Ref.~\cite{BK}). When comparing this result with the coset-model result $\mu_{\rm coset}^2$, Eq.~(\ref{eq3.31}), we see that there is a genuine difference. Even if one were considering a truncated coset model involving only the couplings to the walls entering the type-IX gravity model, the sum in Eq.~(\ref{eq3.31}) should include as roots both the three symmetry roots $\alpha_{12} (\beta) = \beta_2 - \beta_1$, $\alpha_{23} (\beta) = \beta_3 - \beta_2$ and $\alpha_{13} (\beta) = \beta_3 - \beta_1$, and the three gravitational roots $\alpha_{123}^g (\beta) = \beta_1 - \beta_2 - \beta_3 + \Sigma_a \, \beta_a = 2 \, \beta_1$, $\alpha_{231}^g (\beta) = 2 \, \beta_2$ and $\alpha_{312}^g (\beta) = 2 \, \beta_3$. The corresponding $K(AE_3)$ generators, $J_{\alpha} = E_{\alpha} - E_{-\alpha}$, within the {\it spinor} representation of $K(AE_3)$, are known to
be \cite{dBHP,DKN,de Buyl:2005mt,DKN2,DaHi},
\begin{equation}
\label{eq3.38}
J_{\alpha_{12}}^s = \frac{1}{2} \, \gamma^{12} \, , \quad J_{\alpha_{23}}^s = \frac{1}{2} \, \gamma^{23} \, , \quad J_{\alpha_{13}}^s = \frac{1}{2} \, \gamma^{13} \, ,
\end{equation}
for the three symmetry roots, and
\begin{equation}
\label{eq3.39}
J_{\alpha_{123}^g}^s = \frac{1}{2} \, \gamma_0 \, \gamma^{123} \, , \quad J_{\alpha_{231}^g}^s = \frac{1}{2} \, \gamma_0 \, \gamma^{231} \, , \quad J_{\alpha_{312}^g}^s = \frac{1}{2} \, \gamma_0 \, \gamma^{312} \, ,
\end{equation}
for the three gravitational walls. Inserting the expressions (\ref{eq3.38}) into the general coset result (\ref{eq3.31}) yields, in view of the definition (\ref{sigma}) of $\Sigma^{ab}$, precisely the result (\ref{eq3.35}) (including the correct normalization factor). However, a coset model involving the couplings to the walls entering the type-IX gravity model would yield a (quantum) squared-mass of the form
\begin{equation}
\label{eq3.40}
\widehat{\mu}_{\rm coset}^{2 \,({\rm IX}) } = \widehat\mu_{\Sigma}^2 + \frac{3}{2} \left( \frac{i}{2} \, \chi^{\dagger} \gamma_0 \, \gamma^{123} \, \chi \right)^2 \, ,
\end{equation}
where the factor $3$ comes from the fact that the three gravitational-root generators (\ref{eq3.39}) happen to be equal among themselves.
\smallskip
The reason why the coset-results (\ref{eq3.31}) or (\ref{eq3.40}) (for the type-IX truncation) contain more contributions than the original type-IX gravity model, Eq.~(\ref{eq3.35}), is easily understood from the corresponding derivations of these results in \cite{dBHP} and in Section~\ref{sec2} above. Indeed, one sees that each individual contribution $\frac{1}{2} \, (i \, \chi^{\dagger} J_{\alpha}^s \, \chi)^2$ to $\mu^2$ comes for the presence of a corresponding {\it time-derivative} coupling $\sim e \, \dot q \, A$, see Eq.~(\ref{eq2.64}), between gravity and the spinor. For instance, it is because in Eq.~(\ref{eq2.56}) there were (say in the far-wall limit) three Lagrangian-coupling terms,
\begin{equation}
\label{eq3.41}
-\cosh (\beta_1 - \beta_2) \, w^{12} \, \Sigma^{12} + {\rm cyclic} \simeq - \frac{1}{2} \, e^{\alpha_{12} (\beta)} \, w^{12} \, \Sigma^{12} + {\rm cyclic} \, ,
\end{equation}
between the rotational velocities $w^{12} = \dot\varphi \cos \theta + \dot\psi$, {\it etc.} of the off-diagonal metric variables (parametrized by Euler angles), that one ended up with $\widehat{\mu}_{\Sigma}^2$ given by the sum of three symmetry wall contributions $\alpha_{12}$, $\alpha_{23}$, $\alpha_{13}$. By contrast, the couplings corresponding to the three gravitational walls $\alpha_{123}^g$, $\alpha_{231}^g$, $\alpha_{312}^g$, between gravity and the spinor, predicted by the Einstein-Dirac action, were contained in $V_{s \, {\rm grav}}$, Eq.~(\ref{eq2.60}), {\it i.e.} were of the type
\begin{equation}
\label{eq3.42}
V_{s \, {\rm grav}} = - \frac{1}{2} \, C^1 \, _{23} \, e^{-\alpha_{123}^g (\beta)} \left[ \frac{i}{2} \, \chi^{\dagger} \gamma_0 \, \gamma^{123} \, \chi \right] + {\rm cyclic} \, .
\end{equation}
This is very analogous to the coupling (\ref{eq3.41}), with, however, the crucial difference that the rotational velocities $w^{12}$, {\it etc.}, appearing in (\ref{eq3.41}) are replaced in (\ref{eq3.42}) by the structure constants $C^1 \, _{23}$, {\it etc.} In the gravity picture, the structure constants are {\it spatial} derivatives of the metric, and therefore do not contribute a squared-mass term {\it via} the
\begin{equation}
\label{eq3.43}
\frac{(p-e \, A)^2}{2m} = \frac{1}{2} \, \frac{e^2 \, A^2}{m} + \ldots
\end{equation}
mechanism (with $e \, A \propto \frac{i}{2} \, \chi^{\dagger} J_{\alpha}^s \, \chi$) explained above, which was linked to a {\it time-derivative} coupling $\Delta {\mathcal L} = e \, \dot q \, A$. By contrast, the coset action is related to the gravity action by a {\it dualization} of some of the gravity degrees of freedom. In particular, the structure constants $C^a \, _{bc}$ become replaced by the conjugate momenta, {\it i.e.} essentially by the time-derivative, of some dual coset field, namely $\widetilde C^{ab} := \frac 12 \varepsilon^{bcd} \, C^a \, _{cd} \propto \Pi^{ab} \sim \dot\phi_{ab}$ for the $AE_3$ case linked to the $(3+1)$-dimensional Bianchi-IX dynamics (see Section~8 of \cite{DHN2}). This explains why the gravitational roots (and actually all the coset roots) give a contribution to $\mu^2$ within the coset model
(for the spin 1/2 case), but not within the original gravity model.
\smallskip
Let us now turn to the quantum nature of the Bianchi-IX squared-mass term (\ref{eq3.35}). Though each term $(\Sigma^{ab})^2 = \left( \frac{i}{2} \, \chi^{\dagger} \gamma^{ab} \, \chi \right)^2$ is quartic in the quantized spinor field $\chi$, we have shown above that the Clifford-algebra nature of the quantization conditions for a Majorana spinor led to simple final results. More precisely, we see from (\ref{eq3.28}) that each separate $(\Sigma^{ab})^2$ term yields the same result, so that the total quantum $\mu^2$ reads
\begin{equation}
\label{eq3.44}
\widehat{\mu}_{\Sigma}^2 = \frac{3}{8} \, P_- = \frac{3}{8} \, \frac{1-\Gamma_5}{2} \, ,
\end{equation}
where we recall that $\Gamma_5$ denotes the involutive $(\Gamma_5^2 = 1)$ operator (\ref{eq3.11}). The simple result (\ref{eq3.44}) shows that, in the quantum theory, the squared-mass term [in absence of any additional ordering-constant contribution $4c$, as in Eq.~(\ref{eq3.37})] is an operator which has two possible eigenvalues: the eigenvalue $\mu_{\Sigma}^2 = 0$ in the $2$-dimensional space of ``helicity'' $\Gamma_5 = +1$, or the eigenvalue $\mu_{\Sigma}^2 = 3/8$ in the complementary (orthogonal) $2$-dimensional space of ``helicity'' $\Gamma_5 = -1$.
[Note that $\mu_{\Sigma}^2$ vanishes in half of the total ($4$-dimensional) Majorana-spinor Hilbert space.] In terms of $(2j+1)$-dimensional irreducible representations $D_j$ of $SU(2)$, we can view the three components $\widehat\Sigma^{12}, \widehat\Sigma^{23}, \widehat\Sigma^{31}$ of
a single Majorana field as describing the sum $D_0 + D_0 +D_{\frac 12}$.
Let us also note that our results above, also gives the eigenvalues
of $\widehat\Sigma^{12}, \widehat\Sigma^{23}, \widehat\Sigma^{31}$ (and then of $\mu^2$) when considering a complex Dirac spinor. Indeed, this case is obtained by considering $\chi = \chi_R + i \, \chi_I$ as the complex combination of two, independent ({\it i.e.} anticommuting), Majorana spinors, so that each final $\widehat\Sigma^{12}$ etc., is given by a sum,
$\widehat\Sigma^{12}_R + \widehat\Sigma^{12}_I$,
of two commuting spin operators. The corresponding representation space is therefore
the tensor product $ (D_0 + D_0 +D_{\frac 12}) \times (D_0 + D_0 +D_{\frac 12})$,
i.e. the sum $ 5 D_0 + 4 D_{\frac 12} + D_1$.
This shows that the possible eigenvalues of $\widehat\mu^2$ are $0$, $3/8$ and $1$. Moreover, the eigenvalue $\widehat\mu^2 = 0$ is obtained (when assuming $c=0$) in a $5$-dimensional subspace of the $16$-dimensional total Hilbert space.
Note that we can extend this result by considering the case where gravity couples to a sum of
$N$ independent (i.e. anticommuting) Majorana spinor fields. In that case each
spin operator $\widehat\Sigma^{12}$ etc. will be the sum of $N$ independent (commuting)
contributions belonging to a $D_0 + D_0 +D_{\frac 12}$ representation space.
This implies that the eigenvalues of $\widehat\mu_{\Sigma}^2$ will range between $0$ and $N(N+2)/8$.
Decomposing the Hilbert space obtained from $N$ tensorial products of $(D_0 + D_0 +D_{\frac 12})$ into a direct sum of irreducible spin $s$ subspaces $D_s$,
one finds that the number $\Delta(s,N)$ of irreducible spaces $D_s$ appearing in the $4^N$ dimensional product space is given by :
$$\Delta(s,N)= \frac{(2\,s +1)}{N+1}\frac{ (2N+2)!}{ (N+2\,s+2)!\, (N-2\,s)!}.$$
For large $N$ and fixed $s$ this behaves as : $\Delta(s,N)\simeq 4^{N+1}\,(1+ 2\,s)/(\sqrt{\pi}\,N^{3/2})$. Applying this result to
the case $s=0$ shows that the eigenvalue $\mu_{\Sigma}^2=0$ will be realized in a rather large fraction of the total Hilbert space,
namely a subspace of dimension $ (2N+2)!/((N+1)! (N+2)!)\simeq 4^{N+1}/(\sqrt{\pi}\,N^{3/2})$ of the total $4^N$-dimensional space. For any $N$, the mean value of $ \mu_{\Sigma}^2 $ (treating all states as equally probable) is equal to $3 N/16>0$.
The standard deviation of $ \mu_{\Sigma}^2$ is found to be $\sigma_{\mu_{\Sigma}^2} =\frac 1{16}\sqrt{3(2\,N^2+N)}\approx 0.15 \,N $ (for large $N$).
In the large $N$ limit, one would recover the result of the classical treatments of Refs \cite{BK,dBHP}.
\smallskip
Let us also note that, if one were considering the coset-type mass (\ref{eq3.40}), including the gravitational-root contributions, the result (\ref{eq3.23}) shows that one would obtain
\begin{equation}
\label{eq3.45}
\widehat\mu_{\rm coset}^{2 \, ({\rm IX})} = \frac{3}{8} \, \frac{1-\Gamma_5}{2} + \frac{3}{8} \, \frac{1+\Gamma_5}{2} = \frac{3}{8} \, ,
\end{equation}
{\it i.e.} a constant, $c$-number result, all over the Hilbert space. This result is reminiscent of the full coset result (\ref{eq3.31}), whose r.h.s., is the quadratic Casimir invariant of the spinor representation of $K(G)$, which is a $c$-number in any irreducible representation.
Note, however, that, though the representation space of the spinor representation of
the Kac-Moody maximally compact subgroup $K(G)$ is finite dimensional, there is an
infinite number of roots $\alpha$, and a corresponding infinite number of generators
$J^s_\alpha$, that one should sum over in the coset result Eq. (\ref{eq3.31}), so that the
value of $\widehat\mu_{\rm coset}^2$ is formally infinite, or, at least, ill defined
as an operator. This indicates the need to renormalize it by some ordering prescription.
This suggests that, in spite of contrary appearances, the renormalized value
$\widehat\mu_{\rm coset}^2=0$ is an allowed possibility.
\smallskip
Let us come back to the result (\ref{eq3.44}) predicted by the usual Einstein-Dirac action, and further discuss its meaning. Let us first note that the non-zero value (\ref{eq3.44}) of $\widehat\mu^2$ is a quantum effect, which directly comes from the anti-commutation relation (\ref{eq3.7}) of the quantized spinor $\chi$. If we reinstate the $\hbar$ on the r.h.s. of (\ref{eq3.7}), we see that the objects $\Gamma_\alpha$ that satisfy the unit-normalized Clifford algebra (\ref{eq3.10}) are actually related to $\chi$ {\it via}: $\chi_{\alpha} = \frac{1}{2} \, \sqrt\hbar \, \Gamma_{\alpha}$. As $\widehat\mu^2$ is quartic in $\chi$ this shows that $\widehat\mu_{\Sigma}^2$ is actually $\frac{3}{8} \, \frac{1-\Gamma_5}{2} \, \hbar^2$, {\it i.e.} of order $\hbar^2$. In turn, this shows that, when one is in a quasi-classical (WKB-type) regime where the gravitational momenta $\pi_{\beta}$ are large compared to $\hbar$, the mass term will have only a sub-leading effect. This is consistent with the finding above that $\mu^2$ has an intrinsic quantum ambiguity due to ordering problems: indeed, the constant $c$ in Eq.~(\ref{eq3.37}) is also seen to be ${\mathcal O} (\hbar^2)$. On the other hand, when considering the Klein-Gordon-type equation defined by (\ref{eq3.37}) with $\widehat\pi_{\beta} = -i \hbar \, \partial / \partial \beta_a$, we see that a same factor $\hbar^2$ appears in front of the $\beta$-space d'Alembertian $- \, G^{ab} \, \partial_{\beta_a} \partial_{\beta_b}$ and of both the $\Sigma$-generated mass $\widehat\mu_{\Sigma}^2$, and any quantum-ordering contribution $4c$. This means that the mass term $\mu^2$ will be important when the wave function $\Phi (\beta , \sigma)$ has a characteristic scale of variation in $\beta$-space of order unity.
\smallskip
Let us finally note that the ambiguity in the Einstein-Dirac action linked to using a first-order or second-order formalism is equivalent, in view of Eq.~(\ref{eq3.19}), to adding to $\widehat\mu^2$ a term proportional to $- \widehat\mu_{\Sigma}^2 -+ \frac{1}{3} \widehat\mu_{\rm grav}^2$, where $\widehat\mu_{\rm grav}^2$ is the contribution to the coset result (\ref{eq3.31}) coming from the three gravitational roots, {\it i.e.} the second term on the r.h.s. of Eq.~(\ref{eq3.40}) (which is the {\it sum} $\widehat\mu_{\Sigma}^2 + \widehat\mu_{\rm grav}^2$).
\smallskip
This shows that the addition of a suitable multiple of the extra contribution (\ref{eq3.19}) to the Einstein-Dirac action can modify the basic result $\mu_{\Sigma}^2$ in a more general result of the type
\begin{equation}
\label{eq3.46}
(1 -k) \widehat\mu_{\Sigma}^2 + \frac{k}{3} \, \widehat\mu_{\rm grav}^2 + 4c \, ,
\end{equation}
where we took also into account a possible ($c$-number) ordering constant. For instance, if $k = \frac{3}{4}$ this would generate a mass term $\frac{1}{4} \, (\widehat\mu_{\Sigma}^2 + \widehat\mu_{\rm grav}^2) + 4c$. When $c=0$, this would correspond (modulo the factor $\frac{1}{4}$) with the coset-like result (\ref{eq3.45}). On the other hand, if we choose $k = \frac{3}{4}$ and $4c = -3/32$, we end up with a vanishing total squared-mass.
\smallskip
Keeping in mind all these (quantum) ambiguities in the value of $\widehat\mu^2$, we shall now discuss in more detail the quantum dynamics of the Bianchi-IX-spinor system.
\subsection{Quantum dynamics of the Bianchi-IX-spinor system}\label{ssec3.6}
The Wheeler-DeWitt equation for the Bianchi-IX-Majorana-spinor system has the form
\begin{eqnarray}
\label{eq3.47}
\biggl( - \frac{1}{4} \, G^{ab} \, \partial_{\beta_a} \, \partial_{\beta_b} +V_g^{\rm IX} (\beta) + \widehat{V}^{\rm centrif} (\beta , \chi) + \widehat{V}^{\rm IX}_{s \, {\rm grav}} (\beta , \chi) + \nonumber \\
+ \widehat{V}_{s \, {\rm mass}} (\beta , \chi) + \frac{1}{4} \, \widehat{\mu}_{\rm tot}^2 (\chi) \biggl) \Phi (\beta , \sigma) = 0 \, .
\end{eqnarray}
Here, the spin-independent term $V_g^{\rm IX} (\beta)$ is the usual bosonic type-IX potential (\ref{eq2.30}), while there appears three different spin-dependent potentials. Using our results above, their explicit expressions are (for type IX)
\begin{eqnarray}
\label{eq3.48}
\widehat V^{\rm centrif}_{\rm IX}(\beta , \chi) &= &\frac{1}{32} \left( \frac{1}{\sinh^2 (\beta_1 - \beta_2)} + \frac{1}{\sinh^2 (\beta_2 - \beta_3)} + \frac{1}{\sinh^2 (\beta_3 - \beta_1)} \right) \nonumber \\
&&\times \frac{1-\Gamma_5}{2} \, , \\
\label{eq3.49}
\widehat V_{s \, {\rm grav}}^{\rm IX} (\beta , \chi) &= &- \frac{1}{4} \, (e^{-2\beta_1} + e^{-2\beta_2} + e^{-2\beta_3})(i \, \Gamma_1 \Gamma_2) \, \frac{1+\Gamma_5}{2} \, , \\
\label{eq3.50}
\widehat V_{s \, {\rm mass}}^{\rm IX} (\beta , \chi) &= &+ \, m \, e^{-(\beta_1 + \beta_2 + \beta_3)} \, (i \, \Gamma_2 \Gamma_3) \, \frac{1+\Gamma_5}{2} \, .
\end{eqnarray}
The various possible values of $\widehat{\mu}_{\rm tot}^2$ have been discussed in the previous subsection.
\smallskip
The discrete spin label $\sigma$, indicated as argument of $\Phi$, takes four values, and reminds us that the various spin operators $\Gamma_5$, $i \, \Gamma_1 \Gamma_2$ and $i \, \Gamma_2 \Gamma_3$ act upon a $4$-dimensional Hilbert space. In other words, Eq.~(\ref{eq3.47}) is similar to the second-order form of the Dirac equation coupled to an external electromagnetic field, namely
\begin{equation}
\label{eq3.51}
\left[ (i \, \partial_{\mu} - e \, A_{\mu})^2 + \frac{1}{2} \, e \, \sigma^{\mu\nu} \, F_{\mu\nu} + m^2 \right] \psi = 0
\end{equation}
where $\sigma^{\mu\nu} = \frac{i}{2} \, \gamma^{\mu\nu}$ is the spin generator (in the Clifford algebra).
\smallskip
In the case of the usual Dirac equation (\ref{eq3.51}), the spin coupling term $\frac{1}{2} \, e \, \sigma^{\mu\nu} \, F_{\mu\nu}$ (linked to the magnetic moment of the electron) couples the $4$-different components of the (first-quantized) Dirac spinor $\psi$, and embodies physical phenomena such as the precession of the electron spin in an external magnetic field. Similarly, we can think of the wave function of the universe $\Phi (\beta)$ as a column vector of $4$ components $\Phi_{\sigma} (\beta)$, which propagate in the Lorentzian $\beta$-space, and are deflected by the spin-independent potential $V_g^{\rm IX} (\beta)$, together with the three spin-dependent potentials (\ref{eq3.48})--(\ref{eq3.50}) which are analogous to the $\frac{1}{2} \, e \, \sigma^{\mu\nu} \, F_{\mu\nu}$ spin-coupling term in Eq.~(\ref{eq3.51}). In addition, the mass term $\widehat{\mu}_{\rm tot}^2 (\chi)$ is also spin-dependent, as discussed in the previous subsection.
\smallskip
In this work, we shall focus on understanding the dynamics of the universe's multi-component wave function $\Phi (\beta)$ in the BKL regime where one approaches the singularity, {\it i.e.} in the limit where the various $\beta_a$'s all tend to $+ \, \infty$. For concreteness, let us first consider the case where $\widehat\mu^2$ is given by the Bianchi-IX prediction (\ref{eq3.44}) together with a possible ordering constant, but without any further contribution of the type of Eq.~(\ref{eq3.19}). In that case we have
\begin{equation}
\label{eq3.52}
\widehat\mu^2 = \frac{3}{8} \, \frac{1-\Gamma_5}{2} + 4c \, .
\end{equation}
Let us start by noting that the helicity operator $\Gamma_5$ commutes with all the terms in the Wheeler-DeWitt equation (\ref{eq3.47}). This means that the $4$ components of $\Phi$ can be viewed as the superposition of two independent, $2$-component wave functions; say $\Phi_+$ and $\Phi_-$, with
\begin{equation}
\label{eq3.53}
\Phi_{\pm} = \frac{1 \pm \Gamma_5}{2} \, \Phi \, , \quad \Phi = \Phi_+ + \Phi_- \, ,
\end{equation}
and where $\Phi_+$ and $\Phi_-$ undergo two uncoupled evolutions. More precisely $\Phi_+$ (which is a positive-helicity eigenstate $\Gamma_5 \, \Phi_+ = + \Phi_+$) satisfies (with $\Box_{\beta} \equiv G^{ab} \, \partial_{\beta_a} \, \partial_{\beta_b}$)
\begin{eqnarray}
\label{eq3.54}
&&\biggl( - \frac{1}{4} \, \Box_{\beta} + V_g^{\rm IX} (\beta) - \frac{1}{4} \, (e^{-2\beta_1} + e^{-2\beta_2} + e^{-2\beta_3}) \, i \, \Gamma_1 \Gamma_2 \nonumber \\
&&+ \, m \, e^{-(\beta_1 + \beta_2 + \beta_3)} \, i \, \Gamma_2 \Gamma_3 + c \biggl) \Phi_+ (\beta) = 0 \, ,
\end{eqnarray}
while $\Phi_- (\Gamma_5 \, \Phi_- = - \Phi_-)$ satisfies
\begin{eqnarray}
\label{eq3.55}
&&\biggl( - \, \frac{1}{4} \, \Box_{\beta} + V_g^{\rm IX} (\beta) + \frac{1}{32} \Bigl( \frac{1}{\sinh^2 (\beta_1 - \beta_2)} + \frac{1}{\sinh^2 (\beta_2 - \beta_3)} \nonumber \\
&&+ \, \frac{1}{\sinh^2 (\beta_3 - \beta_1)} \Bigl)+ \frac{3}{32} + c \biggl) \Phi_- (\beta) = 0 \, .
\end{eqnarray}
These two sub-equations have rather different structures: (i) they contain a different mass term, (ii) the $\Phi_-$ equation is spin-independent ({\it i.e.} the two independent components of $\Phi_-$ satisfy the same equation), while the two-independent components of $\Phi_+$ are coupled {\it via} the presence of the $i \, \Gamma_1 \Gamma_2$ and $i \, \Gamma_2 \Gamma_3$ spin-coupling terms; and (iii) they contain different spin-averaged potentials (besides the mass term). Concerning the last property, the spin-averaged potential term for $\Phi_+$ (taking into account that $(i \, \Gamma_a \Gamma_b)^2 = + \, 1$, {\it i.e.} that the eigenvalues of $i \, \Gamma_a \Gamma_b$, $a \ne b$, are $\pm \, 1$) is
\begin{eqnarray}
\label{eq3.56}
V_+ (\beta) &= &V_g^{\rm IX} (\beta) = \frac{1}{2} \, \, (e^{-4\beta_1} + e^{-4\beta_2} + e^{-4 \beta_3}) - \nonumber \\
&&-(e^{-2(\beta_1 + \beta_2)} + e^{-2 (\beta_2 + \beta_3)} + e^{-2 (\beta_3 + \beta_1)}) \, ,
\end{eqnarray}
while the spin-averaged potential for $\Phi_-$ is
\begin{eqnarray}
\label{eq3.57}
V_- (\beta) &= &V_g^{\rm IX} (\beta) + \frac{1}{32} \biggl( \frac{1}{\sinh^2 (\beta_1 - \beta_2)} + \frac{1}{\sinh^2 (\beta_2 - \beta_3)} \nonumber \\
&&+ \frac{1}{\sinh^2 (\beta_3 - \beta_1)} \biggl) \, .
\end{eqnarray}
We note that the spin-averaged potential $V_+ (\beta)$ for $\Phi_+$ is the usual type-IX potential for a {\it diagonal} metric, which contains only gravitational walls, but no symmetry (or centrifugal) walls. Classically, this potential (approximately) confines the motion of the $\beta$ particle (when considering the near-singularity limit) to stay within the Lorentzian billiard chamber defined by the condition that the three gravitational wall forms be positive:
\begin{equation}
\label{eq3.58}
w_{123}^g (\beta) = 2 \, \beta_1 \geq 0 \, , \quad w_{231}^g (\beta) = 2 \, \beta_2 \geq 0 \, , \quad w_{312}^g (\beta) = 2 \, \beta_3 \geq 0 \, .
\end{equation}
At the quantum level, this confinement mechanism will be blurred by tunnelling effects. As is well-known the physics of tunnelling in quantum cosmology crucially depends on the choice of boundary conditions in configuration space, see, e.g., \cite{Vilenkin}. Leaving a discussion of more general states to future work, we shall focus here on wave functions which are quantum analogs of the classical billiard motions within the appropriate billiard chamber ({\it e.g.} the chamber (\ref{eq3.58}) when considering the $\Phi_+$ component). For such states, taking the ordering constant $c$ in (\ref{eq3.54}) to its naive value, $c=0$, we can qualitatively describe the quantum evolution of $\Phi_+$ by looking at the various potential terms in (\ref{eq3.54}). First, we note that, when trying to penetrate the gravitational walls (\ref{eq3.58}), {\it i.e.} when exploring regions where some of the wall forms, {\it e.g.}, $2 \, \beta_1$, become negative, the spin-averaged bosonic potential $V_+ (\beta) = V_g^{\rm IX} (\beta)$ will grow like $+ \, \frac{1}{2} \, e^{-4\beta_1}$, and will therefore dominate over the corresponding (non positive-definite) growing gravitational spin-dependent potential $\propto e^{-2\beta_1} \, i \, \Gamma_1 \Gamma_2$. In addition, the spin-dependent potential related to the mass $m$ of the spinor becomes exponentially small in the near-singularity limit where the volume of the universe, $\propto \sqrt g = e^{-(\beta_1 + \beta_2 +\beta_3)}$, tends to zero ({\it i.e.} when the sum $\beta_1 + \beta_2 + \beta_3 \to + \, \infty$). This discussion shows that there will exist quasi-classical spinor-like wave functions $\Phi_+ (\beta)$ (with two independent components) consisting of WKB-like solutions approximately bouncing between the walls of the chamber (\ref{eq3.58}), and decaying in the ``forbidden'' domain, $\beta_1 < 0$, $\beta_2 < 0$, $\beta_3 < 0$. Compared to the usual, pure gravity Wheeler-DeWitt equation for Bianchi-IX, which would be
\begin{equation}
\label{eq3.59}
\left( - \, \frac{1}{4} \, \Box_{\beta} + V_g^{\rm IX} (\beta) \right) \Phi_0 (\beta) = 0 \, ,
\end{equation}
for a scalar-valued (one component) $\Phi_0 (\beta)$, the interesting new feature of the gravity-spinor system is the presence of additional spin-dependent couplings. In the case of the $\Phi_+$ equation (\ref{eq3.54}), these are the terms containing $i \, \Gamma_1 \Gamma_2$ and $i \, \Gamma_2 \Gamma_3$. As these two (Clifford algebra) operators do not commute among themselves, their presence means that the two independent components of $\Phi_+$ will continuously mix under the influence of these terms. We thereby end up with the picture of a quantum fermionic billiard where the various polarization state of a spinor-valued wave function mix as the quantum state bounces within the billiard chamber. We shall discuss below the precise link between such a quantum fermionic billiard, and its Grassmann-valued correspondent, studied in Ref.~\cite{DaHi}.
\smallskip
In the BKL limit $\beta_1 + \beta_2 + \beta_3 \to + \, \infty$, the term proportional to $m$ in the evolution equations for $\Phi$ become negligible. Let us then consider the special case where $m=0$, which should also describe the general behaviour when $\beta_1 + \beta_2 + \beta_3 \to + \, \infty$. For the case $m=0$ (and $c=0$), the evolution equation (\ref{eq3.54}) for $\Phi_+$ further simplifies in that it contains only one spin-dependent operator, namely $i \, \Gamma_1 \Gamma_2$. In that case, the idempotent operator $i \, \Gamma_1 \Gamma_2$ {\it commutes} with the Hamiltonian, so that we can further decompose $\Phi_+$ into two (scalar) components, say
\begin{equation}
\label{eq3.60}
\Phi_+^{(\pm)} := \frac{1 \pm i \, \Gamma_1 \Gamma_2}{2} \, \Phi_+ \, ,
\end{equation}
that evolve independently of each other. Namely, these components satisfy
\begin{equation}
\label{eq3.61}
\left( - \, \frac{1}{4} \, \Box_{\beta} + V_g^{\rm IX} (\beta) \mp \frac{1}{4} \, (e^{-2\beta_1} + e^{-2\beta_2} + e^{-2\beta_3}) \right) \Phi_+^{(\pm)} (\beta) = 0 \, .
\end{equation}
In other words, when $m=0$, we can reduce the dynamics of the multi-component wave function $\Phi (\beta , \sigma)$ to the uncoupled dynamics of several separate components satisfying different scalar, Wheeler-DeWitt equations.
\smallskip
Turning to the dynamics of $\Phi_-$, Eq.~(\ref{eq3.55}), we already noted that it did not contain spin-dependent terms. This means that, similarly to the $m=0 \, , \Phi_+$ case discussed above, the two components of $\Phi_-$ satisfy (even when $m \ne 0$) the same scalar, Wheeler-DeWitt equation (\ref{eq3.55}). The latter equation has two interesting new features compared to the $\Phi_+$ one. First, though the off-diagonal, gauge-like Euler-angle degrees of freedom have been eliminated by the diffeomorphism constraints, they have left behind new centrifugal contributions, $\propto 1/\sinh^2 (\beta_1 - \beta_2) + {\rm cyclic}$, to the potential (see Eq.~(\ref{eq3.57}). These centrifugal terms create infinite potential barriers located at the three symmetry walls: $\beta_1 = \beta_2$, $\beta_2 = \beta_3$ and $\beta_3 = \beta_1$. To examine the effect of these barriers on the quantum wave function, let us consider, say, what happens near the $\beta_1 = \beta_2$ symmetry wall. Locally, we can change coordinates in $\beta$-space and use $x \equiv \beta_1 - \beta_2$ as new coordinate. As the hyperplane $\beta_1 = \beta_2$ is timelike ({\it i.e.} its normal is spacelike), the coordinate $x$ is a spatial coordinate in the Lorentzian $\beta$-space. This means that the $\Phi_-$ equation has, near the $x=0$ hyperplane, the structure
\begin{equation}
\label{eq3.62}
\left( \partial_t^2 - \partial_y^2 - \partial_x^2 + \frac{\alpha^2}{x^2} + \mbox{regular} \right) \Phi_- (t,y,x) = 0 \, ,
\end{equation} where we completed the spatial coordinate $x$ by another spatial one, $y$, together with some time-like variable $t$, and where $\alpha^2$ denotes a positive numerical constant linked to the coefficient of the centrifugal terms. [In the case considered here
we have $\alpha^2=1/16$, but it is instructive to leave its value arbitrary.]
Here, $t,x,y$ are some linear combinations of the three variables $\beta_1 , \beta_2 , \beta_3$, chosen so that the Lorentzian metric $G_{ab} \, d\beta_a \, d\beta_b = -dt^2 + dx^2 + dy^2$. We can separate the motion w.r.t. $t$ and $y$, {\it i.e.} $\Phi_- = e^{iky - i\omega t} \, \varphi (x)$. Finally, the behaviour of $\varphi (x)$ near the singular point $x=0$ is given by $\partial_x^2 \, \varphi (x) \simeq \alpha^2 \, x^{-2} \, \varphi (x)$. This implies that $\varphi (x) \sim x^s$ with a power $s$ satisfying the indicial equation $s(s-1) = \alpha^2$, so that $s_{\pm} = \frac{1}{2} \pm \sqrt{\frac{1}{4} + \alpha^2}$. The important point is that the two possible solutions satisfy $s_+ > 1$ and $s_- < 0$. If we reject the possibility of a singular wave function at $x=0$, this eliminates $\varphi_- (x) \sim x^{s_-}$. Finally, we conclude that the centrifugal terms force the wave function to vanish at each symmetry wall like $\vert \beta_1 - \beta_2 \vert^{s_+}$, $\vert \beta_2 - \beta_3 \vert^{s_+}$ and $\vert \beta_3 - \beta_1 \vert^{s_+}$, with $s_+ > 1$. In other words, the centrifugal terms confine the evolution of $\Phi_- (\beta)$ to six separate chambers, corresponding to the six possible orderings of $\beta_1 , \beta_2 , \beta_3$: namely, $\beta_1 \leq \beta_2 \leq \beta_3$, $\beta_2 \leq \beta_1 \leq \beta_3$, {\it etc.} In each symmetry chamber, the WKB-like solutions for $\Phi_-$ will bounce between two ``hard'' symmetry walls ({\it e.g.} $\beta_1 = \beta_2 < \beta_3$ and $\beta_1 < \beta_2 = \beta_3$ for the chamber $\beta_1 \leq \beta_2 \leq \beta_3$) and the ``soft'' (exponential) gravitational wall ({\it e.g.} the $2 \, \beta_2 = 0$ wall for the $\beta_1 \leq \beta_2 \leq \beta_3$ chamber).
\smallskip
If we use the naive ordering constant $c=0$ (which led to a ``massless'' Wheeler-DeWitt equation for $\Phi_+$) we see in Eq.~(\ref{eq3.55}) that the $\Phi_-$ components have acquired a positive squared-mass $\mu_-^2 = 3/32$. This mass term will strongly affect the near-singularity behaviour of the $\Phi_-$ wave function. Indeed, as there exist wave packets of the free {\it massive} Klein-Gordon equation $(\Box_{\beta} - \mu_-^2) \, \Phi = 0$ that are approximately localized (in $\beta$ space) around, say, a time-like line $\beta_a = v^a \, \tau + \beta_a ^{(0)}$, with $G_{ab} \, v^a \, v^b < 0$, such wave packets yield approximate asymptotic solutions of the Wheeler-DeWitt equation (\ref{eq3.55}), because, as $\beta_1 + \beta_2 + \beta_3 \to + \, \infty$ (within, say, the chamber $\beta_1 \leq \beta_2 \leq \beta_3$) the potential terms will become negligible in the domain where the wave packet is approximately localized. This is the quantum version of the result of \cite{BK,dBHP} that a positive $\mu^2$ ultimately quenches the chaotic billiard motion, in the near singularity limit.
However, let us note that, when $c=0$, the fact that the q-number $\widehat{\mu}^2$
admits the eigenvalue zero means that there will always exist a part of the quantum wavefunction
(namely the one described by its $\Phi_+$ component) which will
exhibit a chaotic behavior near the singularity. In other words, at the quantum level,
the classical result of \cite{BK,dBHP} does not prevent part of the quantum reality to
behave chaotically.
\section{Quantum versus Grassmannian fermionic billiards}\label{sec4}
The two crucial new features brought by coupling gravity to a spinor are: (i) the appearance
of new spin-dependent potential terms; and (ii) the possible presence of a squared-mass term
in the Klein-Gordon-Wheeler-DeWitt equation. As discussed above, though the feature (ii)
is important for knowing whether the mixmaster chaos is affected or not by the coupling
to a quantum spinor, it is also a delicate quantum effect, which is sensitive to the choice
of ordering in the Wheeler-DeWitt equation, as well as to other choices. We shall assume
in this Section either that we have made choices so that the relevant
squared-mass term vanishes\footnote{This would, for instance, be the case for the $\Phi_+$ components of $\Phi$, with the
naive ordering constant $c=0$, see Eq. (\ref{eq3.54}).}, or that we are considering situations
where one can distinguish the effect of the spin-dependent potential from other effects.
Under this assumption, we can focus on the effect of the spin-dependent potential terms,
and contrast a {\it quantum} treatment of these spin-dependent terms, to a treatment
where the spinor variables are considered as Grassmann-valued (G-numbers).
The recent work \cite{DaHi} has studied in detail the fermionic billiards defined
by the chaotic dynamics of G-valued (spin-3/2) spinor fields,
as they arise in the near-singularity limit of supergravity (both in dimension eleven,
and dimension four). The two main results of Ref. \cite{DaHi} that are relevant to our
present study are: (1) the spin-3/2 billiard defined by supergravity factorizes into
a spin-1 (vector) billiard, and a spin-1/2 (Majorana-spinor) one; in view of the results
of Ref. \cite{dBHP}, it is the latter spin-1/2
billiard which is of relevance for us here; and (2) the spin-1/2 billiard consists of a succession
of generalized Weyl reflections, defined in a kind of spin-covering $\mathcal{W}^s$
of the standard Weyl group
of the Kac-Moody algebra associated with the considered supergravity model.
More precisely, it was found that each collision of the universe on a wall labelled
by a Kac-Moody root $\alpha$ (i.e. so that the corresponding wall form is
simply $w(\beta) =\alpha(\beta)$), corresponding to a spin-dependent potential proportional to
$ \exp (- \alpha(\beta) ) \, i \, \chi^{\dagger} J_{\alpha}^s \, \chi$, causes
the G-valued spin-1/2 field $\chi$ to ``rotate'', in the spin-1/2 representation space,
by an angle of $\pi/4$ along the axis defined by $ J_{\alpha}^s$. More precisely,
the value of the Dirac spinor $\chi$ after the collision on the wall $\alpha$ differs
from its incident value by the matrix
\begin{equation}
\label{Ralpha}
\mathcal{R}_{\alpha} = e^{\varepsilon_\alpha \frac{\pi}{2} J_{\alpha}^s} \, .
\end{equation}
Here, $\varepsilon_\alpha = \pm 1$ is a sign which will be defined below, and
the object $J_{\alpha}^s$, which was introduced above, is an anti-hermitian generator which represents (in the spinor representation $s$) the ``rotation'' $E_{\alpha} - E_{-\alpha}$ of the maximally compact subgroup $K$ of $G$ associated to the root $\alpha$. For instance, in the
($3+1$)-dimensional case of relevance here, the generators $J_{\alpha}^s$ associated,
respectively, to the symmetry walls and the leading gravitational walls, were
written down in Eqs. (\ref{eq3.38}), (\ref{eq3.39}). Note that the generators $ J_{\alpha}^s$
all include a factor $1/2$, so that their eigenvalues are $\pm \frac{1}{2}$.
[In conjunction with the
$\frac{\pi}{2}$ prefactor, this implies, as stated, a rotation angle of $\pm \frac{\pi}{4}$.]
Note also that all the generators $ J_{\alpha}^s$ belong to the Clifford algebra defined
by the usual ($3+1$)-dimensional gamma matrices, so that the spinor-wall-reflection matrices
(\ref{Ralpha}) are $4 \times 4$ matrices acting on the usual Dirac spinor space. [More precisely,
they are real, and act on the real Majorana spinor space.]
The picture of fermionic billiards emerging from Ref. \cite{DaHi} is a growing, chaotic
succession of spinor reflections
\begin{equation}
\label{Rproduct}
\mathcal{R}_{\alpha_n} . \mathcal{R}_{\alpha_{n -1}} \cdots \mathcal{R}_{\alpha_2} . \mathcal{R}_{\alpha_1}
\end{equation}
acting on the spinor index of the G-valued spinor $\chi$.
It was found in Ref. \cite{DaHi} that the multiplicative group $\mathcal{W}^s$
defined by spin-1/2
billiard, i.e. by all the matrix products (\ref{Rproduct}), is of finite cardinality,
both in the eleven-dimensional case, and the four-dimensional one. For example,
in the 4-dimensional case of relevance here, the group of products (\ref{Rproduct})
of the basic reflection generators, Eqs. (\ref{eq3.38}), (\ref{eq3.39}), associated with
the symmetry and gravitational walls is a finite group of cardinality $4 \times 48$.
Actually, it happens that the 4-dimensional case is quite special in that the
symmetry generators happen to {\it commute} with the gravitational wall ones
(which, themselves, reduce to only one element as the three generators (\ref{eq3.39})
differ only by a cyclic permutation of 123, which does not change the value of
$\gamma^{123}$). Due to this, the group $\mathcal{W}^s$ is the direct product
of two separate groups.
Having recalled the results concerning the near-singularity limit of the dynamics
of G-valued spinors coupled to a chaotic bosonic cosmology, we wish now to
clarify the connection of these results to the case of quantum spinors, sourcing
a cosmological model. In that case, we must replace the classical evolution
of a G-valued spinor $\chi$, considered along the classical evolution of the
bosonic billiard described by the $\beta$-particle bouncing between symmetry
and gravitational walls, by a solution of the coupled multi-component Wheeler-DeWitt
equation. However, in order to be able to compare the two treatments we must consider
a case where the Wheeler-DeWitt equation do correspond to the spinor coupling
terms studied in Refs. \cite{dBHP,DaHi}. Indeed, the latter works considered
classical Hamiltonians of the form
\begin{equation} \label{Hspincoupling}
H= \frac{1}{4} \, G^{ab} \, \pi_{\beta_a} \, \pi_{\beta_b} + \sum_{\alpha} \frac{1}{2}
e^{-2 \alpha(\beta)} \Pi_\alpha^2 + \sum_{\alpha} \frac{1}{2}
e^{- \alpha(\beta)} \Pi_\alpha \Sigma_\alpha + c_{\rm tot}
\end{equation}
Here, the $\Pi_\alpha$'s are dynamical variables whose meaning is different if one
considers a gravity model or a coset model, the $ \Sigma_\alpha$'s are the spin-coupling
terms corresponding to a general root, namely
\begin{equation} \label{sigmaalpha}
\Sigma_\alpha = i \, \chi^{\dagger} J_{\alpha}^s \, \chi
\end{equation}
where the $J_{\alpha}^s$'s are the products of usual gamma matrices defined
in Eqs. (\ref{eq3.38}), (\ref{eq3.39}) for symmetry and gravitational roots, and
where the constant $c_{\rm tot}$ (defining the squared mass $\mu^2= 4 c_{\rm tot}$)
includes various types of contributions, which depend on the model considered (as discussed above).
The important feature of the spin-couplings included in (\ref{Hspincoupling}) is that
the {\it same} quantity $\Pi_\alpha$ enters the bosonic potential wall
$\frac{1}{2} e^{-2 \alpha(\beta)} \Pi_\alpha^2$, and the corresponding spin-dependent
potential term $\frac{1}{2} e^{- \alpha(\beta)} \Pi_\alpha \Sigma_\alpha $.
One can verify that this is indeed the case for the general Hamiltonian (\ref{eq2.69}),
both for the terms linked to the dominant gravitational walls (with, e.g., $\alpha_{123} = 2 \beta_1$,
$\Pi_{\alpha_{123}}= -C^1_{\ 23}$, and $J_{\alpha_{123}}$ given by (\ref{eq3.39}) ), and to the symmetry walls. However,
in the latter case, as Eq. (\ref{Hspincoupling}) refers to an Iwasawa decomposition of the metric
one must reconsider the far-wall limit of the Gauss-decomposition-based Hamiltonian (\ref{eq2.69}).
In that limit (e.g. when $\beta_2- \beta_1 \gg 1$) one finds that the various hyperbolic
functions in Eq. (\ref{eq2.76}) do yield a structure compatible with the general Iwasawa
result (\ref{Hspincoupling}) (with, e.g., $\Pi_{\alpha_{12}} = \pi_{w^{12}}$).
In order to see more clearly the differences between the Grassmannian treatment
of spinor couplings, and its quantum treatment, let us consider the quantization
of the simplest version of the general Hamiltonian (\ref{Hspincoupling}), namely
the case where there is only {\it one wall}, corresponding to one root $\alpha$.
In that case, the dynamical variable $\Pi_\alpha$ is clearly seen to be a constant
of motion, and can therefore be considered as being a c-number both in a
classical and a quantum treatment. [In the quantum treatment, we consider
eigenstates of the operator $\Pi_\alpha$.] It then remains to quantize $\beta$
and $\chi$. Canonical quantization of the
$\beta$ dynamics yields $ \pi_\beta = - i \partial_\beta$, while the quantization
of $\chi$ is done according to the anticommutation relations (\ref{eq3.5}), or
(\ref{eq3.7}) in the Majorana case that we shall consider here. This yields a quantum
Hamiltonian of the form
\begin{equation} \label{Hquantumalpha}
H= -\frac{1}{4} \, G^{ab} \, \partial_{\beta_a} \, \partial_{\beta_b} + \frac{1}{2}
e^{-2 \alpha(\beta)} \Pi_\alpha^2 + \frac{1}{2}
e^{- \alpha(\beta)} \Pi_\alpha \widehat{\Sigma}_\alpha + c_{\rm tot}
\end{equation}
where $\widehat{\Sigma}_\alpha$ is the quantum operator defined by replacing
$\chi$ in Eq. (\ref{sigmaalpha}) by its quantized version, submitted to (\ref{eq3.7}).
As explained above, the quantum wavefunction $\Phi(\beta, \sigma)$ has both
continuous indices related to the $\beta$ dynamics, and a discrete one, $\sigma$,
related to the Clifford algebra (\ref{eq3.7}) satisfied by the quantum $\chi$.
Let us consider solutions of $ H \Phi(\beta, \sigma)=0$ describing wavepackets
colliding on the $\alpha$ potential wall. One can write such solutions explicitly
by separating the variables in Eq. (\ref{Hquantumalpha}). Namely, as the potential
terms depend only on the combination $\alpha(\beta)$ of the $\beta$'s, we can look
for wavefunctions $\Phi$ of the form
\begin{equation}
\Phi(\beta, \sigma) = e^{i k_{\parallel}(\beta)} F(\beta_{\perp}, \sigma)
\end{equation}
where the linear form $k^{\parallel}$ describes the $\beta$-space momentum
parallel to the wall $\alpha(\beta)=0$ (i.e. $G^{a b} k^{\parallel}_a \alpha_b =0$),
while $ F(\beta_{\perp}, \sigma)$, where $\beta_{\perp} := \alpha(\beta)$,
describes the motion perpendicular to the $\alpha$ wall
(as well as the spin degrees of freedom). Inserting this separated wavefunction in the
Wheeler-DeWitt equation $ H \Phi(\beta, \sigma)=0$ (and considering a root $\alpha(\beta)$
of norm $G^{ab} \alpha_a \alpha_b = 2$, as is the case for all the roots considered here)
yields the following equation for $F$:
\begin{equation} \label{Morse1}
\partial_{\beta_{\perp}}^2 F(\beta_{\perp},\sigma) = (e^{- 2 \beta_{\perp}} \Pi_\alpha^2 + e^{- \beta_{\perp}} \Pi_\alpha \widehat{\Sigma}_\alpha - {\cal Q} ) F(\beta_{\perp},\sigma)
\end{equation}
where $\cal Q$ is a separation constant which involves both $ (k^{\parallel})^2 :=G^{a b} k^{\parallel}_a k^{\parallel}_b$ and the constant $c_{\rm tot}$ in Eq. (\ref{Hquantumalpha}).
If $ \widehat{\Sigma}_\alpha$ were a c-number, this would be the Schr\"odinger equation in a Morse potential, which is well-known to be reducible
to the general confluent hypergeometric equation. In our case, $ \widehat{\Sigma}_\alpha$
is an operator in spin space (i.e. acting on the index $\sigma$ labelling the four
components of the wavefunction $\Phi$). However, we can reduce our problem to a
c-number-valued $ \widehat{\Sigma}_\alpha$ by first considering wavefunctions $\Phi(\sigma)$
that are eigenstates of this operator: say $ \widehat{\Sigma}_\alpha \Phi(\sigma) = {\Sigma}_\alpha \Phi(\sigma)$. We can also simplify the resulting equation by
shifting $\beta_{\perp} $ by a constant so as to absorb the $\Pi_\alpha^2$ factor multiplying
$ e^{- 2 \beta_{\perp}} $. Namely, one defines $\beta'_{\perp}$ such that
$ e^{- 2 \beta'_{\perp}} = \Pi_\alpha^2 e^{- 2 \beta_{\perp}} $
This makes also $\Pi_\alpha$ disappear from the second term,
except for its sign: $\varepsilon_\alpha := \text{sgn}(\Pi_\alpha) = \pm 1$.
When $ {\cal Q} $ is positive, there is a solution of the resulting equation
\begin{equation} \label{Morse2}
\partial_{\beta'_{\perp}}^2 F(\beta'_{\perp},\sigma) = (e^{- 2 \beta'_{\perp}} + e^{- \beta'_{\perp}} \varepsilon_\alpha {\Sigma}_\alpha - {\cal Q}) F(\beta'_{\perp},\sigma)
\end{equation}
which starts as an incident plane wave
$F(\beta'_{\perp},\sigma) \propto \exp - i \sqrt{ {\cal Q}} \beta'_{\perp} $ faraway
from the wall, i.e. when $\beta_{\perp} \gg +1 $, reflects on the wall located
around $\beta'_{\perp} =0$ (with an evanescent wave in the `forbidden' region
$\beta'_{\perp} < 0$), and ends up, with some dephasing, as an outgoing
plane wave $F(\beta'_{\perp},\sigma) \propto \exp + i \sqrt{ {\cal Q}} \beta'_{\perp} $
when $\beta'_{\perp} \to + \infty $. Using the notation $U(a,b,z)$ for the second
Kummer function (see \cite{AS}, chapter 13), the exact solution of Eq. (\ref{Morse2})
describing this scattering process on the combined bosonic+spinorial wall $\alpha(\beta)$
is given by
\begin{equation}
F [ \beta'_{\perp}]=\exp[-e^{-\,\beta'_{\perp}}]\,e^{-\,i\,\sqrt{{\cal Q}}\,\beta'_{\perp}}U[\frac 12+ \frac 12 \varepsilon_\alpha \Sigma_\alpha +i\,\sqrt{{\cal Q}},1+2\,i\,\sqrt{{\cal Q}},2 \, e^{-\,\beta'_{\perp}}]\\
\end{equation}
We can then extract from this exact expression (using the expansion of the Kummer
$U(a,b,z)$ function near $z=0$) the phase factor $e^{i \delta_\alpha}$
between the incident wave and the outgoing one. We find
\begin{equation} \label{q-phase}
e^{i\, \delta_\alpha(\Sigma_\alpha)} = -\frac{\Gamma(\frac 12+ \frac 12 \varepsilon_\alpha \Sigma_\alpha -i\,\sqrt{{\cal Q}})\,\Gamma(\frac 12+2\,i\,\sqrt{{\cal Q}})}{\Gamma(\frac 12+ \frac 12 \varepsilon_\alpha \Sigma_\alpha+i\,\sqrt{{\cal Q}})\,\Gamma(\frac 12 -2\,i\,\sqrt{{\cal Q}})}
\end{equation}
Let us now compare this quantum dephasing with the Grassmannian
description, recalled above, of the reflection of the spinor $\chi_m$ on the
wall $\alpha(\beta)$. [Here, we denote by $m,n, \ldots$
spinor indices, to avoid the confusion with the use of $\alpha$ as a label for the root].
This description was that the value $\chi'$ of the spinor after the interaction
with the wall $\alpha$ is obtained from its value $\chi$ before the interaction via
the following matrix transformation:
\begin{equation} \label{c-phase}
\chi'_m = (\mathcal{R}_{\alpha})_{mn} \chi_n = \left[ e^{\varepsilon_\alpha \frac{\pi}{2} J_{\alpha}^s} \right]_{mn} \chi_n
\end{equation}
where $\varepsilon_\alpha := \text{sgn}(\Pi_\alpha) = \pm 1$ as above.
Similarly to what happens for the usual Dirac equation, we can consider that
the various components, labelled by $m$, of the spinor $\chi_m$ encode
the classical polarization state of $\chi$. This polarization state can then be also encoded
by decomposing the spinor $\chi_m$ onto a basis of eigenstates of the
first-quantized hermitian spin generator $i \, J_{\alpha}^s$. [We recall that $i \, J_{\alpha}^s$ is
$i$ times the product of an even number of ordinary gamma matrices; e.g. for the
symmetry root $\alpha_{12} = \beta_2 -\beta_1$ it is the ordinary first-quantized
spin generator $i \,\frac 12 \gamma^{12}$, whose eigenvalues are $\pm \frac 12$, and
whose eigen-spinors are often used to decompose a general spinor into various spin states.]
We then see that Eq. (\ref{c-phase}) is saying that the classical dephasing, upon
reflection on the wall $\alpha$, of a classical spinor $\chi_m$ polarized so as to
be an eigenstate of $i J_{\alpha}^s$, with eigenvalue $ \sigma_\alpha = \pm \frac 12$
is given by the phase factor $\exp - i \varepsilon_\alpha \frac{\pi}{2} \sigma_\alpha$.
The quantum dephasing (\ref{q-phase}) looks {\sl a priori} quite different from the classical
dephasing $\exp - i \varepsilon_\alpha \frac{\pi}{2} \sigma_{\alpha}$. Let us, however, check that they
agree, as they should, in the quasi-classical, WKB limit. This limit corresponds to
considering a large value of $\sqrt{ {\cal Q} }$, i.e. high-frequency
wavepackets $\sim \exp \pm i \sqrt{ {\cal Q}} \beta'_{\perp} $ . In addition, we need
to decompose the quantum phase $ \delta_\alpha(\Sigma_\alpha)$ in (\ref{q-phase}) into two separate contributions:
(i) a spin-independent part, $ \delta_\alpha(0)$, which can be mathematically defined by
replacing $\Sigma_\alpha$ by zero on the r.h.s. of Eq. (\ref{q-phase}), and (ii)
the spin-dependent contribution, $ \delta_\alpha(\Sigma_\alpha) - \delta_\alpha(0)$.
This yields a spin-dependent dephasing factor given by
\begin{equation} \label{q-spinphase}
e^{i\, (\delta_\alpha(\Sigma_\alpha) - \delta_\alpha(0))} =
\frac{\Gamma(\frac 12+ \frac 12 \varepsilon_\alpha \Sigma_\alpha -i\,\sqrt{{\cal Q}})}
{\Gamma(\frac 12 -i\,\sqrt{{\cal Q}})}
\frac{\Gamma(\frac 12 +i\,\sqrt{{\cal Q}})}{\Gamma(\frac 12+ \frac 12 \varepsilon_\alpha \Sigma_\alpha +i\,\sqrt{{\cal Q}})}
\end{equation}
Using now the fact that, for large values of $z$, one has
$\Gamma(z+a)/\Gamma(z) = z^a (1 + O(1/z))$, we see that, modulo fractional
corrections of order $1/\sqrt{{\cal Q}}$, the spin-dependent part of the quantum phase,
i.e. $\delta_\alpha(\Sigma_\alpha) - \delta_\alpha(0)$, is equal to
$ - \frac{\pi}{2} \varepsilon_\alpha \Sigma_\alpha$, in perfect agreement with the classical result $ - \frac{\pi}{2} \varepsilon_\alpha \sigma_\alpha$.
This result also shows that the generalized spin operators
$\widehat{\Sigma}_\alpha = i \chi^{\dagger} J_{\alpha}^s \chi$ are the {\it second-quantized}
versions of the first-quantized gamma-algebra generators $i J_{\alpha}^s$, and that
classical spinors $\chi_m$ that are eigenstates of $i J_{\alpha}^s$ do correspond,
after quantizing $\chi$ according to Eq. (\ref{eq3.5}), to quantum eigenstates $\Phi$
of $\widehat{\Sigma}_\alpha $. Note also that the fact that the quantum phase
(\ref{q-phase}) agrees, in the quasi-classical limit, with the result obtained
by integrating the classical equation of motion of $\chi$, namely
$$
\partial_t \chi = \frac 12 e^{- \alpha(\beta)} \Pi_\alpha J_\alpha^s \chi
$$
finally rests on the fact that, according to the general result (\ref{eq3.8}),
the Heisenberg-picture quantum operator $\hat{\chi}$ satisfies the same
equation of motion.
We have compared here the quantum and classical descriptions of the
`collision' of the universe on a single (bosonic+spinorial) potential wall $\alpha(\beta)$.
We could discuss, exactly along the same lines, the dynamics of the two $\Phi_+$
components of the wavefunction $\Phi$ in the Bianchi-IX case. Indeed, $\Phi_+$
satisfies the (two-component) equation (\ref{eq3.54}) which (in the case $m=0$)
is closely similar to the Wheeler-DeWitt equation associated to Eq. (\ref{Hquantumalpha}),
except that we now have spin-dependent collisions on three different gravitational
walls. Note, however, that the spin evolution is very simple in this case, as the
three gravitational-wall spin operators $J_{\alpha_{123}}^s, J_{\alpha_{231}}^s, J_{\alpha_{312}}^s$ all coincide. The spin eigenstates being the same for the three
different walls, there is no room for any interesting chaotic behaviour of the
spin polarization, as could happen if one had had three different (and non commuting)
spin operators. As for the two remaining $\Phi_-$ components of the Bianchi-IX $\Phi$
they satisfy the spin-independent equation (\ref{eq3.55}) so that we cannot
compare its dynamics to the discussion of \cite{DaHi}.
There is, however, another Bianchi model for which we can compare the
classical and quantum dynamics, it is the Bianchi-II model, defined by structure constants
$C^a \, _{bc} = \varepsilon_{bcd} \, n^{ad}$ with $n^{ab} = {\rm diag} (1,0,0)$.
The classical dynamics of this model is discussed in Appendix C, using an
Iwasawa fixing of the dreibein. The diagonal metric variables are denoted $\beta_1, \beta_2, \beta_3$, while the Iwasawa off-diagonal degrees of freedom are denoted
$\nu_{12}, \nu_{23}$ and $\nu_{13}$. [Note that indices are not naturally cyclically permuted
when working in an Iwasawa representation.]
The quantization of $\chi$ is done as above,
while the quantization of the metric degrees of freedom is conveniently done
in a $\beta, \nu$ representation, i.e. with momenta conjugated to the three $\beta$'s
defined as $ \widehat\varpi^a = -i \partial_{\beta_a}$,
and with momenta conjugated to the three $\nu$'s defined as
$\widehat\varpi^{ab} = -i \partial_{\nu_{ab}}$.
In the Bianchi-II model, there are only two non trivial diffeomorphism constraints
(classically given by Eqs. (\ref{BIISconst})), in addition to the Hamiltonian constraints (Eq. (\ref{BIILconst})).
The quantum mechanical versions of these two constraints will then be the following
constraints on the quantum state $\Phi$:
\begin{equation}
\widehat\varpi^{12} \Phi=\widehat\varpi^{13} \Phi=0\label{QSconst}
\end{equation}
In the $\beta, \nu$ representation, i.e. for a wavefunction $\Phi(\beta, \nu, \sigma)$
(where, as above, the discrete index $\sigma$ labels the spin degrees of freedom),
the constraints (\ref{QSconst})) imply that $\Phi(\beta, \nu, \sigma)$ does not depend
on $\nu_{12}$ and $\nu_{13}$, but only (besides $\sigma$)
on $\beta_1$, $\beta_2$, $\beta_3$ and $\nu_{23}$.
The remaining constraint is the Hamiltonian one, of the form
\begin{equation} \label{BIIWDW1}
\widehat{\cal H}\Phi=0
\end{equation}
Assuming for simplicity a vanishing Dirac mass, $m=0$, but allowing, as above, for
an ordering constant $c$, the Hamiltonian operator
[whose classical version is (Eq. (\ref{BIILconst})) ] may be written as
\begin{eqnarray} \label{qHII}
\widehat {\cal H}&=& \frac 14\left[\sum_i(\widehat\varpi^i)^2-\frac 12 (\sum_i\widehat\varpi^i)^2\right]+ \frac 12e^{-4\beta_1} \nonumber\\
&&
+ \frac18 \left[ \widehat \Sigma^{[\hat 1\hat 2]}\strut^2+ \widehat \Sigma^{[\hat 3\hat 1]}\strut^2+(\widehat \Sigma^{[\hat 2\hat 3]}+ 2\,e^{(\beta_2-\beta_3)} \widehat{\varpi}^{23})^2 \right] -\frac {e^{-2\beta_1}}4 i \chi^\dagger\gamma_0 \gamma_{123} \chi \nonumber\\
&&+ c
\end{eqnarray}
It is amusing to note that the quantum dynamics of this Bianchi-II model is
simpler to analyze than its classical dynamics. Indeed, as discussed in Appendix C,
it is difficult to solve generically the classical (or Grassmannian) spinor evolution equation
(\ref{SWeq}). By contrast, in the Schr\"odinger-Wheeler-DeWitt picture of Eq. (\ref{BIIWDW1})
there are no fermionic equations to solve. In addition, we have that: (1)
the spin dependence of the Hamiltonian is analog to that of a {\it symmetric} top, i.e. involving the spin operators $\widehat \Sigma^{ab}$
only through $\widehat \Sigma^{[\hat 2\hat 3]}$, and the combination $\widehat \Sigma^{[\hat 1\hat 2]}\strut^2+\widehat \Sigma^{[\hat 1\hat 3]}\strut^2$; and (2) the additional spin-dependent term linked to
the gravitational wall, i.e. $ i \chi^\dagger\gamma_0 \gamma_{123} \chi$, commutes
with the $\widehat \Sigma^{ab}$-dependent terms (see the subsection \ref{ssec3.4}). [Note that this commutation property (2) is valid in the case where one would consider a complex,
Dirac spinor $\chi$, because of the independence of the two real Majorana
components of such a general $\chi$.]
In other words, we can diagonalize the Hamiltonian by imposing that the state $\Phi$
is a simultaneous eigenstate of the following three operators:
\begin{eqnarray}
\widehat \Sigma^{[\hat 2\hat 3]} \Phi = \sigma_{23} \Phi, \nonumber \\
\left( \widehat \Sigma^{[\hat 1\hat 2]}\strut^2+\widehat \Sigma^{[\hat 1\hat 3]}\strut^2+\widehat \Sigma^{[\hat 2\hat 3]}\strut^2 \right) \Phi = {\cal S} ^2 \Phi, \nonumber \\
\frac{i}{4} \chi^\dagger\gamma_0 \gamma_{123} \chi \Phi = {\cal C} _g \Phi
\end{eqnarray}
On each $\sigma$ component of such a state, the Hamiltonian constraint equation
leads to an equation of the form
\begin{equation}
\left[-\frac 14\square_{\beta}+ \frac 12e^{-4\beta_1} -\frac 12 e^{2(\beta_2-\beta_3)}\partial^2_{\nu_{23}}-\frac i2\,{ \sigma_{2 3} }\,e^{\beta_2-\beta_3}\partial_{\nu_{23}}+{\cal C}_g\,e^{-2\beta_1}+\frac 18 {\cal S} ^2 +c \right]\Phi=0\, .\label{QGenEq}
\end{equation}
By using the results of subsection \ref{ssec3.4} the allowed values of the quantum numbers ${ s_{2 3} }$, ${\cal C}_g$ and $ {\cal S} ^2$, and their multiplicity, for a Majorana spinor are,
$$
[\sigma_{23},\, {\cal C}_g,\, {\cal S} ^2]_{\rm mult.}: [0,\pm 1/4,0]_1,[\pm 1/2,0,3/4]_1\, .
$$
while, for a Dirac spinor they are :
$$
[\sigma_{23},\, {\cal C}_g,\, {\cal S} ^2]_{\rm mult.}: [0,0,0]_3, \, [0,\pm 1/2,0], \,
[\pm 1/2,\pm 1/4,3/4]_2, \, [0,0, 2], \, [\pm 1,0,2] \, .
$$
Note that in the Majorana case $\sigma_{23}$ and $ {\cal C} _g$ cannot be both non zero, and we
have the simple link $ {\cal S} ^2 = 3\, \sigma_{23}^2$. These links are relaxed in the Dirac-spinor case.
Eq. (\ref{QGenEq}) can be solved by separation of variables. Indeed, the wave operator
$\square_{\beta}$ depends on three $\beta$ variables, while the potential walls
entering the equation involve only two combinations of the three $\beta$'s,
namely $\alpha_{123}(\beta) = 2 \beta_1$ and $\alpha_{23}(\beta) = \beta_3 - \beta_2$.
Therefore there exists a linear combination of the three $\beta$'s which will be `orthogonal'
(in the Lorentz-$\beta$-space sense) to the two combinations $\alpha_{123}(\beta)$
and $\alpha_{23}(\beta)$. It is easy to see that
$\alpha_0(\beta) := 2 \beta_1 + \beta_2 + \beta_3$ is
such a combination. Actually, one can easily check that the three variables
$\alpha_0, \alpha_{23}, \alpha_{123}$ define an orthogonal coordinate system
in Lorentzian $\beta$-space, with $\alpha_0$ being a time-like coordinate, and
$ \alpha_{23}$, and $ \alpha_{123}$ two space-like coordinates. Finally, a generic
solution can always be expressed by superposing separated solutions of the form:
\begin{equation}\label{mode }
\Phi(\beta, \nu_{23})=e^{i\,p\,\nu_{23}}e^{i\,k\,(2\beta_1+\beta_2+\beta_3)} F_1(2 \beta_1)F_2(\beta_3-\beta_2)
\end{equation}
where the two functions $F_1$ and $F_2$ satisfy exactly the same Morse-potential-type
Schr\"odinger equation that we encountered above (in the single-wall case) namely:
\begin{eqnarray}
F_1^{\prime\prime}[2 \beta_1]&=&\left(e^{-4\,\beta_1}- 2\, {\cal C}_g\,e^{-2\,\beta_1}-{\cal Q}_1\right) F_1[2\beta_1]\\
F_2^{\prime\prime}[ \beta_3-\beta_2]&=&\left(p^2\,e^{2\,(\beta_2-\beta_3)}+ { \sigma_{2 3} }\,p\,e^{(\beta_2-\beta_3)} - {\cal Q} _2 \right) F_2[\beta_3-\beta_2]
\end{eqnarray}
Here the two separation constants $ {\cal Q} _1$ and $ {\cal Q} _2$ must satisfy the following
mass-shell condition
\begin{equation} \label{mass-shell1}
-k^2 + {\cal Q} _1 + {\cal Q} _2 = -\frac 14 {\cal S} ^2 - 2 c = - \frac 12 \mu^2
\end{equation}
where $\mu^2$ is the squared-mass of the Wheeler-DeWitt equation.
If we look, as above, for wave functions that vanish behind the $ \alpha_{23}$, and
$ \alpha_{123}$ walls, we must choose Kummer's $U$-type solution for $F_1$
and $F_2$.
When ${\cal Q}_1 $ and $ {\cal Q} _2$ are positive, the physically relevant solutions are
(when assuming $p>0$ for definiteness):
\begin{eqnarray}
&&F_1 [2 \beta_1]=\exp[-e^{-2\,\beta_1}]\,e^{-2\,i\,\sqrt{{\cal Q}_1}\,\beta_1}U[\frac 12- {\cal C}_g+i\,\sqrt{{\cal Q}_1},1+2\,i\,\sqrt{{\cal Q}_1},2 e^{-2\,\beta_1}] \, , \nonumber \\
&&\\
&&F_2 [ \beta_3-\beta_2]= \exp[-p e^{-(\beta_3-\beta_2)}]\,e^{- i\,\sqrt{ {\cal Q} _2}(\beta_3-\beta_2)}\quad \nonumber \\
&&\phantom{F_2 [ \beta_3-\beta_2]= } \times U[\frac 12(1+ {\sigma_{2 3} })+i\,\sqrt{ {\cal Q} _2},1+2\,i\,\sqrt{ {\cal Q} _2},\, 2 \,p\,e^{-(\beta_3-\beta_2)}] \nonumber\ .\\&&
\end{eqnarray}
Far from the two walls, these modes propagate as plane waves in all the variables, with $\beta$-space momenta $\varpi_a$ of the form
$$
\varpi_a {[\varepsilon_G,\varepsilon_S]}=\{2(k+\varepsilon_G\, \sqrt{{\cal Q}_1}),k+\varepsilon_S\,\sqrt{ {\cal Q} _2},k-\varepsilon_S\,\sqrt{ {\cal Q} _2}\}
$$
satisfying the mass-shell condition
\begin{equation} \label{mass-shell2}
G^{ab} \varpi_a \varpi_b = - \mu^2 = - \frac 12 {\cal S} ^2 - 4 c
\end{equation}
The quantities $\varepsilon_G$ and $\varepsilon_S$ in the momenta $\varpi_a$
denote some $\pm 1$ signs, that flip upon collisions on the walls.
As above, we can also compute the phase shifts of these modes as they reflect on a wall.
More precisely we find that, for given quantum numbers $ {\cal C} _g$, $ {\cal Q} _1$,
$ { \sigma_{2 3} }$ and ${\cal Q}_2$,
the phase shift $\alpha_g$ of $F_1$ as it reflects on a gravitational wall, and the
phase shift $\alpha_s$ of $F_2$ as it reflects on a symmetry wall, are respectively given by:
\begin{eqnarray}
e^{i\,\alpha_g}&=&-\frac{\Gamma(1/2-{\cal C}_g-i\,\sqrt{{\cal Q}_1})\,\Gamma(1/2+2\,i\,\sqrt{{\cal Q}_1})}{\Gamma(1/2-{\cal C}_g+i\,\sqrt{{\cal Q}_1})\,\Gamma(1/2-2\,i\,\sqrt{{\cal Q}_1})}\\
e^{i\,\alpha_s}&=&-\frac{\Gamma(1/2+{ \sigma_{2 3}/2 }-i\,\sqrt{ {\cal Q} _2})\,\Gamma(1/2+2\,i\,\sqrt{ {\cal Q} _2})}{\Gamma(1/2+{ \sigma_{2 3} /2}+i\,\sqrt{ {\cal Q} _2})\,\Gamma(1/2-2\,i\,\sqrt{ {\cal Q} _2})}\\
\end{eqnarray}
As above, we can deduce from these results the intrinsic phase shifts due to the
spin dependence of the walls by subtracting the phase shift of the spin zero mode.
Let us finally note that the existence of a non-zero mass term
$\mu^2 = {\cal S} ^2/2 + 4 c$ in the Wheeler-DeWitt equation can lead
to an interesting phenomenon (whose classical analogue is discussed in Appendix C).
Indeed, a strictly positive $\mu^2$ (e.g. corresponding to $c=0$ and $ {\cal S} ^2 \neq 0$)
forces the classical trajectory of the wavepacket to stay time-like in $\beta$-space,
i.e. prevents it to reach the $\beta$-space light-cone. Therefore, such a mass term
constitutes a kind of potential wall that prevents the wavepacket to reach the light-cone.
We can therefore think of the dynamics described by the Bianchi-II Hamiltonian above
as that of a quantum particle moving in a three-dimensional Lorentzian space,
and confined by {\it three} different walls: the two spacelike walls $\alpha_{123}$
and $\alpha_{23}$ (that prevent the particle from going on the negative sides
of those spacelike walls), and a third effective $\mu^2$ wall that prevents the
particle, after it has bounced on the spacelike walls and aims towards
the light-cone, to reach the light-cone. These three walls thereby define
a kind of waveguide that oblige the particle to move in a time-like direction,
which is somewhere midway between the spacelike walls and the light-cone.
The interesting consequence of this waveguide phenomenon is that it can trap
the particle in a bound state, confined between all these walls.
This happens when both $ {\cal Q} _1$ and $ {\cal Q} _2$ are negative (in view
of the mass-shell condition (\ref{mass-shell1}) this happens only when
$\mu^2> 2 k^2 >0$). In that case, one should no longer consider
scattering states of the Morse-potential equations above, but rather
{\it bound states} in the Morse potentials. It is well known
(see, e.g., \cite{Landau-LifchitzMQ}) that these
bound states occur for the following quantization conditions
\begin{equation}
\sqrt{-{\cal Q}_1}=-n_1-\frac 12 + {\cal C}_g\, ,
\end{equation}
and
\begin{equation}
\sqrt{- {\cal Q} _2}=-n_2-\frac 12(1 +{ \sigma_{2 3} })\, .
\end{equation}
where $n_1$ and $n_2$ must be natural integers (starting with 0).
For instance, in the minimal case where the ordering constant is simply the
naive value $c=0$, there will exist only one such bound state, namely
the ground state $n_1=n_2=0$, with $ {\cal C} _g = -\sigma_{23}/2 = 1/4$, $ {\cal Q} _1= {\cal Q} _2=-1/16$,
and $k=\pm 1/4$. This solution furnishes an interesting example of a quantum
cosmological ground state associated to a Bianchi-II billiard.
Note that we have discussed here a state which is bound simultaneously within
the two separate Morse-potential equations associated to the gravitational
and symmetry walls. There can also exist semi-bound states, i.e. states
which are bound w.r.t, say, the gravitational-wall Morse-potential, but
which represent scattering states w.r.t. the symmetry-wall potential.
\section{Conclusions}\label{Conc}
We have studied the minisuperspace quantization of spatially homogeneous (Bianchi)
cosmological universes sourced by a Dirac (or Majorana) spinor field. In the main text
we used a formulation of the spinor dynamics in which the local $SO(3,1)$ local Lorentz
symmetry of the vielbein is gauge-fixed from the start. [Appendix A compares this
approach to the one where one does not initially fix the vielbein.] In the Bianchi types
IX and VIII (corresponding to simple homogeneity groups $G$) we fixed the $SO(3)$
freedom in the dreibein by using the existence of a three-dimensional automorphism group
of the Lie algebra of $G$. In the Bianchi type II case, we fixed the $SO(3)$
freedom of the dreibein by using an Iwasawa decomposition (which happens to be
compatible with the automorphism group of the corresponding $G$).
In the Bianchi types IX and VIII, the quantum version of the three diffeomorphism constraints
means that the wavefunction does not depend on the three Euler angles parametrizing
the (pseudo-)orthogonal matrix $S \in SO(3) $ (for type IX) or $S \in SO(1,2)$ (for type VIII),
entering the Gauss decomposition $g = S^T {\rm diag} S$ of the metric $g$. This is the
quantum version of the classical possibility of restricting $g$ to a diagonal form,
$ {\rm diag} = {\rm diag}( e^{-2 \beta_1}, e^{-2 \beta_2}, e^{-2 \beta_3} )$ .
The quantization of the homogeneous spinor (denoted $\chi$ after a rescaling)
leads to a finite-dimensional fermionic Hilbert
space, which means that the wavefunction of the universe, which, in the bosonic case, has only one component, becomes multi-component in presence of a spinor field. In addition, in the
Majorana case, the four components of the wavefunction can be identified with the four
components of a spinor of an Euclidean $O(4)$.
The multi-component Wheeler-DeWitt equation satisfied by the wavefunction $\Phi$ is
similar to the second-order form of the Dirac equation coupled
to an external electromagnetic field ($ \left[ (i \, \partial_{\mu} - e \, A_{\mu})^2 + \frac{1}{2} \, e \, \sigma^{\mu\nu} \, F_{\mu\nu} + m^2 \right] \psi = 0$), namely it has a structure
of the form (for types IX and VIII)
\begin{equation} \label{WDWconcl}
\biggl( - \frac{1}{4} \, G^{ab} \, \partial_{\beta_a} \, \partial_{\beta_b} +V_0(\beta) + \sum_\alpha V_\alpha(\beta) \widehat{\Sigma}_\alpha( \chi) + \sum_{\alpha'} V_{2 \alpha'}(\beta) (\widehat{\Sigma}_{\alpha'}( \chi))^2+ \frac{1}{4} \, \widehat{\mu}_{\rm tot}^2 (\chi) \biggl) \Phi (\beta , \sigma) = 0 \, .
\end{equation}
where $\widehat{\Sigma}_\alpha( \chi) = i \chi^{\dagger} J^s_\alpha \chi$,
and $\widehat{\Sigma}_{\alpha'}( \chi)= i \chi^{\dagger} J^s_{\alpha'} \chi $ are
some bilinears in $\chi$, and where $\alpha$ and $\alpha'$ run over some sets of
linear forms in $\beta$ (or `roots'). There exist a limit in $\beta$ space (the far-wall, or BKL, limit) where all the potentials $V_A(\beta)$ tend exponentially towards zero. [The existence
of this limit defines the separation between the $V_A(\beta)$-terms and the squared-mass term $ \frac{1}{4} \, \widehat{\mu}_{\rm tot}^2 (\chi)$.]
The main features of this multi-component Wheeler-DeWitt equation are the following.
The squared-mass term $\widehat{\mu}_{\rm tot}^2 (\chi)$ is a quantum effect, which is of
order $\hbar^2$, and which is affected by several sorts of quantum ordering ambiguities.
We discussed the fact that it was different in the original minisuperspace Einstein-Dirac theory compared to the spin-1/2 Kac-Moody coset proposed in \cite{dBHP}. This suggests that
we should choose an ordering (and additional terms, such as (\ref{eq3.19})) such that
the total squared-mass term vanishes. On the other hand, if we do not do so, but use
instead the naive ordering that looks natural in the quasi-Gaussian spacetime gauge
$N=\sqrt{g}$, one gets a specific prediction for $\widehat{\mu}_{\rm tot}^2 (\chi)$.
One then finds that this quantum (spin-dependent) operator admits the eigenvalue zero
in a fraction of the total Hilbert space. This ensures that a part of the total wavefunction,
i.e. a part of the total quantum reality, will formally continue to behave chaotically
near the singularity, in contrast with the case where the spinor source is treated
classically, where $\mu^2$ is a strictly positive c-number \cite{BK} (see (\ref{eq3.32})).
We discussed in some detail the physical effects linked to the other terms in
the multi-component Wheeler-DeWitt equation (\ref{WDWconcl}). We studied in
particular the spin-dependent terms of the form $\sum_\alpha V_\alpha(\beta) \widehat{\Sigma}_\alpha( \chi) $.
Such terms appear both in the Bianchi IX and VIII cases, and in the Bianchi II one.
In the Bianchi IX and VIII cases the set of roots
$\alpha$ entering these terms are the three gravitational roots, while, in the
Bianchi II case there appear both a gravitational root and a symmetry root
(see (\ref{QGenEq})). When combining these spin-dependent terms with the
corresponding potential terms $\sim e^{- 2 \alpha(\beta)}$ of the spinless potential
$V_0(\beta)$ involving the same root $\alpha$, we found that they lead to a
Schr\"odinger equation in a Morse potential. By studying the quantum scattering
on such spin-dependent Morse potentials we could relate the quantum dynamics
of wavepackets reflecting on them to a previous study of fermionic billiards,
which used Grassmann-valued spinor fields. The quantum spin dynamics
of Bianchi IX and Bianchi VIII happens to be rather trivial because all the corresponding
spin-dependent couplings $\widehat{\Sigma}_\alpha( \chi) $ can be simultaneously
diagonalized. A more interesting spin dynamics would, however, be obtained in more
complicated models (e.g. in higher- dimensions) where the various
$\widehat{\Sigma}_\alpha( \chi) $'s do not commute among themselves.
We also studied in detail the Bianchi II model. For this case we could provide the
exact general solution of the quantum dynamics. It can indeed be decomposed
in separated modes, which can all be expressed in terms of confluent hypergeometric
functions. Some of these solutions describe wavepackets reflecting on the gravitational
and symmetry wall forms $\alpha_{123}(\beta)$ and $\alpha_{23}(\beta)$ of the Bianchi II model,
while other solutions (present when $\mu^2 > 0$) can describe interesting bound
states, trapped between the walls $\alpha_{123}(\beta)$ and $\alpha_{23}(\beta)$,
and the effective wall generated by the positive squared-mass term.
Note finally that the Appendices provide more details about several formal aspects
of the gravity-spinor Hamiltonian dynamics, as well as a study of the classical
limit of this dynamics.
\section*{Acknowledgments}
We thank Marc Henneaux for a clarifying discussion. Philippe Spindel is grateful to IHES,
where an important part of this work was elaborated, for its kind hospitality.
This work was also partially supported in part by IISN-Belgium (convention
4.4511.06), and by ``Communaut\'e fran\c caise de Belgique - Actions de Recherche
Concert\'ees''.
|
\section{Introduction}
Soft Gamma-ray Repeaters (SGRs) are neutron stars that are identified
by the repeated bursts they emit in hard X-rays and soft
gamma-rays. During their burst active phases, SGRs emit anywhere from
a few to thousands of short bursts, typically lasting a fraction of a
second. Energy released during such a short time is very large,
ranging from $\sim 10^{37}$~erg to 10$^{40}$~erg. On rare occasions,
SGRs emit extremely energetic giant flares, lasting for a few hundreds
of seconds and releasing a total energy in excess of $10^{44}$~erg
(Palmer et al. 2005).
SGRs also display persistent bright X-ray emission with luminosities
of the order of L$_{\rm X}\lesssim 10^{35}$~erg~s$^{-1}$, which
significantly exceeds the spindown power that can be generated by
these slowly rotating neutron stars. Both the persistent X-ray
emission and the energetic bursts led to the interpretation of SGRs as
extremely magnetized neutron stars, or magnetars (Duncan \& Thompson
1992; see Woods \& Thompson 2006 for a detailed review). Within the
magnetar paradigm, the decay of a very strong magnetic field
($10^{14}-10^{15}$~G) powers the persistent emission from SGRs
(Thompson \& Duncan 1996), while the observed bursts are attributed
either to cracking of the neutron crust that is strained by magnetic
stress (Thompson \& Duncan 1995) or to magnetic reconnection (Lyutikov
2003).
Large period derivatives, of the order of $\dot{P} \simeq
10^{-11}$~s~s$^{-1}$ measured from numerous SGRs in the past indicated
large inferred dipole spindown fields $B_{\rm dip} = 10^{14} (P/5 {\rm
s})^{1/2} (\dot{P}/10^{-11}\,{\rm s}\,{\rm s}^{-1})^{1/2}$~G, thus
lending further support to the magnetar character of these sources
(e.g., Kouveliotou 1998). When models of high magnetic field neutron
star atmospheres and magnetospheres were used to fit to the continuum
X-ray spectra of magnetars, these analyses also yielded magnetic field
strengths that are comparable to the inferred dipole fields with a
typical range of 2 $ -$ 5 $\times$ 10$^{14}$ Gauss (e.g., G\"uver et
al.\ 2007, 2008; \"Ozel, G\"uver, \& G\"o\u{g}\"u\c{s}\ 2008; Ng et
al.\ 2010; G\"o\u{g}\"u\c{s} et al.\ 2011).
The ubiquitous presence of large period derivatives in SGRs is now
being challenged by the unusually small period derivative measured in
SGR~0418+5729. Early attempts to obtain its period derivative, and therefore,
to establish its inferred dipolar magnetic field strength, were
inconclusive (Woods et al. 2009; Kuiper \& Hermsen 2009; Esposito et
al. 2010). Using multi-satellite observations spanning over 440 days
following the onset of the outburst from this source, Rea et
al. (2010) recently reported a 2$\sigma$ upper limit to the period
derivative, implying that the inferred dipolar magnetic field strength
of SGR~0418+5729\ should be less than $7.5 \times 10^{12}$~G.
SGR 0418+5729 is a “regular” SGR in every other way and a transient
magnetar candidate (Rea \& Esposito 2011): It was discovered on 2009
June 5 by emitting two bursts detected with GBM on board Fermi
Gamma-ray Space Telescope (van der Horst et al. 2010). The energy
released by these two events were modest, totaling $8 \times
10^{36}$~erg and $4 \times 10^{37}$~erg in the 20$-$200 keV band.
Rossi X-ray Timing Explorer observations of the source following the
discovery revealed the 9.07~s spin period (G\"o\u{g}\"u\c{s} et al.
2009). Detectors on board Swift, Fermi, and RXTE have allowed a
growing number of transient magnetars to be discovered in the recent
years, which are first detected via bursting activity and accompanied
persistent flux increase up to several orders of magnitude.
Persistent flux of these transient sources decreases back to quiescent
level ($\approx$ 10$^{-13}$ erg~cm$^{-2}$~s$^{-1}$) on a time scale of
months to years (Rea \& Esposito 2011). X-ray output variation of
SGR~0418+5729\ also very much resembles that of transient magnetar candidates:
the 1$-$10~keV flux rapidly increased in conjunction with bursting and
subsequently decayed by a factor of 10 over a time span of $\sim$150
days (Esposito et al. 2010). SGR~0418+5729\ is located towards the Galactic
anti-center and likely to be in the Perseus arm or the outer-arm of
our Galaxy with a distance of $\approx$2 kpc or higher (van der Horst
et al.\ 2010).
In this paper, we aim to place constraints on the magnetic field
strength of SGR~0418+5729\ using X-ray spectroscopy. To this end, we analyze
the high signal-to-noise X-ray spectrum of SGR~0418+5729\ obtained from
XMM-Newton observations. We use a number of different X-ray spectral
models to constrain the physical properties of the neutron star. In
particular, we employ (i) the phenomenological blackbody plus
power-law model that has been traditionally used to fit SGR spectra,
(ii) the low-to-moderate magnetic field Neutron Star Atmosphere model,
or NSA (Pavlov et al. 1995; Zavlin, Pavlin, \& Shibanov 1996), and
(iii) the high magnetic field Surface Thermal Emission and
Magnetospheric Scattering (STEMS) model (G\"uver, et al. 2007,
2008). We find that the thermal spectrum of SGR~0418+5729\ is best described by
high magnetic field model. We also compare the magnetic field
determined spectroscopically to the dipole magnetic field inferred
from spindown. Because the spectroscopically determined field probes
the strength measured on the neutron star surface, whereas the dipole
field is the one inferred at the light cylinder, this comparison can
be revealing for the field geometry of SGR~0418+5729.
We introduce the XMM-Newton data and our data reduction procedure in
the next section. In \S 3, we fit the spectrum with the three
different continuum models. Finally, we discuss the implications of
the inferred surface magnetic field strength and compare our findings
for this source with those of SGRs in general to address the possible
differences between them in the final section.
\section{Observations and Data Analysis}
In the period following its discovery through its X-ray/soft-gamma-ray
bursts, SGR~0418+5729\ could not be observed immediately with pointing
telescopes due to its unfavorable sky location. In particular, Swift
XRT observations of the source could start about a month after the
discovery, while the XMM-Newton observation was performed about two
months after. Because the X-ray intensity of the source steadily
declined following the outburst (see Figure 1 of Esposito et
al.~2010), only the large collecting area of XMM-Newton was able to
provide sufficiently high quality X-ray spectrum for the measurement
of spectral parameters. We, therefore, use in this study the archival
XMM-Newton observation that took place on 2009 August 12 (Obs ID.
0610000601). We note that SGR~0418+5729\ was still not in quiescence during
the XMM-Newton observation (Esposito et al.~2010). During this
observation, the source count-rate as observed with the EPIC-pn and
MOS detectors were, 1.35 and 0.5 c~s$^{-1}$, respectively. EPIC-pn
and MOS detectors were operated in small and partial window modes with
time resolutions of 6 and 300 ms, respectively, to prevent photon
pile-up.
Part of the XMM-Newton observation, especially towards the end, was
severely affected by large particle flares. We are, therefore, able to
use $\approx$36~ks out of the total exposure time of 65~ks. We
utilize data collected with both the EPIC-pn and EPIC-MOS cameras.
We used the {\it epproc} and {\it emproc} tasks for the EPIC-pn and
EPIC-MOS data with the Science Analysis Software (SAS) version 10.0.0
and the latest available calibration files as of December 2010. We
extracted X-ray spectra using a circular region with a radius of 32
arc-seconds centered on the source. Similarly, background regions
were selected from a source free region with a typical radius of 50
arc-second. We used the {\it rmfgen} and {\it arfgen} tools to create
response and ancillary response files. We rebinned the X-ray spectra
to have at least 30 counts per bin and not to oversample the intrinsic
energy resolution of the detectors by more than a factor of 3.
\section{X-ray spectral analysis and results}
\label{spec}
We fit the data with XSPEC version 12.5.1n (Arnaud, 1996). In all the
fits, we used the {\it tbabs} model and the ISM abundances (Wilms et
al.\ 2000) to account for the effects of interstellar absorption. The
XMM-Newton Calibration team reported that the EPIC-MOS detectors
measure 5-8\% higher fluxes than the EPIC-pn consistently over the
whole energy band\footnote{see, e.g.,
http://xmm.vilspa.esa.es/docs/documents/CAL-TN-0018.pdf.}. However,
the exact amount of this excess flux varies over different energy
ranges. Therefore, in all our fits we allowed the normalization
parameters of the models to be free between the detectors and added a
1\% systematic uncertainty to take into account the possible energy
dependence of the uncertainties in instrumental calibrations. We
calculated the unabsorbed flux using the {\it cflux} model in XSPEC.
Because of the above-mentioned calibration uncertainty, we only report
the flux and/or emitting area values as measured with EPIC-pn. We
assumed a gravitational redshift for the neutron star as 0.25,
corresponding to a neutron star mass of 1.4 M$_{\sun}$ and radius of
11.5 km. We fit all spectra in the 0.5 to 6.5 keV energy range.
Unless otherwise noted, all the uncertainties quoted are for 68\%
confidence interval.
\begin{figure*}
\centering
\includegraphics[scale=0.38, angle=270]{f1a.ps}
\includegraphics[scale=0.38, angle=270]{f1b.ps}
\caption{Best fit fluxed blackbody + power-law (upper left panel) and
neutron star atmosphere (NSA) + power-law (upper right panel) fit to
the X-ray spectrum of SGR~0418+5729\ obtained with the XMM-Newton. The
magnetic field strength in the NSA model is set to $10^{12}$~G.
Blackbody + power-law provides a moderate fit with $\chi^2 / {\rm
dof} = 1.12$ for 346 degrees of freedom, while the NSA + power-law
is a poor fit with $\chi^2 / {\rm dof} = 2.78$ for 346 degrees of
freedom. In both cases, the thermal component dominates (shown with
dashed green lines) over the power-law component (shown with thick
cyan lines) at high photon energies, i.e., in the $>1.5$~keV range,
which is difficult to interpret physically. Lower panels show the
residuals from each fit, respectively.}
\label{unfolded}
\end{figure*}
We first fit the X-ray spectra using a combination of a blackbody
({\it bbodyrad} as defined in XSPEC) and a power-law model. We
performed this analysis only as a phenomenological classification of
the source spectrum because the atmospheric structure even for a weak
or ``zero'' magnetic field has a strong effect on the spectral shape
of the X-ray emission of neutron stars and has been shown to distort
it away from a blackbody (see e.g. Romani\ 1988; Pavlov et al.\
1995). The spectrum and the best-fit model are shown in the left panel
of Figure~\ref{unfolded}. Blackbody plus power-law provided a
moderate fit with a $\chi^{2}$/dof of 1.12 for 346 degrees of
freedom. The best fit model resulted in a hydrogen column density of
$(1.10 \pm 0.05) \times 10^{22}$~cm$^{2}$, a blackbody temperature of
$0.93 \pm 0.006$~keV, and a photon index of the power-law component of
$3.18 \pm 0.19$. The temperature is unusually high for two reasons.
First, the blackbody component dominates over the power-law component
at high photon energies, i.e., in the $>1.5$~keV range, as can be seen
in Figure~\ref{unfolded}, which is physically difficult to interpret
and is also contrary to what is seen in the spectra of other SGRs.
Second, this high temperature corresponds to an extremely small
emitting radius of 0.18~km/2kpc. Such a small emitting radius is
especially hard to interpret given the observed pulsed fraction of
20-40\% in the soft X-rays (Esposito et al.\ 2010).
The current constraints on the period derivative of SGR~0418+5729\ point to a
dipolar magnetic field strength smaller than $7.5 \times 10^{12}$~G.
We, therefore, tried to fit the spectrum of SGR~0418+5729\ with low to
intermediate magnetic field strength neutron-star atmosphere models.
The NSA model (Pavlov et al. 1995; Zavlin, Pavlin, \& Shibanov 1996)
provides a model for the X-ray spectra emitted from a hydrogen
atmosphere of a neutron star at three different magnetic field
strengths: $B < 10^{8} - 10^{9}$~G, $B = 10^{12}$~G, and $B =
10^{13}$~G. The {\bf fit} parameters of the NSA model, in addition to
the magnetic field strength, are the neutron star mass and radius, the
surface temperature, and the normalization, which is a function of the
source distance $1/D^2_{\rm kpc}$.
We consider the $B< 10^8$~G case only for completeness, because it
would be unfeasible to account for the X-ray pulsations observed from
SGR~0418+5729\ if indeed possessed a negligible magnetic field. For this case,
the NSA model does not provide a good fit with a $\chi^{2}$/dof equal
to 1.29 for 350 degrees of freedom and yields a best-fit surface
temperature of 0.73~keV ($8.5 \times 10^6$~K). As with the
blackbody-plus-power-law fit, this rather high temperature corresponds
to a very small best fit normalization of $3.0 \times 10^{-10}$, or
equivalently an emitting radius of $R = 0.4$~km/2~kpc.
At a field strength of $B=10^{12}$~G, the NSA model cannot fit the
spectrum of SGR~0418+5729, yielding a minimum $\chi^{2}/{\rm dof} = 13.88$ for
350 degrees of freedom. The surface temperature also hits $10^7$~K,
which is the maximum of the allowed range in the models. Finally, the
obtained best fit normalization of $8.5 \times 10^{-11}$ translates to
an unphysically small emitting radius of $R = 0.21$~km/2~kpc. This
indicates that the spectrum of SGR~0418+5729\ is inconsistent with a surface
magnetic field strength that is comparable to its dipole field
strength inferred from its period derivative.
We performed a final NSA fit where we set the magnetic field strength
to $B = 10^{13}$~G. The quality of the fit improved compared to the
$10^{12}$~G case but is still not adequate, yielding a $\chi^{2}/{\rm
dof} = 3.88$ for 350 degrees of freedom. The best-fit temperature
again hits the upper limit of the allowed range at $10^7$~K
(0.86~keV), corresponding to a normalization of $1.3 \times 10^{-10}$,
which is an emitting radius of $R=0.3$~km/2~kpc.
In an attempt to obtain statistically better fits and to investigate
whether additional spectral components would have an effect on the
measured effective temperature values, we also modeled the X-ray
spectrum of SGR~0418+5729\ with an absorbed NSA plus a power-law model. We
found that the addition of the power-law component statistically
improved the fits, yielding $\chi^{2}$/dof values of 1.13, 2.78, and
1.49 for 346 degrees of freedom, when the magnetic field strength was
set to 0, 10$^{12}$, and 10$^{13}$~G, respectively. However, the
addition of the power-law component did not decrease the inferred
effective temperature, resulting in values that still reached the
upper temperature limit of the model. In the right panel of
Figure~\ref{unfolded}, we show the X-ray spectra and an example case
for the NSA+PL models, where the strength of the magnetic field was
set to $10^{12}$~G.
We finally modeled the spectrum of SGR~0418+5729\ with magnetar strength fields
in the few $\times 10^{13}-10^{15}$~G range, which may reflect the
higher multipole field strengths present at the stellar surface. To
this end, we used the Surface Thermal Emission and Magnetospheric
Scattering model (STEMS, see G\"uver et al. 2007, 2008). STEMS takes
into account the effects of the ultrastrong magnetic fields on the
fully ionized hydrogen atmospheres of magnetars (\"Ozel 2001, 2003)
and the resonant cyclotron scattering of the surface photons by mildly
relativistic charges in the magnetosphere (Lyutikov \& Gavriil
2006). For the calculation of the surface emission, we follow \"Ozel
(2001, 2003) and solve the radiative transfer equations in a
polarization-mode-dependent manner including the absorption, emission
and scattering processes in the atmosphere. We also take into account
the effects of vacuum polarization resonance and include the
interaction of photons with protons in the plasma, which gives rise to
absorption features at the proton cyclotron energy (\"Ozel 2003). In
the stellar magnetosphere, we incorporate a treatment of resonant
cyclotron scattering, following Lyutikov \& Gavriil (2006). Resonant
cyclotron scattering can take place in a neutron star magnetosphere as
long as there is a sufficient density of moderately relativistic
electrons and the resulting up-scattering shifts the initial spectrum
to higher energies and smears out the proton cyclotron features
(Lyutikov \& Gavriil 2006).
In the STEMS model, we use the emission emerging from the surface as
input for the resonant cyclotron scattering model to obtain the
resulting energy distribution of photons. In total STEMS depends on
four parameters: the surface effective temperature ($kT = 0.1-
0.6$~keV), the strength of the magnetic field at the surface ($B = 0.6
- 50 \times 10^{14}$~G), the optical depth to scattering in the
magnetosphere ($\tau = 1.0 - 12.0$), and the velocity of the particles
in the magnetosphere ($\beta = 0.1 - 0.7$). Note that, even though in
the calculations of Lyutikov \& Gavriil (2006), a high optical depth
is motivated by a twist angle, this is not a necessary assumption.
Such high charge densities are seen in the magnetospheres of even
normal pulsars and the only difference between the SGR~0418+5729 ~and a neutron
star with a magnetar-strength dipole field would be the occurrence of
the resonant layers closer to the neutron star surface. As an
example, based on the current limit on the dipole field of SGR~0418+5729, the
resonant layer would be at 4~R$_{NS}$, whereas for a quadrupole field
that is 10$^{14}$~G at the surface, the resonant layer would be at
$\approx$6~R$_{NS}$. Finally we take into account general
relativistic effects on the propagation of the photons by assuming a
gravitational redshift.
Compared to the blackbody plus a power-law fit, STEMS model provided
an equally good fit to the XMM-Newton spectrum, resulting in a
$\chi^{2}$/dof of 1.18 for 347 degrees of freedom. We show the best
fit model curve and the X-ray spectra in Figure~\ref{spectra}. The
parameters of the best fit model were: the surface magnetic field
strength $B = (1.00_{-0.01}^{+0.02}) \times 10^{14}$~G, the surface
effective temperature $kT~=~0.246_{-0.01}^{+0.003}$~keV, the
scattering optical depth in the magnetosphere $\tau =
8.94_{-0.18}^{+1.18}$ and the particle velocity in the magnetosphere
$\beta = 0.56\pm0.01$. The inferred optical depth and the average
electron velocity are well within the assumptions of the resonant
cyclotron scattering model (see. e.g., Section 3.4 of Lyutikov \&
Gavriil 2006). We also found a hydrogen column density of $(0.74 \pm
0.02) \times 10^{22}$~cm$^{2}$ and the unabsorbed 0.5-6.5 keV source
flux as $(8.53_{-0.06}^{+0.05})\times10^{-12}$~erg~s$^{-1}$~cm$^{-2}$.
The inferred emitting radius corresponding to this effective
temperature and flux is $R=2.98$~km/2~kpc. We note that in this fit
fractional emitting area, i.e., $A_{\rm hot}/A_{\rm NS}$, where
$A_{\rm NS}$ is the entire neutron star surface area, is in the right
range to produce the pulsed fraction of 20-40\%, as observed by
Esposito et al.\ (2010) in the soft X-rays.
We present in Figure~\ref{confidence} the confidence contours of the
best fit effective temperature and the magnetic field strength at the
surface of the neutron star. From the confidence contours it can
clearly be seen that the STEMS model provides a lower limit to the
surface magnetic field strength of the neutron star of $\approx 1
\times 10^{14}$~G. Because our focus in this paper is the surface
magnetic field strength of SGR~0418+5729, we further explored the uncertainty
in this parameter by doing the following: We fit the data by freezing
the magnetic field at 1000 different points within the $0.6-4.0 \times
10^{14}$~G range and allowing the other parameters to vary.
Figure~\ref{confidence} shows the variation of the $\chi^{2}$/dof over
the full magnetic field range investigated. It is evident from the
distinct minimum in the $\chi^2$ in this figure that the magnetic
field strength at the surface of the neutron star is uniquely
constrained and is found to be $1.0 \times 10^{14}$~G.
\begin{figure*}
\centering
\includegraphics[scale=0.35, angle=270]{f2.ps}
\caption{XMM-Newton EPIC-pn (blue) and EPIC-MOS (green and cyan)
spectra of the SGR~0418$+$5729. Best fit STEMS model is also shown
with red thick lines. Residuals from the model for each detector are
shown in lower panels.}
\label{spectra}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale=0.25]{f3a.ps}
\includegraphics[scale=0.25]{f3b.ps}
\caption{{\it Left Panel:} Confidence contours of the best fit surface
effective temperature and the magnetic field strength at the surface
of the neutron star as inferred from the STEMS model. {\it Right
Panel:} Variation of the $\chi^{2} / {\rm dof} $ with the surface
magnetic field strength. Horizontal and vertical dashed lines show
$\chi^{2}$/dof and surface magnetic field strength for the best fit
model.}
\label{confidence}
\end{figure*}
\section{Discussion}
The recent discovery and the subsequent observations of SGR~0418+5729\ have
raised a number of questions on our understanding of the SGRs and
AXPs, especially owing to a lack of secular evolution in its spin
period. A number of explanations, both within and outside the context
of the magnetar model, have been proposed to interpret the peculiarity
of SGR~0418+5729. For example, the effects of a stronger toroidal magnetic
field below the surface on the neutron star crust was employed to
create the observed X-ray bursts independent of strength of the dipole
component (see, e.g., Perna \& Pons 2011; Rea et al.\ 2010). The
existence of a fallback disk has been proposed to evolve the neutron
star to its current spin period with a dipole field as low as
10$^{12-13}$~G (Alpar, Ertan, \& {\c C}al{\i}{\c s}kan 2010; see also
Ertan et al.\ 2007).
In this paper, we analyzed the X-ray spectrum of SGR~0418+5729\ to constrain
its surface magnetic field strength, as well as the temperature and
the magnetospheric parameters of the neutron star. We find that the
empirical blackbody + power-law model provides a moderate fit, but the
emitting area inferred from this fit is unphysically small ($R=
0.18$~km/2~kpc). In addition, contrary to the physical expectations
and in contrast to other AXP and SGR spectra, the blackbody component
dominates over the power-law component in the higher, as opposed to
the lower, energy regime. We also found that realistic atmosphere
models of a neutron star with moderate magnetic field strengths (NSA),
do not describe the spectrum adequately, yielding $\chi^2/{\rm dof}
\geq 3.9$. In contrast, neutron star atmosphere models with
magnetar-strength fields (STEMS) produces a fit with $\chi^{2}$/{\rm
dof} $=$1.18 for 347 degrees of freedom degrees of freedom. The
best fit value of the magnetic field strength is $1.0 \times
10^{14}$~G. Thus, the spectral analysis strongly disfavors X-ray
models where the magnetic field strength at the surface is assumed to
be 10$^{8-9}$, 10$^{12}$, or 10$^{13}$~G.
X-ray observations of a number of other magnetars have been
successfully modeled with neutron star atmospheres with high magnetic
field strengths, including XTE~J1810$-$197 (G\"uver et al.\ 2007),
4U~0142+61 (G\"uver et al.\ 2008), 1E~1048.1$-$5937,
1RXS~J170849.0$-$400910 (\"Ozel et al.\ 2008), 1E~1547.0$-$5408 (Ng et
al.\ 2011) and SGR~1900+14 (G\"o\u{g}\"u\c{s} et al.\ 2011), and were
used to obtain the surface parameters of the neutron stars such as
their magnetic field strengths and effective temperatures. The
spectroscopically inferred magnetic field strength values for these
sources, in units of 10$^{14}$ G, are 2.77$\pm$0.05, 4.75$\pm$0.03,
2.26$\pm$0.05, 3.96$\pm0.17$, 3.1$\pm$0.5, and 5.0$\pm$0.48,
respectively. In Figure \ref{comp}, we compare these
spectroscopically determined field strengths with the previously
reported inferred dipole magnetic field for each source, which we
obtained from the McGill SGR/AXP
Catalog\footnote{http://www.physics.mcgill.ca/~pulsar/magnetar/main.html},
Ng et al.\ (2011), and \"Ozel et al.\ (2008). For all of these
sources, the spectrally inferred surface magnetic field strength {\bf
is} in good agreement with dipole magnetic field estimates,
differing at most by a factor of four. SGR~0418+5729\ is the first magnetar
candidate for which the surface magnetic field strength is
significantly larger than the limit on the dipole magnetic field
($\ge$~15 times larger). It is also interesting that the surface
magnetic field strength we report here for SGR~0418+5729\ is the lowest among
other STEMS measurements for other AXPs and SGRs so far.
\begin{figure*}
\centering
\includegraphics[angle=270,scale=0.5]{f4.ps}
\caption{Comparison of the magnetic field strengths as inferred using
the STEMS model to the dipole fields deduced from the spindown
properties for seven magnetar candidates. The error bars in dipole
field strengths represent the range of measured spindown rates for
each source, while the error bars in the spectroscopic magnetic
field strength represent 2-$\sigma$ statistical
uncertainties. Dashed lines show the relation where
$B_{surf}=B_{dip}$.}
\label{comp}
\end{figure*}
Even though the dipole field strengths inferred from spindown involve
a number of simplifying assumptions, relaxing them in realistic
simulations leads to a factor of two difference in the inferred
magnetic field strengths (Contopoulos \& Spitkovsky 2006). Therefore,
the discrepancy between the field strength inferred from the spindown
of SGR~0418+5729\ and the field strength inferred from both its spectrum and
its bursts is too large to be accounted for in this way. Instead, the
difference points to a complex magnetic field geometry, i.e, it can be
attributed to the presence of higher order multipole components at the
surface of the neutron star, which can shape the characteristics of
the X-ray emission but do not contribute to the spindown. The
multipole fields are likely to play a role in determining the pulse
shapes observed in AXPs and SGRs (\"Ozel 2002). In this
interpretation, the higher order components of the magnetic field can
also cause the fracturing of the neutron star crust, leading to the
observed X-ray bursts. Further comparisons of the surface and dipole
magnetic field strengths of AXPs and SGRs, as well as numerical
simulations of the magnetic field evolution in young neutron stars may
help further constrain the magnetic field strengths and geometries of
X-ray bright neutron stars.
\section*{Acknowledgments}
We thank the anonymous referee for his/her comments that improved the
clarity of the manuscript. This work makes use of observations
obtained with XMM-Newton, an ESA science mission with instruments and
contributions directly funded by ESA Member States and NASA. F.\
\"O. and T.\ G.\ acknowledge support from NSF grant AST-07-08640.
|
\section{Introduction}
\label{s_intro}
It is generally accepted that the vast amount of continuum emission of
type 1 active galactic nuclei (AGNs) in UV and optical wavelengths
originates in the accretion disk surrounding a supermassive black hole,
and the UV-optical variability found at the beginning of AGN studies
has been considered a powerful tool for understanding
the nature of the AGN central engine.
However, the mechanism of this variability is still under discussion.
Many models for variability have been proposed, including
accretion disk instabilities \citep{1984ARA&A..22..471R, 1998ApJ...504..671K},
X-ray reprocessing (Krolik et al. 1991; Kawaguchi in prep.),
star collisions (Courvoisier, Paltani, \& Walter 1996; Torricelli-Ciamponi et al. 2000),
and gravitational microlensing (Hawkins 1993);
however, none of these successfully explained more than a few properties
of UV-optical variability \citep{2004ApJ...601..692V}.
The spectral variability of the UV-optical continuum emission
during flux variation
is a key property for understanding the central engines
and their variability mechanisms in AGNs.
For example,
since the spectral energy distribution (SED)
of a hot spot or a flare in the accretion disk
caused by local enhancement of mass accretion
or disk instabilities would be different from
that of the entire disk,
the spectral shape of continuum emission
is expected to vary with the flux variation
when caused by those mechanisms.
On the other hand,
a change of the global mass accretion rate of the disk
and a certain reprocessing model \citep{1999MNRAS.302L..24C}
would not change the temperature distribution
of the accretion disk at larger radii;
thus, the absence of the spectral variation
of optical continuum emission during flux variation
suggests those variation mechanisms.
Sakata et al. (2010; hereafter Paper I)
addressed the spectral variability of optical continuum emission of AGNs.
Paper I examined the long-term multi-band monitoring data of 11 nearby AGNs
and precisely estimated the contaminated flux of the host galaxy
and the narrow emission lines.
Then, it was found that
the multi-epoch optical flux data in any two different bands
obtained on the same night showed a very tight linear flux-to-flux relationship
for all target AGNs
and that the non-variable component of the host galaxy plus narrow lines
was located on the fainter extension of the linear regression line
of multi-epoch flux-to-flux plots.
From these results, Paper I concluded that the spectral shape
of AGN continuum emission in the optical region ($\sim4400-7900$\AA)
does not systematically change during flux variation
and that the trend of spectral hardening
in which the optical continuum emission becomes bluer as it becomes brighter,
which has been reported by many studies
\citep{1990ApJ...354..446W, 1999MNRAS.306..637G, 2000ApJ...540..652W},
is caused by the contamination of the non-variable component
of the host galaxy plus narrow emission lines,
which is usually redder than AGN continuum emission.
In contrast to optical continuum emission, two opposite claims have not been resolved
for the spectral variability of UV continuum emission.
\citet{2004ApJ...601..692V} statistically examined
the properties of flux variations of QSOs
from the two-epoch photometric observations of
about 25,000 QSOs obtained by the Sloan Digital Sky Survey (SDSS)
and found a larger amplitude of variation
at shorter wavelengths in the UV region of $\lambda < 4000$ \AA,
indicating spectral hardening during flux variation of QSOs.
\citet{2005ApJ...633..638W} obtained a composite
of the differential spectrum of two-epoch spectroscopic observations
for hundreds of SDSS QSOs
and found that the composite differential spectrum
was bluer than the composite spectrum of QSOs,
which also indicates spectral hardening of the UV continuum emission.
On the other hand,
\citet{1996A&A...312...55P}
applied principal component analysis (PCA) to
the multi-epoch UV spectra of 15 nearby AGNs
obtained by the {\it International Ultraviolet Explorer} (IUE) satellite
and concluded that the UV flux variation of the AGNs
consists of a variable component with a constant spectral shape
and a non-variable component.
Based on the decomposition of the multi-epoch spectra to a variable
and a non-variable components, they further concluded that
the variable component has a power-law shape of fixed spectral index
and that the non-variable component can be reproduced using
the sum of a steep Balmer continuum and a \ion{Fe}{2} pseudo-continuum,
which corresponds to the spectral feature of the small blue bump (SBB).
\citet{1995MNRAS.274....1S} and \citet{1997ApJS..110....9R}
observed NGC 4593 and Fairall 9, respectively, using the IUE satellite.
They found that the multi-epoch UV flux data in two different bands
obtained on the same night showed a linear flux-to-flux relationship
and that the contaminated flux of the SBB was located on a fainter extension of
the linear regression line.
These authors concluded that
the UV continuum emission retains a constant spectral shape
during the AGNs' flux variation.
In this paper,
we examine the spectral variability of the UV continuum emission of AGNs,
in the same way as in Paper I,
on the basis of the long-term multi-epoch photometric data
of 10 mid-redshift luminous QSOs obtained by the SDSS.
In Section \ref{s_obsdata},
we describe the selection of target QSOs, their basic properties
such as the central black hole mass and accretion rate,
and present their light curves.
In Section \ref{s_result},
we examine the UV color variability of the target QSOs
from the analysis of the flux-to-flux plots
and the contaminated fluxes of the host galaxies.
In Section \ref{s_discuss},
we compare our results with previous observational studies
about the UV color variability of AGNs
and also with an accretion disk model.
In Section \ref{s_summary}, we summarize our results.
We assume cosmological parameters of
$(h_0, \Omega _0, \lambda_0)=(0.73,0.27,0.73)$ throughout this paper.
\section{Multicolor Light Curve of QSOs}
\label{s_obsdata}
\subsection{SDSS Stripe 82 Data}
\label{s_stripe82data}
Stripe 82 is located in the South Galactic Gap
and was scanned multiple times by the SDSS Legacy Survey
in order to enable a deep co-addition of the data and to find variable objects.
It is defined as the region spanning 8 h in right ascension (RA)
from $\alpha$ = 20$^h$ to 4$^h$ and $2^\circ.5$ in declination (Dec.)
from $\delta$ = $-1^\circ.25$ to $1^\circ.25$,
consisting of two scan regions referred to
as the north and south strips.
Both the north and south strips have been repeatedly imaged
in $u$, $g$, $r$, $i$, and $z$ bands about 80 times on average
by more than 300 nights of observations from 1998 to 2007,
with about 70 percent imaging runs obtained after 2005
since when the SDSS-II Supernova Survey started.
There are about 37,000 QSO candidates in the Stripe 82 region
\citep{2009ApJS..180...67R} of which
spectroscopic data were available for about 8,300 candidates.
\subsection{Target Selection}
\label{s_targetselec}
In order to examine rest-UV continuum color variation of QSOs,
we first selected the spectroscopically identified QSOs in Stripe 82
from the QSO candidate list of \citet{2009ApJS..180...67R}.
Then, we selected the targets using the following criteria:
(1) The target redshift is $z=1.05,1.54,1.71,2.35\pm0.05$
in which strong broad lines such as Ly$\alpha$ and \ion{C}{4} do not contaminate
SDSS filters and in which rest wavelengths of the $u$- or $g$-band
are just longer or shorter than
that of the \ion{C}{4} line (around $1400$ \AA \, or $1730$ \AA).
\footnote{When the \ion{C}{4} line is included in the SDSS $u$ or $g$ bands
at a certain redshift, its flux is estimated
as about $30-50$\% of UV continuum flux based on the typical \ion{C}{4} equivalent width
for SDSS QSOs \citep{2008MNRAS.389.1703X}.}
(2) The signal to noise ratio (S/N) of the photometric data
in two bands used for the analysis is more than 30.
The S/N values are taken from \citet{2009ApJS..180...67R}
and are based on the single-epoch of observation.
Finally, we selected 10 luminous QSOs.
Although wide wavelength coverage is preferable for examining the spectral variability,
we basically used the $i$-band data in place of the $z$-band data as the longest wavelength observation
because the S/N value of the former is much higher than that of the latter for most QSOs.
The targets, their redshifts, and the filter selections are listed in Table \ref{t_filterwave}.
We then calculated the black hole mass and the mean Eddington ratio of
accretion rate for our targets.
The black hole mass was estimated using the scaling relations
for the \ion{C}{4} emission line, or the \ion{Mg}{2} emission line
if the \ion{C}{4} emission line was unavailable,
according to \citet{2009ApJ...700...49T}.
The full width at half maximum (FWHM) of the emission line was estimated from
the standard deviation of the emission-line width
taken from the SDSS database,
which was derived by fitting the Gaussian profile to the emission line,
by multiplying a factor of $2\sqrt{2\ln2}$.
We selected the specific luminosity at $1350$ \AA \,
for the scaling relation of the \ion{C}{4} emission line
and at $3000$ \AA \, for that of the \ion{Mg}{2} emission line
where the effective wavelengths in the rest frame of the photometric
bands are neighbored,
estimated by interpolating or extrapolating the fluxes
of the two photometric bands,
and they were averaged over the light curves of the targets.
Strong absorption features in the \ion{C}{4} emission line
were found in three targets
(J0136$-$0046, J0346$-$0042, and J2134$+$0048),
and the error of the black hole mass estimate
might be larger for them.
The bolometric luminosity was calculated using the equation
$L_{bol}=5\lambda L_{\lambda}(3000$ \AA$)$ \citep{2005ApJ...629...61K}
and the Eddington luminosity was calculated as
$L_{Edd}=\frac{4\pi Gcm_p}{\sigma_e}M_{BH}$,
where $G$ is the gravitational constant, $c$ is light speed, $m_p$ is the mass of a proton,
$\sigma_e$ is the Thomson scattering cross-section,
and $M_{BH}$ is the black hole mass estimated as described above.
Then, the mean Eddington ratio of accretion rate, $L_{bol}/L_{Edd}$ was calculated.
The specific luminosities, the standard deviation of the emission
lines, the black hole masses, and the Eddington ratios are listed
in Table \ref{t_objbasicparm}.
The black hole mass of the targets are about $10^{9-10}M_{\odot}$,
and the mean Eddington ratio is between $0.1-1.2$.
We also examined the radio flux of the targets.
As listed in Table \ref{t_objbasicparm},
8 of the 10 targets were not detected by
the Very Large Array (VLA) FIRST survey (Becker, White, \& Helfand 1995),
and the remaining two targets are not covered by
the data of the FIRST survey.
From these estimations of the black hole mass and the Eddington ratio,
and the examination of the radio flux,
the targets can be characterized as
radio quiet QSOs with very massive black holes.
\subsection{Photometry}
\label{s_gooddataselec}
We used the PSF magnitude corrected with Galactic extinction
obtained from the SDSS database for the photometry of the target QSOs
because the photometric error of the PSF magnitude
was smaller than that of the aperture magnitude.
Then, we further examined those data in order to improve
the photometric accuracy of the light curves of the targets.
First, we evaluated the atmospheric condition for photometry
using the QA value and selected the data after 2004,
when the QA value was available. It was provided by the SDSS database
and gives $1\sigma$ fluctuations by the millimag in the recalibrated
g-band camcol 3 zero point for fields in the run.
If this number is zero, then the night was photometric or
no recalibration was done,
and if this number is non-zero,
a large number corresponds to more variable clouds
and worse photometric calibration can be expected
in general.
Then, we adopted only the data whose QA values were less than 50,
that is, $0.05$ mag error of the photometric calibration,
if the QA value was available.
In addition, we examined the atmospheric condition
using the flux of reference stars located near the target,
which should be constant.
When the flux of the reference star for a target QSO at some epoch
was more or less than $3\sigma$ from its mean flux for all available data,
we did not use the photometric data of the target at that epoch.
Even though the QA value was unavailable for the data before 2004,
the selection procedure based on the flux of the reference stars
efficiently rejected the data obtained in bad atmospheric condition.
Finally, the magnitude of the target QSO was measured
relative to the nearby reference stars,
then the magnitudes of the reference stars are added.
One or two reference stars were selected,
which locate in the same field of the target
and are brighter than the target by more than 2 mag.
This relative photometry would reduce the fluctuation
of the flux caused by changing atmospheric conditions
during the observation and would improve the photometric accuracy
of the light curves of the targets.
The light curves of the target QSOs are presented in
Figures \ref{f_lc1} and \ref{f_lc2},
and the parameters of the light curves
are listed in Table \ref{t_filterwave}.
We found that all targets showed significant flux variations
during 7 years of observation, that is,
a time span of $2-3.5$ years in the rest frame of the target QSOs.
\section{Examination of UV Color Variability}
\label{s_result}
\subsection{Flux-to-flux Plot and Linear Fit}
\label{s_ffplotfit}
In order to examine the color variability of the UV continuum emission of QSOs
with flux variation, we applied the flux-to-flux plot analysis
as applied to the color variability of optical continuum emission
of Seyfert galaxies in Paper I, which was originally proposed by
\citet{1981AcA....31..293C}.
We plotted the flux data in two different bands observed on the same night
in the flux-to-flux diagram,
where the flux in longer wavelength is assigned to the $x$-axis
and the flux in shorter wavelength is assigned to the $y$-axis.
The flux-to-flux plots for all target QSOs are presented
in Figures \ref{f_ff1} and \ref{f_ff2}, and
the band pair and its effective wavelength in the rest frame
for each target are listed in Table \ref{t_filterwave}.
The rest wavelength of the shorter-$\lambda$ filter is either $\sim1400$ \AA \,
or $\sim1730$ \AA \,,
while that of the longer-$\lambda$ is between $2200-3600$ \AA.
As shown in Figures \ref{f_ff1} and \ref{f_ff2},
the rest-frame UV flux data in the two different bands
are distributed linearly for all targets,
and especially for those with large flux variation
such a correlation seems very tight.
In order to examine the linear relationship between
the UV fluxes in the two different bands,
both straight-line fitting and power-law fitting to the data
in the flux-to-flux diagram were carried out
using the fitting code made by \citet{tomitaphd}
following the most generalized multivariate
least square process given by Jefferys (1980, 1981).
The fitting functions adopted for the straight-line fitting
and the power-law fitting are
$F_{\nu 1}=\alpha \times (F_{\nu_{2}}-F_{\nu_{2}0})$
and $F_{\nu 1}=\alpha \times (F_{\nu_{2}}-F_{\nu_{2}0})^\beta $, respectively,
because larger contamination of the host galaxy flux is expected
in the longer-wavelength flux ($F_{\nu_{2}}$)
than in the shorter-wavelength flux ($F_{\nu_{1}}$).
The best-fit parameter values and the reduced $\chi^2$ values
of the straight-line fitting and the power-law fitting
are listed in Table \ref{t_fffit}.
As listed in Table \ref{t_fffit},
the residuals of the straight-line fitting
of the flux-to-flux plots in UV wavelengths
are so small that the reduced $\chi^2$ is near unity,
except for J2045$-$0051.
Indeed, according to the $\chi^2$-test for the straight-line fitting,
the linear relationship is not rejected for 6 of the 10 targets
(J0312$-$0113, J0346$-$0042, J2111$+$0024, J2119$+$0032, J2123$-$0050,
and J2134$+$0048) at a 1\% level of significance.
Although the $\chi^2$-test rejects the linear relationship
at a 1\% level of significance for the remaining four targets,
no clear curvature can be seen in their flux-to-flux plots.
Indeed, no significant improvement of $\chi^2$ values
applying the power-law fitting instead of the straight-line fitting
can be found for two of the four targets (J0136$-$0046 and J2045$-$0051).
The larger $\chi^2$ values of the straight-line fitting
of the four targets are probably caused by
another variable component of fluxes that is
not synchronized to the UV continuum emission,
or by underestimation of the photometric errors.
In fact, as will be described in Section \ref {s_effectsbp},
the reduced $\chi^2$ values of the straight-line fitting
are significantly decreased for three of the four targets
(J0105$-$0050, J0105$+$0040, and J2045$-$0051),
and the linear relationship is not rejected at a 1\% level of significance
for two targets (J0105$-$0050 and J0105$+$0040)
when the variable flux component of the SBB is considered.
Based on these considerations,
we conclude that the UV flux of all of the target QSOs
in the two different bands shows a linear correlation
during the flux variations.
\subsection{Location of Host-Galaxy Component in Flux-to-flux Diagram}
\label{s_effectotagncont}
The x-intercept of the best-fit linear regression line,
$F_{\nu_{2}0}$, is positive for 9 of the 10 targets,
as listed in Table \ref {t_fffit}
(with the exception of J2111$+$0024, for which
$F_{\nu_{2}0}$ is consistent to be zero within $2\sigma $ error).
Hence, their rest-frame UV colors of the fluxes
observed in the two different bands become bluer as they become brighter.
However, as presented in the discussion in Paper I,
if the sum of the non-variable flux components, such as that of the host galaxy,
is located on the fainter extension of the best-fit regression lines
of the flux-to-flux plots,
the rest-frame UV color is almost constant during the flux variation.
In this section, we estimate the host galaxy fluxes of the target QSOs
and examine whether the UV colors of the targets
show the trend of spectral hardening
or remain almost constant during the flux variations.
Since the host galaxies of the target QSOs are not
spatially resolved in the SDSS images,
we first estimated the stellar mass of the host galaxy
from the black hole mass estimated in Section \ref {s_targetselec}
and the Magorrian relation \citep{1998AJ....115.2285M},
a tight correlation between the central black hole mass
and the stellar mass of the bulge of the host galaxy.
Next, the host galaxy fluxes in observing bands were converted
from the bulge mass using a mass-to-light ratio
and an SED of stellar population of the host galaxy.
We assumed bulge luminosity as the whole luminosity of
the host galaxy
because the QSOs at $z\lesssim 2$ are mostly hosted
by the bulge-dominated early-type galaxies
(e.g., Bahcall et al. 1997; Hutchings et al. 2002;
Dunlop et al. 2003; Zakamska et al. 2006)
We used the Magorrian relation presented in \citet{2003ApJ...589L..21M},
\begin{equation}
\log M_{\rm BH}= 8.28 + 0.96 \times \log (M_{\rm bulge } -10.9),
\end{equation}
to estimate the stellar mass of the host galaxy from the black hole mass.
The scatter of the correlation was estimated as 0.21 dex by them.
Including the scatter of the black hole mass estimation
based on the scaling relation for \ion{C}{4} against
UV luminosity 0.32 dex presented in \citet{2006ApJ...641..689V}
we estimated the scatter of the stellar mass of the host galaxy
as 0.38 dex, that is, a factor of 2.4 for $1\sigma $ error
and 5.8 for $2\sigma $ error.
Although the redshift evolution of the Magorrian relation
is still uncertain, \citet{2006ApJ...649..616P} suggested that
the $M_{\rm BH}/M_{\rm bulge}$ mass ratio
increases a factor of $\sim 4$ at $z>1.7$
and \citet{2010MNRAS.402.2453D} also showed that
it increases a factor of $\sim 7$ at $z=3$
compared to that of the local universe.
Therefore, the stellar mass, thus the luminosity of the host galaxy
of the targets estimated here, might be systematically overestimated
by a factor of $\sim 5$.
Contrary to the morphology, recent studies show that
the optical colors of AGN host galaxies at $z\lesssim 1.5$
are slightly bluer than those of quiescent early-type galaxies,
indicating a component of newly formed stellar population
(e.g., Jahnke, Kuhlbrodt, \& Wisotzki 2004b;
S\'{a}nchez et al. 2004; Nandra et al. 2007; Silverman et al. 2008).
In addition, at higher redshift,
\citet{Jahnke2004b} showed that
the rest-frame UV colors of host galaxies of QSOs at $1.8<z<2.75$
are bluer than that expected from an old stellar population with
a formation epoch at $z\sim 5$,
suggesting a young stellar population
of a few percent of the total mass.
Schramm, Wisotzki, \& Jahnke (2008) showed that
the rest-frame $B-V$ colors of the host galaxies of three QSOs
at $z\sim 3$ are close to zero,
indicating a substantial contribution from young stars,
and a stellar mass-to-light ratio below 1.
Since Kiuchi, Ohta, \& Akiyama (2009) demonstrated that
most optical SEDs of host galaxies of type-2 AGNs at $0.5<z<1.15$
follow composite SEDs of E to Sbc galaxies of the local universe,
and since \citet{Barger2003} showed that
the obs-frame $R-HK'$ colors of host galaxies of type-2 QSOs at $z=1-2$
are widely scattered between the K-corrected colors of
composite SEDs of local E and Im galaxies,
we adopt a local Sbc galaxy as a typical model
for estimating the mass-to-light ratio and the SED
of host galaxies of the target QSOs;
we use local E and Im galaxies for the comparison.
We use mass-to-light ratios of 20, 2, and 0.5
for E, Sbc, and Im, respectively
\citep{1976ApJ...204..668F, 2001ApJ...550..212B},
and calculate the fluxes of the host galaxies
in the rest-frame wavelength coverage of the observing filters
using an UV-optical spectrum template of
nearby galaxies for E, Sbc, and Im presented
in \citet{2008ApJ...676..286A}.
Figures \ref{f_ff1} and \ref{f_ff2} present
the location of the host galaxy flux of the target QSOs
in the flux-to-flux diagrams.
The estimated host galaxy fluxes, as described,
are listed in Table \ref{t_host}.
Figures \ref{f_ff1} and \ref{f_ff2} clearly show that
the location of the host galaxy flux is systematically
on the left side of the best-fit regression line.
For a typical case of the Sbc-type galaxy,
the host galaxy fluxes of 9 of the 10 targets
are different from the fainter
extension of the best-fit regression line
by more than $2\sigma $ error of the host galaxy fluxes,
or a factor of 5.8.
The only target for which the host galaxy flux is located
on the extension of the best-fit regression line
within $2\sigma $ error, J2115$+$0024, is in fact
$F_{\nu _{2}0}$ consistent to be zero,
and the host galaxy flux is estimated to be considerably small.
Instead, if we assume the Im-type galaxy for the host galaxy
to model recent star-forming activity,
as reported for high-redshift QSOs by Schramm et al. (2008),
the host-galaxy fluxes of 5 of the 10 targets
(J0105$-$0050, J0105$+$0040, J2045$-$0051, J2111$+$0024, and J2134$+$0048)
are located on the extension of the best-fit regression line
within $2\sigma $ error.
However, if the estimated black hole mass
using the Magorrian relation at the local universe
is systematically smaller by a factor of $\sim 5$
caused by the redshift evolution,
the location of the host galaxy flux
is on the left side of the best-fit regression line
for four of those five targets
(except the only target, again J2115$+$0024).
We conclude that most of the target QSOs show
spectral hardening in UV wavelengths, that is,
the UV continuum emission becomes bluer as it becomes brighter,
in spite of maintaining the linear correlations between the QSOs' two-band flux data.
These results suggest that the UV continuum emission of QSOs
usually shows spectral hardening, which is consistent with
the results reported by \citet{2004ApJ...601..692V}
and \citet{2005ApJ...633..638W}
from the colors of the differential fluxes
or the composite differential spectrum
of two-epoch observations for SDSS QSOs.
\subsection{Effect of Small Blue Bump}
\label{s_effectsbp}
The best-fit parameters and the reduced $\chi ^2$ values
of the straight-line fitting are listed in Table \ref{t_fffit}.
As described in Section \ref{s_ffplotfit},
the reduced $\chi ^2$ value is so large
that the linear relationship is rejected
at a 1\% level of significance for 4 of the 10 targets
(J0105$-$0050, J0105$+$0040, J0136$-$0046, and J2045$-$0051).
These larger $\chi ^2$ values suggest contamination from
another variable component of fluxes
that is not synchronized to the UV continuum emission.
The SBB consisting of \ion {Fe}{2} emission lines
and Balmer continuum emission,
and the \ion {Mg}{2} $\lambda 2798$ emission line,
which more or less contaminate our observing bands
of the longer-wavelengths,
are thought to be the most promising sources
for such additional variable components.
They are considered to originate in the broad emission-line region (BLR),
and
moreover, reverberation mapping observations
have found the lag of the SBB flux variation
behind that of the UV continuum emission
for NGC 5548 \citep{1993ApJ...404..576M}
and the lag of the \ion {Mg}{2} emission-line flux variation
for NGC 4151 \citep{metzroth2006}.
Since the size of the BLR of the target QSOs
is estimated to be more than 200 light days
for H$\beta $ emission line by the luminosity scaling relation
\citep{2005ApJ...629...61K},
the flux variation of the SBB emission and the \ion {Mg}{2} emission line
is not at all synchronized with that of the continuum emission
and would provide a scatter around the linear correlation
between the fluxes of the UV continuum emission in the two different bands.
We estimated the scatter of the linear correlation
in the flux-to-flux plots caused by
the contaminations from the SBB and \ion{Mg}{2} emissions
and re-examined the linear relationship of
the UV continuum emission in the two different bands.
We first fit the target spectrum obtained from the SDSS database
by a power-law model with the spectral windows
of $\lambda=1440-1470$\AA $\,, 1685-1695$\AA, $2200-2230$\AA, and $4000-4020\AA$,
where the spectrum is free from any strong emission lines.
The synthesized flux of the spectrum in the photometric band
was estimated by convolving the spectrum with the filter transmission curve,
and then, the flux of the spectrum was recalibrated by
scaling the synthesized flux by the photometric data
at the observing date of the spectroscopy
that was estimated by interpolating the light curve.
The contaminated flux of the SBB and \ion{Mg}{2} emissions
in the broad-band filter was estimated by subtracting the synthesized flux
of the best-fit power-law spectrum from that of the original spectrum.
The contaminated fluxes of the SBB and \ion{Mg}{2} emissions
in the longer wavelength filter are listed in Table \ref{t_sbb}.
Their uncertainties were derived from the error of the power-law fitting.
Next, the $1\sigma $ scatter of the flux variation
of the SBB and \ion{Mg}{2} emission-line component,
$\sigma _{SBB}$, was estimated as $f_{SBB}\times \sigma_{c}/\bar{f_{c}}$,
where $f_{SBB}$, $\bar{f_{c}}$, and $\sigma_c$ are
the contaminated flux of the SBB and \ion{Mg}{2} emissions,
the average flux, and the $1\sigma $ scatter
over the broad-band light curve, respectively.
The $\sigma _{SBB}$ is also listed in Table \ref{t_sbb}.
Finally, assuming that the flux variation of
the SBB and \ion{Mg}{2} emission line
is totally uncorrelated with that of the continuum emission in time,
we recalculated the reduced $\chi^2$ of
the straight-line fitting for the flux-to-flux plots
by replacing $\sigma _{obs}$ by $\sqrt{\sigma_{obs}^2+\sigma_{SBB}^2}$,
where $\sigma _{obs}$ is the observational error of the photometry.
The new reduced $\chi^2$ values of the straight-line fitting
in which the scatter caused by the contaminations of
the SBB and \ion{Mg}{2} emissions
is considered, are listed in Table \ref{t_sbb}.
We found that
the reduced $\chi^2$ values of the straight-line fitting
were significantly decreased for three targets
(J0105$-$0050, J0105$+$0040, and J2045$-$0051)
out of the four targets,
for which the linear relationship was rejected
by the $\chi ^2$ analysis in Section \ref {s_ffplotfit},
and the linear relationship is now not rejected
at a 1\% level of significance
for two targets (J0105$-$0050 and J0105$+$0040).
These results strongly indicate that
a part of the scatter of the flux-to-flux plot from
the best-fit regression line is caused by
the contaminations from the SBB and \ion {Mg}{2} emission line,
and the UV continuum emissions in two different bands
of the target QSOs show a tight linear correlation
during the flux variations.
\section{Discussion}
\label{s_discuss}
\subsection{Comparison with Previous Studies}
\label{s_compprevstudy}
\subsubsection{Vanden Berk et al. (2004)}
\label{s_vandenberk}
\citet{2004ApJ...601..692V}
examined the relationship between variability amplitude
and the rest-frame wavelength of the UV-optical region
from SDSS two-epoch multi-band observations
of about 25,000 QSOs,
and found a larger amplitude of variation
at shorter wavelengths in the UV region of $\lambda < 4000$ \AA.
In this section, the tracks of the flux-to-flux plot of the target QSOs
obtained from the photometric monitoring data,
which show the hardening trend in the UV region, as presented in the previous sections,
are compared with the hardening trend of the amplitude of variation
presented in \citet{2004ApJ...601..692V}.
Assuming that the wavelength-dependent amplitude of variation
always holds for the flux variation at all times for individual QSOs,
a power-law function of $F_{\nu {1}}=\alpha \times F_{\nu {2}}^\beta $ in a flux-to-flux plot
is obtained, where $\beta =v(\lambda_1)/v(\lambda_2) $ is
derived from the wavelength-dependent amplitude of variation $v(\lambda )$,
presented in Equation 11 of \citet{2004ApJ...601..692V}.
Then, we fit the power-law function,
$F_{\nu {1}}=\alpha_{2} \times F_{\nu _{2}}^{\beta_{2}}$,
to the flux-to-flux plot data of the target QSOs,
fixing the parameter $\beta _{2}$ as $\beta_{2}=v(\lambda_1)/v(\lambda_2)$,
where $\lambda_1$ and $\lambda_2 (\lambda_1 <\lambda_2)$ are
the effective wavelengths of the observing filters in the rest frame.
Figures \ref{f_ff1} and \ref{f_ff2} present the fitting of the power-law
function to the flux-to-flux plots, and the best-fit parameters
and the reduced $\chi ^2$ values are listed in Table \ref{t_vanden}.
The reduced $\chi^2$ value of the power-law fit is
comparable to that of the straight-line fit for most of the target QSOs.
The slope of the power-law fit seems to be
slightly gentler than that of the flux-to-flux plots for J2123$-$0050,
and in fact, the reduced $\chi ^2$ of the power-law fit
is slightly larger than that of the straight-line fit
even if it is still as small as $\sim 1$.
When we refit the power-law function,
$F_{\nu {1}}=\alpha_{2} \times F_{\nu _{2}}^{\beta_{2}}$,
with both $\alpha _{2}$ and $\beta _{2}$ freed,
$\beta _{2}=1.908\pm 0.152$ is derived from the new power-law fit
with the reduced $\chi ^2=0.695$.
The $\beta _{2}$ value is slightly different from $v(\lambda_1)/v(\lambda_2) $
but would be still consistent with \citet{2004ApJ...601..692V}
within the fitting error and the diversity of QSO properties,
and the reduced $\chi ^2$ becomes comparable to that of the straight-line fit.
From these results, we conclude that
our results of the hardening trend during the flux variation
obtained from the flux-to-flux plots in the rest-frame UV wavelengths
are consistent with the hardening trend of
the variability amplitude in the UV region presented in \citet{2004ApJ...601..692V}.
\subsubsection{\citet{1996A&A...312...55P}}
\label{s_paltaniwalter}
\citet{1996A&A...312...55P} observed the UV spectra of 15 nearby AGNs
at different epochs using the IUE satellite.
Applying PCA to the multi-epoch spectra and the decomposition
to a variable and a non-variable components,
they explained the UV variation in almost all their targets
as the sum of a variable component of a constant power-law spectral shape
and a non-variable component of the SBB,
which means that the spectral shape of the UV continuum emission
is almost constant during the flux variation.
This finding is not consistent with our results
that the UV continuum emission becomes bluer
as it becomes brighter for most of the target QSOs.
In fact, \citet{1996A&A...312...55P} did not
estimate the SBB fluxes of the target AGNs
by fitting a spectral model of a power-law continuum plux the SBB
to their UV spectrum in order to examine the consistency
of that method with the conclusion from the decomposition to
a variable and a non-variable components.
Moreover, as discussed in Section \ref{s_effectsbp},
the SBB is thought to originate in the BLR
and its flux was found to be variable.
It is suspected that a sufficient amount of the SBB flux is
constant, which explains the constant spectral shape
of the UV continuum emission during flux variation.
On the other hand, Paltani \& Walter's finding that
one eigenvalue dominated all the others in the PCA of the UV flux variation
is consistent with the linear relationship of the fluxes
in the two different bands found in the flux-to-flux plots of our target QSOs.
The models of the accretion disk predict that
the spectral shape of the UV-optical continuum emission
depends on the mass accretion rate and the black hole mass;
thus, the spectral properties of flux variation
would also depend on those AGN properties.
Therefore, while it is beyond the scope of this paper,
it is very interesting to re-examine
the multi-epoch UV spectra of nearby AGNs of \citet{1996A&A...312...55P}
using the flux-to-flux plot analysis with accurate estimation of
non-variable components such as the host galaxy
and also to discuss the luminosity or the black hole mass
dependency of the spectral variability of the UV continuum emission.
\subsection{Comparison with Accretion Disk Model}
\label{s_accratemodel}
We showed that the spectral shape of the optical continuum emission
is almost constant during the flux variation for 11 Seyfert galaxies in Paper I,
which suggests that the radial temperature profile of
an optical emitting region of an accretion disk
does not change during the flux variation.
On the other hand,
we here find that the UV continuum emission becomes bluer as it brightens
for most of the target QSOs, while the UV fluxes in the two different bands
show a tight linear correlation.
This suggests that
the temperature structure of the UV emitting region of an accretion disk
shows systematic change.
When the mass accretion rate changes in a
standard accretion disk model \citep{1973A&A....24..337S},
the spectral shape of the continuum emission from the accretion disk
changes in UV wavelengths
while it remains almost constant in optical wavelengths for AGNs
whose central black hole mass is larger than $\sim 10^{7-8} M_{\odot}$
(e.g., Kawaguchi, Shimura, \& Mineshige 2001).
That is because the variation of the mass accretion rate
changes the maximum temperature of the accretion disk,
which affects the spectral shape of UV continuum emission,
while it does not change the radial profile of outer accretion disk
of $T(r)\propto r^{-3/4}$ where the optical continuum emission originates,
which would result in an almost constant optical continuum spectral shape.
\citet{2006ApJ...642...87P} presented that
a standard accretion disk model in which
the mass accretion rate changed from one epoch to the next
could be successfully fitted to the composite differential spectrum
of two epochs of observations for hundreds of SDSS QSOs,
which is bluer than the composite spectrum of QSOs
indicating the spectral hardening of UV continuum emission
\citep{2005ApJ...633..638W}.
However, the composite spectrum represents
the average characteristics of QSOs;
moreover, the differential spectrum corresponds
only to the slope of the linear relationship between fluxes
in two different bands in the flux-to-flux plots.
Here, we fit a standard accretion disk model
with a varying mass accretion rate
to the flux-to-flux plots for individual target QSOs,
and examine whether the spectral variation of the UV continuum emission
during flux variations at various epochs
can be described by the standard accretion disk model
with various mass accretion rates and a constant black hole mass.
The inner radius of the standard accretion disk model
is set as $R_{in}=3Rs$, where $Rs$ is the Schwarzschild radius.
The outer radius is set as $R_{out}=1000Rs$,
but it is not important for the UV spectrum.
A face-on view is assumed to calculate the flux from the accretion disk.
Then, a trajectory of the accretion disk model in the flux-to-flux diagram
can be traced by changing the mass accretion rate
with a constant black hole mass.
We fit the trajectory of the accretion disk model
to the flux-to-flux plots of the individual target QSOs
with a free parameter of black hole mass.
Since the observed flux includes the contaminating fluxes of
the host galaxy, the SBB and \ion{Mg}{2} emission line
in addition to the continuum emission,
the Sbc-type host galaxy flux estimated in Section \ref{s_effectotagncont},
and the SBB and \ion{Mg}{2} contaminating flux
estimated in Section \ref{s_effectsbp}
are added to the trajectory of the accretion disk model
before fitting.
Figures \ref{f_ffsim1} and \ref{f_ffsim2} present
the best-fit accretion disk model for the flux-to-flux plots
of the target QSOs, and the best-fit black hole mass and
the reduced $\chi ^2$ are listed in Table \ref{t_simstat}.
The error of the black hole mass in the table is estimated from
the statistical uncertainty of $\chi^2$-fitting of the model.
As shown in Figures \ref{f_ffsim1} and \ref{f_ffsim2},
the flux-to-flux plots can be fitted well
by the standard accretion disk model, changing the mass accretion rate
with a constant black hole mass.
Indeed, the reduced $\chi^2$ values of the standard accretion disk model
are comparable to those of the straight-line fitting
for 9 of the 10 target QSOs, and even for the exception,
J2123$-$0050, the reduced $\chi ^2$ is as small as $\sim 1$.
In addition, as shown in Figure \ref{f_bhmass},
the best-fit black hole mass of the standard accretion disk model
is consistent with the black hole mass estimated from the emission-line width
and the luminosity scaling relation in Section \ref{s_targetselec},
within 0.76 dex for $2\sigma $ error of
the latter black hole mass estimation.
We also calculate the bolometric luminosity of
the best-fit standard accretion disk model whose mass accretion rate
corresponds to the average flux of the light curve.
The ratio of the bolometric luminosity
to the specific luminosity in UV,
the bolometric correction $L_{\rm bol}/\lambda L_{\lambda }(3000{\rm \AA})$,
is listed in Table \ref{t_simstat}.
We find that the bolometric correction
is consistent with those obtained for
hundreds of type 1 QSOs from \citet{2006ApJS..166..470R},
which are distributed between 4 to 9 and whose average value is 5.62.
The black hole mass and the luminosity of the best-fit
standard accretion disk model fitted to the flux-to-flux plots
in UV wavebands are reasonable, and it could be a new technique for
estimating a black hole mass of luminous QSOs although the systematic
uncertainties should be examined.
As presented in Paper I, the optical color of
the best-fit regression line in the flux-to-flux diagram
can be regarded as that of the AGN optical continuum,
because the optical spectral shape remains almost constant
during the flux variation and the colors derived from
the flux-to-flux plot analysis of nearby Seyfert galaxies
were consistent with those of the standard accretion disk model,
$\alpha _{\nu}=1/3$.
Although it has long been known that the standard accretion disk model
apparently shows a contradiction with observations
that the composite spectrum of QSOs is redder than that of
the standard accretion disk model
\citep{1991ApJ...373..465F,2001AJ....122..549V},
the detailed analyses of
the spectral shape of the UV-optical continuum emission of AGNs
using flux variation presented here and in Paper I
strongly support the standard accretion disk.
A possible explanation for the composite spectra being redder is
that host galaxy light contaminates the spectra preferentially
at longer wavelengths.
These results are consistent with the recent analyses
of the spectral shape of the optical to near-infrared continuum emission,
which comes from the accretion disk
by flux variations \citep{2006ApJ...652L..13T}
and polarizations \citep{2008Natur.454..492K}.
However, it has a critical difficulty
of considering the variation
of global accretion rate
of a standard accretion disk
as a primary source of flux variation of QSOs.
The timescale of changes of the global mass accretion rate
is believed to correspond to the viscous timescale
(e.g., Pringle 1981; Frank et al. 2002),
which is about $\sim 10^6$ years for our target QSOs
and much longer than the timescale of the flux variation we observed.
Clearly,
further study of the flux variation mechanism is desired,
which can explain the observed properties
of spectral variations
of the UV-optical continuum emission of AGNs,
suggesting a variation of the characteristic
temperature of an accretion disk.
\section{Summary}
\label{s_summary}
We examined the spectral variability of UV continuum emission of AGNs
based on the multi-epoch photometric data of 10 SDSS QSOs
in the Stripe 82.
The target redshift was $z=1.05,1.54,1.71, 2.35\pm0.05$
in which strong broad lines such as Ly$\alpha$ and \ion{C}{4}
do not contaminate SDSS filters.
The target luminosity was about $\lambda L_{\lambda }(3000{\rm \AA})\sim 10^{46-47}$ erg s$^{-1}$,
the black hole mass was estimated to be about $10^{9-10}M_{\odot}$
based on the luminosity scaling relations
for \ion{C}{4} or \ion{Mg}{2} emission lines and their line width,
and the mean Eddington ratio was between $0.1-1.2$.
All target QSOs showed significant flux variation
within 7 years of observation, or $2-3.5$ years in the rest-frame.
We plotted the flux data in two different bands
(shorter-$\lambda_{rest}\sim 1400$ \AA \, or $\sim 1730$ \AA,
longer-$\lambda_{rest}\sim2200$ \AA $- 3600$ \AA)
observed on the same night in the flux-to-flux diagram
and found a strong linear correlation
between the fluxes in the two different bands
during the flux variations for all target QSOs.
The linear relationship was not rejected for 6 of the 10 targets
at a 1$\%$ level of significance based on the $\chi^2$-test
for straight-line fitting.
Even for the remaining four targets,
no clear curvature could be seen in their flux-to-flux plots;
no significant improvement of $\chi^2$ values by the power-law fitting
than those of the straight-line fitting could be found
for two of the four targets,
and the reduced $\chi^2$ values of the straight-line fitting
were significantly decreased for three of the four targets
when the contamination of the SBB was considered.
We then examined the location of the host galaxy flux in the flux-to-flux diagram.
The stellar mass of the host galaxy
was estimated from the black hole mass
on the basis of the Magorrian relation \citep{2003ApJ...589L..21M};
then, the host galaxy flux was estimated from the stellar mass
using the mass-to-light ratio and the SED of a local Sbc galaxy
\citep{2001ApJ...550..212B,2008ApJ...676..286A}.
We found that the location of the host galaxy flux was systematically on the left side
of the best-fit regression line for 9 of the 10 targets;
that is,
the UV continuum emission becomes bluer as it brightens
for most of the target QSOs
in spite of holding their linear correlations between the two-band flux data.
This trend of spectral hardening in UV wavelengths
is consistent with the findings of \citet{2004ApJ...601..692V}.
Indeed, we found that the flux-to-flux plots of the target QSOs
could be fitted well by a power-law function assuming
the wavelength-dependent variability amplitude of QSOs.
We presented in Paper I that
the spectral shape of the optical continuum emission of AGNs
remains almost constant and that its color is consistent with that of
the standard accretion disk model, $\alpha _{\nu }=1/3$.
Then, in the present paper, we fit a standard accretion disk model
with a changing mass accretion rate and a constant black hole mass
to the UV flux-to-flux plots for the individual target QSOs,
and found that the spectral variation of the UV continuum emission
during flux variations at various epochs
could be described by the standard accretion disk model
with varied mass accretion rates.
The black hole mass and the bolometric luminosity
from the best-fit standard accretion disk model
were reasonable.
Although it has long been known that
the composite spectrum of QSOs is redder than that of
the standard accretion disk model,
the detailed analyses of
the spectral shape of the UV-optical continuum emission of AGNs
using flux variation strongly support the standard accretion disk model.
However,
the variation timescale of the global mass accretion rate
was considered to be too large
to explain the flux variations within a few years,
and further study of the flux variation mechanism is desired.
\acknowledgments
We thank Toshihiro Kawaguchi for useful discussion and comments.
This research has been partly supported by the Grant-in-Aids of Scientific Research (10041110,
10304014, 12640233, 14047206, 14253001, and 14540223) and COE Research (07CE2002)
of the Ministry of Education, Science, Culture and Sports of Japan.
Tomoki Morokuma has been supported by the JSPS (Japan Society for the Promotion of Science)
Research Fellowship for Young Scientists.
\clearpage
|
\section{Introduction}
The radiative heat transfer between macroscopic dielectric bodies is a well-understood
physical phenomenom and the energy transported per unit time and unit area can be
expressed by the Kirchhoff-Planck law~\cite{L.D.Landau_E.M.Lifschitz_1979} in the stationary case, i.e.,
when the bodies are in local thermal equilibrium. But this is only true for distances $d$
which are large compared with the thermal wave length $\lambda_{{\rm th}}$ given
by Wien's law, i.e., for $d \gg \lambda_{{\rm th}}$.
For distances $d \ll \lambda_{{\rm th}}$ the radiative heat transfer increases
strongly due to so-called evanescent modes.
For convenience we consider two semi-infinite bodies with a vacuum gap of distance $d$
between them (see Fig.~\ref{Fig:SemiInfiniteBodies}). This system is invariant under rotations and we chose the $z$-axis as
the axis of rotation. We further assume that the first body at $z < 0$ has a temperature
$T_1 \ne 0$, whereas the second semi-infinite body at $z > d$ has zero temperature, i.e.,
$T_2 = 0$. Due to quantum and vacuum fluctuations of the atoms and electrons in the
first body there will be a fluctuating electromagnetical field inside that body.
Now, the boundary conditions of macroscopic electrodynamics require that the transversal
components of the electric and the magnetic field are continuous at the interface $z = 0$.
Therefore, there will also be a fluctuating electromagnetic field outside the semi-infinite
body produced by fluctuating internal sources. This field can be divided into a
propagating and an evanescent part. This becomes clear if we consider a plane wave solution
$\exp({\rm i} (k_z z - \omega t))$ in the vacuum gap propagating into the positive $z$-direction,
with the modulus of the wave vector given by
\begin{equation}
k_z = \sqrt{\frac{\omega^2}{c^2} - k_\perp^2} =
\begin{cases}
\text{real} &, k_\perp < \frac{\omega}{c}, \\
\text{complex} &, k_\perp >\frac{\omega}{c}.
\end{cases}
\end{equation}
Obviously, the modulus of the wave vector $k_z$ is purely real if the modulus of
the transversal wave vector $k_\perp$ is smaller than the modulus of the vacuum wave
vector, $k_0 = \omega c^{-1}$. In this case we will have an oscillatory solution, i.e.,
propagating waves. On the other hand, $k_z$ is purely imaginary if $k_\perp > k_0$.
In that case the plane wave is exponentially damped in $z$-direction, i.e., the solution
is a so-called evanescent wave.
\begin{figure}
\centering
\includegraphics[angle=0.0, height = 4.3cm]{halfspace2.eps}
\caption{Sketch of the propagating and evanescent modes contributing to
the radiative heat transfer between two semi-infinite bodies at different
temperatures.}
\label{Fig:SemiInfiniteBodies}
\end{figure}
The characteristic wave length of the fluctuating field in the vacuum gap, generated in the semi-infinite
body at $z < 0$, is the thermal wave length $\lambda_{{\rm th}}$. It follows
that for distances $d \gg \lambda$, i.e., in the far-field region, the evanescent field
can not reach into the second body and therefore only propagating modes can contribute
to the radiative energy transfer. This energy transfer --- as mentioned above --- can then be
described with the Kirchhoff-Planck law and is smaller than the energy transfer
between two black bodies, which is given by the Stefan-Boltzmann law~\cite{L.D.Landau_E.M.Lifschitz_1979}
\begin{equation}
S_{{\rm BB}} = \sigma (T_1^4 - T_2^4)
\label{Eq:StefanBoltzmann}
\end{equation}
with $\sigma = 5.67\cdot10^{-8} {\rm W}{\rm m}^{-2}{\rm K}^{-4}$. On the other hand, for distances $d$ between
the two bodies smaller than $\lambda_{{\rm th}}$, i.e., in the near-field region,
the evanescent waves generated in the first body can reach into the second body and propagate there.
This can be seen as some kind of photon tunneling through a vacuum barrier (see Fig.\ \ref{Fig:SemiInfiniteBodies}).
Hence, in the near-field
the propagating and evanescent fluctuating fields both will contribute to the radiative energy transfer. In that
case, the energy transfer can be several orders of magnitude larger than the black body value
(\ref{Eq:StefanBoltzmann}), as we will see later.
\section{Theoretical description}
\vspace{-0.3cm}
\subsection{Fluctuating fields}
The theoretical treatment of radiative heat transfer is usually based on Rytov's fluctuating
electrodynamics~\cite{S.M.Rytov_et_al_1989}. Within this framework the macroscopic Maxwell equations
are augmented by fluctuating source currents $\mathbf{j}$, describing the fluctuating
sources of the electric and magnetic field $\mathbf{E}$ and $\mathbf{H}$ inside
the dielectric dissipative body. The individual
frequency components of these source currents $\mathbf{j}(\mathbf{r},\omega)$ are
considered as stochastic Gaussian variables. Within this treatment the Maxwell
equations become so-called stochastic Maxwell equations
\begin{align}
\nabla\times\mathbf{E}(\mathbf{r},\omega) &= {\rm i} \omega \mu_0 \mathbf{H}(\mathbf{r},\omega), \nonumber \\
\nabla\times\mathbf{H}(\mathbf{r},\omega) &= \mathbf{j}(\mathbf{r},\omega)
- {\rm i}\omega \epsilon(\omega) \mathbf{E}(\mathbf{r},\omega)
\label{Eq:StochasticMaxwell}
\end{align}
with the permeability of the vacuum $\mu_0$ and the permittivity of the body $\epsilon(\omega)$.
By writing the stochastic Maxwell equations in the form of eq.\ (\ref{Eq:StochasticMaxwell}) the dielectric body
containing the source currents is already assumed to be non-magnetic, homogeneous, isotropic and local.
The electric and magnetic fields described by the stochastic Maxwell equations are classical
stochastic processes and can formally be expressed as
\begin{align}
\mathbf{E}(\mathbf{r},\omega) &= {\rm i} \omega \mu_0 \int\!\!{\rm d}^3 r'\,\mathds{G}^E(\mathbf{r,r'}) \mathbf{j}(\mathbf{r'},\omega), \nonumber \\
\mathbf{H}(\mathbf{r},\omega) &= {\rm i} \omega \mu_0 \int\!\!{\rm d}^3 r'\,\mathds{G}^H(\mathbf{r,r'}) \mathbf{j}(\mathbf{r'},\omega)
\label{Eq:StochasticFields}
\end{align}
where the integral has to be taken over the source region, i.e., over the volume of the dielectric body
containing the source currents. The tensors
$\mathds{G}^E$ and $\mathds{G}^H$ are the dyadic electric and the magnetic Green's function. The
electric dyadic Green's function is a solution of the Helmholtz wave equation
\begin{equation}
\bigl(\nabla\times\nabla\times - k^2 \bigr) \mathds{G}^{E}(\mathbf{r,r'})
= \mathds{1} \delta(\mathbf{r - r'})
\end{equation}
with the wave vector $\mathbf{k}$ and the unit matrix $\mathds{1}$. By definition, the electric dyadic Green's
function satisfies the boundary conditions of the electric field. Even though a similiar statement holds
for the magnetic Green's function, it is more convenient to construct the dyadic magnetic Green's from
the electric Green's function by means of the relation
\begin{equation}
\mathds{G}^H (\mathbf{r,r'}) = - \frac{{\rm i}}{\omega \mu_0} \nabla \times \mathds{G}^{E}(\mathbf{r,r'}),
\end{equation}
which is a direct consequence of Maxwell's equations.
The reformulation of the electric and magnetic field $\mathbf{E}$ and $\mathbf{H}$ as an integral
over the source currents in eqs.\ (\ref{Eq:StochasticFields}) makes clear that the fluctuating properties
of the fields are directly determined by the fluctuational properties of the source currents.
In fact, if $\langle \mathbf{j} (\mathbf{r},\omega) \rangle$ and
$\langle \mathbf{j}(\mathbf{r},\omega) \mathbf{j}^*(\mathbf{r'},\omega') \rangle$
are known, where the angular brackets symbolise the ensemble average, the corresponding averages and
correlation functions of the fields can be evaluated. Therefore, it remains to determine the
stochastic properties of the source currents.
From the definition of the source currents as stochastic Gaussian variables describing the thermal and
quantum fluctuations inside a dielectric dissipative body, it should be clear that the ensemble average
of the source currents vanishes. It follows with (\ref{Eq:StochasticFields}) that
\begin{equation}
\langle \mathbf{E}(\mathbf{r},\omega) \rangle = \mathbf{0} \quad \text{and} \quad
\langle \mathbf{H} (\mathbf{r},\omega)\rangle = \mathbf{0}.
\end{equation}
However, the correlations and therewith the fluctuations of the source currents are determined by
the properties of the dielectric dissipative body. This connection between dissipation and fluctuation
can be expressed in a general way by means of the fluctuation-dissipation theorem~\cite{L.D.Landau_E.M.Lifschitz_1979},
which can be applied to our problem~\cite{S.M.Rytov_et_al_1989}. For a homogeneous, isotropic and local
dielectric dissipative body this yields
\begin{equation}
\begin{split}
\langle j_\alpha(\mathbf{r},\omega) j_\beta^*(\mathbf{r'},\omega') \rangle & \\
& \hspace{-2cm} = 4 \pi \omega E(\omega, T) \epsilon''(\omega) \delta_{\alpha\beta}
\delta(\mathbf{r-r'}) \delta(\omega - \omega')
\end{split}
\label{Eq:FDT}
\end{equation}
for the components of the source currents, where we have written $\epsilon = \epsilon' + {\rm i} \epsilon''$, and
$E(\omega, T)$ is the Einstein function
\begin{equation}
E(\omega, T) = \frac{\hbar \omega}{2} + \frac{\hbar \omega}{{\rm e}^{\hbar \omega \beta} - 1}
\end{equation}
with the usual inverse temperature $\beta = 1/k_{\rm B} T$ of the body. The appearance of the
Einstein function is implied by the fluctuation-dissipation theorem, which indeed is
a quantum mechanical operator relation, but the fluctuating fields in the framework of fluctuating
electrodynamics are classical. This is a somewhat weak point in the theory, but it was recently
shown~\cite{M.Janowicz_et_al_2003} that a rigorous quantum electrodynamical treatment for dissipative media
leads to the same correlation functions for the fields as fluctuating electrodynamics. Hence,
it seems to be justified to work within the framework of fluctuating electrodynamics, which should be
classified as a semi-classical theory.
Now, it is possible to evaluate the correlation functions of the fluctuating fields. Using eqs.\
(\ref{Eq:StochasticFields}) together with the fluctuation-dissipation theorem in eq.\ (\ref{Eq:FDT})
yields
\begin{equation}
\begin{split}
\langle E_\alpha (\mathbf{r},t) H_\beta(\mathbf{r},t) \rangle
&= \int_0^\infty\!\!\!{\rm d}\, \omega\, E(\omega, T) \frac{\omega \epsilon''(\omega)}{\pi} (\mu_0^2 \omega^2) \\
&\quad\times\int\!\!{\rm d}^3 r'\, \biggl( \mathds{G}^E {\mathds{G}^H}^\dagger \biggr)_{\alpha\beta} + {\rm c.c.}
\end{split}
\label{Eq:CorrelationEH}
\end{equation}
where the $\dagger$-symbol denotes the adjoint dyadic and c.c. is an abbreviation for the complex conjugated term.
Similiar equations can be derived for $\langle E_\alpha (\mathbf{r},t) E_\beta(\mathbf{r},t) \rangle$ and
$\langle H_\alpha (\mathbf{r},t) H_\beta(\mathbf{r},t) \rangle$, respectively.
The ensemble averages of the physical quantities --- the Poynting vector, the energy density and the stress tensor ---
of a dielectric dissipative body can be evaluated with the help of these field correlation functions, if the
material properties, i.e., the permittivity $\epsilon(\omega)$, and the temperature of the body are known. Furthermore,
it is necessary to calculate the dyadic Green's functions for the given electrodynamical problem, in which
the geometry of the problem is incorporated.
\vspace{-0.3cm}
\subsection{Radiative heat transfer between two slabs}
In principle, we are now able to calculate the radiative heat transfer between dielectric bodies of
arbitrary shape and volume as long these bodies are homogeneous, isotropic, and local dissipative dielectrics,
and can be described macroscopically. Unfortunately, the calculation of the dyadic Green's function
for the most geometries is rather complicated. Therefore, we will consider only the radiative heat transfer
between two bodies of the simplest possible geometry, i.e., two semi-infinite dielectric bodies separated
by a vacuum gap of width $d$ (see Fig.\ \ref{Fig:SemiInfiniteBodies}).
The solution to this problem is well-known: the first derivation in the framework of Rytov's fluctuating
electrodynamics was given by Polder and van Hove~\cite{D.Polder_M.VanHove_1971}. Levin {\itshape et al.}~\cite{M.L.Levin_et_al_1980}
solved the problem with the surface impedance method of Leontovich, and Loomis and Maris~\cite{J.J._Loomis_H.J._Maris_1994}
reinvestigated the problem some years ago. The radiative heat transfer between the two slabs can
be stated as~\cite{D.Polder_M.VanHove_1971}
\begin{equation}
\begin{split}
\langle S \rangle &= \frac{1}{4 \pi^2} \int_0^\infty\!\!\!{\rm d} \omega\, \bigl[ E(\omega,T_1) - E(\omega,T_2)\bigr] \\
&\quad\times\int_0^\infty \!\!\! {\rm d} k_\perp\, k_\perp \bigl\{ T_\parallel^{12} + T_\perp^{12} \bigr\}
\end{split}
\label{Eq:HeatTransfer}
\end{equation}
where $T^{12}_\parallel$ and $T_\perp^{12}$ are the
transmission coefficients from the first body at $z < 0$ to the second body at $z > d$ for the TM- and TE-modes,
respectively. These transmission coefficients can be expressed by means of the usual Fresnel
coefficients~\cite{L.D.Landau_E.M.Lifschitz_1974}, which gives for the propagating modes, i.e., for $k_\perp < k_0$,
\begin{equation}
(T_\parallel^{12})^{\rm{prop}} = \frac{(1 - |r^{1}_\parallel|^2)(1 - |r^{2}_\parallel|^2)}{|1 - r^{1}_\parallel r^{2}_\parallel \rm{e}^{2 \rm{i} k_{z} d}|^2},
\label{Eq:TransPr}
\end{equation}
and for the evanescent modes, i.e., for $k_\perp > k_0$,
\begin{equation}
(T_\parallel^{12})^{\rm{ev}} = \frac{4 \Im(r^{1}_\parallel) \Im(r^{2}_\parallel) \rm{e}^{-2 \gamma d}}{|1 - r^{1}_\parallel r^{2}_\parallel \rm{e}^{- 2 \gamma d}|^2},
\label{Eq:TransEv}
\end{equation}
where the transmission coefficents for the TE-modes can be deduced from these relations by the
substitution $\parallel \rightarrow \perp$.
The mean energy transfer per unit time and unit area between the two semi-infinite bodies at
different temperatures in eq.\ (\ref{Eq:HeatTransfer}) consists of two parts.
The temperature enters only into the first part,
which is given by the difference of the Einstein functions evaluated at the coresponding
temperatures. From this difference it becomes clear that the vacuum term in the Einstein functions
vanishes; there will be no vacuum contribution to the energy transfer. The second part is the integral
over the lateral wave vector $k_\perp$ which counts the modes contributing to the energy transfer.
It can be shown that the above transmission coefficients are always smaller than $1$. Moreover, it
is possible to retrieve the energy transfer between two black bodies, which by definition have zero Fresnel
coefficients. It follows that the transmission coefficient for the evanescent modes (\ref{Eq:TransEv}) becomes zero and
the transmission coefficient for the propagation modes (\ref{Eq:TransPr}) becomes~$1$. Inserting these transmission
coefficients in eq.\ (\ref{Eq:HeatTransfer}) yields the Stefan-Bolzmann law stated in eq.\ (\ref{Eq:StefanBoltzmann}).
From eq.\ (\ref{Eq:HeatTransfer}) together with the transmission coefficents it is possible to
study the radiative heat transfer between the two semi-infinite bodies in detail. At the moment,
we are only interested in the near-field limit of eq.\ (\ref{Eq:HeatTransfer}), i.e., in distances $k_0 d \ll 1$.
A Taylor expansion~\cite{M.L.Levin_et_al_1980} shows that in the near-field
\begin{equation}
\langle S_\parallel \rangle \propto \frac{1}{d^2} \quad\text{and}\quad \langle S_\perp \rangle = {\rm const}.
\label{Eq:TaylorExpansion}
\end{equation}
Therefore, the TM-mode part of the energy transfer dominates in the near-field region,
$\langle S_\parallel \rangle \gg \langle S_\perp \rangle$, but it is not clear a priori at what distance
this domination begins. In fact, we will see by numerical calculations that the distance where the
TM-mode contribution becomes dominant depends on the material properties of the two slabs. Furthermore,
from the near-field limit in eq.\ (\ref{Eq:TaylorExpansion}) two questions arise: i) Is a divergent radiative
energy transfer physically reasonable? ii) Is it possible to measure this power law? Or better, at what
distance does the domination of the TM-mode part begin?
The first question was discussed controversially~\cite{J.L.Pan_2000,J.P.Mulet_et_al_2001,J.L.Pan_2001}.
But it should be clear that the description based on fluctuating electrodynamics
is still a macroscopic one, which means that at distances smaller
than $10^{-8}{\rm m}$ this theory is not valid and therefore does not lead to physically reasonable results.
In fact, the source currents in the fluctuation-dissipation theorem in eq.\ (\ref{Eq:FDT}) are delta-correlated
with respect to $\mathbf{r - r'}$, but this can not be true at distances where the microscopic properties of
the materials make themselves felt. This delta-correlation of source currents seems to be the
source of the divergent heat transfer at small distances, but in order to give a satisfactory answer to the
question stated above a theory is needed that takes the finite correlations of source currents into account.
\begin{figure}
\centering
\includegraphics[angle=0.0, height = 5.5cm]{abcd_300_0.eps}
\caption{Numerical results of eq.\ (\ref{Eq:HeatTransfer}) for different Drude materials
at different temperatures normalised to the black body value $S_{{\rm BB}}$ given
by the Stefan-Boltzmann law in eq.\ (\ref{Eq:StefanBoltzmann}).}
\label{Fig:NumericalResults}
\end{figure}
With this drawback of the macroscopic theory and the second question in mind we will now discuss some numerical
results of the heat transfer (\ref{Eq:HeatTransfer}) between two Drude materials~\cite{N.W.Ashcroft_N.D.Mermin_2001}.
For convenience we set the temperature of the first body to $T_1 = 300 {\rm K}$ and the temperature of the
second body to $T_2 = 0 {\rm K}$. Moreover, we use the same Drude permittivity for both bodies, i.e.,
$\epsilon_1 = \epsilon_2$.
The numerical results for different plasma frequencies $\omega_p$ and relaxation times $\tau$
are given in Fig.\ \ref{Fig:NumericalResults}. For $a$, $b$, and $d$ we used relatively high relaxation
times and high plasma frequencies, so that $a$, $b$, and $d$ are plots for good conductors. On the
other hand, for $c$ we used a relatively small relaxation time and small plasma frequency, which means that
$c$ is a plot for a bad conductor.
Obviously the divergency discussed before, which is associated with the
TM-mode part of the radiative heat transfer $\langle S_\parallel \rangle$, can only be seen for $c$, i.e.,
for the bad conductor. There also is a divergency for good conductors, but it appears only for distances much smaller
than $10^{-9} {\rm m}$. Otherwise, in the region between $10^{-9} {\rm m}$ and $10^{-7} {\rm m}$
the radiative heat transfer for the good conductors $a$, $b$, and $c$ becomes constant. As discussed
before, such a behaviour can be attributed to the TE-mode part of the radiative heat transfer, $\langle S_\perp \rangle$.
As an answer to the second question, it follows that for bad conductors the radiative heat transfer in
the region between $10^{-9} {\rm m}$ and $10^{-7} {\rm m}$, which is accessible to measurements, is dominated by
the TM-mode part of the radiative heat transfer, whereas for good conductors the TE-mode
part dominates.
\section{Summary and outlook}
The radiative heat transfer between macroscopic dielectric bodies can be described
by Rytov's theory of fluctuating electrodynamics. In the near-field the evanescent
modes give the main contribution to the radiative heat transfer. As the numerical
results indicate, for good conductors the TE-mode part dominates the heat transfer
for experimentally accessible distances. In contrast, for bad conductors the
TM-mode part dominates in that region and leads to a divergent radiative heat transfer.
This unsatisfactory feature of the theory can be traced back to the delta correlation in the correlation
function of the source currents. There is a need for a theory which takes a finite
correlation length of the source currents into account, and which should lead to a finite heat
transfer at all distances.
Experiments now being prepared in Oldenburg should answer some questions and thus
provide a basis for such a theory. Some of these questions are: At what distance does
the macroscopic theory fail? What is the finite value of $\langle S \rangle$ for
$d \rightarrow 0$? Is there an appropriate theory describing the sensor-sample geometry?
|
\section{Introduction}
\label{sec:introduction}
``Lately, there has been a lot of fuss about sparse approximation.''
is the beginning of the paper~\cite{tropp2006relax} from 2006 and this
note could have started with the same sentence. Three different
minimization problems have gained much attention. We
follow~\cite{vandenberg2008paretofrontiertbasispursuit} and denote
them as follows: For a matrix $A\in\mathbf{R}^{k\times n}$ and $b\in\mathbf{R}^k$
and positive numbers $\sigma$, $\lambda$ and $\tau$ we define the
Basis Pursuit Denoising (\cite{chen1998basispursuit}) with constraint by
\begin{equation}
\label{eq:BPsigma}
\tag{$\text{BP}_\sigma$}
\min_x \norm[1]{x}\ \text{ subject to }\ \norm[2]{Ax-b}\leq \sigma,
\end{equation}
the Basis Pursuit Denoising with penalty (\cite{chen1998basispursuit}) by
\begin{equation}
\label{eq:QPlambda}
\tag{$\text{QP}_\lambda$}
\min_x \tfrac12\norm[2]{Ax-b}^2 + \lambda\norm[1]{x},
\end{equation}
and the LASSO (least absolute shrinkage and selection
operator~\cite{tibshirani1996lasso}) by
\begin{equation}
\label{eq:LStau}
\tag{$\text{LS}_\tau$}
\min_x \norm[2]{Ax-b}\ \text{ subject to } \norm[1]{x}\leq\tau.
\end{equation}
All three problems are related: if we denote with
$x_{\text{QP}}(\lambda)$ a solution of~\eqref{eq:QPlambda}, this also
solves~\eqref{eq:BPsigma} for
$\sigma=\norm[2]{Ax_{\text{QP}}(\lambda)-b}$ and~\eqref{eq:LStau} for
$\tau = \norm[1]{x_{\text{QP}}(\lambda)}$ (see
e.g.~\cite{vandenberg2008paretofrontiertbasispursuit,loris2009l1performance}).
However, this relation is implicit and relies in general on the
knowledge of the solutions. Hence, it is not totally true that these
problems are equivalent.
One may argue, that~\eqref{eq:BPsigma} is harder than the other
problems since its objective is nonsmooth and shall be minimized over
a complicated convex set (e.g.~projecting on this set is
difficult). Moreover, one may argue, that~\eqref{eq:QPlambda} is
harder than~\eqref{eq:LStau} since the latter has a smooth objective
(to be minimized over a somehow simple convex set) while the first has
a nonsmooth objective. Computational experience with with these
problems lead to the same conclusion.
Recently, minimization problems similar to Basis Pursuit Denoising
have appeared in several contexts, e.g. group sparsity (or joint
sparsity)
\cite{vandenberg2009jointsparserecovery,fornasier2008jointsparsity,mishali2008reduceandboost}
for sparse recovery, nuclear norm minimization for low-rank matrix
recovery~\cite{recht2010nuclearnorm} to name just two.
\subsection{Notation}
With $\norm[p]{x}$ we denote the $p$-norm of a vector $x\in\mathbf{R}^n$,
$A^T$ is the transpose of a matrix $A$, the range of a matrix $A$ is
denoted with $\rg A$ and with $\Sign(x)$ we denote the multivalued
sign, i.e.
\[
y\in\Sign(x) \iff y_i\
\begin{cases}
= 1 & \text{if }\ x_i>0\\
= -1 & \text{if }\ x_i<0\\
\in [-1,1] & \text{if }\ x_i=0
\end{cases}.
\]
\section{Construction of instances with known solution}
\label{sec:constr-probl-with}
In this section we illustrate how instances (i.e.~tuples
$(A,b,\lambda)$) can be generated, such that the solution $x^*$
of~\eqref{eq:QPlambda} is known up to machine precision. This is
achieved by prescribing the solution $x^*$ (and the matrix $A$ and the
value $\lambda$) and computing a corresponding right hand side $b$.
The basis is the following simple observation which has a one-line
proof:
\begin{lemma}
Let $A\in\mathbf{R}^{k\times n}$, $\lambda>0$ and $x^*\in\mathbf{R}^n$ and let
$w\in\rg A^T$ fulfill $w \in\Sign(x^*)$. Then it holds: If $y$ is a
solution to $A^Ty = w$ and $b$ is defined by $b = \lambda y + Ax^*$,
then $x^*$ is a solution of~\eqref{eq:QPlambda}.
\end{lemma}
\begin{proof}
Simply check
\begin{align*}
-A^T(Ax^*-b) & = -A^T(Ax^* - \lambda y - Ax^*)\\
& = \lambda A^Ty = \lambda w \in \lambda\Sign(x^*).
\end{align*}
Hence $x^*$ fulfills the necessary and sufficient condition for
optimality.
\end{proof}
\begin{remark}
The existence of the vector $w$ is exactly the \emph{source
condition} used in sparse regularization of ill-posed
problems. There one shows that a vector $x^\dagger$ for which such a
vector $w$ exists can be reconstructed from noisy measurements
$b^\delta$ with $\norm[2]{Ax^\dagger- b^\delta}\leq \delta$ by
solving~\eqref{eq:QPlambda} with $b^\delta$ instead of $b$ and
$\lambda\asymp\delta$ and that one achieves a linear convergence
rate, i.e.~for the solution $x_\lambda^\delta$ one gets
$\norm[1]{x_\lambda^\delta-x^\dagger} = \mathcal{O}(\delta)$,
see~\cite{grasmair2008sparseregularization,lorenz2010beyondconvergence,grasmair2011conditionsell1}.
\end{remark}
The following corollary reformulates the above lemma in a way which is
more suitable for an algorithmic reformulation.
\begin{corollary}
\label{cor:construct_b}
Let $\{1,\dots,n\}$ be partitioned into sets $\mathcal{I}$, $\AA_+$ and
$\AA_-$ and let $x^*\in\mathbf{R}^n$ be any vector such that
\begin{equation}
\label{eq:compliance_x}
\begin{split}
x_i^*&>0, \quad i\in \AA_+\\
x_i^*&<0, \quad i\in \AA_-\\
x_i^*&=0, \quad i\in \mathcal{I}
\end{split}
\end{equation}
and let $\lambda>0$. Furthermore assume that $y\in\mathbf{R}^k$ fulfills
\begin{equation}
\label{eq:condition_on_y}
\begin{split}
(A^Ty)_i&=1, \quad i\in \AA_+\\
(A^Ty)_i&=-1 , \quad i\in \AA_-\\
|A^Ty|_i&\leq 1, \quad i\in \mathcal{I}
\end{split}
\end{equation}
and define $b = \lambda y + Ax^*$. Then $x^*$ is a solution
of~\eqref{eq:QPlambda}.
\end{corollary}
According to this corollary we can construct an instance
$(A,b,\lambda)$ with known solution $x^*$ as follows:
\begin{enumerate}
\item Specify $A\in\mathbf{R}^{m\times n}$ and a sign-pattern (given by the
partition $\AA_+$, $\AA_-$, $\mathcal{I}$).
\item Construct a vector $y\in\mathbf{R}^m$ which
fulfills~\eqref{eq:condition_on_y}.
\item Choose any $\lambda>0$ and any $x^*\in\mathbf{R}^n$ which complies with
the sign-pattern, i.e.~\eqref{eq:compliance_x} holds.
\item Define $b = \lambda y + Ax$.
\end{enumerate}
The vector $y$ can be constructed by several methods which are outline
in Appendix~\ref{app:algorithm}. These methods have been implemented
in the Matlab package \texttt{L1TestPack} in the function
\verb|construct_bpdn_rhs| \footnote{The package is available at
\url{http://www.tu-braunschweig.de/iaa/personal/lorenz/l1testpack}.}.
One should note that a vector $y$ as in
Corollary~\ref{cor:construct_b} need not to exist. Indeed, for a fixed
matrix $A$ not every sign-pattern of $x^*$ can occur as a minimizer of
any~\eqref{eq:QPlambda}.
\begin{remark}
For injective $A$ everything is much simpler: Since $A^T$ is
surjective, we can just choose some $w\in\Sign(x^*)$, solve
$A^Ty = w$ and set $b = \lambda y+Ax^*$.
\end{remark}
We discuss advantages and disadvantages
of our approach:\\
\textbf{Advantages:}
\begin{itemize}
\item The algorithm is independent of the value of $\lambda$ while the
performance of solvers for~\eqref{eq:QPlambda} usually deteriorates
for smaller $\lambda$, see,
e.g.~\cite{figueiredo2007gradproj,figueiredo2009sparsa} and
Section~\ref{sec:infl-param-lambda}.
\item The algorithm is independent of the dynamic range of the optimal
value $x^*$, however, several experiments have recorded that the
performance of solvers for~\eqref{eq:QPlambda} depends greatly on
the dynamic range, see, e.g.~\cite{becker2009nesta} and Section~\ref{sec:infl-dynam-range}.
\item For square matrices $A$ with full rank, one immediately get a
desired vector $y$ by solving $A^Ty = w$ for some vector
$w\in\Sign(x^*)$. While this setting is unusual, e.g., in compressed
sensing, one encounters such situations in regularization with
sparsity constraints,
see~\cite{daubechies2003iteratethresh,bredies2008harditer,lorenz2008reglp,griesse2008ssnsparsity,denis2009sparseholograms,ramlau2008regproptikhonov}.
\end{itemize}
\textbf{Disadvantages}
\begin{itemize}
\item The construction of $b$ from $x^*$ leads to a specific noise
model, namely, the noise is given by $\lambda y$. Hence, there is no
control about the noise distribution\footnote{However, one observes
that the noise level $\norm{Ax^*-b} = \lambda \norm{y}$ is
proportional to $\lambda$ which, again, motivates that one should
choose $\lambda$ proportional to the noise level.}. This limits the
use of instances constructed in this way to the comparison of
solvers for basis pursuit denoising. For other sparse reconstruction
methods like matching pursuit algorithms they seem to be useless.
\item The algorithm produces one particular element $w\in\Sign(x^*)$
and it is not clear if this has any additional properties. Usually,
several $w\in\Sign(x^*)\cap \rg A^T$ exist and probably the
proposed method favors a particular form of $w$.
\end{itemize}
\section{Illustrative instances}
\label{sec:illustr-inst}
Numerous papers contain comparisons of different solvers for the three
problems \eqref{eq:BPsigma},~\eqref{eq:QPlambda} and \eqref{eq:LStau},
see
e.g.~\cite{vandenberg2008paretofrontiertbasispursuit,figueiredo2007gradproj,hale2008fixedpointcontinuation,yin2008bregmaniterations,beck2009fista,griesse2008ssnsparsity,becker2009nesta,loris2009l1performance}. Hence,
we not aim at yet another comparison of solvers but try to illustrate,
how different features of the measurement matrix and the solution
influence the difficulty of the problem.
From the zoo of available solvers we have chosen four. The choice was
not uniformly at random but to represent four different classes:
fpc~\cite{hale2008fixedpointcontinuation} as a simple tuning of the
basic iterative thresholding algorithm, FISTA \cite{beck2009fista} as
a representative of the ``optimal algorithms'' in the sense of worst
case complexity, GPSR~\cite{figueiredo2007gradproj} as a highly tuned
basic gradient method and YALL1~\cite{yang2009yall1} as a member of
the class of alternating directions methods\footnote{Sources: fpc
version 2.0 \url{http://www.caam.rice.edu/~optimization/L1/fpc/},
GPSR version 6.0 \url{http://www.lx.it.pt/~mtf/GPSR/}, YALL1 version
1.0 \url{http://yall1.blogs.rice.edu/} and an own implementation of
FISTA.}. All these solvers proceed iteratively and use (basically)
one application of $A$ and one of $A^T$ for each iteration. Hence, the
runtime of these algorithms is mainly related to the number of
iterations. We did not include higher order solvers like
fss~\cite{lee2006featuresignsearch} or
ssn~\cite{griesse2008ssnsparsity} and also did not use any variant of
homotopy approaches~\cite{loris2008l1pack}.
For algorithms we overrode the implemented stopping criteria by
the criterion that the relative error in the reconstruction
\[
R_n = \frac{\norm{x_n-x^*}}{\norm{x^*}}
\]
falls below a given threshold.
\newcommand{\setup}[7]{
\begin{description}
\item[Dimensions:]\mbox{}\\[-\baselineskip]
\begin{itemize}
\item $n={#1}$ variables,
\item $k={#2}$ measurements
\end{itemize}
\item[Matrix $A$:]\mbox{}\\
#3
\item[Solution $x^*$:]\mbox{}\\
$s={#4}$ non-zero entries, #5
\item[$\lambda$:]
{#6}
\item[Results:]\mbox{}\\
#7
\end{description}
}
\subsection{Influence of the parameter $\lambda$}
\label{sec:infl-param-lambda}
Here we consider a standard example from compressed sensing, namely a
sensing matrix $A$ which consists of random rows of a DCT matrix. The
setup is as follows: \setup{1000}%
{200}%
{Random rows of a DCT matrix}%
{20}%
{magnitude normally distributed with mean zero and variance one.}%
{$10^{-1}$, $10^{-2}$, $10^{-4}$}%
{In general, all solver slow down for smaller values of
$\lambda$. However, some solvers depend greatly on the size of
$\lambda$, see Figure~\ref{fig:example1}.}
\begin{figure*}
\centering
\begin{tabular}{ccc}
\includegraphics[page=1]{figures_paper/example1}&
\includegraphics[page=2]{figures_paper/example1}&
\includegraphics[page=3]{figures_paper/example1}\\
$\lambda = 10^{-1}$ &
$\lambda = 10^{-2}$ &
$\lambda = 10^{-4}$
\end{tabular}
\includegraphics{figures_paper/legend}
\caption{Results for from Section~\ref{sec:infl-param-lambda} on the
influence of $\lambda$.}
\label{fig:example1}
\end{figure*}
\subsection{Influence of the sparsity level}
\label{sec:infl-spars-level}
While the construction of a test instance is independent of the
parameter $\lambda$, it gets harder for less sparsity. The behavior of
the solvers with respect to the sparsity level is illustrated by this
example:
\setup{2000}%
{200}%
{Bernoulli ensemble, i.e.~random $\pm 1$}%
{4,\,80}%
{respectively; magnitude normally distributed with mean zero and
variance one.}%
{$10^{-1}$}%
{Most solvers take longer for less sparsity; however, surprisingly,
YALL1 is even faster for lower sparsity, see
Figure~\ref{fig:example2}.}
\begin{figure}
\centering
\begin{tabular}{cc}
\includegraphics[page=1]{figures_paper/example2}&
\includegraphics[page=2]{figures_paper/example2}\\
$s=4$ &
$s=80$
\end{tabular}
\includegraphics{figures_paper/legend_stacked}
\caption{Results for from Section~\ref{sec:infl-spars-level} on the
influence of the sparsity level $s$.}
\label{fig:example2}
\end{figure}
\subsection{Influence of the dynamic range of the entries in $x^*$}
\label{sec:infl-dynam-range}
As claimed in the introduction, the \emph{dynamic range}
\[
\Theta(x^*) = \frac{\max\{\abs{x^*}\, :\, x^*\neq 0\}}{\min\{\abs{x^*}\, :\, x^*\neq 0\}}
\]
also influences the performance.
\setup{3000}%
{1000}%
{Union of three orthonormal basis: the identity matrix, the DCT matrix
and an orthonormalized random matrix}%
{50}%
{with a dynamic range of approximately 9, 701 and 55.000,
respectively.}%
{$10^{-1}$}%
{Some solvers dramatically slow down for larger dynamic range, see
Figure~\ref{fig:example3}}
\begin{figure*}
\centering
\begin{tabular}{ccc}
\includegraphics[page=1]{figures_paper/example3}&
\includegraphics[page=2]{figures_paper/example3}&
\includegraphics[page=3]{figures_paper/example3}\\
$\Theta(x^*) \approx 9$ &
$\Theta(x^*) \approx 700$ &
$\Theta(x^*) \approx 55.000$
\end{tabular}
\includegraphics{figures_paper/legend}
\caption{Results for from Section~\ref{sec:infl-dynam-range} on the
influence of the dynamic range.}
\label{fig:example3}
\end{figure*}
\subsection{Influence of the coherence of $A$}
\label{sec:infl-coher-a}
To illustrate that also a large coherence can cause solvers to slow
down, we have chosen the following setup: We considered square matrices
$A\in\mathbf{R}^{n\times n}$ which are zero expect on the diagonal and a
certain number $K$ of lower off-diagonals, scaled to have $\norm{A}=1$:
\begin{align*}
A_K & = c
\begin{bmatrix}
1 & 0 & \cdots & \cdots& \cdots& 0\\
\vdots & \ddots&\ddots & & & \vdots\\
1 & & \ddots & \ddots & & \vdots\\
0 & \ddots & &\ddots & \ddots&\vdots\\
\vdots & \ddots & \ddots& & \ddots& 0\\
0 & \cdots & 0 & 1 & \cdots & 1
\end{bmatrix}.\\[-0.7\baselineskip]
& \hspace*{20ex}\underbrace{\hspace*{11ex}}_{K\ \text{columns}}
\end{align*}
We also considered the extreme case $K = n$, also known as the
Heaviside matrix. Denoting the columns of $A_K$ by $a_j$, we calculate
the coherence of the matrix $A_K$ as
\[
\mu = \max_{i\neq j}\frac{\scp{a_i}{a_j}}{\norm{a_i}\norm{a_j}} = \sqrt{\frac{K-1}{K}}.
\]
\setup{300}%
{300}%
{Increasingly coherent matrices with $K = 5,\, 40,\, 100,\, 300$}%
{30}%
{Bernoulli, i.e. randomly selected $+1$ and $-1$.}%
{$10^{-1}$}%
{This problem, while with an square and invertible matrix, is known
the be notoriously hard. Especially for large $K$ all solvers
deteriorate, see Figure~\ref{fig:example4}.}
\begin{figure*}
\centering
\begin{tabular}{cccc}
\includegraphics[page=1]{figures_paper/example4}&
\includegraphics[page=2]{figures_paper/example4}&
\includegraphics[page=3]{figures_paper/example4}&
\includegraphics[page=4]{figures_paper/example4}\\
$K=5$, $\mu=0.89 $&
$K=40$, $\mu=0.987 $&
$K=100$, $\mu=0.995$ &
$K=300$, $\mu=0.998$
\end{tabular}
\includegraphics{figures_paper/legend}
\caption{Results for from Section~\ref{sec:infl-coher-a} on the
influence of the coherence.}
\label{fig:example4}
\end{figure*}
\appendices
\section{Algorithms}
\label{app:algorithm}
Instead of $y\in\mathbf{R}^m$ we construct a vector $w\in\mathbf{R}^n$ such that
\[
w \in \rg A^T \cap \Sign(x^*)
\]
which can be reformulated as
\[
\begin{split}
w_i&=1, \quad i\in \AA_+\\
w_i&=-1 , \quad i\in \AA_-\\
|w_i|&\leq 1, \quad i\in \mathcal{I}
\end{split}
\]
and $w\in\rg A^T$. Then $y$ can be found by solving $A^Ty = w$.
\subsection{Solution by projection onto convex sets}
\label{sec:solut-proj-onto}
The condition $w\in\rg A^T \cap \Sign(x^*)$ can be seen as a convex
feasibility problem~\cite{bauschke1996convexfeasibility} since both
the sets $\rg A^T$ and $\Sign(x^*)$ are convex. Moreover, the
projection onto each set is computationally feasible: The projection
onto the range of $A^T$ can be calculated explicitly, e.g.~with the
help of QR factorization. If $A^T = QR$ with orthonormal $Q$ and upper
triangular $R$, the projection $P_{\rg A^T}$ is given by $P_{rg A^T} =
Q(:,1:k) Q(:,1:k)^T$. Projecting onto the convex set $\Sign(x^*)$ is
even simpler: Set the fixed components to $\pm 1$ respectively and
clip the others by $x\mapsto\max(\min(x,1)x,-1)$. We done the
projection onto $\Sign(x^*)$ by $P_{\Sign(x^*)}$.
Now we find $w$ by alternatingly project an initial guess onto both
sets, a strategy knows as \emph{projection onto convex sets} (POCS)
\cite{cheney1959pocs,gubin1967pocs}. This is given as pseudo code in
Algorithm~\ref{alg:pocs}.
\begin{algorithm}
\caption{Calculation of $y$ by POCS}
\label{alg:pocs}
\begin{algorithmic}[1]
\REQUIRE Input $A\in\mathbf{R}^{m\times n}$, a partition $\AA_+$, $\AA_-$
and $\mathcal{I}$ of $\{1,\dots,n\}$ (coded as $\Sign(x^*)$), a tolerance
$\epsilon>0$ and an initial guess $w_0$.
\FOR{$i=0,1,\dots$}
\STATE $v^n = P_{\rg A^T}w^n$
\STATE $w^{n+1} = P_{\Sign(x^*)}v^n$
\IF{$\max(\norm{v^n-w^n},\norm{w^{n+1}-v^n})\leq \epsilon$}
\STATE break
\ENDIF
\ENDFOR
\STATE Solve $A^Ty = w$
\RETURN $y$
\end{algorithmic}
\end{algorithm}
\subsection{Solution by quadratic programming}
\label{sec:solut-quadr-progr}
We sketch another approach by quadratic programming: We call $\AA =
\AA_+\cup\AA_-$ the \emph{active set} and $\mathcal{I}$ the \emph{inactive
set} and define $s\in\mathbf{R}^\AA$ by
\begin{equation}
\label{eq:def_s}
\begin{split}
s_i&=1, \quad i\in \AA_+\\
s_i&=-1 , \quad i\in \AA_-.
\end{split}
\end{equation}
Furthermore we denote with $P_\AA:\mathbf{R}^n\to\mathbf{R}^\AA$ the projection
which deletes the ``inactive'' components and with
$P_\mathcal{I}:\mathbf{R}^n\to\mathbf{R}^\mathcal{I}$ the projection which deletes in ``active''
components and the respective adjoint $P_\AA^T$ and $P_\mathcal{I}^T$ which
fill up the vectors be zeros. With this notation, we aim at finding
$w\in\rg A^T$ such that
\[
P_\AA w = s,\ \text{ and }\ \norm[\infty]{P_\mathcal{I} w}\leq 1.
\]
To fulfill the condition $w\in\rg A^T$ we use the orthogonal
projection on $\rg A^T$, denoted by $P_{\rg A^T}$ and require $P_{\rg
A^T}w = w$. Since $w$ is determined on the active set $\AA$ we
rewrite is as
\begin{equation}
\label{eq:def_w}
w = P_\AA^Ts + P_\mathcal{I}^T z
\end{equation}
with a $z\in\mathbf{R}^\mathcal{I}$. Putting this together we have to find a vector
$z\in\mathbf{R}^\mathcal{I}$ such that
\[
(P_{\rg A^T} - \Id)P_\mathcal{I}^T z = (\Id - P_{\rg A^T})P_\AA^Ts,\quad
\norm[\infty]{z}\leq 1.
\]
We the abbreviations
\begin{equation}
\label{eq:def_barP_barv}
\begin{split}
\bar P &= (P_{\rg A^T} - \Id)P_\mathcal{I}^T\\
\bar v &= (\Id -P_{\rg A^T})P_\AA^Ts
\end{split}
\end{equation}
we reformulate this as the optimization problem
\begin{equation}
\label{eq:min_prob_infty}
\min_{z\in\mathbf{R}^\mathcal{I}} \tfrac12\norm{\bar Pz-\bar v}^2\ \text{s.t.}\ \norm[\infty]{z}\leq 1.
\end{equation}
This quadratic programming or constrained regression problem can be
solved by various methods~\cite{boyd2004convexoptimization} including
the simple gradient projection~\cite{goldstein1965descent} or the
conditional gradient
method~\cite{frank1956quadprog,beck2004conditionalgradient}. Note
that we require that the optimal value of~\eqref{eq:min_prob_infty} is
indeed zero.
Algorithm~\ref{alg:calc_y} gives pseudo-code for calculating $y$.
\begin{algorithm}
\caption{Calculation of $y$ by quadratic programming}
\label{alg:calc_y}
\begin{algorithmic}[1]
\REQUIRE Input $A\in\mathbf{R}^{m\times n}$ and a partition $\AA_+$,
$\AA_-$ and $\mathcal{I}$ of $\{1,\dots,n\}$.
\STATE Set $s$ according to~\eqref{eq:def_s}.
\STATE Calculate the projection matrix $P_{\rg A^T}$ (e.g.~by
QR-factorization or singular value decomposition) and define $\bar
P$ and $\bar v$ according to~\eqref{eq:def_barP_barv}.
\STATE Calculate $z$ as a solution of~\eqref{eq:min_prob_infty}.
\IF{$\bar P z = \bar v$}
\STATE calculate $w$ according to \eqref{eq:def_w}
\ELSE
\STATE return ``Error: No solution with this sign-pattern''
\ENDIF
\STATE Solve $A^Ty = w$
\RETURN $y$
\end{algorithmic}
\end{algorithm}
\bibliographystyle{plain}
|
\section{Introduction}
Quite recently, "buckyballs" have been observed for the first time in space by
Astronomers of University of Western Ontario \cite{nasajully10,jcjbsepsem2010}
using NASA's Spitzer Space Telescope. They identified spectral signatures
of fullerenes in a cloud of cosmic dust surrounding a distant star 6500 light
years away. The existence of fullerenes has thus been
confirmed as anticipated after the first observation in laboratory by
Kroto {\it et al.} at Rice University \cite{hwkjrhscobrfcres85}. One may therefore
expect observation of molecular species trapped in fullerenes in interstellar medium.
In this work we investigate on the changes that are induced in the spectral signatures
of species trapped in fullerene in order to determine how these changes can be used
to characterize the environment where fullerenes are observed. From inert matrix
isolation spectroscopy, one knows that infrared (IR) spectra of molecular
species trapped in a site are simplified in that, as rotational structures
are absent in the spectra, a vibrational transition
is observed as one peak blue or red shifted or sometimes two or more in case of
splitting of degenerate levels \cite{cgal89_1,prdibvrjlthcallam99,prdalhcjmc2006}.
Trapping of atoms or ions and small diatomic molecules in fullerene have been studied
both experimentally and theoretically
\cite{dsbrdjjrsmsdvcsy93,hstwdkbjh93,mshajvrjcrjp93,hajvrjcmsrjp94,
mshajvrjcsmmlgdeg94,rsrjcms97,rjcakms2003,jc91,ciwmawlp93,jhrjbjmgl96_1,jhrjbjmgl96_2,
ehtoavdapesw96,rjc2008,mxfszbrlnjt2008_1,mxfszbrlnjt2008_2,zsfulasn2005}.
However, except for CO \cite{jhrjbjmgl96_2,ehtoavdapesw96,rjc2008}
and H$_2$ \cite{mxfszbrlnjt2008_1,mxfszbrlnjt2008_2} molecules, only a few
of the theoretical work has been done in the infrared region. In the case
of NH$_3$ trapped in fullerene, calculations have been performed
from {\it ab-initio} methods to study the trapping
mechanism \cite{zsfulasn2005,selt2003,mdgmmog2009} but no experimental
work has been published. Erko\c{c} and T\"{u}rker \cite{selt2003} have
reported that up to six NH$_3$ molecules can be trapped
in C$_{60}$, on one hand and Ganji {\it et al.} \cite{mdgmmog2009} suggest
that only one NH$_3$ molecule can form a stable
complex NH$_3$@C$_{60}$ on the other hand. While Slanina {\it et al.}
\cite{zsfulasn2005} have reported a binding energy of -1830 cm$^{-1}$ with
the ammonia molecule oriented towards a pair of parallel pentagons.
The aim of this paper is to determine how fullerene modifies the spectral signature
of ammonia molecule, in particular in the region of the vibration-inversion
mode (namely umbrella mode).
A theoretical study of an ammonia molecule trapped in fullerene C$_{60}$ is thus
performed in order to simulate its infrared spectrum in the vibration-inversion
frequency region.
The choice of NH$_3$ is motivated by the fact that it contains
a nitrogen atom. CO$_2$ could otherwise have been chosen for the simulation.
To simulate the infrared spectra of NH$_3$ in fullerene, we use the site
inclusion model successfully applied to analyze spectra of CO$_2$ isotopologues
\cite{prdalhcjmc2006} isolated in rare gas matrix. When the molecule is
trapped in the nano-cage of fullerene, the interaction with carbon atoms in the
short range scale modifies the translational and rotational degrees of freedom
because the molecule is no longer free to move. As a result, the umbrella mode
which is characterized by the inversion doubling through tunneling of the
nitrogen atom across the plane defined by the three H atoms is perturbed
as discussed below.
To calculate the bar-spectra of NH$_3$ trapped in C$_{60}$ nano-cage a
renormalization procedure is applied on the total Hamiltonian of the system. The
dynamical coupling between the molecular degrees of freedom and those of C$_{60}$
nano-cage is then small enough, at least to a first approximation, to justify
an \textit{initial chaos hypothesis}.
As a result, the density matrix operator of the system can be factorized into a
product consisting of two terms, one related to the NH$_3$ renormalized optical
states and the other to the C$_{60}$ bath states spanned by its vibrational modes.
The dynamical line-shifts and line-widths of the bar-spectra can be disregarded and
C$_{60}$ nano-cage considered as rigid.
The interaction potential model used in our calculations is
presented in section 2 and the renormalization method applied to separate
the different motions from each other described. Results of the calculation
to determine the frequency shifts of the vibrational modes and the orientational
level schemes of the molecule are then given. Sections 3 and 4 are devoted to
a presentation of the formalism used for the construction of the infrared
spectra of NH$_3$ in fullerene and to a discussion of the infrared bar-spectrum
in the frequency region of the $\nu_2$ mode.
\section{The interaction potential energy}
\subsection{Potential energy model}
The interaction potential energy $V_{\text{MC}}$ between the trapped ammonia molecule NH$_3$
and the rigid fullerene molecule C$_{60}$ is modelled as the sum of 12-6 Lennard-Jones (LJ)
pairwise atom-atom potentials characterizing the repulsion-dispersion contributions and the
induction part due to the interaction between the permanent electric multipole moments of the
molecule and their images created at the positions of the polarized fullerene carbon atoms.
It can be expressed as
\begin{eqnarray}
V_{\text{MC}} = \sum\limits_{j=1}^{60}\sum\limits_{i=1}^44\epsilon_{ij}\left\{ \left(
\frac{\sigma _{ij}}{\left| \mathbf{r}_{ij}\right| }\right) ^{12}
-\left( \frac{\sigma _{ij}}{\left| \mathbf{r}_{ij}\right| }\right) ^6 \right\}
-\frac 12\sum\limits_{j=1}^{60} \textbf{E}_{\text{M}}^j :\mathbf{\alpha }^j :\textbf{E}_{\text{M}}^j,
\label{eq1}
\end{eqnarray}
\noindent where \textit{i} and \textit{j} denote the \textit{i}th atom of the ammonia
molecule and the \textit{j}th carbon atom of the fullerene molecule, respectively;
$\epsilon_{ij}$ and $\sigma _{ij}$ are the mixed LJ potential parameters, obtained from the
usual Lorentz-Berthelot combination rules $\epsilon _{ij}=\sqrt{\epsilon _{ii}\epsilon _{jj}}$
and $2\sigma_{ij}=\sigma _{ii}+\sigma _{jj}$ and $\textbf{r}_{ij}$ is the distance vector
between the \textit{i}th atom of the molecule and the \textit{j}th carbon atom.
In the second term of Eq. (\ref{eq1}) $\textbf{E}_{\text{M}}^j$ is the field generated by
the molecular permanent electric multipole moments ($\mathbf{\mu }$, $\mathbf{\Theta }$,...)
of the NH$_3$ molecule at position \textbf{r}$_0$ on the \textit{j}th carbon atom of the
fullerene molecule at position \textbf{r}$_j$ with polarizability
tensor $\mathbf{\alpha }^j$. It is expressed as
\begin{eqnarray}
\textbf{E}_{\text{M}}^j = \mathbf{\nabla\nabla }\left( \frac 1{\left| \mathbf{r}_j-\mathbf{r}_0\right|
}\right) \cdotp\mathbf{\mu} + \frac 13 \mathbf{\nabla\nabla
\nabla }\left( \frac 1{\left| \mathbf{r}_j-\mathbf{r}_0\right| }\right)
\mathbf{:\Theta }+... \label{eq2}
\end{eqnarray}
Note that the multipole moments of NH$_3$ and the polarizability tensor of each carbon atom
of C$_{60}$ are usually defined and given with respect to local frames with
their origins at the centre of mass of NH$_3$ (G,\textbf{x},\textbf{y},\textbf{z}) or
each carbon atom (C$_j$,\textbf{x}$_j$,\textbf{y}$_j$,\textbf{z}$_j$) respectively.
In the calculation, it is then necessary to express these quantities in an absolute
frame (O,\textbf{X},\textbf{Y},\textbf{Z}) defined with respect to the fixed C$_{60}$
molecule. We use a transformation rotation matrix following Rose's
convention \cite{mer67} and given in Appendix B.
Figure 1 indicates how to describe the internal degrees of freedom of the
molecule with respect to its frame (G,\textbf{x},\textbf{y},\textbf{z})
and its external orientational and translational degrees of freedom with respect
to the absolute frame, and Table 1 gives the various fullerene
and ammonia characteristic values used in our calculations.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=16cm]{Figure1.eps}
\end{center}
\caption{Geometrical characteristics of a molecule trapped in the
fullerene molecule. (O,\textbf{X},\textbf{Y},\textbf{Z})
and (G,\textbf{x},\textbf{y},\textbf{z}) represent the absolute
frame and the molecular frame, respectively. \textit{q}
and $\beta$ define the equilibrium internal coordinates of the trapped molecule.}
\label{FIG. 1}
\end{figure}
\begin{table}[h!]
\begin{center}
\caption{Numerical parameters for NH$_3$ molecule and
for C$_{60}$ molecule used in our calculations.}
\begin{tabular}{llll}
\hline\hline
NH$_3$ & & C$_{60}$ & \\
\hline
$q$ (\AA ) & 1.0156 & $R^{(a)}$ (\AA ) & 3.556 \\
$\beta$ (deg.) & 68. & d(C$-$C)$^{(a)}$ (\AA ) & 1.455 \\
& & d(C$=$C)$^{(a)}$ (\AA ) & 1.398 \\
$\mu ^{\mathrm{e}}$ (D) & 1.476 & $\alpha _{\bot }^{\mathrm{c}}$
(\AA $^3$) & 1.44 \\
$\Theta ^{\mathrm{e}}$ (D\AA ) & -2.930 & $\alpha _{\Vert }^{\mathrm{c
}$ (\AA $^3$) & 0.41 \\
$B$ (cm$^{-1}$) & 9.941 & & \\
$C$ (cm$^{-1}$) & 6.309 & & \\
\hline
\hline
& C - C$^{(b)}$ & N - N$^{(c)}$ & H - H$^{(c)}$ \\
\hline
$\epsilon$ (cm$^{-1}$) & 19.4 & 26.5 & 17.2 \\
$\sigma$ (\AA ) & 3.40 & 3.38 & 2.53 \\
\hline
\raggedright $(a)$ From Ref. \cite{khlhdbcabhcdrdjmdv91}.
\raggedright $(b)$ From Ref. \cite{hysxwz2009}.
\raggedright $(c)$ From Ref. \cite{al2000}.
\end{tabular}
\end{center}
\label{Table 1}
\end{table}
The distance vector $\mathbf{r}_{ij}$ in Eq.(\ref{eq1}) can be expressed in
terms of the position vectors \textbf{r}$_0$ of the molecular centre of mass (c.m.)
and $\mathbf{r}_j$ of the $j$th carbon atom of the fullerene molecule with respect
to the absolute frame (O,\textbf{X},\textbf{Y},\textbf{Z}) and of the position
vector $\mathbf{\eta }_i$ of the $i$th atom of the molecule with respect to its
associated frame (G,\textbf{x},\textbf{y},\textbf{z}):
\begin{equation}
\mathbf{r}_{ij} = \mathbf{r}_j-\mathbf{r}_0-\mathbf{\eta }_i, \label{eq3}
\end{equation}
The IR spectra of the trapped NH$_3$ molecule are usually given in terms of
molecular vibrational \{$Q$\} degrees of freedom.
Thus, to determine the influence of the surroundings on the molecular
internal motions, the position vectors $\mathbf{\eta }_i$ of the atoms in the
molecule, the molecular dipole moment vector $\mathbf{\mu }$, and the
quadrupole moment tensor $\mathbf{\Theta }$ can all be expressed in a series
expansion with respect to the molecular frame in terms of the vibrational
normal coordinates \{$Q$\}.
Taking only the first order terms gives
\begin{eqnarray}
\mathbf{\eta }_i &=&\mathbf{\eta }_i^{\text{e}}+\sum\limits_\nu \mathbf{a
_i^\nu Q_\nu , \nonumber \\
\mathbf{\mu } &=&\mathbf{\mu }^{\text{e}}+\sum\limits_\nu \mathbf{b}^\nu
Q_\nu , \nonumber \\
\mathbf{\Theta } &=&\mathbf{\Theta }^{\text{e}}+\sum\limits_\nu \mathbf{c
^\nu Q_\nu . \label{eq4}
\end{eqnarray}
In these expressions $\mathbf{\eta }_i^{\text{e}}$ is the vector position of
the \textit{i}th atom in the rigid molecule, and $\mathbf{\mu }^{\text{e}}$
and $\mathbf{\Theta }^{\text{e}}$ are its permanent multipole moments,
while $\mathbf{a}_i^\nu $, $\mathbf{b}^\nu $, and $\mathbf{c}^\nu $ are the
first derivatives of $\mathbf{\eta }_i$, $\mathbf{\mu }$, and $\mathbf{\Theta }$
with respect to the normal coordinate $Q_\nu $ which describes the $\nu $th
molecular vibrational mode with frequency $\omega _\nu $.
Note that the expression of $\mathbf{\mu }$ will also be used in the near and
far infrared spectra calculations.
However, it must be noticed that because of the \textit{dual nature} of the
$\nu_2$ mode of the ammonia molecule, that is, the \textit{high frequency vibration-low
frequency inversion}, this mode will be treated in a particular way.
Using the Born-Oppenheimer approximation (adiabatic approximation) to separate the
high frequency vibrational modes \{$Q$\} of the molecule from its low frequency
external modes, that is, the orientation $\mathbf{\Omega}=(\varphi,\theta,\chi)$ and
c.m. translation \textbf{r}$_0$ motions, the interaction
potential energy $V_{\text{MC}}$ can be written as
\begin{equation}
V_{\text{MC}} = V_{\text{M}}(\mathbf{r}_0,\mathbf{\Omega}) +
W_{\text{M}}(\{Q\}) + \bigtriangleup V_{\text{M}}(\mathbf{r}_0,\mathbf{\Omega},
\{Q\}), \label{eq5}
\end{equation}
\noindent where $V_{\text{M}}(\mathbf{r}_0,\mathbf{\Omega})$ represents the low frequency
motions dependent part of the potential energy for the non vibrating molecule,
and $W_{\text{M}}(\{Q\})$ characterizes the vibrational dependent part for
the molecule at its equilibrium position and orientation, and is generally a small
perturbation which involves only vibrational level shifts and/or splittings.
It can be developed in a Taylor series expansion up to second order with respect to
the vibrational normal coordinates \{$Q$\} as
\begin{equation}
W_{\text{M}}(\{Q\}) = \sum\limits_\nu \frac{\partial
V_{\text{MC}}}{\partial Q_\nu }Q_\nu +\frac 12\sum\limits_{\nu \nu ^{\prime }}\frac{
\partial ^2V_{\text{MC}}}{\partial Q_\nu \partial Q_{\nu ^{\prime }}}Q_\nu Q_{\nu
^{\prime }}+... \label{eq6}
\end{equation}
Finally the third term in Eq.(\ref{eq5}) characterizes the coupling between the
external and the internal modes which can lead to the vibrational energy relaxation.
\subsection{Potential energy surfaces}
\subsubsection{Orientation-translation motions}
The first step of the numerical calculations consists in determining the equilibrium
configuration of the rigid ammonia molecule into the C$_{60}$ cage. The potential
energy $V_{\text{M}}$ (see Eq.(\ref{eq5}) is minimized with
respect to both the c.m. displacement vector \textbf{r}$_0$ and
the molecular orientional coordinates $\varphi$, $\theta$, and $\chi$.
We find that the equilibrium configuration has an energy minimum value of
- 1460 cm$^{-1}$ for the \textbf{z} molecular axis (threefold symetry axis C$_3$)
oriented along the direction connected to each two
symmetrical carbon atoms. The c.m. is then displaced from the site centre by
about 0.184 $\text{\AA}$ in the same direction.
An energy maximum of - 1445 cm$^{-1}$ is obtained when the molecule is oriented and
its c.m. displaced by 0.188 $\text{\AA}$ along the pentagon centre directions.
There is also an intermediate energy value of - 1455 cm$^{-1}$ that is calculated for
the molecule oriented and its c.m. displaced by 0.180 $\text{\AA}$ along the hexagon
centre directions.
Thus, the calculations show that \textit{i}) in all the molecular orientational
configurations, the c.m. displacements \textit{r}$_0$ have always opposite directions
to the molecular \textbf{z} axis, and \textit{ii}) the energy changes are negligibly
small $\lesssim$ 3 cm$^{-1}$ when \textit{r}$_0$ changes by about 0.01 $\text{\AA}$
along these directions. So, in the following we will take the value
\textit{r}$_{0}^{\text{e}}$ = 0.184 $\text{\AA}$ as a mean value for every orientation.
This means that the trajectory of the molecular c.m. can be approximated by the
surface of a sphere of radius \textit{r}$_{0}^{\text{e}}$.
Note, moreover, that the induction and repulsion-dispersion contributions
represent, respectively, 35$\%$ (- 498 cm$^{-1}$) and 65$\%$ (- 963 cm$^{-1}$) of
the total energy.
As a result, the potential energy surface
$V_{\text{M}}(\mathit{r}_0(\mathbf{\Omega}))$ is nearly flat since the barrier
height is always $\lesssim$ 16 cm$^{-1}$ and the c.m. displacement vectors \textbf{r}$_0$
from the cage centre remain always nearly anticollinear to the molecular \textbf{z}
axis \textit{i.e.} \textbf{z}$\cdotp$ \textbf{r}$_0$ $\simeq$ - \textit{r}$_0$.
This is not surprising because of the repulsive character of the hydrogen atoms
of the molecule and reflects a very strong coupling between its orientation
and the location of its centre of mass.
Thus, on one hand, the axis of the molecule undergoes only oscillatory motions
($\varphi,\theta$) of small amplitude around the direction of each displacement vector
\textbf{r}$_{0}^{\text{e}}$. The corresponding potential energy surface which is given in
Figure 2 is a nearly harmonic two dimension librational oscillator.
The associated frequencies are $\omega_{\varphi}$ $\simeq$ 155 cm$^{-1}$ and
$\omega_{\theta}$ $\simeq$ 151 cm$^{-1}$.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=10cm]{Figure2.eps}
\end{center}
\caption{Potential energy surface \textit{versus} ($\varphi,\theta$) angular motions
around the molecular equilibrium configuration
($\mathbf{r}_{0}^{\text{e}}$,$\mathbf{\Omega}$).}
\label{FIG. 2}
\end{figure}
On the other hand, for a given molecular orientation, the radial potential
function $V_{\text{M}}(\textit{r}_0)$ of the translation motion of the molecular c.m.
around the equilibrium position $r_{0}^{\text{e}}$ = 0.184 $\text{\AA}$ is
plotted in Figure 3. The curve appears as that of a nearly harmonic one dimensional
oscillator with a force
constant value \textit{k} = 2.4$\times$10$^4$ cm$^{-1} \ldotp \text{\AA}^{-2}$
corresponding to a frequency $\omega_{\text{rad}}$ $\simeq$ 218 cm$^{-1}$.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=10cm]{Figure3.eps}
\end{center}
\caption{Potential energy function \textit{versus}
molecular c.m. displacement $r_0$ around its equilibrium position
vector \textbf{r}$_{0}^{\text{e}}$.}
\label{FIG. 3}
\end{figure}
On the basis of these features, we can conclude that ammonia should exhibit a
nearly free rotation motion simultaneously around the centre of the cage by describing
a spherical curve with radius \textit{r}$_{0}^{\text{e}}$ = 0.184 $\text{\AA}$, and
around its centre of mass. Note however that the spinning motion $\chi$ remains around
the threefold symmetry axis of the molecule, only.
Finally, it is clear that this rotation-c.m. translation motion is slow with respect to the
librational and the radial motions and can be separated from the latter using the
adiabatic approximation.
\subsubsection{Inversion-translation motions}
It is also clear, from the results above, that when the inversion motion of the molecule
proceeds, its centre of mass moves simultaneously from one position
(\textbf{r}$_{0}^{\text{e}},\mathbf{\Omega}$) to the opposite one
(-\textbf{r}$_{0}^{\text{e}},\mathbf{\Omega}$) with respect to the cage
centre where the molecule lies in its planar configuration.
\subsubsection{Few remarks}
It must be noticed that the interaction potential energy experienced by the ammonia
molecule trapped in the fullerene molecule strongly depends on the geometrical
characteristics such as the radius of the C$_{60}$ cage and the equilibrium
internal coordinates (\textit{q},$\beta$ see Fig. 1) of the ammonia molecule, and
also on the dispersion-repulsion potential parameters especially the
Lennard-Jones $\sigma$ parameters.
For instance, in one hand, the minimum depth is enhanced from - 1460
to - 750 cm$^{-1}$ when the fullerene radius is decreased
by 1$\%$, from 3.556 to 3.519 $\text{\AA}$, and it diminishes
to - 2000 and - 2411 cm$^{-1}$ when the
radius value is increased by 1$\%$ and 2$\%$, to 3.592 and 3.627 $\text{\AA}$,
respectively. The induction energy contribution undergoing only weakly variations.
In the other hand, changes of the $\sigma$ parameters of + 1$\%$ and - 1$\%$
involve, respectively, minimum depths of - 952 and - 1976 cm$^{-1}$.
However, in spite of these changes in the well depth values, the potential energy surface
$V_{\text{M}}(\mathbf{r}_{0}^{\text{e}},\mathbf{\Omega})$ remains nearly flat since
its barrier height does not exceed 24 cm$^{-1}$. Furthermore, the potential energy surfaces
associated with the various motions remain nearly unchanged.
Thus, from the previous analysis of the potential energy experienced by the
ammonia molecule into the fullerene molecule, we will assume that the c.m. position
vector \textbf{r}$_{0}$ remains permanently anticollinear to the \textbf{z} molecular
axis and separate it into a static position vector
\textbf{r}$_{0}^{\text{e}}(\varphi,\theta,Q_2)$ which parametrically depends on
the orientation and inversion degrees of freedom, and a dynamical
coordinate \textbf{u}$_{0}$ characterizing the c.m. vibration
around \textbf{r}$_{0}^{\text{e}}(\varphi,\theta,Q_2)$.
Finally, it is clear that inversion-c.m. translation motions
(\textit{r}$_{0}^{\text{e}}$,$Q_2$), on one hand, and
orientation-c.m. translation motions (\textbf{r}$_{0}^{\text{e}}$,$\mathbf{\Omega}$),
on the other hand, are strongly coupled; their kinetic operator must be classically
treated, as explained in appendix A, before their quantum mechanical treatment as
described below. The interaction potential energy $V_{\text{MC}}$ can be rewritten
as
\begin{equation}
V_{\text{MC}} = V_{\text{M}}(\mathbf{r}_{0}^{\text{e}},\mathbf{\Omega}) +
V_{\text{M}}(\textit{r}_{0}) + V_{\text{M}}(\textit{r}_{0},Q_2) +
W_{\text{M}}(\{Q\}) + \bigtriangleup V_{\text{M}}(\mathbf{r}_0,\mathbf{\Omega},
\{Q\}), \label{eq7}
\end{equation}
\section{Quantum-classical model}
As we have mentioned above, the purpose of this work is to determine the infrared
spectra of the NH$_3$ molecule trapped in a rigid C$_{60}$ molecule. The Hamiltonian
of the optically active molecule considered as quantum mechanical system can
be expressed from the potential energy surfaces calculated above and the quantum
mechanical model for the orientational and vibrational kinetic energy operators.
The latter are obtained from the correspondence principle applied to the classical
formulation described in Appendix A. Disregarding the dynamical couplings between
the orientational and vibrational degrees of freedom the Hamiltonian can thus be
written as
\begin{equation}
H_{\text{a}}^{\text{eff}} = H_{\text{orient}}^{\text{eff}} + H_{\text{vib}}^{\text{eff}};
\label{eq8}
\end{equation}
\noindent where $H_{\text{orient}}^{\text{eff}}$ and $H_{\text{vib}}^{\text{eff}}$
are, respectively, the renormalized Hamiltonians associated with the
orientation and the vibration modes which account for the couplings
(\textbf{r}$_{0}^{\text{e}}$,$\mathbf{\Omega}$)
and (\textit{r}$_{0}^{\text{e}}$,$Q_2$).
\subsection{The orientation modes}
Following expressions derived in Appendix A, the
quantum mechanical renormalized Hamiltonian $H_{\text{orient}}^{\text{eff}}$ for
the non-vibrating molecule can be written as
\begin{equation}
H_{\text{orient}}^{\text{eff}} = T_{\text{rot}}^{\text{eff}} +
V_{\text{M}}(\mathbf{r}_{0}^{\text{e}},\mathbf{\Omega}); \label{eq9}
\end{equation}
\noindent where $T_{\text{rot}}^{\text{eff}}$ is
the effective rotational kinectic operator
\begin{eqnarray}
T_{\text{rot}}^{\text{eff}} = -B^{\text{eff}}\left\{ \frac{\partial ^2}
\partial \theta ^2} + \cot \theta \frac
\partial {\partial \theta } + \frac 1{\sin ^2\theta }\frac{\partial ^2}
\partial \varphi ^2} + \left( \cot ^2\theta +
\frac {C}{B^{\text{eff}}}\right) \frac{\partial ^
}{\partial \chi ^2} - 2\frac{\cot \theta }{\sin \theta }\frac{\partial ^2}
\partial \varphi \partial \chi }\right\}. \label{eq10}
\end{eqnarray}
In this equation $B^{\text{eff}}$ = $\hbar^2/2I_{\text{B}}^{\text{eff}}$ is
the effective rotational constant connected to the rotation-c.m. translation motion of the
C$_3$ symmetry axis for the trapped molecule and $C$ its rotational constant of
the spinning motion around this axis for the gas-phase molecule
$C$ = 6.309 cm$^{-1}$. The value of $B^{\text{eff}}$ is equal
to 7.419 cm$^{-1}$ which is less than the gas-phase value of 9.941 cm$^{-1}$.
Furthermore, since the barrier height associated with the orientation-c.m. translation
motion is always $\lesssim$ 16 cm$^{-1}$, the potential
$V_{\text{M}}(\mathbf{r}_{0}^{\text{e}},\mathbf{\Omega})$ can be neglected.
The renormalized Hamiltonian becomes
$H_{\text{orient}}^{\text{eff}} \simeq T_{\text{rot}}^{\text{eff}}$.
The eigensolutions of this Hamiltonian are then
$E_{JMK}=B^{\text{eff}}J(J+1)+(C-B^{\text{eff}})K^2$ for the eigenvalues
(energy levels) and $|JMK\rangle$ for the eigenvectors where $J$, $M$, and $K$
are the usual quantum numbers connected to the free rotation motion of the
ammonia molecule with the conditions $-J\leq M \leq +J$ and $-J\leq K \leq +J$.
Note that these levels have (2$J$+1)-fold degeneracy on the quantum
number $M$ and twofold degeneracy ($\pm$) on the $K$ one, except when $K=0$ \cite{chtals55}.
The corresponding line structure of the far infrared (FIR) bar-spectrum is then
identical to that of NH$_3$ gas-phase one.
\subsection{The $\nu_2$ vibration-inversion mode}
\subsubsection{Gas-phase NH$_3$}
In the vibration-inversion mode of the NH$_3$ molecule, the nitrogen atom can pass
from one side of the plane defined by the three hydrogen atoms to the opposite one.
This mode is generally described, to a good approximation, as one-dimension
tunneling motion of one particle of mass equal
to the reduced mass $\mu_2$ of NH$_3$ (2.563 g.mol$^{-1}$) moving in a symmetrical
double-well potential function $V_{\nu_2}(Q_2)$ (left (L) and right (R) wells) as
shown in Figure 4a, with a high but finite hindering barrier
of 2047 cm$^{-1}$ \cite{jdsjai62}, such as
\begin{equation}
V_{\nu_2}(Q_2) = \frac 12 k_2 Q_{2}^2 + A_2 (\text{e}^{-a_{2}Q_{2}^2} - 1); \label{eq11}
\end{equation}
\noindent where $k_2$ = 58306 cm$^{-1} \ldotp \text{\AA}^{-2}$, $A_2$ = 12469 cm$^{-1}$ and
$a_{2}$ = 4.8224 $\text{\AA}^{-2}$.
The solution of such a problem leads to vibrational energy levels split into
doublets, ($+$) and ($-$), due to inversion. The vibration-inversion wavefunctions
are then described by $|v_{2}^{(\alpha)}\rangle$ ($\alpha = +,-$), instead of
$|v_2\rangle$, and are expressed as linear combinations (symmetric and
antisymmetric) of the vibrational wavefunctions $|v_{2}^{(\text{L})}\rangle$ and
$|v_{2}^{(\text{R})}\rangle$ associated with the left and right wells, respectively.
Moreover, the only permitted transitions are $(+) \longleftrightarrow (-)$ as shown
in Fig. 4a.
Note that the potential parameters given above were adjusted to give the fundamental
frequency and the level splittings corresponding to the experimental
values, that is, 950 cm$^{-1}$ for the frequency
and 0.8 cm$^{-1}$ for $v_2$ = 0 and 35.8 cm$^{-1}$ for $v_2$ = 1 level splittings.
\subsubsection{Trapped NH$_3$}
When the NH$_3$ molecule is trapped in C$_{60}$ cage, the gas-phase double-well
potential function is modified. In the calculations, the inversion-c.m. translation
potential energy $V_{\text{M}}(r_0,Q_2)$ (see Eq. (\ref{eq7})) is numerically
added to the potential function $V_{\nu_2}(Q_2)$ of Eq. (\ref{eq11}).
The new double-well potential function is shown in Fig. 4b as a function of
the inversion angular coordinate. Its symmetry is preserved but the hindering
barrier is increased by 404 cm$^{-1}$.
Moreover, from expressions given in Appendix A, the effective reduced
mass is equal to $\mu_{2}^{\text{eff}}$ = 6.569 g.mol$^{-1}$ leading to
an increase of 156$\%$ with respect to the gas-phase value of 2.563 g.mol$^{-1}$.
The corresponding Schr\"{o}dinger equation was solved numerically using
a discrete variable representation method \cite{jcliphjvl85}. The computed level
scheme is given in Fig. 4b.
In gas-phase, the $\nu_2$ mode is measured at 931.7 cm$^{-1}$ and 968.3 cm$^{-1}$.
In C$_{60}$, it is shifted by 47.0 cm$^{-1}$ (as given below) and because the
double-well inversion potential is modified with an increase in the hindering
barrier height and the effective reduced mass as discussed above, this
mode is quasi-degenerate and is calculated at 751.6 cm$^{-1}$, that is,
(704.6+47.0) cm$^{-1}$ (see Fig. 4b). This result is to be related to an
increase in the tunneling
time for both the fundamental and the first excited levels in C$_{60}$, since
the values are equal to 20.85 ps and 0.47 ps in gas-phase compared
to 55594 ps and 333.56 ps in C$_{60}$, respectively. We can infer from these
results that in C$_{60}$, localization of the N atom occurs in one of the
wells of the potential function and that the $\nu_2$ mode is quasi-degenerate.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=7cm]{Figure4a.eps}
\includegraphics[width=7cm]{Figure4b.eps}
\end{center}
\caption{Vibration-inversion double-well potential energy of NH$_3$ as a function of
the angle between the N$-$H bond and the plane of the H atoms. The obtained energy
levels and the allowed fundamental transitions are also shown (a) for the
gas-phase and (b) for the trapped one in the fullerene nano-cage.}
\label{FIG. 4}
\end{figure}
\subsection{The vibrational frequency shifts}
The second term of Eq. (\ref{eq8}) is the renormalized vibrational Hamiltonian
which accounts for the vibrational dependent part
$W_{\text{M}}(\{Q\})$ and inversion-c.m. translation part $V_{\text{M}}(r_0,Q_2)$ of the
interaction potential energy
\begin{equation}
H_{\text{vib}}^{\text{eff}} = H_{\text{vib}} + W_{\text{M}}(\{Q\}) + V_{\text{M}}(r_0,Q_2);
\label{eq12}
\end{equation}
\noindent where $H_\text{vib}$ characterizes the gas-phase molecular vibrational
Hamiltonian written as
\begin{equation}
H_{\text{vib}} = \sum\limits_\nu \frac{P_\nu ^2}{2\mu _\nu }+\frac 12 \sum\limits_\nu k_\nu
Q_\nu ^2+ \bigtriangleup V_{\text{vib}}^{\text{anh}}(\{Q\}). \label{eq13}
\end{equation}
In this expression $\mu _\nu$ and $k _\nu$ are, respectively, the reduced mass and
the harmonic force constant connected to the $\nu$th vibrational mode, with associated
normal coordinate $Q_\nu$ and conjugate momentum $P_\nu$, and
$\bigtriangleup V_{\text{vib}}^{\text{anh}}(\{Q\})$ represents the anharmonic part of
the internal potential function. The eigenelements of this Hamiltonian $H_{\text{vib}}$
have previously been obtained from the \textit{ab initio} calculations of
Martin \textit{et al.} \cite{jmlmtjlprt92}.
Let $E_{v_\nu}$ and $|...v_\nu...\rangle$ be the eigenvalues and eigenvectors of
the \textit{v}th level associated with the $\nu$th vibrational mode.
Thus, a first order perturbation treatment allows us to determine the frequency
shift $\Delta \omega _\nu$ associated with the vibrational fundamental transition
$\mid...0_\nu ...\rangle \rightarrow \mid ...1_\nu...\rangle $ for
the trapped molecule from the equation
\begin{equation}
\Delta \omega _\nu = \hslash^{-1}\left[ \langle ...1_\nu...\mid W_{\text{M}}(\{Q\})
\mid ...1_\nu...\rangle -
\langle ...0_\nu ...\mid W_{\text{M}}(\{Q\})\mid ...0_\nu...\rangle \right]; \label{eq14}
\end{equation}
\noindent in which all other modes remain in their fundamental states
and $\hslash = h/2\pi$. The computed values are given in Table 3. We note that each
of the vibrational modes are blue shifted with splitting of the degenerate
levels of $\nu_3$ mode (strong) and $\nu_4$ mode (weak) modelled from static effect.
Whereas for degenerate $\nu_2$ mode the blue shift of 47.0 cm$^{-1}$ from static
effect, is modified to red shift when dynamic effect on 1D tunneling motion
of the N atom is considered as discussed above.
\begin{table}[h!]
\caption{Vibrational frequency shifts (cm$^{-1}$) of NH$_3$ molecule
trapped in the C$_{60}$ cage, together with the gas-phase frequencies. }
\begin{tabular}{ccc}
\hline\hline
\begin{tabular}{c} Vibrational modes \end{tabular} & \begin{tabular}{c} Frequency
shifts (cm$^{-1}$) \end{tabular} & \begin{tabular}{c} Gas-phase frequencies (cm$^{-1}$)
\end{tabular} \\
\hline
$\nu_1$ & \multicolumn{1}{c}{133.0} & \multicolumn{1}{c}{3337} \\
$\nu_2$ & \multicolumn{1}{c}{47.0} & \multicolumn{1}{c}{950} \\
$\nu_{3a}$ & \multicolumn{1}{c}{171.0} & \multicolumn{1}{c}{3448} \\
$\nu_{3b}$ & \multicolumn{1}{c}{69.6} & \multicolumn{1}{c}{} \\
$\nu_{4a}$ & \multicolumn{1}{c}{47.1} & \multicolumn{1}{c}{1627} \\
$\nu_{4b}$ & \multicolumn{1}{c}{34.0} & \multicolumn{1}{c}{} \\
\hline
\end{tabular}
\label{Table 2}
\end{table}
\section{Infrared absorption spectra}
\subsection{General}
The infrared absorption coefficient for the optically active ammonia molecule
trapped in a C$_{60}$ fullerene molecule at temperature \textit{T}, is defined
as the real part of the spectral density, \textit{i.e.}, the Fourier transform of the
time-dependent autocorrelation function $\phi (t)$
\begin{eqnarray}
I\left( \omega \right) &=&\frac{4\pi \mathcal{N}\omega }{3hc}\mathrm{Re
\int\limits_0^\infty dte^{i\omega t}\phi (t), \nonumber \\
\phi (t) &=&\text{Tr}\left[ \rho (0)\mathbf{\mu }_A(0)\mathbf{\mu
_A(t)\right], \label{eq15}
\end{eqnarray}
\noindent where $\omega $ is a running frequency variable
and $c$ the speed of light, $\rho (0)$
characterizes the initial canonical density matrix operator of the system, and
$\mu _A$ is the molecular dipole moment operator defined in the absolute frame
(see Appendix B). The trace (Tr) operation means an average over the initial
conditions at time $t=0$ and over all the possible evolutions of the system
between times $0$ and $t$.
However, since this work is devoted to constructing the infrared bar-spectra
of the trapped ammonia molecule, we assume the \textit{initial chaos hypothesis} for the
density matrix to be valid. This allows us to write the optical wave functions
$\mid v_\nu JMK\rangle$ as products of the renormalized vibrational and
orientational ones $\mid v_\nu \rangle \otimes |JMK\rangle_{v_{\nu}}$.
It must be noticed that the orientational states parametrically depend on the
vibrational states through the moments of inertia. This dependence will be
ignored below.
\subsection{Near infrared bar-spectrum}
Within this framework, the infrared bar-spectrum connected to the $\nu$th
fundamental vibrational mode transition
$\mid 0_\nu\rangle \longrightarrow \mid 1_\nu\rangle $
(all other modes being in their fundamental states) is written in the Lorentzian form as
\begin{eqnarray}
I_\nu (\omega ) = \frac{8\pi ^2\mathcal{N}}{3hc}\omega
\left| \left\langle 0_\nu \left| Q_\nu \right|
1_\nu\right\rangle \right| ^2
\sum\limits_{ {JMK}{J^{\prime }M^{\prime }K^{\prime }}}\frac
e^{-\beta E_{0JMK}}-e^{-\beta E_{1J^{\prime }M^{\prime }K^{\prime}}}}Z
\left| \left\langle JMK\left| \frac{\partial \mathbf{\mu }_
}{\partial Q_\nu }\right| J^{\prime }M^{\prime }K^{\prime
\right\rangle \right| ^2 \nonumber\\
\times \delta \left(\omega -\omega _{0JMK\rightarrow
1J^{\prime }M^{\prime }K^{\prime }}-\Delta \omega _{\nu}\right).
\label{eq16}
\end{eqnarray}
\noindent In this equation the $\left\langle ...\right\rangle $ brackets
refer to vibrational transition elements of the normal coordinate $Q_\nu $
and to the orientational transition elements of the first derivative of the
molecular dipole moment with respect to this coordinate. The numerical values
connected to all the NH$_3$ modes are given in Appendix B. Moreover, $Z$ defines the
orientation canonical partition function associated with the fundamental
vibrational level at temperature $T$, $\beta=(k_{B}T)^{-1}$ where $k_B$ is the
Boltzmann's constant, $\delta$ is the Dirac's function, and $\omega
_{0JMK\rightarrow 1J^{\prime }M^{\prime }K^{\prime }}=\hbar
^{-1}\left( E_{1J^{\prime }M^{\prime }K^{\prime }}-E_{0JMK}\right)$
is the transition frequency. Note that we use below the usual spectroscopic
nomenclature to designate the line transitions connected to the rotational
selection rules, that is, Q($J_K$) for $\Delta J = 0$, R($J_K$) for $\Delta J = +1$,
and P($J_K$) for $\Delta J = -1$.
\subsection{$\nu_2$ mode results and discussion}
In Figure 5, are given the near infrared bar-spectra of NH$_3$ trapped in fullerene
in the $\nu_2$ mode absorption region for three different temperatures, 10 K, 30 K
and 100 K. The spectra are calculated by taking into account the frequency shift and
allowed rovibrational transition elements following selection rules.
For the latter, as the first derivative of the molecular dipole moment with respect
to the $Q_2$ coordinate does not depend on the rotational spinning angle $\chi$,
$\Delta K= 0$ for the $K$ quantum number.
The infrared bar-spectrum consists of a Q line located at the pure
vibration-inversion frequency 751.6 cm$^{-1}$, and a set of
rotation-vibration-inversion lines P and R 14.8 cm$^{-1}$ apart, corresponding
to 2$B^{\text{eff}}$, on both sides of the Q branch.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=15cm]{Figure5a.eps}
\end{center}
\begin{center}
\includegraphics[width=15cm]{Figure5b.eps}
\end{center}
\begin{center}
\includegraphics[width=15cm]{Figure5c.eps}
\end{center}
\caption{Near infrared bar-spectrum in the $\nu_2$ vibration-inversion frequency region for
NH$_3$ trapped in the fullerene nano-cage at temperatures 10, 30 and 100 K.}
\label{FIG. 5}
\end{figure}
At $T$ = 10 K, only the lowest few rotational states are populated. As a
result, the bar-spectrum shown in Figure 5a exhibits only three lines:
\textit{i}) a Q line at 751.6 cm$^{-1}$ which is in fact a superposition of
the three transitions Q($1_1$), Q($2_1$) and Q($2_2$),
\textit{ii}) a R($0_0$) line at 766.4 cm$^{-1}$
and \textit{iii}) a R line, superposition of R($1_0$) and R($1_1$) lines
at 781.3 cm$^{-1}$. There is only one P($1_0$) line
located at 736.8 cm$^{-1}$.
As the temperature increases, more and more rotational states
are populated and as result, more lines appear in the spectra with a different
distribution in terms of intensities. At 30 K, the Q line is the most intense
consisting of a superposition of Q($J_K$) lines for $J$ = 1 to 4
and $1\leq K \leq J$. Moreover, the intensity of lines R($0_0$)
and R($1_K$) decreases leading to the increase of the R($2_K$) line which is located
at 796.1 cm$^{-1}$. More lines appear in the R and P branches
(R($3_K$) and P($2_K$)).
At 100 K, the Q branch consists of a superposition of 6 lines for $J$ varying
from 1 to 6. The R and P branches are broader with lines of weaker intensities
than those at 10 and 30 K in particular for the maxima.
The modification of the spectra near $\nu_2$ with temperature shows that the spectral
signature of the trapped molecule can be used to sense the environmental
temperature of fullerene C$_{60}$. As the coupling with the vibrational modes
of C$_{60}$ is weak, the signal to noise ratio may be high enough for all the
lines to be observed. In our opinion this result can be extrapolated to other
types of simple molecules trapped in C$_{60}$.
Finally, the dynamical coupling of the vibration-rotation states of the NH$_3$
molecule with the vibrational modes of fullerene C$_{60}$ involves well separated
states for the latter. As a result, this coupling is weak and the energy from
local translational modes cannot redistribute itself in the thermal bath of C$_{60}$.
The linewidths are then negligible.
In conclusion, this work shows that the simulation of the spectra of NH$_3$ can
be used to probe the temperature of the media in which fullerene is observed.
This result can probably be extended to other molecules trapped in fullerene, in
particular when the coupling with the vibration modes of fullerene is weak, because
signal to noise ratio should then be high enough for the rotation lines to be observed.
\newpage
\textbf{ACKNOWLEDGEMENTS}
The authors thank the Regional Councels of Franche-Comt\'{e} and Ile de France,
the General Councel of the 78$^{\text{th}}$ district of Yvelines, the DGCIS
from Ministry of Industry in France, the MOVEO cluster for financing the FUI
project MEMOIRE under the label of cluster MOVEO, CNRS for financing support under
the Interdisciplinary Research Program of the INSU \textquotedblleft Planetary
Environments and Life Origins (EPOV)\textquotedblright, and also Camille Lakhlifi
for figures help.
|
\section{Introduction and Main Results} \label{S1}
We introduce a new elementary method for the study of certain solutions to reaction-diffusion equations with {\it Kolmogorov-Petrovskii-Piskunov (KPP) type} non-linearities. We use it to prove existence of transition front solutions for very general spatially inhomogeneous KPP reaction-diffusion equations in one dimension as well as some special ones in several dimensions, and to obtain very good estimates on these solutions. Our method is based on relating the solutions of the original non-linear equation to those of its linearization at $u=0$.
Let us first consider the reaction-diffusion equation
\begin{equation} \label{1.1}
u_t=u_{xx} + f(x,u)
\end{equation}
with $x\in {\mathbb{R}}$ and $f$ an {\it inhomogeneous KPP reaction function}.
That is, we assume that $f$ is Lipschitz, $a(x)\equiv f_u(x,0)>0$ exists,
\begin{equation} \label{1.2}
f(x,0)=f(x,1)=0 \qquad\text{and} \qquad a(x)g(u)\le f(x,u)\le a(x)u \quad \text{for $(x,u)\in{\mathbb{R}}\times[0,1]$,}
\end{equation}
where $g\in C^1([0,1])$ is such that
\begin{equation} \label{1.3}
g(0)=g(1)=0, \qquad g'(0)=1, \qquad \text{and} \qquad 0<g(u)\le u \quad \text{for $u\in(0,1)$.}
\end{equation}
We will also assume
\begin{equation} \label{1.3a}
\int_0^1 \frac{u-g(u)}{u^2} du<\infty \qquad\text{and}\qquad g'(u)\le 1 \quad \text{for $u\in(0,1)$.}
\end{equation}
We define $a_-\equiv \inf_{x\in{\mathbb{R}}} a(x)\ge 0$ and also assume existence of $a_+<\infty$ such that
\begin{equation} \label{1.4}
a(x)\le a_+ \qquad \text{for $x\in{\mathbb{R}}$}.
\end{equation}
A (right-moving)
{\it transition front} for \eqref{1.1} is an {\it entire} (global-in-time) solution $0\le u\le 1$ connecting 0 and 1 in the sense of
\begin{equation} \label{1.6}
\lim_{x\to-\infty}u(t,x)=1 \qquad\text{and}\qquad \lim_{x\to+\infty}u(t,x)=0
\end{equation}
for each $t\in{\mathbb{R}}$. It models an {\it invasion} of the unstable state $u\equiv 0$ by the asymptotically stable state $u\equiv 1$. Moreover, we also require that for any $\varepsilon>0$ there is $L_\varepsilon<\infty$ such that
\begin{equation} \label{1.7}
\sup_{t\in{\mathbb{R}}} {\rm diam}\left\{x\in{\mathbb{R}} \,|\,\varepsilon\le u(t,x)\le 1-\varepsilon\right\} \le L_\varepsilon,
\end{equation}
that is, the width of the transition region between $\varepsilon$ and $1-\varepsilon$ is uniformly bounded in time. This definition of transition fronts has first appeared in \cite{BH2,Matano}.
It has been well known since the seminal works of Fisher \cite{Fisher} and Kolmogorov-Petrovskii-Piskunov \cite{KPP} that in the homogeneous case $f(x,u)=f(u)$, there exist transition fronts where constant-in-time speed and profile. More specifically, \eqref{1.1} has solutions of the form $u(t,x)=U(x-ct)$ with $U(-\infty)=1$ and $U(\infty)=0$ precisely when the front speed $c\ge c^*_f$, with $c^*_f\equiv 2\sqrt{f'(0)}$ is the {\it minimal front speed}. These fronts have a constant-in-time profile $U$ with $U'<0$, are unique for each $c$ up to a translation, and are usually called {\it traveling fronts}. There are also other transition fronts in this case \cite{HN1}, which are obtained as a combination of two or more traveling fronts with different speeds (we will discuss this in more detail below). Later, existence of KPP transition fronts with time-periodic profiles (called {\it pulsating fronts}) was proved for $x$-periodic reactions $f$, again for all speeds $c\ge c^*_f$ with some $c^*_f>0$ \cite{BH}.
Very recently, existence of transition fronts was first time proved for some non-periodic inhomogeneous KPP reactions \cite{NRRZ} (see \cite{MNRR,MRS,NolRyz,ZlaGenfronts} for results on ignition reactions, and \cite{ZlaGenfronts} for results on some non-KPP non-negative reactions). Specifically, if $a_->0$ and $a(x)-a_-$ is {\it compactly supported}, then transition fronts exist when $ \lambda_0\equiv \sup \sigma[\partial^2_{xx}+a(x)]$, the supremum of the spectrum of the operator $\partial^2_{xx}+a(x)$, satisfies $ \lambda_0<2a_-$ (note that always $\lambda_0\ge a_-$). These fronts do not have a constant profile but for each $c\in(2\sqrt{a_-}, \lambda_0 (\lambda_0 -a_-)^{-1/2})$ there is a front which has a {\it mean speed}
\begin{equation} \label{1.7a}
\lim_{|t-s|\to\infty} \frac{X(t)-X(s)}{t-s}
\end{equation}
equal to $c$, where $X(t)$ is the rightmost point such that $u(t,X(t))=\tfrac 12$. Moreover, no transition fronts exist when, in addition, $a(x)\ge a_-$ and $ \lambda_0>2a_-$ \cite{NRRZ}; this is the first {\it non-existence-of-fronts} result.
We consider here the question of existence of transition fronts in general inhomogeneous media without the assumption of compact support of $a(x)-a_-$ (in which case no constant or mean speed fronts exist in general) and answer it in the affirmative again when $ \lambda_0<2a_-$. We achieve this by using a new and elementary method which exploits the close connection between the equation \eqref{1.1} and its linearization
\begin{equation} \label{1.8}
v_t=v_{xx} + a(x)v
\end{equation}
at $u=0$.
Such a connection is well known, in particular, when $f(x,u)=f(u)$ and so $a(x)\equiv a=f'(0)$ is constant. Then \eqref{1.8} has traveling-front-like solutions $e^{-\gamma(x-c_{a,\gamma}t)}$ with $\gamma>0$ and speed $c_{a,\gamma}\equiv \gamma+a\gamma^{-1}\ge 2 \sqrt a=c_f^*$.
It turns out \cite{Uchi} that if $c>2 \sqrt a$ and $\gamma<\sqrt a$ is such that $c=c_{a,\gamma}$, then the traveling front for \eqref{1.1} with speed $c$ also has asymptotic decay $e^{-\gamma(x-c_{a,\gamma}t)}$ as $x\to\infty$, while for $c=2 \sqrt a$, the asymptotic decay is $(x-2\sqrt a\, t)e^{-\sqrt a(x-2\sqrt a\, t)}$ as $x\to\infty$ (fronts for \eqref{1.8} with $\gamma>\sqrt a$ do not give rise to fronts for \eqref{1.1}).
This means that if $U_{f,\gamma}$ is a traveling front profile for \eqref{1.1} corresponding to speed $c_{a,\gamma} \ge c_f^*$ with $\gamma\le \sqrt a$, and the function $h:[0,\infty)\to[0,1)$ is given by $U_{f,\gamma}(x)=h(e^{-\gamma x})$ (so that $h(0)=0$ and $\lim_{v\to\infty} h(v)=1$), then $h'(0)=1$ when $\gamma< \sqrt a$ and $\lim_{v\to 0}h(v)(-v\ln v)^{-1}=1$ when $\gamma= \sqrt a$, after an appropriate translation of $U_{f,\gamma}$ in $x$.
The above shows that for $f(x,u)=f(u)$ and for faster-than-minimal speed $c>c^*_f$, the ``tails'' of the corresponding traveling fronts for \eqref{1.1} and \eqref{1.8} are asymptotically the same. We will show that this still holds for some transition fronts in general inhomogeneous media when $ \lambda_0<2a_-$. We will in fact show that {\it the study of these fronts for \eqref{1.1} is essentially equivalent to the study of the corresponding front-like solutions for the simpler equation \eqref{1.8}}.
Similarly to the compactly supported $a(x)-a_-$ setting in \cite{NRRZ}, examples of the latter can be found in the form $v_{\lambda}(t,x)\equiv e^{\lambda t}\phi_\lambda(x)$, where $\phi_\lambda(x)>0$ is a
solution of the Schr\" odinger generalized eigenfunction equation
\
\phi_\lambda''+a(x)\phi_\lambda=\lambda\phi_\lambda,
\]
with $\lim_{x\to\infty} \phi_\lambda(x)=0$ and $\phi_\lambda(0)=1$. Notice that if $a$ is constant, then $v_\lambda(t,x)=e^{\lambda t-\sqrt{\lambda-a}\,x}=e^{-\gamma(x-c_{a,\gamma}t)}$ with $\gamma\equiv \sqrt{\lambda-a}$.
Sturm oscillation theory shows that such $\phi_\lambda>0$ exists and is unique precisely when $\lambda> \lambda_0$. Moreover, $\phi_\lambda$ grows exponentially as $x\to -\infty$ (see \eqref{2.8}). Then $v_{\lambda}$ is a super-solution of \eqref{1.1} and we will show that for any $\lambda\in(\lambda_0,2a_-)$ there is $h:[0,\infty)\to[0,1)$ such that $w_{\lambda}(t,x)\equiv h(v_{\lambda}(t,x))$ is a sub-solution (rather than an outright solution, as in the homogeneous case). Moreover, $\lambda<2a_-$ will ensure $h(v)\le v$ so it will follow that there exists a transition front $u\in[w_{\lambda}, v_{\lambda}]$ for \eqref{1.1}. We note that this construction cannot be expected to work for $\lambda\ge 2a_-$ in general because in the homogeneous case this translates to $\gamma\ge\sqrt a$, which either gives rise to no front for \eqref{1.1} when $\gamma>\sqrt a$ or violates $h(v)\le v$ when $\gamma = \sqrt a$.
There is, in fact, a larger class of positive entire solutions of \eqref{1.8}, of which the $v_\lambda$ are the extremal points. Indeed, if $\mu$ is a finite non-negative non-zero Borel measure on $(\lambda_0,\infty)$ with a bounded support, then Harnack inequality shows that
\begin{equation} \label{1.8b}
v_\mu(t,x)\equiv \int_{\mathbb{R}} v_{\lambda}(t,x) d\mu(\lambda) = \int_{\mathbb{R}} e^{\lambda t} \phi_{\lambda}(x) d\mu(\lambda)
\end{equation}
is well-defined, and it is obiously an entire solution of \eqref{1.8}. We will show that $v_\mu$ also gives rise to an entire solution of \eqref{1.1} provided $ \sup \text{\rm{supp}}(\mu)<2a_-$.
Finally, our result extends to and will be stated for the more general PDEs
\begin{equation} \label{1.9}
u_t=(B(x)u_{x})_{x} + q(x) u_x + f(x,u)
\end{equation}
and
\begin{equation} \label{1.10}
v_t=(B(x)v_{x})_{x} +q(x) v_x + a(x)v
\end{equation}
with $B,q$ Lipschitz and satisfying
\begin{equation} \label{1.11}
0<B_-\le B(x)\le B_+<\infty \qquad \text{and} \qquad |q(x)|\le q_+<\infty \quad \text{\,\,for $x\in{\mathbb{R}}$}.
\end{equation}
Let us define
\begin{equation} \label{1.11a}
\lambda_0\equiv \sup_{\psi \in H^1({\mathbb{R}})} \frac{\int_{\mathbb{R}} [ - B(x)\psi'(x)^2 + q(x)\psi'(x)\psi(x) + a(x)\psi(x)^2] dx}{\int_{\mathbb{R}} \psi(x)^2 dx} \quad(\ge a_-).
\end{equation}
Note that when $q\equiv 0$, then the Rayleigh quotient formula for self-adjoint operators gives
\[
\lambda_0 = \sup \sigma \left[ \partial_x(B(x)\partial_{x})+a(x) \right] .
\]
As we show below, for $\lambda> \lambda_0$ there is again a unique $\phi_\lambda>0$ such that
\begin{equation} \label{1.12}
(B(x)\phi_\lambda')' + q(x)\phi_\lambda' +a(x)\phi_\lambda=\lambda\phi_\lambda,
\end{equation}
$\lim_{x\to\infty} \phi_\lambda(x)=0$ and $\phi_\lambda(0)=1$.
\begin{theorem} \label{T.1.1}
Assume \eqref{1.2}--\eqref{1.4} and \eqref{1.11}, let $ \lambda_0$
be as in \eqref{1.11a} and for $\lambda> \lambda_0$ let $\phi_\lambda$ be as in \eqref{1.12}.
Let $(aB)_-\equiv \inf_{x\in{\mathbb{R}}} [a(x)B(x)]$, and assume also that $q_+\le 2\sqrt{(aB)_-}$ and
\begin{equation} \label{1.13}
\lambda_0< \lambda_1\equiv \inf_{x\in{\mathbb{R}}} \left\{ a(x) + \sqrt{(aB)_-} \left[ \sqrt{(aB)_-} - |q(x)| \right] B(x)^{-1} \right\}.
\end{equation}
Let $\mu$ be a finite non-negative non-zero Borel measure on $(\lambda_0,\lambda_1)$ with $\mu_0\equiv \inf \text{\rm{supp}}(\mu)$ and $\mu_1\equiv \sup \text{\rm{supp}}(\mu)$, and define $v_\mu$ as in \eqref{1.8b}.
(i) If $\mu_1<\lambda_1$, then there is an increasing function $h:[0,\infty)\to[0,1)$ with $h(0)=0$, $h'(0)=1$, $\lim_{v\to\infty} h(v)=1$, and an entire solution $u_\mu$ of \eqref{1.9} satisfying \eqref{1.6}, $(u_\mu)_t>0$,
\begin{equation} \label{1.14}
h \left( v_\mu \right) \le u_\mu \le \min \left\{ v_\mu,1 \right\}.
\end{equation}
In fact, we can choose $h= h_{g,\alpha}$ from \eqref{2.1} below, with any $\alpha\in ( 1-(\lambda_1-\mu_1)a_+^{-1},1)$.
(ii) If $\lambda_0<\mu_0\le \mu_1<\lambda_1$, then $u_\mu$ from (i) is a transition front (i.e., satisfying also \eqref{1.7}), with $L_\varepsilon$ depending only on $g,a_+,B_\pm,\varepsilon$ and $\zeta$, provided $\min\{\mu_0-\lambda_0, \lambda_1-\mu_1\} \ge\zeta>0$.
\end{theorem}
{\it Remarks.} 1. Condition \eqref{1.13} is sharp in this generality, as exhibited by the previously mentioned non-existence of transition fronts in the case of $B\equiv 1$, $q\equiv 0$, and compactly supported $a(x)-a_-$ with $a(x)\ge a_->0$ and $\lambda_0>2a_-$ \cite{NRRZ}.
\smallskip
2. The properties of $h$ give $\lim_{x\to\infty} u_\mu(t,x) v_\mu(t,x)^{-1}=1$ for each $t\in{\mathbb{R}}$.
\smallskip
3. Note that $a_- + \sqrt{(aB)_-} [ \sqrt{(aB)_-} - q_+ ] B_+^{-1} \le \lambda_1 \le 2a_-$, so \eqref{1.13} is satisfied when $ \lambda_0< a_- + \sqrt{(aB)_-} [ \sqrt{(aB)_-} - q_+ ] B_+^{-1}$.
In the case $B\equiv 1$ and $q\equiv 0$ we have $\lambda_1=2a_-$, so \eqref{1.13} simplifies to $ \lambda_0<2a_-$, the condition mentioned above.
\smallskip
4. Of course, an identical result holds for solutions moving to the left, with $\psi_\lambda$ defined as $\phi_\lambda$ but satisfying instead $\lim_{x\to -\infty} \psi_\lambda(x)=0$. In addition, a combination of two solutions of \eqref{1.10} from (i), moving in opposite directions, gives an entire solution of \eqref{1.9} whose spatial infimum converges to 1 as $t\to\infty$.
\smallskip
5. The borderline case $\mu=\delta_{\lambda_1}$, which corresponds to the traveling front with the minimal speed $c_f^*$ and maximal decay $\sim e^{-\sqrt{f'(0)}x}$ when $f(x,u)=f(u)$, is not covered by our result (because then $\alpha=1$ in Lemma \ref{L.2.1} below). It is an open question whether a {\it maximal decay transition front} exists in the inhomogeneous setting.
\smallskip
6. The nonlinearity $f$ can in addition depend on time, as long as $f_u(t,x,0)$ is time independent. This is also the case for the other results in this paper.
\smallskip
7. Finally, we note that all our results continue to hold if in \eqref{1.2} one does not necessarily require $f(x,1)=0$. In that case we drop the lower bound on $f$ in \eqref{1.2} for $u>1$, consider solutions $u\ge 0$ (rather than $0\le u\le 1$) not necessarily converging to 1 as $x\to-\infty$, and the upper bound in \eqref{1.14} becomes just $u_\mu(t,x) \le v_\mu(t,x)$.
\smallskip
Although the ``extremal'' fronts $v_{\delta_\lambda}(=v_\lambda)$ have a constant speed in homogeneous media, one cannot expect them to have a constant or even a mean speed in general. However, if the medium is random and stationary ergodic, they do have (almost surely) a deterministic {\it aymptotic speed}
\begin{equation} \label{1.15}
c\equiv \lim_{|t|\to\infty} \frac{X(t)}{t} >0.
\end{equation}
with $X(t)$ as in \eqref{1.7a}.
\begin{theorem} \label{T.1.2}
Consider a probability space $(\Omega,\mathcal{F},{\mathbb{P}})$ and assume that a measurable function $p\equiv (a,B,q):\Omega\to L^\infty_{\rm loc}({\mathbb{R}})^3$ is Lipschitz in $x$ and satisfies \eqref{1.4} and \eqref{1.11}, uniformly in $\omega\in\Omega$. In addition, assume that $p$ is stationary ergodic. That is, there is a group $\{\pi_y\}_{y\in {\mathbb{R}}}$ of measure preserving transformations acting ergodically on $\Omega$ such that $p({\pi_y\omega};x)=p(\omega;x+y)$. Then $\lambda_0,\lambda_1$ from Theorem \ref{T.1.1} are constant in $\omega$, except on a measure zero set. If $\lambda_0<\lambda_1$ and a reaction $f(\omega;x,u)$ satisfies \eqref{1.2}--\eqref{1.3a} for almost all $\omega\in\Omega$, then for each $\lambda\in(\lambda_0,\lambda_1)$ there is $c_\lambda>0$ such that the transition front $u_{\delta_\lambda}({\omega};t,x)$ from Theorem~\ref{T.1.1}(ii) has asymptotic speed $c_\lambda$ in the sense of \eqref{1.15} for almost all $\omega\in\Omega$.
\end{theorem}
{\it Remarks.} 1. Notice that $f$ itself need not be stationary ergodic.
\smallskip
2. If $B\equiv 1$ and $q\equiv 0$, the condition $\lambda_0<\lambda_1$ again becomes $\lambda_0<2a_-$, which is guaranteed, for instance, when $a_+<2a_-$, regardless of the structure of the randomness.
\smallskip
3. It is conceivable that, in general, transition fronts exist {\it almost surely} even if $\lambda_0\ge 2a_-$. We do not know the answer to this question at this time and pose it as an open problem.
\smallskip
We also provide applications of our method in several spatial dimensions, to the study of solutions of the reaction-diffusion equation
\begin{equation} \label{2.4a}
u_t = \nabla\cdot(B(x)\nabla u) + q(x)\cdot\nabla u + f(x,u)
\end{equation}
on ${\mathbb{R}}\times{\mathbb{R}}^d$, where $f,B,q$ are again as above but with $B$ a matrix field and $q$ a vector field.
Let us start with the special case
\begin{equation} \label{1.16}
u_t=\Delta u + f(x,u)
\end{equation}
with $f_u(x,0)\equiv\aaa >0$ independent of $x$.
The corresponding linear PDE
\begin{equation} \label{1.17}
v_t=\Delta v + \aaa v
\end{equation}
has ``extremal'' solutions $v_{0} (t,x)\equiv e^{\aaa t}$ and
\
v_{\gamma\eta} (t,x)\equiv e^{ -\gamma\eta\cdot x + (\gamma^2+\aaa )t } = e^{-\gamma(x\cdot\eta - c_{\aaa ,\gamma} t)},
\]
with $\gamma>0$, $\eta\in{\mathbb{R}}^d$ a unit vector, and as before,
\[
c_{\aaa ,\gamma} = \gamma+ \aaa \gamma^{-1} \ge 2\sqrt{\aaa }.
\]
From the one-dimensional case mentioned above it immediately follows that each traveling front for \eqref{1.16} of the form $u(t,x)=U(x\cdot\eta-ct)$ has the same decay (as $x\cdot\eta\to\infty$) as a multiple of $v_{\gamma\eta}$ for some $\gamma\in(0, \sqrt{\aaa }]$ (with an extra factor $x\cdot\eta - 2\sqrt{\aaa}\, t$ if $\gamma=\sqrt{\aaa }\,$), and then $c=c_{\aaa ,\gamma}$.
Both $u$ and $v_{\gamma\eta}$ travel with speed $c_{\aaa ,\gamma}$ in the direction $\eta$.
We will therefore only consider $\gamma\le\sqrt{\aaa }$ and let $Y\equiv \overline{B}(0,\sqrt{\aaa })$ be the closed ball in ${\mathbb{R}}^d$ with radius $\sqrt{\aaa }$ and centered at 0, with topology inherited from ${\mathbb{R}}^d$. If $\mu$ is a finite non-negative non-zero Borel measure on $Y$, then we let
\begin{equation}\label{1.17a}
v_\mu(t,x)\equiv \int_Y v_{\xi}(t,x) d\mu(\xi) = \int_Y e^{ -\xi\cdot x + (|\xi|^2+\aaa )t } d\mu(\xi)
\end{equation}
(i.e., $v_{\delta_\xi}=v_\xi$). Notice that $v_\mu(t,x)\le e^{\sqrt{\aaa }|x|+ \aaa(3+\sgn(t)) t/2}$ and it is a positive entire solution of \eqref{1.17}.
Also, $Y$ becomes an analog of $[-\lambda_0,-\lambda_1]\cup [\lambda_0,\lambda_1]$ in Theorem \ref{T.1.1} (the latter set supports measures corresponding to solutions from Remark 4 after Theorem \ref{T.1.1}), after recalling that for homogeneous reactions, $\lambda_0=\aaa $, $\lambda_1=2\aaa $, and $\gamma=\sqrt{\lambda - \aaa }$.
Part (i) of our next result shows that each $v_\mu$ gives rise to an entire solution $u_\mu$ of \eqref{1.16}. Moreover, in parts (ii) and (iii) we address the questions when this solution connects 0 and 1 and when does the transition zone between $\varepsilon$ and $1-\varepsilon$ have a bounded width (in some sense) for each $\varepsilon>0$.
To this end, let us define the {\it convex hull} of a measure $\mu$ on ${\mathbb{R}}^d$ to be
\
\ch(\mu) \equiv \{ \zeta\in{\mathbb{R}}^d \,|\, \zeta={\mathbb{E}}(\nu) \text{ for some measure } 0<\nu\le \mu \},
\]
with ${\mathbb{E}}(\nu)\equiv \nu({\mathbb{R}}^d)^{-1}\int_{{\mathbb{R}}^d} \xi d\nu(\xi)$. Then $\ch(\mu)$ is convex because
\[
{\mathbb{E}}(\beta \nu + (1-\beta)\nu') = [ \beta \nu({\mathbb{R}}^d) + (1-\beta) \nu'({\mathbb{R}}^d) ]^{-1} \left[ \beta \nu({\mathbb{R}}^d) {\mathbb{E}}(\nu) + (1-\beta) \nu'({\mathbb{R}}^d) {\mathbb{E}}(\nu') \right]
\]
but not necessarily closed. We note that $\ch(\mu)$ is also the intersection of convex hulls of all essential supports of $\mu$, that is, sets $A\subset{\mathbb{R}}^d$ such that $\mu(A)=\mu({\mathbb{R}}^d)$ and $\mu(A')<\mu(A)$ whenever $A'\subset A$ and $A\setminus A'$ has a positive Lebesgue measure (see the remark after the proof of Theorem \ref{T.1.3}), although $\ch(\mu)$ itself need not be an essential support of $\mu$ (e.g., if $B\subset{\mathbb{R}}^d$ is an open ball and $\mu$ the uniform measure on the sphere $\partial B$, then $\ch(\mu)=B$).
\begin{theorem} \label{T.1.3}
Assume \eqref{1.2}--\eqref{1.3a} for $x\in{\mathbb{R}}^d$ and with $a(x)\equiv a>0$. Let $\mu$ be a finite non-negative non-zero Borel measure with support in the open ball $B(0,\sqrt{\aaa })$ and let $v_\mu$ be as in \eqref{1.17a}.
(i) There is an increasing function $h:[0,\infty)\to[0,1)$ with $h(0)=0$, $h'(0)=1$ and $\lim_{v\to\infty}h(v)=1$, and an entire solution $u_\mu$ of \eqref{1.16} such that $(u_\mu)_t>0$ and \eqref{1.14} holds.
In fact, we can choose $h= h_{g,\alpha}$ from \eqref{2.1} below,
provided $\mu$ is supported in $\overline{B}(0,\sqrt{\alpha \aaa })$.
Also, $u_\mu \not\equiv u_{\mu'}$ when $\mu\neq\mu'$.
(ii) We have
\begin{equation} \label{1.19a}
\inf_{x\in{\mathbb{R}}^d} u_\mu(x,t) = 0 \qquad\text{and}\qquad \sup_{x\in{\mathbb{R}}^d} u_\mu(x,t) = 1
\end{equation}
for each $t\in{\mathbb{R}}$ (equivalently, for some $t\in{\mathbb{R}}$) if and only if $0\notin \ch(\mu)$.
(iii) If $0\notin \text{\rm{supp}} (\mu)$, then for each $\varepsilon,\theta>0$ there is $L_{\varepsilon,\theta}$ (depending also on $\dist(0,\text{\rm{supp}}(\mu))$, $f$, and $\alpha$ from (i)), such that the following holds. If $u_\mu(t,x)\ge \varepsilon$, then there is a unit vector $\eta_{t,x}\in{\mathbb{R}}^d$ such that $u_\mu(t,x+y)\ge 1-\varepsilon$ whenever $\eta_{t,x}\cdot y|y|^{-1} \ge \theta$ and $|y|\ge L_{\varepsilon,\theta}$.
\end{theorem}
Part (i) of this result is closely related to a result of Hamel-Nadirashvili \cite[Theorem~1.2]{HN2}. Under the additional assumptions of $f$ being independent of $x$, concave in $u$, and \hbox{$f\in C^2([0,1])$,} they prove the existence of an infinite-dimensional manifold of entire solutions of \eqref{1.16}. These solutions are parametrized by measures supported on the 1-point compactification $X$ of \hbox{${\mathbb{R}}^d\setminus B(0,2\sqrt{\aaa })$}, where distance from origin denotes the front speed $c\ge 2\sqrt{\aaa }$ rather than $\gamma\le \sqrt{\aaa }$. The mapping $\gamma\mapsto c_{\aaa ,\gamma}$ yields a natural identification of $Y$ and $X$ (we consider the former a slightly more natural parameter space for our method than the latter), so one could ask what is the relationship of the two sets of entire solutions.
Under the above additional assumptions on $f$, it is also shown in \cite[Theorem 1.4]{HN2} that any entire solution $0<u<1$ which satisfies
\begin{equation} \label{1.20}
\lim_{t\to -\infty} \sup_{|x|<(2\sqrt{\aaa }+\varepsilon)|t|}u(t,x) = 0
\end{equation}
for some $\varepsilon>0$, is from their manifold. This gives a characterization of all entire solutions satisfying \eqref{1.20}. Our $u_\mu$ satisfies \eqref{1.20} with some $\varepsilon(\alpha)>0$ as well as the properties of the solution from \cite[Theorem 1.2]{HN2} corresponding to the measure obtained from $\mu$ under the above-mentioned identification of $Y$ and $X$. Since these
properties uniquely define a solution in the manifold, it follows that for $f\in C^2([0,1])$, independent of $x$, and concave in $u$, the two solutions coincide; and the solutions from Theorem \ref{T.1.3}(i) are all the entire solutions of \eqref{1.16} satisfying \eqref{1.20}.
Moreover, the manifold in \cite[Theorem 1.2]{HN2} also contains solutions corresponding to some measures supported in $X$ but not in its interior (which we do not construct in Theorem~\ref{T.1.3}), namely, those whose restriction to $\partial B(0,2\sqrt{\aaa })$ is a finite sum of Dirac masses.
However, besides proving the existence of this manifold of solutions, \cite{HN2} only obtains certain claims about the $t\to -\infty$ asymptotic behavior of each of them, with better control only for those corresponding to measures $\mu$ which are finite sums of Dirac masses \cite[Theorem~1.1]{HN2}. The contribution of Theorem \ref{T.1.3}(i) is therefore not only in proving the existence of these entire solutions for more general (and even inhomogeneous) KPP reactions, but also in obtaining the explicit estimate \eqref{1.14}, valid for all times and yielding the new results in (ii) and (iii). Moreover, the usage of our method (from Lemma \ref{L.2.1} below) makes the proof immediate and elementary, while the proof of \cite[Theorem~1.2]{HN2} is 30 pages long.
In fact, Theorem \ref{T.1.3} extends to some periodic $(a,B,q)$ ($f$ need not be periodic in $x$ and can even be time-dependent, as mentioned above). Now
\
v_{\xi} (t,x)\equiv e^{ -\xi\cdot x + \kappa_\xi t } \theta_\xi(x),
\]
where $(\theta_\xi,\kappa_\xi)$ is the unique solution of
\begin{equation} \label{2.4e}
\nabla\cdot (B(x)\nabla\theta) + (q(x) - 2 B(x)\xi) \cdot\nabla\theta + [\xi\cdot B(x)\xi - \nabla \cdot (B(x)\xi) - q(x)\cdot\xi + a(x)] \theta = \kappa \theta
\end{equation}
on the unit cell of periodicity ${\mathcal C}$ (satisfying periodic boundary conditions) with $\theta_\xi>0$ and $\int_{T^{d}}\theta_\xi(x)dx=1$. Again
\begin{equation}\label{2.4d}
v_\mu(t,x)\equiv \int_Y v_{\xi}(t,x) d\mu(\xi)
\end{equation}
solves
\
v_t = \nabla\cdot(B(x)\nabla v) + q(x)\cdot\nabla v + a(x)v
\]
when $\mu$ is as above. Finally, let $S_\alpha$ be the set
of all $\xi\in{\mathbb{R}}^d$ such that
\begin{equation}\label{2.4f}
\left\| \left( \frac{\nabla\theta_\xi}{\theta_\xi} -\xi \right) \cdot \frac Ba \left( \frac{\nabla\theta_\xi}{\theta_\xi} -\xi \right) \right\|_{L^\infty({\mathcal C})} \le \alpha.
\end{equation}
\begin{theorem} \label{T.1.5}
Assume \eqref{1.2}--\eqref{1.3a} for $x\in{\mathbb{R}}^d$ and with $(a,B,q)$ periodic. Let $\mu$ be a finite non-negative non-zero Borel measure supported on $S_\alpha$ for some $\alpha<1$, and let $v_\mu$ be as in \eqref{2.4d}.
Then Theorem \ref{T.1.3}(i)--(iii) hold with $h=h_{g,\alpha}$ from \eqref{2.1} below, except possibly the last statement in (i).
\end{theorem}
{\it Remark.} We note that in general, all $S_\alpha$ for $\alpha<1$ may be empty. However this is not the case when $B-I$ is small in $C^{1,\delta}({\mathbb{T}}^{d})$ and $a-\bar a,q$ (with $\bar a\equiv \int_{{\mathbb{T}}^{d}} a(x)dx$) are small in $C^\delta({\mathbb{T}}^{d})$ for some $\delta>0$. Indeed, in that case we obtain a uniform (in norms of $B-I,a-\bar a,q$ in the respective spaces) bound on $\theta_\xi$ in $C^{2,\delta}({\mathbb{T}}^d)$ for all $|\xi|\le 1$. If now $(a-\bar a,B-I,q)\in C^{1,\delta}\times C^{\delta}\times C^{\delta}$ is small enough, then $\kappa_\xi-|\xi|^2-\bar a$ is also small, so $a(x)+|\xi|^2-\kappa_\xi$ is small in $C^\delta$ and \eqref{2.4e} can be rewritten as
\begin{align*}
\Delta\theta_\xi +2\xi\cdot\nabla\theta_\xi
= & - \nabla\cdot [(B(x)-I)\nabla\theta_\xi] - [q(x) - 2 (B(x)-I)\xi] \cdot\nabla\theta_\xi \\
& - [\xi\cdot (B(x)-I)\xi - \nabla \cdot (B(x)\xi) - q(x)\cdot\xi + a(x)+|\xi|^2-\kappa_\xi] \theta_\xi ,
\end{align*}
with the right-hand side uniformly small in $C^\delta$ for all $|\xi|\le 1$. Thus $\theta_\xi- \int_{T^{d-1}}\theta_\xi(x)dx = \theta_\xi-1$ is uniformly small in $C^{2,\delta}$. This means that for each $\beta<1$, \eqref{2.4f} holds for $\alpha\equiv \tfrac 12(1+\beta)$ and all $|\xi|\le\beta$ provided $(a-\bar a,B-I,q)$ is sufficiently small in $C^{1,\delta}\times C^{\delta}\times C^{\delta}$.
\smallskip
We end this introduction with an application of our method to obtaining explicit bounds on certain solutions $u$ of \eqref{1.16} with constant $f_u(x,u)=a$, in terms of the solutions of the {\it heat equation} $\tilde u_t=\Delta \tilde u$ with the same initial condition (in which case $\tilde u\le u \le e^{\aaa t}\tilde u $). Of course, the latter is just
\begin{equation}\label{1.20a}
\tilde u(t,x)=(4\pi t)^{-d/2} \int_{{\mathbb{R}}^d} e^{-|x-y|^2/4t} u(0,y) dy.
\end{equation}
\begin{theorem} \label{T.1.4}
Assume \eqref{1.2}--\eqref{1.3a} for $x\in{\mathbb{R}}^d$ and with $a(x)\equiv a>0$. Let $0\le u\le 1$ solve \eqref{1.16} on ${\mathbb{R}}^+\times{\mathbb{R}}^d$. If $\tilde u$ from \eqref{1.20a} satisfies
\begin{equation} \label{1.21}
|\nabla \tilde u(t_0,x)| \le \sqrt{\alpha \aaa }\, \tilde u(t_0,x)
\end{equation}
for some $t_0\ge 0$, $\alpha<1$, and all $x\in{\mathbb{R}}^d$, then
\begin{equation} \label{1.22}
h_{g,\alpha} \left( e^{\aaa (t-t_0)} \tilde u(t,x) \right) \le u(t,x) \le \min\{e^{\aaa t} \tilde u(t,x), 1\}
\end{equation}
for all $(t,x)\in {\mathbb{R}}^+\times{\mathbb{R}}^d$, with $h_{g,\alpha}$ from \eqref{2.1} below (in particular, $h_{g,\alpha}'(0)=1=h_{g,\alpha}(\infty)$).
\end{theorem}
We prove Theorems \ref{T.1.1}--\ref{T.1.4} in the next section, after introducing our main tool, Lemma~\ref{L.2.1}.
Finally, we note that existence of transition fronts for \eqref{1.1} with very general $f$ (including KPP) is claimed in the paper \cite{SLL}. This statement is false in the full generality claimed there (in particular, it contradicts the non-existence result in \cite{NRRZ}), and its proof is also incorrect. The latter is a direct adaptation of the existence-of-fronts proof for ignition reactions from \cite{MRS} which, however, does not extend to non-ignition reactions. In particular, various claims in \cite{SLL}, such as the one between (2.22) and (2.23), Corollary 2.6(i), and Proposition 2.7, are made without a proof and are, in fact, false for general non-ignition reactions.
The author would like to thank Fran\c cois Hamel for pointing out the argument in the remark after Theorem \ref{T.1.5}. He also acknowledges partial support by NSF grants DMS-1113017 and DMS-1056327, and by an Alfred P. Sloan Research Fellowship.
\section{The Key Lemma and the Proofs of Theorems \ref{T.1.1}--\ref{T.1.4}} \label{S2}
Our main tool is the following lemma, which constructs sub-solutions $w=h(v)$ of \eqref{1.9} from certain solutions $v$ of \eqref{1.10} (which are also super-solutions of \eqref{1.9}). Here the function $h=h_{g,\alpha}:[0,\infty)\to[0,1)$ depends on $g\in C^1([0,1])$ satisfying \eqref{1.3}, \eqref{1.3a} and also on an additional parameter $\alpha\le 1$. Specifically, $h_{g,\alpha}(0)=0$ and
\begin{equation} \label{2.1}
h_{g,\alpha}(v) \equiv U_{g,\sqrt{\alpha}}(-\alpha^{-1/2} \ln v)
\end{equation}
for $v>0$, where $U_{g,\sqrt{\alpha}}$ is the traveling front profile for the homogeneous PDE
\begin{equation} \label{A.1}
u_t=u_{xx} + g(u)
\end{equation}
corresponding to speed $c_{1,\sqrt{\alpha}}\equiv \alpha^{1/2} + \alpha^{-1/2}\ge 2$. That is, $U_{g,\sqrt{\alpha}}(-\infty)=1$, $U_{g,\sqrt{\alpha}}(\infty)=0$, $U_{g,\sqrt{\alpha}}'<0$, and
\begin{equation} \label{A.1a}
U''_{g,\sqrt{\alpha}}+c_{1,\sqrt{\alpha}} U'_{g,\sqrt{\alpha}}+g(U_{g,\sqrt{\alpha}})=0
\end{equation}
on ${\mathbb{R}}$. Notice that the $\lim_{v\to\infty} h_{g,\alpha}(v)=1$ and \eqref{A.1a} implies
\begin{equation} \label{2.5d}
\alpha v^2 h_{g,\alpha}''(v) - vh_{g,\alpha}'(v) + g(h_{g,\alpha}(v))=0.
\end{equation}
It is well known that $U_{g,\sqrt{\alpha}}$ is unique up to translation and if $\alpha<1$, then there is a unique translation such that $\lim_{x\to\infty} U_{g,\sqrt\alpha}(x) e^{\sqrt\alpha x}=1$ \cite{Uchi}. With this choice of $U_{g,\sqrt\alpha}$ we obtain $h_{g,\alpha}'(0)=1$ for $\alpha<1$. It then also follows that
\begin{equation}\label{2.6c}
h_{g,\alpha}(v)\le v
\end{equation}
for $v\in[0,\infty)$ because $h_{g,\alpha}''<0$ (see the proof of Lemma \ref{L.2.1} below).
For $\alpha=1$ we instead have $\lim_{v\to 0} h_{g,\alpha}(v)(-v\ln v)^{-1}=1$, provided the first condition in \eqref{1.3a} is replaced by $\int_0^1 [u-g(u)] |\ln u| u^{-2} du<\infty$ \cite{Uchi}.
We state the lemma in a more general form, with time-dependent coefficients.
\begin{lemma} \label{L.2.1}
With $f,a,B,q$ Lipschitz and time-dependent ($B$ a matrix and $q$ a vector field), assume \eqref{1.2}--\eqref{1.4} and \eqref{1.11} for $(t,x)\in(t_0,t_1)\times{\mathbb{R}}^d$ (where $a(t,x)\equiv f_u(t,x,u)$ and \hbox{$-\infty< t_0<t_1\le\infty$}). Let $v>0$ be a solution of
\
v_t = \nabla\cdot(B(t,x)\nabla v) + q(t,x)\cdot\nabla v + a(t,x) v
\]
on $(t_0,t_1)\times{\mathbb{R}}^d$.
If for some $\alpha<1$,
\begin{equation} \label{2.3}
\nabla v(t,x) \cdot B(t,x) \nabla v(t,x) \le \alpha a(t,x) v(t,x)^2
\end{equation}
holds for all $(t,x)\in (t_0,t_1)\times{\mathbb{R}}^d$, then $v$ and $w\equiv h_{g,\alpha}(v)$ are a super- and sub-solution of
\begin{equation} \label{2.4}
u_t = \nabla\cdot(B(t,x)\nabla u) + q(t,x)\cdot\nabla u + f(t,x,u)
\end{equation}
on $(t_0,t_1)\times{\mathbb{R}}^d$.
Therefore, if $0\le u\le 1$ solves \eqref{2.4} with $w(t_0,x)\le u(t_0,x)\le v(t_0,x)$ for all $x\in{\mathbb{R}}^d$, then for all $(t,x)\in (t_0,t_1)\times{\mathbb{R}}^d$ we have
\begin{equation} \label{2.5}
w(t,x)\le u(t,x)\le \min\{v(t,x),1\}.
\end{equation}
\end{lemma}
{\it Remark.} Of course, the crucial hypothesis here is \eqref{2.3}.
\begin{proof}
Obviously $v$ is a super-solution of \eqref{2.4}, giving the second inequality. We also have
\begin{align*}
w_t - \nabla\cdot(B\nabla w) - q\cdot\nabla w & = h'(v) [v_t - \nabla\cdot(B\nabla v) - q\cdot\nabla v] - h''(v) \nabla v \cdot B \nabla v
\\ & =h'(v) a v - h''(v) \nabla v \cdot B \nabla v
\\ & \le a [ h'(v) v - \alpha h''(v) v^2 ].
\end{align*}
In the last inequality we used \eqref{2.3} and $h''<0$. The latter is due to \eqref{2.5d} and Lemma \ref{L.A.1} from the Appendix with $\gamma\equiv \sqrt\alpha$, which yield
\[
\alpha v^2 h''(v) = vh'(v) - g(h(v)) = -\alpha^{-1/2} U_{g,\sqrt\alpha}'(-\alpha^{-1/2} \ln v) - g(U_{g,\sqrt\alpha}(-\alpha^{-1/2} \ln v))<0.
\]
Thus \eqref{2.5d} and \eqref{1.2} give
\[
w_t - \nabla\cdot(B(t,x)\nabla w) - q(t,x)\cdot\nabla w \le a(t,x)g(h(v)) \le f(t,x,w),
\]
so $w$ is a sub-solution of \eqref{1.9}, and the first inequality in \eqref{2.5} follows as well.
\end{proof}
\begin{proof}[Proof of Theorem \ref{T.1.4}]
The comparison principle, together with \eqref{1.2} yields the upper bound, as well as $\tilde u\le u$. Then let $v(t,x)\equiv e^{\aaa (t-t_0)}\tilde u(t,x)$ and note that $r\equiv\nabla v v^{-1} = \nabla\tilde u\tilde u^{-1}$ satisfies
\[
r_t=\Delta r + \nabla(|r|^2)
\]
because
\[
(\ln \tilde u)_t = {\Delta \tilde u} {\tilde u^{-1}} = \Delta(\ln \tilde u) + |r|^2.
\]
Thus $\rho\equiv |r|^2$ satisfies
\[
\rho_t=\Delta \rho +2r\cdot \nabla \rho -2|\nabla r|^2,
\]
so \eqref{1.21} and the maximum principle give $\rho(t,x) \le \alpha \aaa $ for $(t,x)\in(t_0,\infty)\times{\mathbb{R}}^d$. Then Lemma~\ref{L.2.1} yields the lower bound in \eqref{1.22}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{T.1.1}]
(i) Let us start with the proof of existence of $\phi_\lambda$ from \eqref{1.12}, for $\lambda>\lambda_0$. With ${\mathcal L}$ the operator on the left-hand side of \eqref{1.12} and $\lambda_0$ from \eqref{1.11a}, we have
\
\int_{\mathbb{R}} \psi(x)[(\lambda-{\mathcal L})\psi](x) dx \ge (\lambda-\lambda_0) \int_{\mathbb{R}} \psi(x)^2dx
\]
for $\psi\in H^2({\mathbb{R}})$, after integrating by parts.
Thus $(\lambda-{\mathcal L})^{-1}:L^2({\mathbb{R}})\to H^2({\mathbb{R}})$ exists and if $0\not\equiv \psi\in L^2({\mathbb{R}})$ is compactly supported in ${\mathbb{R}}^-$, then $0\not\equiv \phi\equiv (\lambda-{\mathcal L})^{-1}\psi\in H^2({\mathbb{R}})$. Since $\phi$ also satisfies \eqref{1.12} on ${\mathbb{R}}^+$, Harnack inequality shows that $\lim_{x\to\infty} \phi(x)=0$. Let $\tilde \phi(x)\equiv \phi(x)$ for $x\ge 0$ and extend it onto ${\mathbb{R}}^-$ so that it solves \eqref{1.12}. Then $\tilde \phi$ has no roots because if $\tilde\phi(x_0)=0$, then plugging the function $\tilde\phi|_{[x_0,\infty)}$, extended by 0 on $(-\infty,x_0)$, into \eqref{1.11a} would yield $\lambda_0\ge\lambda$. Thus we have $\phi_\lambda(x) = \tilde\phi(x)\tilde\phi(0)^{-1}$. Uniqueness follows from existence of $\psi_\lambda$ with the same properties but with $\lim_{x\to -\infty} \psi_\lambda(x)=0$ (by a reflected argument), from $\lim_{x\to -\infty} \phi_\lambda(x)=\infty$ (by \eqref{2.8} below), and the fact that the space of solutions of \eqref{1.12} is two-dimensional.
Next, choose $\alpha<1$ such that
\
m\equiv \inf_{x\in{\mathbb{R}}, \beta\ge\alpha} \left\{ a(x) + \sqrt{\beta (aB)_-} \left[ \sqrt{\beta (aB)_-} - |q(x)| \right] B(x)^{-1} \right\} - \mu_1>0.
\]
Any $\alpha\in( 1-(\lambda_1-\mu_1)a_+^{-1},1)$ works because the derivative of the expression in the brackets with respect to $\beta$ is bounded above by $(aB)_-B(x)^{-1}\le a_+$ and is positive for $\beta> 1$ (the latter due to $q_+\le 2\sqrt{(aB)_-}$). Now let $w_\mu(t,x)\equiv h_{g,\alpha}(v_\mu(t,x))$ and notice that $w_\mu\le v_\mu$ by \eqref{2.6c}. Then Lemma~\ref{L.2.1} will be applicable to $v_\mu,w_\mu$ once we establish
\begin{equation} \label{2.6}
B(x) \phi_\lambda'(x)^2 \le \alpha a(x) \phi_\lambda(x)^2
\end{equation}
for all $\lambda\in(\lambda_0,\mu_1]$ and $x\in{\mathbb{R}}$. Indeed, \eqref{2.6} and $\phi_\lambda>0$ then yield \eqref{2.3} for $v_\mu$. To this end, we need to show
\begin{equation} \label{2.7}
|\psi(x)| \le \sqrt{\alpha a(x)B(x)}
\end{equation}
for $x\in {\mathbb{R}}$, with $\psi \equiv B\phi_\lambda'/\phi_\lambda$ and $\lambda\in(\lambda_0,\mu_1]$.
Let us assume that $\psi(x_0)\ge \sqrt{\alpha (aB)_-}$ for some $x_0$. We have $\psi'=\lambda-a-\psi(\psi+q)B^{-1}$ on ${\mathbb{R}}$, so $\psi'(x_0)\le \lambda-m-\mu_1\le -m$.
But then $\psi$ must be decreasing on $(-\infty,x_0]$ with $\psi'\le-m$ there.
From this and $\psi'=\lambda-a-(\psi^2+q\psi)B^{-1}$ it follows that $\psi$ must blow up at some $x_1\in (-\infty, x_0)$, a contradiction. We obtain the same conclusion when assuming $\psi(x_0)\le - \sqrt{\alpha (aB)_-}$ (because $\psi'=\lambda-a-|\psi|(|\psi|-q)B^{-1}$ when $\psi<0$), with blowup at some $x_1\in (x_0,\infty)$. It follows that $\| \psi \|_\infty\le \sqrt{\alpha(aB)_-}$, which gives \eqref{2.7}, so Lemma \ref{L.2.1} applies to $v_\mu,w_\mu,\alpha$.
A standard limiting argument (see, for instance, \cite{FM}) now recovers an entire solution to \eqref{1.9} between $\min\{v_\mu,1\}$ and $w_\mu$. Indeed, we let $u_k$ be the solution of \eqref{1.9} on $(-k,\infty)\times{\mathbb{R}}$ with initial datum $u_k(-k,x)\equiv w_\mu(-k,x)$. Then by Lemma \ref{L.2.1} we have
\[
w_\mu(t,x) \le u_k(t,x)\le \min\{v_\mu(t,x),1\}
\]
on $(-k,\infty)\times{\mathbb{R}}$. By parabolic regularity, there is a locally uniform (on ${\mathbb{R}}^2$) limit $u_\mu\in[w_\mu,\min\{v_\mu,1\}]$ of $u_k$ (along a subsequence if needed), which is an entire solution of \eqref{1.9}. Since $(w_\mu)_t\ge 0$, the same is true for $u_k$ and thus $u_\mu$, by the maximum principle. The strong maximum principle then gives $(u_\mu)_t>0$ because $(u_\mu)_t\not\equiv 0$.
Finally, \eqref{1.6} follows from \eqref{1.14} and $v_\mu(-\infty)=\infty$, the latter being due to \eqref{2.8} below.
(ii) The fact that $u_\mu$ is a transition front with a bounded width in the sense of \eqref{1.7} when $\lambda_0<\mu_0\le \mu_1<\lambda_1$ will follow from the existence of $L>0$ such that
\begin{equation} \label{2.8}
\phi_\lambda(c) \ge 2 \phi_\lambda(d)
\end{equation}
whenever $\lambda\in[\mu_0,\mu_1]$ and $c\le d-L$. Indeed, we will show that such $L$ depends only on $a_+,B_\pm,\zeta$, provided $\mu_0-\lambda_0\ge\zeta>0$. Then \eqref{2.8} holds with the same $L$ for $v_\mu$ in place of $\phi_\lambda$. Therefore, if now $\min\{\mu_0-\lambda_0, \lambda_1-\mu_1\} \ge\zeta>0$, then this and (i) gives \eqref{1.7} with $L_\varepsilon$ depending only on $g,a_+,B_\pm,\varepsilon,\zeta$.
We are left with proving \eqref{2.8}.
If in \eqref{1.11a} we take
\[
\psi(x)\equiv
\begin{cases}
\phi_\lambda (x) & x\in(c,d),
\\ \phi_\lambda(c)(x-c+1) & x\in[c-1,c],
\\ \phi_\lambda(d)(d+1-x) & x\in[d,d+1],
\\ 0 & x\in{\mathbb{R}}\setminus [c-1,d+1]
\end{cases}
\]
for some $c<d$, then we obtain using \eqref{2.6} and $\alpha<1$,
\begin{align*}
\int_ {\mathbb{R}} & [- B(x)\psi'(x)^2 + q(x)\psi'(x)\psi(x) + a(x)\psi(x)^2 ] dx
\\ \ge & \int_c^d [- B(x)\phi_\lambda'(x)^2 + q(x)\phi_\lambda'(x)\phi_\lambda(x) + a(x)\phi_\lambda(x)^2 ] dx - (B_+ + q_+)(\phi_\lambda(c)^2+\phi_\lambda(d)^2)
\\ \ge & \int_c^d [(B(x)\phi_\lambda'(x))' + q(x)\phi_\lambda'(x) + a(x)\phi_\lambda(x) ]\phi_\lambda(x) dx
\\ & \qquad - (B_+ + q_+)(|\phi_\lambda'(c)|\phi_\lambda(c) + |\phi_\lambda'(d)|\phi_\lambda(d) + \phi_\lambda(c)^2+\phi_\lambda(d)^2)
\\ \ge & \lambda \int_c^d \phi_\lambda(x)^2 dx - (B_+ + q_+)(1+a_+^{1/2}B_-^{-1/2})( \phi_\lambda(c)^2+\phi_\lambda(d)^2).
\end{align*}
This and \eqref{1.11a}
give
\[
\lambda_0 \int_c^d \phi_\lambda(x)^2dx \ge \lambda \int_c^d \phi_\lambda(x)^2 dx - [\lambda_0 + (B_+ + q_+)(1+a_+^{1/2}B_-^{-1/2})]( \phi_\lambda(c)^2+\phi_\lambda(d)^2),
\]
which after setting $M\equiv [\lambda_0 + (B_+ + q_+)(1+a_+^{1/2}B_-^{-1/2})](\lambda-\lambda_0)^{-1}$ reads
\begin{equation} \label{2.10}
\int_c^d \phi_\lambda(x)^2dx \le M( \phi_\lambda(c)^2+\phi_\lambda(d)^2).
\end{equation}
By the Harnack inequality, there is $N>0$ such that $\phi_\lambda(y)\le N \phi_\lambda(x)$ if $|x-y|\le 2M$. Set $L\equiv 6MN^2$ and assume \eqref{2.8} is violated for some $c\le d-L$ (notice that $L$ depends only on $a_+,B_\pm,\zeta$ if $\mu_0-\lambda_0\ge\zeta>0$, because $\lambda_0\le a_+$ and $q_+\le 2\sqrt{a_+B_+}$). Then there must be $x\in[c,d]$ such that $\phi_\lambda(x)\le N^{-1}\phi_\lambda(d)$ because otherwise
\[
\int_c^d \phi_\lambda(x)^2dx \ge 6M\phi_\lambda(d)^2 > M( \phi_\lambda(c)^2+\phi_\lambda(d)^2),
\]
contradicting \eqref{2.10}. Let $y$ be the rightmost point such that $y<d$ and $\phi_\lambda(y)= N^{-1}\phi_\lambda(d)$, and $z$ the leftmost point such that $z>d$ and $\phi_\lambda(z)= N^{-1}\phi_\lambda(d)$. Then $y\le d-2M$, $z\ge d+2M$, and $\phi_\lambda(x)\ge N^{-1}\phi_\lambda(d)$ for any $x\in[y,z]$. But this contradicts \eqref{2.10} with $y,z$ in place of $c,d$, so \eqref{2.8} is proved and we are done.
\end{proof}
{\it Remark.} The argument in (i) works even for $\mu_1=\lambda_1$, with $\alpha=1$ and $m=0$. Then $w_\mu\equiv h_{g,1}(v_\mu)$ will again be a sub-solution of \eqref{1.9} but this time $w_\mu \not\le v_\mu$ so we cannot recover a solution between them.
\begin{proof}[Proof of Theorem \ref{T.1.2}]
From \eqref{1.11a}
we know that $\lambda_0:L^\infty_{\rm loc}({\mathbb{R}})^3\to{\mathbb{R}}$ is lower semi-continuous, which together with measurability of $p:\Omega\to L^\infty_{\rm loc}({\mathbb{R}})^3$ means that $A_\zeta\equiv \{\omega\in\Omega\,|\, \lambda_0(\omega)>\zeta\}$ is a measurable set. Obviously $\pi_y A_\zeta=A_\zeta$ for all $y\in{\mathbb{R}}$, so ${\mathbb{P}}(A_\zeta)\in\{0,1\}$ for each $\zeta\in{\mathbb{R}}$. This means that $\lambda_0$ is almost constant on $\Omega$. The same follows for $\lambda_1$, using its upper semi-continuity as a function on $L^\infty_{\rm loc}({\mathbb{R}})^3$, which follows from its definition.
Let us replace $\Omega$ by its full-measure subset on which $\lambda_0,\lambda_1$ are constant. Next fix any $\lambda\in(\lambda_0,\lambda_1)$ and let $u_{\delta_\lambda}(\omega;t,x)$ be the corresponding random transition front. The remark after the proof of Theorem \ref{T.1.1} shows that there is $L$ such that \eqref{2.8} holds for any $\omega\in\Omega$ and $c\le d-L$. Therefore also $L_\varepsilon$ in that proof is uniform in $\omega$, which means that if $Y(\omega;t)$ is the rightmost point such that $e^{\lambda t}\phi_\lambda(\omega;Y(\omega;t))=\tfrac 12$ and $X(\omega;t)$ the rightmost point such that $u_{\delta_\lambda}(\omega; t, X(\omega;t))=\tfrac 12$, then $|X(\omega;t)-Y(\omega;t)|$ is uniformly bounded on $\Omega\times{\mathbb{R}}$. Thus we only need to prove \eqref{1.15} for $Y$ in place of $X$.
Notice that if $r_\lambda(\omega)\equiv \phi_\lambda'(0)$, then $r_\lambda:\Omega\to{\mathbb{R}}$ is measurable because $p:\Omega\to p(\Omega)$ is measurable and $r_\lambda:p(\Omega)\to {\mathbb{R}}$ is continuous when $p(\Omega)$ is equipped with $L^\infty_{\rm loc}({\mathbb{R}})^3$-induced topology. The latter follows from \eqref{2.8} and the fact that any solution of \eqref{1.12} with $\phi(0)=1$ and $\phi'(0)\neq r_\lambda(\omega)$ grows exponentially as $x\to\infty$ (by \eqref{2.8} applied to the solution $\psi_\lambda$ converging to 0 as $x\to-\infty$ and the fact that $\phi_\lambda,\psi_\lambda$ are a basis of the set of all solutions).
Therefore $\phi_\lambda(\cdot;x)$ is measurable for any fixed $x$. Since $\phi_\lambda(\pi_y\omega;\cdot)=\phi_\lambda(\omega;y)^{-1}\phi_\lambda(\omega;y+\cdot)$, we have $\phi_\lambda(\omega;y+x)=\phi_\lambda(\omega;y)\phi_\lambda(\pi_y\omega;x)$. So from ergodicity of $\{\pi_y\}_{y\in{\mathbb{R}}}$ and Oseledec theorem it follows that for almost all $\omega\in\Omega$,
\[
\lim_{x\to \pm\infty} \frac 1x \ln \phi_\lambda(\omega; x) = -\tau_\pm
\]
for some $\tau_\pm\in{\mathbb{R}}$ (and $\tau_\pm>0$ by \eqref{2.8}). Moreover, $\tau_+=\tau_-$. Otherwise,
there exists $\Omega'\subset\Omega$ and $M<\infty$ such that ${\mathbb{P}}(\Omega')>\tfrac 12$ and
\[
\left |\frac 1{\pm M} \ln \phi_\lambda(\omega; {\pm M}) -\tau_\pm \right| < \frac{|\tau_+-\tau_-|}2
\]
for all $\omega\in\Omega'$. But then
\[
\left |\frac 1M \ln \phi_\lambda(\pi_{-M} \omega; M) -\tau_- \right| < \frac{|\tau_+-\tau_-|}2
\]
for all $\omega\in\Omega'$, so $\Omega'\cap\pi_{-M}\Omega'=\emptyset$, a contradiction with ${\mathbb{P}}(\pi_{-M}\Omega')={\mathbb{P}}(\Omega')>\tfrac 12$.
Then $\tau_+=\tau_-$ and \eqref{2.8} give
\[
\lim_{|t|\to\infty} \frac{Y(\omega; t)}t = \frac \lambda{\tau_\pm} \equiv c_\lambda
\]
and the result follows.
\end{proof}
\begin{proof}[Proof of Theorem \ref{T.1.3}]
(i) The proof of all the claims, with the exception of the last one, is identical to the proof of Theorem \ref{T.1.1}(i), with $\alpha<1$ from the statement of Theorem \ref{T.1.3}(i), and \eqref{2.6} replaced by
\[
|\nabla v_\xi(t,x)|^2 = |\xi|^2 v_\xi(t,x)^2 \le\alpha \aaa v_\xi(t,x)^2
\]
for all $|\xi|\le \sqrt{\alpha \aaa }$.
The last claim is an easy consequence of $u_\mu(t,x)v_\mu(t,x)^{-1} \to 1$ as $v_\mu(t,x)\to 0$ and of
\[
\left( \frac{|t|}\pi \right)^{d/2} v_\mu(t,2t\zeta) e^{(|\zeta|^2-\aaa )t} d\zeta \rightharpoonup d\mu(\zeta)
\]
as $t\to -\infty$.
The latter statement, similar to one in \cite{HN2}, follows from
\[
\left( \frac{|t|}\pi \right)^{d/2} v_\mu(t,2t\zeta) e^{(|\zeta|^2-\aaa )t} = \int_Y \left( \frac{|t|}\pi \right)^{d/2} e^{-|\xi-\zeta|^2|t|} d\mu(\xi)
\]
for $\zeta\in{\mathbb{R}}^d$ and $t<0$.
(iii) If $u_\mu(t,x)\ge \varepsilon$, then $v_\mu(t,x)\ge h^{(-1)}(\varepsilon)$ with $h$ from (i). Then there is a unit vector $\eta=\eta_{x,t}$ such that
\[
\int_{Y_{\eta,\theta}} e^{-\xi\cdot x+(|\xi|^2+\aaa )t} d\mu(\xi) \ge \frac \theta{2\pi} h^{(-1)}(\varepsilon),
\]
where
\[
Y_{\eta,\theta} \equiv \left\{\xi\in Y \,\bigg|\, \arccos \frac {-\eta\cdot\xi} {|\xi|} \le \frac\theta 2 \right\}.
\]
If now $\eta\cdot y|y|^{-1} \ge \theta$, then $\arccos(\eta\cdot y|y|^{-1}) \le \tfrac \pi 2-\theta$, and so $\arccos(-\xi\cdot y|y|^{-1}|\xi|^{-1}) \le \tfrac {\pi-\theta} 2$ for any $\xi\in Y_{\eta,\theta}$. Therefore
\[
v_\mu(t,x +y) \ge \int_{Y_{\eta,\theta}} e^{-\xi\cdot (x+y)+(|\xi|^2+\aaa )t} d\mu(\xi) \ge \frac \theta{2\pi} h^{(-1)}(\varepsilon) |y| \dist(0,\text{\rm{supp}}(\mu)) \cos \frac {\pi-\theta} 2
\]
and the result follows from \eqref{1.14} with
\[
L_{\varepsilon,\theta}\equiv \left[\frac \theta{2\pi} h^{(-1)}(\varepsilon) \dist(0,\text{\rm{supp}}(\mu)) \cos \frac {\pi-\theta} 2 \right]^{-1} h^{(-1)}(1-\varepsilon).
\]
(ii) Assume first that $0\in\ch(\mu)$ and $\nu(Y)^{-1}\int_{Y} \xi d\nu(\xi) = 0$ for some $0<\nu\le\mu$. Then
\[
v_\mu(t,x) \ge \int_{Y} e^{-\xi\cdot x + \aaa(3-\sgn(t)) t/2} d\nu(\xi) \ge \nu(Y) e^{- \nu(Y)^{-1}\int_{Y} \xi d\nu(\xi)\cdot x} e^{\aaa(3-\sgn(t)) t/2} = \nu(Y) e^{\aaa(3-\sgn(t)) t/2}
\]
by Jensen's inequality. This and \eqref{1.14} yield $\inf_{x\in{\mathbb{R}}^d} u_\mu(x,t) >0$ for each $t\in{\mathbb{R}}$.
Now assume that $0\notin\ch(\mu)$ and define $\hat\mu_d\equiv \mu$. The second claim in \eqref{1.19a} follows from $\mu>0$ and (i) so let us prove the first claim. Since $\ch(\mu)$ is a convex set, it must be contained in a closed half-space with 0 on its boundary. Assume without loss it is ${\mathbb{R}}^{d-1}\times{\mathbb{R}}^+_0$, and let $\mu_d\equiv \hat\mu_d|_{{\mathbb{R}}^{d-1}\times{\mathbb{R}}^+}$ and $\hat \mu_{d-1}\equiv \hat\mu_d|_{{\mathbb{R}}^{d-1}\times\{0\}} = \hat\mu_d-\mu_d$. Now $\ch(\mu)\cap ({\mathbb{R}}^{d-1}\times\{0\})$ must be contained in a closed half-space of ${\mathbb{R}}^{d-1}\times\{0\}$ with 0 on its boundary. Assume without loss it is ${\mathbb{R}}^{d-2}\times{\mathbb{R}}^+_0\times\{0\}$, and let $\mu_{d-1}\equiv \hat\mu_{d-1}|_{{\mathbb{R}}^{d-2}\times{\mathbb{R}}^+\times\{0\}}$ and $\hat \mu_{d-2}\equiv \hat\mu_{d-1}-\mu_{d-1}$. Continue in this way until obtaining $\mu_1 = \hat\mu_1$ supported in ${\mathbb{R}}^+\times\{0\}^{d-1}$ (because $\hat\mu_0=\mu|_{\{0\}}=0$).
Since $\mu=\mu_1+\dots+\mu_d$ and $u_\mu\le v_\mu$, it is sufficient to show that for any $\varepsilon>0$ there is $x\in{\mathbb{R}}^d$ such that for $k=1,\dots,d$ we have
\begin{equation} \label{2.11}
\int_{Y} e^{-\xi\cdot x} d\mu_k(\xi) \le \frac \varepsilon d
\end{equation}
(the extra factor $e^{(|\xi|^2+\aaa )t}\le e^{\aaa(3+\sgn(t)) t/2}$ from the definition of $v_\mu$ can be absorbed in $\varepsilon$). For $k=1$, the set of $x\in{\mathbb{R}}^d$ satisfying \eqref{2.11} contains some half-space $[\rho_1,\infty)\times{\mathbb{R}}^{d-1}$. For each $k= 2,\dots ,d$ and any $r_k>0$, it contains $\bar B_{r_k}(0)\times[\rho_{k,r_k},\infty)\times{\mathbb{R}}^{d-k}$ for some $\rho_{k,r_k}>0$, where $\bar B_{r_k}(0)$ is the closed ball in ${\mathbb{R}}^{k-1}$ with radius $r_k$ and center 0. If we choose $r_2\ge \rho_1$ and then recursively $r_k\ge r_{k-1}+\rho_{k-1,r_{k-1}}$ for $k=3,\dots,d$, the corresponding $k$ sets
all contain the point $x=(\rho_1,\rho_{2,r_2},\dots,\rho_{d,r_d})$. So \eqref{2.11} holds for this $x$ and we are done.
\end{proof}
{\it Remark.} We have $\ch(\mu)\subseteq\ches(\mu)$, the intersection of convex hulls of all essential supports of $\mu$. This is because if $A$ is an essential support of $\mu$ and $\ch(A)$ its convex hull, then ${\mathbb{E}}(\nu)=\nu({\mathbb{R}}^d)^{-1}\int_A\xi d\nu(\xi)\in\ch(A)$ when $0<\nu\le\mu$. The opposite inclusion follows from the construction at the end of the previous proof applied to any $\zeta\notin\ch(\mu)$ instead of 0. Indeed, for any such $\zeta$, one can again find open half-spaces $S_d,\dots,S_1$ of dimensions $d,\dots,1$ whose boundaries contain $\zeta$ (without loss these can be assumed to be $S_k=\zeta + {\mathbb{R}}^{k-1}\times{\mathbb{R}}^+\times\{0\}^{d-k} $) and measures $\mu_k$ on $S_k$ ($k=d,\dots,1$) such that $\mu=\mu_1+\dots+\mu_d$. Thus $S\equiv \bigcup_{k=1}^d S_k$ is an essential support of $\mu$ and $\zeta\notin S$, which yields $\ch(\mu)\supseteq\ches(\mu)$.
\begin{proof}[Proof of Theorem \ref{T.1.5}]
This is identical to the previous proof, using that \eqref{2.4f} yields \eqref{2.3} for $v_\xi$ when $\xi\in S_\alpha$, and thus also for $v_\mu$ because $v_\xi>0$.
\end{proof}
\medskip
|
\section{Introduction}
Low-dimensional superconductors are systems displaying a surprisingly
complex and rich physics, allowing the study of paradigmatical phenomena
in condensed matter physics, like quantum phase transitions and quantum
critical behavior, electronic localization, Coulomb blockade, etc.
\citep{schoen90_review_ultrasmall_tunnel_junctions,fazio01_review_superconducting_networks}
In particular, an intriguing superconductor-insulator phase transition
(SIT) was observed experimentally in superconducting films \citep{Haviland89_Onset_of_superconductivity_in_the_two-dimensional_limit_PhysRevLett.62.2180,Mason99_Dissipation_effects_on_the_SIT_in_2D,Mason02_SIT_in_a_capacitively_coupled_dissipative_environment},
wires \citep{bezryadin00,arutyunov08_superconductivity_1d_review,Bezryadin08_QPS_review},
and in ultrasmall-capacitance Josephson junction arrays (JJAs) in
two \citep{vanderZant96_QPT_in_2D_Experiments_in_JJas,rimberg97_dissipation_driven_sit_2D_josephson_array,Takahide00_SIT_in_2DJJAs}
and one dimensions \citep{Chow98_Length_scale_dependence_of_SIT_in_1D_array_J,Kuo01_Magnetic_induced_transition_in1DJJA,Miyazaki02_QPT_in_1D_arrays_of_JJs,Takahide06_SIT_2D_1D_crossover},
giving rise to an intense theoretical activity.\citep{Fisher1989,Fisher1990}
In this transition, as one of the parameters is varied (e.g., the
normal-state resistance of the film, the thickness of the wire, the
Josephson coupling $E_{J}$ in the array, etc.) the groundstate of
the system changes from superconducting to insulator.
In one-dimensional (1D) superconductors, one particular kind of excitation,
the so-called quantum-phase slip (QPS) processes, have been recently
the focus of an intense research.\citep{Bezryadin08_QPS_review} The
interest is based both on the putative role of QPS in the SIT in 1D\citep{zaikin97},
as well as for their potential uses in novel qubit architectures\citep{Mooij05_Phase-slip_qubit},
a fact that has stimulated recent interesting experimental research
in 1DJJAs.\citep{Pop10_QPS_in_JJA,Manucharyan10_Coherent_QPS_in_JJA}
A phase-slip is a discrete process occuring in a 1D superconductor,
in which the amplitude of the order parameter vanishes temporarily
at a particular point, allowing the phase of the order parameter to
change abruptly in units of $2\pi$. In particular, a\textit{ }QPS
is a phase-slip excitation originated in macroscopic quantum tunneling
of the phase of the order parameter. \citep{giordano94}
On the other hand, since the seminal works by Caldeira and Leggett\citep{caldeira&leggett81},
it has been known that dissipation in macroscopic quantum systems
plays a central role. For instance, in a two-dimensional JJA capacitively
coupled to a proximate two-dimensional electron gas (2DEG), Rimberg
\textit{et al.} observed a tunable-SIT upon variation of the backgate
voltage $V_{g}$ applied to the 2DEG.\citep{rimberg97_dissipation_driven_sit_2D_josephson_array}
In that work, it was shown that $V_{g}$ has the effect of tuning
the sheet resistance $R_{\square}$ in the 2DEG through the modulation
of its electronic density, a fact that in turn modifies the electromagnetic
environment of the JJA. It was argued later by Wagenblast \textit{et
al.} \citep{Wagenblast97_SIT_in_a_Tunable_Dissipative_Environment}
that due to the incomplete screening of the Coulomb interaction provided
by the 2DEG, the 2D plasma mode in the array is overdamped and the
charging energy in the junction $E_{C}$ is renormalized to higher
values, producing a SIT whenever the ratio $E_{J}/\tilde{E}_{C}\sim1$,
with $\tilde{E}_{C}$ the renormalized $E_{C}$. While this scenario
is reasonable in a 2D geometry, in a 1DJJA capacitively coupled to
a 2DEG, the screening provided by the metal is typically very efficient,
and a significant damping of the 1D propagating plasma mode is not
expected.\citep{Lobos10_Dissipative_phase_fluctuations} This leads
to the naive conclusion that in the 1DJJA/2DEG geometry, a dissipation-driven
SIT should not occur.
\begin{figure}[h]
\includegraphics[bb=75bp 50bp 700bp 270bp,clip,scale=0.4]{jja_system_2}\caption{\label{fig:system}Scheme of the system under study. The 1DJJA is
capacitively coupled to the metallic film, which controls the electromagnetic
environment. A gate voltage allows to modify the sheet-resistance
$R_{\square}$ of the film, modifying the dissipation in the 1DJJA.}
\end{figure}
However, in a closely related Luttinger-liquid system placed in proximity
to a metallic plane, a dissipation-driven quantum phase transition
was predicted to occur.\citep{cazalilla06_dissipative_transition}
This transition is driven by backscattering events originated in the
Luttinger liquid under the effect of the dynamically screened Coulomb
interaction. It is therefore interesting to study to what extent the
same dissipative processes will affect the dynamics of QPS in 1D superconductors
in proximity of a diffusive metallic plane. Indeed, the question of
dissipation in 1D superconductors is an active area of research, and
theoretical predictions point towards the important role of intrinsic
and/or extrinsic dissipation mechanisms in determining their $T=0$
phase diagram.\citep{goswami06_josephson_array,refael07_SN_transition_in_grains&nanowires}
In this article we explore the possibility of a dissipation-driven
SIT in a 1DJJA capacitively coupled to a proximate diffusive 2DEG.
We concentrate in particular on the low-temperature phase diagram
and on the transport properties of the array. Using a bosonization
approach,\citep{giamarchi_book_1d} we derive the dissipative effective
action from a microscopic Hamiltonian, and we elucidate the role of
dissipation in the SIT. One important conclusion in our work is that
for weak dissipation, the transition occurs always between a superconducting
and an insulating phase, in contrast to other works predicting quadrupolar
and normal phases.\citep{goswami06_josephson_array,refael07_SN_transition_in_grains&nanowires}
We believe this is a consequence of a different kind of dissipation
in the model. We find that, except in the experimentally challenging
situation in which the Cooper-pair density in the superconducting
islands is not commensurate, the SIT is always of the Berezinskii-Kosterlitz-Thouless
(BKT) type,\citep{Berezinskii1971,Kosterlitz1973} and is originated
in the unbinding of QPS/anti-QPS pairs. Dissipation stabilizes the
insulating groundstate through the introduction of friction in the
dynamics of the 1D superfluid density, a fact that could be observed
experimentally in the dc-resistivity of the 1DJJA. Specifically, we
predict a resistivity of the form $\varrho\left(T\right)\sim A_{1}T^{\nu_{1}}+A_{2}T^{\nu_{2}}$
in the superconducting phase, and $\varrho\left(T\right)\sim\eta Te^{\Delta/T}/\Delta$,
in the insulating phase, with $\Delta$ the insulating gap and $\eta$
the dissipation parameter.
The paper is organized as follows. In Section \ref{sec:model} we
derive the effective model for a 1DJJA coupled to a 2DEG, in Section
\ref{sec:phase-diagram} we derive the $T=0$ phase-diagram as a
function of the parameters of the model, Section \ref{sec:transport}
is devoted to the study of the experimental consequences of our results,
and finally in Section \ref{sec:summary} we present a summary and
our conclusions.
\section{\label{sec:model}Model }
We start the analysis by considering an ideally isolated JJA, with
length $L\rightarrow\infty$. To simplify the analysis, we neglect
in the following the fermionic degrees of freedom forming the Cooper-pairs
at a microscopic level. This {}``boson-only'' approximation is belived
to describe correctly the critical properties of a JJA at temperatures
$T\ll T_{c}$, with $T_{c}$ the superconducting critical temperature
in the bulk of the superconducting island.\citep{Fisher1989} The
usual description of the isolated, infinite 1DJJA is given in terms
of the quantum phase model \citep{fazio01_review_superconducting_networks}\begin{align}
H_{\text{JJA}} & =\frac{1}{2}\sum_{i,j}\left(n_{i}-\bar{n}\right)v_{ij}\left(n_{j}-\bar{n}\right)+\sum_{\left\langle ij\right\rangle }E_{J}\left(1-\cos\theta_{i}-\theta_{j}\right).\label{eq:H_JJA}\end{align}
The dynamical variables of this model are the number of Cooper pairs
$n_{i}$ and the phase of the superconducting order parameter $\theta_{i}$
at every site $i$ in the array. These variables obey the usual phase-number
commutation relation in the BCS groundstate, i.e., $\left[\theta_{i},n_{j}\right]=i\delta_{ij}.$\citep{Tinkham}
The first term in Eq. (\ref{eq:H_JJA}) represents the charging energy,
with $v_{ij}$ the unscreened Coulomb interaction {[}cf. Eq. (\ref{eq:Coulomb_potential}){]}
between the excess charges at sites $i$ and $j$, and $\bar{n}$
corresponds to an average charge imposed, e.g., by external gate voltages.
The second term is the Josephson energy contribution, parametrized
by $E_{J}$. In the following we use the convention $\hbar=k_{B}=1$.
The critical properties of model Eq. (\ref{eq:H_JJA}) are more conveniently
studied using a field-theoretical approach, valid for fluctuations
of wavelengths much larger that the lattice parameter of the array
$a$.\citep{giamarchi_book_1d} We therefore introduce the coarse-grained
superfluid density $\delta\rho\left(x\right)$, defined as $\delta\rho\left(x_{i}\right)=\left(n_{i}-\bar{n}\right)/a$,
and we expand the Josephson term as $E_{J}\cos\left(\theta_{i}-\theta_{j}\right)\simeq E_{J}a^{2}\left[\nabla\theta\left(x_{i}\right)\right]^{2}$.
At low temperatures, the continuum limit of Hamiltonian Eq. (\ref{eq:H_JJA})
reads\begin{align}
H_{\text{JJA}} & =\frac{1}{2}\int dxdx^{\prime}\;\delta\rho\left(x\right)v\left(x-x^{\prime},0\right)\delta\rho\left(x^{\prime}\right)\nonumber \\
& +\frac{1}{2}E_{J}a\int dx\;\left(\nabla\theta\left(x\right)\right)^{2}.\label{eq:H_JJA_continuum}\end{align}
Here the 1D superfluid interacts via the \textit{bare} Coulomb potential,
which we define for convenience as \begin{align}
v\left(\mathbf{r},z\right) & =\frac{e^{2}}{\epsilon_{\text{r}}}\frac{1}{\sqrt{r^{2}+z^{2}+a^{2}}},\label{eq:Coulomb_potential}\end{align}
where $r=\left|\mathbf{r}\right|$ and $z$ are, respectively, the
distance in the $xy$-plane and along the azimuthal direction between
two point-charges (cf. Fig. \ref{fig:system}). Here the lattice parameter
$a$ acts as the short-distance regularization of the interaction
and $\epsilon_{\text{r}}$ is the permitivity of the insulating medium
surrounding the islands. Note that in Eq. (\ref{eq:H_JJA_continuum})
we do not assume an \textit{a priori} short-ranged, screened interaction
as is usually done when dealing with JJAs.\citep{Wagenblast97_SIT_in_a_Tunable_Dissipative_Environment}
This will result as a natural consequence of the interaction with
the 2DEG (see below). One problem of this field-theoretical approach
is that Mott-instabilities (crucial when the superfluid density is
commensurate with the lattice) are lost in Eq. (\ref{eq:H_JJA_continuum})
after taking the continuum limit. One way to cure this problem is
to introduce a phenomenological term $H_{1}=-\int dx\; V_{l}\left(x\right)\rho\left(x\right)$,
where the effective superfluid density $\rho\left(x\right)$ {[}cf.
Eq. (\ref{eq:density_bosonized}){]} couples to the phenomenological
potential $V_{l}\left(x\right)$, having the same periodicity of the
lattice.\citep{giamarchi_book_1d}
The electrons in the 2DEG are described by the Hamiltonian\begin{align}
H_{\textrm{2D}} & =\int d^{2}\mathbf{r}\;\sum_{\sigma}\left[-\frac{1}{2m}\eta_{\sigma}^{\dagger}\nabla^{2}\eta_{\sigma}+V_{\textrm{imp}}\eta_{\sigma}^{\dagger}\eta_{\sigma}\right]+\nonumber \\
& +\frac{1}{2}\int d^{2}\mathbf{r}d^{2}\mathbf{r}^{\prime}\;\delta\rho_{\textrm{2D}}\left(\mathbf{r}\right)v\left(\mathbf{r}-\mathbf{r}^{\prime},0\right)\delta\rho_{\textrm{2D}}\left(\mathbf{r}^{\prime}\right),\label{eq:h_n}\end{align}
where the fermionic field-operator $\eta_{\sigma}^{\dagger}\equiv\eta_{\sigma}^{\dagger}\left(\mathbf{r}\right)$
creates an electron in the 2DEG with spin projection $\sigma$ at
spatial position $\mathbf{r}\equiv\left(x,y\right)$, and $V_{\textrm{imp}}\equiv V_{\textrm{imp}}\left(\mathbf{r}\right)$
represents a weak static impurity potential which provides a finite
resitivity and dissipation in the metal. In terms of $\eta_{\sigma}^{\dagger}\left(\mathbf{r}\right),\eta_{\sigma}\left(\mathbf{r}\right)$,
the density-operator $\rho_{\textrm{2D}}\left(\mathbf{r}\right)$
in the 2DEG writes $\rho_{\textrm{2D}}\left(\mathbf{r}\right)\equiv\sum_{\sigma}\eta_{\sigma}^{\dagger}\left(\mathbf{r}\right)\eta_{\sigma}\left(\mathbf{r}\right)$,
and $\delta\rho_{\textrm{2D}}\left(\mathbf{r}\right)\equiv\rho_{\textrm{2D}}\left(\mathbf{r}\right)-\rho_{0,\textrm{2D}}$,
with $\rho_{0,\textrm{2D}}$ the average density in the metal.
Finally, the interaction between the 1DJJA and the 2DEG placed at
a distance $d$ (cf. Fig. \ref{fig:system}) is described by the Hamiltonian\begin{align}
H_{\textrm{int}} & =\int d^{2}\mathbf{r}dx^{\prime}\;\delta\rho\left(x^{\prime}\right)v\left(x^{\prime}-\mathbf{r},d\right)\delta\rho_{\textrm{2D}}\left(\mathbf{r}\right).\label{eq:h_int}\end{align}
Our goal in this Section is to derive an effective model for the 1DJJA
capacitively coupled to the 2DEG. To that end we introduce the partition
function of the system\citep{negele_book}
\begin{align*}
Z & =\int\mathcal{D}\left[\rho,\theta\right]\mathcal{D}\left[\bar{\eta},\eta\right]\; e^{-S},\end{align*}
where $S$ is the Euclidean action of the problem\begin{align}
S & =S_{\textrm{JJA}}+S_{\textrm{2D}}+S_{\textrm{int}},\label{eq:S_eff_total}\end{align}
where \begin{align*}
S_{\text{JJA}} & =\int_{0}^{\beta}d\tau\int dx\; i\partial_{\tau}\theta\left(x,\tau\right)\rho\left(x,\tau\right)+\int_{0}^{\beta}d\tau\; H_{\text{JJA}}\left(\tau\right),\\
S_{\text{2D}} & =\int_{0}^{\beta}d\tau\left[\int d^{2}\mathbf{r}\;\bar{\eta}\left(\mathbf{r},\tau\right)\left(\partial_{\tau}-\mu_{\text{2D}}\right)\eta\left(\mathbf{r},\tau\right)+H_{\text{2D}}\left(\tau\right)\right],\\
S_{\text{int}} & =\int_{0}^{\beta}d\tau\; H_{\text{int}}\left(\tau\right).\end{align*}
Here $\mu_{\textrm{2D}}=k_{\text{F}}^{2}/2m-eV_{g}$ is the effective
chemical potential in the metal, with $k_{\text{F}}=\left|\mathbf{k}_{\text{F}}\right|$
the Fermi wavevector, and $V_{g}$ the gate voltage applied to the
2DEG, which allows to change the value of $\rho_{0,\textrm{2D}}$,
and therefore, the sheet-resistance $R_{\square}$.
The first step in the derivation of an effective model for the array
is to integrate out the fermionic degrees of freedom $\bar{\eta}\left(\mathbf{r},\tau\right),\eta\left(\mathbf{r},\tau\right)$
in the 2DEG. Assuming that the term $S_{\textrm{int}}$ can be treated
perturbatively (we check the consistency of this assumption later),
the integration of the fermionic degrees of freedom in the metal yields
\begin{align}
S_{\textrm{eff}} & \simeq S_{\textrm{JJA}}-\frac{1}{2}\int d\tau d\tau^{\prime}\int dxdx^{\prime}\;\delta\rho\left(x,\tau\right)\nonumber \\
& \times v_{\textrm{scr}}\left(x-x^{\prime},\tau-\tau^{\prime}\right)\delta\rho\left(x^{\prime},\tau^{\prime}\right).\label{eq:S_2nd_order}\end{align}
We do not provide the details of this derivation here, and we refer
the interested reader to Refs. \onlinecite{Lobos10_Dissipative_phase_fluctuations}
and \onlinecite{cazalilla06_dissipative_transition}. In Eq. (\ref{eq:S_2nd_order})
we have introduced the 1D effective screening potential $v_{\textrm{scr}}\left(x-x^{\prime},\tau-\tau^{\prime}\right)$,
which encodes all the screening effects provided by the 2DEG. This
quantity writes more conveniently in Fourier representation as\citep{Lobos10_Dissipative_phase_fluctuations}
\begin{align}
v_{\textrm{scr}}\left(k,\omega_{m}\right) & \equiv\frac{1}{L_{\perp}}\sum_{k_{\perp}}\frac{\left[v_{\textrm{2D}}\left(\mathbf{k},d\right)\right]^{2}\chi_{0,\textrm{2D}}\left(\mathbf{k},\omega_{m}\right)}{1+v_{\textrm{2D}}\left(\mathbf{k},0\right)\chi_{0,\textrm{2D}}\left(\mathbf{k},\omega_{m}\right)},\label{eq:v_scr}\end{align}
where $\omega_{m}=2\pi m/\beta$ are the bosonic Matsubara frequencies,\citep{mahan2000}
and $\mathbf{k}=\left(k,k_{\perp}\right)$ is the wavevector in 2D,
where we have made explicit the component $k_{\perp}$ in the 2DEG,
perpendicular to the 1DJJA. The quantity $v_{\textrm{2D}}\left(\mathbf{k},d\right)=\left(2\pi e^{2}/\epsilon_{\text{r}}\right)\exp\left(-\left|\mathbf{k}\right|\sqrt{d^{2}+a^{2}}\right)/\left|\mathbf{k}\right|$
is the 2D Fourier transform of the Coulomb potential Eq. (\ref{eq:Coulomb_potential}).
We assume that the length of the array is $L<\xi_{\text{loc}}$, with
$\xi_{\text{loc}}$ the Anderson localization length in the 2DEG,
a condition well fulfilled in practice. In that case, the density-density
response function in the 2DEG, averaged over disorder configurations,
writes $\chi_{0,\textrm{2D}}\left(\mathbf{k},\omega_{m}\right)=2\mathcal{N}_{2D}^{0}D\mathbf{k}^{2}/\left(D\mathbf{k}^{2}+\left|\omega_{m}\right|\right)$,
where $D$ and $\mathcal{N}_{2D}^{0}$ are, respectively, the diffusion
constant and the density of states (at the Fermi energy) per spin
projection.\citep{akkermans}
We can now define the total effective retarded interaction
\begin{align}
v_{\textrm{eff}}\left(k,\omega_{m}\right) & =v_{\textrm{1D}}\left(k,0\right)-v_{\textrm{scr}}\left(k,\omega_{m}\right),\label{eq:v_eff}\end{align}
where $v_{\textrm{1D}}\left(k,0\right)=2e^{2}K_{0}\left(\left|k\right|a\right)/\epsilon_{\text{r}}$
is the Fourier transform of Eq. (\ref{eq:Coulomb_potential}) in 1D,
and $K_{0}\left(x\right)$ is the zeroth-order modified Bessel function.\citep{abramowitz_math_functions}
Physically, the effective potential $v_{\textrm{eff}}\left(k,\omega_{m}\right)$
describes the interaction among charges in the array, both via the
instantaneous interaction $v_{\textrm{1D}}\left(k,0\right)$ arising
from the direct intrawire Coulomb interaction, as well as indirectly
via the coupling to the diffusive modes in the 2DEG, which corresponds
to the retarded interaction $v_{\textrm{scr}}\left(k,\omega_{m}\right)$
Eq. (\ref{eq:v_scr}).
We now introduce a more convenient representation of the superfluid
density in the 1DJJA. To motivate our approach, we first note that
in the absence of Josephson coupling {[}i.e., $E_{J}=0$ in Eq. (\ref{eq:H_JJA}){]},
the Cooper-pair occupation number $n_{i}$ is a good quantum number
in each island, fixed by $\bar{n}$ via the application of an external
gate-voltage. Increasing $E_{J}$ will evidently introduce fluctuations
in $n_{i}$ due to the transfer of Cooper-pairs between neighboring
islands, and $n_{i}$ is no longer a good quantum number. However,
we expect that in the experimentally interesting regime $E_{J}/E_{0}\sim1$,
where $E_{0}$ is the characteristic charging energy in the island,
fluctuations $\Delta n_{i}\equiv n_{i}-\bar{n}$ will be of order
$\Delta n_{i}\simeq\pm1$, and that all other charging states such
that $\left|\Delta n_{i}\right|\gg1$ will be energetically forbiden.
We therefore truncate those states from our description and focus
on charge-fluctuations of $\Delta n_{i}=\pm1$. In terms of a continuous
field $\phi\left(x\right)$, which is slowly varying on the scale
of $a$, the superfluid density in this effective model can be more
conveniently written as\citep{Haldane1981}
\begin{align}
\rho\left(x\right) & =\left[\rho_{0}-\frac{1}{\pi}\nabla\phi\left(x\right)\right]\sum_{p}e^{i2p\left(\pi\rho_{0}x-\phi\left(x\right)\right)},\label{eq:density_bosonized}\end{align}
where the parameter $\rho_{0}$ is defined as $\rho_{0}\equiv1/a$
in the commensurate case. Note that $\rho_{0}$ is an effective parameter
of our model, and cannot be interpreted as the total physical density,
in contrast to truly 1D systems.\citep{giamarchi_book_1d} Only the
\textit{fluctuations} $\delta\rho\left(x\right)\equiv\rho\left(x\right)-\rho_{0}$
have a physical meaning in our model.
In order to obey the phase-number commutation relations in the BCS-groundstate,\citep{Tinkham}
note that the field $\phi\left(x\right)$ must verify the new commutation
relation\begin{align}
\left[\theta\left(x\right),\nabla\phi\left(x^{\prime}\right)\right] & =i\pi\delta\left(x-x^{\prime}\right).\label{eq:commutation_relation_theta_phi}\end{align}
The contribution in squared brackets in Eq. (\ref{eq:density_bosonized})
describe long-wavelength density fluctuations around the average value
$\rho_{0}$, while in the last term, each contribution describes low-energy
density fluctuations of momentum $k\sim2p\rho_{0}$, where $p$ is
an integer. When replaced into the effective action Eq. (\ref{eq:S_2nd_order})
we obtain the following effective model\begin{align}
S_{\text{eff}}\left[\phi\right] & =S_{0}\left[\phi\right]+S_{1}\left[\phi\right]+S_{2}\left[\phi\right],\label{eq:S_eff}\end{align}
where the contribution $S_{0}$ corresponds to a Luttinger liquid
model\citep{giamarchi_book_1d} \begin{align}
S_{0}\left[\phi\right] & =\frac{1}{2\pi\beta L}\sum_{k,\omega_{m}}\left[\frac{\omega_{m}^{2}}{uK}+\frac{uk^{2}}{K}+\frac{\eta\left|\omega_{m}\right|\left|k\right|}{2\pi c}\right]\left|\phi\left(k,\omega_{m}\right)\right|^{2}.\label{eq:S_0}\end{align}
resulting from the slow fluctuations of the density $\delta\rho\left(x\right)\sim-\nabla\phi\left(x\right)/\pi$
and from the hydrodynamic (i.e., $\left\{ k,\omega_{m}\right\} \rightarrow0$)
sector of $v_{\textrm{eff}}\left(k,\omega_{m}\right)$. $u$ and $K$
are respectively the velocity of the 1D plasmon and the interaction
Luttinger parameter
\begin{align}
K & \equiv\pi\sqrt{\frac{E_{J}}{E_{0}}},\label{eq:parameter_K}\\
u & \equiv a\sqrt{E_{J}E_{0}},\label{eq:parameter_u}\end{align}
where $E_{0}\equiv e^{2}/2C_{0}$ is the charging energy with respect
to the ground, with $C_{0}=\epsilon_{\text{r}}a/4\ln\left(2d/a\right)$
the effective ground capacitance of the Josephson junction. In our
treatment, due to the screening provided by the 2DEG, the static effective
potential $v_{\textrm{eff}}\left(k,0\right)$ is effectively short-ranged
for distances $x\gg d$, and therefore the Luttinger parameter $K$
is a constant.\citep{giamarchi_book_1d}
\footnote{Indeed, in the static limit $\omega_{m}=0$, the effective screening
potential writes $v_{\textrm{scr}}\left(k,0\right)=2e^{2}K_{0}\left(2kd\right)/\epsilon_{\text{r}}$
and in the limit $kd\rightarrow0$ compensates the logarithmic divergence
of $v_{\textrm{1D}}\left(k,0\right)$ {[}cf. Eq. (\ref{eq:v_eff}){]},
and yields $\lim_{k\rightarrow0}v_{\textrm{eff}}\left(k,0\right)=2e^{2}\ln\left(2d/a\right)/\epsilon_{\text{r}}$.\citep{Lobos10_Dissipative_phase_fluctuations}
}In terms of the capacitance matrix $C_{ij}$ of the 1DJJA, this amounts
to neglecting the interjunction capacitance $C$, since this contribution,
although relevant for density fluctuations of momentum $k\sim a^{-1}$,
drops off in the long-wavelength sector $k\rightarrow0$ (i.e. the
interaction is screened in a length $L_{\text{scr}}\sim a\sqrt{C/C_{0}}$).\citep{fazio01_review_superconducting_networks}
In the present context, the Luttinger parameter $K$ physically represents
the competition between coherence and charging effects in the array
{[}cf. Eq. (\ref{eq:parameter_K}){]}. Therefore, a large parameter
$K$ favors superconducting correlations, while a small value of $K$
tends to destroy superconductivity due to strong charging effects.\citep{fazio01_review_superconducting_networks}
The dissipative parameter $\eta$ is defined as\begin{align}
\eta & \equiv\frac{c}{\epsilon_{\text{r}}8\pi}\frac{R_{\square}}{R_{Q}},\label{eq:eta_parameter}\end{align}
where $R_{\square}$ is the sheet-resistance of the 2D film and $c$
is a numerical constant of order $c\sim\mathcal{O}\left(1\right)$.
Eq. (\ref{eq:S_0}) with a non-vanishing $\eta$ describes a 1D plasmon-mode
with a finite lifetime $\Gamma\sim\left|k\right|/\eta$.\citep{Lobos10_Dissipative_phase_fluctuations}
Physically, a broadening of the 1D plasma mode occurs due to coupling
to the diffusive modes in the 2DEG. The term $\sim\eta\left|\omega_{m}\right|\left|k\right|$
in Eq. (\ref{eq:S_0}) is the result of combining the leading contribution
in powers of $\left|\omega_{m}\right|/Dk_{\text{TF}}\left|k\right|$
(with $k_{\text{TF}}$ the Thomas-Fermi momentum in the 2DEG) in the
expansion of the retarded potential $v_{\text{eff}}\left(k,\omega_{m}\right)$,
and the long-wavelength fluctuations of the density $\sim\left(\nabla\phi\right)^{2}$,
which contributes a term $\sim k^{2}\left|\phi\left(k,\omega_{m}\right)\right|^{2}$
in Fourier representation. Note that since the scaling dimension of
the term $\sim\left|k\right|\left|\omega_{m}\right|$ is $2$, the
critical properties of the system are not modified. Moreover, for
a metallic plane {[}cf. Ref. \onlinecite{rimberg97_dissipation_driven_sit_2D_josephson_array}{]}
with $R_{\square}\sim0.1R_{Q}$, $\eta\simeq10^{-2}\ll1$ and it can
be effectively ignored, allowing us to write\begin{align}
S_{0}\left[\phi\right] & \simeq\frac{1}{2\pi\beta L}\sum_{k,\omega_{m}}\left[\frac{1}{uK}\omega_{m}^{2}+\frac{u}{K}k^{2}\right]\left|\phi\left(k,\omega_{m}\right)\right|^{2}.\label{eq:S_0-1}\end{align}
This assumption greatly simplifies the analysis, since the action
$S_{0}$ recovers Lorentz-invariance in space-time.
The next term $S_{1}$ in Eq. (\ref{eq:S_eff}) originates in the
phenomenological potential $V_{l}\left(x\right)$, which has the same
periodicity of the array. Therefore, it can be decomposed in Fourier
components as $V_{l}\left(x\right)=\sum_{n}V_{n}\cos\left(Qnx\right),$
with $Q=2\pi/a$. In general, all terms other than $p=n=0$ in Eq.
(\ref{eq:density_bosonized}) are rapidly oscillating and vanish under
the integral sign. However, if $Qn=2\pi p\rho_{0}$, or equivalently
$\rho_{0}a=n/p$ (i.e., the average density of bosons is commensurate
with the lattice), then the term $\sim\int dx\; V_{l}\left(x\right)\rho\left(x\right)$
yields a term $e^{-i\left(Qn-2\pi p\rho_{0}\right)x}=1$ which is
not oscillating and, in addition to the term $p=n=0$, we have the
additional term \begin{align}
S_{1}\left[\phi\right] & =-\frac{\lambda}{a\tau_{0}}\int dxd\tau\;\cos\left(2\phi\left(x,\tau\right)\right),\label{eq:S_1}\end{align}
where we have only kept the most important commensurability ($p=1$),
and where we have defined the dimensionless parameter $\lambda$ \begin{align}
\lambda & \equiv V_{1}\tau_{0}.\label{eq:lambda_parameter}\end{align}
\begin{comment}
Note that $\lambda$ is dimensionless. Then the phase-slips rate is
\begin{align*}
\Gamma & \equiv\frac{\lambda}{\tau_{0}}=\frac{a\rho_{0}V_{1}}{\hbar}.\end{align*}
Then \begin{align*}
\lambda & =\Gamma\tau_{0}\\
& =\frac{\Gamma a}{u}\\
& =\frac{\hbar\Gamma}{\sqrt{E_{J}E_{C}}}\\
& =\frac{10^{-34}J.s\times10^{9}s^{-1}}{K.k_{B}\sqrt{3.3\times0.17}}\\
& =\frac{10^{-34}J.s\times10^{9}s^{-1}}{K.1.38\times10^{-23}J.K^{-1}\sqrt{3.3\times0.17}}\\
& =\frac{10^{-34}\times10^{9}}{1.38\times10^{-23}\sqrt{3.3\times0.17}}\\
& \simeq10^{-34+23+9}\sim0.01\end{align*}
\end{comment}
{}and the short-time cutoff $\tau_{0}\equiv a/u$. The term $V_{0}$
can be reabsorbed in a redefinition of the chemical potential of the
externally imposed charge $\bar{n}$, so in the following we will
not consider it. Physically, the dimensionless parameter $\lambda$
is related to the QPS rate in the Josephson junction by $\Gamma_{\text{QPS}}=\lambda/\tau_{0}$.
Estimated experimental values for $\Gamma_{\text{QPS}}$ are in the
order of $\sim1\;\text{GHz}$\citep{Manucharyan10_Coherent_QPS_in_JJA},
which yields $\lambda\simeq0.06$.
The final term in Eq. (\ref{eq:S_eff}) comes from the dissipative
part of $v_{\textrm{scr}}\left(k,\omega_{m}\right)$. Due to the strongly
oscillating factors $\sim e^{-i2\pi p\rho_{0}x}$ in Eq. (\ref{eq:density_bosonized}),
it results in the local dissipative term \begin{align}
S_{2}\left[\phi\right] & =-\frac{\eta}{a}\int dxd\tau d\tau^{\prime}\;\sum_{p>0}\frac{1}{p}\frac{\cos2p\left[\phi\left(x,\tau\right)-\phi\left(x,\tau^{\prime}\right)\right]}{\left(\tau-\tau^{\prime}\right)^{2}},\label{eq:S_2}\end{align}
This contribution is consistent with that of Ref. \onlinecite{cazalilla06_dissipative_transition},
obtained in the context of Luttinger liquids capacitively coupled
to diffusive metals. In spite of the small magnitude of $\eta$, we
will show that this contribution has important consequences for the
critical properties of the 1DJJA, in contrast to the term proportional
to $\eta$ in Eq. (\ref{eq:S_0}).
In the following we study the critical properties and phases of the
model obtained in Eq. (\ref{eq:S_eff}).
\section{\label{sec:phase-diagram}Phase diagram}
\subsection{\label{sub:RG}Weak-coupling renormalization group analysis}
We first focus on the phases of the 1DJJA at $T=0$. To that end,
we perform a weak-coupling renormalization group (RG) analysis of
the model Eq. (\ref{eq:S_eff}), assuming that $S_{1}$ and $S_{2}$
in Eqs. (\ref{eq:S_1}) and (\ref{eq:S_2}) respectively are weak
perturbations to the Luttinger liquid $S_{0}$ in Eq. (\ref{eq:S_0-1}).
Since the action $S_{0}$ is Lorentz-invariant in space and imaginary
time, we adopt an RG procedure that rescales homogenously space and
time. As usual, we assumme that the original theory is defined up
to a certain momentum cutoff $\Lambda\left(l\right)=\Lambda_{0}e^{-l}$
(with $\Lambda_{0}\sim a^{-1}$), and we study how the action $S_{0}$
is renormalized upon integration of high-energy modes in a window
between $\Lambda\left(l\right)/s<\left|\begin{gathered}\mathbf{q}\end{gathered}
\right|<\Lambda\left(l\right)$, with $s=e^{dl}$, where we have employed the compact notation $\mathbf{q}\equiv\left\{ k,-\frac{\omega_{m}}{u}\right\} $
and $\mathbf{x}\equiv\left\{ x,u\tau\right\} $.
We obtain the perturbative RG-flow equations of the model by performing
a one-loop correction in $S_{2}$ and a two-loop correction in $S_{1}$,
and requiring that the term $S_{0}$ is invariant upon scaling.\citep{Shankar}
We obtain the RG-flow equations
\begin{align}
\frac{dK\left(l\right)}{dl} & =\left[-2\pi\eta\left(l\right)-\left(2\pi\right)^{2}K\left(l\right)\lambda^{2}\left(l\right)C\right]K^{2}\left(l\right),\label{eq:RG_eq_K}\\
\frac{du\left(l\right)}{dl} & =-2\pi\eta\left(l\right)u\left(l\right)K\left(l\right),\label{eq:RG_eq_u}\\
\frac{d\lambda\left(l\right)}{dl} & =\left[2-K\left(l\right)\right]\lambda\left(l\right),\label{eq:RG_eq_lambda}\\
\frac{d\eta\left(l\right)}{dl} & =\left[1-2K\left(l\right)\right]\eta\left(l\right),\label{eq:RG_eq_eta}\end{align}
where the numerical constant $C$ is of order unity.
Note that both $S_{1}$ and $S_{2}$ tend to destroy superconducting
correlations in the Luttinger liquid phase, a fact that is reflected
in Eq. (\ref{eq:RG_eq_K}) where the Luttinger parameter $K\left(l\right)$
is renormalized to \textit{smaller} values, meaning that charging
effects are enhanced. This can be interpreted as an effective increase
of the charging energy $E_{0}$ in Eq. (\ref{eq:parameter_K}). In
addition, since $S_{2}$ is the only term that breaks the Lorentz
invariance of the theory, note that the plasmon velocity $u\left(l\right)$
is proportional only to $\eta\left(l\right)$, and is independent
of $\lambda\left(l\right)$.
\begin{figure}[h]
\includegraphics[clip,scale=0.9]{rg_flow}\caption{Schematic phase diagram of the 1DJJA in the $K-\eta$ plane, obtained
from the integration of the RG-flow Eqs. (\ref{eq:RG_eq_K})-(\ref{eq:RG_eq_eta}),
with the initial parameter $\lambda_{0}\equiv\lambda\left(l=0\right)=0.01$.
An increase of $R_{\square}$ in the 2DEG, and consequently, of the
dissipative parameter $\eta$, can induce a SIT. Note that, in absence
of dissipation, the critical value $K_{c}=K_{c}\left(\lambda_{0}\right)\simeq2.1$
is slightly shifted with respect to the value $K_{c}\left(\lambda_{0}\rightarrow0\right)\rightarrow2$.\label{fig:phase_diagram}}
\end{figure}
When $K\left(l\right)<2$, the perturbative parameter $\lambda\left(l\right)$
flows to strong-coupling {[}cf. Eq. (\ref{eq:RG_eq_lambda}){]}, and
the perturbative RG procedure is no longer valid. In the limit $\eta\rightarrow0$
we recover the usual Mott-transition of the BKT-type described by
the sine-Gordon model, and below the critical value $K_{c}=2$, the
1DJJA is in the insulating phase.\citep{bradley_josephson_chain,glazman_josephson_1d,fazio01_review_superconducting_networks,giamarchi_book_1d}
Using Eq. (\ref{eq:parameter_K}), this means that in absence of dissipation,
the SIT occurs for $E_{J}/E_{0}=\left(2/\pi\right)^{2}$.\citep{fazio01_review_superconducting_networks}
Note that our situation corresponds strictly to the case when the
superfluid density in the 1DJJA is commensurate to the lattice, and
is in clear distinction to the non-commensurate situation (i.e., $\lambda=0$),
where dissipation (i.e., the term $S_{2}$) becomes relevant for $K\left(l\right)<1/2$,
inducing a different kind of non-superconducting groundstate.\citep{cazalilla06_dissipative_transition}
In the present case, the scaling dimension of the dissipative parameter
$\eta\left(l\right)$ is always smaller than that of $\lambda\left(l\right)$,
which means that for $K\left(l\right)\simeq2$, $S_{1}$ is a stronger
perturbation as compared to $S_{2}$. Therefore, one would expect
the nature of the non-superconducting groundstate to be determined
essentially by $S_{1}$. However, based on this fact, one could naively
conclude that the term $S_{2}$ is unimportant near the SIT, a conclusion
we prove incorrect. In fact, a more detailed analysis reveals the
importance of the term $S_{2}$ near the SIT. Physically, the coupling
to the diffusive degrees of freedom in the 2DEG quenches charge-fluctuations
in the 1DJJA, resulting in an enhanced effective charging energy $E_{0}^{*}$.
This phenomenon is more precisely described by the RG-flow equation
for $K\left(l\right)$ {[}cf. Eq. (\ref{eq:RG_eq_K}){]}, where $K\left(l\right)$
is renormalized to \textit{lower} values by $\eta\left(l\right)$.
Indeed, near the SIT, a small increase in the initial value $\eta_{0}\equiv\eta\left(l=0\right)$
(i.e., an increase in $R_{\square}$) can effectively control the
RG-flow of $K\left(l\right)$ and therefore, that of $\lambda\left(l\right)$,
inducing the SIT. We illustrate this point in Fig \ref{fig:phase_diagram},
where the schematic phase diagram obtained by integration of the RG-flow
Eqs. (\ref{eq:RG_eq_K})-(\ref{eq:RG_eq_eta}), with initial parameter
$\lambda_{0}\equiv\lambda\left(l=0\right)=0.01$. Note the stabilization
of the insulating groundstate due to Ohmic dissipation induced by
the coupling to the 2DEG.
In a first approximation, this effect is similar to the dissipation-driven
SIT observed in 2DJJAs capacitively coupled to a diffusive 2DEG.\citep{rimberg97_dissipation_driven_sit_2D_josephson_array,Wagenblast97_SIT_in_a_Tunable_Dissipative_Environment,Vishwanath04_Screening_and_dissipation_at_the_SIT_induced_by_a_metallic_ground_plane}
However, important differences appear with respect to the 2D case.
In that case, it was argued that dissipation produced a renormalization
of the effective parameters $E_{J}$ and $E_{C}$ of the array due
to the incomplete screening of the Coulomb interaction in a certain
frequency-regime.\citep{Wagenblast97_SIT_in_a_Tunable_Dissipative_Environment}
Physically, the slow diffusive response of the 2DEG cannot follow
the faster dynamics of the 2D plasma mode, and cannot screen it efficiently.
However, in the 1D geometry the 1D plasmon is effectively very well
screened by the 2DEG\citep{Lobos10_Dissipative_phase_fluctuations},
and it could be naively concluded that no dissipation-driven SIT should
be observed. However, this screening effect is compensated by the
presence of strong backscattering occuring in 1D {[}i.e., action $S_{2}$,
Eq. (\ref{eq:S_2}){]}, and originated in the retarded interaction
$v_{\text{eff}}\left(x,\tau\right)$. The net result is that the dissipation-driven
SIT is restored in 1D.
Although one expects the nature of the non-superconducting groundstate
to be of the Mott-insulating type, by analogy with the well-known
results for the sine-Gordon model\citep{bradley_josephson_chain,glazman_josephson_1d,fazio01_review_superconducting_networks,giamarchi_book_1d},
strictly speaking we cannot extrapolate the results in this Section
to the strong-coupling situation, and a different method is needed
in that regime.
\subsection{\label{sub:SCHA}Self-consistent harmonic approximation}
To gain more insight into the phase in which the parameter $\lambda\left(l\right)$
flows to strong-coupling, in this Section we make use of the variational
self-consistent harmonic approximation\citep{feynman_statmech}. This
method consist in finding the optimal propagator $g_{\text{tr}}^{-1}\left(\mathbf{q}\right)$
of a Gaussian trial action of the 1DJJA
\begin{align}
S_{\text{tr}} & =\frac{1}{2\beta L}\sum_{\mathbf{q}}g_{\text{tr}}^{-1}\left(\mathbf{q}\right)\left|\phi_{\mathbf{q}}\right|^{2},\label{eq:S_tr}\end{align}
where $\phi_{\mathbf{q}}$ is the Fourier transform of $\phi\left(x,\tau\right)$.
Here we have introduced the compact notation $\mathbf{q}=\left(k,-\omega_{m}/u\right)$.
The idea is to minimize the variational free-energy
\begin{eqnarray}
F_{\text{var}} & \equiv & F_{\text{tr}}+\frac{1}{\beta}\left\langle S_{\text{eff}}-S_{\text{tr}}\right\rangle _{\text{tr}},\label{eq:F_var}\end{eqnarray}
where the {}``trial'' free-energy $F_{\text{tr}}$ is \begin{align}
F_{\text{tr}} & =-\frac{1}{2\beta}\sum_{\mathbf{q}}\log\left[\beta Lg_{\text{tr}}\left(\mathbf{q}\right)\right]\label{eq:F_tr}\end{align}
The factor $1/2$ in Eqs. (\ref{eq:S_tr}) and (\ref{eq:F_tr}) come
from the constraint $\phi^{*}\left(\mathbf{q}\right)=\phi\left(-\mathbf{q}\right)$
since $\phi\left(x,\tau\right)$ is a real field, a fact that reduces
the number of independent degrees of freedom .
\begin{figure}[h]
\includegraphics[bb=40bp 85bp 420bp 340bp,clip,scale=0.6]{scha}\caption{Dimensionless SCHA parameters $\Delta$ and $\zeta$ as a function
of $K$, calculated for the parameters $\eta=0.01$, $\lambda=0.05$.
We obtain non-vanishing values only in the region $K\lesssim2$ (note
the abrupt increase in that region), consistent with the results of
the RG-analysis. \label{fig:scha}}
\end{figure}
The minimization of $F_{\text{var}}$ in Eq. (\ref{eq:F_var}) with
respect to $g_{\text{tr}}\left(\mathbf{q}\right)$ yields the self-consistent
equation for $g_{\text{tr}}\left(\mathbf{q}\right)$\begin{align}
g_{\text{tr}}^{-1}\left(\mathbf{q}\right) & =\frac{1}{\pi uK}\omega_{m}^{2}+\frac{u}{\pi K}k^{2}+\frac{4\lambda}{a\tau_{0}}e^{-\frac{2}{\beta L}\sum_{\mathbf{q}^{\prime}}g_{\text{tr}}\left(\mathbf{q}^{\prime}\right)}\nonumber \\
& -\frac{8\eta}{a}\int_{0}^{\beta}d\tau\;\frac{\cos\left(\omega_{m}\tau\right)-1}{\tau^{2}}\times\nonumber \\
& \times e^{-4\frac{1}{\beta L}\sum_{\mathbf{q}^{\prime}}\left[1-\cos\left(\omega_{m}^{\prime}\tau\right)\right]g_{\text{tr}}\left(\mathbf{q}^{\prime}\right)}.\label{eq:self_consistent_scha}\end{align}
In general, the solution of this equation has to be found numerically.
However, for small $\lambda$ and $\eta$ the analytical solution
\begin{align}
g_{\text{tr}}^{-1}\left(\mathbf{q}\right) & =g_{LL}^{-1}\left(\mathbf{q}\right)+\frac{\zeta}{a}\left|\omega_{m}\right|+\frac{\Delta}{a\tau_{0}},\label{eq:g0_combined}\end{align}
is obtained. Here $g_{LL}^{-1}\left(\mathbf{q}\right)=\frac{1}{\pi uK}\omega_{m}^{2}+\frac{u}{\pi K}k^{2}$
is the Luttinger liquid propagator, corresponding to the action Eq.
(\ref{eq:S_0}) (for $\eta=0$). Physically, this propagator describes
an insulator (given by a non-vanishing gap or {}``mass'' term $\Delta$)
with Ohmic-dissipative dynamics (encoded in a non-vanishing $\zeta$).
Note that dissipation dominantes for frequencies $\left|\omega_{m}\right|>\Delta/\zeta\tau_{0}$.
These parameters are found solving the following set of non-linear
equations \begin{align}
\zeta & =8\pi\eta\left(\frac{\zeta K\pi+2\sqrt{K\pi\Delta}}{4}\right)^{2K},\label{eq:scha_zeta}\\
\Delta & =4\lambda\left(\frac{\zeta K\pi+2\sqrt{K\pi\Delta}}{4}\right)^{K},\label{eq:scha_Delta}\end{align}
obtained replacing the solution Eq. (\ref{eq:g0_combined}) back into
Eq. (\ref{eq:self_consistent_scha}). Starting from the self-consistent
solution of Eqs. (\ref{eq:scha_zeta}) and (\ref{eq:scha_Delta})
for $\Delta$ in absence of dissipation (i.e. $\eta=0$), we can study
the regime $\eta\ll\lambda\ll1$ perturbatively in $\eta$, and we
obtain the following estimate for the gap increase due to dissipative
effects \begin{align}
\delta\Delta & \simeq2\pi^{2}\frac{\eta K\Delta_{0}^{2}}{\lambda}.\label{eq:gap_increase}\end{align}
This result is consistent with the fact that dissipation in the density
(i.e., field $\phi$) quenches charge-fluctuations and therefore favors
an insulating groundstate.
In Fig. \ref{fig:scha} we show numerical results for $\Delta$ and
$\zeta$ as a function of $K$ for the values $\lambda=0.05$ and
$\eta=0.0\text{1}$. Note the sharp increase of both $\Delta$ and
$\zeta$ for $K<2$. This result is consistent with the RG-analysis,
which predict the breakdown of the Luttinger liquid phase for $K<2$
in the weak-coupling regime. Within the SCHA, the physics of the strong-coupling
fixed point is encoded in non-vanishing values of $\zeta$ and $\Delta$,
providing a complementary description to the RG-analysis.
\section{\label{sec:transport}Transport properties}
In this section we concentrate on the dc-conductivity of the 1DJJA,
a quantity of central interest in experiments.\citep{schoen90_review_ultrasmall_tunnel_junctions,fazio01_review_superconducting_networks}
We first focus on the current-density $j\left(x\right)$. Since the
field $\nabla\theta\left(x\right)/\pi$ is the momentum of Cooper-pairs
{[}cf. Eq. (\ref{eq:H_JJA_continuum}){]}, the usual minimal coupling
procedure $\nabla\theta\left(x\right)/\pi\rightarrow\left[\nabla\theta\left(x\right)-2eA\left(x\right)\right]/\pi$
(with $e$ the electron charge and $A$ the vector potential) in Hamiltonian
Eq. (\ref{eq:H_JJA_continuum}) allows to obtain the current as $j\left(x\right)\equiv-\delta H_{JJA}/\delta A\left(x\right)$.
In our problem, it explicitly reads \citep{giamarchi_book_1d} \begin{align}
j\left(x\right) & =uK\left(\frac{2e}{\pi}\right)\left[\nabla\theta\left(x\right)-2eA\left(x\right)\right].\label{eq:current_density}\end{align}
The conductivity along the wire is obtained from the Kubo formula
\citep{mahan2000,giamarchi_book_1d} \begin{align}
\sigma\left(\omega\right) & \equiv\frac{\chi_{jj}^{R}\left(0,\omega\right)}{i\left(\omega+i\delta\right)},\label{eq:conductivity}\end{align}
where $\chi_{jj}^{R}\left(k,\omega\right)\equiv\lim_{i\omega_{m}\rightarrow\omega+i\delta}\chi_{jj}\left(\mathbf{q}\right)$
is the retarded current-current correlation function and $\chi_{jj}\left(\mathbf{q}\right)\equiv\left\langle j^{*}\left(\mathbf{q}\right)j\left(\mathbf{q}\right)\right\rangle =\left.\delta^{2}\ln Z/\delta A\left(\mathbf{q}\right)\delta A^{*}\left(\mathbf{q}\right)\right|_{A=0}$
is the current-current correlation function obtained in the linear-response
regime. It is convenient to express this correlator as $\chi_{jj}\left(\mathbf{q}\right)=\chi_{jj}^{d}+\chi_{jj}^{p}\left(\mathbf{q}\right)$,
where $\chi_{jj}^{d}\equiv-\left(2e\right)^{2}uK/\pi$ is the diamagnetic
contribution and
\begin{align}
\chi_{jj}^{p}\left(\mathbf{q}\right) & \equiv\left(\frac{2e}{\pi}\right)^{2}\left(uK\right)^{2}k^{2}\left\langle \theta\left(\mathbf{q}\right)\theta\left(-\mathbf{q}\right)\right\rangle ,\label{eq:xi_p}\end{align}
is the paramagnetic term.\citep{mahan2000} In absence of current-decaying
mechanisms {[}i.e., $\lambda=\eta=0$ in Eq. (\ref{eq:S_eff_total}){]},
the conductivity writes\begin{align*}
\sigma_{0}\left(\omega\right) & =\frac{\left(2e\right)^{2}}{\hbar}uK\left[\delta\left(\omega\right)+i\text{P}\left(\frac{1}{\pi\omega}\right)\right],\end{align*}
where we have restored the Planck constant and where we have used
that $\chi_{jj}^{p}\left(\mathbf{q}\right)\rightarrow0$ in the limit
$k=0$.\citep{giamarchi_book_1d} Note that the real part of $\sigma_{0}\left(\omega\right)$
consists of a Drude-peak at $\omega=0$, as expected for a superconductor.
This result can be understood from the fact that the total charge
current $J_{e}=\int dx\; j\left(x\right)$ is a conserved quantity
in absence of QPS and dissipation processes, i.e., it commutes with
the hamiltonian $H_{\text{JJA}}$.
The effect of a finite $\eta$ in the Gaussian sector of the theory
{[}cf. Eq. (\ref{eq:S_0}){]} has been studied in Ref. \onlinecite{Lobos10_Dissipative_phase_fluctuations},
and produces a broadening of the plasmon peak, whose width $\Gamma$
vanishes as $\Gamma\sim\left|k\right|$. Consequently, only taking
into account this effect, a well-defined Drude-peak in $\sigma\left(\omega\right)$
for $\omega=0$ is recovered, and the system should behave as a perfect
conductor.
Let us now study the effects of the terms $S_{1}$ and $S_{2}$. When
$\lambda$ and $\eta$ are irrelevant perturbations (in the RG sense),
their effects on the conductivity can be studied within the theoretical
framework of the memory function formalism.\citep{gotze_fonction_memoire}
In this approach, the central assumption is that the Kubo formula
for the conductivity Eq. (\ref{eq:conductivity}) can be recasted
as\citep{giamarchi_book_1d} \begin{align}
\sigma\left(\omega,T\right) & =\frac{i\left(2e\right)^{2}}{\pi\hbar}\frac{uK}{\omega+M\left(\omega,T\right)},\label{eq:sigma_omega}\end{align}
where $M\left(\omega,T\right)$ (i.e., the memory function) is a meromorphic
function depending on the terms in the Hamiltonian responsible for
degrading the current, and hence producing a finite resistivity. Current-decay
originated in QPS and in the coupling to the dissipative modes in
the 2DEG induce finite resistivity in the 1DJJA for all temperatures
$T<T_{c}$. In particular for temperatures $T\ll T_{c}$, and perturbatively
in $\lambda$ and $\eta$, we obtain\begin{align}
\varrho\left(T\right) & =\frac{\hbar}{a\left(2e\right)^{2}}\left[A_{1}T^{2K-3}+A_{2}T^{2K}\right]\label{eq:resistivity}\end{align}
where
\begin{align}
A_{1} & \equiv\lambda^{2}4\pi^{3}\left[\cos\left(\frac{\pi K}{2}\right)B\left(\frac{K}{2},1-K\right)\right]^{2}\left(\frac{2\pi a}{u}\right)^{2K-3},\label{eq:A_1}\\
A_{2} & \equiv\eta32\pi^{3}\cos\left[\left(1+K\right)\pi\right]B\left[1+K,-1-2K\right]\left(\frac{2\pi a}{u}\right)^{2K},\label{eq:A_2}\end{align}
where the function $B\left(x,y\right)$ is defined as $B\left(x,y\right)\equiv\Gamma\left(x\right)\Gamma\left(y\right)/\Gamma\left(x+y\right)$,
and $\Gamma\left(x\right)$ is the standard Euler's Gamma function.\citep{abramowitz_math_functions}
The term $\sim T^{2K-3}$ in Eq. (\ref{eq:resistivity}) is the contribution
due to QPS processes, consistent with former theoretical predictions.\citep{giamarchi_attract_1d,zaikin97}
The second term $\sim T^{2K}$ originates in backscattering effects
induced by dissipation, and is consistent with the behavior predicted
by Cazalilla \textit{et al.}\citep{cazalilla06_dissipative_transition}
This last effect can be interpreted as a frictional drag produced
by the diffusive modes in the 2DEG.\citep{rojo99_coulomb_drag_review}
Note that at lowest order in $\lambda$ and $\eta$, the two contributions
add up independently, indicating that for temperatures $T^{*}<T\ll T_{c}$,
where $T^{*}\equiv\sqrt[3]{A_{1}/A_{2}}/2\pi\tau_{0}$, the resistivity
in the 1DJJA is dominated by frictional drag, while for $T<T^{*}\ll T_{c}$
the effect of QPS takes over.
The non-trivial effects due to the renormalization of the bare couplings
can be taken into account integrating the RG-flow Eqs. (\ref{eq:RG_eq_K})-(\ref{eq:RG_eq_eta}),
and injecting them in the above Eqs. (\ref{eq:resistivity}), (\ref{eq:A_1})
and (\ref{eq:A_2}). We integrate the RG-flow up to the scale given
by the temperature $a\left(l\right)=a\left(0\right)e^{l}=u\left(l\right)/2\pi T$,
and we use formula Eq. (\ref{eq:resistivity}) with the parameters
of the model calculated at the scale $a\left(l\right)$. This allows
to obtain $\varrho\left(T\left(l\right)\right)$ vs $T\left(l\right)$.
In Fig. \ref{fig:resistivity_log_log} we show the resistivity $\varrho\left(T\right)$
of the 1DJJA, calculated for different values of the parameter $K$
and using the estimations for the bare parameters $\lambda_{0}=0.01$
and $\eta_{0}=0.01$. The results are normalized to a {}``high-temperature''
resistivity $\varrho\left(T_{0}\right)$, where $T_{0}\simeq a\left(0\right)/u=\tau_{0}$,
represents a high-temperature cutoff in the theory (e.g., $T_{c}$).
\begin{figure}[h]
\includegraphics[bb=25bp 450bp 500bp 820bp,clip,scale=0.47]{resistivity_log_log}\caption{\label{fig:resistivity_log_log}Dc-resistivity $\varrho\left(T\right)$
of the 1DJJA, normalized to a {}``high-temperature'' value $\varrho\left(T_{0}\right)$,
as a function of $T/T_{0}$, calculated for the parameters $\lambda_{0}=0.01$
and $\eta_{0}=0.01$, and for different values of $K=\pi\sqrt{E_{J}/E_{0}}$.
A low-temperature upturn of $\varrho\left(T\right)$ signals the formation
of the insulating phase.}
\end{figure}
Note that for the values $K=2.5$ and $K=2.3$, the resistivity shows
a monotonically decreasing behavior, indicating a superconducting
groundstate and consistent with the RG-analysis of Sec. \ref{sub:RG}.
We also note a small kink around $T^{*}\sim0.4\; T_{0}$, signalling
the aforementioned crossover from dissipation-dominated to QPS-dominated
resistivity. For $K=2.1$, the resistivity first decreases and then
shows a low-temperature upturn, indicating that the array is near
the quantum critical point $K_{c}$. Finally, for lower values of
$K$, the insulating behavior in the 1DJJA is clear. Since both the
integration of the RG-flow equations and the calculation of the memory-function
formulas are perturbative in $\lambda$ and $\eta$, the calculation
of the resistivity must be stopped whenever $\lambda\left(l\right)$
or $\eta\left(l\right)$ become of order unity.
In Fig. \ref{fig:resistivity_eta}, we show the resistivity as a function
of $T/T_{0}$, calculated for fixed $K=2.3$ and $\lambda=0.01$,
and for different values of parameter $\eta$. We see that for $\eta=0$
(i.e., $R_{\square}=0$ in the 2DEG), the array shows superconducting
behavior, and the resistivity due to QPS processes is well described
by the predicted power-law $\varrho\left(T\right)\sim T^{2\tilde{K}-3}$,
with $\tilde{K}=K\left(l\rightarrow\infty\right)\simeq2.2$ the renormalized
value predicted by Eq. (\ref{eq:RG_eq_lambda}). Upon increasing the
parameter $\eta$, the resistivity of the array increases, developing
the aforementioned kink, but most importantly, the low-temperature
resistivity develops an upturn, indicating a dissipation-driven phase
transition to the insulating phase.
\begin{figure}[h]
\includegraphics[bb=30bp 30bp 450bp 350bp,clip,scale=0.55]{resistivity_eta}\caption{\label{fig:resistivity_eta}Resistivity of the 1DJJA (in units of
$h/e^{2}a$ ) as a function of $T/T_{0}$, calculated for parameters
$K=2.3$ and $\lambda=0.01$ and for different values of $\eta$.
Although in absence of dissipation, the array is in the superconducting
phase, a dissipation-driven SIT occurs upon increasing $\eta$, consistent
with the results in Fig. \ref{fig:phase_diagram}. The curve $\varrho\left(T\right)\sim T^{2K-3}$
is shown for comparison.}
\end{figure}
More insight into the insulating phase can be obtained using the Luther-Emery
refermionization solution for $K=1$.\citep{luther_emery_backscattering,giamarchi_book_1d}
In absence of dissipation (i.e., $\eta=0$) an exact solution is obtained
in terms of non-interacting fermions, with a gap $\tilde{\Delta}\equiv\pi a\lambda$
in their spectrum of excitations. Using the Kubo formula, one obtains
the following expression for the dc-conductivity at low temperatures
$T\ll\tilde{\Delta}$
\begin{align*}
\sigma\left(\omega\right) & \approx\frac{e^{2}}{\hbar}u\sqrt{\frac{2\pi T}{\tilde{\Delta}}}e^{-\tilde{\Delta}/T}\delta\left(\omega\right).\end{align*}
This contribution arises from the excited quasiparticles above the
gap $\tilde{\Delta}$, which have an exponential population at low
enough temperatures. This infinite conductivity occurs because in
absence of dissipation, excited quasiparticles are infinitely long-lived.
Using the memory-function approach for the refermionized problem in
the regime $\eta\ll\lambda,T\ll\tilde{\Delta}$, we find the analytical
result
\begin{align*}
\sigma\left(\omega=0\right) & =\frac{e^{2}}{\hbar}\frac{c_{2}}{\eta}\frac{1}{\tau_{0}^{2}\tilde{\Delta}^{2}}\frac{\tilde{\Delta}}{T}e^{-\tilde{\Delta}/T},\end{align*}
where $c_{2}$ is a numerical coefficient $c_{2}\simeq\mathcal{O}\left(1\right)$.
As expected, dissipation introduced a finite lifetime in the quasiparticles,
and a finite resistivity is obtained at $\omega=0$.
\section{\label{sec:summary}Summary and conclusions}
We have investigated the properties of a linear JJA capacitively coupled
to a diffusive 2DEG placed in close proximity. Using a bosonization
approach, we have derived an effective model for the 1DJJA, and have
obtained its critical properties and phases at $T=0$. Our main result
is the possibility to observe a SIT tuned by the parameter $\eta\sim R_{\square}/R_{Q}$
{[}cf. Eq. (\ref{eq:eta_parameter}){]}. This setup could be used
to investigate the superconductor-insulator transition in a 1DJJA
under better controlled experimental conditions as compared to other
setups used in the past.\citep{Chow98_Length_scale_dependence_of_SIT_in_1D_array_J,Kuo01_Magnetic_induced_transition_in1DJJA,Miyazaki02_QPT_in_1D_arrays_of_JJs,Takahide06_SIT_2D_1D_crossover}
Our work could shed some light on the understanding of other 1D superconducting
systems showing a similar behavior, such as ultra-thin superconducting
wires built by molecular templating\citep{bezryadin00,lau01,Bezryadin08_QPS_review}
or by e-beam lithography\citep{Chen09_Magnetic-Field-Induced_Superconducting_State_in_Zn_Nanowires_Driven_in_the_Normal_State_by_an_Electric_Current}
techniques.
We have shown that besides the more or less trivial static screening
effect, the presence of a 2DEG induces dissipative effects in the
quantum dynamics of the 1DJJA due to backscattering processes induced
by the dynamically screened Coulomb interaction, and explicitly depend
on the sheet-resistance $R_{\square}$ of the 2DEG. These dynamical
effects play an important role in the quantum phase diagram of the
1DJJA. This situation is different from previous approaches in higher
dimensions.\citep{Wagenblast97_SIT_in_a_Tunable_Dissipative_Environment,Vishwanath04_Screening_and_dissipation_at_the_SIT_induced_by_a_metallic_ground_plane}
Indeed, in 1D the plasmon mode is almost statically screened\citep{Lobos10_Dissipative_phase_fluctuations},
and this would lead to the naive conclusion that dynamical effects
are not important. However, a more careful analysis shows that backscattering
originated in the dynamically screened Coulomb potential has the effect
of restoring the SIT. In our system, these dynamical effects have
important consequences for the critical properties of the array, and
should be possible to observe them in dc-transport measurements. Physically,
the coupling to diffusive modes in the metal induces charging effects
which are local in space (i.e., of the order of the lattice parameter
$a$ of the 1DJJA) but which are non-local in time (i.e., Ohmic dissipation
effects), and tend to quench charge fluctuations, rendering superconductivity
weaker.
By the means of a weak-coupling RG-analysis and a variational approach,
we predict a SIT driven by the presence of dissipation in the 2DEG.
This SIT is of the BKT-type and mediated by unbinding of QPS/anti
QPS pairs, like in the dissipationless case.\citep{bradley_josephson_chain}
Near the critical line the effects of QPS are stronger than those
originated in dissipation and results in a SIT. This scenario is corroborated
by a subsequent variational analysis of action Eq. (\ref{eq:S_eff}),
which suggests the formation of a gap $\Delta$ in the spectrum of
excitations of the 1DJJA {[}cf. Eq. (\ref{eq:g0_combined}){]}.
Our results suggest that dissipation renormalizes the QPS-rate to
\textit{higher} values and the ratio $\sqrt{E_{J}/E_{0}}$ to \textit{lower}
values {[}cf. Eqs. (\ref{eq:RG_eq_K})-(\ref{eq:RG_eq_eta}){]}, rendering
superconductivity in the 1DJJA weaker. Eventually, an increase of
$R_{\square}$ {[}and therefore of $\eta$, in view of Eq. (\ref{eq:eta_parameter}){]},
could drive the system into the insulating phase, as can be seen in
Figs. \ref{fig:phase_diagram} and \ref{fig:resistivity_eta}. This
phenomenon is different to the case studied by Cazalilla \textit{et
al.}, where QPS processes were absent, and it was dissipation \textit{itself}
that drove the quantum phase transition for the critical value $K_{c}=1/2$.\citep{cazalilla06_dissipative_transition}
We have also studied the consequences on the temperature-dependent
dc-resistivity of the array $\varrho\left(T\right)$. We have shown
that a non-vanishing $R_{\square}$ induces a rich behavior of $\varrho\left(T\right)$.
In particular in the superconducting phase, where the 1DJJA is in
the Luttinger liquid universality class, and the effects of QPS and
dissipation are perturbative, the resistivity of the array $\varrho\left(T\right)$
follows a power-law behavior $\varrho\left(T\right)=A_{1}T^{\nu_{1}}+A_{2}T^{\nu_{2}}$,
with exponents $\nu_{1}=2K-3$ and $\nu_{2}=2K$ {[}cf. Eq. (\ref{eq:resistivity}){]}
generated by QPS and dissipation, respectively. Therefore, the results
of this paper could be relevant in the interpretation of experimental
results of transport through superconducting circuits subject to dissipative
effects. In the insulating phase, the low-temperature dc-resistivity
is expected to show thermally-activated behavior.\citep{giamarchi_book_1d,fazio01_review_superconducting_networks}
In particular for $K=1$, the resulting model can be studied analytically
with a refermionization approach, and results in a resistivity $\varrho\left(T\right)\sim\eta Te^{\Delta/T}/\Delta$.
Quite importantly, note in this expression that the resistivity depends
also implicitly on $\eta$ via a renormalization of the gap $\eta$.
\begin{acknowledgments}
We acknowledge useful discussions with I. Pop, W. Guichard and F.
Hekking. This work was supported by the Swiss National Foundation
under MaNEP and division II.
\end{acknowledgments}
\bibliographystyle{apsrev}
|
\section{Introduction}
\label{s_intro}
Cosmic rays are widely believed to originate in supernova remnant (SNR)
shock waves, because the cosmic-ray energy spectrum agrees with model predictions,
because power-law distributions of energetic electrons are seen in
SNRs, and because the power required to maintain the cosmic ray population
could be supplied by about 10\% of kinetic energy of Galactic supernovae.
The standard theory for the process is Diffusive Shock Acceleration (DSA),
which is a first order Fermi process requiring that particles scatter
between a gasdynamic subshock and plasma turbulence in a shock precursor.
Evidence for non-linear DSA comes
from curved synchrotron spectra
\citep{reynolds92,vink06,allen}, evidence for high compression factors
\citep{warren05,cassam08} and evidence for lower than expected downstream
temperatures \citep{hughes00,helder09}.
However, all this evidence is based on observation of downstream
properties. The effects of precursor physics on the H$\alpha$ emission
described here offer a direct probe of the properties
of the precursor.
A crucial parameter for these models is the diffusion coefficient
$\kappa$, as it determines the precursor scale length, which is typically
$\kappa$ divided by the
shock speed $V_S$. Gas is compressed in the precursor and accelerated
to a fraction of the shock speed, and this compression is related
to $V_S$, to the efficiency of particle acceleration and to the
escape of energetic particles from the region \citep{bykov05, vink10}.
Neutrals can impede the acceleration
process by damping the turbulence needed to scatter particles
back to the shock. However, \citet{drury} found that the
acceleration efficiency can be high as long as the density and
neutral fraction are not too large, though the maximum particle energy
is reduced.
One set of diagnostics for the physics of collisionless shocks is
based on the emission from particles in the narrow ionization zone
just behind a nonradiative shock \citep{ray91, heng10}. In particular,
H$\alpha$ photons from a nonradiative shock in partly neutral gas originate
very close to the shock, and Coulomb collisions
do not have time to erase such signatures as unequal electron and ion
temperatures or non-Maxwellian velocity distributions \citep{laming96,
ghavam01, ray08, ray10}. In the optical these shocks are seen as pure Balmer line
filaments whose profiles show a narrow component characteristic of
the pre-shock kinetic temperature and a broad component closely
related to the post-shock proton temperature \citep{cr78, heng10,
vanA}. The intensity ratio of the broad and narrow components
is determined by the electron to ion temperature ratio at the shock
\citep{ghavam01, vanA, helder10t}.
The Balmer line profiles also contain signatures of shock precursors.
In general, the narrow component line widths are 40 to 50 $\rm km~s^{-1}$,
indicating temperatures around 40,000 K. If that were the ambient ISM
temperature, there would be no neutrals to create the Balmer line
filament, so the width is interpreted as an indication of heating in
a narrow precursor too thin to completely ionize the hydrogen
\citep{smith94, hester94, lee07, sollerman}. Faint emission
ahead of the sharp filament is interpreted as emission from the
compressed and heated precursor gas \citep{hester94, lee07, lee10}.
This paper considers the role of neutrals in heating the precursor
plasma and computes the properties of precursor H$\alpha$ emission.
While cosmic ray pressure in the precursor can compress, heat and
accelerate ions and electrons by means of plasma turbulence and
magnetic fields, the neutrals only interact with the precursor
by means of collisions with protons and electrons.
If the density is very high, neutrals and protons are
tightly coupled by charge transfer. In that case, the neutrals are compressed
along with the protons and adiabatically heated. They also share
in any other heating of the protons, such as dissipation of Alfv\'{e}n
waves generated by cosmic-ray streaming. On the other hand, if the
density is very low, neutrals pass through the precursor and the
shock without interacting at all, preserving their pre-shock
velocity distribution.
The intermediate case is more complex. A shock that efficiently
accelerates cosmic rays is strongly modified, and gas reaches a
significant fraction of the shock speed in the precursor \citep{vladimirov, wagner09}.
If neutrals and ions are fairly well coupled, they can be described
as fluids whose relative speed gives a frictional heating similar
to that in C-shocks \citep{drainemckee}. If a
neutral encounters this high speed compressed plasma without having
been brought gradually up to speed by many previous charge transfers,
it can be ionized and become a pickup ion \citep{ray08, ohira} like those observed in
the solar wind \citep{moebius85}. It can then have an energy on the order
of 1 keV, which it can share with the other protons. Electron heating
is more uncertain, but it can occur by means of Lower Hybrid waves
\citep{cairnszank}. If the electrons are heated they
can excite and ionize H atoms, changing the H$\alpha$ profile and the
broad-to-narrow line ratio used as an electron temperature diagnostic \citep{ghavam01}.
In this paper we compute the proton, neutral and electron temperatures
in the precursors for a variety of parameters, along with the
ionization and excitation of H atoms. We consider the effects of
these processes on Balmer line diagnostics currently in use.
\citet{ohira} considered the effects of neutrals on the velocity structure of
the precursor, the compression ratio and the acceleration process. They
found that the pickup ions can reduce the compression by the subshock and
enhance proton injection into the acceleration process.
\citet{morlino} self-consistently computed the particle acceleration
and heating due to neutrals, but within the fluid approximation for
both neutrals and ions. In this paper we emphasize the effects on
the H$\alpha$ line profile.
\section{Model Calculations}
\label{s_model}
We parameterize
the precursor structure in a relatively simple manner. We
assume that the precursor accelerates and compresses the interstellar
gas over a length scale $\kappa$/$V_S$, where $\kappa$ is the
diffusion coefficient for cosmic rays near the cutoff.
Effective cosmic-ray acceleration requires $\kappa$ on the
order of $10^{24}~\rm cm^2~s^{-1}$, and estimates based on
the scales of H$\alpha$ precursors are 2 to 4$\times 10^{24}~\rm cm^2~s^{-1}$
\citep{lee07, lee10}. We do not consider the second order
effects of momentum
and energy deposition by the neutrals on
the precursor length scale.
We assume an exponential form, so that the compression is given by
\begin{equation}
\chi ~=~ 1 + (\chi_1 -1) ~~ e^{(x V_S / \kappa)}
\end{equation}
\noindent
where x is negative ahead of the shock and $\chi_1$ is
the compression ratio just upstream of the subshock.
It is related to the fractional pressure of cosmic rays behind the
shock, w=$P_{CR} / (P_G + P_{CR})$, by equation 9 of \citet{vink10}. We simplify this
equation with the
assumption that for w$<$0.8 the compression in the gas subshock equals 4, so that
\begin{equation}
\chi_1~=~ (1-w/4)/(1-w)
\end{equation}
\noindent
Mass conservation implies velocities $V = V_S / \chi$ in the frame of the shock.
To compute the proton and electron temperatures we include
adiabatic compression, Coulomb energy transfer between
protons and electrons, energy losses due to ionization and
excitation of Hydrogen, and heating terms.
We assume that any neutral that interacts with the plasma at
position $x_i$ joins the proton flow at that position. If the interaction
was charge transfer, a new neutral is formed with the bulk speed and
thermal speed of the protons at $x_i$. Thus the neutrals arriving at $x_i$
are those that last went through charge transfer at all upstream positions
$x_j$, and they have the speeds, $v_j$, of the plasma at $x_j$. Each ionization
of a neutral from $x_j$ at $x_i$ deposits energy
$0.5 m_p (v_j-v_i)^2$. We assume that the energy is quickly
thermalized among the protons, unlike \citet{ohira}, who assumed
a pickup ion velocity distribution. The thermalization time scale is
very uncertain, because full kinetic calculations have not been carried out.
However, in the highly
turbulent precursor there are many wave modes besides Alfv\'{e}n
waves that can thermalize the protons, in particular those
associated with bump-on-tail, mirror and firehose-like instabilities
\citep{winske, gary, sagdeev}. We also ignore heating
due to Alfv\'{e}n wave damping or shocks excited by the cosmic-ray
pressure gradient in order to isolate the effect of the neutrals.
Therefore, we compute a lower limit to the heating.
For electrons, we follow \citet{cairnszank},
who found that ionization of fast neutrals forms a ring beam, in which all
the particles gyrate around the magnetic field with the same speed but
different phases. The ring beam is unstable, and provided that the beam
velocity (in this case the relative velocity of bins i and j) is less
than 5 times the Alfv\'{e}n speed, it transfers a significant
amount of heat to electrons via Lower Hybrid waves. We follow
\citet{cairnszank} in taking this fraction to be 10\%. Again,
to isolate the effects of neutrals we ignore any heating of
electrons by Lower Hybrid waves generated by cosmic-ray
streaming \citep{ghavam07, rakow08}.
Charge transfer rates are taken from \citep{schultz} using the
quadrature sum of the thermal speed and the ion-neutral relative speed. Ionization
and excitation rates are computed from cross sections from \citet{janev} by
integrating over the electron velocity distribution
including the relative electron-neutral flow speed.
\section{Results}
\label{s_results}
Figure 1 shows a set of models for a shock speed of
2000 $\rm km~s^{-1}$ with $\kappa = 2.0 \times 10^{24}~\rm cm^2~s^{-1}$,
a pre-shock density of 0.2 $\rm cm^{-3}$ and a neutral fraction of
0.2. The four models have ratios of cosmic-ray partial pressure to total
pressure behind the subshock, w, of 0.1, 0.3, 0.5 and 0.7. The compression
ratios just ahead of the subshock are 1.0833, 1.3214, 1.75 and 2.75.
In the high $V_S$, high efficiency models the neutrals are
not compressed to this level, because of collisional ionization
and because some pass through without charge transfer.
The protons and electrons are strongly heated in the more efficient
models, but the electrons are much cooler than the protons. The
drop in heating just before the subshock in the 70\% efficient model
results from the reduced number of neutrals.
Figure 2 shows the velocity distributions of the neutrals perpendicular
to the shock just before the subshock.
Note that the w=50\% model shows a narrow component
due to neutrals that last experience charge transfer far upstream, along with
a broader component of particles that undergo charge transfer close to
the subshock.
Figure 3 shows a grid of models of the neutral velocity distribution at
the shock for a range of shock speeds and cosmic-ray
partial pressures. The panels show the FWHM measured directly from the
computed velocity distribution and the kurtosis, which would be 3.0 for
a Gaussian distribution. Kurtosis is a problematic statistical moment
for real data because it is sensitive to noise far from the
line center and the choice of background level. However, for the
theoretical profiles computed here it highlights cases in which
some neutrals undergo charge transfer close to the subshock and others
do not. We also show the fraction of incident neutrals that survive up to
the subshock and the average number of excitations to the n=3 level per
incident H atom. Not all of these excitations will result in H$\alpha$ photons because
some Ly$\beta$ photons escape, but this is a convenient comparison to
the 0.2 to 0.25 H$\alpha$ photons per H atom produced in the post-shock region.
\section{Discussion}
\label{s_discussion}
The interaction of neutral hydrogen with the ionized plasma in cosmic-ray
precursors described above offers an important tool to measure
the properties of cosmic-ray precursors.
The outcome of DSA is very much influenced by
physical processes in the precursor, which are not well determined.
For example, non-adiabatic heating and magnetic field amplification due the presence
of cosmic rays tend to decrease the overall compression factor
from $\chi_{12}>>20$ \citep[e.g.][]{berezhko,blasi05} to
$7 \lesssim \chi_{12}\lesssim 15$ \citep{vladimirov, caprioli}. In addition, if the Alfv\'en
waves in the precursor have some drift velocity this will affect the
cosmic-ray pressure profile \citep{zirakashvili08}, which
limits the escape of energy from the shock region. A lower
energy escape automatically implies a
lower downstream cosmic-ray pressure \citep{vink10}.
As shown here, neutrals will influence the physics of the precursor.
\citet{morlino} treated the neutrals as a fluid coupled to the ions
by charge transfer for a unified treatment
of the heating and dynamics of the precursor, but the fluid approximation
is only appropriate if neutrals and ions are coupled fairly well.
They obtained a FWHM of 46 $\rm km~s^{-1}$ for
the H$\alpha$ line in a 2000 $\rm km~s^{-1}$ shock with modest efficiency,
a pre-shock density of 1 $\rm cm^{-3}$ and 50\% neutral fraction. For
similar parameters we
find a non-Maxwellian profile with smaller FWHM and broader wings.
Neutrals can also damp plasma waves, which limits the efficiency of cosmic-ray
acceleration. This damping is caused by the central processes
described above: charge exchange and ionization.
\citet{drury} found that the maximum particle energy, and therefore the
maximum acceleration efficiency, is considerably
higher than suggested by \citet{drainemckee}. In addition, the heating
due to neutrals penetrating the precursor is a form of non-adiabatic heating.
Energy dissipated in the precursor
limits the amount of free energy available for shock acceleration. If the
neutrals ionized in the precursor behave as pickup ions rather than
thermalizing with the protons, the injection efficiency and particle
spectrum will be affected \citep{ohira}. In any case, the heating of
electrons in the precursor is poorly know, and that will stongly affect the
ionization and excitation of H atoms, which in turn will affect the
intensity ratio of the broad and narrow components as well as the
narrow component line width.
The physics of neutral-ion coupling means these processes
are not only sensitive to cosmic-ray pressure and the structure
of the precursor, but also to the pre-shock density and neutral fraction.
For example, the protons and neutrals in the Cygnus Loop nonradiative shocks
\citep{salvesen} are tightly coupled, and they behave nearly
adiabatically, while the pre-shock density in SN1006 is so low \citep{acero07}
that neutrals pass straight through it. This may explain the narrow line
width seen by \citet{sollerman} in SN\,1006
in comparison to broader narrow lines observed for other young SNRs.
The number of charge transfer events for an average neutral in the precursor
can be estimated roughly as
\begin{equation}
N_{\rm CT} = nL\sigma \chi_1 / V_S ,
\end{equation}
\noindent
where L is the precursor length scale, and the charge transfer cross section, $\sigma$,
declines slowly with velocity below about 2000 km s$^{-1}$, then very rapidly.
Comparing our results to observations, it is obvious that the observed narrow-line
H$\alpha$ widths are in general smaller than predicted by our calculations for efficient
shock acceleration (i.e., $w>0.5$). This may indicate that none of the shocks investigated so far
accelerate particles efficiently. However, more work is needed before such a conclusion
can be drawn, as the line width depends also on pre-shock density and shock velocity.
For very high shock velocities combined with low densities the neutrals hardly interact in the precursor,
leading to narrow line widths. Such may be the case for the northeastern region of RCW 86, for which
\citet{helder09} reported a high cosmic-ray acceleration efficiency
\citep[$w \gtrsim 0.5$, see also][]{vink10}. For this
region the pre-shock density may be as low as $n\lesssim 0.1$ \citep{vink06}, which, combined
with the high velocity ($V_s\gtrsim 3000$~km/s), gives few interactions in the precursor and
widths $\lesssim 100$~km/s. Note that this is smaller than could be measured given the moderate
spectral resolution of the measurement.
Perhaps the most striking result from these calculations is that for efficient shocks near 1000
$\rm km~s^{-1}$ a substantial number of H$\alpha$ photons will be produced in the precursor. These
will usually be included in the narrow component, potentially affecting the electron temperature
estimate based on the broad-to-narrow intensity ratio \citep{ghavam01}. Narrow component
emission from the precursor could explain the broad-to-narrow intensity ratios that
cannot be fit by models of post-shock emission \citep{vanA, rakow09}. In extreme cases,
emission from the precursor might also contribute to the broad component, possibly accounting for the
non-Maxwellian profile seen in Tycho's SNR \citep{ray10} and generally leading to an underestimate
of the shock speed. Both of these conclusions depend on the diffusion coefficient and the electron
heating, however.
Other H$\alpha$ line measurements show that the narrow lines are broader than one might expect
for temperatures of typical HII regions, but smaller than 50~km/s \citep{sollerman}.
(Not all of the narrow line emission comes from the precursor, but the narrow line emission
downstream is determined by the velocity distribution in the precursor.)
Another effect of charge exchange in the precursor is that neutrals enter the downstream shock
region with a velocity offset with respect to the local interstellar medium, as seen in
Tycho's supernova remnant (Lee et al., 2007). For shocks observed
face on this should produce a narrow line offset, which for the combined front and back side
of the remnant should lead to two narrow lines. The spectra of several LMC remnants \citep{smith94}
do not show such an effect. For one of the remnants in this set, SNR 0509-67.5, the
cosmic-ray acceleration efficiency was estimated to be $w\approx 0.2$ \citep{helder10}.
We note that several improvements should be made to the calculations
presented here. Additional heating due to wave dissipation can heat
the protons, resulting in larger narrow component line widths, or it can
heat electrons, increasing the H$\alpha$ narrow component intensity and
reducing the number of neutrals that reach the shock, especially if the
electron velocity distribution is non-Maxwellian \citep{laming07}. In addition, ionization and excitation by proton and helium ion
impact are important at high relative velocities \citep{laming96}, and at high shock speeds
the velocity distributions of particles are anisotropic \citep{hengmccray, heng07, vanA}.
Amplification of the magnetic field may also be important, and radiative transfer calculations in the Ly$\beta$ line
must be done to compute the H$\alpha$ emission. We plan to address these issues in future work.
\acknowledgements
This work was carried out while JCR was visiting the Astronomical Institute Utrecht
as Minnaert Professor. It was supported by NASA grant GO-11184.01-A-R to
the Smithsonian Astrophysical Observatory. JV and EH are supported by the
VIDI grant awarded to JV by the Netherlands Science Foundation (NWO).
\bibliographystyle{apj}
|
\section{Introduction}
Let $\fq$ be a finite field of $q = p^n$ elements and $\fq^*$ be the
set of all nonzero elements. Any mapping from $\fq$ to itself can be uniquely
represented by a polynomial of degree at most $q-1$. The degree of
such a polynomial is called the {\it reduced} degree. A multiset $M$ of size $q$ of field elements is called
the {\it range} of the polynomial $f(x) \in \fq[x]$ if $M = \{ f(x) : x \in \fq \}$ as a multiset
(that is, not only values, but also multiplicities need to be the
same). Here we use the set notation for multisets as well. We refer the readers to \cite{gac} for more details. In the study of polynomials
with prescribed range, G\`{a}cs et al. recently proposed the
following conjecture.
\begin{Con} [Conjecture 5.1, \cite{gac}]\label{conjecture1}
Suppose $M=\{a_1,a_2,\ldots ,a_q\}$ is a multiset of $\fq$
with $a_1+\ldots +a_q=0$, where $q=p^n$, $p$ prime. Let
$k<\sqrt{p}$. If there is no polynomial with range $M$ of degree
less than $q-k$, then $M$ contains an element of multiplicity at
least $q-k$.
\end{Con}
We note that Conjecture~\ref{conjecture1} is equivalent to
\begin{Con}\label{contrapositive}
Suppose $M=\{a_1,a_2,\ldots ,a_q\}$ is a multiset of $\fq$
with $a_1+\ldots +a_q=0$, where $q=p^n$, $p$ prime. Let
$k<\sqrt{p}$. If multiplicities of all elements in $M$ are less than
$q-k$, then there exist a polynomial with range $M$ of the degree
less than $q-k$.
\end{Con}
In the case $k=2$, Conjecture~\ref{conjecture1} holds by
Theorem~2.2.\ in \cite{gac}. In particular, Theorem~2.2 in
\cite{gac} gives a complete description of $M$ so that there is no
polynomial with range $M$ of reduced degree less than $q-2$. In
this paper, we study the above conjecture for $k \geq 3$.
Suppose we take a prescribed range $M$ such that the highest
multiplicity in $M$ is $\ell =q-k-1$, if the above conjecture were
true then it follows that there exist a polynomial, say $g(x)$, with
range $M$ and the degree of $g(x)$ is less than $q-k$. On the other
hand, If $a\in M$ is the element with multiplicity $\ell$ then
polynomial $g(x)-a$ has $\ell$ roots and thus
the degree of $g(x)$ is at least equal to the
highest multiplicity $\ell$ in $M$. Therefore the degree of $g(x)$
must be $\ell =q-k-1$. This means that, if
Conjecture~\ref{contrapositive} were true, then for every multiset
$M$ with the highest multiplicity $\ell =q-k-1$ where $1\leq k<
\sqrt{p}$ there exists a polynomial with range $M$ of the degree
$\ell $. Note that $k<\sqrt{p}$ implies $\ell
=q-k-1>q-\sqrt{p}-1\geq \frac{q}{2}$ when $q > 5$. Also $3\leq k\leq
\sqrt{p}$ implies $p>9$.
Let $M=\{a_1,a_2,\ldots a_q\}$ be a given multiset. We consider
polynomials $f(x):\fq \rightarrow M$, with the least degree. Denote
by $\ell $ the highest multiplicity in $M$ and let $\ell +m=q$. If
$a\in M$ is an element with multiplicity $\ell $ then the polynomial
$f(x)-a$ has the same degree as $f(x)$ and $0$ is in the range of
$f(x)-a$ such that $0$ has the same highest multiplicity $\ell$.
Therefore, we only consider multisets $M$ where $0$ has the highest
multiplicity for the rest of paper.
In particular, we prove the following theorem.
\begin{Them} \label{main}
Let $\fq$ be a finite field of $q =p^n$ elements with $p > 9$. For
every $\ell$ with $q-\sqrt{p}-1\leq \ell < q-3$ there exists a
mutiset $M$ with $\sum_{b\in M} b =0$ and the highest mutiplicity
$\ell$ achieved at $0\in M$ such that every polynomial over the
finite field $\fq$ with the prescribed range $M$ has degree greater
than $\ell $.
\end{Them}
In particular, for any $p > 9$, if we take $\ell =q-k-1\leq q-4$,
i.e., $k\geq 3$, then Theorem~\ref{main} implies that
Conjecture~\ref{contrapositive} fails.
\section{Proof of Theorem~\ref{main}}
In this section, we prove Theorem~\ref{main}. Let $\ell $ be fixed
and $q-\sqrt{p}-1 \leq \ell <q-3$. Because $p > 9$, $\sqrt{p} > 3$
and such $\ell$ exists.
Let $M$ be a multiset such that
$0\in M$ has the highest multiplicity $\ell $ and $\sum_{b\in
M}b=0$. Note that $\ell \geq \frac{q}{2}$ implies that multipilicity
of any nonzero element in $M$ is less than $m := q-\ell \leq
\frac{q}{2}$ (indeed the highest multiplicity is achieved at $0$).
Let $f:\fq \rightarrow M$. Let $U\subseteq \fq$ such that
$f(U)=\{0^\ell \}$ (the multiset of $\ell$ zeros) and $T=\fq \setminus U$, i.e., $x\in T$ implies
$f(x)\neq 0$. Then $|U|=\ell $ and $|T|=m$ and $M=f(U)\cup f(T)$.
Then polynomial $f:\fq\rightarrow M$ can be written in the form
$f(x)=h(x)P(x)$ where $P(x)=\prod_{s\in U}(x-s)$ and $h(x)\neq 0$
has no zeros in $T$. Then $\deg(f)\geq \deg(P) = \ell $. We note
that there is a bijection between polynomials with range $M =\{a_1, \ldots, a_q\}$ and the
ordered sets $(b_1, \ldots, b_q)$ (that is, permutations) of $\fq$: a
permutation corresponds to the function $f(b_i) = a_i$. For each
$U$, there are many different $h(x)$'s corresponding to different
ordered sets $(b_1, \ldots, b_q)$ such that $f(b_i) =0$ for all $b_i
\in U$. However, if $h(x)=\lambda\in \fq^*$ then $f(x)$ is a
polynomial of the least degree and each polynomial $f(x)$ is
uniquely determined by a set $T$ and a nonzero scalar $\lambda $.
Thus we denote $f(x)$ by
\begin{equation} \label{poly}
f_{(\lambda ,T)}(x)=\lambda \prod_{s\in \fq\setminus T}(x-s).
\end{equation}
Therefore its range $M$ is also uniquely determined by $T$ and
$\lambda $. Denote by $\mathcal{T}$ the family of all subsets of
$\fq $ of cardinality $m$, i.e.,
$$\mathcal{T}=\{T \mid T\subseteq \fq, |T|=m\}.$$
Denote by $\mathcal{M}$ the family of all multisets $M$ of order $q$
containing $0$, having the highest multiplicity $\ell $ achieved at
$0$ and whose sum of elements in $M$ is equal to the $0$, i.e.,
$$
\mathcal{M}=\{M \mid 0\in M, \text{ multiplicity}(0)=\ell,
\sum_{b\in M}b=0\}.
$$
Equation~(\ref{poly}) uniquely determines a mapping
$$\mathcal{F}:\fq^* \times \mathcal{T}\rightarrow \mathcal{M}$$
where $$(\lambda ,T)\mapsto \text{range}(f_{\lambda, T}(x)).$$
Now by Equation~(\ref{poly}) it follows
that for every $\hat{s}\in T$ we have
\begin{equation}\label{basicequations}
f_{\lambda, T}(\hat{s})=\lambda P(\hat{s})=\lambda \prod_{s\in
\fq,s\neq \hat{s}}(\hat{s}-s)\Bigl(\prod_{s\in T,s\neq
\hat{s}}(\hat{s}-s)\Bigr)^{-1}=-\lambda \Bigl(\prod_{s\in T,s\neq
\hat{s}}(\hat{s}-s)\Bigr)^{-1}.\end{equation} (Note that this
equation does not hold for $x\in \fq\setminus T$). In the following
we find an upper bound of $| \text{range}(\mathcal{F}) |$ and a
lower bound of $| \mathcal{M} |$ and show that $|\mathcal{M}| > |\text{range}(\mathcal{F})|$. This implies that
Theorem~\ref{main} holds.
First of all we observe
\begin{Lem} Let $\lambda $ and $T$ be given. For any $c\in
\fq^*$ and any $b\in \fq$, we have
$$f_{(\lambda,T)}(\hat{s})=f_{(c^{m-1}\lambda, cT+b)}(c\hat{s}+b),\qquad {\text for}~\hat{s}\in T$$
i.e.,
$$\mathcal{F}(\lambda, T)=\mathcal{F}(c^{m-1}\lambda ,cT+b).$$
\end{Lem}
\begin{proof}
We use notation $cT+b=\{cs+b \mid s\in T\}$. Substituting in
(\ref{basicequations}), we obtain $f_{(c^{m-1}\lambda,
cT+b)}(c\hat{s}+b)=-c^{m-1}\lambda (\prod_{s\in T,s\neq
\hat{s}}((c\hat{s}+b)-(cs+b)))^{-1}=-\lambda (\prod_{s\in T,s\neq
\hat{s}}(\hat{s}-s))^{-1}=f_{(\lambda, T)}(\hat{s})$.
\end{proof}
Now we use Burnside's Lemma to find an upper bound of the
cardinality of $\text{range}(\mathcal{F})$.
\begin{Lem}\label{countingRange} Let $m< \sqrt{p}+1$ and let $d = \gcd(q-1, m-1)$. Then
$$|\text{range}(\mathcal{F})|\leq \frac{(q-1)(q-2)\ldots
(q-m+1)}{m!}+\sum_{{i\mid d \atop i >1}
}\phi(i)\binom{\frac{q-1}{i}}{\frac{m-1}{i}}.$$
\end{Lem}
\begin{proof}
Let $\mathcal{G}$ be group of all (nonzero) linear polynomials in
$\fq[x]$ with the composition operation. Indeed, $\mathcal{G}$ is a
subgroup of the group of all permutation polynomials because the
composition of two linear polynomials is again a linear polynomial,
the identity mapping is a linear polynomial, and the inverse of a
linear polynomial is again a linear polynomial. We use notation
$cT+b=\{cs+b \mid s\in T\}$ again. Then $\mathcal{G}$ acts on the
set $\fq^*\times \mathcal{T}$ with $\Phi :
\mathcal{G}\times(\fq^*\times \mathcal{T}) \rightarrow \fq^*\times
\mathcal{T}$, where
$$\Phi : (cx+b, (\lambda,T))\mapsto ( c^{m-1}\lambda ,cT+b).$$
The elements of the same orbit
$$\mathcal{G}(\lambda, T)=\{(c^{m-1}\lambda ,cT+b) \mid cx+b\in\mathcal{G}\}$$
are all mapped to the same element $M\in \mathcal{M}$ by Lemma 1. By
Burnside's Lemma the number of orbits $N$ is given by
$$ N =\frac{1}{|\mathcal{G}|}\sum_{g\in \mathcal{G}}|(\fq^*\times
\mathcal{T})_g|,$$ where $g(x)=cx+b$, and
$$(\fq^*\times \mathcal{T})_g=\{(\lambda ,T) \mid (\lambda ,T)\in
\fq^*\times \mathcal{T}, (c^{m-1}\lambda ,cT+b)=(\lambda ,T)\}.$$
The equation $cx+b=x$ over $\fq$ is equivalent to $(c-1)x=-b$, which
has exactly one solution if $c\neq 1$; no solutions if $c=1$ and
$b\neq 0$; $q$ solutions if $c=1$ and $b=0$. If $c\neq 1$ and $i
:=ord(c)\mid q-1$, then this linear polynomial has one fixed
element and $\frac{q-1}{i}$ cycles of length $i$. If $c=1$ and
$b\neq 0$ then $g^p(x)=x+pb=x$ and thus $g(x)$ has cycles of length
$p$ where $p = char(\fq)$.
Assume $T=cT+b$. Let $s\in T$. Then $g(s)\in cT+b=T$. So the cycle
$s,g(s),g^2(s),\ldots ,g^i(s)=s$ is contained in $T$.
This means that, under the assumptions of $c\neq 1$ and $T=cT+b$,
either $T$ has one fixed element and $\frac{m-1}{i}$ cycles of the
length $i$ which are defined by permutation $g(x)$, or $T$ has
$\frac{m}{i}$ cycles length $i$ which are defined by permutation
$g(x)$. In the latter case, the fixed element of $g(x)$ is in $\fq
\setminus T$.
In the former case, if $c\in \fq^* \setminus \{1\}$ satisfies $i=ord(c) \mid
d = \gcd(q-1,m-1)$ then there are
$\binom{\frac{q-1}{i}}{\frac{m-1}{i}}$ sets fixed by $g(x)$.
Moreover, $c^{m-1}=(c^i)^\frac{m-1}{i}=1$. Hence, for each set $T$
fixed by $g(x)$ and any $\lambda \in \fq^*$ we must have
$(c^{m-1}\lambda ,cT+b)=(\lambda,T)$. This implies that
$$|(\fq^*\times
\mathcal{T})_g|=(q-1)\binom{\frac{q-1}{i}}{\frac{m-1}{i}}.$$
If $c\in \fq^*$ satisfies $i = ord(c) \mid \gcd(q-1,m)$ then there
are $\binom{\frac{q-1}{i}}{\frac{m}{i}}$ sets $T$ fixed by $g(x)$.
But for each $T$ fixed by $g(x)$, $c^{m-1}=c^{-1}\neq 1$ and thus
$(c^{m-1}\lambda ,cT+b)\neq (\lambda, T)$. Therefore
$$|(\fq^*\times
\mathcal{T})_g|=0.$$
If $c=1$ and $b=0$ then $g(x) =x$. So $|(\fq^*\times
\mathcal{T})_g|=(q-1)\binom{q}{m}$. If $c=1$ and $b\neq 0$ then
$cT+b \neq T$. Otherwise, it implies that $T$ contains elements of
the cycles of the length $p$ which contradicts to $m<\sqrt{p}+1$.
Since $d = \gcd(q-1, m-1)$, we obtain
\begin{eqnarray*}
N&=&\frac{1}{|\mathcal{G}|}\sum_{g\in \mathcal{G}}|(\fq^*\times
\mathcal{T})_g|\\
&=&\frac{1}{q(q-1)}\biggl( (q-1)\cdot
\binom{q}{m}+\sum_{\begin{array}{c}\scriptscriptstyle c\in \fq^*\setminus \{1\}\\
\scriptscriptstyle i=ord(c)\mid
d\\
\scriptscriptstyle b\in \fq \end{array}} (q-1)\binom{\frac{q-1}{i}}{\frac{m-1}{i}} \biggr),\\
&=&\frac{1}{q(q-1)}\biggl((q-1)\cdot
\binom{q}{m}+q(q-1)\sum_{\begin{array}{c} \scriptscriptstyle c\in \fq^*\setminus \{1\}\\
\scriptscriptstyle i=ord(c)\mid d
\end{array}} \binom{\frac{q-1}{i}}{\frac{m-1}{i}} \biggr)\\
&=&\frac{(q-1)(q-2)\ldots (q-m+1)}{m!}+\sum_{i\mid
d}\phi(i)\binom{\frac{q-1}{i}}{\frac{m-1}{i}},
\end{eqnarray*}
where $\phi(i)$ is the number of $c$'s such that the order of $c$
is $i > 1$.
Since two orbits could possibly be mapped to the same
multiset $M\in \mathcal{M}$ we finally have an inequality
\begin{equation}|\text{range}(\mathcal{F})|\leq \frac{(q-1)(q-2)\ldots
(q-m+1)}{m!}+\sum_{i\mid
d}\phi(i)\binom{\frac{q-1}{i}}{\frac{m-1}{i}}.\end{equation}
\end{proof}
\par Now we find a lower bound of the cardinality of
$\mathcal{M}=\{\stackrel{l\quad \text{ times}}{\overbrace{0,0\ldots
,0}},b_1,b_2,\ldots ,b_m\}$ such that $b_i \neq 0$ for $i=1,
\ldots, m$ and
\begin{equation}\label{imagesum}
b_1+b_2+\ldots +b_m=0.
\end{equation}
Although we can find a simpler exact formula for the number of solutions to Equation~(\ref{imagesum}), we prefer the following lower bound for $|\mathcal{M}|$ which has the same format as the upper bound of $|\text{range}(\mathcal{F})|$ in order to compare them directly.
\begin{Lem}\label{countingM}
Let $A=1$ if $m-1\mid q-1$ and $A=0$ otherwise. If $m\geq 6$ then
$$
|\mathcal{M}|\geq \frac{(q-1)\ldots
(q-m+2)(q-2)}{m!}+$$
$$\sum_{\begin{matrix}
\scriptscriptstyle 1<i<m-1\\
\scriptscriptstyle i\mid
\gcd(q-1,m-1)\end{matrix}}\frac{[(q-1)\ldots
(q-\frac{m-1}{i}+2)][(q-\frac{m-1}{i}+1)(q-\frac{m-1}{i}-1)+\frac{m-1}{i}]}{\Bigl(\frac{m-1}{i}\Bigr)!}+A(q-1).$$
If $m=4$ and $3\mid q-1$ then
$$|\mathcal{M}|\geq \frac{(q-1)(q-2)^2}{4!}.$$
If $m=5$ then
$$|\mathcal{M}|\geq \frac{(q-1)(q-2)^2(q-3)}{5!}+ A(q-1).$$
\end{Lem}
\begin{proof}
In order to give a lower bound of $|\mathcal{M}|$, we count two
different classes of families of multisets $M$. The first class
contains families of those multisets $M$ such that almost all
nonzero elements $b_i$'s have the same multiplicities greater than
one except the last element $b_m$. And the second family class
contains those multisets $M$ such that almost all nonzero elements
$b_i$'s have multiplicities one except that the last two elements
$b_{m-1}$ and $b_{m}$.
First, we count those multisets $M$ such that almost all nonzero
elements $b_i$'s have the same multiplicities greater than one
except the last element $b_m$. That is, for any $i$ such that
$1<i<m-1$ and $i\mid \gcd(q-1,m-1)$,
we want to choose $\frac{m-1}{i}$ pairwise distinct nonzero
elements each of multiplicity $i$ so that
$\displaystyle\sum_{j=1}^{\frac{m-1}{i}}ib_j\neq 0$ (the sum being
equal to zero would imply $b_m=0$, a contradiction). For each such
$i$, we denote the family of these multisets by $\mathcal{M}_i$.
We note that each multiset $M\in \mathcal{M}_i$ can be written as
$$M=\{\stackrel{\ell \quad \text{ times}}{\overbrace{0,0,\ldots
,0}},\stackrel{i\quad \text{ times}}{\overbrace{b_1,\ldots
,b_1}},\ldots ,\stackrel{i\quad \text{
times}}{\overbrace{b_{\frac{m-1}{i}},\ldots
,b_{\frac{m-1}{i}}}},b_m\}.$$
Obviously each multiset is invariant to the ordering. However, let
us first consider the ordered tuples $(b_1,\ldots
,b_{\frac{m-1}{i}-1})$ satisfying that $b_i$'s are nonzero and
pairwise distinct. Out of a total of $(q-1)\ldots
(q-\frac{m-1}{i}+1)$ such ordered tuples, there are
$(q-1)\ldots
(q-\frac{m-1}{i}+2)\frac{m-1}{i}$ choices such that
$-\displaystyle\sum_{j=1}^{\frac{m-1}{i}-1}b_j\in \{0,b_1,\ldots
,b_{\frac{m-1}{i}-1} \}$ and $(q-1)\ldots (q-\frac{m-1}{i}+2)(q- 2
\frac{m-1}{i}+1)$ ordered tuples such that
$-\displaystyle\sum_{j=1}^{\frac{m-1}{i}-1}b_j\not \in
\{0,b_1,\ldots ,b_{\frac{m-1}{i}-1} \}$. If
$-\displaystyle\sum_{j=1}^{\frac{m-1}{i}-1}b_j\in \{0,b_1,\ldots
,b_{\frac{m-1}{i}-1} \}$ then $b_{\frac{m-1}{i}}$ can be chosen in
$ q-\frac{m-1}{i}$ ways and otherwise it can be chosen in
$q-\frac{m-1}{i}-1$ way. Because the element $b_m$ is uniquely
determined by $i\sum_{j=1}^{\frac{m-1}{i}}b_j$, we have in total
$$(q-1)\ldots
(q-\frac{m-1}{i}+2)\frac{m-1}{i}(q-\frac{m-1}{i})$$
$$+ (q-1)\ldots
(q-\frac{m-1}{i}+2)(q-2\frac{m-1}{i}+1)(q-1-\frac{m-1}{i})$$
$$= \left( (q-1)\ldots(q-\frac{m-1}{i}+2)\right) \left((q-\frac{m-1}{i}+1)(q-\frac{m-1}{i}-1)+\frac{m-1}{i}\right)$$
ordered tuple $(b_1\ldots ,b_{\frac{m-1}{i}})$ satisfying
Equation~(\ref{imagesum}) and that $b_i \neq 0$ for $i=1, \ldots, m$
and each element is of multiplicity $i$ except that last element.
Since there are $\Bigl(\frac{m-1}{i}\Bigr)!$ permutations of the
ordered tuples $(b_1,\ldots ,b_{\frac{m-1}{i}})$, there are
$$\frac{[(q-1)\ldots(q-\frac{m-1}{i}+2)][(q-\frac{m-1}{i}+1)(q-\frac{m-1}{i}-1)+\frac{m-1}{i}]}{\Bigl(\frac{m-1}{i}\Bigr)!}$$
elements in
$\mathcal{M}_i$.
Similarly, if $m-1\mid q-1$, we denote by $\mathcal{M}_{m-1}$ the
set of multisets $M$ such that all $b_i$'s are the same nonzero
element for $i=1, \ldots, m-1$ and their sum together with $b_m$ is
zero. It is easy to see that there are $q-1$ such $M$'s, i.e.,
$|\mathcal{M}_{m-1}|=q-1$.
Now we show that $\mathcal{M}_i\cap \mathcal{M}_j =\emptyset$ for $1
< i\neq j \leq m-1$. We prove this by contradiction and we use
heavily the fact that, for each $i$, there are $\frac{m-1}{i}+1$
distinct elements in $M\in \mathcal{M}_i$ if $b_m\neq b_k$ for
$1\leq k\leq \frac{m-1}{i}$ and
there are $\frac{m-1}{i}$ distinct elements in $M$ if $b_m = b_k$ for some $k$. Assume that $\mathcal{M}_i\cap \mathcal{M}_j \neq \emptyset$. Obviously,
$\frac{m-1}{i}\neq \frac{m-1}{j}$ because $i \neq j$. Hence
either $\frac{m-1}{i}+1 = \frac{m-1}{j}$ or $\frac{m-1}{j}+1 =
\frac{m-1}{i}$.
Without loss of generality, we assume now $\frac{m-1}{i}+1=
\frac{m-1}{j}$ and $M\in \mathcal{M}_i\cap \mathcal{M}_j$. Then in
the multiset $M$, we have $\frac{m-1}{i}$ elements of multiplicity
$i$ and one element of multiplicity $1$ since $M\in \mathcal{M}_i$.
Moreover, the number of elements of multiplicity $j$ is
$\frac{m-1}{j}-1$ and there is one element of multiplicity $j+1$
since $M\in \mathcal{M}_j$. Because $i >j$, we must have $i=j+1$ and
$j=1$ by comparing the multipicities. However, this implies we must
have $\frac{m-1}{i}=1$ and $\frac{m-1}{j}-1=1$. Hence $i = m-1$ and
$j=\frac{m-1}{2}$, contradicts to $i =j+1$ when $m > 3$.
Therefore $\mathcal{M}_i\cap \mathcal{M}_j\neq \emptyset $ for all $
1 < i\neq j \leq m-1$. Now for $m \geq 4$ we have
$$|\bigcup_{\begin{matrix}\scriptscriptstyle 1<i\leq m-1\\
\scriptscriptstyle i\mid \gcd(q-1,m-1)\end{matrix}}\mathcal{M}_i|=
A(q-1) +$$
$$\sum_{\begin{matrix}\scriptscriptstyle 1<i<m-1\\
\scriptscriptstyle i\mid
\gcd(q-1,m-1)\end{matrix}}\frac{[(q-1)\ldots
(q-\frac{m-1}{i}+2)][(q-\frac{m-1}{i}+1)(q-\frac{m-1}{i}-1)+\frac{m-1}{i}]}{\Bigl(\frac{m-1}{i}\Bigr)!}.$$
Next we count those multisets $M$ such that almost all nonzero
elements $b_i$'s have multiplicities one except that the last two
elements $b_{m-1}, b_{m}$. That is, $b_1,\ldots, b_{m-2}$ are
pairwise distinct nonzero elements, $b_{m-1} \neq 0$ is chosen in a
way such that $\displaystyle\sum_{j=1}^{m-1}b_j\neq 0$,
and $b_m$ is uniquely
determined by $\displaystyle\sum_{j=1}^mb_j=0$. The family of such
multisets is denoted by $\mathcal{M}_0$. We note that $b_{m-1}$ and
$b_m$ could be same as one of $b_j$'s where $j=1, \ldots, m-2$. So
the highest mulitiplicity is at most $3$.
Consider all $(q-1)\ldots (q-m+2)$ different ordered tuples
$(b_1,\ldots ,b_{m-2})$. If $-\displaystyle\sum_{j=1}^{m-2}b_j\neq
0$
we can choose $b_{m-1}$ in $q-2$ ways and
otherwise there are $q-1$ choices for $b_{m-1}$. Thus in total there
are at least $(q-1)\ldots (q-m+2)(q-2)$
ordered tuples $(b_1,\ldots ,b_m)$.
Let $S_1$ be the number of such ordered tuples without repetition,
$S_2$ be the number of ordered tuples with exactly one repeated
element, $S_3$ be the number of arrays with exactly two pairs of
repeated elements, and $S_4$ be the number of tuples with exactly
one element repeated $3$ times. Because multisets are invariant to
the ordering, there are at least
$$\frac{S_1}{m!}+\frac{S_2}{(m-1)!}+\frac{S_3}{(m-2)!2!}+\frac{S_4}{(m-2)!}\geq\frac{(q-1)\ldots
(q-m+2)(q-2)}{m!}$$
such multisets in $\mathcal{M}_0$, i.e., $$|\mathcal{M}_0|\geq \frac{(q-1)\ldots
(q-m+2)(q-2)}{m!}.$$
We note that each multiset from $\mathcal{M}_0$ contains at least $m-2$ distinct elements and each multiset from $\mathcal{M}_i$ with $i>1$ contains at most $\frac{m-1}{i}+1\leq \frac{m-1}{2}+1$ distinct
elements. Since $\frac{m-1}{2}+1<m-2$ for $m\geq 6$ we have that
$\mathcal{M}_0\cap\mathcal{M}_i=\emptyset $ as long as $m\geq 6$.
Therefore we can conclude that for $m\geq 6$ we have
$$|\mathcal{M}|\geq |\mathcal{M}_0|+|\bigcup_{\begin{matrix}\scriptscriptstyle 1<i\leq m-1\\
\scriptscriptstyle i\mid
\gcd(q-1,m-1)\end{matrix}}\mathcal{M}_i|\geq \frac{(q-1)\ldots
(q-m+2)(q-2)}{m!}+$$
$$\sum_{\begin{matrix}\scriptscriptstyle 1<i<m-1\\
\scriptscriptstyle i \mid
\gcd(q-1,m-1)\end{matrix}}\frac{[(q-1)\ldots
(q-\frac{m-1}{i}+2)][(q-\frac{m-1}{i}+1)(q-\frac{m-1}{i}-1)+\frac{m-1}{i}]}{\Bigl(\frac{m-1}{i}\Bigr)!}+A(q-1).$$
Let $m=4$. If $i>1$ and $i\mid
\gcd(m-1,q-1)$ then $i=3$. Thus in this case
$\mathcal{M}_3\cap \mathcal{M}_0=\{ \{a,a,a,b\} \mid a\in \fq^*, b=-3a\neq
a \}$ since $p> 9$.
By the principle of the inclusion-exclusion we obtain
$$|\mathcal{M}|\geq \frac{(q-1)(q-2)(q-2)}{4!}+(q-1)-(q-1)=\frac{(q-1)(q-2)(q-2)}{4!}.$$
If $m=5$, then $i>1$ and $i\mid \gcd(4, q-1)$ imply $i=2$ or $i=4$.
Obviously $\mathcal{M}_0 \cap \mathcal{M}_4 = \emptyset$ because
each element in a multiset of $\mathcal{M}_0$ has multipliciy at
most $3$. Similarly, any multiset in both $\mathcal{M}_0$ and
$\mathcal{M}_2$ must contain $\frac{m-1}{i}+1=m-2=3$ distinct
elements, two of them come in pairs. That is,
$$\mathcal{M}_0\cap \mathcal{M}_2=\Bigl\{\{a,a,b,b,c\} \mid a,b,c\in
\fq^*,a\neq b,a\neq c,b\neq c\Bigr\}.$$ If $a$ is chosen in $q-1$
ways then $b\not\in\{0,a,-a\}$ and we can choose $b$ in $q-3$ ways.
Since multisets are invariant to the ordering we have
$$|\mathcal{M}_0\cap \mathcal{M}_2|=\frac{(q-1)(q-3)}{2!}.$$
Again the principle of inclusion-exclusion implies
\begin{eqnarray*}
|\mathcal{M}| &\geq & |\mathcal{M}_0|+|\mathcal{M}_2|+|\mathcal{M}_4|- |\mathcal{M}_0 \cap \mathcal{M}_2| \\
&= & \frac{(q-1)(q-2)(q-3)(q-2)}{5!}+ \frac{(q-1)(q-3)}{2!}+A(q-1)-\frac{(q-1)(q-3)}{2!} \\
&= & \frac{(q-1)(q-2)(q-3)(q-2)}{5!}+A(q-1).
\end{eqnarray*}
\end{proof}
We need the following simple result to compare the bounds of
$\mathcal{M}$ and $|\text{range}(\mathcal{F})|$ in order to complete
the proof of Theorem~\ref{main}.
\begin{Lem} \label{inequality}
(i) For $m\geq 4$, we have
$$\frac{(q-1)(q-2)\ldots
(q-m+1)}{m!}<\frac{(q-1)\ldots (q-m+2)(q-2)}{m!}.$$
(ii) If $1<i<m-1$ and $i\mid \gcd(q-1,m-1) $ then
$$\phi(i)\binom{\frac{q-1}{i}}{\frac{m-1}{i}}<\frac{(q-1)\ldots
(q-1-\frac{m-1}{i}+2)[(q-\frac{m-1}{i}+1)(q-\frac{m-1}{i}-1)+\frac{m-1}{i}]}{(\frac{m-1}{i})!}.$$
(iii) If $i=m-1\mid q-1$ then
$$\phi(m-1)\binom{\frac{q-1}{i}}{\frac{m-1}{i}}<q-1.$$
\end{Lem}
\begin{proof}
\par (\emph{i}) Clearly, $q-m+1<q-2$ for $m\geq 4$.
\par (\emph{ii}) The inequality
$$\phi(i)\binom{\frac{q-1}{i}}{\frac{m-1}{i}}<\frac{(q-1)\ldots
(q-\frac{m-1}{i}+2)[(q-\frac{m-1}{i}+1)(q-\frac{m-1}{i}-1)+\frac{m-1}{i}]}{(\frac{m-1}{i})!}$$
is equivalent to
$$
\phi(i)\left(\frac{q-1}{i}\bigl(\frac{q-1}{i}-1\bigr)\ldots
\bigl(\frac{q-1}{i}-\frac{m-1}{i}+1\bigr)\right) $$
$$<(q-1)(q-2)\ldots (q-\frac{m-1}{i}+2) \left((q-\frac{m-1}{i}+1)(q-\frac{m-1}{i}-1)+\frac{m-1}{i}\right).$$
Using $\phi(i)\frac{q-1}{i}<q-1,$ $\frac{q-1}{i}-j<\frac{q-1}{i}$
for $j=1,\ldots ,\frac{m-1}{i}-2$ and $\frac{q-1}{i} -\frac{m-1}{i}
+ 1 < q-1 - \frac{m-1}{i}$ (since $i >1$), we have
\begin{eqnarray*}
&& \phi(i)\frac{q-1}{i}\bigl(\frac{q-1}{i}-1\bigr)\ldots
\bigl(\frac{q-1}{i}-\frac{m-1}{i}+1\bigr) \\
&<&(q-1)\bigl(\frac{q-1}{i}-1\bigr)\ldots
\bigl(\frac{q-1}{i}-\frac{m-1}{i}+1\bigr) \\
&<& (q-1)(q-2)\ldots
(q-\frac{m-1}{i}+2)(q-\frac{m-1}{i}+1) \bigl(\frac{q-1}{i}-\frac{m-1}{i}+1\bigr)\\
&<& (q-1)(q-2)\ldots
(q-\frac{m-1}{i}+2)(q-\frac{m-1}{i}+1)(q- 1 - \frac{m-1}{i}) \\
&<&(q-1)(q-2)\ldots (q-\frac{m-1}{i}+2) \left(
(q-\frac{m-1}{i}+1)(q-\frac{m-1}{i}-1)+\frac{m-1}{i}\right).
\end{eqnarray*}
\par (\emph{iii}) If $i=m-1\mid q-1$ then $\phi(m-1)\frac{q-1}{m-1}<q-1$.
\end{proof}
\textbf{Proof of Theorem~\ref{main}:}
If $m\geq 6$ it follows directly from Lemmas~\ref{countingRange}, ~\ref{countingM}, ~\ref{inequality}.
Note that $m\leq \sqrt{p}+1$. If $m=5$ then $5 \leq \sqrt{p} + 1$
implies that $p >16$. Hence we have
\begin{eqnarray*}
|\text{range}(\mathcal{F})| &\leq&
\frac{(q-1)(q-2)(q-3)(q-4)}{5!}+\phi(2)\binom{\frac{q-1}{2}}{2}+A\phi(4)\frac{q-1}{4}\\
& = & \frac{(q-1)(q-2)(q-3)(q-4)}{5!}+\frac{(q-1)(q-3)}{8}+A\frac{q-1}{2}\\
& < & \frac{(q-1)(q-2)(q-3)(q-4)}{5!}+\frac{2(q-2)}{15} \frac{(q-1)(q-3)}{8}+A\frac{q-1}{2}\\
&\leq& \frac{(q-1)(q-2)(q-3)(q-2)}{5!}+A(q-1)\\
&\leq & |\mathcal{M}|.
\end{eqnarray*}
If $m=4$ and $3\nmid q-1$ then the result follows directly from
Lemmas~\ref{countingRange}, \ref{countingM}, and \ref{inequality}
(i). If $3\mid q-1$ then
$$ |\text{range}(\mathcal{F})|\leq
\frac{((q-1)(q-2)(q-3)}{4!}+\phi(3)\frac{q-1}{3}<\frac{(q-1)(q-2)^2}{4!}<|\mathcal{M}|$$
holds for $q>18$. Note that $m\leq \sqrt{p}+1$ implies $p\geq 9$. By the
assumption of $p > 9$, we must have $p\geq 11$.
The only possible prime power $q \leq 18$ such that $p\geq 11$ and $3 \mid q-1$ is $q=13$. It is easy
to compute that the number of all the possible solutions to Equation~(\ref{imagesum}) with desired properites over $\mathbb{F}_{13}$ is $|\mathcal{M}|=105$ by a computer program.
For $q=13$, then $\gcd(q-1, m-1)= 3$
and thus
$|\text{range}(\mathcal{F})| \leq 63 < 105 = |\mathcal{M}|$.
Hence the proof is complete. $\Box$
If $m=2$ and $m=3$ these polynomials satisfying the conjecture do
exist. Indeed, if $m=2$ and
$b_2=-b_1$, then we can construct the minimum degree polynomial $f(x) = \lambda \prod_{s \in \fq \setminus T} (x-s)$
with the prescribed range $M=\{0, \ldots, 0, b_1, -b_1\}$ by letting $T= \{b_1^{-1}, 0\}$ and $\lambda =1$.
\par For the case $m=3$, for any multiset $M = \{0,\ldots ,0,b_1,b_2,b_3\}$ with $b_1+b_2+b_3=0$ such that $b_1, b_2,
b_3$ are all nonzero there exists a polynomial
$f(x)=\lambda\prod_{s\in \fq\setminus T}(x-s)$ of the least degree
with range $M$. Indeed, let $T=\{b_2,-b_1,0\}$ and $\lambda
=b_1b_2b_3$. Then using $b_3=-(b_1+b_2)$ we obtain
$$f(b_2)=b_1b_2b_3\frac{-1}{(b_2+b_1)b_2}=b_1;$$
$$f(-b_1)=b_1b_2b_3\frac{-1}{(-b_1-b_2)(-b_1)}=b_2;$$
$$f(0)=b_1b_2b_3\frac{-1}{(-b_1)(b_2)}=b_3.$$
|
\section{Introduction}
A closed curve in $\mathbb{R}^3$ is called ideal if it minimizes its ropelength $\scriptL[\cdot]/\Delta[\cdot]$
-- i.e. its length divided by its thickness -- within its knot class \cite{GM99}.
In this paper we will focus on the simplest of all proper ideal knots,
namely the trefoil knot. Various numerical approximations of this specific
knot are available. It is not trivial to define what the properties of a ``good'' approximation are.
Quantities like ropelength, functions such as curvature and torsion, or the contact set for a given knot are
can all be used to assess whether a knot is close to ideal. There exist
several algorithms to compute ideal knot shapes, which use
different approximations for the curve description\cite{L98,bookchapter,P98,CPR05,ACPR10}. These
numerical computations are expected to lead to a better
understanding of ideal knots. In this sense a numerical shape is ``good'',
if it leads to more insight about properties of ideal knots.
A curve is in contact with itself at the points $p,q$ if the distance between $p$ and $q$
is precisely two times the thickness of the curve and the line segment between them is orthogonal
to the curve at both ends. \cite{S04, bookchapter} define a robust sense of contact with a tolerance and their computations
suggest that each point on the ideal trefoil in $\mathbb{R}^3$ is in contact with two other points .
Starting
from a point $p_0$, it is in contact with a point $p_1$ that itself is again in contact with a point $p_2\neq p_0$ and
so on. Does this sequence close to a cycle?
In this article we observe that computations suggest that the ideal trefoil
knot has a periodic, and attracting nine-cycle of contact chords,
as illustrated in Figure \ref{fig:billiard}.
A similar construction of periodic cycles, but in each point of the curve,
helped to construct the ideal Borromean rings \cite{Sta03,CFKSW06}.
The existence of this cycle is significant because it partitions the trefoil in such a way
that, using the apparent symmetries, it can be re-constructed from two unknown small pieces of
curves mutually in contact.
\begin{figure}
\begin{center}
\includegraphics[width=.6\textwidth]{billiard.pdf}
\caption{ The parameters $s_i := \sigma^i(0)$ of the nine-cycle $b_9$ partition the trefoil in 9 curves:
$\beta_i$ and $\tilde{\beta_i}$ are all congruent as are $\alpha_i$ ($i=1,2,3$). The contact function
$\sigma$ maps the parameter
interval of each curve bijectively to the parameter interval of another curve (see Figure
\ref{fig:connect}).
} \label{fig:billiard}
\end{center}
\end{figure}
We approximated the ideal trefoil using a Fourier representation described in \cite{CG10}.
The numerical computations were carried out with \texttt{libbiarc} \cite{LB10}
and the data is available from \cite{Data10}.
The numerical Fourier trefoil is not the best known in ropelength sense, but -- to our knowledge --
the best shape to observe the closed cycle, probably because we can enforce specific
symmetries.
Another interesting discovery is that, if we follow the contact chords starting
at an arbitrary point on the trefoil, we always end up at
the previously mentioned cycle, in other words, it is a global attractor.
In order to present this closed cycle on the trefoil we first
review the notions of global radius of curvature \cite{GM99}
and contact of a curve \cite{PP02, bookchapter} in Section \ref{sec:grc}.
Section \ref{sec:billiards} introduces contact billiards and cycles. Then
we present and discuss a candidate for a cycle in the trefoil
and show numerically that it seems to act as an attractor for all the billiards
on the trefoil.
\section{The Ideal Trefoil -- Its Contact Chords and Symmetries}\label{sec:grc}
A knot is a closed curve $\gamma \in C^1(\mathbb{S},\mathbb{R}^3)$
where $\mathbb{S}:=\mathbb{R}/\mathbb{Z}$ is the unit interval with the endpoints identified, isomorphic to the unit circle.
We use the global radius of curvature to assign a thickness $\Delta$
to $\gamma$.
\begin{definition}[Global radius of curvature] \cite{GM99} \label{def:grc}
For a $C^0$-curve $\gamma: \mathbb{S} \longrightarrow \mathbb{R}^3$ the global radius of
curvature at $s \in \mathbb{S}$ is
\begin{equation}
\rho_G[\gamma](s) := \inf_{\sigma,\tau \in \mathbb{S}, \sigma \not= \tau, \sigma\not=s, \tau\not=s}
R(\gamma(s),\gamma(\sigma), \gamma(\tau)).
\end{equation}
Here $R(x,y,z) \ge 0$ is the radius of the smallest circle
through the points $x,y,z \in \mathbb{R}^3$, i.e.
\[
R(x,y,z) :=
\left\{
\begin{array}{cl}
\frac{|x-z|}{2 \sin \measuredangle(x-y,y-z)} &
\mbox{$x,y,z$ not collinear}, \\
\infty &
\mbox{$x,y,z$ collinear, pairwise distinct}, \\
\frac{{\rm diam}(\{x,y,z\})}{2} & \mbox{otherwise}.
\end {array}
\right.
\]
where $\measuredangle(x-y,y-z) \in [0,\pi/2]$ is the smaller angle between
the vectors $(x-y)$ and $(y-z) \in \mathbb{R}^3$, and
\[
{\rm diam}(M):= \sup_{x,y \in M} |x-y| \; \mbox{for} \; M \subset \mathbb{R}^3
\]
is the diameter of the set $M$.
The {\it thickness of $\gamma$}, denoted as
\begin{equation} \label{1.1} \Delta[\gamma] := \inf_{s \in \mathbb{S}} \rho_G[\gamma](s)
= \inf_{s,\sigma,\tau \in \mathbb{S}, \sigma \not= \tau,\sigma\not=s,\tau\not=s}
R(\gamma(s),\gamma(\sigma), \gamma(\tau)),
\end{equation}
is defined as the infimum of $\rho_G$.
\end{definition}
A curve that minimizes arclength over thickness is called an ideal
knot \cite{KBMSDS96,GM99}.
Already \cite{GM99} showed in a $C^2$-setting that for a knot to be ideal,
$\rho_G$ around a parameter\footnote{The proof from \cite{GM99} only requires the curve to be $C^2$ on a neighborhood of the parameter, not everywhere.}
is either constant and equal to
the infimum, or the curve is locally a straight line.
A proof of this necessary condition for $C^{1,1}$ curves
is not known yet.
Assume for a moment\footnote{So far, the numerical shapes suggest that most ideal knots are $C^2$ except for a finite number of points.},
that $\gamma$ is ideal and $C^2$, then for each $t \in \mathbb{S}$ we distinguish
the following three cases \cite{GM99,LSDR99}:
\begin{enumerate}
\item[(A)] $\rho_G(t) > \Delta[\gamma]$ and there
exists $\varepsilon>0$ such that $\gamma(\{s: |s-t|<\varepsilon\})$ is a straight line.
\item[(B)] $\rho_G(t) = \Delta[\gamma]$ and the curvature of $\gamma$ at $t$ is $1/\Delta[\gamma]$.
\item[(C)] $\rho_G(t) = \Delta[\gamma]$ and
there exists a $s \in \mathbb{S}$ with $|\gamma(s)-\gamma(t)|=2\Delta[\gamma]$ and
$\langle\gamma'(s), \gamma(s)-\gamma(t)\rangle = \langle\gamma'(t), \gamma(s)-\gamma(t)\rangle=0$.
\end{enumerate}
In case (B) we say that global curvature is attained locally, or that curvature is active, while in case (C) we say
that the contact is global.\footnote{For $C^{1,1}$-curves the situation is less clear but the cases (B) and (C) remain interesting.}
The global contact is realized by a contact chord.
\begin{definition}[Contact Chord]\label{def:contact_chord}
Let $\gamma \in C^1(\mathbb{S},\mathbb{R}^3)$ be a regular, i.e. $|\gamma'(s)|>0$, curve with $\Delta[\gamma]>0$ and
let $s,t \in \mathbb{S}$ be such that $c(s,t) := \gamma(t)-\gamma(s)$ has length
\[
|c(s,t)|=2\Delta[\gamma],
\]
and $c(s,t)$ is orthogonal to $\gamma$, i.e.
$\langle\gamma'(s), c(s,t))\rangle = \langle\gamma'(t), c(s,t)\rangle=0,$
then we call $c(s,t)$ a {\it contact chord}.
If such $s$ and $t$ exist, we say $\gamma$ has a
contact chord connecting $\gamma(s)$ and $\gamma(t)$ or the parameters
$s$ and $t$ are (globally) in contact.
The set
\[
\{\gamma(s)+h c(s,t): h \in [0,1]\} \subset \mathbb{R}^3
\]
will also be called a contact chord. Being in contact is a symmetric
relation.
\end{definition}
\begin{figure}
\begin{center}
\begin{tabular}{|ll|}
\hline
Name & \verb:k3_1:\\
Degrees of freedom & 165 \\
Biarc nodes & 333 \\
Arclength $\scriptL$ & 1 \\
Thickness $\Delta$ & 0.030539753 \\
Ropelength $\scriptL/\Delta$ & 32.744208 \\
\hline
\end{tabular} \\
\vspace{.4cm}
\begin{tabular}{ccc}
\includegraphics[width=.3\textwidth]{billiard_struts.png} &
\includegraphics[width=.3\textwidth]{k3_1_sym_1.png} &
\includegraphics[width=.3\textwidth]{k3_1_sym_2.png} \\
(a) & (b) & (c)
\end{tabular}
\end{center}
\caption{In the top box we describe the data for the trefoil used
for the computations.
In the bottom row we have (a) the trefoil with a
few contact chords. The dark chords visualize the closed cycle
in the contact chords of the trefoil. The pictures (b) and
(c) illustrate two different views of plausible symmetry axes.
The $120^\circ$ rotation axis and the three $180^\circ$ rotation axes are decorated with a prism and an ellipsoid at
the end, respectively.
\label{fig:billiard_struts} }
\end{figure}
For the rest of the article, we will restrict ourselves to the ideal trefoil\footnote{It is widely assumed that the ideal
trefoil is unique but it remains to be proven rigorously.} $\gamma_{3_1}$.
The trefoil data used in this article was computed as in \cite{CG10}.
A Fourier representation of the knot makes enforcing symmetries
natural. The specific symmetries are proposed in Conjecture \ref{con:sym}.
They significantly reduce the number of independent Fourier parameters in
simulated annealing \cite{L98}, while the computation
of the thickness is done by interpolating the Fourier knot
with biarcs \cite{bookchapter,CG10}. The Fourier coefficients
and the point-tangent data for the trefoil are available online
\footnote{Data is available at \cite{Data10}.
\texttt{k3\_1.3} with MD5 sum \texttt{cf5e2f8550c4c1e91a2fd7f5e9830343}
and \texttt{k3\_1.pkf} with sum \texttt{531492b73b2ec4be2829f6ab2239d4d5}.
} (see also Figure \ref{fig:billiard_struts}).
The numeric approximations of the ideal trefoil suggest that every point is globally in contact
with two other points on the trefoil \cite{bookchapter}.
We can sort the contact chords in a continuous fashion, such that each point has an incoming and an outgoing
contact.
In our numerical computations of the contact chords we used the point-to-point
distance function
\[
{\rm pp}(s,\sigma):=|\gamma(s)-\gamma(\sigma)|.
\]
The general belief is that the ${\rm pp}$-function of the ideal trefoil has an extremely flat double valley away from the diagonal
\cite{bookchapter} (see Figure \ref{fig:pp3d} for a 3D version of ${\rm pp}(s,\sigma)$ for the trefoil).
For a sampling $s_i:=i/n$, $i=0,\ldots,n$, we compute $\sigma_i$ as the minimum of ${\rm pp}(s_i,\cdot)$
restricted to the region $[\sigma_{i-1}-\varepsilon,\sigma_{i-1}+\varepsilon]$,
$1\gg\varepsilon>0$.
The initial
value $(s_0,\sigma_0)$ is computed as the local minimum away from the diagonal.
We now choose one of the two valley floors. By staying close to the previously computed minimum,
we never cross over to the second valley. We then
linearly interpolate between the $(s_i,\sigma_i)$ pairs to obtain an approximation of the so called contact
function $\sigma$ (also see Figure \ref{fig:connect} below for a top view of $\sigma$):
\begin{figure}
\begin{center}
\includegraphics[width=.8\textwidth]{distance_plot.png}
\end{center}
\caption{The distance function ${\rm pp}(s,\sigma)$ for $s,\sigma\in \mathbb{S}$ goes to $0$ on the diagonal and forms a large valley with two very shallow sub-valleys.
The dotted lines marked with arrows in the valley indicate the two local minima, i.e. $\sigma(s)$ and
$\tau(s)$.}
\label{fig:pp3d}
\end{figure}
\begin{definition}[Contact functions]\label{def:sigma}
Let $\sigma:\mathbb{S}\longrightarrow\mathbb{S}$ be a continuous, bijective and orientation preserving function, such that
$\gamma_{3_1}(s)-\gamma_{3_1}(\sigma(s))$ is a contact chord for every $s \in \mathbb{S}$.
The inverse function of $\sigma$ is $\tau:=\sigma^{-1}$.
\end{definition}
As mentioned previously, numerics point out that the trefoil is
symmetric with respect to a specific symmetry group \cite{bookchapter,BPP08,CG10}.
These symmetries have helped to identify the closed cycle proposed
later in this article.
\begin{conjecture}[Symmetry of the ideal trefoil]\label{con:sym}
The ideal trefoil has symmetries as shown in Figure \ref{fig:billiard_struts}.
\end{conjecture}
The symmetries of the trefoil are also apparent in its contact functions $\sigma$
and $\tau$. The relations are listed in the following lemma.
\begin{lemma}[Symmetry of $\sigma,\tau$]\label{lem:sigma_sym}
Assume Conjecture \ref{con:sym} about the symmetry of the constant speed parameterized trefoil
$\gamma_{3_1}: \mathbb{S} \longrightarrow \mathbb{R}^3$ is true. Then the contact
functions $\sigma,\tau: \mathbb{S}\longrightarrow \mathbb{S}$ have the following properties:
\begin{eqnarray}
\sigma(s^*+t) &\stackrel{\text{in\;} \mathbb{S}}{=}& -\tau(s^*-t), \quad \forall t\in \mathbb{S} \\
\sigma(t+1/3) &\stackrel{\text{in\;} \mathbb{S}}{=}& \sigma(t)+1/3,\quad \forall t\in \mathbb{S} \\
\tau(s^*+t) &\stackrel{\text{in\;} \mathbb{S}}{=}& -\sigma(s^*-t), \quad \forall t\in \mathbb{S} \\
\tau(t+1/3) &\stackrel{\text{in\;} \mathbb{S}}{=}& \tau(t)+1/3, \quad \forall t\in \mathbb{S}
\end{eqnarray}
where $s^*\in\mathbb{S}$ is a parameter such that $\gamma_{3_1}(s^*)$ is on a $180^\circ$ rotation axis. \qed
\end{lemma}
\section{Closed Cycles} \label{sec:billiards}
\begin{comment}
This section deals only with the ideal trefoil $\gamma_{3_1}$.
While some ideas may carry over to other ideal knots, the available numerics are not sufficiently converged
to be compelling.
\end{comment}
Recall from the previous section that numerics suggest that every point on the ideal trefoil $\gamma_{3_1}$
is in contact with two other points and we assume to be able to define a contact
function $\sigma$ as in Definition \ref{def:sigma}.
Is there a finite tuple of points such that each parameter is in contact with, and only with, its cyclic
predecessor and successor?
Inspired by {\it Dynamical Systems} \cite{B27} we call a sequence of parameters
that are in contact with each predecessor a billiard.
If a billiard closes, we call it a cycle:
\begin{definition}[Cycle]
For $n \in \mathbb{N}$ let $b:=(t_0,\ldots,t_{n-1}) \in \mathbb{S}\times\cdots\times\mathbb{S}$ be an $n$-tuple.
We call $b$ an $n$-cycle if $\sigma(t_i)=t_{i+1}$ for $i=0,\ldots,n-2$ and $\sigma(t_{n-1})=t_0$, where
$\sigma$ is defined as in Definition \ref{def:sigma}.
The cycle $b$ is called minimal if all $t_i$ are pairwise distinct.
\end{definition}
Each cyclic permutation of a cycle is again a cycle. Basing the definition of cycles on the
continuous function $\sigma$ instead of closed polygons in $\mathbb{R}^3$ makes it slightly easier to find them numerically:
\begin{remark}
The $\gamma_{3_1}$ has an $n$-cycle iff there exists some $t \in \mathbb{S}$ such that
\[
\sigma^n(t):=\underbrace{\sigma\circ\cdots\circ\sigma}_{n \text{ times }}(t)=t.
\]
The cycle is then $b:=(t,\sigma^1(t),\ldots,\sigma^{n-1}(t))$. \hfill$\Box$
\end{remark}
All parameters of a minimal $n$-cycle are pairwise distinct so each minimal $n$-cycle corresponds
to $n$ points in the set $\{t \in \mathbb{S}: \sigma^n(t) = t\}$. Since there are $n$ cyclic permutations of an $n$-cycle
and since minimal $n$-cycles that are not cyclic permutations must be point-wise distinct this leads to:
\begin{lemma}[Counting Cycles]\label{lem:counting}
Define the set of intersections of $\sigma^n$ with the diagonal $I := \{t \in \mathbb{S}: \sigma^n(t) = t\}$.
If there is a finite number of minimal $n$-cycles then
\begin{eqnarray*}
\#I &\ge& (\text{Number of distinct minimal $n$-cycles})\cdot n \\
& = & (\text{Number of minimal $n$-cycles}).
\end{eqnarray*}
\qed
\end{lemma}
In Figure \ref{fig:sigmatable} we compiled small plots of $\sigma^n$ for $n=1,\ldots,27,144$.
For $n=2$ the
function $\sigma^2$ comes close to the diagonal for the first time, but can not touch it in less than three points
by Lemma \ref{lem:sigma_sym}. If it would touch it would have to touch at least six times by
Lemma \ref{lem:counting}, which does not seem to be the case.\par
By similar arguments, we exclude the possibility of cycles for $n=7,$ $11,$ $13,$ $16,$ $20$ and $25$.
On the other hand the
case $n=9$ looks promising (see Figure \ref{fig:sigma9}). It seems to touch the diagonal precisely nine times which
suggests the existence of a single minimal cycle $b_9$ and its cyclic permutations. With our parameterization
the cycle $b_9$ happens to start at $0$ and we compute a numerical error of only $\sigma^9(0) = 0.0007 \approx 0$.
Consequently the cases $n=18,27$ would also touch the diagonal, but the corresponding cycle would not be minimal.
We studied the plots till $n=100$, but did not find any other promising candidates (apart from $n=k\cdot 9$ for $k\in\mathbb{N}$).
Keep in mind that the numerical error increases with $n$, but even for $n=144$ the graph looks reasonable.
\begin{figure}
\begin{center}
\begin{tabular}{ccccc}
\includegraphics[width=\smallpic]{sigma_fold_001.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_002.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_003.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_004.pdf} \\
\includegraphics[width=\smallpic]{sigma_fold_005.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_006.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_007.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_008.pdf} \\
\includegraphics[width=\smallpic]{sigma_fold_009.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_010.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_011.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_012.pdf} \\
\includegraphics[width=\smallpic]{sigma_fold_013.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_014.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_015.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_016.pdf} \\
\includegraphics[width=\smallpic]{sigma_fold_017.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_018.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_019.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_020.pdf} \\
\includegraphics[width=\smallpic]{sigma_fold_021.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_022.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_023.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_024.pdf} \\
\includegraphics[width=\smallpic]{sigma_fold_025.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_026.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_027.pdf} &
\includegraphics[width=\smallpic]{sigma_fold_144.pdf}
\end{tabular}
\end{center}
\caption{
Plots of $\sigma^n:\mathbb{S}\longrightarrow\mathbb{S}$. In plot 09 the function $\sigma^9$ seems to touch the
diagonal (see also Figure \ref{fig:sigma9} for an enlarged plot).
Accordingly $\sigma^i$ touches the diagonal in
plot $i$ for $i \in \{9,18,27,\dots\}$. Note that in plot 144 the numerical errors have added up so that
$\sigma^{144}$ no longer touches the diagonal.
} \label{fig:sigmatable}
\end{figure}
We believe that $b_9$ is indeed a cycle (see Figure \ref{fig:billiard}):
\begin{conjecture}[Existence of nine-cycle] \label{con:ninebilliard}
Let $\gamma_{3_1}: \mathbb{S} \longrightarrow \mathbb{R}^3$ be the ideal trefoil,
parameterized with constant speed such that $\gamma_{3_1}(0)$ is the outer point of the trefoil on a symmetry axis.
Then $b_9=(s_0,\ldots,s_8)$ with $s_i:=\sigma^i(0)$ is a nine-cycle. Numerics suggest that $\gamma_{3_1}$ passes
from 0 to 1 through $s_i$ in the sequence: $s_0, s_7, s_5, s_3, s_1, s_8, s_6, s_4, s_2$.
\end{conjecture}
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{sigma9.pdf}
\caption{The plot shows graph number $9$ from Figure \ref{fig:sigmatable}
rotated by $45^\circ$. It seems to touch the diagonal 9 times
in the points $s_0=0$, $s_7=0.159$, $s_5=0.175$,
$s_3=0.334 \approx 1/3$, $s_1= 0.492$, $s_8=0.508$,
$s_6=0.667 \approx 2/3$, $s_4=0.826$, $s_2=0.841$.
This indicates the existence of a nine-cycle $b_9=(0,\sigma^1(0),\ldots,\sigma^8(0))$.
} \label{fig:sigma9}
\end{center}
\end{figure}
Note that $b_9$ partitions the trefoil in 9 parts (see Figure \ref{fig:billiard}): Three curves
\begin{eqnarray*}
\beta_1 &:=& \gamma_{3_1}|_{[s_0,s_7]}, \\
\beta_2 &:=& \gamma_{3_1}|_{[s_6,s_4]}, \\
\beta_3 &:=& \gamma_{3_1}|_{[s_3,s_1]}, \\
\text{or} \; \beta_i &:=& \gamma_{3_1}|_{[s_{6(i-1)},s_{6(i-1)-2}]} \; \text{with} \; s_k=s_{k+9}, \\
\end{eqnarray*}
which are congruent by $120^\circ$ rotations around the $z$-axis.
Another three curves
\begin{eqnarray*}
\tilde{\beta_1} &:=& \gamma_{3_1}|_{[s_2,s_0]}, \\
\tilde{\beta_2} &:=& \gamma_{3_1}|_{[s_8,s_6]}, \\
\tilde{\beta_3} &:=& \gamma_{3_1}|_{[s_5,s_3]}, \\
\text{or} \; \tilde{\beta_i} &:=& \gamma_{3_1}|_{[s_{6(i-1)+2},s_{6(i-1)}]} \; \text{with} \; s_k=s_{k+9}, \\
\end{eqnarray*}
which are again congruent by $120^\circ$ rotations and with each $\tilde{\beta_i}$ congruent to $\beta_i$ by a
$180^\circ$ rotation.
And finally three curves
\begin{eqnarray*}
\alpha_1 &:=& \gamma_{3_1}|_{[s_1,s_8]}, \\
\alpha_2 &:=& \gamma_{3_1}|_{[s_7,s_5]}, \\
\alpha_3 &:=& \gamma_{3_1}|_{[s_4,s_2]}, \\
\text{or} \; \alpha_i &:=& \gamma_{3_1}|_{[s_{6(i-1)+1},s_{6(i-1)-1}]} \; \text{with} \; s_k=s_{k+9}, \\
\end{eqnarray*}
which are congruent by rotations of $120^\circ$ and self congruent by a rotation of $180^\circ$.
Because $b_9$ is a cycle, each piece of the curve gets mapped one-to-one to another piece of the curve.
\begin{lemma}[Piece to piece] \label{lem:piece2piece}
Assume that the ideal trefoil admits a contact function $\sigma$ as in Definition
\ref{def:sigma} and
Conjectures \ref{con:sym}, \ref{con:ninebilliard} about symmetry and the
existence of a nine cycle $b_9=(s_0,\ldots,s_8)$ hold. Then $\sigma$ maps each parameter
interval $[s_i,s_j]$ to $[s_{i+1},s_{j+1}]$.
In particular: Following the contact in $\sigma$ direction
we get the sequence
$\alpha_1\to\tilde{\beta_1}\to\beta_3\to \alpha_3\to\tilde{\beta_3}\to\beta_2\to
\alpha_2\to\tilde{\beta_2}\to\beta_1 (\to \alpha_1)$.
Each piece is in one-to-one contact with the next in the sequence
(see also Figure \ref{fig:connect}).
\end{lemma}
\begin{proof} By definition $s_i$ is mapped to $s_{i+1}$ and by Definition
\ref{def:sigma} the contact function $\sigma$ is continuous and orientation
preserving so the interval $[s_i,s_j]$ gets mapped to
$[\sigma(s_i),\sigma(s_j)] = [s_{i+1},s_{j+1}]$.
\end{proof}
One further remark about the plots in Figure \ref{fig:sigmatable}. If $s_i \in \mathbb{S}$ is a solution of $\sigma^9(s_i)=s_i$ then by Lemma
\ref{lem:sigma_sym} the parameter $r=s_i+k/3$ is also a solution of $\sigma^9(r)=\sigma^9(s_i+k/3)=s_i+k/3=r$ for $k \in \{0,1,2\}$. Since there are
presumably only nine solutions $s_i$, they happen to fall in three classes represented by $s_0, s_1, s_2$ and with $s_{i+3k} = s_i+k/3$ the remaining six
are defined. Consequently we find that $\sigma^3(s_i)=s_{i+3}=s_i+1/3$, i.e. there are nine solutions
of $\sigma^3(s)=s+1/3$ which can be seen in plot number 3 of Figure \ref{fig:sigmatable}.
Similarly, there are nine solutions of $\sigma^6(s)=s+2/3$ in plot number 6 and so on.
We now briefly discuss the relationship between particular points in
the curvature plot and the closed-cycle points $(s_0,\cdots,s_8)$,
i.e. the partitioning introduced above. In Figure \ref{fig:curvature}
we show the curvature plot scaled by the thickness $\Delta$ on the interval $[0,1/3] = [s0,s3]$.
Since curvature is confined in $[0,1/\Delta]$ for thick knots this always gives a comparable graph.
Due to the $3$-symmetry, the plots on the intervals $[1/3,2/3]$ and $[2/3,1]$ are identical.
The $180$-degree rotation symmetry shows up in the plot as a symmetry
around $(s_7+s_5)/2$, the center of a self-congruent piece $\alpha_2$.
The curvature profile is close to constant $.5$ on the major part
$\beta_1$ and $\tilde{\beta}_3$.
A significant change occurs at the transition points
between $\alpha_i$ and $\beta_i$, where it reaches its maximum
at the junction points $s_7$ and $s_5$, where curvature is believed to
be active\cite{bookchapter, BPP08}.
The spikes of our computation do not achieve the maximal value, and there is a
local maximum at the center of an $\alpha_i$ piece. We believe these deviations from
earlier observations are numerical artefacts due to the Fourier
representation used to compute this trefoil.\footnote{In fact the curvature function needs not even to
converge, as one approaches an ideal shape \cite[Section 2.5]{Ge10}}.
The alignment of the closed-cycle points and the points where
curvature seems active only enforces that all the numerical pieces
fit together nicely, which is a good indication, that these are not
numerical artefacts.
\begin{figure}
\begin{center}
\includegraphics[width=0.6\textwidth]{connect.pdf}
\caption{Plotting the $\sigma$ function with a grid of the partition points $s_i$ we can read off which piece of
the curve is in contact (in the $\sigma$ direction) with which other piece. For example starting at the
top with $\alpha_1$ we see that $\sigma$ maps its parameter-interval to the parameter-interval
of $\tilde{\beta_1}$ on the right, which itself gets mapped from top $\tilde{\beta_1}$ to $\beta_3$
on the right and so on (see Lemma \ref{lem:piece2piece}).
} \label{fig:connect}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{curvature}
\end{center}
\caption{The curvature $\kappa$ of $\gamma_{3_1}$ is confined to $[0,1/\Delta]$. Consequently the graph shows $\kappa\cdot\Delta$.
Since $\kappa$ is three-periodic, for greater detail we show only one third of the interval.
The maximal curvature is attained at $s_7$ and $s_5$.
}
\label{fig:curvature}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{attractor.pdf}
\caption{
Since numerical experiments find that $\sigma^{i+9}-\sigma^i$ converges point-wise to $0$
for $i\to\infty$ we conjecture that
the cycle $b_9$ acts as an attractor, i.e. $(\sigma^{9i}(s), \sigma^{9i+1}(s), \ldots,\sigma^{9i+8}(s))$ converges
to $b_9$ up to a cyclic permutation. Notice that the convergence is only point-wise
and cannot be uniform since $\sigma$ is continuous; with enough samples we would
see large spikes after each $s_i$ as behind $s_6$ and $s_8$ above.
} \label{fig:attractor}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{attractor_lines.pdf}
\caption{The cycle $b_9$ seems to be an attractor: starting at the arbitrary point $t=0.05$ the sequence $\sigma^{9i}(t)$
seems to converge to $s_7$ for growing $i$. After $i>16$ iterations, presumably, errors in the
approximation and the numerics add up and $\sigma^{9\cdot 18}(t)$ is in the next
interval $(s_7,s_5)$. Starting from $\sigma^{k}(t)$ in other intervals shows a similar behavior.
}
\label{fig:attractorlines}
\end{center}
\end{figure}
Taking a second look at Figure \ref{fig:sigmatable} it looks like $\sigma^n$ is approaching a
step function as $n$ increases. What are the accumulation points of the sequence $\{\sigma^i(t)\}_i$ as a function of
$t\in\mathbb{S}$? Looking at $(\sigma^{i}(t), \sigma^{i+1}(t), \ldots, \sigma^{i+8}(t))$ for arbitrary $t \in \mathbb{S}$
it seems to converge to $b_9$ up to a cyclic permutation for $i$ large enough, i.e. $b_9$ contains the
accumulation-points of the above sequence. Figure \ref{fig:attractor} shows some numeric values
of $\sigma^{i+9}-\sigma^i$ which seems to converge
point-wise
to $0$ for $i\to\infty$. An arbitrary point $t$ between neighboring points $s_l$ and $s_r$ gets by each application of
$\sigma^9$ repelled from the left by $s_l$ and attracted to the right by $s_r$ (see
Figure \ref{fig:attractorlines}).\footnote{We would like to thank E. Starostin for encouraging
us to take a closer look at this issue.}
Note that the attactor has a direction that is induced by the chirality of the trefoil
(left or right-handed) and the choice of the contact function $\sigma$ made in Definition \ref{def:sigma}.
\begin{conjecture}[Attractor]\label{con:attractor}
Let $b_n \in (\mathbb{S})^n$ be a cycle. We call $b_n$ an attractor if
for any $t \in \mathbb{S}$ and fixed $k \in \{0,\ldots,n-1\}$ the $n$-tuple
$(\sigma^i(t), \sigma^{i+1}(t), \ldots,\sigma^{i+n-1}(t))$
converges to a cyclic permutation of $b_n$ for
$j\to\infty$ with $i=nj+k$.
The $b_9$ cycle of Conjecture \ref{con:ninebilliard} is an attractor.
\end{conjecture}
The existence of an attractor rules out the existence of other cycles:
\begin{lemma}\label{lem:one_attractor}
Assume $\gamma_{3_1}$ has a contact function $\sigma$ as in Definition \ref{def:sigma} and
let $b_n \in (\mathbb{S})^n$ be a $n$-cycle as in Conjecture \ref{con:ninebilliard}.
Then $b_n$ is an attractor in the sense of Conjecture \ref{con:attractor}
iff there is no other $n$-cycle on $\gamma_{3_1}$.
\end{lemma}
\begin{proof}
Assume that $c_n=(t_0,\sigma(t_0),\ldots,\sigma^{n-1}(t_0))$ is a $n$-cycle different from $b_n$.
Then $\sigma^{n}(t_0)=t_0$ and $b_n$ cannot be an attractor.
To prove the converse, let $b_n=(s_0,\ldots,s_{n-1})$ be an $n$-cycle and
consider $\sigma^n$ as a continuous, injective and orientation preserving
map from $[s_k,s_l] \subset \mathbb{R}$ to itself, where $l$ and $k$
are such that $s_l$ and $s_k$ are neighboring, i.e. $s_i \not\in (s_k,s_l)$ for all $i$.
The cycle $b_n$ is an attractor iff for all $x \in (s_k,s_l]$ the sequence
$x_i:=\sigma^{ni}(x)$ converges to $s_l$ as $i \to \infty$.
Since $\sigma$ is orientation preserving we have $x_i \le x_{i+1}$, i.e. the sequence is monotone.
Assume that $b_n$ is not an attractor, i.e. for some $x$ the sequence
$\{\sigma^{ni}(x)\}_i$ is bounded away from $s_l$, then
it must converge to some smaller value $c < s_l$. By continuity of $\sigma^n$ it follows that $c$ is a fixed point.
Therefore $c_n:=(c,\sigma^1(c),\sigma^2(c),\ldots,\sigma^n(c))$ is a closed $n$-cycle different from $b_n$.
\end{proof}
As mentioned above, by concatenation, a minimal cycle gives rise to a series of larger, non-minimal cycles:
For an $n$-cycle $a_n=(x_0,\ldots,x_{n-1})$ we define an $nk$-cycle
\[
a^k_{n} := (\underbrace{x_0,\ldots,x_{n-1},\ldots,x_0,\ldots,x_{n-1}}_{k \text{ times}}).
\]
\begin{proposition}\label{prop:one_cycle}
Assume $\gamma_{3_1}$ has a contact function $\sigma$ as in Definition \ref{def:sigma} and
let $b_n \in (\mathbb{S})^n$ be a minimal $n$-cycle as in Conjecture \ref{con:ninebilliard} and an
attractor in the sense of Conjecture \ref{con:attractor}.
Then $b_n$ is the only minimal cycle and all other cycles are multiples of $b_n$.
\end{proposition}
\begin{proof}
Let $c_m$ be a $m$-cycle different from $b_n$.
We claim $c_m=b^k_n$ and $m=kn$ for some $k \in \mathbb{N}$.
If $b_n$ is an attractor, then $b^i_n$ is also an attractor for $i \in \mathbb{N}$.
Let $g$ be the greatest common divisor of $n$ and $m$, so $l = mn/g$ is their least common multiple.
Then $b_n^{m/g}$ is an attractor and an $l$-cycle, but $c_m^{n/g}$ is an $l$-cycle as well and Lemma \ref{lem:one_attractor}
implies $b^{m/g}_n=c^{n/g}_m$. Since $b_n$ was minimal, this is only possible if $c_m=b^k_n$ for some $k\in\mathbb{N}$.
\end{proof}
Lemma \ref{lem:one_attractor} and Proposition \ref{prop:one_cycle} fit well with our numerical observation. We find only one possible cycle and it seems to
be an attractor.
\section{Conclusion}
We have presented numerical and esthetical compelling evidence for the existence of
a closed nine-cycle in the contact chords of the
ideal trefoil knot. Enforcing symmetry based on a Fourier representation
turned out to be essential to observe this feature. The cycle leads to a
partitioning of the trefoil. Only two segments of the curve have to
be considered, the remaining parts of the trefoil can be reconstructed
by symmetry. For other contact chord paths, after enough iterations,
it seems that they eventually converge to the nine-cycle. So the
closed cycle acts as an attractor for all other billiards.
Preliminary numeric experiments by E. Starostin suggest closed cycles
in ideal shapes of knots with a higher number of crossings as well.
The interesting cases remain however inconclusive,
since these knot shapes are believed to be much less ideal than the trefoil.
Closed cycles in these knots
might then also suggest a natural partitioning of the curves, therefore
improving the understanding of these knots.
In the $\mathbb{S}^3$ setting \cite{Ge10} suggested a candidate trefoil for
ideality to the problem of maximizing thickness.
Each point on the $\mathbb{S}^3$ trefoil is in contact with two other
points on the curve. Following the contact great-arcs (in $\mathbb{S}^3$)
five times forms a circle, i.e. a 5-cycle.
The numerical computations suggest at least two new challenges: First, can we get new insights about the ideal trefoil assuming the existence of a nine-cycle?
And second, can we prove, under some reasonable hypothesis, that the
ideal trefoil or even every ideal knot has closed contact cycles?
\section{Acknowledgements}
Research supported by the Swiss National Science Foundation SNSF No. 117898
and SNSF No. 116740.
We would like to thank E. Starostin and J.H. Maddocks for
interesting discussions and helpful comments.
|
\section{Introduction}
The use of feedback to stabilize desirable structures or behaviors that would otherwise be unobservable is a topic of considerable interest in classical dynamics \cite{KCon1}.
Feedback concepts have been extended into the quantum realm \cite{QCon1,QCon2}, and have been shown to be effective in e.g. prolonging the life of Schr\"odinger cat states \cite{HK97}, increasing the rate of state reduction under measurement \cite{QCon4a}, and for state purification \cite{QCon4b}. Feedback schemes have also been proposed in the solid state, both theoretically \cite{QCon4c} and in a recent experiment \cite{QCon4d}.
In these latter examples, a quantum transport device (e.g. quantum point contact, single-electron transistor) acts as a detector of the system (typically a qubit) which is to be controlled. By feeding information gained from the measurement device into the system in a control loop, one can purify the quantum state of the qubit.
In this work, we take a different approach and couple information gathered about the charge transport {\em through} a device back into the device itself.
More specifically, we monitor the flow of electrons from the source lead to a device in the Coulomb blockade regime, as in full counting statistics (FCS)\cite{FCS_exp,FCS1, FCS2, FCS3}, and conditioned on electron jumps, perform unitary operations on the device. Using this scheme, we show how such feedback can lead to a purification of the device wave function, despite it being in non-equilibrium (this contrasts with the situation without control, where the state of the system is decidedly mixed, as one might expect). We view this behavior as a genuine state stabilization as these pure states are implicit in the behavior of the device without feedback; but only with the feedback do they become stable.
Furthermore, we show how successful purification of the transport state also leads to a distinctive signature in the statistics of electron flow through the device. We describe how this information may be used in the feedback loop to optimize the stabilization.
After presenting a general theory of state stabilization of non-equilibrium quantum transport, we consider the example of a double quantum dot (DQD) (Fig.~\ref{Fig1}) and show that the stabilization of coherent delocalized states is possible.
\begin{figure}[top]
\includegraphics[width=0.48\textwidth]{Fig1.eps}
\caption{\label{Fig1} Main part: After an electron jumps into the double quantum dot (DQD) (measured e.g. with a quantum point contact QPC) a pulsed control operation modifies e.g. the gate voltages $V_G$ and rotates the electron into a different state. In an effective model this control operation modifies the left jump operator.
The DQD is described by $\epsilon\mathord=\epsilon_L\mathord-\epsilon_R$ the detuning and $T_C$ the coupling strength between the dots. The electrons tunnel into/out the left/right dot with the rate $\Gamma_{L/R}$.
Inset: For sufficient control operations the DQD is stabilized. The transport can then be described by a single resonate level model with an effective right tunnel rate $\gamma_R^{(j)}$.
}
\end{figure}
\section{General theory of the control scheme}
\subsection{Transport model}
We start by considering a general transport Hamiltonian,
$H=H_\mathrm{S}+H_\mathrm{L}+H_\mathrm{T}$,
composed of
system (e.g. quantum dots (QD)s), leads and tunnel-coupling parts. We consider two leads with non-interacting Hamiltonian
$
H_\mathrm{L} = \sum_{k,\alpha}\epsilon_{\alpha k} c^\dagger_{\alpha k} c_{\alpha k}
$
with $c_{\alpha k}$ the annihilation operator of an electron in lead $\alpha=L,R$ (left, right) with quantum numbers $k$.
We assume strong Coulomb blockade and a bias configuration such that at most one excess electron can be in the system at any given time. The state-space of the system is thus spanned by the empty state $\ket{0}$ and $N$ single electron states $\ket{n}$, typically not eigenstates of system Hamiltonian $H_\mathrm{S}$. Finally we assume that the leads are each coupled to one and only one localized system level ($\ket{L}$ and $\ket{R}$ respectively) such that the tunnel Hamiltonian reads
$
H_\mathrm{T} = \sum_{ k}
t_{R k} c_{R k}^\dagger D_{R}
+t_{L k} c_{L k}^\dagger D_{L}^\dag
+\mathrm{H.c.}
$
with $D_{L}^\dagger\mathord=\op{L}{0}$ and $D_{R}^\dagger \mathord=\op{0}{R}$.
In the high-bias limit with tunnel rates $\Gamma_{\alpha} =2\pi \sum_{k} |t_{\alpha k}|^2 \delta(\epsilon-\epsilon_{\alpha k})$ assumed constant, the behavior of the system density matrix can be described with a Markovian master equation of Lindblad form
\cite{Gur}:
\begin{eqnarray}\label {master0}
\dot\rho = {\cal W}\rho = \Big( {\cal W}_0
+
{\cal J}_{L}
+
{\cal J}_{R} \Big) \rho
,
\end{eqnarray}
with jump super-operators
$
{\cal J}_{\alpha}\rho\mathord
= \Gamma_\alpha D^\dag_{\alpha} \rho D_{\alpha}
$, which transfer an electron to/from lead $\alpha = R/L$, and free Liouvillian ${\cal W}_0$, which describes the evolution of the system without electron transfer. This latter assumes the form
$ {\cal W}_0 \rho =
-i
\left\{
\widetilde{H} \rho - \rho \widetilde{H}^\dag
\right\}
\label{free-ev}
$ (we set $\hbar=1$),
where
\begin{eqnarray}
\widetilde{H} = H_S
- i \frac{1}{2} \sum_\alpha \Gamma_\alpha D^\dag_\alpha D_\alpha
, \label{newH}
\end{eqnarray}
is an effective {\it non-Hermitian} `Hamiltonian' operator for the system.
\subsection{Electron counting and control }
The evolution of the density matrix under \eq{master0} can be interpreted in terms of trajectories\cite{qu-traj}. The density matrix at time $t$ is then given as a sum of terms such as
\begin{align}
\int^t_0 dt_2
\int^{t_2}_0 dt_1
\Omega_0(t-t_2) {\cal J}_{R} \Omega_0(t_2-t_1) {\cal J}_{L}\Omega_0(t_1) \rho(t_0), \label{traj}
\end{align}
which describe all trajectories in which an electron jumps into the system at time $t_1$ ($ {\cal J}_{L}$), out again at time $t_2$ ($ {\cal J}_{R}$), with evolution in between described by the no-jump propagator $\Omega_0(t) = e^{{\cal W}_0 t}$.
As in FCS experiments, we assume that our measurement device is sensitive to single electron jumps. Conditioned then on the detection of an electron jump from the left lead, we set the control electronics to implement a control operation on the system which we consider here to act as an instantaneous unitary operation on the system \cite{QCon1}. In an experimental setup for the transport through QDs these control operations could be achieved, for example, with pulsed gate voltages.
This scheme modifies the quantum trajectories such that each left jump operator is followed by a control operation; we thus replace
${\cal J}_{L}$ with ${\cal J}_{L}^C={\cal C}_{L} {\cal J}_{L}$ in all trajectories like \eq{traj}. Re-summing all the trajectories, we arrive at an effective Liouvillian for the evolution of the density matrix with control that is the same as ${\cal W}$ in \eq{master0} but with ${\cal J}_{L} $ replaced by ${\cal J}_{L}^C$.
\subsection{Feedback Stabilization}
We now describe how this control schema can be used to stabilize certain states.
The effective ``Hamiltonian'', \eq{newH}, has right and left eigenstates
$ \widetilde{H} \ket{\psi_j}= \varepsilon_j \ket{\psi_j} $ and
$ \bra{\widetilde{\psi_j}} \widetilde{H}= \varepsilon_j \bra{\widetilde{\psi_j}}$,
%
which, in general, are non-adjoint: $\bra{\widetilde{\psi_j}} \ne \rb{\ket{\psi_j}}^\dag$,
and can be used to write $\widetilde{H}$ in its diagonal basis as
\begin{eqnarray}
\widetilde{H}
=
- i \frac{1}{2} \Gamma_{L} \op{0}{\widetilde{0}}
+\sum_{j=1}^N\varepsilon_j \op{\psi_j}{\widetilde{\psi_j}}
,
\end{eqnarray}
where we have separated the empty state since it is not coupled to other states by $\widetilde{H}$.
The eigen-operators of the free Liouvillian are
$\rho_{jk}\mathord = \op{\psi_j}{\psi_k}$ such that
${\cal W}_0 \rho_{jk}\mathord = -i \rb{\varepsilon_j - \varepsilon_k^*}\rho_{jk}$.
The diagonal matrices $\rho_{jj}$ represent pure states and obey
$
{\cal W}_0 \rho_{jj}\mathord =2 \Im( \varepsilon_j) \rho_{jj}
$.
The empty state is one such eigen-operator with
$
{\cal W}_0\rho_{00} = -\Gamma_L \rho_{00}
$.
The remaining pure density matrices $\rho_{jj};~j>0$ are states that we will seek to stabilize with our feedback scheme.
To effect the state stabilization, we choose the control operation ${\cal C}_{L}$ to rotate state $\ket{L}$ into the desired eigenstate of $\widetilde{H}$.
The effective left jump operator therefore acts as
%
$
{\cal J}^C_{L} \rho_{00} = \Gamma_L \rho_{jj},
$
where the $\Gamma_L$ forefactor arises from the normalization and the fact that the empty state decays with rate $\Gamma_L$ with or without control.
Once in state $\rho_{jj}$, no transitions out of this state are induced by ${\cal W}_0$. Rather, the only thing that happens to an electron in this state is that it leaves at a rate
$\gamma_R^{(j)} = - 2 \Im(\varepsilon_j)$.
By jumping directly into one of the eigenstates of the free Liouvillian $\ket{\psi_j}$ the dynamic of the system is then determined solely by the vacuum $\ket{\psi_0}$ and $\ket{\psi_j}$; the other states are decoupled. In the basis $\{\rho_{00},\rho_{jj}\}$ the Liouvillian of the feedback controlled system in the above parametrization can be written as
\begin{eqnarray} \label{Leff}
{\cal W}^{(j)}_C =
\rb{
\begin{array}{cc}
-\Gamma_{L} & \gamma_R^{(j)} \\
\Gamma_L & - \gamma_R^{(j)}
\end{array}
}.
\end{eqnarray}
This Liouvillian corresponds to an effective single resonant level, with the effective tunnel rate $\gamma_R^{(j)}$ and the stationary state
\begin{eqnarray}
\rho_\mathrm{stat} =
\frac{1}{\Gamma_{L} + \gamma_R^{(j)}}
\rb{
\gamma_R^{(j)}\op{0}{0}
+
\Gamma_{L}\op{\psi_j}{\psi_j}
}.
\end{eqnarray}
In the limit $\Gamma_{L} \gg \gamma_R^{(j)}$ the stationary state reduces to
\begin{eqnarray}
\lim_{\Gamma_{L}/\gamma_R^{(j)} \to \infty}\rho_\mathrm{stat} =\op{\psi_j}{\psi_j},
\end{eqnarray}
and the system is thus {\em stabilized} in the pure state $\ket{\psi_j}$.
In general then, a system with $N$ internal states has $N$ pure states that can be prepared in this manner. These states depend on the internal system parameters as well as on $\Gamma_R$. In the limit of $\Gamma_R\mathord\rightarrow 0$, the stabilized states will be eigenstates of $H_\mathrm{S}$. For finite $\Gamma_R$, however, these states are, in some sense, non-equilibrium pure states, since they depend on the rate $\Gamma_R$ and can be quite different from the eigenstates of the bare Hamiltonian $H_\mathrm{S}$. Needless to say, the states are unstable without control.
\section{Feedback stabilization of a double quantum dot}
In the following, we discuss the $N\mathord=2$ example of a charge qubit with Hamiltonian
$
H_\mathrm{S}=\frac{1}{2}\epsilon \sigma_z +T_C\sigma_x ,
$
with $\sigma_z \mathord=\op{L}{L}\mathord-\op{R}{R}$, $\sigma_x\mathord= \op{L}{R}\mathord+\op{R}{L}$, $\epsilon\mathord=\epsilon_L\mathord-\epsilon_R$ the detuning and $T_C$ the coupling strength between the dots.
In the infinite bias limit \cite{Gur,DQD1,Bra05} the transport through the DQD is described by \eq{master0}
with $D_L^\dagger \mathord= \op{L}{0}$ and $D_R^\dagger \mathord= \op{0}{R}$. The corresponding Liouvillian of the DQD is shown in appendix~\ref{App1}.
\subsection{Zero detuning $\epsilon=0$}
First we consider the case without detuning $\epsilon\mathord=0$ for simplicity.
The right-eigenvalues of $\widetilde{H}$ are $\varepsilon_0\mathord=\mathord-i \frac{\Gamma_L}{2}$
with eigenvector $\ket{\psi_0}\mathord=\ket{0}$ and
\begin{eqnarray} \label{eigendot}
\varepsilon_{\mp}=\frac{1}{4}\big(-i\Gamma_R\mp \kappa \big),\quad \kappa\mathord=\sqrt{16T_C^2-\Gamma_R^2},
\end{eqnarray}
with the eigenvectors
%
$\ket{\psi_\mp}=\mathcal N_{\mp}^{-1}\big[ (i\Gamma_R\mp \kappa) \ket{L}+4T_C\ket{R}\big]$,
normalized with
$\mathcal N_{\mp}\mathord=\sqrt{16T_C^2+|i\Gamma_R \mp \kappa |^2}$.
In the following, we consider the simplification $\Gamma_L \gg \Gamma_R,T_C $ which allows us to effectively project out the empty
dot state $\ket{\psi_0}$ and to assess the purity of the transport qubit in the usual Bloch representation
with $\ew{\sigma_x}\mathord= 2\Re(\rho_{LR})$,
$\ew{\sigma_y}\mathord= 2\Im(\rho_{LR})$ and $\ew{\sigma_z}\mathord= \rho_{LL}\mathord-\rho_{RR}$.
The pure-state density matrices $\rho_{\mp\mp}\mathord=\op{\psi_\mp}{\psi_\mp}$ to be stabilized correspond to two of the occupied eigenstates of the free Liouvillian ${\cal W}_0$.
Fig.~\ref{Fig2}(a) depicts these stabilisable pure states on the surface of the Bloch sphere.
For $\Gamma_R\mathord>4 T_C$, they describe partly localized states
with
\begin{align}
\ew{\sigma_z}\mathord=& \mp \frac{\tilde{\kappa}}{\Gamma_R},&
\ew{\sigma_x}\mathord=&0\quad &&\text{and} &
\ew{\sigma_y}\mathord=&\frac{4T_C}{\Gamma_R}, \label{smallTc}
\intertext{and for $\Gamma_R\mathord<4 T_C$
completely delocalized states with}
\ew{\sigma_z}\mathord=& 0,&
\ew{\sigma_x}\mathord=&\mathord \mp \frac{\kappa }{4T_C} &&\text{and}&
\ew{\sigma_y}\mathord=&\frac{\Gamma_R}{4T_C}, \label{bigTc}
\end{align}
with $\tilde{\kappa}\mathord=\sqrt{\Gamma_R^2-16T_C^2}$.
At $\Gamma_R\mathord=4T_C$ the eigenvalues $\varepsilon_\mp$ are degenerate and the corresponding eigenstates cross each other with $\ew{\sigma_z}\mathord=0$, $\ew{\sigma_x}\mathord=0$ and $\ew{\sigma_y}\mathord=1$.
\begin{figure}[top]
\includegraphics[width=0.48\textwidth]{Fig2.eps}
\caption{
(a) Stabilisable double quantum dot states $\rho_{\mp\mp}$ on the surface of the Bloch sphere, and stationary state $\rho_{\rm wo}$ without control, for tunnel coupling $T_C\in \{0,\infty\}$ . The triangle marks the state which is stabilized in (b).
(b)
Time evolution of Bloch vector length $|\ew{\mbox{\boldmath$\sigma$}}|$ with control operation switched on at $t=8 \Gamma_R^{-1}$ for different temperatures $T$. Parameters: $T_C=2\Gamma_R$, phonon coupling $g=8\cdotp 10^{-4}$, Debye-cutoff $\omega_c=500\Gamma_R$ and control parameters $\theta=\arccos(3 \sqrt{7/127})$ and
$\theta_C=\text{arcsec}(8 \sqrt{2})$.
\label{Fig2}}
\end{figure}
Without control an electron simply tunnels into the left dot (and thus into a localized, pure state).
The goal of the control operation is to rotate this state into one of the pure states $\rho_{\pm\pm}\mathord \equiv \op{\psi_\pm}{\psi_\pm}$.
In general, we can consider every qubit rotation as a possible control operation ${\cal C}_L \mathord=e^{\theta_C \bld{n}_0\cdot \vec{\Sigma}}$ where $\vec{\Sigma}$ corresponds to the Pauli matrices in Liouville space $\vec{\Sigma}\rho\mathord=\mathord-i2[ \mbox{\boldmath$\sigma$},\rho]$ with the usual Pauli matrix vector $\mbox{\boldmath$\sigma$}$,
the unit vector $\bld{n}_0\mathord=\{n_x,n_y,n_z\}\mathord=\{\sin{\theta},0,\cos{\theta}\}$ that determines the angle of the rotation in the $\sigma_x$-$\sigma_z$ plane, and the control strength parameter $\theta_C$.
A straightforward calculation (see appendix \ref{App2}) leads to the control parameters $\theta\mathord=\pi/2$ and $\theta_{C}\mathord=\arccos\rb{\sqrt{{(\Gamma_R \pm \tilde{\kappa})}/{2\Gamma_R}}}$ for $\Gamma_R\mathord>4 T_C$, and
$\theta\mathord=\arccos\rb{\mathord\pm\tilde{\kappa}/\sqrt{{\Gamma_R^2-32T_C^2}}}$ and
$\theta_{C}\mathord=\text{arcsec}\rb{{4\sqrt{2}T_C}/{\Gamma_R}}$ for $\Gamma_R\mathord<4 T_C$, corresponding to the stabilization of $\rho_{\pm\pm}$. The upper left plot of Fig.~\ref{Fig3n} shows these control parameter on the control plane.
The fidelity of our control scheme is characterized by the length of the Bloch vector $|\ew{\mbox{\boldmath$\sigma$}}|$, obtained by solving \eq{master0} for the DQD model.
Fig.~\ref{Fig2}(b)(solid line) demonstrates the effect of the control operation $\mathcal{C}_L$ on the DQD Bloch vector length $|\ew{\mbox{\boldmath$\sigma$}}|$ --- switching on the control (at time $t=8\Gamma_R^{-1}$ in the plot) leads to a rapid purification of the transport qubit state, even though electrons continuously tunnel through the DQD.
\begin{figure}[top]
\includegraphics[width=0.48\textwidth]{Fig3.eps}
\caption{\label{Fig3n} (left)
The control operations needed to stabilize states at different detunings $\epsilon$ for $T_C\in \{0,\infty \}$ are shown on the control plane. The control strength $\theta_C$ is determined by the length of a vector between a point on the control plane and the origin. The angle between the positive part of the $\theta_C n_z$-axis and the vector is the control angle $\theta$.
(right) The corresponding stabilizable states
are shown as projection on the $\ew{\sigma_z}$-$\ew{\sigma_x}$- plane of the Bloch sphere. }
\end{figure}
\subsection{Finite detuning $\epsilon\neq 0$}
With finite detuning it is possible to stabilize all states on the Bloch sphere which have a positive $\ew{\sigma_y}$ component. The exact expressions for the stabilizable eigenstates with $\epsilon\neq0$ can be found in appendix \ref{App2}.
Fig.~\ref{Fig3n} (right) shows the projection of the stabilizable eigenstates on the $\ew{\sigma_z}$-$\ew{\sigma_x}$- plane of the Bloch sphere for different detunings for $T_C\in\{0,\infty\}$. Without detuning ($\epsilon=0$) the stabilizable states lie on the axis of the plane which can be also seen in Fig.~\ref{Fig2}(a).
Since all stabilized states have a Bloch vector with length one and a positive $\ew{\sigma_y}$- component, the size of $\ew{\sigma_y}$ is maximal when the states in the $\ew{\sigma_z}$-$\ew{\sigma_x}$- plane have the shortest distance to the origin.
For increasing detuning these states move asymptotically towards the unit circle, meaning the maximal size of the $\ew{\sigma_y}$ component decreases for increasing $|\epsilon|$. Each of the four quadrants of the plane is covered by a different solution: The I- quadrant by the $\rho_{++}$ solution for $\epsilon>0$, the II- quadrant by the $\rho_{--}$ solution for $\epsilon<0$, the III- quadrant by the $\rho_{--}$ solution for $\epsilon>0$ and the IV- quadrant by the $\rho_{++}$ solution for $\epsilon<0$.
The control operation necessary to stabilize the states for each detuning are shown on the left side of Fig.~\ref{Fig3n}.
\subsection{Phonons}
In any realistic experimental setup, our ideal coherent feedback control scheme will be disturbed by the coupling to a dissipative environment. In order to make quantitative predictions, we therefore introduce an additional coupling of the DQD to a bath of thermal phonons \cite{DQD3} with wave vector $Q$ and energy $\omega_Q$. We describe the electron-phonon coupling
via a (pseudo) spin-boson Hamiltonian \cite{Bra05}
\begin{align}
H_{\text{phon}} =& \sum_{Q}\Big[ \frac{1}{2} \sigma_z g_Q(a_{-Q}+a^\dagger_Q)+\omega_Q a^\dagger_Q a_Q \Big]
\end{align}
that leads to a modification of the free Liouvillian ${\cal W}_0$ in our master equation \eq{master0}. These modifications are included in the free Liouvillian shown in appendix~\ref{App1}. We assume that an electron in the DQD couples to a phonon bath with an ohmic spectral density with exponential cut-off at Debye frequency $\omega_c$.
Without detuning ($\epsilon=0$) the corresponding dephasing rates at temperature $T$ (we set the Boltzmann constant $k_B=1$) are
\begin{eqnarray}\label{gammap}
\gamma_p = \gamma \coth \frac{T_C}{T},\quad \gamma \equiv g \pi T_C e^{\frac{-2T_C}{\omega_c}}, \quad \gamma_b=0,
\end{eqnarray}
where $g\ll 1$ is a dimensionless coupling parameter.
Fig.~\ref{Fig2}(b) also shows the evolution of the fidelity with the inclusion of the phonon bath (dashed and dotted line) at finite temperature. The lengths of the stabilized steady-state Bloch-vectors with phonons is below one $|\ew{\bld{\sigma}}|<1$ since the phonon dephasing limits the stabilization of the pure states.
\begin{figure}[top]
\includegraphics[width=0.48\textwidth]{Fig4.eps
\caption{\label{Fig4} Loss of stabilization of the control stabilized pure states $\rho_{\mp\mp}$ due to phonons illustrated by the length of the Bloch vector as a function of $T_C$ for different detunings (a) $\epsilon=0$ (without detuning), (b) $\epsilon=\pm\frac{1}{20}\Gamma_R$ and (c) $\epsilon=\pm10\Gamma_R$. Without control ($\theta_C=0$) the length of the Bloch vector decays to zero for large tunnel couplings $T_C$.
Parameters: $\Gamma_L\rightarrow \infty$, phonon coupling $g=4\cdotp 10^{-4}$, $\omega_c/\Gamma_R=500$, $k_B T=50\Gamma_R$. }
\end{figure}
Fig.~\ref{Fig4}(a) shows the fidelity of the steady state for $\epsilon=0$
as a function of tunnel coupling $T_C$. The stationary state without control ($\theta_C\mathord=0$) is always mixed (except for trivially at $T_C=0$ where $|\ew{\mbox{\boldmath$\sigma$}}|\mathord = 1$ as then an electron tunnels into the left dot and never leaves it), with $|\ew{\mbox{\boldmath$\sigma$}}|\mathord\to 0$ at very large $T_C$.
In contrast the control-stabilized states $\rho_{\pm\pm}$ remain pure in absence of phonons ($|\ew{\mbox{\boldmath$\sigma$}}|\mathord=1$ for $g=0$).
These pure states are differently affected by the phonon dephasing.
This striking behaviour of the two phonon branches for the control-stabilized states $\rho_{\mp\mp}$ in Fig.~\ref{Fig4}(a) can be traced back to the character of the eigenvalues $\varepsilon_{\mp}$, \eq{eigendot},
and the corresponding eigenstates $\ket{\psi_\mp}$ of $\widetilde{H}$. Starting from large $T_C\gg \Gamma_R/4$, the $\varepsilon_{\mp}$ have a much larger real than imaginary part and $\ket{\psi_-}$ ($\ket{\psi_+}$) essentially corresponds to the ground (excited) state of the isolated DQD. In this regime of large level splitting, spontaneous emission of phonons from $\ket{\psi_+}$ to $\ket{\psi_-}$ dominates phonon absorption,
with the purity of $\rho_{++}$ strongly reduced as compared with the one of $\rho_{--}$, which is shown by the explicit expressions for the fidelities $|\ew{\mbox{\boldmath$\sigma$}_{\pm}}|$ of the two states,
\begin{align}
|\ew{\mbox{\boldmath$\sigma$}_{\pm}}|\approx1-\rb{\frac{2\gamma_p}{\Gamma_R}-\frac{4\gamma_p^2}{\Gamma_R^2}}
\left(1\pm \tanh\!\frac{T_C}{T}\right). \label{appTcbig}
\end{align}
In particular, the shape of the $\rho_{++}$ branch for large $T_C$ and at low temperatures is entirely determined by the spectral density of the phonons $\propto T_C e^{-2T_C/\omega_c}$ and can therefore be used
as a spectroscopic tool to determine the dephasing rate $\gamma_p$, \eq{gammap}.
Moving to smaller tunnel coupling $T_C$, the two broadened levels of the DQD start to overlap at $T_C \mathord= \Gamma_R/4$ at which point the $\varepsilon_{\mp}$, \eq{eigendot}, become purely imaginary, and the two merged branches split again at still smaller $T_C$. In absence of phonons, our control scheme coherently rotates the DQD into the pure states $\ket{\psi_\pm}$, i.e.,
$|\ew{\mbox{\boldmath$\sigma$}_{\pm}}|\mathord=1$ for $g\mathord=0$. The presence of phonons ($g\mathord> 0$), however, leads to an un-controlled decay of $\ket{\psi_\pm}$ that can be
described by $\Gamma_R\to \Gamma_R\mathord+2\gamma_p$ in the imaginary eigenvalues \eq{eigendot}, which for $T_C \mathord\ll \Gamma_R$ then become $\varepsilon_{+} \mathord\approx 0 $ (very slow decay) and $\varepsilon_{-}\mathord \approx -i(\Gamma_R/2+\gamma_p) $ (fast decay). This `repulsion' of imaginary levels is analogous to the Dicke spectral line effect \cite{Bra05,Dic53}, i.e. a splitting into a sub- and a superradiant decay channel, and is again confirmed by the fidelities
\begin{eqnarray}
|\ew{\mbox{\boldmath$\sigma$}_{+}}|\approx 1,\quad
|\ew{\mbox{\boldmath$\sigma$}_{-}}|\approx 1-\frac{4\gamma_p}{\Gamma_R}-\frac{8\gamma_p^2}{\Gamma_R^2},
\end{eqnarray}
valid
for $T_C \mathord\ll \Gamma_R$.
The fidelity with finite detuning as a function of $T_C$ is shown in Fig.~\ref{Fig4}(b) and (c). The main features of the plots are the same as in the case without detuning Fig.~\ref{Fig4}(a). However if the detuning between the dots is positive $\epsilon>0$ the $\rho_{++}$- and the $\rho_{--}$- branch cross each other, while for negative detunings $\epsilon<0$ they preform an anti-crossing. In general the stabilizable states (see Fig.~\ref{Fig3n}(right)) for small couplings $T_C\rightarrow0$ lie close to $\ew{\sigma_z}=\pm 1$ and for large $T_C\rightarrow\infty$ close to $\ew{\sigma_x}=\pm1$. The phonon dephasing has the strongest effect on the states which either lie close to $\ew{\sigma_z}=-1$ or on the states with $\ew{\sigma_x}\lesssim1$.
The control operation in the first case rotates the electron that originally tunnels in the left dot into a state which is almost entirely localized in the right dot and in the second case the electron is rotated into a state which corresponds to the excited state of the isolated system (see discussion for $\epsilon=0$). In both cases the states are strongly effected by dephasing. In the first case this effect vanishes for decreasing temperatures but in the second case spontaneous emission even at zero temperature destabilizes the states, with the exception of states above the Debye cut-off (these states are only weakly effected by the phonon dephasing). This knowledge explains the anti-crossing and crossing in Fig.~\ref{Fig4}(b) and (c). For $\epsilon<0$ the solution stronger destabilized by phonons is the same ($\rho_{++}$) for small and for large $T_C$ which leads to an anti crossing and for $\epsilon>0$ the solution stronger affected by dephasing switches form $\rho_{--}$ for small $T_C$ to $\rho_{++}$ for large $T_C$ resulting in a crossing.
\begin{figure}[top]
\includegraphics[width=0.5\textwidth]{Fig5.eps}
\caption{Fano factors $F^{(n)}$ for the double quantum dot with zero phonon coupling
as a function of the control strength $\theta_C$ at fixed qubit rotation angle $\theta=\frac{\pi}{2}$.
Stabilization of a pure state
occurs at the value of $\theta_C$ where $F^{(n)}=1; \forall n>0$, corresponding to Poissonian statistics in the limit of $\Gamma_L\gg T_C,\Gamma_R$ considered here. ($\epsilon=0$)
\label{Fig5}}
\end{figure}
\subsection{Full counting statistics}
Collecting all trajectories with $n$ left-jumps as partial density matrix $ \rho^{(n)}$, \eq{master0} implies the number-resolved master equation for the system with control
\begin{eqnarray} \label{nres}
\dot{\rho}^{(n)} =
( {\cal W}_0 + {\cal J}_{R})\rho^{(n)}
+ {\cal J}^C_{L}\rho^{(n-1)}.
\end{eqnarray}
With the Fourier transform $\rho(\chi) = \sum_{n} e^{i n \chi} \rho^{(n)}$ we obtain the $\chi$-resolved master equation
\begin{eqnarray}\label {master1}
\dot\rho(\chi)
= \Big( {\cal W}_0 +
{\cal J}_{R}
+
{\cal J}^C_{L} e^{i\chi }
\Big) \rho(\chi).
\end{eqnarray}
Using the current cumulant generating function\cite{FCS1, FCS2, FCS3}
\begin{align}
{\cal F}(\chi,t)= \ln\Big(\text{Tr}_D\big\{ e^{({\cal W}_0 +
{\cal J}_{R}
+
{\cal J}^C_{L} e^{i\chi })(t-t_0)}\rho(t_0)\big\} \Big),
\end{align}
where $\text{Tr}_D\{\cdots\}$ corresponds to the trace of density matrix, we calculate the $n$-th order zero frequency current correlation
\begin{align}
\ew{S^{(n)}}=\frac{\text{d}}{\text{d}t}\frac{\partial^n}{\partial (i\chi)^n}
{\cal F}(\chi,t) |_{\chi=0,t\rightarrow\infty}.
\end{align}
We define the zero-frequency Fano factors as
\begin{align}
F^{(n)}=\frac{\ew{S^{(n)}}}{\ew{S^{(1)}}},
\end{align}
the $n$-th order current correlation functions, normalized by the stationary current \cite{FCS3} ($\ew{S^{(1)}}=\ew{I}$).
Fig. \ref{Fig5} shows the zero-frequency Fano factors $F^{(n)}$ with $n=2,3,4$ for our DQD example as a function of control strength $\theta_C$ with fixed static dot parameters and control angle $\theta=\frac{\pi}{2}$ (again in the limit $\Gamma_L \to \infty$). Whilst the cumulants exhibit a rather complex dependence on $\theta_C$ at exactly the point where the pure qubit state is stabilized, all Fano factors are simply $F^{(n)}=1$.
This can be understood by recalling that at the stabilization point, the system is equivalent to a single resonant level in its transport properties, cf. \eq{Leff}. The effective outgoing rate for our DQD example for $\epsilon=0$ reads
\begin{eqnarray}
\gamma_R^{(\mp)}&=&
\left\{
\begin{array}{cc}
\frac{1}{2}\Gamma_R & \Gamma_R\mathord<4T_C\\
\frac{1}{2}(\Gamma_R\pm\sqrt{\Gamma_R^2-16T_C^2} ) & \Gamma_R\mathord>4T_C
\end{array}
\right. .
\end{eqnarray}
The cumulant generating function of this effective model in the long time limit is well-known \cite{FCS2,FCS3}
\begin{align}
\mathcal F(\chi,t)=& \frac{t}{2}\Big(-\mathord\Gamma_L\mathord-\gamma_R^{(j)}\mathord+\sqrt{(\Gamma_L-\gamma_R^{(j)})^2 +4\Gamma_L \gamma_R^{(j)}e^{i\chi}}\Big). \label{SET-FCS}
\end{align}
Which in the limit $\Gamma_L \to \infty$ reduces to the cumulant generating function of a Poissonian process for which all Fano factors are equal to unity.
This information that the stabilized state has such a distinctive FCS can be used to locate the control operation required for stabilization {\em even in the absence of knowledge of the precise model parameters}. Knowledge of the electron jumps permits the FCS of the transport to be calculated. For a given set of, presumably non-stabilizing, control parameters, this FCS can be compared with an ideal Poissonian signal as a target. The control parameters can then be adjusted and the FCS compared again. This can be implemented in an {\em in situ} (classical) closed feedback loop to optimize the control operation such as to permit stabilization.
With electron jumps occurring, e.g., on time scales of milli-seconds in state-of-the-art FCS experiments \cite{Flietal09}, no ultra-fast electronics is necessary for information processing.
The optimization should also work if the condition of large $\Gamma_L$ can not be assumed, in which case the target signal would no longer be Poissonian but given by the cumulants
of the two coupled Poisson processes (left and right tunnel barrier) of the resonant level model \eq{SET-FCS}, Ref.~\cite{FCS2,FCS3}.
\section{Conclusions}
We have presented a general theory for the stabilization of pure states in a non-equilibrium electron transport setup. The stabilized states have the features of a single resonant level including the FCS, which can be used to detect these pure states.
We demonstrated this stabilization on the example of a DQD and have shown that half of the states on the Bloch sphere can be stabilized.
Moreover the interaction with an dissipative environment due to the coupling to a phonon bath was discussed.
Further generalization of the control stabilization to systems with many body states and more leads enable to use this stabilization scheme for arbitrary Coulomb blockade systems.
We are grateful to G.~Kie{\ss}lich, G.~Schaller, A.~Beyreuther and
K.~Mosshammer for useful discussions. Financial support by DFG projects BR 1528/7, BR 1528/8, GRK 1558 and SFB 910 is acknowledged.
|
\section{}
There is a "depletion zone" in flowing solutions of macromolecules which is characterized by a drastic reduction in the concentration of the polymer near a surface.\protect\cite{Fang:2007p7340,Ma:2005p5986, Jendrejack:2004p2674,CriadoSancho:2000p2666} Over the past decade it has become clear, after early work by Jendrejack, et al., that this behavior is due to a hydrodynamic lift force, which is a direct consequence of polymer stretching under the influence of an external fluid flow.\protect\cite{Ma:2005p5986,Hoda:2007p16,Jendrejack:2004p2674,Jendrejack:2002p824,Hoda:2007p7378,HernandezOrtiz:2006p7379, JendrejackdePablo} The origin of this lift-force arises from a no-slip boundary condition at a solid surface that breaks the fluid flow symmetry around a chain under tension, and the chain feels a resulting upwards force. This force has been widely investigated and applied to a variety of contexts, in particular the description of the aforementioned depletion zone and the case of polymer adsorption.\protect\cite{Ma:2005p5986,Hoda:2007p16,Jendrejack:2004p2674,Jendrejack:2002p824,HernandezOrtiz:2006p7379,Sendner:2007p1461,Hoda:2007p15,Hoda:2007p7378,Sendner:2008p6216,Serr:2009p6217, NETZ, Watari} Ma and Graham were the first to consider this behavior in the context of adsorption, and showed how a dumbell polymer model could be developed that displayed the qualitative characteristics seen in experiment.\protect\cite{Ma:2005p5986} This model was further developed by Hoda and Kumar who extended this model to incorporate the effects of electrostatic interactions.\protect\cite{Hoda:2007p7378} Theoretical and simulation work by Sendner and Netz has further investigated this phenomenon in the context of more refined bead-spring models that more explicitly incorporate the correct hydrodynamics.\protect\cite{Serr:2009p6217,Sendner:2008p6216,Sendner:2007p1461, NETZ} Despite these advances in the theory, comparison of many of these results with experiment reveal qualitative, but not quantitative agreement. Larson, et al. has indicated that the depletion zone near a surface is considerably smaller than predicted by theory, by as much as an order of magnitude.\protect\cite{Fang:2007p7340}
In this Letter we provide a general theory that describes the hydrodynamic lift behavior of a polymer at an arbitrary distance from the surface. Our theory provides an explanation of the discrepancies seen between experiment, simulation, and theory for surface depletion, and also an understanding of the lift forces that are relevant for polymer adsorption and desorption. In particular we note that previous models addressing this force, such as the dumbbell model, uses an explicit far-field assumption, the validity of which has been questioned by Hernandez-Ortiz, et al. and Hoda, et al.. However, a complete quantitative description of this regime has not yet been developed.\protect\cite{HernandezOrtiz:2006p7379,Hoda:2007p7378} Here, we provide analytical results that capture the fundamental physical picture in both the far-field and near-surface regimes. Our analytical theory supposes no underlying chain model, and relies only on geometric parameters and flow conditions. It shows for the first time that the lift force on an extended object, like an extended polymer, can become non-monotonic, increasing linearly with the distance from the surface. In the far-field regime we recover the well known results that the force decreases quadratically with the distance. Of further interest is the fact that the crossover from the near-surface to the far-field regime where the lift force is maximal occurs at a height $Z^* \sim L$, where $L$ corresponds to the polymer extension length. Since $L$ can be extremely large, particularly in strong flows, the near field regime may extend far away from the surface.
To develop this model, we only need to consider the geometry of a polymer chain near a surface. Here we provide a description similar to the blob model of a polymer chain under tension by representing the chain as a series of hydrodynamic beads whose length is equal to the ratio of the average length to the average width of the elongated polymer chain $2N = \langle \Delta X \rangle / \langle \Delta Z \rangle$ (see Figure~\ref{Fig1}a for a schematic).\protect\cite{DeGennes} This description of the polymer chain is amenable to the incorporation of molecular theory; for example a straightforward calculation of this geometry in shear flow based on the work of DeGennes is performed in the Supplemental Material (SM), and yields the relationship $2N = \sqrt{1+ (\dot{\gamma} \tau_Z)^2/(2E^2)}$, where $\dot{\gamma} \tau_Z = Wi$ is the shear rate rendered dimensionless through the use of the chain relaxation time $\tau_Z$ and $E$ is Peterlin's chain stretching parameter that is $1$ at low extension and diverges at high extensions.\protect\cite{DEGENNES:1974p2152} We introduce a single length scale $2a = \langle \Delta Z \rangle$ to describe the chain dimension normal to the wall that effectively corresponds to the bead radius. This consequently allows us to represent our system as a bead-spring chain that lies parallel to the surface, in the same spirit as Sendner and Netz (see Figure~\ref{Fig1}b for a schematic).\protect\cite{Serr:2009p6217,Sendner:2008p6216} We can then determine an analytical form for the hydrodynamic lift force of this fully elongated polymer using the Blake-Oseen tensor, which has been shown to provide a good description of hydrodynamics near surfaces in the regime of interest ($Z > a$).\protect\cite{BLAKE:1974p2245, Karrila} This geometry conveniently enables us to use a simplified version of the hydrodynamic interaction tensor, and captures the essential physics of the lift force in the limit of high stretching. This approach is also the relevant case for a desorbing or adsorbing polymer, which exhibits extended conformations as it interacts with both the external fluid flow and the surface.\protect\cite{Serr:2009p6217,Ladoux:2000p5497} Understanding a polymer in such a limit is also important when considering the behavior of a polymer in channels that are characterized by small length scales.
The lift force on a polymer, or any other deformable component dispersed in a liquid medium near a surface, is due to the hydrodynamic interactions between individual components and the no-slip boundary condition at the surface. The simplest manifestation of the hydrodynamic interaction between an entity and its surroundings is given by the Oseen mobility tensor $\mu_{ij,O}(\mathbf{r}_i, \mathbf{r}_j)$, which describes the entity as a single point force.\protect\cite{DeGennes,BLAKE:1974p2245} This mobility tensor describes the effect of a point force $\mathbf{F}_j (\mathbf{r}_j)$ on the surrounding velocity field $\mathbf{v}_i (\mathbf{r}_i)$ through the equation $\mathbf{v}_i (\mathbf{r}_i) = \mu_{ij,O}(\mathbf{r}_i, \mathbf{r}_j) \cdot \mathbf{F}_j (\mathbf{r}_j)$. To account for the effect of a nearby surface, where a no-slip boundary condition must be maintained, Blake introduced an image system that accounted for the aforementioned boundary condition to produce a new mobility tensor $\mu_{ij,B}$.\protect\cite{BLAKE:1974p2245} See the SM for more discussion of the mobility tensors.
\begin{figure}
\includegraphics[width=150mm]{Fig1NEWNEW.png}
\caption{(a) Simulation data showing the accumulation of bead positions over 4800 $\tau$ at $\dot{\gamma} \tau= 2.1$ for a $\theta$-chain of $100$ beads. We approximate this geometry using a bead-rod model, which groups portion of the chain into beads (orange circles). This geometry is shown in (b), and has $2N = \langle \Delta X \rangle / \langle \Delta Z \rangle$ beads of radius $a = \langle \Delta Z \rangle/2$ parallel to the surface at height $Z$. Individual dumbbell-pairs feel a net force $F_f$ which is equivalent to the gradient of the tensile force along the chain in the continuum limit. All pairs of dumbbells contribute to an overall lift force $F_L$. \label{Fig1} \vspace{-15pt}}
\end{figure}
In our model we consider a finite chain of $2N$ hydrodynamic beads at a distance $Z$ from the surface, each of radius $a$. Using these parameters, we can render all variables dimensionless (rescaled variables denoted by tildes) in terms of distances $a$, energies $kT$, and times $\tau_0 = 6 \pi \eta a^3/(kT)$, where $\eta$ is the solvent viscosity. This geometry is shown schematically in Figure~\ref{Fig1}b. The tension force $F_T$ along the chain is considered to only be in the $x$-direction, such that the chain is in equilibrium in this configuration. Since we are primarily concerned with the lift forces in the direction normal to the surface, we only need to consider the resulting velocity field in the $z$-direction. The relevant component of the mobility tensor is then just $\mu^{xz}_{ij,B}(\mathbf{r_x}) = (-3 \mathbf{r}_x Z^3)/(2 \pi \eta (\mathbf{r}^2_x + 4Z^2)^{5/2})$. To evaluate the overall force, one can just define the effect of a given dumbell centered at the mid point of the chain that is separated by a distance $4na$ on the velocity field at a dumbell that is separated by a distance $4ma$, where $n$ and $m$ are just indexing parameters (see Fig. 1). The resulting lift force $F_{d}$ from the $n$ dumbbell on the $m$ dumbbell can be writen as:
\begin{eqnarray}
F_{d}(n,m,F_f,\tilde{Z})
= -12 \pi \eta a F_f \left[ \mu_{n,m,B}^{zx} + \mu_{n,-m,B}^{zx} \right] \nonumber \\
= \frac{9F_f \tilde{Z}^3}{8} \left[ \frac{(n-m)}{[(n-m)^2 + \tilde{Z}^2]^{5/2}} + \frac{(n+m)}{[(n+m)^2 + \tilde{Z}^2]^{5/2}} \right]
\end{eqnarray}
where $F_f$ is the magnitude of the force on the $n$th bead and the rescaled height $ \tilde{Z}=Z/a$. To determine the overall lift force due to a given dumbell on the entire chain $F_{L,d} (n,F_f,\tilde{Z}$ (including the dumbell itself), we sum over all dumbells $m = 1$ to $N$:
\begin{eqnarray}
F_{L,d} (n,F_f,\tilde{Z})= \sum_{m=1}^N {F_{d} \approx \int_0^N F_{d} (n,m,F_f,\tilde{Z}) dm} = \nonumber \\
= \frac{3 F_f \tilde{Z}^3}{16} \left[ \left[ (N-n)^2 + \tilde{Z}^2 \right]^{-3/2} - \left[ (N+n)^2 + \tilde{Z}^2 \right]^{-3/2} \right]
\label{Sum1}
\end{eqnarray}
where we have used the approximation that $N$ is large enough that we can replace the sum by a continuous integral from $0$ to $N$. To find the total lift force, we perform the summation over all dipoles $n$ and again replace it with an integration:
\begin{equation}
F_L = \sum_{n=1}^N F_{L,d} \approx \int_0^N{ F_{L,d}} dn
= \frac{3\tilde{Z}^3}{16} \int_0^N{\left( \frac{\partial F_T}{\partial n} \right) \left[ \left[ (N-n)^2 + \tilde{Z}^2 \right]^{-3/2} - \left[ (N+n)^2 + \tilde{Z}^2 \right]^{-3/2} \right] dn}
\label{Sum2}
\end{equation}
Since $F_f$, the overall force on the dumbbell, is the difference of the tension on either side of the bead, we make the replacement of this force with its continuum analogue, $(\partial F_T/ \partial n)$. This represents a "kernel" in which the form of the load profile is input into the theory. In this paper we consider a profile that corresponds to shear flows with the extended polymer at a small angle $\theta$ (another example is given in the SM). We incorporate the result $(\partial F_T / \partial n) \approx 6 \pi \eta a^2 \dot{\gamma} \sin{ (2 \theta)} n$. Performing the integration, we get the result:
\begin{equation}
\tilde{F}_{L,shear} = \frac{3}{16} \dot{\gamma} \tau \sin{ (2 \theta)} \left[ \frac{ 2N^2 \tilde{Z} + \tilde{Z}^3}{\sqrt{4N^2+\tilde{Z}^2}} - \tilde{Z}^2 \right]
\label{SHEAR}
\end{equation}
For the far-field, we retain the proper $\tilde{Z}^{-2}$ scaling: $\tilde{F}_{L,shear} \approx \frac{3}{8} \dot{\gamma} \tau \sin{(2 \theta)} \frac{N^4}{\tilde{Z}^2}$, however below a crossover height $\tilde{Z}^* \sim N$ there is a near-surface regime where the lift force becomes non-monotonic and the full equation ~\ref{SHEAR} must be used. This theory can now be related to any microscopic theory of choice, with the geometrical parameters ($N$, $a$, $\tau$) corresponding to chain parameters. We show a straightforward example of this in the SM using Gaussian dumbbells to obtain the well-known far-field scaling relationship for the depletion length, $Z_{dep} \sim Wi^{2/3} n^{1/2}$.\protect\cite{Ma:2005p5986, Jendrejack:2004p2674,Hoda:2007p7378}
\begin{figure}
\includegraphics[width=150mm]{Fig2NEWNEW.png}
\caption{(a) Load profile that corresponds to a chain in shear flow, which is placed into the equation~\ref{Sum2} to yield equation~\ref{SHEAR}. It is described by $\tilde{\mu} (\partial \mathbf{\tilde{F}_{f} (\tilde{r}_i)}/\partial n) = \dot{\gamma} \tau \sin{2\theta} (\mathbf{\tilde{r}_{i,x}} - \sum_j^{2N}{\mathbf{\tilde{r}_{j,x}}}/(2N))$. (b)Graph of $\tilde{F}_L$ versus $\tilde{Z}$ for a shear load profile. Solid lines show theoretical results, dashed lines show the effect of bead-discretization at low values of $\tilde{Z}$ (see SM), and dotted lines demonstrate the far-field dumbbell results. Simulation data is also shown, with filled symbols representing data without fluctuations and open symbols representing data including fluctuations. These results collapse onto a single curve (inset) using the scaling $\tilde{Z} \rightarrow \tilde{Z}/N$ and $\tilde{F} \rightarrow \tilde{F}/N^2$. \label{Fig2} \vspace{-15pt}}
\end{figure}
Or results are confirmed by Brownian Dynamics simulations (details in the SM) on chains that are extended parallel to a surface. These bead-spring models represent the effective blobs that make up an extended polymer chain. For these simulations, we incorporate the appropriate hydrodynamics by using the Rotne-Prager-Blake tensor and consider the chain with both the absence and presence of thermal fluctuations.\protect\cite{ROTNE:1969p5862, Karrila} For these simulations we use the force loading profile that corresponds to shear flow, which is given by $\tilde{\mu} (\partial \mathbf{\tilde{F}_{f} (\tilde{r}_i)}/\partial n) = \dot{\gamma} \tau \sin{2\theta} (\mathbf{\tilde{r}_{i,x}} - \sum_j^{2N}{\mathbf{\tilde{r}_{j,x}}}/(2N))$ (shown in Figure~\ref{Fig2}a).
\begin{figure}
\includegraphics[width=75mm]{LiftFig3R.png}
\caption{(a) Concentration profiles highlighting the difference between the elongated chain model (solid lines), the dumbbell model (dashed lines), and the far-field dumbbell model (dotted lines) for a number of different shear rates near a surface with $N=10$ (given in terms of geometric parameters). The depletion region is often considerably smaller when near-surface hydrodynamics and elongated chain geometries are considered. While direct comparison of these results to Weissenberg numbers is dependent on the specific chain model, a simple scaling model (see SM) using these values suggests they correspond to $Wi \sim 1-10$. All three cases correspond to the same far-field behavior. (b) Probability distribution function for chains in a sheared slit flow. The near-surface hydrodynamics, which are considered in the elongated chain model, have the effect of greatly decreasing surface depletion in channel widths on the order of the chain contour length. $\dot{\gamma} \tau_{eff} = 0.01$ and $N=10$, roughly corresponding to $Wi \sim 5-10$. All three cases correspond to the same far-field behavior. Finite concentrations at $\tilde{Z} = 0$ in (a) and (b) occur due to a lack of divergence in equation~\ref{SHEAR} and the neglect of a short-range excluded-volume potential at the surface. (c) Diagram demonstrating roughly where the far-field approximation is valid and where the near-surface effects need to be considered (extended-chain theory). \label{Fig3} \vspace{-15pt}}
\end{figure}
In Figure~\ref{Fig2}b we plot the result of equation~\ref{SHEAR} for a shear flow for chains of $2N = 20$, $50$, and $80$ with an effective shear rate of $\dot{\gamma} \tau \sin{2 \theta} = \dot{\gamma} \tau_{eff}= 0.1$. We plot two types of simulations; open symbols indicate simulations where fluctuations are turned off, and filled symbols indicate simulations with thermal fluctuations included. It is clear that, especially in the limit of large $N$, the theory and simulation match well. This plot demonstrates a non-monotonous lift force with two characteristic regimes for this loading profile; the near-surface regime demonstrates an increasing $\tilde{F}_L$ with $\tilde{Z}$, and the far-field regime demonstrates a $\tilde{F}_L \sim \tilde{Z}^{-2}$ decay that is expected from the far-field dumbbell model (plotted as dotted lines in Figure~\ref{Fig2}a). We can rescale this graph for all values of $N$ by rescaling $\tilde{Z} \rightarrow \tilde{Z}/N$ and $\tilde{F}_L \rightarrow \tilde{F}_L/N^2$ to collapse all of the curves for a given value of $\dot{\gamma} \tau_{eff}$ onto a single curve. This is shown in the inset of Figure~\ref{Fig2}. There are deviations between theory and simulation at small values of $\tilde{Z}$ due to the discretized representation of the chain in the simulations as opposed to continuous representation of the chain in the theory. A discretized version of our theory can be developed that only requires a few terms at low $\tilde{Z}$ (see SM) which more accurately reflects the low $\tilde{Z}$ behavior in our simulations, and is shown as a dashed line in Figure~\ref{Fig2}b. We expect that a real chain would more closely resemble the continuous case, as the discretization of the chain is arbitrarily done to develop a convenient simulation model.
The size and shape of the depletion layer near a surface can now be calculated using equation~\ref{SHEAR}. This is done by describing a potential of mean force $\tilde{U}_L = - \int_{\infty}^{\tilde{Z}}{\tilde{F}_L d \tilde{Z}}$. If the only force acting on the polymer in the $\tilde{Z}$-direction is the lift force, we can write the concentration $c(\tilde{Z})$ of a dilute polymer solution at height $\tilde{Z}$ from the surface as $c(\tilde{Z}) = c(\infty)e^{-\tilde{U}_L}$. We plot sample profiles for our model, for the dumbbell model (where we use both the far-field result and the dumbbell result given by Hoda and Kumar) for shear flow in Figure~\ref{Fig3}a.~\protect\cite{Hoda:2007p7378} There is a striking difference between these two profiles, with drastic differences appearing due to the significantly lower lift forces close to the surface using our elongated chain model rather than the dumbbell model. We also predict a finite concentration at the wall, at least until steric or other short-range wall forces become significant, since the effective lift force potential does not diverge at $\tilde{Z} = 0$. These characteristics have been noticed before in experiments, with depletion layer widths that are smaller than the predicted ones by as much as an order of magnitude, and which are often accompanied by finite surface concentrations of polymer.\protect\cite{Fang:2007p7340} Our model clearly provides an explanation of the origin of the difference between experiment and previous theories, which do not adequately account for near-surface hydrodynamics.
We can also consider flows in a slit, where two walls are present. If the slit height $H$ is on the same order of magnitude as the polymer contour length $Na$, we demonstrate extremely large differences between existing theories and our elongated chain model. Since we have introduced an expression for $\tilde{F}_L(\tilde{Z}, N, \tilde{F}_T)$ that is based off a single image chain, we simply need to consider an infinite series of image chains. For a slit of height $H$, this becomes $\tilde{F}_{L,slit} = \sum_{j = 0}^{\infty} \tilde{F}_L(jH/a+\tilde{Z},N, \tilde{F}_T) - \tilde{F}_L((j+1)H/a-\tilde{Z},N, \tilde{F}_T) $. We plot a sample distribution function in a slit for our chain elongation model and both the far-field and non-far-field dumbbell models in Figure~\ref{Fig3}b, where the far-field behavior is held constant. While the two dumbbell cases are essentially identical in this regime, there is a marked difference in the distribution function that results from considering the elongated chain geometry.
These depletion layer calculations emphasize where our elongated-chain theory provides a meaningful improvement over dumbbell-based models, which is at high-shear and near-surface conditions. Higher shear rates result in a longer effective chain, which increases the height $\tilde{Z}^*$ where there is a crossover from the near-surface to the far-field regimes. This is demonstrated schematically in Figure~\ref{Fig3}c, which indicates the extent of the near-surface regime as a function of flow rate (shown here in terms of the Weissenberg number $Wi = \dot{\gamma} \tau_Z$, for ease of comparison with traditional literature).
In summary, we have developed an analytical expression for the lift force on a polymer near a surface that significantly improves upon existing theories by accounting for extended geometries. Computer simulations incorporating the hydrodynamic forces were performed and confirmed these analytical expressions, which give rise to the appearance of non-monotonic lift force behavior. This non-monotonic lift force behavior appears in a near-surface, highly extended regime, and it is not captured by a dumbbell theory. Furthermore, our results have important implications for slit flows in microfluidics and polymer adsorption, where the distance of the polymer from the surface and its contour length are on the same order of magnitude.\protect\cite{NETZ} Figure~\ref{Fig3}b suggests that this result could be verified experimentally using the visualization of the cross-channel distribution fluorescently-labelled polymers by applying strong shear flows in narrow channels where this effect is particularly large. Cross slit flows may also provide a tool to directly examine the lift-force on individual polymers.
We acknowledge funding from the National Defense Science and Engineering Graduate Fellowship
\section{Supplemental Information}
\subsection{Hydrodynamic Mobility Tensors}
To describe the effect of a force applied to a bead at $\mathbf{r}_j$ on the surrounding fluid flow at $\mathbf{r}_i$, we define a mobility tensor $\mu_{ij} (\mathbf{r}_i, \mathbf{r}_j)$ that satisfies the equation $\mathbf{v}_i (\mathbf{r}_i ) = \mu_{ij} (\mathbf{r}_i, \mathbf{r}_j) \cdot \mathbf{F}_j (\mathbf{r}_j )$. There are multiple forms for this tensor, depending on the assumptions made and the system of interest. The simplest version, called the Oseen tensor, describes the effect of a single point force on the surrounding fluid flow in an infinite medium:\protect\cite{BLAKE:1974p2245}
\begin{equation}
\mu_{ij,O} (\mathbf{r}_i - \mathbf{r}_j = \mathbf{r}) = \frac{1}{8 \pi \eta r} \left[ \mathbf{I} + \mathbf{\hat{r} \hat{r}}\right]
\end{equation}
where we define $r$ as the magnitude of $\mathbf{r}$ and $\mathbf{\hat{r}} = \mathbf{r}/r$. To consider beads of finite size, we can instead use the multipole expansion of the Oseen tensor, known as the Rotne-Prager-Yamakawa tensor:\protect\cite{ROTNE:1969p5862, Karrila}
\begin{equation}
\frac{\mu_{ij,RPY} (\mathbf{r})}{\mu_0 } =
\frac{3a}{4r_{ij}} \left[ \left( 1 + \frac{2a^2}{3r_{ij}^2} \right) \mathbf{I} + \left( 1 - \frac{2a^2}{r_{ij}^2} \right) \mathbf{\hat{r} \hat{r}} \right]
\end{equation}
where $\mu_0 = 1/(6 \pi \eta a)$ is the Stoke's mobility of a bead of radius $a$.
Both the Oseen and Rotne-Prager-Yamakawa tensors assume an infinite medium. If there is a surface present, additional terms need to be added to these tensors such that they account for the no-slip boundary condition. The resulting tensor is called the Blake tensor, and is given (for the Oseen-based Blake tensor) by:\protect\cite{BLAKE:1974p2245}
\begin{eqnarray}
\mu_{ij,B} (\mathbf{r}) = \mu_{ij,O}(\mathbf{r}) - \mu_{ij,O} (\mathbf{r} + \mathbf{r}_{j,Z}) + \nonumber \\ + 2 \mathbf{r}_{j,Z}^2 \mu_{ij,PD} (\mathbf{r} + \mathbf{r}_{j,Z}) - 2 \mathbf{r}_{j,Z} \mu_{ij,SD}(\mathbf{r} + \mathbf{r}_{j,Z})
\end{eqnarray}
where $\mathbf{r}_{j,i,Z} = 0$ at the surface and:
\begin{equation}
\mu_{ij,PD} (\mathbf{r}) = (1-2\delta_{x_i,Z}) \left( \frac{\mathbf{I} - 3 \mathbf{\hat{r} \hat{r}}}{r_{ij}^3} \right)
\end{equation}
and
\begin{equation}
\mu_{ij,SD} (\mathbf{r}) = \mathbf{r}_{Z} \mu_{ij,PD} (\mathbf{r}) +(1-2\delta_{x_i,Z}) \left[ \frac{\delta_{x_j,Z} \mathbf{r}_{i} - \delta_{x_i,Z} \mathbf{r}_j}{ \mathbf{r}_{ij}^3}\right]
\end{equation}
\subsection{Simulation Methods}
To simulate a single chain in the geometry specified in the paper, we use Brownian Dynamics (BD) simulations. Our polymer is represented by a bead-spring model, which is composed of $2N$ beads $i$ at positions $\mathbf{r}_i$ that are held together by a harmonic potential:
\begin{equation}
\tilde{U}_S = \frac{\tilde{\kappa}}{2} \sum_{i=1}^{2N-1}{(\tilde{r}_{i,i+1}-2)^2}
\end{equation}
The beads interact with other monomers through a Lennard-Jones potential:
\begin{equation}
\tilde{U}_{LJ} = \tilde{\epsilon} \sum_{i,j}^{2N} \left[ \left( \frac{2}{\tilde{r}_{ij}} \right)^{12} - 2 \left( \frac{2}{\tilde{r}_{ij}} \right)^6 \right]
\end{equation}
There is also a harmonic potential that fixes the polymer at a given height $\tilde{Z}$ from the surface:
\begin{equation}
\tilde{U}_Z = \frac{(\sum_j^{2N}{\mathbf{\tilde{r}_{j,z}}}/(2N) - \tilde{Z})^2}{10}
\end{equation}
where all values designated with a tilde are dimensionless, with distances normalized by the bead radius $a$, energies normalized by $kT$, and times normalized by the characteristic diffusion time $\tau = 6 \pi \eta a^3/(kT)$. $\tilde{r}_{ij}$ is the distance between beads $i$ and $j$, $2N$ is the overall number of beads in the chain, $\tilde{\kappa} = 500$ is the bead-bead spring constant, and $\tilde{\epsilon}$ is a bead interaction parameter that controls the strength of bead-bead attraction. For this paper we use the value $\tilde{\epsilon} = 0.41$, which is typical for a $\theta$- polymer. Beads move through this potential via integration of the Langevin equation:
\begin{equation}
\frac{\partial}{\partial \tilde{t}}\mathbf{\tilde{r}_i} =\mathbf{\tilde{v}_\infty} (\mathbf{\tilde{r}_i)} - \sum_j \left( \tilde{\mu}_{ij,B} \nabla_{\mathbf{r_j}} \tilde{U}_{tot} (\tilde{t}) + \nabla_{\mathbf{r_j}} \cdot \mathbf{\tilde{D}}_{ij} \right) + \xi_i (\tilde{t})
\end{equation}
where $\mathbf{\tilde{v}_{\infty} (\tilde{r}_i)}$ is the undisturbed solvent flow profile, $\mu_{ij}$ is the Rotne-Prager-Blake mobility matrix, $\tilde{U}_{tot} = \tilde{U}_S + \tilde{U}_{LJ} + \tilde{U}_{Z}$, $\mathbf{D}_{ij} = k_B T \mu_{ij}$ is the diffusion tensor, and $\xi_i$ is a random velocity that satisfies $\langle \xi_i \rangle = 0$ and $\langle \xi_j (t) \xi_i (t') \rangle = 2k_B T \mu_{ij} \delta (t-t')$. Two cases are considered in this letter for $\mathbf{\tilde{v}_{\infty} (\tilde{r}_i)}$. We represent the force loading corresponding to a shear flow with $\mathbf{\tilde{v}_{\infty, x} (\tilde{r}_i)} = \dot{\gamma} \tau (\mathbf{r_{i,x}} - \sum_j^{2N}{\mathbf{r_{j,x}}}/(2N))$, which captures the linear flow increase away from the center of mass of the chain. The pulling case is given by $\mathbf{\tilde{v}_{\infty, x} (\tilde{r}_i)} = \tilde{F}_0 \delta(i-2N) - \tilde{F}_0 \delta(i)$, which essentially only places a force in the $x$-direction on the first and last beads in a chain. The Langevin equation is discretized by a time step of $\Delta \tilde{t} = 10^{-4}$, and for a given set of conditions $10^7$ simulation steps are used. For a given value of $\tilde{Z}$, the average lift force $\tilde{F}_L$ on a polymer is given by:
\begin{equation}
\tilde{F}_L = \frac{\sum_j^{2N}{\mathbf{\tilde{r}_{j,z}}}/(2N) - \tilde{Z}}{5}
\end{equation}
\subsection{Polymer Dimensions in Shear Flow}
The landmark work by DeGennes on chain behavior in strong flow fields allows us to calculate chain dimensions as a function of shear rate.\protect\cite{DeGennes} We desire a relation that yields an approximate "aspect ratio" of a polymer chain in shear flow that serves as the basis for the respective number of beads that would be used in an analogous bead-spring representation of the chain.
We begin by introducing a series of equations that DeGennes derives from the Peterlin Dumbell formalism:\protect\cite{DEGENNES:1974p2152}
\begin{eqnarray}
1 - EC_{xx} + \dot{\gamma} \tau_Z C_{zx} = 0 \label{DG1} \\
-EC_{xz} + \frac{1}{2} \dot{\gamma} \tau_Z C_{zz} = 0 \label{DG2} \\
EC_{zz} = EC_{yy} = 1
\end{eqnarray}
where $E$ is a chain extension parameter that is $1$ at low extensions and diverges at high extensions, $C_{ij} = 3 \langle r_i r_j \rangle / Na^2$ is a symmetric matrix describing the polymer dimensions, and $\dot{\gamma} \tau_Z$ is the dimensionless shear rate (also known as the Weissenberg number) based on the Zimm relaxation time $\tau_Z$. Equations~\ref{DG1} and~\ref{DG2} can be reorganized to give the result:
\begin{equation}
C_{xx} = \frac{1}{E} + \frac{(\dot{\gamma} \tau_Z)^2}{2E^3}
\end{equation}
The aspect ratio of the chain geometry, which will become $2N$, can be then obtained using the relationship:
\begin{equation}
2N = \left( \frac{C_{xx}}{C_{zz}} \right)^{1/2} = \left[ 1+\frac{(\dot{\gamma} \tau_Z)^2}{2E^2} \right]^{1/2}
\end{equation}
This equation yields intuitive behavior - at $\dot{\gamma} \tau_Z < 1$, the polymer is roughly spherical. As $\dot{\gamma} \tau_Z$ becomes greater than $1$, there becomes a regime of linear scaling of the aspect ratio with the shear rate. This regime flattens out when $E>1$ at high extension, and the value of $2N$ asymptotically approaches the contour length $4N$ of the polymer.
\subsection{Derivation of Lift Force for an Elongated Chain (Pulling Scenario)}
In this letter we have derived the behavior using the kernel $(\partial F_T/ \partial n) \approx 6 \pi \eta a^2 \dot{\gamma} \sin{(2 \theta)} n$, however others could be applicable. For example, if there is an abundance of mass at the chain ends, it may be more appropriate to use a pulling profile $(\partial F_T/\partial n) = F_0 \delta(n - N)$. This is essentially the dumbbell result, only we now integrate over the entire chain contour in the same fashion as before:
\begin{eqnarray}
F_L = \sum_{n=1}^N F_{L,d} \approx \int_0^N F_{L,d} dn
= \frac{3\tilde{Z}^3}{16} \int_0^N{\left( \frac{\partial F_T}{\partial n} \right) \left[ \frac{1}{\left[ (N-n)^2 + \tilde{Z}^2 \right]^{3/2}} - \frac{1}{\left[ (N+n)^2 + \tilde{Z}^2 \right]^{3/2}} \right] dn}
\label{Sum2}
\end{eqnarray}
Performing the integration, we get the result:
\begin{equation}
\tilde{F}_{L,pull} = \frac{3 \tilde{F}_0 \tilde{Z}^3}{16} \left[ \tilde{Z}^{-3} - \frac{1}{\left( 4N^2 + Z^2 \right)^{3/2}}\right]
\end{equation}
This simplifies to the result as $\tilde{Z}$ goes to zero:
\begin{eqnarray}
\tilde{F}_{L,pull} \approx \frac{3 \tilde{F}_0}{16}
\end{eqnarray}
and as $Z$ goes to infinity:
\begin{eqnarray}
\tilde{F}_{L,pull} \approx \frac{9 \tilde{F}_0 N^2}{8 Z^2}
\end{eqnarray}
We can still define a characteristic height $\tilde{Z}^*$ where there is a transition between these two behaviors. Solving for the maximum of the shear behavior, we obtain the result $\tilde{Z}^* \approx N$. Rearrangement of the result for the pulling case can be plotted as $\tilde{F}$ versus $\tilde{Z}/N$. Results for this are shown in Figure 1 (SM), which is analogous to Figure 2 in the letter.
\begin{figure}
\includegraphics[width=75mm]{LiftFigS1R.png}%
\caption{Graph of $\tilde{F}_L$ versus $\tilde{Z}$ for a shear load profile. Solid lines show theoretical results, dashed lines show the effect of bead-discretization at low values of $\tilde{Z}$, and dotted lines demonstrate the far-field dumbbell results. Simulation data is also shown, with filled symbols representing data without fluctuations and open symbols representing data including fluctuations. These results collapse onto a single curve (inset) using the scaling $\tilde{Z} \rightarrow \tilde{Z}/N$.}
\end{figure}
\subsection{Low-Z Regime}
The replacement of the summation in equation 2 with an integral is valid if the chain is continuous or if the chain is in the far-field limit. If the chain is discrete and close to the wall, like in our simulations, this approximation does not hold. If we retain the summation in the low-$\tilde{Z}$ limit, we can match our simulation results and demonstrate that this is the reason for the disparity between simulation and theory in Figures 2a and 2b. The discretized version of equation 2 yields the result:
\begin{equation}
F_{L, disc} = \frac{9 \tilde{Z}^3}{16 N} \sum_{n,m = 0}^N{\left[ \frac{(n-m)}{F_{T,n}\left[ (n-m)^2+\tilde{Z}^2\right]^{5/2}} + \frac{(n+m)}{\left[ (n+m)^2+\tilde{Z}^2\right]^{5/2}} \right]}
\end{equation}
We can simplify this for small $\tilde{Z}$ by considering that the effect of a given dipole on its surroundings is both antisymmetric and quickly approaches zero. Therefore, beads in the center of the chain will not exert a strong net lift force. Only the beads close to the chain ends contribute non-negligibly to the lift force, which we express as a truncated series:
\begin{equation}
F_{L, disc, shear} \approx \frac{9 \tilde{Z}^3 N \dot{\gamma} \tau}{4}\sum_{i = 0}^l{\left[ \frac{i^2}{\left[ (i)^2+\tilde{Z}^2\right]^{5/2}} \right]}
\end{equation}
for shear and:
\begin{equation}
F_{L, disc, shear} \approx \frac{9 \tilde{Z}^3 F_0}{8}\sum_{i = 0}^l{\left[ \frac{i}{\left[ (i)^2+\tilde{Z}^2\right]^{5/2}} \right]}
\end{equation}
for pulling. $l$ is a small number representing the point at which the summation is truncated (we use $l = 3$). This result agrees with the values obtained for simulation data at small $\tilde{Z}$, as shown by the dotted lines in Figures 2 and 1(SM). This indicates that the main reason for the disparity between theory and simulation is due to the discretization of the chain in the simulations, so therefore we expect that a more realistic chain would more closely agree with the original, continuous theory.
\subsection{Translating From Geometric to Chain Parameters}
The theory presented as such is entirely based on geometric parameters, such as the polymer aspect ratio $N$ and the vertical length scale $a$. In this respect, no particular chain model is assumed and the theory is very general. It is possible to incorporate chain characteristics into such a theory through a change of variables. We take the De Gennes model of a polymer in flow as a simple example, and we use scaling to show how we can obtain the traditional scaling relationship found for the far-field dumbbell model.~\protect\cite{DEGENNES:1974p2152} We use the far-field result for shear flow, $\tilde{F}_{L,shear} \approx \frac{3}{8} \dot{\gamma} \tau \sin{(2 \theta)} \frac{N^4}{\tilde{Z}^2}$, and make the following scaling replacements:
\begin{eqnarray}
N \sim Wi \\
a \sim \frac{n^{1/2} b}{Wi^{1/2}} \\
\sin{(2 \theta)} \sim Wi^{-1}
\end{eqnarray}
which are appropriate at intermediate values of $Wi$.~\protect\cite{DEGENNES:1974p2152} $n$ is the degree of polymerization, and $b$ is the monomer size. The translation of $\tau$ to $\tau_Z$ is done by noting that $\tau \sim \eta a^3 / (kT)$ and $\tau_Z \sim \eta b^3 n^{3/2} / (kT)$. This leads to the result:
\begin{equation}
\tilde{F}_L \sim \frac{Wi^2 n^{1/2}}{\tilde{Z}^2}
\end{equation}
where all distances are now scaled by $b$. The depletion zone can be calculated by counterbalancing the lift force with the diffusive force pushing the polymer back towards the surface, $F_S$. Since there is a no-flux condition at the surface, we expect the concentration profile to be of the form $c(\tilde{Z}) = c(0) + A \tilde{Z}^2$ where $A$ is an arbitrary constant. The diffusive force is thus given by $\tilde{F}_S \sim \nabla c(\tilde{Z}) \sim -\tilde{Z}$. There is then a critical height $\tilde{Z}^*$ where these forces balance, which is the depletion length:
\begin{equation}
\tilde{Z}^* \sim Wi^{2/3} n^{1/2}
\end{equation}
which is the relationship given in the literature.~\protect\cite{Ma:2005p5986} We add the caveat that these scaling relationships are only appropriate in a finite range of Weissenberg numbers, and is thus not universal. Separate analysis would have to be performed to obtain this sort of relationship under different conditions, however the underlying geometric theory is completely general.
In the letter, order of magnitude calculations are done to relate the geometric parameters to the chain parameters. This is done using the relationship (from the above scaling analysis):
\begin{equation}
\frac{2 Wi^2 n^{1/2} b}{Z} \sim \frac{\dot{\gamma} \tau_{eff} N^4 a}{Z}
\end{equation}
Using the values and relationships $N = 10$, $\dot{\gamma} \tau_{eff} = 0.01$, and $n^{1/2} b/a \sim Wi^{1/2}$, we obtain the result:
\begin{equation}
Wi \sim 5
\end{equation}
This is an order of magnitude result, but it demonstrates that the results presented in this paper are in the relevant shear rate regime to see the effects indicated (lift and depletion).
\bibliographystyle{plainnat}
|
\section{Introduction}
All graphs considered in this paper are simple, finite and
undirected. We follow the notation and terminology of
\cite{Bondy-Murty}. An edge-colored graph $G$ is rainbow connected
if any pair of distinct vertices are connected by a path whose edges
have distinct colors. Clearly, if a graph is rainbow edge-connected,
then it is also connected. Conversely, any connected graph has
trivial edge coloring that makes it rainbow edge-connected; just
color each edge with a distinct color. The rainbow connection number
of a connected graph $G$, denoted by $rc(G)$, is the minimum number
of colors that are needed in order to make $G$ rainbow connected,
which was introduced by Charrand et al. Obviously, we always have
$diam(G)\leq rc(G)\leq n-1$, where $diam(G)$ denotes the diameter of
$G$. Notice that $rc(G)=1$ if and only if $G$ is a complete graph,
and that $rc(G)=n-1$ if and only if $G$ is a tree.
In \cite{KY}, Krivelevich and Yuster proposed the concept of rainbow
vertex-connection. A vertex-colored graph is rainbow
vertex-connected if any pair of distinct vertices are connected by a
path whose internal vertices have distinct colors. The rainbow
vertex-connection of a connected graph $G$, denoted by $rvc(G)$, is
the minimum number of colors that are needed to make $G$ rainbow
vertex-connected. An easy observation is that if $G$ is a connected
graph with $n$ vertices then $rvc(G)\leq n-2$. We note the trivial
fact that $rvc(G)=0$ if and only if $G$ is a complete graph. Also,
clearly, $rvc(G)\geq diam(G)-1$ with equality if the diameter is $1$
or $2$.
A Nordhaus--Gaddum-type result is a (tight) lower or upper bound on
the sum or product of a parameter of a graph and its complement. The
name Nordhaus--Gaddum-type is used because in 1956 Nordhaus and
Gaddum \cite{Nordhaus-Gaddum} first established the following
inequalities for the chromatic numbers of graphs, they proved that
if $G$ and $\overline{G}$ are complementary graphs on $n$ vertices
whose chromatic numbers are $\chi(G)$, $\chi(\overline{G})$
respectively, then
$$2\sqrt{n}\leq \chi(G)+\chi(\overline{G}) \leq n+1.$$ Since then, many
analogous inequalities of other graph parameters are concerned, such
as domination number \cite{Harary-Haynes}, Wiener index and some
other chemical indices \cite{Zhang-Wu}, and so on.
In \cite{LCL}, the authors considered Nordhaus--Gaddum-type result
for the rainbow connection number. In this paper, we are concerned
with analogous inequalities involving the rainbow vertex-connection
number of graphs. We prove that $$2\leq rvc(G)+rvc(\overline{G})\leq
n-1.$$
The rest of this paper is organized as follows. Section 2 contains
the proof of the sharp upper bound. Section 3 contains the proof of
the sharp lower bound.
\section{Upper bound on $rvc(G)+rvc(\overline{G})$}
We begin this section with two lemmas that are needed in order to
establish the proof of the upper bound.
\begin{lem}\label{thm1}
Let $G$ be a nontrivial connected graph of order $n$, and
$rvc(G)=k$. Add a new vertex $v$ to $G$, and make $v$ be adjacent to
$q$ vertices of $G$, the resulting graph is denoted by $G'$. Then if
$q\geq n-k$, we have $rvc(G')\leq k.$
\end{lem}
\begin{pf}Let $c: V(G) \rightarrow \{1, 2, \cdots, k\}$ be a rainbow
$k$-vertex-coloring of $G$, $X=\{x_1, x_2, \cdots, x_q\}$ be the
vertices that are adjacent to $v$, $V\backslash X=\{y_1, y_2,
\cdots, y_{n-q}\}$. We can assume that there exists some $y_j$ such
that there is no rainbow vertex-connected-path from $v$ to $y_{j}$;
otherwise, the result holds obviously. Because $G$ is a rainbow
$k$-vertex-coloring, there is a rainbow vertex-connected-path $P_i$
from $x_i$ to $y_j$ for every $x_i$, $i\in \{1,2,\cdots, q\}$.
Certainly, $P_i \bigcap P_j$ may not be empty. We notice that no
other vertices of $\{x_1,x_2,\cdots,x_q\}$ different from $x_i$
belong to $P_i$ for each $1\leq i \leq q$. If so, let ${x_i}'$ be
the last vertex in $\{x_1,x_2,\cdots,x_q\}$ which belongs to $P_i$,
denote $P_i$ by $x_i{P_i}'{x_i}'Q_iy_j$, then $v{x_i}'Q_iy_j$ is a
rainbow vertex-connected-path, a contradiction to our assumption.
Since $v$ and $y_j$ are not rainbow vertex-connected, for each
$P_i$, there is some $y_{k_i}$ such that $c(x_i)=c(y_{k_i})$. That
means that the colors that are assigned to $X$ are among the colors
that are assigned to $V\backslash X $. So $rvc(G)=k\leq n-q$. By the
hypothesis $q\geq n-k$, we have $rvc(G)= n-q$, that is, all vertices
in $V\backslash X$ have distinct colors. Now we construct a new
graph $G'=P_1\bigcup P_2\bigcup \cdots \bigcup P_q$. For every $y_t$
not in $G'$, there is a $y_s \in G'$ such that $y_ty_s \in E(G)$. Or
$N(y_t)\subseteqq \{x_1,x_2,\cdots, x_q\}$. Since $G$ is rainbow
$k$-vertex-connected, there is a rainbow vertex-connected path from
$y_t$ to $y_j$, denoted by $y_tx_kQy_j$, where $x_k \in N(y_t)$.
Thus $vx_kQy_j$ is a rainbow vertex-connected path, a contradiction.
It follows that $G[y_1,y_2,\cdots,y_{n-q}]$ is connected. Certainly,
$G[y_1,y_2,\cdots,y_{n-q}]$ has a spanning tree $T$, and $T$ has at
least two pendant vertices. Then there must exist a pendant vertex
whose color is different from $x_1$, and we assign the color to
$x_1$. It is easy to check that $G$ is still rainbow
$k$-vertex-connected, and there is a rainbow vertex-connected path
between $v$ and $y_j$. If there still exists some $y_j$ such that
$v$ and $y_j$ are not rainbow vertex-connected, we do the same
operation, until $v$ and $y_j$ are rainbow vertex-connected for each
$j\in\{1,2,\cdots, n-q\}$. Thus $G'$ is rainbow vertex-connected. It
follows that $rvc(G')\leq k.$
\end{pf}
\begin{qed}
\end{qed}
\begin{lem}\label{lem1}
Let $G$ be a connected graph of order $5$. If $\overline{G}$ is
connected, then $rvc(G)+rvc(\overline{G})\leq 4$.
\end{lem}
\begin{pf}We consider the situations of $G$.
First, if $G$ is a path, then $rvc(G)=3$. In this case
$diam(\overline{G})=2$, and then $rvc(\overline{G})=1$.
Second, if $G$ is a tree but not a path, then $rvc(G)<3$. Since $G$
is a bipartite graph, then $\overline{G}$ consists of a $K_2$ and a
$K_3$ and two edges between them. So we assign color $1$ to the
vertices of $K_2$ and color $2$ to the vertices of $K_3$, and this
makes $\overline{G}$ rainbow vertex-connected, that is,
$rvc(\overline{G})\leq 2$.
Finally, if both $G$ and $\overline{G}$ are not trees, then
$e(G)=e(\overline{G})=5$. If $G$ contains a cycle of length $5$,
then $G=\overline G=C_5$, thus $rvc(G)=rvc(\overline{G})=1$. If $G$
contains a cycle of length $4$, there is only one graph $G$ which is
showed in Figure 1, we can color $G$ and $\overline G$ with 2 colors
to make them rainbow vertex-connected, see Figure 1. If $G$ contains
a cycle of length $3$, then $G$ and $\overline G$ are showed in
Figure 2. By the coloring showed in the graphs, we have
$rvc(G)+rvc(\overline{G})=4.$
\setlength{\unitlength}{1.2mm}
\begin{center}
\begin{picture}(80,60)
\put(0,45){\circle*{1}} \put(10,45){\circle*{1}}
\put(0,55){\circle*{1}} \put(10,55){\circle*{1}}
\put(20,55){\circle*{1}} \put(50,45){\circle*{1}}
\put(50,55){\circle*{1}} \put(60,50){\circle*{1}}
\put(70,50){\circle*{1}} \put(80,50){\circle*{1}}
\put(-0.1,45){\line(0,1){10}} \put(0,45){\line(1,0){10}}
\put(0,55){\line(1,0){20}}\put(10,45){\line(0,1){10}}
\put(50,45){\line(0,1){10}}\put(60,50){\line(1,0){20}}
\put(60,50){\line(-2,1){10}}\put(50,45){\line(2,1){10}}
\put(5,13){\circle*{1}} \put(15,13){\circle*{1}}
\put(5,20){\circle*{1}} \put(15,20){\circle*{1}}
\put(10,27){\circle*{1}} \put(55,13){\circle*{1}}
\put(65,13){\circle*{1}} \put(55,20){\circle*{1}}
\put(65,20){\circle*{1}} \put(60,27){\circle*{1}}
\put(5,13){\line(0,1){7}} \put(15,13){\line(0,1){7}}
\put(5,20){\line(1,0){10}}\put(5,20){\line(3,4){5}}
\put(15,20){\line(-3,4){5}}\put(55,13){\line(0,1){7}}
\put(65,13){\line(0,1){7}}
\put(55,20){\line(1,0){10}}\put(55,20){\line(3,4){5}}
\put(65,20){\line(-3,4){5}}
\put(-0.5,56){\tiny{$1$}}\put(10,56){\tiny{$2$}}
\put(60,51){\tiny{$1$}}\put(70,51){\tiny{$2$}}
\put(4.5,21){\tiny{$1$}}\put(15,21){\tiny{$2$}}
\put(54.5,21){\tiny{$1$}}\put(65,21){\tiny{$2$}}
\put(10,40){\small$G$}\put(60,40){\small$\overline{G}$}
\put(10,32){Figure $1$: $G$ contains a cycle of length 4.}
\put(10,8){\small$G$}\put(60,8){\small$\overline{G}$}
\put(8,2){Figure $2$: $G$ contains a cycle of length 3.}
\end{picture}
\end{center}
By these cases, we have $rvc(G)+rvc(\overline{G})\leq 4.$
\end{pf}
\begin{qed}
\end{qed}
From the above lemmas, we have our first theorem.
\begin{thm}
$rvc(G)+rvc(\overline{G})\leq n-1$ for all $n\geq 5$, and this bound
is best possible.
\end{thm}
\begin{pf}
We use induction on $n$. By Lemma \ref{lem1}, the result is evident
for $n=5$. We assume that $rvc(G)+rvc(\overline{G})\leq n-1$ holds
for complementary graphs on $n$ vertices. To the union of a
connected graph $G$ and its $\overline{G}$, which forms the complete
graph on these $n$ vertices, we adjoin a vertex $v$. Let $q$ of the
$n$ edges between $v$ and the union be adjoined to $G$ and the
remaining $n-q$ edges to $\overline{G}$. If $G'$ and $\overline{G'}$
are the graphs so determined (each of order $n+1$), then
$$rvc(G')\leq rvc(G)+1, \ \ rvc(\overline{G'})\leq rvc(\overline{G})+1.$$
These inequalities are evident from the fact that if given a rainbow
$rvc(G)$-vertex-coloring ($rvc(\overline{G})$-vertex-coloring) of
$G$ ($\overline{G}$), we assign the new color to the vertex which is
adjacent to $v$ and keep other vertices unchanged, the resulting
coloring makes $G'$ ($\overline{G'}$) rainbow vertex-connected. Then
$rvc(G')+rvc(\overline{G'})\leq rvc(G)+rvc(\overline{G})+2\leq n+1.$
And $rvc(G')+rvc(\overline{G'})\leq n$ except possibly when
$$rvc(G')=rvc(G)+1, \ \ rvc(\overline{G'})=rvc(\overline{G})+1.$$
In this case, by Lemma \ref{thm1}, $q\leq n-rvc(G)-1, n-q\leq
n-rvc(\overline{G})-1$, thus $rvc(G)+rvc(\overline{G})\leq n-2,$
from which $rvc(G')+rvc(\overline{G'})\leq n$. This completes the
induction.
The following example shows that the bound established is sharp for
all $n \geq 5$: If $G$ be a path of order $n$, then $rvc(G)= n-2$.
It is easy to obtain $\overline{G}$, and check that
$diam(\overline{G})=2$. Then $rvc(\overline{G})= 1$, and so we have
$rvc(G)+rvc(\overline{G})= n-1$.
\end{pf}
\begin{qed}
\end{qed}
\section{Lower bound on $rvc(G)+rvc(\overline{G})$}
As we note that $rvc(G)=0$ if and only if $G$ is a complete graph.
Thus if we want both $G$ and $\overline{G}$ are connected, and so
$rvc(G)\geq 1, rvc(\overline{G})\geq 1$. Then $rvc(G)+
rvc(\overline{G})\geq 2.$ Our next theorem shows that the lower
bound is sharp for all $n \geq 5$.
\begin{thm}\label{thm2}
For $n \geq 5$, the lower bound of $rvc(G)+ rvc(\overline{G})\geq 2$
is best possible, that is, there are graphs $G$ and $\overline{G}$
with $n$ vertices, such that $rvc(G)= rvc(\overline{G})=1.$
\end{thm}
\begin{pf}We only need to prove that for $n\geq 5$, there are graphs $G$ and
$\overline{G}$ with $n$ vertices, such that $diam(G)=
diam(\overline{G})=2.$
We construct $G$ as follows: if $n=2k+1$,
$$
V(G)=\{v,v_1,v_2,\cdots,v_k,u_1,u_2,\cdots,u_k\}
$$
$$
E(G)=\{vv_i|1\leq i\leq k\}\bigcup \{v_iu_i|1\leq i\leq k\}\bigcup
\{u_iu_j|1\leq i,j\leq k\};
$$
if $n=2k$,
$$
V(G)=\{v,v_1,v_2,\cdots,v_k,u_1,u_2,\cdots,u_{k-1}\}
$$
$$
E(G)=\{vv_i|1\leq i\leq k\}\bigcup \{v_iu_i|1\leq
i<k\}\bigcup\{v_ku_{k-1}\}\bigcup \{u_iu_j|1\leq i,j\leq k-1\}.
$$
We can easily check that $diam(G)= diam(\overline{G})=2$.
\end{pf}
\begin{qed}
\end{qed}
|
\subsection*{Motivation}
Consider a dynamical system $f:X\to X$ on a proper metric space $X$
having a global attractor, i.e. a compact invariant set $A$ that
attracts all its (bounded) neighborhoods. We will not consider the
map $f$ itself, but the map $f_{*}:\mathcal{P}_{1}(X)\to\mathcal{P}_{1}(X)$
defined via push-forward map on the space of probability measures.
If $f$ is {}``nice'' then $f_{*}$ is continuous and has the global
attractor $K=\mathcal{P}(A)$, i.e. the probability measures supported
on the global attractor of $f$.
Markov-type noise can be considered as a perturbation $\tilde{F}$
of $f_{*}$, i.e. instead of $\delta_{x}\mapsto\delta_{f(x)}$ we
have $\delta_{x}\mapsto p(dy|x)$ where $p(dy|x)$ and $\delta_{f(x)}$
uniformly close w.r.t. the Wasserstein distance $w_{1}$ for all $x$.
This can be seen as a smearing of the image $f(x)$ or some uncertainty
about the actual image. For example, if $f$ is the time-$1$ map
of a flow generated by the ODE $\dot{x}=g(x)$ then $\tilde{F}$ could
be the time-one map of the flow of distributions of the SDE $dx=g(x)dt+\epsilon dW_{t}$,
i.e. additive Gaussian noise with small variance.
If $\tilde{F}$ and $f_{*}$ are sufficiently close then $\tilde{F}$
has a (weak) global attractor (in $\mathcal{P}_{1}(X)$) which is
close to $K$ w.r.t. $w_{1}$. Hence stability of the global attractor
holds in the Wasserstein space $\mathcal{P}_{1}(X)$.
The following example is inspired by Crauel, Flandoli - {}``Additive
Noise Destroys a Pitchfork Bifurcation'' \cite{Crauel1998} and could
be decribed as {}``Additive Noise Destroys Attractors''. The noise
will be worse than white noise used by Crauel and Flandoli, but can
still be considered as small.
\begin{example*}
[Generic collapse under "small" noise](1) Suppose $f:X\to X$ has
a global attractor and at least one fixed point $x_{0}$ (the argument
works equally well with general attractors). Take any noise level
$\epsilon>0$ and let $P_{\epsilon}:\mathcal{P}(X)\to\mathcal{P}(X)$
be the Markov map induced by \[
x\mapsto(1-\epsilon)\delta_{f(x)}+\epsilon\delta_{x_{0}}.\]
This map is (weakly) close to the unperturbed system $f_{*}:\mathcal{P}(X)\to\mathcal{P}(X)$.
Namely, if $d_{LP}$ is the Levy-Prokhorov distance (which metrizes
$\mathcal{P}(X)$) then \[
\sup_{\mu\in P(X)}d_{LP}(P_{\epsilon}(\mu),f_{*}(\mu))\le\epsilon.\]
But $P$ has exactly one invariant measure, namely $\delta_{x_{0}}$,
and all others converge to this measure weakly.
(2) Now we want to show that this can also happen in any Wasserstein
space $\mathcal{P}_{p}(X)$ for $1\le p<\infty$ ($\mathcal{P}_{\infty}(X)$
only allows bounded noise which when sufficiently small cannot destroy
local attractors and thus a global attractors with at least two sinks
never collapses, see \cite{Kell2011b}). Suppose $a:X\to[0,1]$ is
a continuous function. Define $q_{a}(dy|\cdot):X\to\mathcal{P}_{p}(X)$
by \[
x\mapsto q_{a}(dy|x)=(1-a(x))\delta_{f(x)}+a(x)\delta_{x_{0}},\]
which is obviously continuous. Thus \[
w_{p}(q_{a}(dy|x),\delta_{f(x)})^{p}=a(x)d(f(x),x_{0})^{p}.\]
So if we define \[
a(x)=\frac{\epsilon^{p}}{1+d(f(x),x_{0})^{p}}\]
then \[
w_{p}(q_{a}(dy|x),\delta_{f(x)})^{p}=\frac{\epsilon^{p}\cdot d(f(x),x_{0})^{p}}{1+d(f(x),x_{0})^{p}}\le\epsilon^{p}.\]
So, in particular, the induced MW-map $Q_{a}:\mathcal{P}_{p}(X)\to\mathcal{P}_{p}(X)$
of order $p$ relative to $f$ has noise level $\epsilon$. Furthermore,
the only invariant measure of $Q_{a}$ is $\delta_{x_{0}}$ and all
other measures converge to it.
\end{example*}
The example above should make clear that using arbitrary unbounded
noise even when it is small can have strange effects on the global
attractor. Although we have some {}``attracting'' invariant measures
of the perturbed system the attractor might look very different from
the original one, in our case it might be just one fixed point and
this one can even be the {}``most'' unstable one of the original
attractor. Therefore, stochastic stability of attractors under arbitrary
{}``small'' noise should not be referred to a single invariant measure
but to all of them, even though we can speak of stochastic stability
if the type of noise is more restricted, besides of being sufficiently
{}``small''.
\section{Discrete-time Conley theory for stable invariant sets}
In this section we will use Conley theory, that is continuation methods
from Conley index theory without using the topological (or (co)homological)
Conley index. We will prove a continuation for a positive invariant
neighborhood of a stable isolated invariant set of a time discrete
dynamical system. The result will not require a compactness assumption
(called admissibility) of the perturbed system and is a different
type of continuation than \cite{MroRyb1991}. Our proof will follow
the proof of {\cite[Theorem 12.3]{Rybakowski1987}} which is the continuation
for semiflows. In particular, the results stated here and in the next
sections also hold for semiflows if we assume that they do not explode
on a given neighborhood.
In both cases the Wa{\.z}ewski principle for the index pair of the
perturbed system does not apply. But we can use other assumptions
to show that attractors continue, e.g. the map is weakly continuous
and closed $\delta$-neighborhoods of compact set are weakly compact,
which is the case for the Wasserstein space on proper metric spaces.
We will now give the definitions used in \cite{MroRyb1991} and \cite{Rybakowski1987}
to prove the existence of an index pair for certain isolated invariant
sets. Our setting will be a complete separable metric space $Y$ and
a dynamical system, i.e. a continuous map $f:Y\to Y$. A full left
solution of $f$ in $N$ is a sequence $\{x_{-n}\}_{n\in\mathbb{N}}\subset N$
such that $f(x_{n-1})=x_{n}$ for $n\le0$. Define the following sets\begin{eqnarray*}
A^{+}(N) & = & \{x\in N\,|\, f^{k}(x)\in N\,\mbox{for all }k\ge0\}\\
A^{-}(N) & = & \{x\in N\,|\,\exists\mbox{\,\ full left solution \ensuremath{\{x_{-n}\}_{n\in\mathbb{Z}}}\,\ in \,\ensuremath{N}\,\ through \ensuremath{x_{0}=x}}\}\\
A(N) & = & A^{+}(N)\cap A^{-}(N).\end{eqnarray*}
These are called the maximal positive invariant (resp. negative invariant,
resp. invariant) set in $N$. If $N$ is unbounded then $A(N)$ usually
denotes only the bounded invariant orbits instead of all of them.
A set $K$ is called invariant if $A(K)=K$. If there is a closed
neighborhood $N$ of an invariant set $K$ with $A(N)=K$ then $K$
is called isolated with isolating neighborhood $N$.
For $l,m\in\mathbb{N}$ and $l\le k$ define \[
f^{[l,m]}(x)=\{y\,|\, f^{k}(x)=y\,\mbox{for some }k\in\mathbb{N}\cap[l,m]\}\]
If $f_{n}:Y\to Y$ is a sequence of continuous maps such that $f_{n}\to f$,
i.e. $f_{n}(x_{n})\to f(x)$ whenever $x_{n}\to x$ as $n\to\infty$,
then we say that a closed bounded set $N$ is $\{f_{n}\}$-admissible
if for any sequence $\{x_{n}\}_{n\in\mathbb{N}}$ with $f_{n}^{[0,m_{n}]}(x_{n})\subset N$
and $m_{n}\to\infty$ the sequence of endpoints $\{f_{n}^{m_{n}}(x_{n})\}_{n\in\mathbb{N}}$
is precompact. In case this property holds for $f_{n}\equiv f$ then
we just say $N$ is $f$-admissible.
\begin{rem*}
Later on, we deal with dynamical systems on the space of probability
measures $Y=\mathcal{P}(X)$ for some metric space $X$. An invariant
measure for that system is invariant w.r.t. the definition above.
In particular, periodic measures will be called invariant. A fixed
point for these systems will be called stationary measure.
\end{rem*}
In the following we will use several ideas from \cite{MroRyb1991}:
Let $N,N'$ be two $f$-admissible isolating neighborhood for some
isolated invariant set $K$ with \[
N\subset\operatorname{int}N'\cap f^{-1}(\operatorname{int}N').\]
The authors in {\cite[4.4]{MroRyb1991}} used a so called Lyapunov
pair $(\phi,\gamma)$ which is continuous on a small neighborhood
$W\subset N$ of $K$ and has the following properties: $K\subset\gamma^{-1}(0)$,
$\phi$ (resp. $\gamma$) is decreasing (resp. increasing) along orbits
and $\phi(x)=0$ with $x\in W$ implies \[
x\in A^{-}(N)\cup\partial N'.\]
Because $K$ is compact we can choose $d(W,\partial N')>0$ and assume
$x\in A^{-}(N)$ whenever $\phi(x)=0$. Furthermore, it is shown that
if $\phi(x_{n})\to0$ then $x_{n}$ admits a convergent subsequence.
\begin{thm}
\label{thm:stable}Suppose $N$ is an isolating neighborhood for $K$
such that the assumptions above hold and \[
A^{-}(N)=A(N)=K\ne\varnothing.\]
Then there exists an admissible isolating neighborhood $B\subset N$
which is positive invariant, i.e. no trajectories exit $B$.\end{thm}
\begin{rem*}
This is the discrete time version of {\cite[I-5.5]{Rybakowski1987}}
using the theory of \cite{MroRyb1991}. The proof is essentially copied
from Rybakowski using the Lyapunov pair above.\end{rem*}
\begin{proof}
Define \begin{eqnarray*}
P_{1}^{\epsilon} & = & N\cap\operatorname{cl}\{x\in\operatorname{int}N'\,|\,\phi(x)<\epsilon\}\\
P_{2}^{\epsilon} & = & P_{1}^{\epsilon}\backslash\{x\in\operatorname{int}N'\,|\,\gamma(x)<\epsilon\}.\end{eqnarray*}
it was shown {\cite[4.4]{MroRyb1991}} that $P_{1}^{\epsilon}\subset W$
is a neighborhood of $K$ for sufficiently small $\epsilon>0$ and
whenever $x\in P_{i}^{\epsilon}$ and $f(x)\in N$ then $x\in P_{i}^{\epsilon}$
and if $x\in P_{1}^{\epsilon}$ and $f(x)\notin N$ then $x\in P_{2}^{\epsilon}$,
i.e. $P_{2}^{\epsilon}$ is the exit ramp for $P_{1}^{\epsilon}$.
Now fix a sufficiently small $\epsilon>0$ and let $0<\delta\le\epsilon$
then $P_{1}^{\delta}\subset P_{1}^{\epsilon}$. Define $\tilde{P}_{2}^{\delta}:=P_{1}^{\delta}\cap P_{2}^{\epsilon}$
then because $\phi$ is decreasing along orbits $\tilde{P}_{2}^{\delta}$
is still an exit ramp for $P_{2}^{\delta}$.
If $A^{-}(N)=A(N)=K$ then we claim that there is a $\delta>0$ such
that $\tilde{P}_{2}^{\delta}=\varnothing$ which implies that $P_{1}^{\delta}$
is positive invariant. If this does not hold then there is a sequence
$x_{n}\in P_{1}^{\delta_{n}}\cap P_{2}^{\epsilon}$ with $\delta_{n}\to0$.
Thus $\phi(x_{n})\to0$ and $\gamma(x_{n})\ge\epsilon$ which implies
that there is a subsequence $x_{n'}\to x\in N$ such that $\phi(x)=0$.
Hence \[
x\in A^{-}(N)=K\]
and $\gamma(x)=0$ by assumption. But $\gamma$ is continuous and
$x_{n'}\to x$ implies $\epsilon\le\gamma(x_{n'})\to\gamma(x)=0$
which is a contradiction. This proofs our claim and thus the theorem.
\end{proof}
In the following we will assume that $K$ satisfies the assumption
of the theorem and that $B:=P_{1}^{\delta_{0}}$, $\phi$ and $\gamma$
are given as in the proof. It is obvious that $P_{1}^{\delta}$ is
positive invariant w.r.t. $f$ for any $0<\delta\le\delta_{0}$. Furthermore,
suppose $f_{n}\to f$.
\begin{thm}
\label{thm:cont}Assume $N'$ (see above) is $\{f_{n_{m}}\}$-admissible
for each subsequence of $\{f_{n}\}_{n\in\mathbb{N}}$. Set $\tilde{U}=\operatorname{int}B$
and define \[
V(a)=\{x\in\tilde{U}\,|\,\phi(x)<a\}.\]
Then for some $a_{0}>0$, $N:=\operatorname{cl}V(a_{0})\subset\tilde{U}$.
Furthermore, for some sufficiently small $\epsilon_{0}>0$ and all
$0<\epsilon\le\epsilon_{0}$ there is an $n_{0}=n_{0}(\epsilon)$
such that for all $n\ge n_{0}$ there is a positive $f_{n}$-invariant
closed $N_{n}(\epsilon)$ and \[
K_{n}\subset V(\epsilon)\subset N_{n}(\epsilon)\subset N.\]
\end{thm}
\begin{rem*}
The complete continuation theorem for index pairs does not hold for
discrete time dynamical systems in general. A proof would require
that there is a neighborhood such that the exit time is continuous
in $\tilde{U}$ , i.e. $\omega_{n}^{+}(x_{n})\to\omega^{+}(x_{0})$
whenever $x_{n}\to x_{0}$ in $\tilde{U}$, which holds for semiflows
only for so called isolating blocks. These blocks do not necessarily
exist for continuous maps.\end{rem*}
\begin{proof}
By {\cite[3.9]{MroRyb1991}} there is an $a_{0}>0$ such that $N=\operatorname{cl}V(a_{0})\subset\tilde{U}$.
And similar to {\cite[I-4.5]{Rybakowski1987}} we can show that for
$0<\epsilon\le a_{0}$ and all $n\ge n_{0}(\epsilon)$ \[
K_{n}\subset V(\epsilon).\]
Define \[
N_{n}(\epsilon)=N\cap\operatorname{cl}\{y\,|\,\mbox{ \ensuremath{\exists x\in V(\epsilon)}, \ensuremath{m\ge0}\,\ s.t.\,\ \ensuremath{f_{n}^{[0,m]}(x)\subset\tilde{U}\,}and \ensuremath{f_{n}^{m}(x)=y}}\}.\]
Following the proof of {\cite[I-12.5]{Rybakowski1987}} we can show
that $N_{n}(\epsilon)$ satisfies the following properties for $n\ge n_{0}(\epsilon)$
\begin{itemize}
\item $x\in N_{n}(\epsilon)$ and $f_{n}(x)\in N$ implies $f_{n}(x)\in N_{n}(\epsilon)$
\item $K_{n}\subset V(\epsilon)\subset N_{n}(\epsilon)$
\end{itemize}
We claim that for small $\epsilon_{0}>0$ whenever $\epsilon\le\epsilon_{0}$
and $n\ge n_{0}(\epsilon)$ then $N_{n}(\epsilon)$ is positive invariant
w.r.t. $f_{n}$. If this is not true then there is a sequence $\epsilon_{m}\to0$
and \[
y_{m}\in N_{n_{m}}(\epsilon_{m})\]
with $f_{n_{m}}(y_{m})\notin N$. By definition of $N_{n_{m}}(\epsilon_{m})$
there is a sequence $\tilde{y}_{m}\in Y$, $x_{m}\in V(\epsilon_{m})$
and $k_{m}\ge0$ such that $d(y_{m},\tilde{y}_{m})<2^{-m}$, $f_{n_{m}}^{[0,k_{m}]}(x_{m})\subset\tilde{U}$
and $\tilde{y}_{m}=f_{n_{m}}^{k_{m}}(x_{m})$. Because $\phi(x_{m})\to0$
and $A_{f}^{-}(B)=A_{f}(B)$ we can assume w.l.o.g. that $x_{m}\to x_{0}\in A_{f}(B)$.
Admissibility and $f_{n_{m}}\to f$ imply the sequence $\{f_{n_{m}}^{k_{m}}(x_{m})\}_{m\in\mathbb{N}}$
has a convergent subsequence and w.l.o.g. $\tilde{y}_{m}=f_{n_{m}}^{k_{m}}(x_{m})\to y_{0}\in A_{f}^{-}(N')=A_{f}(N')\subset\operatorname{int}N$
and thus $y_{m}\to y_{0}$.
Since $f_{n_{m}}(y_{m})\notin N\subset\operatorname{int}N'\cap f^{-1}(\operatorname{int}N')$
and $f_{n}\to f$ we have $f_{n_{m}}(y_{m})\to f(y_{0})\in N'\backslash\operatorname{int}N$.
But $y_{0}\in A_{f}(N')$ implies $f(y_{0})\in A_{f}(N')$ which is
a contradiction because $A_{f}(N')$ and $N'\backslash\operatorname{int}N$
are disjoint.\end{proof}
\begin{cor}
Under the assumption above for all $n\ge n_{0}$ we can find positive
$f_{n}$-invariant $N_{n},N_{n}^{'}$ such that \[
N_{n}\subset U_{\delta}(K)\subset N_{n}^{'}\subset B\]
for some $\delta$-neighborhood of $K$ denoted by \[
U_{\delta}(K)=\{x\in Y\,|\, d(x,y)<\delta\,\mbox{for some}\: y\in K\}.\]
Furthermore, we have $K\subset\operatorname{int}N_{n}$ and there
is an $\epsilon>0$ such that \[
U_{\epsilon}(K_{n})\subset N_{n}^{'}.\]
\end{cor}
\begin{proof}
Applying the previous theorem we get \[
K\cup K_{n}\subset V(\tilde{\epsilon})\subset N_{n}^{'}\subset B.\]
Recalling the definition of $\phi$ it is obvious that because $A_{f}^{-}(B)=A_{f}(B)=K$
for small $0<\epsilon'<1$ and $x\in P_{1}^{\epsilon'}$ \[
d(x,K)\le\epsilon'.\]
Because $K$ is compact and $V(\tilde{\epsilon})$ a neighborhood
of $K$ the $\delta$-neighborhood $U_{\delta}(K)$ of $K$ is contained
in $V(\tilde{\epsilon})$ for $\delta$ sufficiently small. Furthermore,
we can find an $\epsilon'>0$ with $\epsilon'<\delta$ such that $P_{1}^{\epsilon'}\subset U_{\delta}(K)\subset V(\tilde{\epsilon})$.
Applying the theorem again for $P_{1}^{\epsilon'}$ instead of $B$
we get for $n\ge n_{0}$ \[
N_{n}\subset P_{1}^{\epsilon'}\subset U_{\delta}(K)\subset N_{n}^{'}\subset B\]
and $K\subset V(\epsilon')\subset N_{n}$.
To show that $U_{\epsilon}(K_{n})\subset N_{n}$ we need another positive
$f_{n}$-invariant neighborhood $N_{n}^{''}$. First note that there
is a $\delta'>0$ such that \[
d(x,K)\ge\delta'\]
for all $x\in\partial P^{\epsilon'}$. So if we choose $0<2\epsilon<\delta$
then \[
U_{\epsilon}(P_{1}^{\epsilon})\subset P_{1}^{\epsilon'}.\]
Applying the previous theorem again we get a positive $f_{n}$-invariant
isolating $N_{n}^{''}$ of $K_{n}$ inside of $P_{1}^{\epsilon}$.
Hence \[
U_{\epsilon}(K_{n})\subset U_{\epsilon}(N_{n}^{''})\subset U_{\epsilon}(P_{1}^{\epsilon})\subset P_{1}^{\epsilon^{'}}\subset N_{n}^{'}.\]
\end{proof}
Now we are able to continue the attractor. Instead of an admissibility
assumption for the perturbed map we will use weak compactness of close
$\delta$-neighborhoods of compact sets.
\begin{defn}
[weak attractor]Suppose $Y$ has a weaker (Hausdorff) topology (i.e.
$x_{n}\to x$ strongly implies $x_{n}\rightharpoonup x$ weakly) and
$f$ is continuous and weakly continuous. An isolated invariant set
$K$ is called a weak attractor if it admits a positive $f$-invariant
isolating neighborhood $N$ such that $\omega^{\tiny\mbox{weak}}(N)\subset K$
where $\omega^{\tiny\mbox{weak}}(N)$ is defined as \[
\omega^{\tiny\mbox{weak}}(N)=\{y\in Y\,|\,\exists x_{n}\in N,m_{n}\to\infty\,\mbox{s.t. }\, f^{m_{n}}(x_{n})\rightharpoonup y\}.\]
\end{defn}
\begin{rem*}
(1) Our definition of weakness of an attractor is w.r.t. the weaker
topology and is different from one defined in \cite{Hurley2001}.
Even our definition of a (strong) attractor is weaker than the one
used there because we only require the existence of a positive invariant
isolating neighborhood of the invariant set. But there might be a
connection to Ochs' weak random attractor \cite{Ochs1999}.
(2) A Conley theory with weak-admissibility instead of admissibility
might not make sense since the continuation proof requires continuity
of the metric and usually the metric is only lower semicontinuous
w.r.t. weak convergence.
(3) A weakly continuous function might not be continuous and vice
versa (see counterexample in the proof of theorem \ref{thm:wasserstein-cont})\end{rem*}
\begin{thm}
\label{thm:weak-continuation}Under the assumption of the previous
theorem, suppose there is a weaker (Hausdorff) topology on $Y$ and
that (strongly) closed $\delta$-neighborhoods of compact sets are
weakly (sequentially) compact, i.e. $\operatorname{cl}U_{\delta}(C)$
is weakly compact for compact $C$. If $f_{n}$ is weakly continuous
then $K_{n}$ is non-empty and a weakly compact weak attractor w.r.t.
$f_{n}$ for all $n\ge n_{0}$. Furthermore, $U_{\epsilon}(K_{n})\subset N_{n}^{'}$
for some $\epsilon>0$ and positive $f_{n}$-invariant $N_{n}^{'}$.\end{thm}
\begin{proof}
Applying the previous corollary we get \[
N_{n}\subset U_{\delta}(K)\subset N_{n}^{'}\subset B\]
and \[
U_{\epsilon}(K_{n})\subset N_{n}^{'}.\]
Because $f_{n}$ is weakly continuous, $N_{n}$ is positive $f_{n}$-invariant
and $\operatorname{cl}U_{\delta}(K)\subset N_{n}^{'}$ is closed and
thus weakly compact the set \[
\omega_{n}^{\tiny\mbox{weak}}(x)=\{y\in Y\,|\, f_{n}^{n_{k}}(x)\rightharpoonup y\,\mbox{for some}\, n_{k}\to\infty\}\]
for $x\in N_{n}$ is non-empty and weakly compact. This implies $K_{n}\ne\varnothing$
and in particular \[
\omega_{n}^{\tiny\mbox{weak}}(x)\subset K_{n}\subset N_{n}.\]
Similarly weak compactness of $\operatorname{cl}U_{\delta}(K)$ implies
$\omega_{n}^{\tiny\mbox{weak}}(N_{n})\subset A_{n}(N_{n}^{'})=K_{n}\subset N_{n}$
and thus $K_{n}$ is a weak attractor. Obviously $K_{n}$ is weakly
closed and contained in the weakly compact set $\operatorname{cl}U_{\delta}(K)$
and is therefore weakly compact as well.
\end{proof}
\section{Wasserstein spaces}
Now we will introduce some notation and results for Wasserstein spaces
of a metric space, general references are \cite{AmbGigSav2008} and
\cite{Villani2009}.
Let $(X,d)$ be a complete separable metric space, also called Polish
space. We call it proper if every bounded closed set is compact. In
particular, this implies that $X$ is locally compact. The metric
of a non-compact proper metric space is necessarily unbounded.
The space of probability measures on the Borel $\sigma$-algebra of
$X$ is denoted by $\mathcal{P}(X)$. This space is given the weak
topology, i.e. $\mu_{n}\rightharpoonup\mu$ if $\int fd\mu_{n}\to\int fd\mu$
for all bounded continuous functions $f$ . Let $x_{0}$ be an arbitrary
point of $X$ and define $\mathcal{P}_{p}(X)$, the Wasserstein space
(of order $p$), by \[
\mathcal{P}_{p}(X)=\{\mu\in\mathcal{P}(X)\,|\,\int d(x_{0},x)^{p}d\mu(x)<\infty\}.\]
Furthermore, define for $\mu,\nu\in\mathcal{P}_{p}(X)$\[
w_{p}(\mu,\nu)=\left(\inf_{\pi\in\Pi(\mu,\nu)}\int d(x,y)^{p}d\pi(x,y)\right)^{\frac{1}{p}}\]
where $\pi\in\Pi(\mu,\nu)\subset\mathcal{P}(X\times X)$ with $\pi(A\times X)=\mu$
and $\pi(X\times A)=\nu$ for all Borel sets $A$ and $B$. Then $(\mathcal{P}_{p}(X),w_{p})$
is a complete separable metric space. This topology is usually stronger
than the induced subspace topology of $\mathcal{P}_{p}(X)\subset\mathcal{P}(X)$.
If $X$ is compact so is $\mathcal{P}_{p}(X)$. And $\mathcal{P}_{p}(X)$
is local compact only if $X$ is compact. A counterexample for non-proper
metric spaces is given in {\cite[7.1.9]{AmbGigSav2008}}. We will
adjust their example to non-compact proper metric spaces by showing
that the closed $\epsilon$-ball $B_{\epsilon}^{w_{p}}(\delta_{x_{0}})$
around $\delta_{x_{0}}$ in $\mathcal{P}_{p}(X)$ cannot be compact
for any $\epsilon>0$ and thus $\mathcal{P}_{p}(X)$ cannot be locally
compact.
\begin{example*}
Assume $X$ is non-compact and proper and define \[
\mu_{n}=m_{n}\delta_{x_{n}}+(1-m_{n})\delta_{x_{0}}\]
for some sequence $\{x_{n}\}_{x\in\mathbb{N}}\subset X$. Then \[
w_{p}(\mu_{n},\delta_{x_{0}})^{p}=m_{n}d(x_{n},x_{0})^{p}.\]
Suppose $d(x_{n},x_{0})\ge\epsilon>0$ and set $m_{n}=\epsilon^{p}\cdot d(x_{n},x_{0})^{-p}$
then $\mu_{n}\in\partial B_{\epsilon}^{w_{p}}(\delta_{x_{0}})$. If
$\{m_{n}\}_{n\in\mathbb{N}}$ stays bounded away from $0$ then $\{d(x_{n},x_{0})\}_{n\in\mathbb{N}}$
is bounded and thus $\{x_{n}\}_{n\in\mathbb{N}}$ and $\{m_{n}\}_{n\in\mathbb{N}}$
have convergent subsequences $x_{n'}\to x_{\infty}$ and $m_{n'}\to m\in(0,1]$
and $\mu_{n}\to m\delta_{x_{\infty}}+(1-m)\delta_{x_{0}}$ strongly
in $\mathcal{P}_{p}(X)$. But if we assume $d(x_{n},x_{0})\to\infty$
then $m_{n}\to0$ and thus $\mu_{n}\rightharpoonup\delta_{x_{0}}$
weakly. Because strong convergence requires that $w_{p}(\mu_{n},\delta_{x_{0}})=\epsilon$
converges to $0$ the sequence cannot converge strongly in $\mathcal{P}_{p}(X)$.
\end{example*}
Even though Wasserstein spaces are in general not locally compact
we can still show that the following holds for proper metric spaces.
The result is probably known or at least implicitly used in case $X=\mathbb{R}^{n}$.
Because it will be our main reason why the {}``weak'' Conley theory
is applicable and because we couldn't find any reference, we will
prove it completely.
\begin{thm}
\label{thm:weakly-proper}If $(X,d)$ is a proper metric space then
all closed $\delta$-neighborhoods of compact sets in $(\mathcal{P}_{1}(X),w_{1})$
are weakly compact, where the weak topology of $\mathcal{P}_{1}(X)$
is the induced subspace topology $\mathcal{P}_{1}(X)\subset\mathcal{P}(X)$.
A space, e.g. $\mathcal{P}_{1}(X)$, having this property may be called
weakly proper.\end{thm}
\begin{cor}
For $1\le p<q$ closed $\delta$-neighborhoods of compact sets in
$(\mathcal{P}_{q}(X),w_{q})$ are compact in $(\mathcal{P}_{p}(X),w_{p})$,
i.e. $\mathcal{P}_{q}(X)$ is weakly proper w.r.t. the subspace topology
induced by $\mathcal{P}_{q}(X)\subset\mathcal{P}_{p}(X)$. \end{cor}
\begin{rem*}
This is stronger then a compact embedding $i:(\mathcal{P}_{q}(X),w_{q})\to(\mathcal{P}_{p}(X),w_{p})$
because if $\mu_{n}\in\mathcal{P}_{q}(X)$ is bounded then w.l.o.g.
$i(\mu_{n})\to\mu_{*}$ in $\mathcal{P}_{p}(X)$ and necessarily $\mu_{*}\in\mathcal{P}_{q}(X)$,
i.e. bounded sequences never {}``leave'' the space.\end{rem*}
\begin{proof}
[Proof of theorem]We will show that \[
B_{r}^{w}:=\{\nu\in\mathcal{P}_{1}(X)\,|\, w_{1}(\nu,\delta_{x_{0}})\le r\}\]
is weakly compact for all $r\ge0$. Since $B_{r}^{w}$ is closed
and $\mu_{n}\rightharpoonup\mu$ implies $w_{1}(\mu_{n},\delta_{x_{0}})\le\liminf_{n\to\infty}w_{1}(\mu_{n},\delta_{x_{0}})\le r$
we only need to show that $B_{r}^{w}$ is tight.
Tightness of a subset $\mathcal{K}\subset\mathcal{P}_{1}(X)$ means
for all $\epsilon>0$ there is a compact $K_{\epsilon}$ such that
for all $\mu\in\mathcal{K}$\[
\mu(X\backslash K_{\epsilon})\le\epsilon.\]
For $\mu\in B_{r}^{w}$ we have\[
\int_{X}d(x,x_{0})d\mu=w_{1}(\mu,\delta_{x_{0}})\le r.\]
Now choose $K_{\epsilon}=B_{\frac{r}{\epsilon}}(x_{0})$, the closed
ball around $x_{0}$ with radius $\frac{r}{\epsilon}$, which is compact
because $X$ is proper. Then we have \[
\mu(X\backslash K_{\epsilon})=\int_{X\backslash K_{\epsilon}}d\mu(x)\le\frac{\epsilon}{r}\int_{X\backslash K_{\epsilon}}d(x,x_{0})d\mu(x)\le\epsilon.\]
Thus the closed ball in $\mathcal{P}_{1}(X)$ around $\delta_{x_{0}}$
is weakly compact.
Let $K\subset\mathcal{P}_{1}(X)$ be a compact set , e.g. $K=\{\mu\}$,
then the closed $R$-neighborhood around $K$ is defined as \[
N_{R}^{w}(K)=\{\nu\in\mathcal{P}_{1}(X)\,|\, w_{1}(\mu,\nu)\le R\,\mbox{ for some }\mu\in K\}.\]
This set is closed and bounded and for some $\tilde{R}$ we have
\[
N_{R}^{w}(K)\subset B_{\tilde{R}}^{w}.\]
Let $\nu_{n}\in N_{R}^{w}(K)$ be an arbitrary sequence. Then there
are $\mu_{n}\in K$ with $w_{1}(\mu_{n},\nu_{n})\le R$. Because $B_{\tilde{R}}^{w}$
is weakly compact and $K$ is compact there are $\nu_{\infty}\in B_{\tilde{R}}^{w}$
and $\mu_{\infty}\in K$ such that for some subsequence (also denoted
by $\mu_{n}$, resp. $\nu_{n}$) \begin{eqnarray*}
\mu_{n} & \to & \mu_{\infty}\\
\nu_{n} & \rightharpoonup & \nu_{\infty}.\end{eqnarray*}
Since $w_{1}(\cdot,\cdot)$ is weakly lower semicontinuity we have
\[
w_{1}(\mu,\nu)\le\liminf_{n\to\infty}w_{1}(\mu_{n},\nu_{n})\le R,\]
i.e. $\nu\in N_{R}^{w}(K)$ which implies weak compactness.
\end{proof}
\begin{proof}
[Proof of corollary] We only show that $B_{r}^{w_{q}}(\delta_{x_{0}})$
is weakly compact w.r.t. the induced subspace topology $\mathcal{P}_{q}(X)\subset\mathcal{P}_{p}(X)$
for $1\le p<q$. The rest will follow by the same arguments used above.
Assume $\{\mu_{n}\}_{n\in\mathbb{N}}\subset B_{r}^{w_{q}}(\delta_{x_{0}})$.
Since $w_{q}\le w_{1}$ the previous theorem implies w.l.o.g. $\mu_{n}\rightharpoonup\mu_{\infty}$
for some $\mu_{\infty}\in\mathcal{P}_{1}(X)$. Because \[
w_{q}(\mu_{\infty},\delta_{x_{0}})\le\liminf_{n\to\infty}w_{q}(\mu_{n},\delta_{x_{0}})\le r\]
we actually have $\mu_{\infty}\in B_{r}^{w_{q}}(\delta_{x_{0}})\subset\mathcal{P}_{q}(X)$.
Because $1\le p<q$ \[
\int_{X\backslash B_{R}}d(x,x_{0})^{p}d\mu_{n}(x)\le\frac{1}{R^{q-p}}\int_{X\backslash B_{R}}d(x,x_{0})^{q}d\mu_{n}(x)\le\frac{r}{R^{q-p}}.\]
Hence \[
\lim_{R\to\infty}\limsup_{n\to\infty}\int_{X\backslash B_{R}}d(x,x_{0})^{p}d\mu_{n}(x)\le\lim_{R\to\infty}\frac{r}{R^{q-p}}=0.\]
This and the $\mu_{n}\rightharpoonup\mu_{\infty}$ weakly show that
$\mu_{n}\to\mu_{\infty}$ in $\mathcal{P}_{p}(X)$ (see {\cite[6.8]{Villani2009}}).
\end{proof}
In the following assume that $X$ is proper. Suppose now $f:X\to X$
is a continuous map having a global (set) attractor, i.e. there is
a compact $f$-invariant $A\subset X$ such that for all bounded sets
$B$\[
\lim_{n\to\infty}\tilde{d}(f^{n}(B),A)=0,\]
where $\tilde{d}$ is the semi-Hausdorff metric induced by $d$ such
that $\tilde{d}(A,B)=0$ iff $A\subset\operatorname{cl}B$. The map
$f$ induces a continuous map $f_{*}:\mathcal{P}(X)\to\mathcal{P}(X)$
with $f_{*}(\mu)(B)=\mu(f^{-1}(B))$ for all Borel set $B$. Furthermore,
under slightly stronger assumptions $f_{*}$ has a global attractor
\[
\mathcal{P}(A)=\{\mu\in\mathcal{P}(X)\,|\,\mu(A)=1\}.\]
\begin{rem*}
For compact $X$ the global attractor is always $X$ itself. In particular,
since $\mathcal{P}_{p}(X)=\mathcal{P}(X)$ for $1\le p<\infty$ is
compact the global attractor of $\mathcal{P}(X)$ is the space itself
and we don't get new information. The whole theory is only interesting
for non-compact proper metric spaces $X$.
\end{rem*}
Since the Wasserstein space includes distance the behavior of $f$
at infinity becomes important.
\begin{thm}
\label{thm:wasserstein-cont}The map $f_{*}$ induces a continuous
map $(\mathcal{P}_{p}(X),w_{p})\to(\mathcal{P}_{p}(X),w_{p})$ (also
denoted by $f_{*}$) if and only if for some $x_{0}\in X$ \[
\sup_{x\in X}\frac{d(f(x),x_{0})}{1+d(x,x_{0})}<\infty.\]
\end{thm}
\begin{rem*}
For a semiflow $\pi$ we need that $\sup_{x\in X}\nicefrac{d(x\pi t,x_{0})}{1+d(x,x_{0})}=M_{t}<\infty$
for $t\in[0,T_{0}]$. Which means, in particular, that there has to
be a global lower bound on the blow-up time and thus there cannot
be blow-ups at all, i.e. $\pi$ has to be a global semiflow. Which
implies that the induced semiflow $\pi_{*}$ on $\mathcal{P}_{p}(X)$
is also a global semiflow. A necessary requirement for the existence
of a global attractor is an upper bound on $M_{t}$ for all $t\ge0$.
The requirements in {\cite[Chapter 8]{AmbGigSav2008}} are sometimes
too strong. A sufficient condition is that a one-sided Lipschitz condition
holds globally (e.g. $v(x)=x-x^{3}$ is an unbounded vector field
and only locally Lipschitz, but satisfies a one-sided Lipschitz condition).\end{rem*}
\begin{proof}
Suppose first that \[
M=\sup_{x\in X}\frac{d(f(x),x_{0})}{1+d(x,x_{0})}<\infty.\]
We can assume w.l.o.g. $M>0$, otherwise $f|_{X}\equiv x_{0}$ and
the result is obvious. For $\mu\in\mathcal{P}_{p}(X)$ we have \begin{eqnarray*}
\int d(x,x_{0})^{p}df_{*}\mu(x) & = & \int d(f(x),x_{0})^{p}d\mu(x)\\
& \le & M^{p}\int(1+d(x,x_{0}))^{p}d\mu(x)<\infty,\end{eqnarray*}
i.e. $f_{*}(\mu)\in\mathcal{P}_{p}(X)$. So we only need to show
continuity.
Suppose $\mu_{n}\to\mu$ in $\mathcal{P}_{p}(X)$ then $f_{*}\mu_{n}\rightharpoonup f_{*}\mu$.
Since $d(f(\cdot),x_{0})^{p}$ is continuous and grows at most like
$d(\cdot,x_{0})^{p}$ it follows that\begin{eqnarray*}
\int d(x,x_{0})^{p}df_{*}\mu_{n}(x) & = & \int d(f(x),x_{0})^{p}d\mu_{n}(x)\\
& & \longrightarrow\int d(f(x),x_{0})^{p}d\mu(x)=\int d(x,x_{0})^{p}df_{*}\mu(x).\end{eqnarray*}
Thus $f_{*}\mu_{n}\to f_{*}\mu$ (see {\cite[6.8]{Villani2009}})
which shows that $f_{*}:(\mathcal{P}_{p}(X),w_{p})\to(\mathcal{P}_{p}(X),w_{p})$
is (strongly) continuous.
It remains to show that $f_{*}$ is not continuous if there is a sequence
$\{x_{n}\}_{n\in\mathbb{N}}\subset X$ such that \[
d_{n}=\frac{d(f(x_{n}),x_{0})}{1+d(x_{n},x_{0})}\to\infty.\]
Because $X$ is proper and $f$ continuous we must have $d(x_{n},x_{0})\to\infty$.
For large $n$ we can assume $0<\frac{1}{d_{n}}<d(x_{n},x_{0})$.
Set $c_{n}=\frac{1}{d_{n}}$ then \[
\mu_{n}=c_{n}^{p}\frac{1}{d(x_{n},x_{0})^{p}}\delta_{x_{n}}+(1-c_{n}^{p}\frac{1}{d(x_{n},x_{0})^{p}})\delta_{x_{0}}\in\mathcal{P}_{1}(X)\]
and $w_{p}(\mu_{n},\delta_{x_{0}})^{p}=c_{n}^{p}\to0$, i.e. $\mu_{n}\to\delta_{x_{0}}$
strongly in $\mathcal{P}_{p}(X)$. We have \[
f_{*}\mu_{n}=c_{n}^{p}\frac{1}{d(x_{n},x_{0})^{p}}\delta_{f(x_{n})}+(1-c_{n}^{p}\frac{1}{d(x_{n},x_{0})^{p}})\delta_{f(x_{0})}\]
and therefore \[
w_{p}(f_{*}\mu_{n},f_{*}\delta_{x_{0}})^{p}=c_{n}^{p}\frac{d(f(x_{n}),x_{0})^{p}}{d(x_{n},x_{0})^{p}}=\frac{1+d(x_{n},x_{0})^{p}}{d(x_{n},x_{0})^{p}}\to1\]
which implies that $f_{*}$ cannot be (strongly) continuous in $\delta_{x_{0}}$.
\end{proof}
The following results will hold for any $\mathcal{P}_{p}(X)$. To
simplify the notation and some of the proofs we will just state them
for $\mathcal{P}_{1}(X)$. Furthermore, we assume from now on that
$f_{*}:\mathcal{P}_{1}(X)\to\mathcal{P}_{1}(X)$ is strongly continuous
(for short just continuous) and whenever we speak about $f_{*}$ we
mean the map $f_{*}:\mathcal{P}_{1}(X)\to\mathcal{P}_{1}(X)$. Since
$f_{*}(\mathcal{P}_{1}(X))\subset\mathcal{P}_{1}(X)$ (for $f_{*}:\mathcal{P}(X)\to\mathcal{P}(X)$)
this also implies that $f_{*}$ is weakly continuous in $\mathcal{P}_{1}(X)$.
Similarly we could say that $f_{*}:\mathcal{P}_{q}(X)\to\mathcal{P}_{q}(X)$
is continuous and {}``weakly'' continuous in $\mathcal{P}_{q}(X)$
w.r.t. the induced subspace topology of $\mathcal{P}_{q}(X)\subset\mathcal{P}_{p}(X)$
for any $1\le p<q$.
\begin{example*}
Having a global attractor does not imply that $f_{*}$ is strongly
continuous, even finite time compactness is not sufficient: Let $X$
be $\mathbb{R}$ with the Euclidean metric $|\cdot|$. Define $f:\mathbb{R}\to\mathbb{R}$
by \[
f(x)=\begin{cases}
0 & x\ge0\\
x^{2} & x<0.\end{cases}\]
Then $f$ is continuous and $f^{2}\equiv0$ but for $x_{n}=-n$ \[
\frac{d(f(x_{n}),0)}{1+d(x_{n},0)}=\frac{n^{2}}{1+n}\to\infty,\]
i.e. $f_{*}$ is not continuous on $(\mathcal{P}_{p}(X),w_{p})$.
\end{example*}
If $K\subset\mathcal{P}_{1}(X)$ is invariant w.r.t. $f_{*}$ then
it is invariant w.r.t. $f_{*}:\mathcal{P}(X)\to\mathcal{P}(X)$. Which
implies that all measures in $K$ are supported on the global attractor
$A$ of $f$. Since $\mathcal{P}(A)\subset\mathcal{P}_{1}(X)$ the
maximal invariant set of $\mathcal{P}_{1}(X)$ is $\mathcal{P}(A)=\mathcal{P}_{1}(A)$.
Suppose $f$ is finite time compact, i.e. there is an $m$ such that
$f^{m}(X)\subset B_{R}$ for some compact set $B_{R}$. It should
be obvious that this implies $K=\mathcal{P}_{1}(A)$ is the global
attractor of $f_{*}$. Furthermore, we have the following:
\begin{prop}
Suppose for some $m>0$, \[
f^{m}(X)\subset B_{R}(x_{0}).\]
Then $f_{*}$ is finite time compact and thus any closed set $B^{w}\subset\mathcal{P}_{1}(X)$
is $f_{*}$-admissible. \end{prop}
\begin{proof}
Let $\{\nu_{n}\}_{n\in\mathbb{N}}$ be any sequence in $f_{*}^{m}(\mathcal{P}_{1}(X))$.
Then there is a sequence $\{\mu_{n}\}_{n\in\mathbb{N}}$ such that
$\nu_{n}=f_{*}^{m}(\mu_{n})$. $f^{m}(X)\subset B_{R}(x_{0})$ implies
$\operatorname{supp}\nu_{n}\subset B_{R}(x_{0})$. Thus $\nu_{n}$
is tight and \[
\int_{d(x,x_{0})\ge R+\epsilon}d(x,x_{0})d\nu_{n}(x)=0,\]
i.e. $\{\nu_{n}\}_{n\in\mathbb{N}}$ has uniformly integrable first
moments. Which means that $\{\nu_{n}\}_{n\in\mathbb{N}}$ has a convergent
subsequence. Therefore, $f_{*}^{m}(\mathcal{P}_{1}(X))$ is compact
which easily implies admissibility for any closed $B^{w}$.\end{proof}
\begin{prop}
Suppose there is an $R_{0}$, $0\le c<1$ and $m>0$ such that for
all $R\ge R_{0}$\[
f^{m}(B_{R})\subset B_{cR}.\]
Then any bounded closed set $B^{w}\subset\mathcal{P}_{1}(X)$ is $f_{*}$-admissible.\end{prop}
\begin{proof}
Let $\{\mu_{n}\}_{n\in\mathbb{N}}$ be a sequence in $B^{w}$ such
that $f_{*}^{[0,m_{n}]}(\mu_{n})\subset B^{w}$ for some $m_{n}\to\infty$.
Because $B^{w}$ is bounded we have\[
\int_{X}d(x,x_{0})d\mu_{n}(x)\le M.\]
First assume $m_{n}=k_{n}\cdot m$ for an unbounded sequence $k_{n}\in\mathbb{N}$.
Then for $R\ge R_{0}$ \begin{eqnarray*}
\int\chi_{X\backslash B_{R}}(x)\cdot d(x,x_{0})df_{*}^{m_{n}}\mu_{n}(x) & = & \int\chi_{X\backslash B_{R}}(f^{m_{n}}(x))\cdot d(f^{m_{n}}(x),x_{0})d\mu_{n}(x)\\
& \le & \int\chi_{X\backslash B_{c^{-k_{n}}R}}(x)\cdot c^{k_{n}}d(x,x_{0})d\mu_{n}(x)\\
& \le & c^{k_{n}}M\to0,\end{eqnarray*}
which shows that $\{f_{*}^{m_{n}}(\mu_{n})\}_{n\in\mathbb{N}}$ has
uniformly integrable first moments which implies that the sequence
has a convergent subsequence.
If $m_{n}\not\equiv0(\operatorname{mod}m)$ then for some $0\le l_{n}<m$
we have $m_{n}-l_{n}\equiv0(\operatorname{mod}m)$. Therefore, if
we set $\nu_{n}=f_{*}^{l_{n}}(\mu)$ then the argument above applies
to $\nu_{n}$ and the sequence of endpoints (which is equal to $\{f_{*}^{m_{n}}(\mu_{n})\}_{n\in\mathbb{N}}$)
has a convergent subsequence.\end{proof}
\begin{rem*}
Kifer used in {\cite[Theorem 1.7]{Kifer1988}} linear attraction instead
of exponential. This might not be sufficient for admissibility. Nevertheless,
later we will assume that a Markov-type perturbation of $f_{*}$ is
small in the Wasserstein distance which is stronger than Kifer's assumption
and thus an invariant (probability) measure exists for the perturbation
by the same theorem. But that theorem does not imply that the perturbed
and unperturbed invariant measures are close w.r.t. the Wasserstein
distance, the perturbed invariant measures might not even be in the
Wasserstein space. So our result improves this sufficiently.
\end{rem*}
Before we show how to use Conley theory for small Markov-type noise
applied to $f$ we give a sufficient condition such that bounded sets
in $\mathcal{P}_{1}(X)$ are $\{F_{n}\}$-admissible for $F_{n}\to f_{*}$.
\begin{prop}
Let $B^{w}$ be closed and bounded and $U$ be a $\delta$-neighborhood
of $B^{w}$ with $f_{*}$-admissible closure. Suppose $F_{n}\to f_{*}$
uniformly on some $U$, i.e. \[
\sup_{\mu\in U}w_{1}(f_{*}(\mu),F_{n}(\mu))=\epsilon_{n}\to0.\]
If $f_{*}$ is uniformly continuous in $U$ then $B^{w}$ is $\{F_{n}\}$-admissible.\end{prop}
\begin{rem*}
(1) The idea is to use the uniform convergence and uniform continuity
to construct longer and longer orbits of $f_{*}$ close the the last
part of the orbits of $F_{n}$, i.e. the orbit $f^{[0,m_{n}-k_{n}]}(y_{n})$
and $F_{n}^{[k_{n},m_{n}]}(x_{n})$ should be closer and closer and
$m_{n}-k_{n}\to\infty$ for $y_{n}=F_{n}^{k_{n}}(x_{n})$.
(2) Uniform continuity of $f_{*}$ and uniform convergence of $F_{n}\to f_{*}$
are the assumptions Benci \cite{Benci1991} used to prove his continuation
theorem for the Conley index. Besides having continuous time dynamical
systems he also needs invertibility. \end{rem*}
\begin{proof}
Let $\mu_{n}\in N$ and $m_{n}\to\infty$ be sequences with $F_{n}^{[0,m_{n}]}(\mu_{n})$.
Uniform convergence of $F_{n}\to f_{*}$ and uniform continuity of
$f$ imply that for some $\epsilon(\epsilon_{n})\to0$ as $\epsilon_{n}\to0$
\begin{eqnarray*}
w_{1}(F_{n}^{2}(\mu),f_{*}^{2}(\mu)) & \le & w_{1}(F_{n}^{2}(\mu),f_{*}(F_{n}(\mu))+w_{1}(f_{*}(F_{n}(\mu)),f_{*}^{2}(\mu))\\
& \le & \epsilon_{n}+\epsilon(\epsilon_{n})=:\epsilon_{n,2}\to0\,\mbox{as \,}n\to\infty.\end{eqnarray*}
Similarly we can show that there are $\epsilon_{n,k}\to0$ as $n\to\infty$
such that \[
w_{1}(F_{n}^{k}(\mu),f_{*}^{k}(\mu))\le\epsilon_{n,k}\to0.\]
Therefore, there is a sequence $k_{n}\ge0$ with $m_{n}-k_{n}\to\infty$
such that $F_{n}^{[0,m_{n}]}(\mu_{n})\subset B^{w}$ implies that
\[
f_{*}^{[0,m_{n}-k_{n}]}(\nu_{n})\subset U=U_{\delta}(B^{w})\]
for $\nu_{n}=F_{n}^{k_{n}}(\mu_{n})$. Furthermore, we can choose
$k_{n}$ such that \[
\delta_{n}=\max_{k\in[0,m_{n}-k_{n}]}\epsilon_{n,k}\to0\]
and therefore \[
w_{1}(F_{n}^{m_{n}}(\mu_{n}),f_{*}^{m_{n}-k_{n}}(\nu_{n}))\le\delta_{n}.\]
Because the closure of $U$ is $f_{*}$-admissible, the sequence of
endpoints $\{f_{*}^{m_{n}-k_{n}}(\nu_{n})\}_{n\in\mathbb{N}}\subset U$
has a convergent subsequence which implies that $\{F_{n}^{m_{n}}(\mu_{n})\}_{n\in\mathbb{N}}$
has a convergent subsequence, in particular the limit point is in
$B^{w}$.
\end{proof}
This proposition applies in particular to Lipschitz continuous functions
$f:X\to X$ because the induce map $f_{*}:\mathcal{P}_{1}(X)\to\mathcal{P}_{1}(X)$
is Lipschitz continuous as well.
E.g. suppose $X=\mathbb{R}^{n}$ and $f$ is the time $h$ map of
a flow generated by an ODE $\dot{x}=g(x)$ such that $g$ satisfies
the one-sided Lipschitz condition for some $M\in\mathbb{R}$\[
\langle x-y,g(x)-g(y)\rangle\le M\|x-y\|^{2}\]
then $f$ is Lipschitz continuous with constant $e^{Mh}$.
\section{\label{sec:MW-maps}Markov-Wasserstein maps}
\begin{defn}
A Markov-Wasserstein map (MW-map) of order $p$ is a continuous map
$P:\mathcal{P}_{p}(X)\to\mathcal{P}_{p}(X)$ which is convex linear,
i.e. for $\mu,\nu\in\mathcal{P}_{p}(X)$ and $a\in[0,1]$\[
P(a\mu+(1-a)\nu)=aP(\mu)+(1-a)P(\nu).\]
Suppose $P$ is induced by a kernel $p(dy|x):X\to\mathcal{P}_{p}(X)$
(necessarily continuous), i.e. \[
P:d\mu(y)\mapsto\int p(dy|x)d\mu(x).\]
The map $P=P_{f}^{M}$ is called an MW-map of order $p$ relative
to $f$ with noise level (at most) $M$ if\[
\sup_{x\in X}w_{p}(p(dy|x),\delta_{f(x)})\le M.\]
This implies by {\cite[4.8]{Villani2009}} \[
w_{p}(P(\mu),f_{*}(\mu))^{p}\le\int w_{p}(p(dy|x),\delta_{f(x)})^{p}d\mu(x)\le M^{p}.\]
\end{defn}
\begin{rem*}
As in the sections before, the results also hold for semiflows and
a suitable definition for MW-semiflows, i.e. a continuous semigroups
$(P_{t})_{t\ge0}$ on $\mathcal{P}_{p}(X)$. The noise level model
can be stated similarly, but we only require that it is uniformly
small for all {}``small'' $t$. We will focus here only on maps,
resp. MW-maps, because the intuition behind these is easier, but all
results also hold for MW-semiflows if the noise level is sufficiently
small.
\end{rem*}
MW-maps (resp. MW-semiflows) appear naturally in the theory of Markov
chains (resp. processes). The map $p(dy|\cdot):X\to\mathcal{P}_{p}(X)$
is the Markov transition probability function, whereas the map $P:\mathcal{P}_{p}(X)\to\mathcal{P}_{p}(X)$
almost never has a name. If $P=f_{*}$ for some dynamical system $f:X\to X$
then $P$ is sometimes called transfer map. The following will show
that we only need continuity of $p(dy|\cdot)$ to ensure that $P$
is continuous in $\mathcal{P}_{p}(X)$ (and thus for any $\mathcal{P}_{q}(X)$,
$1\le q<p$ and for $\mathcal{P}(X)$).
\begin{thm}
Let $p(dy|\cdot):X\to\mathcal{P}_{p}(X)$ be a Markov kernel, i.e.
a measure-valued map, continuously depending on $x$. If $M=\sup_{x\in X}w_{p}(p(dy|x),\delta_{x})<\infty$
then $P$ defined by \[
P:d\mu(y)\mapsto\int p(dy|x)d\mu(x)\]
is an MW-map of order $p$ relative to $\operatorname{id}:X\to X$
with noise level $M$.\end{thm}
\begin{rem*}
For MW-semiflows weak continuity of $p_{t}(dy|\cdot)$ corresponds
to Feller continuity of the corresponding stochastic process. In fact,
if the initial distribution of $(X_{t}^{n})_{t\ge0}$ is $\delta_{x_{n}}$,
i.e. $X_{0}^{n}=x_{n}$, then by our continuity requirement if $x_{n}\to x_{0}$
then $\mu_{t}^{n}=P_{t}\delta_{x_{t}}\to P_{t}\delta_{x_{0}}=\mu_{t}^{0}$
which implies that \[
u(x_{n})=\mathbb{E}g(X_{t}^{n})=\int g(x)d\mu_{t}^{n}(x)\to\int g(x)d\mu_{t}^{0}(x)=\mathbb{E}g(X_{t}^{0})=u(x_{0})\]
for all bounded continuous function $g(x)$, i.e. the stochastic
process generated by $(P_{t})_{t\ge0}$ is Feller continuous. The
continuity of the moments implies that, in addition, the moments are
also continuous. This condition could be called $p$-Feller continuous.
This type of continuity does not need $t\in\mathbb{R}$ and thus applies
equally to Markov chains, i.e. discrete time stochastic processes.\end{rem*}
\begin{proof}
Continuous dependency implies that $P:\mathcal{P}_{p}(X)\to\mathcal{P}(X)$
is continuous. So we only need to show that $\mu_{n}\to\mu$ in $\mathcal{P}_{p}(X)$
implies that $P\mu_{n}\to P\mu$ in $\mathcal{P}_{p}(X)$
Because $w_{p}(p(dy|x),\delta_{x})\le M$ for some $M<\infty$ we
have\begin{eqnarray*}
\int d(y,x_{0})^{p}p(dy|x) & = & w_{p}(p(dy|x),\delta_{x_{0}})^{p}\\
& \le & (w_{p}(p(dy|x),\delta_{x})+w_{p}(\delta_{x},\delta_{x_{0}}))^{p}\\
& \le & 2^{p-1}(M^{p}+d(x,x_{0})^{p})\end{eqnarray*}
This implies that $g(x)=\int d(y,x_{0})^{p}p(dy|x)$ grows at most
like $d(x,x_{0})^{p}$. Furthermore, $g$ is continuous because $x\mapsto p(dy|x)$
and $\mu\mapsto w_{p}(\mu,\delta_{x_{0}})^{p}$ are. Thus by {\cite[6.8]{Villani2009}}
\[
\int d(y,x_{0})^{p}dP\mu_{n}(x)=\int g(x)d\mu_{n}(x)\to\int g(x)d\mu(x)=\int d(y,x_{0})^{p}dP\mu(y),\]
i.e. $P\mu_{n}\to P\mu$ in $\mathcal{P}_{p}(X)$.\end{proof}
\begin{example*}
(1) Bounded noise can be modeled by Markov maps with \[
M=\sup_{x\in X}d(x,\operatorname{supp}p(dy|x))<\infty.\]
Then $w_{p}(p(dy|x),\delta_{x})\le M$ and thus continuity of $p(dy|\cdot):X\to\mathcal{P}_{p}(X)$
for some $p$ implies that of $P:\mathcal{P}_{p}(X)\to\mathcal{P}_{p}(X)$.
This could also be used to model multi-valued perturbations, i.e.
maps $f_{n}:X\to2^{X}$ with $\sup_{x\in X}d(f(x),f_{n}(x))\le M$.
(2) Let $X=\mathbb{R}^{n}$ with its Euclidean distance. If $\nu$
is the standard normal distribution then $\nu=\rho(x)dx$, where $dx$
is the Lebesgue measure on $\mathbb{R}^{n}$. Any normal distribution
with mean $x$ and variance $\sigma^{2}$ can be modeled as follows
\[
\mu_{x,\sigma^{2}}=\delta_{x}*\rho_{\sigma}\]
where $\rho_{\epsilon}(x)=\epsilon^{-n}\rho(x/\epsilon)$. For $\sigma=0$
we set $\mu_{x,0}=\delta_{x}$.
Since $m_{p}=\int|x|^{p}\rho(x)dx<\infty$ for all $p$ this implies
(see {\cite[7.1.10]{AmbGigSav2008}}) that \[
w_{p}(\delta_{x},\mu_{x,\sigma^{2}})\le\sigma m_{p}.\]
Thus Gaussian noise with uniformly small variance is uniformly small
in all Wasserstein spaces (although the noise level diverges to $\infty$
as $p\to\infty$).\end{example*}
\begin{cor}
If $f:X\to X$ induces a continuous self map on $\mathcal{P}_{p}(X)$
and $p(dy|\cdot):X\to\mathcal{P}_{p}(X)$ is continuous with \[
M=\sup_{x\in X}w_{p}(p(dy|x),\delta_{f(x)})<\infty\]
then $P:\mathcal{P}_{p}(X)\to\mathcal{P}_{p}(X)$ defined as above
is an MW map of order $p$ relative to $f$ with noise level $M$.
\end{cor}
A random perturbation can now be modeled as a composition of $f$
followed by a smearing via $p(dy|\cdot)$, i.e. $f_{*}$ followed
by $P=P_{\operatorname{id}}^{M}$. This corresponds to additive noise
depending only on the image, whereas a general MW-map relative to
$f$ might smear the image $f(x)$ and $f(y)$ for $x\ne y$ differently
even if $f(x)=f(y)$.
If for some sequence $\tilde{p}_{n}(dy|\cdot)$ the noise level\[
\sup_{x\in X}w_{p}(\tilde{p}_{n}(dy|x),\delta_{f(x)})=\epsilon_{n}\]
converges to zero then $\tilde{P}_{n}$ converges to $f_{*}$ uniformly
on $\mathcal{P}_{p}(X)$, i.e.\[
\sup_{\mu\in\mathcal{P}_{p}(X)}w_{p}(\tilde{P}_{n}(\mu),f_{*}(\mu))\to0.\]
An MW-chain relative to $f$ satisfies the Markov property, i.e. future
behavior only depends on the current state. Furthermore, this models
only time-independent random perturbations. Time-dependent perturbations
can be modeled with the result of \cite{Kell2011b}. There it is shown
that a local attractor can be continued if the non-autonomous perturbations
is uniformly small. Translated into this framework this means \[
\sup_{\mu\in\mathcal{P}_{p}(X),k\in\mathbb{Z}}w_{p}(P(\mu,k),f_{*}(\mu))<\epsilon\]
for the non-autonomous dynamical system ($\approx$ inhomogeneous
Markov map)\[
(\mu,k)\mapsto(P(\mu,k),k+1).\]
Instead of using the semi-admissibility argument to show that the
invariant set $K_{n}$ is non-empty we can use a weak compactness
argument to get the same result.
\begin{example*}
Consider the ODE with $\dot{x}=x-x^{3}$. This satisfies the one-side
Lipschitz condition with $M=1$ and generates a global semiflow that
attracts in finite time. Thus the time-one map induces a Lipschitz
continuous map $f_{*}$ on $\mathcal{P}_{p}(X)$ which attracts in
finite time, too. Therefore, any bounded closed set in $\mathcal{P}_{p}(X)$
is $\{F_{n}\}$-admissible for $F_{n}\to f_{*}$ uniformly. In particular,
the MW-map of order $p$ for small noise level has an attractor close
to the original w.r.t. the Wasserstein metric $w_{p}$. \end{example*}
\begin{thm}
\label{thm:weak-cont-wasserstein}Suppose $f:X\to X$ induces a dynamical
system $f_{*}$ on $\mathcal{P}_{p}(X)$ having a global attractor
and that $f_{*}$ is uniformly continuous in a neighborhood of the
global attractor. If $P_{n}\to f_{*}$ is a sequence of MW-maps of
order $p$ relative to $f$ with noise level $\epsilon_{n}\to0$.
Then for $n\ge n_{0}$ there is a positive $P_{n}$-invariant isolating
neighborhood $N_{n}$ such that $K_{n}=A_{P_{n}}(N_{n})$ is non-empty
and a weakly compact weak attractor which contains all bounded $P_{n}$-invariant
measures, i.e. $A_{P_{n}}(P_{p}(X))=K_{n}$. Furthermore, there is
at least one stationary measure in $K_{n}$.\end{thm}
\begin{rem*}
Suppose $p>1$. Whenever $K$ is (strongly) compact in $\mathcal{P}_{p}(X)$
then $\operatorname{cl}U_{\delta}(K)$ is weakly compact w.r.t. the
weaker subspace topology of $\mathcal{P}_{p}(X)$ induced by $\mathcal{P}_{p}(X)\subset\mathcal{P}_{q}(X)$
for any $1\le q<p$. Thus for all $\mu\in N_{n}$ there is a $\mu_{K}\in K_{n}$
\[
P_{n}(\mu)\overset{w_{q}}{\longrightarrow}\mu_{K}.\]
This means that, although the $p$-moment may not converge, any $q$-moment
converges for $1\le q<p$, but the convergence may get worse the closer
$q$ comes to $p$.\end{rem*}
\begin{proof}
Everything but the existence of a stationary measure and $A_{P_{n}}(\mathcal{P}_{p}(X))=K_{n}$
follows from theorem \ref{thm:weak-continuation}. Since $P_{n}$
is convex linear $K_{n}$ must be convex. This implies that for any
$\mu\in K_{n}$ the sequence \[
\left\{ \frac{1}{m}\sum_{k=0}^{m-1}P_{n}^{k}(\mu)\right\} _{m\in\mathbb{N}}\]
is in $K_{n}$ and thus weakly converging to some $\nu\in K_{n}$
and by the Krylov-Bogolyubov theorem it must be a fixed point of $P_{n}$,
i.e. $\nu$ is a stationary measure of $P_{n}$. Furthermore, if $R$
is the distance from the global attractor of $f$ then its mass must
decay as $R^{-p}$.
The invariant measures must all be contained in the interior of $N_{n}$.
Otherwise take $\mu\in A_{P_{n}}(\mathcal{P}_{p}(X))\backslash K_{n}$.
If $\mu$ is $P_{n}$-stationary then the argument is as follows:
For $t\in[0,1]$ and some $\mu_{0}\in K_{n}$ the graph of $t\mapsto t\mu+(1-t)\mu_{0}$
is stationary and intersects $\partial N_{n}$, which implies $K_{n}=A_{P_{n}}(N_{n})$
intersects $\partial N_{n}$. This contradicts the isolatedness of
$K_{n}$.
For the general case assume $\sigma:\mathbb{Z}\to\mathcal{P}_{p}(X)$
is a bounded full solution through $\mu$, i.e. $P_{n}(\sigma(k))=\sigma(k+1)$
and $\sigma(0)=\mu$. Now define the function $g:\mathbb{Z}\to[0,1]$
\[
g:k\mapsto\sup\{t\in[0,1]\,|\, s\sigma(k)+(1-s)\mu_{0}\in N_{n}^{'}\,\mbox{for all}\, s\in[0,t]\}.\]
Because $N_{n}^{'}$ is a positive invariant neighborhood of $K_{n}$,
$\mu_{0}$ stationary, $P_{n}$ convex linear and $\{\sigma(k)\}_{k\in\mathbb{Z}}$
bounded and not entirely in $N_{n}^{'}$ we have \[
0<\delta\le g(k)\le g(k+1).\]
This implies that \[
T=\inf_{k\le0}g(k)\ge\delta.\]
Because $\sigma(0)\notin N_{n}^{'}$ we have $T<1$ and thus by definition
of $g$ \[
w_{p}(T\sigma(k)+(1-T)\mu_{0},\partial N_{n}^{'})\to0\quad\mbox{as\,}k\to-\infty\]
and thus there is a $T_{1}\le T$ such that $\tilde{\sigma}(k)=T_{1}\sigma(k)+(1-T_{1})\mu_{0}$
is a full solution in $N_{n}^{'}$ with \[
w_{p}(\tilde{\sigma}(k_{1}),\partial N_{n}^{'})\le\frac{\epsilon}{2}\]
for some $k_{1}\le0$. But $\tilde{\sigma}(k_{1})\in K_{n}$ and $U_{\epsilon}(K_{n})\subset N_{n}^{'}$
which implies that \[
\epsilon\le w_{p}(\sigma(k_{1}),\partial N_{n}^{'})\le\frac{\epsilon}{2}.\]
This is a contradiction and thus $A_{P_{n}}(\mathcal{P}_{p}(X))\backslash K_{n}=\varnothing$,
i.e. $K_{n}$ contains all bounded invariant measures in $\mathcal{P}_{p}(X)$.\end{proof}
\begin{cor}
The positive $f_{*}$-invariant isolating neighborhood $B$ and the
positive $P_{n}$-invariant isolating neighborhood $N_{n}$ can be
chosen convex, i.e. if $\mu_{i}\in B$ (resp. $\mu_{i}\in N_{n}$)
for $i=0,1$ then $\mu_{t}\in B$ (resp. $\mu_{t}\in N_{n}$) for
$\mu_{t}=t\mu_{0}+(1-t)\mu_{1}$ and $t\in[0,1]$.\end{cor}
\begin{proof}
Let $\nu_{i},\mu_{i}\in\mathcal{P}_{p}(X)$ for $i=0,1$ and define
$\mu_{t}=t\mu_{1}+(1-t)\mu_{0}$ and $\nu_{t}=t\nu_{1}+(1-t)\nu_{0}$.
Assume \[
w_{p}(\mu_{i},\nu_{i})<\epsilon.\]
Then there are optimal transference plans $\pi_{i}\in\Pi(\mu_{i},\nu)$
such that\[
\int d(x,y)^{p}d\pi_{i}(x,y)<\epsilon^{p}.\]
The plan $\pi_{t}=t\pi_{1}+(1-t)\pi_{0}$ is a transference plan for
the pair $(\mu_{t},\nu_{t})$ and thus\begin{eqnarray*}
w_{p}(\mu_{t},\nu_{t})^{p} & \le & \int d(x,y)^{p}d\pi_{t}(x,y)\\
& \le & t\int d(x,y)^{p}d\pi_{1}(x,y)+(1-t)\int d(x,y)^{p}d\pi_{0}(x,y)\\
& < & t\epsilon^{p}+(1-t)\epsilon^{p}=\epsilon^{p}.\end{eqnarray*}
The construction of $B$ is done via a Lyapunov pair $(\phi,\gamma)$
essentially measuring a weighted distance of the forward orbit of
a point, i.e. \[
F_{N'}:\mu\mapsto\min\{1,w_{p}(\mu,A^{-}(N')\cup\partial N')\}\]
and\[
\phi:\mu\mapsto\sup\{(2n+1)F_{N'}(f_{*}^{n}(x))/(n+1)\,|\, n\in\mathbb{N},n\le\omega_{N'}(x)\}\]
Let $P_{1}^{\epsilon}$ be defined as in the proof of theorem \ref{thm:stable}.
We can assume that $w_{p}(P_{1}^{\epsilon},\partial N')>2\epsilon$
for some for sufficiently small $\epsilon>0$. Because $A^{-}(N')=A(N')$
is convex, $\phi(\mu_{i})<\epsilon$ for $i=0,1$ implies \[
\phi(\mu_{t})<\epsilon.\]
Hence $P_{1}^{\epsilon}$ is convex and we can choose $B=P_{1}^{\delta}$
for some small $\delta>0$. Similarly $V(a)$ defined in theorem \ref{thm:cont}
is convex.
The set $N_{n}$ was defined as \[
N_{n}(\epsilon)=N\cap\operatorname{cl}\{y\,|\,\mbox{ \ensuremath{\exists x\in V(\epsilon)}, \ensuremath{m\ge0}\,\ s.t.\,\ \ensuremath{P_{n}^{[0,m]}(x)\subset\tilde{U}\,}and \ensuremath{P_{n}^{m}(x)=y}}\}\]
where $\tilde{U}=\operatorname{int}B$ and $N=\operatorname{cl}V(\epsilon_{0})$
are convex sets. Hence $N_{n}(\epsilon)$ is convex .\end{proof}
\begin{cor}
Under the assumption of the previous theorem if $p=1$ then all orbits
of $P_{n}$ are bounded for $n\ge n_{0}$, i.e. $P_{n}^{[0,\infty]}(\mu)\subset B_{R}^{w}(\mu_{0})$
for some $R\ge0$ and some fixed $\mu_{0}$. In particular, $K_{n}$
is the global weak attractor of $P_{n}$.\end{cor}
\begin{rem*}
The idea of the proof is to control the distance of $\mu_{1}$ and
$\mu_{0}$ by the distance of $\mu_{t}$ and $\mu_{1}$ where $\mu_{0}$
will be some stationary measure and $t\in(0,1]$ is sufficiently small.\end{rem*}
\begin{proof}
Using the Kantorovich-Rubinstein formula we have the following equality
for $\mu,\nu\in\mathcal{P}_{1}(X)$\[
w_{1}(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)}\int d(x,y)d\pi(x,y)=\sup_{\|\phi\|_{\operatorname{Lip}}\le1}\left\{ \int\phi d\mu-\int\phi d\nu\right\} ,\]
i.e. there is a sequence $\phi_{k}$ with $\|\phi_{k}\|_{\operatorname{Lip}}\le1$
such that $\int\phi_{k}d\mu-\int\phi_{k}d\nu\nearrow w_{1}(\mu,\nu)$.
Furthermore, there exist an optimal plan $\pi\in\Pi(\mu,\nu)$ such
that the infimum is actually attained.
Choose $\mu_{0}\in\mathcal{P}_{1}(X)$ and define $\mu_{t}=t\mu_{1}+(1-t)\mu_{0}$
for $t\in[0,1]$ and $\mu_{1}\in\mathcal{P}_{1}(X)$. We claim \[
w_{1}(\mu_{t},\mu_{0})=tw_{1}(\mu_{t},\mu_{0}).\]
Suppose $\pi\in\Pi(\mu_{1},\mu_{0})$ is the optimal plan and $\phi_{k}$
the sequence of Lipschitz maps as above. Then $\tilde{\pi}=t\pi+(1-t)(\operatorname{id},\operatorname{id})_{*}\mu_{0}$
is in $\Pi(\mu_{t},\mu_{0})$. Thus\[
w_{1}(\mu_{t},\mu_{0})\le\int d(x,y)d\tilde{\pi}(x,y)=tw_{1}(\mu_{1},\mu_{0}).\]
Furthermore, we have \[
w_{1}(\mu_{t},\mu_{0})\ge\int\phi_{k}d\mu_{t}-\int\phi_{k}d\mu_{0}=t\left(\int\phi_{k}d\mu_{1}-\int\phi_{k}d\mu_{0}\right).\]
Because the left hand side converges monotonically to $tw_{1}(\mu_{t},\mu_{0})$
we have proved our claim.
Now fix some stationary measure $\mu_{0}\in K_{n}$. Since $N_{n}$
is a neighborhood of $K_{n}$ there is a $t\in(0,1]$ for all $\mu_{1}\in X$
such that $\mu_{t}$ as defined above is in $N_{n}$. Thus $P_{n}^{[0,\infty)}(\mu_{t})$
is in $N_{n}$ and bounded, i.e. $w_{1}(P_{n}^{k}(\mu_{t}),\mu_{0})\le R$
for some $R$. Because $P_{n}$ is convex linear and $\mu_{0}$ stationary
we have $P_{n}^{k}(\mu_{t})=tP_{n}^{k}(\mu_{1})+(1-t)\mu_{0}$ and
hence \begin{eqnarray*}
w_{1}(P_{n}^{k}(\mu_{1}),\mu_{0}) & = & \frac{1}{t}w_{1}(P_{n}^{k}(\mu_{t}),\mu_{0})\le\frac{R}{t},\end{eqnarray*}
which implies that the orbit of $\mu_{1}$ is bounded.
\end{proof}
\bibliographystyle{amsalpha}
|
\section{Introduction}
\label{sec:intro}
\setcounter{footnote}{0}
\setcounter{section}{1}
\setcounter{subsection}{0}
Massive stars explode via core collapse and ejection of their
surrounding layers \citep[e.g.][and references therein]{arn89}. The
extent to which core-collapse supernovae (CCSNe) are, or have been, a
major source of dust in the Universe is of great interest. Of
particular concern is the evidence of enormous amounts of dust
($\gtrsim 10^{8}\,{\rm M}_\odot$) in galaxies at high redshifts
($z\gtrsim 5$). This comes from a variety of observations such as
sub-mm and near-infrared (NIR) studies of the most distant quasars
\citep{ber03,maio04}, obscuration by dust of quasars in damped
Ly-$\alpha$ systems \citep{pei91}, and measurements of metal
abundances in these systems \citep{pet97}. Until recently, the
scenario of dust from AGB stars tended to be rejected since it was
thought that their progenitors would not yet have evolved off the main
sequence. However, \citet{val09} and \citet{dwe11} have argued that,
under certain circumstances, asymptotic giant branch (AGB) stars may
make some contribution to the dust budget at high redshifts. Both
studies nevertheless cannot rule out a supernova contribution. In
this paper, we examine the supernova option through observations of a
nearby core-collapse event.\\
CCSNe arising from short-lived Population III stars might seem to be a
viable alternative. It is estimated that each supernova (SN) must
produce 0.1--1\,${\rm M}_\odot$ of dust to account for the
high-redshift observations \citep{dwe07,mei07}. Such masses have been
predicted in models of dust formation in CCSNe \citep{tod01,noz03},
although more recent calculations by \citet{che10} revise such
estimates downward by a factor of $\sim$5. Perhaps even more
problematic is the fact that actual dust-mass measurements in CCSNe
and SN remnants yield values not exceeding, respectively,
$10^{-3}\,{\rm M}_\odot$ and $10^{-2}\,{\rm M}_\odot$, although only a
handful of such measurements exist. (For a summary of this topic see,
for example, \citealt{kot09}, \S1.) The {\it Spitzer Space Telescope}
\citep[hereafter, {\it Spitzer};][]{wer04} provided an excellent
opportunity for us to test the ubiquity of dust condensation in a
larger number of CCSNe.\\
Newly-condensed dust in CCSNe can be detected by its attenuating
effects on optical/NIR light and/or via thermal emission from the
grains in the ejecta. Prior to {\it Spitzer}, the only evidence of
dust condensation in {\it typical} CCSNe was in the Type II-plateau
(IIP) SN~1988H \citep{tur93} and SN~1999em \citep{elm03}. However the
light curve data used to type SN~1988H was sparse. In the case of
SN~1999em, \citeauthor{elm03} used optical line suppression to infer a
dust mass lower limit of about $10^{-4}$~M$_{\odot}$. With the launch
of {\it Spitzer}, we were at last provided with a facility for
high-sensitivity spectroscopy and imaging of nearby CCSNe over the
mid-infrared (MIR) range, covering the likely peak of the dust thermal
emission spectrum. This can provide a superior measure of the total
flux, temperature, and possibly dust emissivity than can be achieved
at shorter wavelengths. Moreover, the longer-wavelength coverage of
{\it Spitzer} allow us to detect cooler grains and see more deeply
into dust clumps than was previously possible for typical nearby
CCSNe. In this paper, we present our late-time {\it Spitzer}
observations of the Type~IIP SN~2004dj. We use these observations to
study the dust production in this supernova.\\
The paper is arranged as follows. In \S1.1 we summarize and discuss
previous observations of SN~2004dj. In \S2 we present MIR ({\it
Spitzer}) photometric and spectroscopic observations of SN~2004dj,
extending to more than 3~years after the explosion. This MIR coverage
is one of the most extensive ever achieved for a SN~IIP. We also
present late-time optical and NIR photometry and spectra of
SN~2004dj. In \S3 we analyze these data. Corrections are derived in
\S3.1 for the effects of the line-of-sight cluster S96, and in \S3.2
the mass of ejected $^{56}$Ni is determined. We compare the data with
blackbodies in \S3.3, in order to assess the likely number and nature
of the contributing sources. In \S3.4 the origins of the IR radiation
are examined in detail. The work is then summarized in \S4. \\ \\ \\
\subsection{SN~2004dj}
\label{sec:04dj}
SN~2004dj was discovered in the nearby spiral galaxy NGC~2403 on 2004
July 31 by \citet{nak04} and was classified as a normal Type~IIP SN by
\citet{pat04}. It was the nearest such event in over three decades;
the host galaxy lies within the M81 group. In \citet{kot05} we adopted
a distance to NGC~2403 (and the SN) of $3.13\pm0.15$~Mpc (statistical
errors only), this being the Cepheid-derived, zero metallicity value
reported by \citet{fre01} using the \citet{uda99} period-luminosity
slopes. We continue to use this distance in the work presented here.
\citet{vin06} have estimated the distance to SN~2004dj using a
combination of the Freedman et al. value plus their own expanding
photosphere method (EPM) and standard candle estimates. This yields
an average distance of $3.47\pm0.29$~Mpc, implying that the SN
luminosity could be $\sim$20\% larger than the values used herein.\\
The progenitor of SN~2004dj was almost certainly a member of the
compact star cluster Sandage~96 (S96)
\citep{san84,bon04,fil04,mai04,wan05,vin06}. The cluster age is
variously estimated to be $14\pm2$~Myr \citep{mai04}, $\sim$20~Myr
\citep{wan05}, and 10--16~Myr \citep{vin09}. The main-sequence mass
of the progenitor is estimated at $\sim$15~M$_{\odot}$
\citep{mai04,kot05}, $\sim$12~M$_{\odot}$ \citep{wan05}, and
12--20~M$_{\odot}$ \citep{vin09}. \citet{mai04} and \citet{kot05}
favor a red supergiant (RSG) progenitor.\\
\citet{gue04} used echelle spectroscopy of Na~I~D absorption lines in
NGC~2403 along the line of sight to SN~2004dj to infer a heliocentric
velocity of $+164.8\pm0.1$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. This is somewhat larger than the
heliocentric velocity of the nuclear region of NGC~2403 of
$\sim$130~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ given in the SIMBAD and NED databases, but this is not
surprising given the likely dispersion of velocities within the host
galaxy. Indeed, \citet{vin06} point out that H~I mapping of NGC~2403
\citep{fra01} suggests that the true radial velocity of the SN~2004dj
barycenter is about +221~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. We adopt this value here. \\
There is no firm consensus about the value of the reddening to
SN~2004dj. Stellar population fitting for S96 yields total (Galactic +
host) $E(B-V)$ values of $0.17\pm0.02$~mag \citep{mai04},
$0.35\pm0.05$~mag \citep{wan05}, and $0.1\pm0.05$~mag \citep{vin09}.
Direct color comparisons of SN~2004dj with other CCSNe yield $E(B-V)$
values of $\sim$0~mag \citep{zha06}, 0.06~mag \citep{chu06}, and
$0.07\pm0.1$~mag \citep{vin06}. Perhaps most significantly, the
resolved Na~I~D observations of \citet{gue04} yield a host-only $E(B-V)$
value of just $0.026\pm0.002$~mag. The smaller values of $E(B-V)$
obtained using direct measurements toward SN~2004dj suggests that the
SN actually lies near the front of S96. Based on the
extinction maps of \citet{sch98}, \citet{chu05} find a Galactic
reddening of $E(B-V)=0.062$~mag while \citet{zha06} report
$E(B-V)=0.04$~mag (the same value as obtainable from NED). We
therefore adopt a Galactic $E(B-V)=0.05\pm0.01$~mag. If we add the
\citet{gue04} host value to the Galactic value, we obtain
$E(B-V)=0.076\pm0.01$~mag, or $A_V=0.24\pm0.03$~mag for a
\citet{car89} extinction law with $R_V = 3.1$. Given the range of
published values, in the present work we used a total extinction
$A_V=0.31$~mag, the same as that preferred by \citet{vin09}. Adoption
of even the largest published value of $E(B-V)$ would increase our
shortest MIR wavelength (3.6~$\mu$m) flux by just a few percent. \\
Estimates of the explosion date of SN~2004dj vary by several weeks.
On the basis of an early-time spectrum, \citet{pat04} placed the explosion
at about 2004 July 14. This is consistent with the date obtained by
\citet{bes05}, who used the radio $L_{peak}$ versus rise-time relation
\citep{wei02} to yield an explosion date between 2004 July 11 and July
31. On the other hand, based on the EPM method, \citet{vin06} derive
an explosion date as early as 2004 June 30. On the assumptions that
the light curve of SN~1999gi was typical of SNe~IIP and that
the SN~2004dj plateau was of similar length, \citet{chu05} obtained an
even earlier explosion date: 2004 June 13. Likewise, assuming similar
evolution between SN~2004dj and SN~1999em, \citet{zha06} find an
explosion date of 2004 June 11. Nevertheless, partly as a compromise
with the later \citet{pat04} value, \citet{chu06} subsequently adopted
2004 June 28 as the date of the explosion. \citet{chu07} also used
this explosion date. Given the weight of evidence for a later
explosion date, we reject that preferred by \citet{zha06}. In
\citet{kot05} we adopted an explosion date of 2004 July 10, or
MJD = 53196.0. We use the same explosion date here but recognize that
there is an uncertainty of about $\pm7$~d. All epochs will be with
respect to MJD = 53196.0 (i.e., $t=0$~d). \\
Optical light curves of SN~2004dj are presented by
\citet{kor05,chu05,leo06, vin06,zha06,vin09}. The $VR$ light curves
fell by 10\% and 90\% of the total decline from the plateau to the
start of the radioactive tail at, respectively, epochs $70 \pm 5$~d
and $96 \pm 3$~d. At the start of the radioactive tail the SN
luminosity was about 15\% of the plateau value. During the early
nebular phase (up to $\sim$300~d) the $V$~band declined at about 1.1
mag (100~d)$^{-1}$ \citep{vin06}, which is typical for a SN~IIP. On
the basis of the light curves, the mass of ejected $^{56}$Ni has been
estimated at 0.02--0.03~M$_\odot$ \citep{chu05,kot05,vin06,zha06}.
Notwithstanding, in the present work we argue that these authors have
overestimated the $^{56}$Ni mass and that the true mass is more like
0.01~M$_\odot$. \\
At the end of the plateau phase, SN~2004dj exhibited a remarkable and
rapid change in some of its prominent optical lines, especially
H$\alpha$. At 89~d the H$\alpha$ profile still had a typical P~Cygni
morphology with a symmetric peak blueshifted by only about
$-270$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ after correction for the heliocentric velocity of the SN.
Yet, by just $\sim$10~d later the profile had developed a strong
asymmetric profile with a peak at $-1610$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ \citep{chu05}. As time
went by during the first year, this blueshift gradually decreased and
the asymmetry became less pronounced. \citet{chu06} interpreted this
unusual behavior as being due to the gradual emergence of an
asymmetric, bipolar jet whose more massive component is travelling
towards the observer. They propose that the lines are driven by the
radioactive decay of spherical fragments of $^{56}$Co cocooned in
metals and helium, lying within the core. Using spectropolarimetry
\citet{leo06} also found evidence for departure from spherical
symmetry. The polarization was observed to increase dramatically at
the end of the plateau phase, implying the presence of significant
asphericity in the inner regions of the ejecta. \\
Early-time evidence of a significant circumstellar medium (CSM) around
the progenitor of SN~2004dj has also been reported. Radio emission
was detected by \citet{sto04} at 23~d, by \citet{bes05} between 26~d
and 145~d, and by \citet{cha04} at 33~d and 43~d. The SN was also
detected at X-ray wavelengths \citep{poo04} at 30~d. The X-ray
luminosity was about three times that of the Type~IIP SN 1999em and
nearly fifteen times that of the Type~IIP SN 1999gi at similar epochs.
\citet{bes05} point out that both types of emission arise from a
shocked CSM. \citet{chu07} have used the presence of a high-velocity
absorption component in the H$\alpha$ line during the late
photospheric phase to deduce the presence of a cool dense shell (CDS),
with a mass of $3.2\times10^{-4}$~M$_\odot$, produced by interaction
of the ejecta with the pre-existing CSM. To this evidence for a CDS
we add our observation of an early-time IR echo in SN~2004dj (see
\S3.4.2). \\
In \citet{kot05} we presented MIR photometric and spectroscopic
observations of SN~2004dj at epochs 97--137~d after explosion. Simultaneous
modelling of the fundamental (1--0) and first overtone (2--0) of CO
was carried out. The results favor a 15~M$_\odot$ RSG progenitor and
indicate post-explosion CO formation in the range 2000--4000~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi.
\citet{kot05} also noted an underlying NIR continuum. A possible
origin in CSM dust was suggested, but RSGs in S96 were favored as the
more likely cause. Nevertheless, in the present work (\S3.2.2) we find
that the bulk of the early-time NIR and MIR continuum is most
plausibly explained as an IR echo from CDS dust. The presence of
$1.7\times10^{-4}$~M$_\odot$ of Ni$^+$ in the ejecta was also deduced
by \citet{kot05}. \\
In summary, while the early-time optical light curves and spectra of
SN~2004dj are typical of a Type~IIP event, its early radio, MIR, and
X-ray behavior point to an exceptionally strong ejecta/CSM
interaction. Moreover its earlier nebular-phase spectra imply an
atypically asymmetric core. SN~2004dj is only ``typical'' in some
respects. \\
\section{Observations}
\subsection{Mid-Infrared Photometry}
Imaging at 3.6, 4.5, 5.8, and 8.0\,$\mu$m was obtained with the
Infrared Array Camera (IRAC) \citep{faz04}, at 16 and 22\,$\mu$m with
the Peak-up Array (PUI) of the Infrared Spectrograph (IRS)
\citep{hou04}, and at 24\,$\mu$m with the Multiband Imaging Photometer
for Spitzer (MIPS) \citep{rie04}. Imaging observations spanned epochs
89.1 to 1393.3~d, plus four observations, at 3.6~$\mu$m only, covering
1953.9--2143.5~d. Aperture photometry was performed on the images
using the Starlink package GAIA \citep{dra02}. A circular aperture of
radius $\sim3\farcs7$ was used for the photometry. The background
flux was measured and subtracted by using a concentric sky annulus
having inner and outer radii of 1.5 and 2.2 times the aperture radius,
respectively. These parameters were chosen as a compromise between
maximizing the sampled fraction of source flux and minimizing the
effects of the bright, complex background. The aperture radius
corresponds to $\sim$55~pc at the distance of SN~2004dj. The aperture
was centered according to the SN WCS coordinates. Aperture
corrections were derived from the IRAC and MIPS point-response
function frames available from the Spitzer Science Center, and ranged
from $\times$1.16 at 3.6\,$\mu$m to $\times$2.79 at 24\,$\mu$m. A
2$\sigma$ clipped mean sky estimator was used, and the statistical
error was estimated from the variance within the sky annuli. Fluxing
errors due to uncertainties in the aperture corrections are about
$\pm5\%$. \\
The MIR photometry is presented in Table~\ref{tab1}. The {\it
Spitzer} programs from which the imaging data were taken are listed in
the caption. The pre-explosion MIR flux of S96 has not been measured,
so the tabulated values are uncorrected for S96. An estimate of the
S96 contribution is given at the bottom of the Table. The effect of
emission from S96 is discussed in \S3.1. A temporally varying point
source at the SN position is clearly visible in all bands.
Figure~\ref{fig1} shows a sequence of images at 8.0~$\mu$m at 257,
621, and 996~d. The SN is clearly brighter at 621~d (about $\times2$
relative to 257~d). This is due to the epoch being close to the peak
of the thermal emission from the dust (see \S3.4.4.2). It can also be
seen that the SN lies in a region of relatively bright, complex
background emission. The MIR photometry is displayed as light curves
and spectral energy distributions (SEDs) in Fig.~\ref{fig2} and
Fig.~\ref{fig3}, respectively. \\
In Fig.~\ref{fig2}, a rapid initial decline is seen at wavelengths
3.6--8.0~$\mu$m. With the exception of 4.5~$\mu$m, all of the light
curves (with sufficient temporal coverage) exhibit secondary maxima,
with the peak emission occurring at epochs of $\sim$450~d at
3.6~$\mu$m to $\sim$850~d at 24~$\mu$m. As will be discussed later,
these second maxima constitute strong evidence of dust formation.
The absence of a delayed peak at 4.5~$\mu$m is due to the earlier
appearance and dominance of CO fundamental emission in this
band. After $\sim$1000~d, the 24~$\mu$m light curve starts to climb
again. The evolution of the 24~$\mu$m flux is complex, as it is a
combination of the detailed behaviors of emission from the ejecta
dust and from an interstellar (IS) IR echo (see \S3.4.3.1). In
Fig.~\ref{fig3} we see a steady reddening of the MIR SED with time.
It is argued below (\S3.4.4.2) that this effect also constitutes strong
evidence of dust formation and cooling in the SN ejecta. The
large peak at 4.5~$\mu$m at 114~d and 257~d is due to the
aforementioned dominance of CO fundamental emission in this band. As
in Table~\ref{tab1}, neither Fig.~\ref{fig2} nor Fig.~\ref{fig3} have
been corrected for S96.
\subsection{Optical and Near-Infrared Photometry}
\label{sec:nirphot}
NIR imaging of SN~2004dj was obtained using LIRIS (Long-slit
Intermediate Resolution Infrared Spectrograph) on the 4.2~m William
Herschel Telescope (WHT), La Palma, and at an effectively single epoch
(spanning two days) with the OSU-MDM IR Imager/Spectrograph on the
2.4~m Hiltner Telescope of the MDM Observatory, Arizona. The wavebands
are $Z$ (1.033~$\mu$m), $J$ (1.250~$\mu$m), $H$ (1.635~$\mu$m), and
$K_s$ (2.15~$\mu$m). The LIRIS data were reduced using standard IRAF
routines.\footnote{IRAF is written and supported by the IRAF
programming group at the National Optical Astronomy Observatories
(NOAO) in Tucson, Arizona, which are operated by the Association of
Universities for Research in Astronomy (AURA), Inc., under cooperative
agreement with the National Science Foundation (NSF).} The jittered
on-source exposures were median-combined to form sky frames. In each
band the sky-subtracted frames were then aligned and
median-combined. \\
Aperture photometry was performed on the reduced
images using the Starlink package GAIA \citep{dra02} with the same
aperture and sky annuli as for the MIR photometry. The aperture was
centered by centroiding on the sources. The magnitudes at $J$, $H$, and
$K_s$ were obtained by comparison with four field stars lying within
$\sim100^{\prime\prime}$ of SN~2004dj. The field-star magnitudes were
acquired by measurement of 2MASS images \citep{skr97}. For the single
$Z$-band measurement, the magnitude was obtained by comparison with the
four field stars with their $JHK_s$ SEDs extrapolated to the
$Z$~band. The resulting SN photometric measurements are listed in
Table~\ref{tab2} and plotted in Fig.~\ref{fig4}.
Errors shown include uncertainties in the magnitudes of the four 2MASS
comparison field stars. Pre-explosion $JHK_s$ fluxes of S96 were
measured from the 2MASS survey (see Table~\ref{tab2}) and used to
correct the $JHK_s$ light curves. Also shown for comparison is the
3.6~$\mu$m light curve from the present work. In $H$ and $K_s$ the
slopes flatten after $\sim$300~d accompanied by a {\it rise} at
3.6~$\mu$m. This is suggestive of radiation from warm, newly-forming
dust.\\
Optical photometry was taken from \citet{vin06,vin09}. The optical
light curves are displayed in Fig.~\ref{fig4} and have been corrected
for emission from the S96 cluster using the $BVRI$ magnitudes given by
\citet{vin06}. In addition, the $V$ and $R$ points of \citet{zha06}
around the end of the plateau were added to enhance the detail of this
phase. Also shown for comparison (labelled ``Rad'') is the temporal
evolution of the radioactive energy deposition for SN~1987A as
specified by \citet{li93} (0 to 1200~days) and \citet{tim96} (500 to
3500 days) with the addition of the early-time contribution of
$^{56}$Ni decay assuming complete absorption. The radioactive isotopes
include $^{56}$Ni, $^{56}$Co, $^{57}$Co, $^{60}$Co, $^{22}$Na, and
$^{44}$Ti. (In subsequent use of the \citeauthor{li93} and
\citeauthor{tim96} deposition specifications, our addition of the
early-time contribution of $^{56}$Ni decay is assumed.) In the
optical light curves, for about 35~d (115--150~d) after the end of the
fall from the plateau the decline rate matches the radioactive
deposition quite closely, indicating that this was powering the
emission during this phase. The optical decline rates then flatten
during $\sim$150--250~d, indicating the emergence of an additional
power source (see below). After about 250~d, the optical light curves
exhibit a steepening (possibly also present in the $J$~band) which
increases up to the final observations. \\
\subsection{Mid-Infrared Spectroscopy}
\label{sec:spec}
Low-resolution ($R \approx 60$--127) MIR spectroscopy between 5.2 and
14.5~$\mu$m was acquired at nine epochs between 106.3~d and 1393.3~d
with the IRS in low-resolution mode. Long-Low (LL; 14--38~$\mu$m)
observations were also attempted. Unfortunately, the LL observations
were unusable. The LL slit lies at $90^\circ$ to the SL slit. This
meant that, given the scheduling constraints, the LL slit always lay
across the host galaxy, resulting in heavy contamination. The MIR
spectroscopic observations were drawn from the MISC programs plus one
epoch at 1207~d from the SEEDS program. The log of spectroscopic
observations is given in Table~\ref{tab3}.\\
The data were processed through the Spitzer Science Center's pipeline
software, which performs standard tasks such as ramp fitting and
dark-current subtraction, and produces Basic Calibrated Data (BCD) frames.
Starting with these data, we produced reduced spectra using both the
SPICE and SMART v6.4 software packages. We first cleaned individual
frames of rogue and otherwise ``bad'' pixels using the IRSCLEAN
task. The first and last five pixels, corresponding to regions of
reduced sensitivity on the detector, were then removed. The individual
frames at each nod position were median-combined with equal
weighting on each resolution element. Sky background was removed from
each combined frame by subtracting the combined frame for the same
order taken with the other nod position. We also experimented with
background removal by subtracting the adjacent order. In general,
nod-nod subtraction was preferred, as the background sampled in this
way is expected to most closely represent the background underlying
the SN. Any residual background was removed by fitting low-order
polynomials to regions immediately adjacent to the SN position.\\
One-dimensional spectra were then extracted using the optimal
extraction tool within the SPICE software package, with default
parameters. We found that in all cases the source was point-like, with
a full width at half-maximum intensity (FWHM) that was never wider
than the point-spread function (PSF). This procedure results in
separate spectra for each nod and for each order. The spectra for
each nod were inspected; features present in only one nod were treated
as artifacts and removed. The two nod positions were subsequently
combined. The nod-combined spectra were then merged to give the final
spectrum for each epoch. Overall, we obtained excellent continuum
matches between different orders. \\
Despite our careful reduction procedure, the fluxes of the IRS spectra
and the IRAC photometry were not completely consistent. This was due
to (a) differences in the fixed sizes of the spectrograph aperture
slits and the circular apertures used for the image photometry, and (b)
the fact that the spectra were generally taken some days before or
after the imaging data, during which time the SN flux changed. We
therefore recalibrated the IRS spectra against contemporaneous
photometry in the 8~$\mu$m band obtained by interpolation of the light
curve. This band was chosen since it was completely spanned by the
short-low (SL) spectrum. For each epoch, the IRAC 8~$\mu$m
transmission function was multiplied by the MIR spectra and by a model
spectrum of Vega\footnote{The R. Kurucz Stellar Atmospheres Atlas,
1993, ftp://ftp.stsci.edu/cdbs/grid/k93models/standards.}. The
resulting MIR spectra for the SN and for Vega were integrated over
wavelength. The total SN spectral flux in the 8~$\mu$m band was then
obtained from the ratio of the two measurements using a zero (Vega)
magnitude of 64.1~Jy (IRAC Data Handbook, Table~5.1). These were then
compared with the 8~$\mu$m photometry to derive scaling factors by
which the spectra were multiplied. The spectra are plotted in
Fig.~\ref{fig5}, together with contemporary photometric data.\\
The MIR spectra comprise both continua and emission features. Up to at
least 281~d, strong emission from the CO fundamental was present in
the IRS spectra and IRAC photometry. This had disappeared by
500~d. Strong lines of H~I, [Ni~I], [Ni~II], [Co~II], and [Ne~II] were
also present during the first year, but by 500~d only
[Ni~II]~6.64~$\mu$m, [Ni~I]~7.51~$\mu$m, and [Ne~II]~12.81~$\mu$m were
still relatively strong (see Fig.~\ref{fig5}). Apart from the CO
region during 106--281~d, the 5--14~$\mu$m region was dominated by
continuum emission. Moreover, simple extrapolation below and above
the spectral coverage to the limits of the photometric coverage
suggests that the continuum dominated over at least 3.6--24~$\mu$m.
\subsection{Optical and Near-Infrared Spectroscopy}
We acquired optical spectra using ISIS on the WHT, La Palma, and
DEIMOS \citep{faber03} on the 10~m Keck~2 telescope, Hawaii. The
895~d Keck spectrum has already been presented by \citet{vin09}. We
also made use of earlier post-plateau optical spectra obtained by
\citet{vin06} at 89~d and 128~d, \citet{leo06} at 95~d, and
\citet{chu05} at 100~d. NIR spectroscopy of SN~2004dj was obtained
using LIRIS on the WHT and with the OSU-MDM IR Imager/Spectrograph on
the 2.4~m Hiltner Telescope of the MDM Observatory, Arizona. The data
were reduced using standard procedures in Figaro \citep{sho02} and
IRAF. The observing log for the optical and NIR spectra is given in
Table~\ref{tab4}. The post-100~d spectra are plotted in
Figs.~\ref{fig6} and \ref{fig7}.\\
The earlier post-plateau spectra still exhibited pronounced P~Cygni
features in H$\alpha$, He~I~5876~\AA\ + Na~I~D, He~I~10830~\AA,
He~I~20581~\AA, and O~I~7771~\AA\ + K~I~7665/99~\AA. By 461/467~d the
absorption components had largely vanished, with a broad-line emission
spectrum now being observed. A few lines persisted to as late as the
final optical spectroscopy epoch at 925~d. We examined in detail the
evolution of the more isolated of these lines, specifically H$\alpha$,
Pa$\beta$, [O~I]~6300~\AA, [Fe~II]~7155~\AA, and [Fe~II]~12567~\AA.
Table~\ref{tab5} lists the line luminosities (dereddened) versus
epoch, together with the radioactive deposition power specified by
\citet{li93} and \citet{tim96} for SN~1987A, but scaled down to an
initial 0.0095~M$_{\odot}$ of $^{56}$Ni (see \S3.2). The evolution of
the luminosities is plotted in Fig.~\ref{fig8}.
This indicates that from just after the plateau phase to $\sim$460~d,
the H$\alpha$ and Pa$\beta$ luminosities declined at a rate roughly
comparable to that of the radioactive deposition. In contrast, from
their earliest observation at $\sim$300~d, the [O~I] and [Fe~II] lines
decline significantly more slowly than the radioactive rate.
Moreover, by 895~d the summed luminosity of just the H$\alpha$,
[O~I]~6300~\AA, and [Fe~II]~7155~\AA\ lines exceeds that of the
radioactive input by $\sim$40\%, rising to over 60\% by 925~d.
Thus, as with the optical light curves, we deduce the appearance
of an additional source of energy, possibly earlier than 300~d.
\\
Table~\ref{tab6} lists profile parameters expressed as velocities for
the more isolated lines over a range of epochs, shifted to the
center-of-mass rest frame of SN~2004dj. Also listed (Col. 7) for
461--925~d are the maximum blue-wing velocities derived from profile
model matches (see \S3.4.4.1). Preliminary inspection indicated that
for lines within a given element (e.g., H$\alpha$ and Pa$\beta$), the
velocities exhibited similar values and evolution. Therefore, in order
to improve the temporal coverage and sampling, H$\alpha$ and Pa$\beta$
were grouped together, as were [Fe~II]~7155~\AA\ and
[Fe~II]~12570~\AA. The evolution of the line velocities (shifted to
the center-of-mass rest frame of the SN) is plotted in
Fig.~\ref{fig9}, and in more detail in Fig.~\ref{fig10}. These plots
reveal a complex velocity evolution.\\
Up to 138~d, of the three elements considered, only hydrogen lines
could be reliably identified. At 89~d (corresponding to about
half-way down the plateau-edge), the half width at half-maximum (HWHM)
velocity was $1820\pm60$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi, although the red and blue wings
extended to much higher values. In addition, the peak emission
exhibited a blueshift of $-450\pm50$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. Then, as already described
in \S1, a strong asymmetry rapidly developed, this being attributed to
the emergence of an asymmetric, bipolar core \citep{chu06}. Rather
than being entirely due to the bulk motion of the ejecta however, some
of the width of the H$\alpha$ and Pa$\beta$ lines may also have been
produced by scattering from thermal electrons as in the cases of
SN~1998S \citep{chu01} and SN~2006gy \citep{smi10}, but this effect is
unlikely to have a significant influence at later epochs. By the time
of the next observation at 283~d, the asymmetry had diminished, with
the blueshift of the peak now only $-730\pm50$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. By this time the
lines of [O~I] and [Fe~II] had emerged, also with asymmetric
blueshifted profiles. As the SN continued to evolve, the line widths
narrowed and by 461~d we see the first signs of a sharp suppression of
the red wing. By the time of the next season's observations (895~d,
925~d), this suppression is very pronounced in all three species.
This phenomenon suggests dust formation resulting in the obscuration
of the far side of the ejecta and will be examined in more detail in
\S3.4.4.1.\\
By 895~d and 925~d a second, weaker peak redshifted by $\sim$170~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi
had also appeared in the H$\alpha$ and [O~I]~6300~\AA\ profiles
(Fig.~\ref{fig10}). (The redshifted peak can also be seen in the
weaker [O~I]~6364~\AA\ component lying at $\sim$+3000~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi in the
[O~I]~6300~\AA\ rest-frame plots.) The previous observations of these
lines were at 467~d when the line luminosities were about a factor of
20 greater (Table~\ref{tab5}) as well as being much wider
(Table~\ref{tab6}). Therefore an underlying, weaker, redshifted
component could have been present at 467~d or earlier, but was swamped
by the main component of the line. The 895~d spectrum has a blueshift
in the main peak in H$\alpha$ of $-140\pm10$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ with the minor peak
showing a redshift of about $+160\pm10$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. In [O~I]~6300~\AA\ the
corresponding velocities are, respectively, $-210\pm10$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ and
$+170\pm10$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. Thus, the main (blueshifted) peak and the weaker
(redshifted) peak lie roughly symmetrically about the local zero
velocity. This suggests that a minor fraction of the line flux
originates in an emission zone centered on the SN and having the
geometry of an expanding ring, jet, or cone. The [Fe~II]~7155~\AA\
line also shows a secondary peak but at a much larger redshift of
$+480\pm10$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. There may actually be a peak also at around
+170~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi, similar to those seen in H$\alpha$ and [O~I]~6300~\AA, but
which is swamped by the stronger peak at the larger redshift. The
profiles are analyzed in \S3.4.4.1.\\
We do not consider the MIR line-profile kinematics due to the much
lower resolution ($R_{\rm FWHM} \approx 100$). The formation and evolution
of the MIR lines will be analyzed in a future paper.
\section{Analysis}
\label{sec:analysis}
The evolution of the MIR spectral continuum indicates IR emission from
dust playing a major role in the post-plateau flux distribution of
SN~2004dj. We now make use of the observations described above to
explore the origin, location, distribution, energy source(s), and
nature of these grains. This will be done by comparison of a variety
of simple models with the observations.
\subsection{Correction of the Supernova Flux for S96.}
The position of SN~2004dj coincides with that of the compact star
cluster S96 \citep{san84}, and it seems likely that the progenitor was
a member \citep{wan05}. Regardless of whether or not this is
the case, it is still important to correct for the contribution of the
cluster to the photometry and spectra, especially at the later epochs,
before we embark on modelling the observed SED. To do this, in the
optical region we made use of the pre-explosion optical photometry
compiled in table~4 of \citet{vin09}. NIR photometry of S96 was
obtained from 2MASS (see Table~\ref{tab2}). Unfortunately there are
no pre-explosion MIR images of S96 and so its contribution to the flux
had to be assessed indirectly. \\
We found that longward of $\sim$0.4~$\mu$m the optical/NIR
photometric points could be fairly represented by a combination of two
blackbodies, reddened according to the \citet{car89} law with
$E(B-V)=0.1$~mag and $R_V = 3.1$. This is illustrated in
Fig.~\ref{fig11}.
In this representation, the fluxes longward of $\sim$1~$\mu$m are
dominated by a component with a temperature of 3500~K. At shorter
wavelengths, the hotter component (50,000~K) becomes increasingly
important. This hot blackbody is not intended as an explanation for
the shortwave radiation, but rather it simply serves as a means of
representing and extrapolating the optical SED. The contribution of
S96 to the MIR photometric points was then obtained by extrapolation
of the cooler blackbody. (We did not make use of the \citet{vin09}
models to correct for S96 as it was unclear how they should be
extrapolated into the MIR region.) Some support for the effectiveness
of our estimation method, at least for the shorter wavelengths, comes
from four serendipitous 3.6~$\mu$m images of the SN~2004dj field
spanning 1954--2143~d (Table~\ref{tab1}), obtained in {\it Spitzer}
program 61002 (PI W.~Freedman). These show that, by this period, the
light curve at this wavelength had levelled out at a mean value of
$0.32\pm0.02$~mJy which in good agreement with our estimate for S96 of
$0.28\pm0.05$~mJy (see Fig.~\ref{fig2}).\\
It is possible that the above procedure could underestimate a
contribution from cooler material but it is unlikely that S96 would be
the source of such emission. \citet{mai04} estimate a cluster age of
13.6~Myr and point out that by this age its parent molecular cloud
should have been dispersed by stellar winds and SN explosions.
\citet{wan05} find an age of $\sim$20~Myr while \citet{vin09} obtain
$\sim10-16$~Myr. This suggests that the flux contribution from S96 to
longer MIR wavelengths would be small. On the other hand a significant
SN-driven IR echo from the general IS dust of the host
galaxy is quite likely (see \S3.4.3.1). \\
We conclude that the 3500~K blackbody extrapolation provides a
reasonable estimate of the MIR flux contributions from S96. The
inferred S96 fluxes in the $3.6-24~\mu$m range are shown in
Table~\ref{tab1}. The values are insensitive to the extinction over
the ranges of $E(B-V)$ ($0.06-0.35$~mag) suggested in the literature. It
can be seen that the contribution of S96 at 3.6~$\mu$m is significant
as early as $\sim$250~d, and dominates by $\sim$1000~d. As we move to
longer wavelengths, the effect of S96 declines, becoming negligible
for wavelengths longward of $\sim$10~$\mu$m even at the latest epochs.
In the optical-NIR-MIR continuum modelling (see \S3.4.4.2), we use the S96
blackbody representation to correct for its contribution to the SED.
\subsection{Mass of $^{56}$Ni in the Ejecta}
It is important to establish the mass of $^{56}$Ni in the ejecta of
SN~2004dj since this will allow us to test for the presence of energy
sources, other than radioactive decay, which might be responsible for
the SN luminosity.
In Fig.~\ref{fig12} we show the bolometric light curves (BLCs) of
\citet{vin06} (open circles) and \citet{zha06} (open triangles). The
phases of these have been shifted to our adopted explosion date of
MJD=53196.0. This date is nearly one month later than that of
\citeauthor{zha06} reducing their derived $^{56}$Ni mass by about
25\%. The explosion date of \citeauthor{vin06} is about 10~d earlier
than ours but it is not clear if this would significantly
affect the $^{56}$Ni mass they derived. In addition to the phase
shifts, the BLCs of these authors have been scaled downward to our
adopted distance of 3.13~Mpc. This has the effect of reducing the
$^{56}$Ni masses of \citeauthor{vin06} and \citeauthor{zha06} by 19\%
and 10\% respectively. We scaled the \citet{vin06} BLC by a further
$\times1.1$ in order to allow approximately for the higher total
extinction ($A_V=0.31$~mag) adopted by \citet{vin09} and the present
work. We scaled the BLC of \citet{zha06} by a further factor of
0.68 to force agreement with our adjusted version of the
\citeauthor{vin06} BLC. The need for this was due to the much larger
extinction, $A_V=1.02$~mag, adopted by \citeauthor{zha06}, compared
with the $A_V=0.31$~mag adopted in the present work. It was found
that these adjustments brought our 89~d, 106~d and 129~d blackbody
total luminosities (see \S3.3 and Table~\ref{tab7}, col.~11.) into
fair coincidence with the other two BLCs (see Fig.~\ref{fig12}).\\
\citet{vin06} constructed their BLC by integrating observed
fluxes in the $BVRI$ bands and then extrapolating linearly from the
$B$ and $I$ fluxes assuming zero flux at 3400~\AA\ and
23,000~\AA. Their BLC extended to 307~d. \citet{zha06} simply
integrated the observed fluxes in 12 narrow bands between 4000~\AA\
and 10,000~\AA. Their BLC extended to 154~d, with five additional
points to 180~d obtained by interpolation within a reduced number of
bands. Thus, neither of these BLCs included unobserved excess flux
beyond about 1~$\mu$m. However, for 89--129~d the unobserved MIR flux
made up no more than 10\% of the total luminosity (see
\S3.4.2). Moreover, the optical/NIR region was dominated by continuum
emission at this time. Consequently, the hot+warm continuum
luminosities obtained from the present work via blackbody matching
(see \S3.3 and Table~\ref{tab7} col.~12.) and plotted in
Fig.~\ref{fig12} (solid squares) agree well with the two adjusted
BLCs. (We exclude the cold component because, as argued in \S3.4.3.1,
it is due to an IS IR echo which was predominantly powered by the
peak luminosity of the SN prior to the earliest epoch of observation.)
By 251/281~d our hot+warm continuum luminosities make up only
$\sim$60\% of the adjusted \citeauthor{vin06} BLC. This is due to the
relatively strong contribution of line emission to the total
luminosity during this time. Line emission luminosity was not included
in our blackbody matches (see \S3.3).\\
In Fig.~\ref{fig12}, the SN BLCs are compared with the radioactive
deposition power in SN~1987A, as specified by \citet{li93} \&
\citet{tim96}. These radioactive decay light curves are scaled to,
respectively, 0.0095\,M$_{\odot}$ (solid line) and 0.016\,M$_{\odot}$
(dashed line) of $^{56}$Ni. Also shown (red) is the total radiocative
luminosity in the case of 0.0095\,M$_{\odot}$. We also show (dotted
lines) the actual UV-augmented BLCs of SN~1987A \citep{pun95} derived
from observations at ESO and CTIO, scaled to an initial $^{56}$Ni mass
of 0.0095\,M$_{\odot}$. It can be seen that the 0.0095\,M$_{\odot}$
case provides a good match to the SN~2004dj BLC during $115-150$~d,
just after the end of the plateau phase. After 150~d, unlike SN~1987A,
the SN~2004dj BLC begins to exceed the luminosity of the
0.0095\,M$_{\odot}$ case with the discrepancy growing steadily with
time. Indeed, even the total radioactive luminosity of the scaled
SN~1987A is exceeded by the SN~2004dj light curve, implying that
an additional source of energy has appeared. \\
Our 0.016\,M$_{\odot}$ deposition plot corresponds approximately to
the 0.02\,M$_{\odot}$ case of \citet{vin06} in their Fig.~18. We
agree that this case provides a fair match to the BLC during
$\sim260-310$~d. Nevertheless, viewed within the context of the whole
BLC, it can be seen that this ``match'' is actually due to an
inflection section during the growth of the BLC excess relative to the
true radioactive deposition. Adoption of the 0.016\,M$_{\odot}$ case
would imply an unexplained BLC {\it deficit} during $\sim95-250$~d.
Given the phase of the event and the unexceptional progenitor mass it
is difficult to see how such a discrepancy would come about.
Consequently, we argue that only during the $115-150$~d phase was the
BLC of SN~2004dj actually dominated by radioactive decay. Beyond this
period, and as deduced also in \S2.2 and \S2.3, an additional
luminosity source emerged. We therefore reject the $^{56}$Ni mass
deduced by \citet{vin06}.\\
We also reject larger $^{56}$Ni masses reported by other authors.
\citet{kot05} used the $V$~band exponential tail method of
\citet{ham03} to derive a $^{56}$Ni mass of $\sim 0.022\,$M$_{\odot}$.
This was based on the $V$ magnitude at 100~d which the subsequently
more complete database shows was not quite yet on the radioactive tail
(see Fig.~\ref{fig4}), thus leading to an overestimate of the
$^{56}$Ni mass. This method was also one of those used by
\citet{zha06} who obtained $0.025\pm0.010$~M$_{\odot}$ of $^{56}$Ni.
As already indicated, their larger value was due mostly to their much
earlier explosion epoch and much larger extinction, neither of which
we view as likely. \citet{chu05} obtained $0.020\pm0.006$~M$_{\odot}$
of $^{56}$Ni based on comparison of the $V$ magnitude at 200~d with
that of SN~1987A. The difficulty here is that by this epoch (as also
in the \citet{vin06} case) an additional power source had appeared in
SN~2004dj, biasing the derived $^{56}$Ni mass to higher values. In
addition, \citet{chu05} used an exceptionally early explosion date,
pushing their result even higher. Finally we note that both
\citet{chu05} and \citet{zha06} also used the $V$-light curve
``steepness'' method of \citet{elm03}, which is insensitive to
distance and extinction uncertainties. \citeauthor{chu05} obtained
$0.013\pm0.004$~M$_{\odot}$, consistent with our result.
\citeauthor{zha06} applied the same method to a number of wavebands,
including $V$, but obtained a larger $0.020\pm0.002$~M$_{\odot}$.
Their steepness parameter at just $V$ yields about $0.019$~M$_{\odot}$
suggesting that their use of multiple bands is not the cause of the
apparent disagreement with \citeauthor{chu05} However, the difference
between the \citeauthor{chu05} and \citeauthor{zha06} determinations
is only at the level of $\sim1.5\sigma$ significance. \\
We conclude that, taking into account the uncertainties in fluxing,
adopted distance, extinction and explosion epoch (see \S1.1) the mass
of $^{56}$Ni ejected by SN~2004dj was
$0.0095\pm0.002$\,M$_{\odot}$. We adopt this value for the rest of the
paper.
\subsection{Comparison of Observed Continua with Blackbody Radiation}
Here we begin to consider the location and energy source of the SN
continuum, especially longward of 2~$\mu$m where thermal emission from
dust would appear. To take an initially neutral standpoint on the
interpretation, we compared optical, NIR and MIR spectra and
photometry with blackbody continua. This provides us with the minimum
radii of the emitting surfaces. The epochs were selected primarily as
those for which MIR spectra were available, although the earliest such
epoch, 106~d, was already during the nebular era. The earliest MIR
photometry was acquired at 89~d when the SN light curve was only about
half way down the fall from the plateau to the nebular level. Given
the potential interest of this epoch we began our model comparisons at
this epoch despite the lack of an MIR spectrum. In addition, to
compensate for the large gap between 859~d and 1207~d we also
considered the 996~d SED based on MIR photometry only. Optical
photometry was taken from \citet{vin06,vin09,zha06}. Details about the
sources of the other data are given in Tables~\ref{tab1}--\ref{tab4}.
Apart from 89~d and 996~d, all the optical and NIR data plus the MIR
photometry were flux-scaled by interpolation of the light curves to
the epochs of contemporaneous MIR spectra. \\
The contribution of S96, represented by a 3500~K blackbody of radius
$1.5\times10^{14}$~cm (see \S3.1), was first subtracted from all the
data. To model the resulting $0.4-24~\mu$m SN continuum it was found
to be necessary to use three blackbodies (``hot'': $5300-10000$~K,
``warm'': $320-1750$~K and ``cold'': $\sim$200~K). These were
reddened and then matched visually to the continua. The hot blackbody
was first added and adjusted to match the optical continuum. While the
hot blackbody provides some information about the energy budget of the
shorter wavelength part of the spectrum, the main reason for its
inclusion in this study was to allow correction for its effect on the
net continuum in the NIR where, up to about 500~d, it is comparable in
strength to the warm component. The warm blackbody was then added and
adjusted to match the $\sim5-10~\mu$m continuum plus the long
wavelength end ($2-2.4~\mu$m) of the NIR continuum. It was found
that, starting at the earliest epoch, as the SN evolved the hot+warm
blackbody flux longward of 10~$\mu$m increasingly fell below that of
the observations. Therefore a cold blackbody was added and adjusted
to provide the final match. For epochs where photometry but no
spectra were available we used the temporally nearest spectral matches
to indicate the likely position of the underlying continuum. The
expansion velocities of the blackbody surfaces, $v_{hot}$, $v_{warm}$
and $v_{cold}$, and temperatures, $T_{hot}$, $T_{warm}$ and
$T_{cold}$, for these matches are tabulated in Table~\ref{tab7}. The
warm blackbody radii, $R_{warm}$ are also listed. The model matches
are displayed in Figs.~\ref{fig13} ($89-500$~d) and \ref{fig14}
($652-1393$~d). In Table~\ref{tab7} we also show the luminosities,
$L_{hot}$, $L_{warm}$ and $L_{cold}$, of the three blackbody
components together with the sum of the hot and warm components
($L_{total}$). We exclude the cold component because, as will be
argued in \S3.4.3.1, it is due to an IS IR echo which was predominantly
powered by the peak luminosity of the SN prior to the earliest epoch
of observation. In the final column is listed the radioactive
deposition power corresponding to the ejection of 0.0095\,M$_{\odot}$
of $^{56}$Ni, scaled from the SN~1987A case specified by \citet{li93}
\& \citet{tim96}.\\
We stress that the blackbody matches were to the underlying spectral
{\it continua} where this could be reasonably judged, and not to the
photometric points which also contained flux from line emission.
Thus, the models sometimes lie below the average level of the spectra.
This is particularly so for later epochs at wavelengths shortward of
2~$\mu$m where the spectra are dominated by broad, blended emission
lines. This tends to mask most of the underlying thermal continuum,
leading to a possible overestimation of the hot
continuum. Nevertheless, the blackbody luminosities tend to {\it
underestimate} the total luminosity as they do not allow for the total
line emission. This is particularly the case around $200-500$~d when
the relative contribution of nebular line emission to the total
luminosity is at a maximum. There is less of a problem before this
era when the hot continuum is relatively strong, or afterwards when
the warm/cold continuum increasingly dominates. Also, by 652~d, the
relative weakening of the hot continuum means that it has a negligible
effect at wavelengths longward of 2~$\mu$m. In the $2-14~\mu$m region
the true continuum level is easier to judge. By 251~d and later, the
total $2-14~\mu$m flux exceeds that of the continuum model by no more
than 25\%. \\
The hot continuum declined monotonically and dominated the total SN
continuum luminosity up to about a year post-explosion. It presumably
arose from hot, optically-thick ejecta gas. At 89~d most of the hot
continuum was probably still driven by the shock-heated photosphere,
with the remainder being due to radioactive decay (see
Table~\ref{tab7} and Fig.~\ref{fig12}). After 500~d the hot continuum
became relatively weak and of low S/N. In addition, the NIR spectral
dataset ended on 554~d. For subsequent epochs the strength of the hot
component was estimated by extrapolation. At 500~d the hot blackbody
match was achieved with a temperature of 10,000~K. Between 251~d and
500~d the blackbody velocity declined exponentially with an e-folding
time of $\sim$130~d and so, for epochs after 500~d the hot blackbody
temperature was fixed at 10,000~K and the velocity obtained by
extrapolation of the earlier exponential behavior (see
Table~\ref{tab7}). While this is likely to be increasingly inaccurate
with time, it is unlikely to be a serious source of error in
determining the warm continuum; for example, by 500~d the hot
blackbody contributed barely 1\% of the flux at
3.6~$\mu$m. Consequently, and in order to illustrate the MIR behavior
in more detail, the optical/NIR region is not shown in the later plots
(Fig.~\ref{fig14}).\\
The warm component luminosity declined monotonically to 281~d but
then, unlike the hot component, increased by a factor of 3.5
by 500~d. By that time the hot+warm SN continuum luminosity exceeded
that of the radioactive energy input by a factor of 4 and this
excess continued to increase with time. The cause of the growing
excess was the warm component luminosity. We also note that
$R_{warm}$ more than doubled between 281~d and 500~d, but remained
roughly constant thereafter. In fact, as we show later (\S3.4.4.1,
\S3.4.4.2), after 500~d slow shrinkage in the size of the warm emission
region occurred. \\
The cold component is primarily defined by the 24~$\mu$m point, and
can be fairly reproduced using a range of temperatures ($150-300$~K)
and velocities. Its luminosity remained roughly constant throughout
the observations. In Table~\ref{tab7} we show the case with the
temperature fixed at 200~K. \\
It is interesting that, even as early as 89~d, warm and cold
blackbodies had to be included to achieve a fair match to the observed
fluxes in the NIR-MIR region. This will be discussed in \S3.4.2 and
3.4.3.1. By 106~d, the nebular phase was just beginning and by 129~d
the hot+warm luminosity was driven predominantly by radioactive decay.
By 251~d and 281~d, the radioactive deposition exceeded the sum of the
hot+warm blackbody luminosity by about 15\%. This excess probably went
into powering the line emission not included in the blackbody matches.
As already indicated, the nebular line emission was particularly
strong at this time. Indeed it was responsible for a $\sim50\%$
excess of the total bolometric luminosity relative to the radioactive
deposition (Fig.~\ref{fig12}), indicating the emergence of an
additional energy source, probably the reverse shock (\S3.4.4.2). The
appearance of an additional power source has already been indicated by
the flattening of the optical light curves (\S2.2). Indeed, after
$\sim$150~d the true BLC exhibited a steadily growing excess relative
to the radioactive deposition (see Fig.~\ref{fig12}). \\
The hot and warm blackbody velocities never exceeded 1750~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi,
indicating continuum emission consistent with an origin in the ejecta
or ejecta/CSM interface. The cold component exhibits velocities
between 4500~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ and 8500~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ during the earlier phase pointing
to an origin more likely to be outside the ejecta, specifically an IR
echo from pre-existing dust. \\
We conclude that the IR continuum comprised at least two components.
The temperatures and temporal variation of these components point to
thermal emission from dust whose energy source is ultimately the
supernova. The surge in the luminosity of the warm component by 500~d
suggests the emergence of an additional source of radiation --- i.e.,
that the warm component was driven by different energy sources at,
respectively, early and late times. It also raises the possibility
that distinct dust populations were responsible for the early and
late-time warm components. \\
\subsection{Origin of the Infrared Radiation}
We now explore the origin of the IR continuum radiation from
SN~2004dj, especially the warm component. To do this we have
constructed a model continuum comprising hot gas, warm local dust, and
cold IS dust. We also made use of spectral line red-wing suppression
in the study of the warm dust. The continuum model was adjusted to
provide visual matches to the observations.
\subsubsection{The Hot Component}
As in \S3.3, the hot continuum component, presumably due to
optically-thick ejecta gas, was represented using a hot blackbody
having a temperature of 5300--10000~K. This blackbody radius and
temperature was adjusted to obtain a match to the optical-NIR
continuum. The warm component model (see \S3.4.2 and \S3.4.4.2) was
then added and adjusted to match the $\sim$5--10~$\mu$m continua plus
the long-wavelength end (2--2.4~$\mu$m) of the NIR continuum. As with
the pure blackbody matches it was found that, at all epochs, the warm
component model flux longward of 10~$\mu$m increasingly fell below
that of the observations. This excess is attributed to an IS IR echo
(see \S3.4.3.1).
\subsubsection{The Warm Component: Early Phase IR Excess due to a CDS}
As pointed out above, a striking result from the hot-blackbody matches
is that an NIR-MIR excess (relative to the hot component flux) was
present as early as 89~d post-explosion. (For brevity we henceforth
refer to the NIR-MIR continuum excess as the ``IR excess''.) The MIR
spectra from 106~d onward show that this was primarily due to
continuum emission. We have no reason to suspect that the IR excess in
the 89~d SED was not also due to continuum emission. Indeed, the CO
peak at 4.5~$\mu$m is noticeably suppressed at this epoch, compared
with 106~d (Fig.~\ref{fig3}). The obvious interpretation of the IR
excess is that it arose from warm dust heated by the supernova. The
very early appearance of this emission, when the H-recombination front
had not yet reached the He-metal core, argues against an origin in
newly-formed ejecta dust. In addition, SN-ejecta dust-formation
models \citep[e.g.][]{tod01,noz03} suggest that dust formation in a
CCSN is unlikely to occur until after one year post-explosion. A
second possibility of direct shock-heating of pre-existing
circumstellar dust is also ruled out. The SN UV flash would evaporate
dust out to $(0.5-1.0)\times10^{17}$~cm ($0.016-0.032$~pc) (see
\S3.4.3.1). To reach this distance by 89~d would require a shock
velocity of at least 65,000~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. Yet the velocities of the H$\alpha$
trough \citep{kor05,vin06} as well as of the extreme blue-profile edge
(our measurements of the spectra of \citet{kor05,vin06}) during the
period $25-90$~d suggests that the bulk of the ejecta never exceeded
velocities much more than $\sim$15,000~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. Indeed \citet{chu07}
adopt a velocity of 13,000~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi as the boundary velocity in their
treatment of SN~2004dj. A third possibility, CSM dust heating by
X-rays, is also implausible. The luminosity of the warm continuum
component at 89~d was $1.3\times10^{40}$~\ifmmode{\rm erg\,s^{-1}}\else\hbox{$\rm erg\,s^{-1}$}\fi (Table~\ref{tab7}) but
the X-ray luminosity at 30~d was only about 1\% of this \citep{poo04}.
In the case of dust in a CDS, collisional heating is ruled out as the
energy available in the CDS is about a factor of $10^5$ less than that
required to account for the observed IR excess. The heat capacity
of the grains is also insufficient to account for required energy. \\
The most likely explanation for the early-phase IR excess is an IR
echo of the SN early-time luminosity from circumstellar dust. There
are two possible scenarios here. In the first of these, the IR echo is
from pre-existing dust in the CSM. Such early IR echoes have been
suspected before in other CCSNe. \citet{woo93} found an IR excess in
SN~1987A as early as 260~d and possibly also at just 60~d. They
hypothesised that the origin of the excess was warm, SN-heated dust in
the CSM. \citet{woo97} reiterates that the cause was CSM dust
``echoing the light curve''. \citet{fas00} reported a strong $K-L'$
excess in the emission from the Type~IIn SN~1998S at 130~d. They
attributed this to pre-existing CSM dust heated either by the SN
luminosity (a conventional IR echo) or by X-rays from the CSM-shock
interaction. \citet{poz04} argued in favor of the former scenario.\\
In the second IR echo scenario, as the fast moving ejecta collides
with the CSM, a CDS forms between the forward and reverse shocks. As
noted in \S1, \citet{chu07} have used the H$\alpha$ spectrum to deduce
the existence of a CDS in SN~2004dj. Within the CDS conditions can
allow new dust to form \citep{poz04}. CDS dust has been invoked as the
origin of the early-time IR excess of SN~2006jc
\citep{smi08,dic08,mat08}. \citet{mat08} showed that the IR emission
was probably an IR~echo from the CDS dust. IR excesses in CCSNe at
later times have also been attributed to emission from newly-condensed
CDS dust viz. in the Type~IIn SN~1998S \citep{poz04} and the Type~IIP
SN~2007od \citep{and10}. \\
We explored the possibility that the early IR excess was due to an IR
echo of the SN luminosity from spherical distributions of either
pre-existing circumstellar dust or newly-formed CDS dust. Details of
the model are given in \citet{mei06}. It fully allows for the effects
of light-travel time across the dust distribution. Versions of this
model have also been used in \citet{mei07,mat08,bot09,kot09}. The
model assumes a spherically symmetric cloud of grains centered on the
SN, with a concentric dust-free cavity at the center. The SN is
treated as a point source. For simplicity, a single grain radius, $a$,
is adopted. For ease of computation, we assumed that the grain
material was amorphous carbon where, for wavelengths longer than $2
\pi a$, the grain absorptivity/emissivity can be well-approximated as
being proportional to $\lambda^{-1.15}$ \citep{rou91}. For shorter
wavelengths, an absorptivity/emissivity of unity was used. The
material density is 1.85\,g\,cm$^{-3}$ \citep{rou91}. Free parameters
are the grain size, grain number density, radial density law and
extent. The input luminosity is a parametrized description up to
550~d of the BLC shown in Fig.~\ref{fig12} viz.: \\
$L_{\rm bol} = L_0 {\rm exp}(-t/\tau)$, where \\
$L_0=57.0$, $\tau=1000.0$~d for $t \le 0.2$~d,\\
$L_0=1.70$, $\tau=23.4$~d for $0.2 < t \le 23.0$~d,\\
$L_0=0.66$, $\tau=141.9$~d for $23.0 < t \le 76.4$~d,\\
$L_0=50.7$, $\tau=15.9$~d for $76.4 < t \le 112.0$~d.\\
$L_0=0.079$, $\tau=174.8$~d for $112.0 < t \le 550.0$~d \\
$L_0=0$ for $t > 550.0$~d \\
\noindent $L_0$ is in units of $10^{42}$~\ifmmode{\rm erg\,s^{-1}}\else\hbox{$\rm erg\,s^{-1}$}\fi. The brief but
highly luminous first term represents the energy in the UV flash. The
second term, which covers the pre-discovery phase, was estimated by
using the BLC of SN~1999em \citep{elm03}, adjusted so that the epoch
and luminosity of the beginning of its plateau phase coincided
approximately with the earliest observed point on the SN~2004dj BLC.\\
For the third, fourth and fifth terms, the adjusted BLCs of
\citet{vin06} and \citet{zha06} were used (see \S3.2). As already
pointed out, these BLCs did not include excess flux beyond about
1~$\mu$m, nor any flux beyond 2.3~$\mu$m; virtually all of the warm
and cold components were excluded. Thus, use of the above parametrized
description avoids ``double-counting'' of a putative CSM/CDS IR echo.
Indeed the BLC description slightly underestimates the true SN
luminosity input as it excludes all IR emission longward of 2.3~$\mu$m,
as well as all UV emission shortward of 0.34~$\mu$m. Consequently the
input luminosity was scaled by a factor of about 1.1 to allow
for unobserved UV and IR fluxes. \\
For the case of pre-existing CSM dust the model was adjusted to
reproduce the IR-excess SEDs for the three earliest epochs (89, 196
and 129~d). The outer limit of the circumstellar dust was initially
set at 10 times that of the cavity radius, and a $r^{-2}$ (steady
wind) density profile was assumed. However, it was impossible to
reproduce the quite rapid temporal decline of the IR~excess without
raising the temperature of the hottest grains to above their
evaporation temperature. While a better match to the SED shape and
evolution was obtained by setting the density profile to steeper than
$r^{-4}$, the best match was achieved with a discrete, thin shell. We
therefore adopted this configuration, setting the shell thickness at
$\times0.1$ the cavity radius. \\
We investigated a range of cavity radii. For pre-existing CSM dust
the minimum size of the concentric dust-free cavity is fixed by the
extent to which the dust was evaporated by the initial UV flash from
the supernova. In this scenario, while the bolometric light curve (BLC)
dominates the heating of the surviving dust, the size of the dust-free
cavity is determined by the luminosity peak of the UV flash, with
$r_{evap} \propto L_{peak}^{0.5}$ \citep{dwe83}. For a Type~IIP SN the
flash luminosity is estimated to peak at about $10^{45}$~\ifmmode{\rm erg\,s^{-1}}\else\hbox{$\rm erg\,s^{-1}$}\fi
\citep{kle78,tom09} although it has never been observed directly. A
similar peak luminosity is estimated for the Type~IIpec SN~1987A
\citep{ens92}. \citet{dwe83} provides an approximate estimate of the
flash-evaporated cavity size for a Type~II supernova. More recently,
in a detailed study \citet{fis02} determined that for SN~1987A the UV
flash would have totally evaporated graphite CSM dust out to a radius
of $(0.5-0.9)\times10^{17}$~cm ($0.016-0.029$~pc). We therefore adopt
the \citeauthor{fis02} estimates and apply them to the case of
SN~2004dj. In any case, matches to the data are fairly insensitive to
the cavity radii. Fair matches to the early-time IR-excess SEDs were
obtainable with cavity radii $(0.5-1.3)\times10^{17}$~cm
($0.016-0.042$~pc) and corresponding grain radii of
$0.1-0.04~\mu$m. For cavity radii exceeding $\sim$0.05~pc the model
continuum slopes were inconsistent with the observations.\\
As an example of a pre-existing CSM dust model match we have: \\
\noindent cavity radius: $0.7\times10^{17}$~cm (0.023~pc), \\
shell thickness: $0.7\times10^{16}$~cm (0.0023~pc),\\
grain size: 0.07~$\mu$m, \\
grain number density: $6.0\times10^{-9}$~cm$^{-3}$,\\
total grain mass: $0.38\times10^{-5}$~M$_{\odot}$. \\
The optical depth through the shell in the optical band is 0.006,
which is easily encompassed within the observed total extinction of
$A_V=0.3$~mag. For a dust/gas mass ratio of 0.005 (see \S3.4.3.1), the
dust mass corresponds to a total shell mass of
$0.76\times10^{-3}$~M$_{\odot}$. For the adopted shell thickness and
a typical RSG wind velocity of 20~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi, this mass would be produced by
a mass loss event about 1100~years ago, lasting for 110~years, with a
mass loss rate rate of $7\times10^{-6}$~M$_{\odot}$~yr$^{-1}$.
However, while this mass loss and loss rate are plausible for RSGs,
such discrete events are not thought to occur in this type of star.
This prompts us to seek a more natural explanation for the thin dust
shell. An obvious candidate is the CDS inferred by \citet{chu07}.\\
For the CDS scenario the size of the cavity is essentially the radius
of the CDS. From \citet{chu07}, the CDS radius is given by
\begin{equation}
r_{CDS}=5.2\times10^{15}\times((t+11.5)/64.0)^{6/7}~cm
\end{equation}
\noindent where $t$ (in days) is with respect to our explosion epoch. Thus,
at the earliest of our epochs, 89~d, the CDS radius was just
$7.7\times10^{15}$~cm, or about one tenth of the pre-existing CSM dust
cavity. This is only 3~light days implying that light travel time
effects are small. Nevertheless, for convenience and ease of comparison
with the pre-existing dust case, we applied the IR~echo model to the
CDS case. Although the \citeauthor{chu07} study terminates at just
99~d, we assume that the CDS radius continues to increase as described
in equation~(1) until at least 500~d, by which time its
contribution to the IR excess is small. \\
We compared the CDS case of the IR~echo model with the observations,
with the dust lying in a thin shell of radius $r_{CDS}$ as given
above. We found that setting the dust mass at a constant value
produced a poor match to the observations. With a match at 106~d, the
model at 89~d yielded a continuum which matched the observed IR excess
at 8~$\mu$m but exceeded the IR excess by nearly $\times2$ at
3.6~$\mu$m. On the other hand, at 129~days, while the model continuum
slope was similar to that of the observed IR excess, it significantly
underproduced it with the deficit being as large as $\sim$35\% at
8~$\mu$m. We propose that this problem is due to the unjustified
assumption of a fixed CDS dust mass. CDS dust could not form until
the supernova flux at the CDS had faded sufficiently for proto-dust
material to cool below the condensation temperature. This occurred at
about 50~d, assuming amorphous carbon dust. Therefore, following
\citet{mat08} we allowed the CDS dust mass to grow as
$M_d=M_0(1-exp(-(t-t_0)/t_d))$ where $t$ is time, $t_0$ is the time at
which dust condensation began (set at 50~d), $t_d$ is the
characteristic grain growth timescale and $M_0$ is the
asymptotically-approached final mass. (We note that
\citeauthor{mat08} also deduced an epoch of 50~d for the start of the
CDS dust condensation in SN~2006jc.) No attempt was made to simulate
the growth of individual grains which were assumed to appear
instantaneously at their final size. Owing to the light travel time
differences across the CDS, the grain condensation is seen to commence
during the epochs ($t_0-(r_{CDS}/c)$) to ($t_0+(r_{CDS}/c)$) days.
Yet, even as late as 500~d, $(r_{CDS}/c)$ was only about 12~light days
and so this effect was ignored. \\
In Fig.~\ref{fig15}, we compare the CDS model light curves and spectra
with, respectively, the observed MIR excess fluxes at 3.6~$\mu$m and
8.0~$\mu$m (LH panel), and with the 89, 106 and 129~d MIR excess SEDs
(RH panel). The free parameters for the CDS case were the dust mass
scaler, the grain radius and the grain growth timescale. Satisfactory
reproduction of the IR excess was achieved for all five epochs
spanning 89--281~d with $t_d=50$~d, a grain radius of 0.2~$\mu$m and a
final dust mass of $M_0=0.33\times10^{-5}$~M$_{\odot}$, similar to the
dust mass derived in the pre-existing CSM case above. The CDS mass is
$3.2\times10^{-4}$~M$_{\odot}$ \citep{chu07} indicating a plausible
final dust/gas mass ratio of 0.01. The hottest dust ranged from 1330
K at 89 days declining to about 640 K at 281 days. The optical depth
in the UV/optical range was 0.077 at 89~d falling to 0.011 by 500~d
--- i.e., consistent with the observed total extinction of
$A_V=0.3$~mag. By 500~d the MIR excess exceeds that of the model by a
significant factor (especially at longer wavelengths)
(Fig.~\ref{fig15}) implying the appearance of an additional energy
source. The CDS contribution to the SN continua for 89--500~d is
plotted in Fig.~\ref{fig18}. The luminosity contribution of the CDS
IR~echo, $L_{CDS}$, is listed in Table~9, Col.~3. For epochs after
500~d the CDS component was negligible and so was ignored. \\
We conclude that the early-time IR~excess was probably due primarily
to an IR~echo from newly-formed dust lying within the CDS. The rapid
decline of the IR excess flux, especially between 89~d and
106~d, is due largely to the ongoing fall from the BLC plateau,
tempered by the growth of CDS dust during this period. Owing to the
small size of the CDS, light travel time effects are small. In
contrast, in the pre-existing CSM dust scenario, the rapid decline of
the input luminosity is tempered by the larger size of the dusty shell
which produces significant light travel time effects, smoothing the
observed IR excess light curve over longer timescales. Nevertheless,
given the natural explanation by the CDS model of the required thin
shell dust distribution, in the completion of the analysis below we
use the CDS scenario. \\
\subsubsection{The Cold Component}
\paragraph{{\it An Interstellar Echo\/}}
The cold component is defined primarily by the 24~$\mu$m data and can
be fairly reproduced using a range of temperatures ($150-300$~K) and
velocities. As described above (\S3.3), the luminosity of the cold
component remained roughly constant throughout the observations. In
addition, the cold blackbody velocities were high, arguing against an
origin in ejecta dust. An obvious alternative is an IS IR echo. The
possibility of detecting the reflection of SN optical light from IS
grains was first suggested by \citet{vdb65}. In the IR a potentially
much more important phenomenon is the absorption and re-radiation by
the grains of the SN BLC energy (the ``IR echo''). The possibility of
detecting an IS IR echo from a SN was first proposed by
\citet{bod80}.\footnote{\citet{wri80} also considered an IR echo from
a SN but only in the more restricted case of an explosion within a
molecular cloud.} The occurrence of cold dust IS IR echoes should be
relatively common for SNe occurring in dusty, late-type galaxies. The
SEDs of such echoes tend to peak in the $20-100~\mu$m region, allowing
echo detection in nearby galaxies by {\it Spitzer} during its cold
mission. {\it Spitzer}-based evidence of this phenomenon in the Milky
Way Galaxy have been presented for the Cassiopeia~A SN
\citep{kra05,kim08,dwe08}. \citet{mei07} showed that an IS IR echo
provided a natural explanation for the strength and decline of the
24~$\mu$m flux between $670-681$~d and 1264~d in SN~2003gd in the
SA(s)c galaxy NGC~628 (M74). \citet{kot09} showed that the cold
component of the SN~2004et SED was most likely due to an IS IR echo in
the SAB(rs)cd host galaxy NGC~6946. The host galaxy of SN~2004dj,
NGC~2403, is of type SAB(s)cd and so there is a good likelihood of a
similar IS IR echo occurring. Therefore, we included an IS IR echo
component in our modelling of the SED. \\
Our IS IR echo model is the same as that used for the
early-time CDS IR echo (\S3.4.2). Only the dust distribution and
grain radii are different. We ignore the CDS dust emission derived in
\S3.4.2 since this is already invoked in the CDS echo model and
included in the continuum modelling up to 500~d. We recognise that
a more extended, lower density CSM may also have existed. However, we
found that the addition of yet another model component was unnecessary
to provide plausible matches to the continua and so for simplicity the
possible effects of an extended CSM were ignored.\\
As described in \S3.4.2, pre-existing dust surrounding the SN would
have been evaporated by the SN flash out to a distance of about
0.025~pc. At sufficiently late epochs, enlargement of this cavity can
be produced by the forward shock. Assuming that the shock velocity is
comparable to the highest ejecta velocity viz. $\sim$15,000~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ (see
\S3.4.2), and that there was no deceleration, the edge of the
UV-flash-determined cavity would be reached after about 600~d. After
this time, the SN shock would evaporate the dust and enlarge the
dust-free cavity. Therefore, for epochs earlier than 600~d we fixed
the inner limit of the IS dust at 0.025~pc. For later epochs the
inner limit of the IS dust increases from 0.028~pc (34~light days) at
662~d to 0.058~pc (70~light days) by 1393~d. It might be objected
that this ignores the possibility of shock deceleration. However, the
IR~echo contribution which the shock-evaporated dust would otherwise
have made, for $600>t>1393$~d, to the cold component is negligible
since the cavity radius was never more than 70~light days i.e. during
this late period the SN peak luminosity would have long since passed
by the dust near the cavity. \\
The outer limit of the IS dust is much less certain. \citet{ben10}
have used {\it Spitzer} imaging of NGC~2403 at 70~$\mu$m and
160~$\mu$m to map the dust column density via its thermal emission.
They also used the far-IR images in conjunction with H~I observations
to map the gas/dust ratio. The objection to using such measurements in
the present work is that they only provide total densities through the
disk (or to the mid-plane assuming symmetry), and not directly to
SN~2004dj. While it was argued above that SN~2004dj actually lies
towards the front of S96, we are still faced with the uncertainty of
the depth of S96 in the galaxy plane. Our approach, therefore, is to
assume a spherically symmetric dust distribution centered on the SN,
with an outer limit of 100~pc. While not appropriate in principle for
the outer limits of the dust in the galactic disk, for the early era
being considered spherical symmetry provides a good approximation; at
this epoch, the echo ellipsoid is extremely elongated and so the
region of the spherical model outer surface intercepted by the
ellipsoid is small. The plane of NGC~2403 appears to be tilted at an
inclination of about $55^{\circ}$ roughly doubling the face-on column
density, and so the adopted outer limit is equivalent to a $\sim$60~pc
scale height for the IS dust with the host galaxy face-on.\\
The free parameters were (i) the grain size, which influenced the dust
temperatures, and (ii) the grain number density, which determined the
luminosity. These parameters were adjusted in conjunction with the
warm dust models (see \S3.4.2 and \S3.4.4.2) to provide a match to the
longwave excess. It was found that satisfactory matches at all epochs
were obtained with a grain radius of 0.1~$\mu$m and a dust number
density of $2.6\times10^{-13}$~cm$^{-3}$ (i.e., an IS gas number
density of 0.24~cm$^{-3}$ and a dust/gas mass ratio of 0.005 -
\citealt{ben10}.) This is comparable to the typical value of
$5\times10^{-13}$~cm$^{-3}$ for the Milky Way \citep{all73} and
$7.0\times10^{-13}$~cm$^{-3}$ obtained by \citet{kot09} for NGC~6946.
In the wavelength region ($\lambda >15~\mu$m)
where the cold component makes a significant contribution ($>20\%$) to
the total flux, the IS IR echo model (15--150~$\mu$m) maintained a
near-constant luminosity of $\sim2.0\times10^{38}$~\ifmmode{\rm erg\,s^{-1}}\else\hbox{$\rm erg\,s^{-1}$}\fi between 89~d
and 1393~d. \\
The \citeauthor{ben10} dust column density map of NGC~2403 indicates
0.05~M$_{\odot}$pc$^{-2}$ at the position of SN~2004dj. From the above
dust number density from the echo model and integrated over 100~pc we
obtain just 0.003~M$_{\odot}$pc$^{-2}$. The absorption opacities used
by \citeauthor{ben10} are consistent, to within a factor of $\sim2$,
with the opacity law used in the present work. Thus, we have a
$\sim\times$8 discrepancy in column density between \citeauthor{ben10}
and the present work. Part of this discrepancy may be that the
adopted dust outer limit is too small. The derived IS dust
density is fairly insensitive to the extent of the outer limit. For
example, with a 75\% increase in the outer limit the model match to
the longwave continuum is retained with a reduction of just $\sim$10\%
in the IS dust density. Thus, at least some of the discrepancy could
be removed by simply increasing the dust outer limit. An explanation
for the remaining discrepancy is that the SN and S96 actually lie
significantly above the mid-plane of NGC~2403. This is confirmed as
follows. The optical depth to UV/optical photons yielded by the model
is about 0.026. We can use this to estimate independently the host
extinction to SN~2004dj. The optical depth translates to an
absorption-only extinction of 0.028 magnitudes. Assuming an albedo of
about 0.4 \citep{dra03}, we obtain a total host extinction of
0.071~mag. This is reasonably consistent with the host-only value of
$A_V=0.081$~mag ($R=3.1$) obtained via the Na~I~D observations of
\citet{gue04}. A similar result is obtainable from our study of
SN~2004et \citep{kot09}. For SN~2004et, the IS IR echo model yields an
absorption-only extinction of 0.081~mag implying a total host
extinction of about $A_V=0.20$~mag. High resolution spectra of Na~I~D
lines to SN~2004et gave a total (host plus Galaxy) $E(B-V)=0.41$~mag
\citep{zwi04}, or $A_V=1.27$~mag ($R_V=3.1$). Subtracting the
estimated Galactic contribution of $A_V=1.06$~mag \citep{sch98,mis07}
yields a host-only value of $A_V=0.21$~mag, in excellent agreement
with the IS IR echo-based value. \\
One possible objection to the IS IR echo interpretation of the cold
component of SN~2004dj is that the steady component of the flux around
24~$\mu$m could be due to cool IS dust in S96, insufficiently
corrected for by the 3500~K blackbody extrapolation (\S3.1). We
regard this as unlikely. As already indicated, \citet{mai04} argue
that the parent molecular cloud of S96 should have been dispersed by
stellar winds and SN explosions. In addition we have found that (a) a
single set of IS IR echo parameters provided a fair match {\it
throughout the $89-1393$~d} covered and (b) the derived dust column
density is comparable to that obtained from IS Na~I~D spectroscopy.\\
A second possible objection to the IS IR echo interpretation is that
the cold component is actually due to free-free emission. This is
discussed and dismissed in \S3.4.3.2. \\
We conclude that the cold component of SN~2004dj was due to an IS IR
echo. As originally suggested by \citet{bod80}, the study of IS IR
echoes from SNe can provide an independent method of measuring the IS
extinction in galaxies. \\
\paragraph{{\it Free-Free and Free-Bound Radiation\/}}
In their MIR study of SN~1987A, \citet{woo93} suggest that during the
$60-415$~d period, a significant proportion of the MIR flux longward
of $\sim$20~$\mu$m was due to free-free (ff) emission, with ff and
free-bound (fb) emission also contributing at shorter wavelengths. For
SN~2004dj we have looked at the possible contribution of fb-ff
radiation. It is particularly important to consider ff emission as
this rises in flux towards longer wavelengths, thus potentially
reducing or even dismissing an IS IR echo contribution. For
consistency in this analysis fb emission must also be included. For
the wavelengths covered this is strongest in the optical/NIR region.
Free-bound radiation has a strong ``sawtooth'' structure and the
extent to which this structure is undetectable in the optical/IR
continuum places a limit on its strength. For a particular
temperature this, in turn, constrains the strength of the ff emission
($F_{ff}/F_{fb}\propto \sqrt T / \nu^3$). \\
The fb-ff radiation was assumed to arise in a hydrogen envelope
centered on the supernova. \citet{koz98} modelled the time-dependent
behavior of the temperature and ionization of SN~1987A, including the
hydrogen envelope, for epochs after 200~d. Therefore, we made use of
their study to model the ff and fb emission from SN~2004dj at epoch
251~d and later. We adopted a hydrogen envelope of a few solar masses
and used their density profile:
\begin{equation}
\rho=9.1\times10^{-16}(t/(500~d))^{-3}(v/(2000~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi)^{-2}~g~cm^{-3}.
\end{equation}
By 251~d, the highest velocity of observable hydrogen in SN~2004dj was
no more than $\sim3100$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ (Table~\ref{tab8}). By 925~d the
maximum H velocity observed in SN~2004dj was no more than 1400~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi.
There were no optical or NIR spectra covering the final three MIR
epochs. We therefore assumed that the maximum hydrogen velocity at
these late epochs remained at 1400~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. The hydrogen inner limit is
less certain; a value of 1000~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi was adopted for all epochs. In any
event, the fb-ff luminosity from 500~d onwards was completely
negligible for any plausible hydrogen velocity limits (see below).\\
We estimated the fb-ff emission from the hydrogen within the velocity
limits. This was done using the escape probability formalism
\citep{ost89}, although in practice this was unnecessary as the
matches to the data always showed that the hydrogen was very optically
thin. The mass of the fb-ff-emitting hydrogen fell from
$\sim$4~M$_{\odot}$ at 251~d to 0.75~M$_{\odot}$ at 996, 1207 and
1393~d. Using modest extrapolation of figure~7 in \citet{koz98} to
the velocities observed in SN~2004dj, we deduced that the hydrogen at
the observed velocity limits stayed at a temperature of $T\sim$5000~K
up to 859~d and then declined to $\sim$2000~K by 1393~d. The
fractional ionization, $\chi_e$, was obtained from figure~9 of
\citeauthor{koz98} by interpolation. This indicated $\chi_e\sim0.02$
at 251~d, falling to $\chi_e\sim0.0005$ by 1393~d. At each epoch, the
$T$ and $\chi_e$ values were assumed to apply to all the hydrogen
within the velocity limits. The \citeauthor{koz98} study did not
extend to epochs as early as the earliest MIR observations (89, 106,
129~d) of SN~2004dj. We therefore adopted 6000~K as a plausible
temperature for these epochs and allowed the ionization to take the
largest value consistent with the overall match to the observed
continuum. Plausible values were obtained, viz: $\chi_e=0.003$ at
89~d rising to $\chi_e=0.005$ at 129~d. We used Gaunt factors
tabulated by \citet{hum88}. For the fb radiation we used the continuum
recombination coefficients of \citet{erc06}. \\
The hydrogen maximum velocity, temperature and free electron density
for each epoch are listed in Table~\ref{tab8}. The fb-ff luminosities
are listed in Table~\ref{tab9}, col.~4. It can be seen that by
500~d the fb-ff contribution to the total luminosity is small,
becoming increasingly negligible at later epochs The fb-ff components
for 89--281~d are plotted in Fig.\ref{fig18}. At 500~d the fb-ff flux
is too weak to appear on the 500~d plot which has been scaled to allow
easy comparison with the earlier epochs. A more comprehensive
estimation of the fb-ff flux, taking into account the temperature and
ionization gradients, is beyond the scope of this paper. \\
As noted in \S3.4.3.1, in the wavelength region ($\lambda >15~\mu$m)
where the cold component makes a significant contribution ($>20\%$) to
the total flux, the IS IR echo maintained a near-constant luminosity
(15--150~$\mu$m) of $\sim2.0\times10^{38}$~\ifmmode{\rm erg\,s^{-1}}\else\hbox{$\rm erg\,s^{-1}$}\fi between 89~d and
1393~d. Moreover, in setting the IS IR model to reproduce the
longwave excess at the latest epochs, when the fb-ff flux was
negligible, it was found that a satisfactory match to the longwave
($\lambda >15~\mu$m) excess was automatically achieved for {\it all}
epochs. Consequently, the fb-ff flux could make, at most, only a minor
contribution even at the earliest epochs. We conclude that any
free-free emission was too weak to account for the cold component at
any epoch. \\\\
\subsubsection{The Warm Component}
\paragraph{{\it Late-Time Optical and NIR Line Profiles\/}}
In \S2.4 we described the evolution of optical and NIR line profiles
in a number of species. In particular, we noted the development of a
sharp suppression in the red wing, suggesting dust formation causing
obscuration of the far side of the ejecta. Could this dust also be
responsible for the warm component at later epochs? In this and the
next sections we explore this possibility. Here we examine the
distributions of dust that might give rise to the late-time
optical/NIR line profiles. In \S3.4.4.2 we shall then test the
hypothesis that the same dust distributions are responsible for the
warm component. We modelled the line profiles of H$\alpha$,
Pa$\beta$, [O~I]~6300~\AA, [Fe~II]~7155~\AA\ and [Fe~II]~12567~\AA.
The period $461-925$~d was studied since this was mostly
overlapped by the $500-1393$~d during which the blackbody
analysis and more detailed studies of the MIR emission (\S3.4.4.2)
suggest that substantial quantities of dust were present in the
ejecta. We note (Table~\ref{tab6}) that the H$\alpha$ line widths
towards the end of the bright plateau phase greatly exceeded those
observed on 461~d and later, implying that there was a negligible
contribution of any light echo of the early phase profiles to the
late-time profiles.\\
We did also examine the blue asymmetry of the H$\alpha$ and
[Fe~II]~7155~\AA\ profiles at a much earlier epoch (283~d) but were
unable to achieve a satisfactory match using the model adopted for
later epochs. In any case, as already demonstrated in \S3.4.2, we
were able to account for the 251/81~d IR excess as thermal emission
from dust formed in the CDS. Moreover, \citet{chu05} and \citet{chu06}
successfully explained the H$\alpha$ profile up to $\sim$300~d by
invoking the emergence of an intrinsic asymmetric, bipolar core
(i.e. no dust involved). We conclude that there is no evidence for
the formation of new ejecta dust earlier than 461~d. \\
The line-profile red wings exhibit increasingly abrupt declines after
$\sim$1~year (Fig.~\ref{fig10}) suggesting the condensation of
attenuating dust. In contrast, the extended blue wings exhibit little
sign of developing suppression. This behavior points to the
attenuating dust lying at low line-of-sight velocities, such as a
face-on disk-like distribution centered on the SN center of mass.
Nevertheless, in our initial considerations of possible dust
configurations we included the case of a spherically symmetric dust
sphere. Our line-profile model comprises a homologously expanding
sphere of gas responsible for most of the observed line flux, with the
emission being attenuated by a dust zone lying concentrically with the
gas sphere. The dust zone is also assumed to participate in the
expansion. The gas emissivity is assumed to have a power-law
dependence with radius. \\
Several arrangements of dust were examined. These were (i) a uniform
opaque sphere (Fig.~\ref{fig16}, upper panels), (ii) a face-on opaque
ring, and (iii) a thin face-on disk (Fig.~\ref{fig16}, lower panels)
whose opacity was uniform or varied radially as an $r^{-\delta}$ power
law along the disk plane but was always uniform normal to the disk
plane. (For the disks, ``opacity'' refers to the value normal to the
disk plane.) For the radially-varying version of (iii), a power law
index of $\delta=1.9$ was chosen as typical of the power law indices
determined for the gas emission profiles. Configurations (i)--(iii)
were initially tested against the 895~d optical profiles since these
were (a) of the highest available resolution, and (b) corresponded to
a time by which the SN spectral continuum was overwhelmingly dominated
by thermal emission from dust. Configurations (i)--(ii) could not
reproduce the observed line profiles. This is illustrated in
Fig.~\ref{fig16} (upper panels) where we show two examples of the line
profile produced by an opaque, concentric dust sphere, compared with
the observed [O~I]~6300~\AA\ profile at 895~d. Only configuration
(iii) - the dust disk - was able to reproduce the observed profiles.
Within this configuration we examined cases where the dust
distribution (a) extended beyond the gas limit, (b) ceased abruptly at
a given velocity within the gas sphere, or (c) extended uniformly to a
given velocity, $v_{duni}$, and then declined as a power law to the
edge of the gas sphere. Case (a) failed to reproduce the observed
line profiles. Cases (b) and (c) are discussed below. \\
As described in \S2.4, by 895~d and 925~d, a ``secondary'', weaker,
redshifted peak had also appeared in the H$\alpha$ and [O~I]~6300~\AA\
profiles, suggesting that a fraction of the line flux originated in an
emission zone centered on the SN and having the geometry of an
expanding ring, jet or cone. The second peak had a redshift of only
+160 to +170~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. Given the preference for a thin face-on disk of
attenuating dust to account for the asymmetry of the line profiles, a
natural explanation for the low velocity of the second peak is that
its source was actually at a velocity comparable to that of the gas
sphere, but moving in a way which was roughly coplanar with the dust
disk. Consequently the low inclination angle led to the low observed
redshift in the second peak. The simplest geometry which accounts for
the second peak is a ring-like emission zone whose plane lay at a
small angle to the ``face-on'' plane. A cone or jet geometry is more
problematic. If we insisted on maintaining axial symmetry then a
cone/jet would have to be normal to the disk but with a mysteriously
low velocity. Alternatively, a cone/jet lying near the disk plane,
while possibly allowing a high intrinsic velocity, would have the
unattractive feature of breaking the axial symmetry. Therefore, to
account for the second peak, we added a thin, near-coplanar ring of
emitting gas to the line-profile model \citep[cf.][for the case of
SN~1998S]{ger00,fra05}. The ring was assumed to participate in the
overall homologous expansion of the gas. \\
We set the intrinsic ring velocity equal to the fastest moving gas
observed viz. hydrogen at 2400~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi as indicated by the H$\alpha$ blue
wing on 895 and 925~d, although higher velocities could have been
used. Consequently the ring emission is unattenuated by the dust.
For ease of computation, we retained a face-on disk ($i=0^{\circ}$).
Tilting the disk to become exactly coplanar with the ring would have
had only a small effect on the line profile. \\
In matching the model to the observed profiles the free parameters for
the gas are the maximum (outer) velocity of the gas sphere,
$v_{gmax}$, the gas emissivity scaling factor and power-law index
$\beta$, the inclination, $i_r$, of the ring component assuming an
intrinsic expansion velocity of 2400~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi, and the ring component
emissivity assumed uniform. In addition, the overall wavelength
positions of the observed profiles were allowed to vary by small
amounts (see below). For ease of comparison, the dust-disk dimensions
were also expressed as velocities within the homologous expansion.
Thus, for the case (b) dust disk the free parameters are the maximum
radial velocity of the dust disk, $v_{\rm dmax}$, the maximum
velocity, $v_{dth}$, perpendicular to the disk plane and the magnitude
of the dust radial density power-law index, $\delta$. Multiplication
of $v_{\rm dmax}$ and $2v_{\rm dth}$ by the epoch yields,
respectively, the disk radius and thickness at that time. For case
(c) $v_{\rm dmax}$ is replaced with $v_{\rm duni}$. We adopted
amorphous carbon as the grain material (see \S3.4.4.2) and a power law
grain-size distribution, index $m=3.5$ \citep{mat77}, with
$a_{(min)}=0.005~\mu$m and $a_{(max)}=0.05~\mu$m. The optical depth
normally through the disk was then calculated. For example, for case
(b) the optical depth for emission from gas lying behind the disk at
wavelength $\lambda$, epoch $t$ and radial velocity $v_d$,
$\tau_{\lambda}(t,v_d)$, is given by:
\begin{equation}
\tau_{\lambda}(t,v_d)=\frac{4}{3}\pi \rho\kappa_{\lambda}k\frac{1}{4-m}
[a^{4-m}_{(max)}-a^{4-m}_{(min)}]v_d^{-\delta}2v_{dth}t^{-2}
\end{equation}
where $k$ is proportional to the grain number density, at a fiducial
epoch, to the radial velocity and to the grain radius, $\rho$ is the
grain material density and $\kappa_{\lambda}$ is the mass absorption
coefficient of the grain material at wavelength $\lambda$. For
emission from gas lying within the disk the same formula is used but
replacing $2v_{\rm dth}$ with the velocity equivalent to the path from the
emission point to the near side of the disk. The disk thickness was
always $<20\%$ of the disk diameter and so edge-effects at the disk's
outer limit were assumed to be negligible. \\
The gas emission profiles (i.e., those for the attenuated sphere and
unattenuated ring) were individually convolved with the appropriate
slit PSF. The PSF FWHM values were 6.5, 4.2, 6.5, 1.25, and 1.75~\AA\
for 461, 467, 554, 895, and 925~d, respectively. The two components
were then summed to form the final model profile which was compared
with the observed profile. The profile model (case (b) disk) was
adjusted to match the observed profiles as follows. The parameters
$v_{\rm gmax}$, $\beta$ and the gas sphere emissivity scaling factor
were adjusted to provide a match to the blue wing of the
profile. Then, for a given value of $\delta$ (0 or 1.9), $v_{\rm dth}$
(equivalent to the disk half-thickness) together with the grain number
density scaler, $k$, were adjusted to reproduce the sharp-decline
section of the profile. The velocity $v_{\rm dmax}$ was varied to
reproduce the suppressed red wing. The gas ring emissivity and
inclination were then adjusted to reproduce the secondary peak. Some
iteration of all the model parameters was necessary to reach the final
profile match. For the profile model (case (c) disk) $v_{\rm dmax}$
was replaced with $v_{\rm duni}$, and $\delta$ was allowed to vary to
arbitrarily large steepness. \\
The offset of the sharp-decline section of the profiles from zero
velocity was only about $-100$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ after correction for the +221~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\
of the heliocentric velocity. Consequently, the match to the
sharp-decline section of the profile (i.e., determination of the disk
half-thickness) could be significantly affected by errors in
individual wavelength measurements or intrinsic variations in the
distributions of different gas species. Forcing a match in each
case by letting the wavelength position of each observed profile to
vary by small amounts, we were able to estimate the disk
half-thickness and its uncertainty. For a heliocentric velocity of
+221~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ we obtained a disk half-thickness equivalent to
$v_{\rm dth}=+112\pm18$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi. There was no significant variation in
$v_{\rm dth}$ during the period covered by the profile study. Note that a
change in the heliocentric velocity estimate would yield the same
change in the disk half-thickness velocity. \\
As indicated above, we first carried out model matches to the 895~d
profiles. It was found that the best case~(b) matches were obtained
when the disk was highly opaque (say, $\tau>5$) at all radial locations.
For case~(c), comparably good matches were achieved by setting the
uniform zone at a high opacity ($\tau>5$) and $\delta$ at values
steeper than $\sim10$ --- i.e., the dust density beyond the uniform zone
had to be extremely steep. Less steep declines tended to suppress the
visibility of the central minimum. We conclude that the dust disk was
highly opaque in the optical region but terminated abruptly at an
approximately fixed radius (see \S3.4.4.2). Given that this
condition was indicated by both cases (b) and (c), we abandoned the
more complicated case (c) model and completed the analysis for all the
461--925~d spectra using only case (b) --- i.e., a dust density which was
uniform or declined radially as $r^{-1.9}$ and which terminated
abruptly at $v_{\rm dmax}$. \\
The models for 895~d were initially adjusted to determine the minimum
opacity -- that is, the minimum dust mass that was needed to provide a
satisfactory match to the observed profile.
The dust masses were obtained from the following:
\begin{equation}
M_d= \frac{8}{3}\pi^2\rho\kappa_{\lambda}k\frac{1}{4-m}
[a^{4-m}_{(max)}-a^{4-m}_{(min)}]\frac{1}{2-\delta}v_{\rm dmax}^{2-\delta}
2v_{\rm dth}.
\end{equation}
At 895~d, with a uniform disk ($\delta=0$), the dust extended to
$v_{\rm dmax}=590\pm25~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi$ i.e. $r_{\rm
dmax}=(4.6\pm0.2)\times10^{15}$~cm. The disk half-thickness was
19\% of this. Simultaneous matches to all three profiles (H$\alpha$,
[O~I]~6300~\AA, [Fe~II]~7155~\AA) required a minimum $\tau$ of 13 at
6300~\AA, corresponding to a minimum dust mass of
$0.27\times10^{-4}$~M$_{\odot}$. Note that the high optical depth was
demanded by the very sharp decline, unresolved even at the
1.25~\AA\ resolution of 895~d. Other parameters are:
$r_{\rm gmax}=7.65\times10^{15}$~cm, $\beta=1.9$, (ring flux)/(total flux)
= 0.07 and ring tilt = $6^{\circ}$. \\
In using the dust-disk model to reproduce the contemporary MIR
continua (\S3.4.4.2) it was found that the opacities and dust masses for
all lines and epochs had to be somewhat higher than the minimum values
required to match the observed optical and NIR line profiles. For
example, for [O~I]~6300~\AA\ at 895~d and a uniform disk it was
necessary to set $\tau \ge 18\pm3$, corresponding to a minimum dust
mass of $(0.38\pm0.06)\times10^{-4}$~M$_{\odot}$. In the rest of this
subsection, therefore, we present and consider results for profile
matching incorporating dust-disk masses obtained by interpolating to
the profile epochs the values obtained from the MIR continuum
modelling. A sketch of the case (b) [O~I]~6300~\AA\ uniform
($\delta=0$) disk model at 895~d, is shown in Fig.~\ref{fig16} (lower
LH panel) together with an illustration (lower RH panel) of the model
match (solid red line) to the observed profile (blue). Also shown are
the intrinsic line-profile contributions from the attenuated gas
sphere (green) and unattenuated gas ring (cyan) components, as well as
the final profile (dotted red line) which would result if the
attenuating dust were removed. In Fig.~\ref{fig17} we show all the
individual matches to the line profiles for 461--925~d. In general,
good matches were achieved for epochs 554~d, 895~d, and 925~d. Poorer
matches were obtained at 461~d and 467~d. \\
On 895~d and 925~d the model did not reproduce the extended red wing
seen in the [Fe~II]~7155~\AA\ profile. This suggests an additional,
faster moving component of iron, and may be indicative of $^{56}$Ni
asymmetry or ``bullets'' in the initial explosion
\citep[e.g.][]{bur95}. At the earliest two epochs (461~d and 467~d)
the model matches are also poorer. Specifically, if the model was
adjusted to match the full suppression of the extreme red wing, then
it also over-attenuated the less redshifted portion of the red wing;
in other words, the steep red decline in the observed profile is
generally less pronounced at these earliest epochs. Alternatively,
matching to the less redshifted portion of the red wing meant that the
full suppression of the extreme red wing was not reproduced. These
points suggest that, during the 461--467~d period, dust formation was
less complete and was not yet fully opaque over the whole disk. We
therefore adjusted the models to reproduce the suppression of the
extreme red wing, but recognise that the derived dust mass lower
limits for 461--467~d may well be overestimated. \\
Between 554~d and 925~d the minimum dust mass increased from
$0.25\times10^{-4}$~M$_{\odot}$ to $0.4\times10^{-4}$~M$_{\odot}$.
During this time the radius of the dust disk appeared to actually {\it
decrease} slightly from $5.3\times10^{15}$~cm to
$4.6\times10^{15}$~cm. We explored the possibility that this was due
to a declining optical depth as the disk expanded --- i.e., that while
the dust disk continued to expand, the optically-thick/thin boundary
declined in radius. We found that the requirement of a steep, sharp
cutoff at the disk limit (see above) ruled out this explanation
implying that the decrease in the dust radius was real. This issue
will be considered further in \S3.4.4.2. \\
The above analysis was repeated with $\delta=1.9$. Similar disk radii
were obtained, but the minimum dust masses were $\sim\times$20 larger
than for $\delta=0$. This is as one would expect. To maintain the
line-profile matches with $\delta=1.9$ required the maintenance of a
high optical depth at the disk edge in the optical/NIR wavelength
range. If we demand that the optical depths be the same at $r_{\rm
dmax}$ for both $\delta=0$ and $\delta>0$ then it may be shown that
$M_d(\delta)/M_d(0)=2/(2-\delta)$, where $M_d(\delta)$ is the dust
mass for a given value of $\delta$ and $M_d(0)$ is the mass for
$\delta=0$. Thus, for $\delta=1.9$, $M_d(\delta)/M_d(0)=20$. We note
that this assumes that the power law continues to the center, which is
unlikely to be the case. Indeed for $\delta=2$ or more, the mass would
be infinite. A more plausible behavior would be that the density law
flattens toward the center. For example, if we assume that within
$0.05r_{\rm dmax}$ the dust density became uniform, with $\delta=1.9$
outside $0.05r_{\rm dmax}$, then the total minimum dust mass would
only be about a factor of 3 larger than for the totally uniform disk
case --- that is, about $10^{-4}$~M$_{\odot}$ \\
The model gas sphere expansion velocity remained at around 2400~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\
in hydrogen at all epochs. Similar velocities were obtained in
[Fe~II]~12567~\AA\ at 461~d and 554~d, and in [O~I]~6300~\AA\ and
[Fe~II]~7155~\AA\ at 467~d. By 895~d and 925~d the expansion
velocities in [O~I]~6300~\AA\ and [Fe~II]~7155~\AA\ had fallen to less
than 1000~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi, perhaps suggesting some stratification by element. At
554~d and earlier epochs, the ring flux was negligible. At the later
epochs, the ring inclination is $6^{\circ}$ (i.e., close to face-on).
The fractional contribution of the ring to the total observed flux
never exceeded 20\%. Indeed, were it not for the dust disk, it is
unlikely that the relatively weak ring emission would have been
detected since it would have been swamped by the emission from the gas
sphere (see Fig.~\ref{fig16}). We therefore regard the ring component
as an interesting but minor effect. \\
We conclude from the line-profile analysis that, in the 461--925~d
period, the mass of the dust disk and its radial extent can be small
enough to be consistent with an origin as newly-condensed ejecta dust.
In \S3.4.4.2 we shall show that the same dust disk can account for the
MIR continua.\\
\paragraph{{\it Later Phase IR Excess and Ejecta Dust Condensation\/}}
We have argued that, up to at least 281~d, the warm
component could be fully accounted for in terms of an IR echo from a
CDS. By 500~d the warm component luminosity exceeded the CDS
contribution by a factor of $\sim20$. This ``late IR excess'',
together with the steady shift of the peak emission of the SED to
longer wavelengths in the period $\sim250-1200$~d (Figs.~\ref{fig3},
\ref{fig5}) suggests the appearance of an additional population of
warm, but cooling dust. What is the location, distribution and
heating mechanism of this dust? Such IR emission can arise from (i)
heating, by a number of possible mechanisms (see below), of new dust
formed either in the ejecta, or in a late surge of dust growth within
the CDS or (ii) heating of pre-existing circumstellar dust by the
early-time SN luminosity (a conventional IR echo), or possibly the
forward shock travelling into undisturbed CSM beyond the CDS.
We can dismiss immediately hypothesis~(ii):\\
(a) An IR echo does not naturally produce the long delay (at least
$\sim300$~d) before the commencement of the warm component excess flux
rise. To account for this within an IR echo scenario, we would have
to invoke an {\it ad hoc} asymmetry in the CSM. \\
(b) IR-echo or forward-shock heating would not account for the
observed late-time red wing suppression in the spectral line profiles
discussed in \S3.4.4.1. In contrast, red-wing attenuation can easily be
produced by dust formation in the ejecta. New CDS dust could
conceivably also have produced such an effect provided that the dust
did not lie completely outside the optical/IR emission zone.\\
(c) Pre-existing CSM dust, whether heated by the SN luminosity or a
forward shock, would not produce steepening of the optical-NIR decline
rate simultaneously with the appearance of the late IR excess. In
contrast, new dust in the ejecta or CDS can be indicated by
optical-NIR steepening. In Fig.~\ref{fig4}, after $\sim$470~d the
$VRIJ$ light curves show weak evidence of steepening, providing a
minor additional argument against a CSM IR echo origin for the later
warm component, although the errors on the latest points are large.
We also note that an alternative explanation for such steepening could
be a faster-than-expected decrease in the $\gamma$-ray absorption
relative to that invoked by \citet{li93} for the radioactive energy
deposition in SN~1987A (Fig.~\ref{fig4}). \\
The above points, especially (a) and (b), leaves us with
hypothesis~(i) viz. that the late IR excess is due to emission from
new dust formed in either the ejecta or CDS. The later-era appearance
of the red wing suppression coincides roughly with the emergence of
the late IR excess, suggesting that the same dust could have been
responsible for both the optical/NIR attenuation effects and the rise
of the MIR emission. That this dust formed in the ejecta rather than
the CDS is indicated by the fact that the extent of the attenuating
dust derived from late-time line-profile analysis was comparable to
that of the MIR-emitting dust as derived from blackbody analysis; for
example, at 859~d the blackbody radius is $3.3\times10^{15}$~cm compared
with a profile-derived radius of $(4.4\pm0.2)\times10^{15}$~cm at
895~d. Moreover, the line-profile analysis demonstrates that the dust
was of high opacity in the optical/NIR, and distributed as a near
face-on disk lying concentrically with the SN center of
mass. Attenuation by a CDS could not have produced the late-time line
profile behavior. \\
In view of the above points, we shall now give detailed consideration
to the scenario where the warm component emission originated in
newly-formed ejecta dust. Such dust could be heated by a number of
mechanisms including radioactive decay, ionization freeze-out effect,
embedded pulsar or reverse-shock radiation. In the case of SN~2004dj
we can rule out radioactivity as the principal heating mechanism of
any putative new ejecta dust. Inspection of Table~\ref{tab7} shows
that, as early as 500~d, the warm component flux exceeded the total
deposited radioactive luminosity by a factor of 2 growing to a factor
of $\sim450$ by 1393~d. Indeed, by the latest epoch, the warm
component luminosity was still as high as $\sim4\times10^{38}$~\ifmmode{\rm erg\,s^{-1}}\else\hbox{$\rm erg\,s^{-1}$}\fi.
We also rule out the freeze-out effect as the main energy source
\citep{cla92,fra93} since, at $\sim1400$~d, its contribution would
only be $\sim5\times10^{36}$~ergs.\\
Power input from an embedded pulsar, via a pulsar wind nebula, is
another possibility \citep{woo89,che92}. \citeauthor{che92} predicted
that a distinctive feature of such a source would be the presence of
certain high ionization lines. The earliest epoch they studied was at
1500~d. The nearest of our optical/NIR spectra to this epoch are those
at 895~d and 925~d. Of the high ionization lines for which
\citeauthor{che92} make luminosity predictions, only
[O~III]~4959/5007~\AA\ was covered by these spectra. We co-added the
spectra at the two epochs and searched for the stronger
[O~III]~5007~\AA\ component. A broad, low S/N emission feature was
detected with a dereddened luminosity of
$(0.8\pm0.2)\times10^{36}$~\ifmmode{\rm erg\,s^{-1}}\else\hbox{$\rm erg\,s^{-1}$}\fi. This is about half the luminosity
predicted by \citeauthor{che92}. A serious difficulty is that the
redshift-corrected position of the feature lies about +400~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ from
where it would be if due to an [O~III]~5007~\AA\ line subject to the
same dust attenuation deduced in other lines (\S3.4.4.1). Even without
dust attenuation, the feature would still be +200~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi\ too far to the
red. This, together with the low S/N, leads us to conclude that there
is no persuasive evidence for the presence of pulsar-driven
high-ionization features in the latest spectra of SN~2004dj. This
leaves us with reverse-shock heating of the ejecta following the
ejecta-CSM collision as the most promising mechanism for the late-time
energy source \\
To test further the hypothesis that the same dust was responsible for
the late-time line-profile red-wing suppression and the MIR emission
we modelled the IR emission over a range of late-time epochs using the
same dust-disk configuration as was employed in the line-profile
analysis. The emission was derived following a similar procedure to
that used in the isothermal dust model by \citet{mei07} and
\citet{kot09}. The resulting flux was then added to the other
continuum components and the net model compared with the observed
continua.\\
We considered the thermal radiation from a warm, isothermal, face-on
disk of dust located symmetrically about the SN center of mass. We
shall refer to this as the IDDM (isothermal dust-disk model). The
disk radius is $r_{\rm dmax}$ and, as in \S3.4.4.1, the dust
number density declines as $r_d^{-\delta}$ where $\delta$ is set as 0
or 1.9. As mentioned above the disk thickness was $<20\%$ of the disk
diameter and so we judged that a thin-disk treatment would provide an
adequate means of estimating the flux from the disk (i.e., edge effects
are ignored). The observed IDDM flux, $dF_{\lambda}(t,r_d)$, at
wavelength $\lambda$ and time $t$ from an elemental ring lying between
radii $r_d$ and $r_d+dr_d$ is:
\begin{equation}
dF_{\lambda}(t,r_d)= 2 \pi r_d dr_d D^{-2} B_{\lambda}(t)
(1-exp(-\tau_{\lambda}(t,r_d)))
\end{equation}
where $D$ is the distance of the SN, $B_{\lambda}(t)$ is the Planck
function, and $\tau_{\lambda}(t,r_d)$ is the optical depth
perpendicularly through the disk. The total flux is then found by
integrating from $r_d=0$ to $r_d=r_{\rm dmax}$. The dust mass is
obtained from equation~(4). Amorphous carbon grains were assumed
(silicate grains are discussed below). As explained before, during
the $89-281$~d period ejecta dust formation was unlikely and in any
case the early IR excess was explainable as emission from CDS dust.
Therefore the IDDM was introduced at 500~d and used at all subsequent
epochs. By 652~d the CDS IR emission was negligible and so was not
included in the modelling of this or subsequent phases. We first
describe the results with $\delta=0$ in the IDDM. The matches
together with the observations are shown in Figs.~\ref{fig18} and
\ref{fig19}. The overall model parameters are listed in
Table~\ref{tab8}.\\
The three MIR observation epochs 500~d, 652~d and 859~d lay within the
timespan of the line-profile analysis. We therefore imposed the
constraint that the disk radius had to be consistent with those values
derived from the line profiles -- that is, $r_{\rm dmax} \approx
5\times10^{15}$~cm and $r_{\rm dth} = v_{\rm dth}\times t$ where
$v_{\rm dth}=+112$~\ifmmode{\rm km\,s^{-1}}\else\hbox{$\rm km\,s^{-1}$}\fi, with $k$ (and therefore $M_d$) being
constrained by the demand that the disk have a high optical depth in
the optical/NIR region (\S3.4.4.1). In practice, matching the IDDM to
the MIR continuum demanded optical depths which were higher than the
minimum values obtained from the line-profile analysis. The disk
parameters for the specific MIR epochs were set by linear
interpolation of the line-profile-derived values for $r_{\rm dmax}$
and $r_{\rm dth}$. The $k$ parameter could take values at or above
those set in the line-profile analysis. Only $T_d$ was a completely
free IDDM parameter. By 500~d the MIR continuum up to $\sim$20~$\mu$m
was overwhelmingly due to the warm dust disk, with the CDS component
yielding no more than $\sim$8\% of the total flux at any wavelength.
For $\lambda\gtrsim20~\mu$m, the IS IR echo dominated. Also from
500~d onwards the hot continuum and fb-ff contributions to the MIR
were negligible. In other words, from 500~d onwards, only the dusty
disk and IS IR echo made significant contributions to the total MIR
continuum. Consequently, and in order to show the MIR behavior in more
detail, the optical/NIR region is not shown in the plots for
652--1393~d (Fig.~\ref{fig19}). \\
In spite of the line-profile constraints on $r_{\rm dmax}$ and $r_{\rm
dth}$, for epochs 500~d, 652~d, and 859~d we were nevertheless able to
obtain fair matches to the observed continuum (Figs.~\ref{fig18},
\ref{fig19}). It was found that the matches required the dust to be
optically thick ($\tau>1$) in the optical/NIR region, but optically
thin in the MIR region. A blackbody spectrum matched to the shorter
MIR wavelengths overproduced the flux at longer wavelengths. This is
illustrated in Fig.~\ref{fig18} (500~d) and Fig.~\ref{fig19} as dotted
cyan lines. This restriction, together with the fixed values for
$r_{\rm dmax}$ and $r_{\rm dth}$ allowed us to obtain specific values,
{\it not} limits, for $T_d$, $\tau$ and $M_d$. At 500~d the ejecta
dust mass was $(0.22\pm0.02)\times10^{-4}$~M$_{\odot}$, with
$T_d=650\pm15$~K. The optical depths through the disk were
$\sim$10 in the $V$ band and $0.45\pm0.05$ at 10~$\mu$m. By 859~d the
dust mass had increased to $(0.33\pm0.05)\times10^{-4}$~M$_{\odot}$,
while $T_d$ fell to $570\pm15$~K. A particular value of this
500--859~d study is that it demonstrates that the same dust-disk
parameters can account for the line profiles {\it and} the MIR
continuum. This adds considerable weight to our contention that the
two disks are one and the same. Moreover specific values, not limits,
for the dust masses and temperatures were determined for the
500--859~d period. \\
As the SN aged between 500~d and 996~d it was found that the IDDM
steadily approached the blackbody case. Moreover, there were no line
profiles available after 925~d. Linear extrapolation to 996~d of the
line-profile-derived values for $r_{\rm dmax}$ gave $r_{\rm
dmax}\sim(4\pm1)\times10^{15}$~cm. The model match yielded dust with
a high optical depth in the optical/NIR but with $\tau\sim1$ in the
MIR. The dust mass was $M_d \approx 0.5\times10^{-4}$~M$_{\odot}$ and
$T_d = 520\pm30$~K. \\
By epochs 1207~d and 1393~d we found that the MIR continuum was best
reproduced by allowing a continuing increase in the IDDM optical depth
such that the disk was optically thick at all observed wavelengths
(see Table~\ref{tab8}). Indeed, given the uncertainties in the
observed fluxes, we cannot rule out a totally opaque disk at all
wavelengths covered. Consequently, the IDDM could provide only dust
mass lower limits of $1.0\times10^{-4}$~M$_{\odot}$ at 1207~d and
$1.5\times10^{-4}$~M$_{\odot}$ at 1393~d. The 1207~d and 1393~d plots
shown in Fig.~\ref{fig19} are with $r_{\rm dmax}$ and $T_d$ set at the
limiting values (Table~\ref{tab8}). \\
In Table~\ref{tab9} we show the luminosities of the IDDM ($\delta=0$)
components compared with the radioactive deposition power of
0.0095\,M$_{\odot}$ of $^{56}$Ni, including $^{56}$Ni decay. The IS
IR echo component is excluded since it is powered primarily by the SN
peak luminosity. The luminosity of the thin disk, $L_{IDDM}$, is
approximated by:
\begin{equation}
L_{IDDM} \approx 2 \pi R_d^2 \pi B_{\nu} (1-exp(-2\tau_{\nu})).
\end{equation}
As already indicated by the BLC analysis (\S3.2) the post-30~d
evolution of SN~2004dj falls into three phases. At 89~d, (about
half-way down the plateau-edge) the luminosity is still dominated by
the shock-ionized ejecta, with radioactive decay contributing a small
proportion of the total. There is then a short period just after the
end of the plateau when radioactive decay deposition dominated the
luminosity. This is supported by the fact that at 129~d, the
radioactive luminosity, $L_{rad}$, is highly similar to the total
continuum model luminosity, $L_{total}$. At the nebular phases of
251~d and 281~d, the small excess in $L_{rad}$, relative to
$L_{total}$, presumably went into powering the line emission which is
not included in the model. By 500~d, in spite of the exclusion of
much of the line luminosity from the model, $L_{total}$ is $\times 2.5
L_{rad}$, rising to $\times 200 L_{rad}$ by 1393~d. As argued above,
the most plausible additional power source available at this stage is
a reverse shock. \\
Model matches were also carried out with $\delta=1.9$ in the IDDM
component. Similar dust parameters were found to those obtained for
$\delta=0$. The only significant difference was that the dust masses
were $\times3$ larger, assuming a uniform density distribution within
the inner $0.05r_{\rm dmax}$ (cf. \S3.4.4.1). These larger values are
due to the growing proportion of dust mass concentrated in the
optically-thick region of the disk. \\
We also considered a dusty disk of warm silicate grains, with the
radii and thickness determined by the line-profile analysis as before.
The demand that the disk should be of optical depth in the optical/NIR
up to its edge meant that the replacement of amorphous carbon dust
with silicate dust had no effect on these dimensions. The problem
with silicate dust continuum matching is the absence of the
$8-14~\mu$m silicate feature in the observed continua during the
period when the dust was optically thin in the MIR, up to 996~d.
Attempts to suppress the feature in the IDDM by increasing the dust
mass and hence the optical depth yielded a continuum that was too
bright. Only for 1207~d and 1393~d was silicate dust able to reproduce
the observed continuum. This is not surprising since, by these two
epochs, the warm ejecta dust was close to being opaque over the MIR
range observed. Between 500~d and 996~d the proportion of silicate
grains by mass could have been no more than 20\%, and was usually
significantly less than this. Indeed, the data were always consistent
with there being no silicate grains at all. We also note the absence
of the SiO feature at $7.5-9.3$~$\mu$m. SiO formation is a necessary
step on the way to silicate grains \citep{tod01,noz03}.\\
The points made in the previous paragraph argue against a substantial
amount of silicate dust in the ejecta of SN~2004dj. The absence of
silicate grains is consistent with the presence of strong CO
fundamental and first overtone emission in the period $\sim$100~d to
$300-500$~d. A high C/O ratio in the ejecta could result in most of
the oxygen being absorbed as CO leaving behind an excess of carbon to
provide carbon grains, but little oxygen to provide silicate
grains. However, in the SN environment, the net grain population can
be affected by factors in addition to the C/O ratio. These include
molecule destruction by high-energy electrons, charge transfer
reactions and ejecta density \citep{liu96,noz03,den06}. Nevertheless,
we conclude that the dust grains in SN~2004dj were predominantly
composed of non-silicate material.\\
The success in reproducing the MIR continua using the same dust disk
as was invoked to explain contemporary line profiles tends to support
the curious result from \S3.4.4.1 that, rather than expanding, the
radius of the dust disk actually shrunk by a small amount. Indeed, if
we include the disk radii for 1207~d and 1393~d derived from the IDDM
matches (Table~\ref{tab8}, col.~7) we find a shrinkage of 27\% since
500~d. In SN~2004et \citep{kot09} it was found that the dust radius
remained roughly constant. The explanation offered was that the dust
was contained within an optically-thin cloud of optically-thick,
pressure-confined clumps. But the availability of late-time
optical/NIR spectra for SN~2004dj and the profile modelling presented
here shows that a clumping explanation for the non-expanding dust-disk
radius would not work. We note that, during the 500--859~d period in
SN~2004dj, the product of the blue-wing, half-maximum (BHM) velocities
and epoch for the line profiles, $R_{BHM}$, yields a roughly constant
value of $\sim 4\times10^{15}$~cm (see Table \ref{tab6}) - similar to
that obtained for the dust-disk radius. This coincidence may imply
that the extent of both the dust and the bulk of the ejecta gas
emission was physically constrained within a radius of $\sim
5\times10^{15}$~cm.\\
We suggest that the apparent shrinkage of the dust-disk radius may
have been due to the presence and inward motion of the reverse shock
such that the shock position defined the disk radius. Dust formation
could have continued within the disk with the dust taking part in the
overall ejecta outflow, but as the ejecta passed through the reverse
shock it would have been destroyed. There would have been a net
increase in dust mass with time, as observed, if the dust growth rate
within the disk exceeded the destruction rate. A possible difficulty
with this scenario is that the dust mass appeared to continue to grow
right up to the final epoch at 1393~d. This is rather later than dust
formation studies suggest. \citet{tod01} find that all grain
condensation would be complete by 800~days. Moreover, both
\citet{tod01} and \citet{noz03} find carbon dust condensation is
complete in not much more than one year. An alternative explanation,
therefore, might be that the dust mass in SN~2004dj did not increase
after $\sim$400~d, but rather that the density gradient of the outer
region of the disk was actually steeper than $r^{-2}$. Thus, as the
ejecta expanded, the dust density at a fixed location in the outer
region would have grown, yielding an apparently higher total dust mass
--- i.e., more dust emerged from the optically-thick inner regions. \\
Radiation from the reverse shock could have been primarily responsible
for heating the dust. As more and more ejecta passed through, the
shock and its radiation would have weakened causing the dust to cool,
as observed. The reverse shock may also have been responsible for the
approximate constancy of the radius of the line emitting gas sphere.
Further examination of the reverse-shocked dust-disk hypothesis is
beyond the scope of this paper. We note that in a recent
observational study of Cassiopeia~A \citet{del10} deduce a flattened
ejecta distribution or ``thick disk'' containing all the ejecta
structures. They also deduce the occurrence of a roughly spherical
reverse shock. \\
Intrinsic axial asymmetry in the form of a bipolar jet biased towards
the observer has been invoked by \citet{chu05} to account for the
H$\alpha$ profile in SN~2004dj up to about 1~year. We note that the
jet makes an angle of only $\sim15^\circ$ to the normal to the dust
disk plane derived from our line-profile analysis, perhaps indicating
a physical connection. However, line-profile asymmetry in the earlier
(pre-$\sim$1~year) nebular spectra of other CCSNe also tend to be
blue-biased. This implies that, in general, an intrinsic bias towards
the observer cannot be the explanation for such line blueshifts
\citep[cf.][]{mil10}. At later (post-$\sim$1~year) nebular epochs,
line blueshifts are also often seen
\citep{luc89,spy90,tur93,fes99,ger00,leo00,fas02,elm03,poz04}
and these are usually attributed to attenuation by newly-formed dust
in the ejecta or CDS. Our dust-disk model invokes this scenario.
This has the advantage of explaining the red-wing suppression without
invoking an intrinsic observer-biased axial asymmetry in the SN.
Moreover, our model uses the same dust disk to simultaneously account
for the line-profile attenuation {\it and} the MIR emission. We
consider it unlikely that the \citeauthor{chu05} model can provide a
superior alternative explanation for the line profiles presented and
analysed in \S3.4.4.1. Further examination of the relationship between
our model and that of \citeauthor{chu05} is beyond the scope of this
paper. \\
We conclude that dust formation in the ejecta of SN~2004dj had
commenced, and may even have been completed, by 500~d with a
near-face-on, disk-like distribution. {\it This dust was responsible
both for the late-time line-profile red-wing suppression and the bulk
of the MIR luminosity up to $\sim20 \mu$m.} The main source of dust
heating was probably reverse-shock radiation. Assuming $\delta=0$,
the dust mass was at least $10^{-4}$~M$_{\odot}$. For $\delta=1.9$,
the lower limit rises to about $3\times 10^{-4}$~M$_{\odot}$. The
value of $\delta$ is poorly constrained. We reject silicates as the
grain material. \\
\section{Conclusions}
We have presented optical, NIR, and MIR observations of the Type~IIP
SN~2004dj. The combination of wavelength and temporal coverage
achieved makes this SN one of the most closely studied of such
events. In the present work we have analyzed the SN continuum over a
period spanning 89--1393~d, augmented by a line-profile analysis over
461--925~d. Our conclusions are as follows. \\
(1) A mass of $0.0095\pm0.002$~M$_\odot$ of $^{56}$Ni was ejected
from SN~2004dj, which is less than that reported by other authors. The
period during which the radioactive tail dominated the bolometric
light curve lasted for an unusually short period of only
$\sim$35~d. Subsequently, a different energy source dominated;
we suggest reverse-shock heating. \\
(2) At early times the optical/NIR (``hot'') part of the continuum
provided most of the SN luminosity. This emission is attributed to hot,
optically-thick ejecta gas. At later epochs the optical/NIR luminosity
was increasingly due to nebular emission, and formed only a minor,
ultimately negligible, proportion of the BLC. \\
(3) At both early and late times, the long-wave portion of the MIR
continuum (the ``cold'' component) was primarily due to an IS IR echo.
Free-free radiation made only a minor contribution. As originally
suggested by \citet{bod80}, the analysis of SN IR echoes may provide a
useful way of studying IS dust in nearby galaxies. \\
\noindent
Subsequent to the submission of this paper, \citet{sza11} reported
their findings on SN 2004dj, making use of some of the data presented
in this work. They dismiss the IS IR echo, and argue against
pre-existing dust since the estimated extinction of SN 2004dj was
lower than that of other SNe. We do not concur. We have shown that an
extinction of only Av=0.02 is all that is required to account for an
adequate IR echo. \citet{sza11} also argue that the early UV/X-ray
flash would create a dust-free cavity of up to $10^{17}$ cm and that
OB-stars would have expelled the ISM/dust from the cluster. In fact
most of the IS IR echo longward of $\sim10~\mu$m comes from dust lying
at considerably more than 1~parsec. We therefore strongly favour a
scenario whereby the long-wave component was primarily due to an IS IR
echo.\\
(4) The early-time NIR/MIR (``warm'') component was probably due to
thermal emission from non-silicate dust formed in the CDS. The CDS
dust growth began at about 50~d, reached 90\% of maximum by 165~d, and
approached a maximum of $0.33\times10^{-5}$~M$_{\odot}$; the dust mass
produced in this way was small. Heating of the dust by the
contemporary optical-NIR BLC completely accounts for the strength and
evolution of the early-time warm component. \citet{sza11} did not
present contemporary optical or NIR data and consequently did not
identify the early warm component. \\
(5) The late-time warm component was dominated by the luminosity of
newly-formed, non-silicate dust in the ejecta. The dust growth
commenced at some time between 281~d and 461~d. The same dust was
responsible for the late-time red-wing attenuation of optical and NIR
spectral line profiles. The dust was distributed as a nearly face-on
disk, within a spherical cloud of emitting gas. During 500--996~d the
disk was effectively opaque in the optical/NIR region, but was
optically thin at longer wavelengths. The dust mass appeared to grow
during this period, attaining $(0.5\pm0.1)\times10^{-4}$~M$_{\odot}$
by 996~d for a uniform density disk, or a few times more than this for
an $r^{-1.9}$ gradient outside $0.05r_{\rm dmax}$ and uniform within.
However, it may be that the dust mass ``growth'' was really due to the
emergence of previously formed (i.e., pre-500~d) dust from
optically-thick regions of a disk with an even steeper density
gradient. For the latest two epochs (1207~d and 1393~d) the dust was
optically thick at all wavelengths and only lower limits could be
obtained for a given gradient --- for example, $>10^{-4}$~M$_{\odot}$
for a flat gradient and a factor of 3 higher limit for an $r^{-1.9}$
gradient outside $0.05r_{\rm dmax}$ and uniform within. This is
broadly in agreement with \citet{sza11}. For a smooth distribution,
they found a dust mass of $\sim10^{-5}-10^{-4}$~M$_{\odot}$ though,
with clumping, up to $\sim10^{-3}$~M$_{\odot}$ would also be
possible. \\
(6) Rather than expanding, the dust-disk radius appeared to
slowly shrink. This may have been due to the dust extent being
confined by the reverse shock, which also heated the grains
radiatively. \\
(7) While the latest epochs provide only lower limits to the
mass of dust produced by SN~2004dj, these limits are at least a factor
of 100 below the 0.1~M$_{\odot}$ of grains per SN required to account
for the dust observed at high redshifts. Moreover, measurements as late
as 996~d yield actual dust masses of only $\sim10^{-4}$~M$_{\odot}$.
While not completely ruling out the possibility that typical CCSN
ejecta are major contributors to cosmic dust production, this study
does suggest that, at least for SN~2004dj, the dust-mass
production was small.
\acknowledgements We thank J. Vink\'o for providing us with digitized
versions of his optical spectra. The work presented here is based on
observations made with the {\it Spitzer Space Telescope}, the
W. M. Keck Observatory, the William Herschel Telescope (WHT), and the
2.4~m Hiltner telescope of the MDM Observatory. The {\it Spitzer
Space Telescope} is operated by the Jet Propulsion Laboratory,
California Institute of Technology, under a contract with NASA. The
W. M. Keck Observatory is operated as a scientific partnership among
the California Institute of Technology, the University of California,
and NASA; it was made possible by the generous financial support of
the W. M. Keck Foundation. We wish to extend special gratitude to
those of Hawaiian ancestry on whose sacred mountain we are privileged
to be guests. The WHT is operated on the island of La Palma by the
Isaac Newton Group in the Spanish Observatorio del Roque de los
Muchachos of the Instituto de Astrof\'isica de Canarias. Financial
support for this research was provided by NASA through an award issued
by JPL/Caltech (specifically grant number 1322321 in the case of
A.V.F.). A.V.F. gratefully acknowledges additional support from NSF
grant AST-0908886 and the TABASGO Foundation. P.A.H. was supported by
NSF grants AST-1008962 and 0708855. S.M. acknowledges support from
the Academy of Finland (project 8120503). J.S. is a Royal Swedish
Academy of Sciences Research Fellow supported by a grant from the Knut
and Alice Wallenberg Foundation. J.C.W gratefully acknowledges
support from NSF grant AST-0707769. The Dark Cosmology Centre is
funded by the Danish National Research Foundation.
\newpage
|
\section{Introduction}
The study of the two-body problem of variable masses began practically with the work of \citet{gyl}, who wrote the differential equations of motion for the problem. The first integrable case to Gylden's equation was given by \citet{mes}, for a specific mass variation law. This mass variation law, and its following generalization \citep{mes2}, are known as \textit{Mestschersky laws}. After Mestschersky's contribution, the physical meaning of the problem became clear and it is known as \textit{Gylden-Mestschersky problem}.
\citet{jeans}, by studying the orbits of binary stars, found a more general mass variation law that was based on the relation between mass and luminosity of the stars presented by \citeauthor{eddington} in the same year. Mestschersky's laws are special cases of Jeans' law.
\citet{gel} considered a different mass variation law. \cite{berkovic} investigated the problem using a differential equation transformation method. Also a number of approximate analytic solution were found, e.g., \cite{prieto}. \cite{luk3} studied the particular problem where the total mass is constant, which can be applied to conservative mass transfer in close binary systems \citep{luk4}.
The Gylden-Mestschersky problem can also be generalized to include the restricted three-body problem.
In this approach, it is assumed that the two heavier bodies have their motion determined by the Gylden-Mestschersky equations. Thus, one have to deal only with the motion of the third body, which does not affect the main bodies motion. It was shown \citep{xx} that this problem presents particular solutions that are analogous to the stationary solutions of the classical problem of constant masses: the three collinear solutions $L_{1}$ to $L_{3}$ and the two triangular solutions $L_{4}$ and $L_{5}$. Many other particular solutions have also been found \citep{bekov,bekov2,luk,luk2}. Since then, further characteristics of this problem have been studied, for example, \citet{luk5,singh2}.
Besides the Gylden-Mestschersky problem, there are many different cases of two-body problems with variable mass \citep{razb}. These can be classified according to the presence or not of reactive forces, to the variation of the mass of just one or both of the bodies, to whether the bodies move in an inertial frame or not and so on.
In this paper, we consider the specific case where the three bodies move in an inertial frame, within a static atmosphere, from which they absorb mass or to which they lose mass. We search for the five particular solutions analogous to the classic case ones, $L_1$ to $L_5$.
In the restricted three-body problem, the motion of the two primary body is determined \textit{a priori}. In the papers mentioned above, this motion is determined by the Gylden-Mestschersky problem \citep{mes2}, whereas in the present paper, the primary bodies move in a static atmosphere, from which they absorb or lose mass. Therefore, reactive forces are present. Fact not considered in the Gylden-Mestschersky problem.
\section{Particular solutions}
The equation of motion for a body whose mass depend on time, $m\left(t\right)$, see for instance \citet{som}, is
\begin{eqnarray}
\textbf{F}=m\dot{\textbf{v}}+\left(\textbf{v}-\textbf{u}\right)\dot{m},\label{pbm1}
\end{eqnarray}
where $\textbf{F}$ is the sum of all the forces acting on it and $\textbf{v}$ is its velocity, both measured in an inertial coordinate system. Also, $\textbf{u}$ is the velocity of the center of mass of the absorbed mass immediately before its union with the body (or of the ejected mass immediately after its ejection). The overdot denotes, as usual, derivation with respect to the time variable.
In the present context, there are two special cases of equation (\ref{pbm1}) to be considered. The first one is when the mass is ejected with the same velocity of the body at any moment ($\textbf{u}=\textbf{v}$), i.e. mass ejection does not produce reactive forces. This case can be used to study the motion of a body ejecting mass isotropically (or radiating energy, since the total reactive momentum would be zero). In this case Eq. (\ref{pbm1}) takes the traditional form $\textbf{F}=m\dot{\textbf{v}}$. This case includes the Gylden-Mestschersky problem.
The second case, which we considered in the present paper, is the one where $\textbf{u}=\textbf{0}$, i.e. the particles are at rest in an inertial coordinate system. This case can be used to study the orbits of a star moving through a static atmosphere, whose particles attach or detach to the star as it moves. In this case, equation (\ref{pbm1}) reduces to
\begin{eqnarray}
\textbf{F}=m\dot{\textbf{v}}+\dot{m}\textbf{v}=\frac{d}{dt}\left(m\textbf{v}\right).\label{pbm2}
\end{eqnarray}
According to (\ref{pbm2}), the equation of motion for a problem of two bodies of varying masses $m_{1}\left(t\right)$ and $m_{2}\left(t\right)$, exchanging mass with a static atmosphere surrounding them, is
\begin{eqnarray}
&&\frac{d}{dt}\left(m_{1}\dot{\textbf{r}}_{1}\right)=-\frac{Gm_{1}m_{2}}{r^{3}}\textbf{r},\nonumber\\
&&\frac{d}{dt}\left(m_{2}\dot{\textbf{r}}_{2}\right)=\frac{Gm_{1}m_{2}}{r^{3}}\textbf{r},\label{pbm3}
\end{eqnarray}
where $\textbf{r}_{1}$ and $\textbf{r}_{2}$ are the position vectors of the two bodies, $\textbf{r}=\textbf{r}_{1}-\textbf{r}_{2}$, $r=|\textbf{r}|$ and $G$ is the gravitational constant. We note that the atmosphere around the bodies does not cause any drag forces on them. The only appreciable effect of this atmosphere is to work as a source or sink of mass for the bodies. From Eqs. (\ref{pbm3}) we get,
\begin{eqnarray}
\frac{d}{dt}\left(m_{1}\dot{\textbf{r}}_{1}+m_{2}\dot{\textbf{r}}_{2}\right)=0.\label{pbm3.5}
\end{eqnarray}
The quantity in parenthesis can be set equal to zero in an appropriated inertial frame. It follows, $\dot{\textbf{r}}_{1}=\left(m_{2}/M\right)\dot{\textbf{r}}$, where $M=m_{1}+m_{2}$, can be used to cast equation (\ref{pbm3}) as
\begin{eqnarray}
\frac{d}{dt}\left(\mu \dot{\textbf{r}}\right)=-\frac{G\mu M}{r^{3}}\textbf{r},\label{pbm4}
\end{eqnarray}
where $\mu\left(t\right)=m_{1}m_{2}/M$ is the reduced mass of the problem. In the classification table of the different two-body problems with variable masses by \citet{razb}, the present physical problem is listed as number 3. This problem also coincides mathematically with problems 14 and 15 (the later, known as Gelfgat-Omarov problem). Eq. (\ref{pbm4}) can be decomposed, in polar coordinates $\left(r,\theta\right)$, in the relations:
\begin{eqnarray}
\frac{d}{dt}\left(\mu\dot{r}\right)&=&\frac{k^{2}}{\mu r^{3}}-\frac{GM\mu}{r^{2}},\label{pbm5}\\
\mu r^{2}\dot{\theta}&=&k.\label{pbm6}
\end{eqnarray}
The coordinate $r$ is the distance between the primary bodies, $\theta$ is the angle between the line passing through their centers and a fixed line in the plane of motion, and $\dot{\theta}$ is the corresponding angular velocity. The constant $k$ is the total angular momentum of the system. Equation (\ref{pbm3.5}), together with the definition of the center of mass position vector, $M\textbf{R}_{cm}=m_{1}\textbf{r}_{1}+m_{2}\textbf{r}_{2}$, also yields the relation for the time dependence of $\textbf{R}_{cm}$
\begin{eqnarray}
\dot{\textbf{R}}_{cm}=\frac{d}{dt}\left(\frac{m_{1}}{M}\right)\textbf{r}\label{pbm7}
\end{eqnarray}
which does not perform, in general, an inertial motion.
We shall restrict ourselves to the case where the masses $m_{1}$ and $m_{2}$ vary arbitrarily, but their ratio remains constant. In this case the center of mass of the system moves inertially.
Following \citet{luk2}, the time dependence of the masses is described by a positive function $u\left(t\right)$:
\begin{eqnarray}
m_{1}=m_{10}u\left(t\right), m_{2}=m_{20}u\left(t\right),
M=M_{0}u\left(t\right).
\end{eqnarray}
Furthermore, we have:
\begin{equation}
\mu=\mu_{0}u\left(t\right).
\end{equation}
\label{pbm8}
The constants $m_{10}$, $m_{20}$, $M_{0}$ and $\mu_{0}$ are all positive, with $M_{0}=m_{10}+m_{20}$ and $\mu_{0}=m_{10}m_{20}/M_{0}$.
Now we introduce a barycentric rectangular coordinate system that rotates comoving with the main bodies, with angular velocity $\boldsymbol\omega=\omega\hat{z}$. The main bodies move in the $xy$ plane, in such a way the $x$ axis always passes through the center of the main bodies, at $x=x_{1}$ and $x=x_{2}$.
In such a coordinate system the equations of motion for the third (lighter) body are,
\begin{eqnarray}
\frac{d}{dt}\left(m\dot{x}\right)&=&F_{x}+\dot{m}\omega y+2m\omega\dot{y}+m\dot{\omega}y+m\omega^{2}x,\nonumber\\
\frac{d}{dt}\left(m\dot{y}\right)&=&F_{y}-\dot{m}\omega x-2m\omega\dot{x}-m\dot{\omega}x+m\omega^{2}y,\nonumber\\
\frac{d}{dt}\left(m\dot{z}\right)&=&F_{z},\label{pbm9}
\end{eqnarray}
where $m\left(t\right)$ is the third body mass and $F_{x}$, $F_{y}$ and $F_{z}$ are the $x$, $y$ and $z$ components of the gravitational forces $\textbf{F}$, due to $m_{1}$ and $m_{2}$. It is instructive to
compare these two last equations with the corresponding equations for the motion of the third body in references like \citeauthor{bekov} (\citeyear{bekov} and \citeyear{bekov2}) and \citeauthor{luk} (\citeyear{luk} and \citeyear{luk2}). Equations (\ref{pbm9}) have the additional terms $\dot{m}\omega y$ and $-\dot{m}\omega x$. These terms arise from the fact that the motion of the third body obeys equation (\ref{pbm4}). Equations (\ref{pbm9}) also appear in \citet{bekov3}.
Following \citet{luk2}, we first write the distance between the primary bodies as $r=r_{0}R\left(t\right)$, where $R\left(t\right)$ is a positive function, so we have $\omega=\omega_{0}/uR^{2}$ (equations (\ref{pbm6}) and (\ref{pbm8})). Then, we define the mass parameter $\nu=m_{20}/M_{0}$ ($0<\nu\leq1/2$, without losing generality), and choose units such that $r_{0}=1$, $M_{0}=1$ and $\omega_{0}=1$. Therefore,
\begin{eqnarray}
x_{1}&=&-\nu R\left(t\right),\nonumber\\
x_{2}&=&\left(1-\nu\right)R\left(t\right).\nonumber
\end{eqnarray}
Furthermore, we shall assume that the mass of the third body also varies as $m=m_{0}u\left(t\right)$, where $m_{0}$ is a positive constant. In the new units described above, the equations of motion (\ref{pbm9}) reduce to,
\begin{eqnarray}
\frac{d}{dt}\left(u\dot{x}\right)&=&-\frac{Gu^{2}}{r_{1}^{3}}\left(1-\nu\right)\left(x+\nu R\right)\nonumber\\
&&-\frac{Gu^{2}}{r_{2}^{3}}\nu\left(x-\left(1-\nu\right)R\right)+\nonumber\\
&&+\frac{2\dot{y}}{R^{2}}-\frac{2y\dot{R}}{R^{3}}+\frac{x}{uR^{4}},\nonumber\\
\frac{d}{dt}\left(u\dot{y}\right)&=&-\frac{Gu^{2}}{r_{1}^{3}}\left(1-\nu\right)
-\frac{Gu^{2}}{r_{2}^{3}}\nu y\nonumber\\
&&-\frac{2\dot{x}}{R^{2}}+\frac{2y\dot{R}}{R^{3}}+\frac{y}{uR^{4}},\nonumber\\
\frac{d}{dt}\left(u\dot{z}\right)&=&-\frac{Gu^{2}}{r_{1}^{3}}\left(1-\nu\right)z-\frac{Gu^{2}}{r_{2}^{3}}\nu z,\label{pbm10}
\end{eqnarray}
where $r_{1}=\sqrt{\left(x+\nu R\right)^{2}+y^{2}+z^{2}}$ and\\ $r_{2}=\sqrt{\left(x-\left(1-\nu\right) R\right)^{2}+y^{2}+z^{2}}$. The constant $m_{0}$ cancels and it does not appear in equations (\ref{pbm10}).
We look for stationary solutions to equations (\ref{pbm10}) of the form $x=\xi R\left(t\right)$, $y=\eta R\left(t\right)$ and $z=\zeta R\left(t\right)$, where $\xi$, $\eta$ and $\zeta$ are constants. By substituting these three relations for $x$, $y$ and $z$ into equations (\ref{pbm10}), we get
\begin{eqnarray}
\xi\frac{d}{dt}\left(u\dot{R}\right)&=&-\frac{Gu^{2}}{R^{2}}\frac{\left(1-\nu\right)\left(\xi+\nu\right)}{\rho_{1}^{3}}\nonumber\\
&&-\frac{Gu^{2}}{R^{2}}\frac{\nu\left(\xi+\nu-1\right)}{\rho_{2}^{3}}+\frac{\xi}{uR^{3}},\nonumber\\
\eta\frac{d}{dt}\left(u\dot{R}\right)&=&-\frac{Gu^{2}}{R^{2}}\frac{\left(1-\nu\right)\eta}{\rho_{1}^{3}}\nonumber\\
&&-\frac{Gu^{2}}{R^{2}}\frac{\nu\eta}{\rho_{2}^{3}}+\frac{\eta}{uR^{3}},\nonumber\\
\zeta\frac{d}{dt}\left(u\dot{R}\right)&=&-\frac{Gu^{2}}{R^{2}}\frac{\left(1-\nu\right)\zeta}{\rho_{1}^{3}}-\frac{Gu^{2}}{R^{2}}\frac{\nu\zeta}{\rho_{2}^{3}},\label{pbm11}
\end{eqnarray}
where $\rho_{1}=\sqrt{\left(\xi+\nu\right)^{2}+\eta^{2}+\zeta^{2}}$ and\\ $\rho_{2}=\sqrt{\left(\xi+\nu-1\right)^{2}+\eta^{2}+\zeta^{2}}.$
The relative motion of $m_{1}$ and $m_{2}$ is given by $R\left(t\right)$. For this class of solutions the geometrical configuration of the system is similar at any moment.
In order to find the time derivative of $u\dot{R}$, Eq. (\ref{pbm5}) is written in the new units as,
\begin{eqnarray}
\frac{d}{dt}\left(u\dot{R}\right)=\frac{1}{uR^{3}}-\frac{Gu^{2}}{R^{2}}.\label{pbm12}
\end{eqnarray}
Finally, we obtain the form of equations (\ref{pbm11}) that allows us to find $\xi$, $\eta$ and $\zeta$
\begin{eqnarray}
\xi-\frac{\left(1-\nu\right)\left(\xi+\nu\right)}{\rho_{1}^{3}}-\frac{\nu\left(\xi+\nu-1\right)}{\rho_{2}^{3}}&=&0, \nonumber\\
\eta\left(1-\frac{1-\nu}{\rho_{1}^{3}}-\frac{\nu}{\rho_{2}^{3}}\right)&=&0, \nonumber\\
\frac{\zeta}{uR^{3}}\left[1-GRu^{3}\left(1-\frac{1-\nu}{\rho_{1}^{3}}-\frac{\nu}{\rho_{2}^{3}}\right)\right]&=&0.\label{pbm13}
\end{eqnarray}
To find the first two equations, we have divided by the nonzero quantity $\frac{Gu^{2}}{R^{2}}$.
If we take $\zeta=0$ (solutions in the plane of motion of the primaries), the first two equations are exactly the corresponding equations for the restricted three-body problem of constant masses. This means that all the stationary solutions of the constant masses problem are also present in the problem of variable masses discussed here, namely the three collinear solutions $L_{1}$ to $L_{3}$ (the three masses aligned) and the two triangular solutions $L_{4}$ and $L_{5}$ (the three masses at the vertices of an equilateral triangle), but now the relative distance of the bodies change with time at the same rate.
In order to look for coplanar solutions, we take $\eta=0$. Then, equations (\ref{pbm13}) reduce to:
\begin{eqnarray}
\xi-\frac{\left(1-\nu\right)\left(\xi+\nu\right)}{\rho_{1}^{3}}-\frac{\nu\left(\xi+\nu-1\right)}{\rho_{2}^{3}}&=&0, \nonumber\\
GRu^{3}\left(1-\frac{1-\nu}{\rho_{1}^{3}}-\frac{\nu}{\rho_{2}^{3}}\right)-1&=&0.\label{pbm14}
\end{eqnarray}
The existence of coplanar stationary solutions in the equations above requires
\begin{equation}
GRu^{3}=constant=\kappa. \label{pbm15}
\end{equation}
In other words, the primary bodies must perform a particular motion, determined by a particular time variation rate for the masses masses. In \cite{luk2}, a similar restriction is found. Whereas in that work the restriction leads the masses to vary according to Mestschersky unified law, here, in general, one must solve numerically equations (\ref{pbm15}) and (\ref{pbm12}). From equation (\ref{pbm14}) we have that the possible values for $\kappa$ are restricted by the relation
$\kappa>1$.
Once equation (\ref{pbm15}) holds (with $\kappa>1$), equations (\ref{pbm14}) determine the existence of coplanar solutions \citep{luk}: $L_{6}$ and $L_{7}$, and also $L_{8}$ to $L_{11}$,
when the value of $\kappa$ is less or equal to a certain value that depends on $\nu$. In the limiting case $\kappa\rightarrow 1$, coplanar solutions do not exist, but the infinitely remote solutions $L_{\pm\infty}$ appear, with $\xi=\eta=0$ and $\zeta=\pm\infty$. Solutions $L_1$ to $L_5$ are still present.
\section{Ring solutions}
From our previous relations it is straightforward to verify the existence of ring ($L_{0}$) solutions. We consider now the collinear three-body problem with variable masses \citep{bekov2}. This is the particular case, where the primary bodies move in a straight line passing through their centers. The motion of the primaries is still given by equation (\ref{pbm4}) and the ratio of their masses is still constant in time. We can obtain the radial equation of motion by setting $k=0$ in equations (\ref{pbm5}) and (\ref{pbm6}):
\begin{equation}
\frac{d}{dt}\left(\mu\dot{r}\right)=-\frac{GM\mu}{r^{2}},\label{pbm16}
\end{equation}
with $\theta=constant$.
Since there is no need for the rotating frame of reference used before, we obtain the equations of motion for the third body by setting $\omega=0$ in equations (\ref{pbm9}), thus returning to the inertial frame. Hence, the analogous of equations (\ref{pbm10}) and (\ref{pbm11}) will present only the first two terms of the right-hand side of each of them. Furthermore, we notice, by symmetry, that it is enough to restrict ourselves to the plane $z=0$ and rotate the found solutions around the $x$-axis for all the spatial solutions. In the units used before, equation (\ref{pbm16}) is cast as:
\begin{eqnarray}
\frac{d}{dt}\left(u\dot{R}\right)=-\frac{Gu^{2}}{R^{2}}\label{pbm17}
\end{eqnarray}
By substituting the above equation in the analogous of equation (\ref{pbm11}), we get exactly the first two of equations (\ref{pbm13}). Therefore, we confirm the existence of solutions $L_{1}$ to $L_{3}$ and also solutions $L_{0}$ (a ring around the $x$-axis, generated by rotation of solutions $L_{4}$ and $L_{5}$).
\section{Conclusions}
The presence of a static atmosphere (i.e., a source or sink of mass) in this three-body problem allows the center of mass and the relative motions of two main bodies to be different from that ones of the Gylden-Mestschersky problem. These motions can only be determined if the time variation of the masses is known.
In summary, we found that our variable mass three body problem
has also the five Lagrange solutions of the classic three-body problem, $L_1$ to $L_5$, but now the relative distance of the bodies
change with time at the same rate. In the particular case in which the problem is collinear, triangular solutions $L_4$ and $L_5$ produce a ring solution $L_{0}$.
The sufficient condition for the existence of these solutions is that the ratio of their masses be constant in time. Obviously, this restriction excludes many physically reasonable models for the mass variation, e.g., the ones where the rate of change of the masses depends on the masses themselves.
The coplanar solutions $L_6$ to $L_{11}$ and the infinitely remote solutions $L_{\pm\infty}$, that appear when the motion of the primaries is given by the Gylden-Mestschersky problem, are also present, under the very same additional conditions on the parameter $\kappa$ found in that case.
\acknowledgments
T. S. Amancio thanks FAPESP and CAPES for financial support.
P.S. Letelier acknowledges FAPESP and CNPq for partial financial support.
|
\section{Introduction}
One of our cherished hopes for the LHC is that it will discover an elementary particle that constitutes the dark matter permeating our Universe. Such a particle would necessarily carry neither electric nor colour charge and would be invisible in detectors, its presence being inferred from an excess of events with measured missing energy.
Though the discovery of a new invisible particle at the LHC would surely be serendipitous in itself, making the subsequent connection between such a particle and the dark matter in the cosmos presents a formidable challenge. To do so, one would need to measure the basic properties of the particle, such as its mass, spin, and couplings, in the laboratory setting. Such measurements are inevitably complicated by the fact that the particle, along with information about the energy and momentum that it carries, is lost.
The fact that information is lost does not render the situation hopeless, however. Indeed, in doing experimental data analysis, one is often faced with the situation that parameters are either poorly measured, or not measured at all: the remedy is simply to marginalize with respect to such parameters when computing the likelihood of some hypothesis. But in order to do so, one needs to have a well-defined hypothesis. This strategy has worked rather well for Standard Model physics, where, for example, the presence of invisible neutrinos in leptonic decays of the top quark has not prevented us from measuring the mass of the latter in that channel. But it is not obvious that the strategy will work well when it comes to new physics, beyond the Standard Model. One can, of course, simply impose a hypothesis by {\em fiat},
in the form of an explicit Lagrangian,
but then one runs the risk that the hypothesis may turn out be wrong, in which case inferences based upon it should not be trusted.
Another strategy is to hope that Nature is benevolent enough to supply us with new physics, within the reach of the LHC, that allows us to make measurements whilst only assuming a much more general hypothesis.\footnote{Even with a more general hypothesis, a likelihood based analysis may bear fruit \cite{Allanach:2004ub,Webber:2009vm}.} Preferably, one would like to make the hypothesis minimal, for example
assuming that new particles are produced in pairs and that each decays to the same
invisible particle.\footnote{Alternatively, one could add limited dynamic assumptions, in the form of an effective Lagrangian or ``simplified model'' \cite{Alves:2011wf}.} This assumption typically holds in models where
dark matter is stabilized by a parity symmetry, e.g. supersymmety
with R-parity and universal extra dimensions with KK-parity. With this assumption, we can perform mass measurements purely based on kinematics [4.5];\footnote{For a review of kinematic methods for mass determination, see \cite{Barr:2010zj}.}
for other measurements, one might need to go further and reconstruct energies and momenta event-by-event. Examples discussed previously include measurements of spin \cite{Cheng:2010yy} and $CP$-violation \cite{MoortgatPick:2009jy}.
One could imagine doing so in a theory that predicts many new particles, with masses roughly degenerate.
In such a theory, we expect that heavier new particles, once produced, will decay via a series of cascade two-body decays (which have greater phase space available than decays with three or more bodies, at least in the limit that the masses of the decay products may be neglected), terminating with the neutral, dark matter candidate particle.
Each decay along the chain imposes a mass-shell constraint on the kinematics of the event. If there are enough constraints, all energies and momenta may be reconstructed. However, one still faces a number of difficulties in reconstructing such events at the LHC:
\begin{enumerate}[(1)]
\item The new particles result from collisions between constituent partons of the incoming protons, whose momenta are unknown.
\item By definition, the processes of interest involve invisible final-state particles.
\item Some of the decay products may be coloured partons, which manifest themselves in the detector as hadronic jets. There are intrinsic uncertainties in reconstructing the kinematics of the parent partons from their associated jets -- the jet energy and angle resolution of the detector, the treatment of jet masses, hadronization and underlying event effects, etc.
\item There are often combinatorial ambiguities in assigning final-state objects to the decay chains.
\end{enumerate}
In an ideal world with only difficulties (1) and (2), mass-shell constraints plus missing transverse momentum measurements (for single events or multiple events of the same process) can suffice to reconstruct full event kinematics. However, even in this case equations for unknown masses or momentum components are typically polynomials, with multiple solutions, only one of which is correct. The question then arises: how should we choose between the solutions?
These polynomials have real coefficients so their roots must be real or else complex-conjugate pairs. Often the polynomials are of even degree, in which case there is at least one incorrect real root accompanying the correct one.
Some of the polynomial roots do not correspond to solutions of the original kinematics. This is obviously the case for complex roots, but may also arise for real roots.
There are two reasons for this. One is the logical abhorrence, familiar to all of us from our schooldays, that, while $x = y$ implies $x^2 = y^2$, $x^2 = y^2$ implies either $x = y$ or $x= -y$. Thus, while a root of the constraint equations is also a root of a polynomial equation that is obtained from them by a process of squaring operations, the converse is not necessarily true. The other reason is that roots may be physically unacceptable on other grounds, for example if a reconstructed energy exceeds the centre of mass energy available in a collision.
But often there are multiple acceptable real roots. In any case, in the presence of the uncertainties (3) the correct root could become complex. These effects can be regarded as perturbations of the coefficients of the polynomial. Then the only way the correct root can become complex is if, as the perturbation is increased, it collides with an incorrect real root and they both move into the complex plane in conjugate directions. This requires that the correct and incorrect roots are close together in the absence of the perturbation.
Remarkably, it often happens that correct and incorrect real roots are indeed close together. This can happen if the process involves a sequential decay chain with a large mass hierarchy, or conversely an approximate degeneracy.
To be explicit, consider the problem of reconstructing the mass of some particle at the head of a cascade decay chain.
(a) If there is a large hierarchy, a mass in the chain is approximately zero on the scale of masses higher up the chain. If $p$ and $q$ are the 4-momenta of decay products of a zero-mass object, which must themselves have zero mass, the mass-shell condition $(p+q)^2=0$ implies $p\propto q$, which represents two more constraints than $(p+q)^2=m^2$, so the number of solutions is reduced. This means that roots of the polynomial must coalesce (or move to infinity, but then they cannot be genuine solutions of the kinematics). So for an approximately zero mass, there may be an incorrect real root ``close to" the correct root.
(b) Similarly if the decay product with 4-momentum $q$ has mass equal to the parent mass $m$, then $p$ must be infinitely soft and the parent must also have 4-momentum $q$, which represents additional constraints. So again roots must coalesce in this limit, and be ``close" near this limit.
In fact, since correct and incorrect solutions must be perfectly correlated (in that they coincide) at both extremities of the range of possible intermediate mass values, it turns out that there is a high degree of correlation between correct and incorrect solutions for any values of the intermediate masses.
We shall see however that the ``closeness" of solutions (or, equivalently, their degree of correlation) is difficult to define quantitatively. For certain kinematic configurations, divergence from the limit can be very rapid as the hierarchy or degeneracy is broken. Nevertheless it means that in these circumstances even the incorrect roots will be more densely distributed near the correct value. And in the presence of effects (3) the real parts of complex roots will be also tend to be close to the correct value, with small imaginary parts. Therefore it can make sense simply to plot the real values of all solutions, with a consequent gain in statistics.
In the presence of combinatorial ambiguities (4), we cannot in general expect to get any real roots from wrong combinations. The only general feature is that the polynomial coefficients are still real and so the roots must be real or occur in complex-conjugate pairs. However, if there is a hierarchy or near degeneracy there will be approximate permutation symmetries that imply that the corresponding wrong combinations have roots close to those of the right combination:
(a) When there is a mass hierarchy, visible objects further down the chain are approximately collinear and therefore permuting their momenta will not significantly affect the reconstruction of kinematics higher up the chain.
(b) When there is an approximate degeneracy, some visible momenta will be soft and permutation of these will also not significantly affect reconstruction.
As before, the correlation between right and wrong combinations, which is perfect at either end of the interval of allowed intermediate masses, persists throughout the interval of intermediate masses.
Thus again in these cases it can make sense to take the real parts of all solutions for all combinations. There will be a peaking around the true solution when the combination is right, and also when the combination corresponds to an approximate permutation symmetry, plus a ``background" due to non-symmetric wrong combinations.
In a later section, we shall present an abstract discussion of these phenomena, showing that they have a natural formulation in terms of the theory of Riemann surfaces.
We shall also investigate, via a combination of analysis and numerical simulations, several examples. Before doing that, we would like to whet the reader's appetite by means of an illustrative example, which is not only simple enough that the behaviour may be understood without too much effort, but also is relevant for collider physics today. The example concerns measuring the mass of a top quark decaying in the leptonic channel.
As this example makes clear, our insights are not limited to new physics, beyond the Standard Model. Indeed, experimental analyses involving event reconstruction techniques are ubiquitous in collider physics. As an example, whenever one observes a lepton in association with missing energy at a hadron collider, one has the option of using the known mass of the $W$ boson to reconstruct the four-momentum of a hypothesized $W$-boson in the event. This information might then be used to study, for example, spin correlations or asymmetries (charge or forward-backward) in pair production of top quarks, or to reconstruct a resonance in the $WW$ channel (such as the Higgs). Similarly, whenever one observes a $\tau$ candidate, one may reconstruct the momentum of the $\tau$ by assuming that the neutrino emitted in the $\tau$ decay is collinear with the visible decay products \cite{Ellis:1987xu}. Until recently, this method was employed by both the ATLAS \cite{Aad:2009wy,ATLAS1} and CMS \cite{CMS1} collaborations in their strategies for searching for Higgs bosons. However, requiring that the reconstructed momenta be physical forces one to discard up to half of the events \cite{Aad:2009wy,ATLAS1}, in the presence of detector resolution, and this strategy has been abandoned in recent studies \cite{ATLAS2,CMS2}.
We present a different method for reconstructing events, using the information that comes from the secondary vertex in $\tau$ decays. We argue that in this case it makes sense to retain unphysical solutions, with a consequent gain in statistics.
Moreover, even these examples involving SM neutrinos have applications in new physics searches: reconstruction of the $W$ mass in this way was used recently to look for a resonance in the dijet plus $W$ channel that might explain the recent anomalous excess observed by CDF \cite{Aaltonen:2011mk}. It has also been suggested as a way to discover (and distinguish between) a new $tt$ or $t\overline{t}$ resonance in the di-leptonic channel \cite{Bai:2008sk}.
\subsection{The top quark example}\label{sec:top}
Consider a top quark, $t$, decaying to a bottom quark, $b$ and a $W$-boson, which in turn decays to a lepton, $l$ and an invisible neutrino, $\nu$ in $3+1$ spacetime dimensions, with the neutrino momentum in the two directions transverse to the beam inferred from the missing transverse momentum in the event.\footnote{For pair produced top quarks, we assume that the other top quark decays to visible hadrons.} We denote the mass and four momentum of particle $i$ by $m_i$ and $p_i^\mu = (E_i, \mathbf{p}_i,q_i)$ where $\mathbf{p}$ are the momentum components in the two directions transverse to the beam. The mass shell constraints then read
\def\mathbf{p}{\mathbf{p}}
\def\slashed{\vp}{\slashed{\mathbf{p}}}
\begin{align}
m_t^2 &= (p_\nu + p_l + p_b)^2, \\
m_W^2 &= (p_\nu + p_l)^2,\\
m_\nu^2 &= p_\nu^2, \\
\mathbf{p}_\nu &= \slashed{\vp},
\end{align}
where we have enforced conservation of four-momentum and where $\slashed{\vp}$ is the inferred missing transverse momentum. Now, assuming the masses other than $m_t$ are already known, these constitute five equations in five unknowns, namely $ p_\nu$ and $m_t$. Thus one can hope to reconstruct both the top mass and all the particles' four-momenta in an event.
A little algebra shows that the last four equations can be reduced to a quadratic equation in either the energy or longitudinal momentum of the neutrino.
Hence, using the first equation, one may obtain a quadratic equation in $m_t^2$, with two real solutions, one of which must have the correct value of $m_t^2$. Neglecting the masses of the $b$ quark, the lepton and the neutrino, the difference between the two solutions is given by
\begin{align}
\label{eq:top}
E_l\Delta E_\nu &= q_l\Delta q_\nu = \frac{E_l q_l}{\mathbf{p}_l^2}\sqrt{(m_W^2 + 2\mathbf{p}_l \cdot \slashed{\vp})^2 - 4 \mathbf{p}_l^2 \slashed{\vp}^2 },\\
\Delta m_t^2 &= 2(E_b \Delta E_\nu - q_b \Delta q_\nu).
\end{align}
This simple expression for the difference between the correct and incorrect solutions for the top mass already contains much information. Firstly, we see that, as the mass of the $W$ increases towards $m_t$, such that the $b$ becomes soft, the difference between correct and incorrect solutions for $m_t$ (though not for $E_\nu $ and $q_\nu$) vanishes. Secondly, we see that the differences all vanish as the mass of the $W$ decreases to zero, since in this limit the lepton and neutrino become collinear, such that $(\mathbf{p}_l \cdot \slashed{\vp})^2 - \mathbf{p}_l^2 \slashed{\vp}^2 \rightarrow 0$. Therefore we expect that near either of these limits, wrong solutions for the top mass will be densely distributed over many events near the right solutions. Thirdly, we see that if one starts at large enough values of $m_W$ (near $m_t$) and decreases $m_W$, the wrong solution will always begin by moving away from the right solution, eventually turning around and coming back towards it at small $m_W$. The turnaround point depends on the kinematics of a particular event, but it tells us that, at a fixed, small value of $m_W$ but with multiple, random events, we can expect that the wrong solutions will
still be more densely distributed near the right ones. Nevertheless, sometimes the wrong solution will be rather far away from the right solution, leading to large tails in our distributions. Indeed, for the extreme case of events that have $\mathbf{p}_l = \mathbf{0}$, we see that the wrong solution lies infinitely far away from the right solution. These events form a set of measure zero, but nevertheless, we learn that very large tails can arise.
We shall return to this example in Section \ref{sec:single}, where we shall provide a simple geometric explanation of the above phenomena and identify further interesting properties of the solutions.
\section{Generalities and connection with Riemann surfaces}
In this section, we give a more abstract discussion
which, although (we hope) illuminating, is not necessary to understand the examples given in later Sections and may be skipped by readers who wish to avoid mathematical niceties.
Let us consider, then, some cascade decay or decays, in which the unknowns, corresponding variously to energies or momenta that go unmeasured (for example those of invisible particles such as neutrinos or dark matter candidates) or {\em a priori} unknown masses, are equalled or outnumbered by the constraints, coming from the mass-shell conditions and measurements of total ``missing'' momenta, inferred from global momentum conservation in an event.
For the time being, we assume that there are no combinatorial ambiguities and that all quantities are well-measured.
This set of equations then has at least one solution (the right solution), but may also possess wrong solutions, which for a generic event will lie in a finite set.
As we have already remarked, there may exist limits of the parameters in which the number of constraints is effectively increased.
Now, it may be the case that these extra constraints are redundant, in the sense that they are already implied by the other constraints on the system. If they are not,
then the number of solutions will be reduced.
This reduction in the number of solutions begs the question: what happens to the other solutions as the limit is taken? In particular, where do the other solutions lie when one is close to the limit? Two possibilities suggest themselves. One is that the wrong solutions become larger and larger and eventually go to infinity. The other possibility is that multiple solutions coalesce in the limit, such that the differences between solutions are small close to the limit. If this is the case, then we have an effect whereby wrong solutions may lie close to right solutions, leading to an apparent correlation between the two in samples of multiple events.
Unfortunately, it is rather difficult to see explicitly from these generic arguments which of the two qualitative possibilities obtains; nor is it easy to decipher quantitatively, simply by staring at the system of constraints, how the number of solutions changes. To do that, it is convenient to reduce the set of constraints to a single equation in a single variable. Since the constraints involve, at worst, the square root operation, one can, by repeated squaring operations, always write this single equation as a polynomial equation in the single unknown, for which we wish to solve. In what follows, we would like to study the behaviour of the solutions (or roots) of this unknown as another parameter in the system (an intermediate input mass, say) is varied. We can, by further squarings, always write the single equation as a polynomial in this parameter too, such that we arrive at a polynomial equation in two variables. This naturally leads us into a discussion of Riemann surfaces.
Before that, we remind the reader that the process of squaring operations just described introduces an unpleasant complication: solutions of the polynomial need not be solutions of the original constraint equations. We shall see explicitly that this can happen in one of our later examples. One should always check explicitly that solutions obtained from the polynomial are indeed {\em bona fide} solutions of the original system of multiple equations.
\subsection{Connection with Riemann surfaces\label{sec:riemann}}
Let us now consider our polynomial equation in two variables: one, say, an unknown mass $w$ (we choose the notation for this Section to match that of complex variable theory), and the other, say, an input mass $z$ of an intermediate particle somewhere further down the chain.
We seek the values of $w$, possibly complex, that result from real input values of the known mass $z$. But the discussion will be clearer if we allow both to take complex values. So we have a polynomial, $P(w,z)=0$ of degree $(n,m)$, say. Ultimately, we wish to solve this for $w$ given some input value $z$, but for now, let us just consider it as a polynomial in two variables (or, an algebraic curve).
Since this is an analytic constraint on two complex variables, it manifestly defines a Riemann surface, viz, a 1-complex-dimensional, analytic, manifold, $\mathcal{M}^g$, of genus $g$.\footnote{For a generic $P$, there exists a beautiful way to compute the genus of $\mathcal{M}^g$ directly from $P$ using the Newton polytope; sadly, we shall not need it here.}
We may also find two less explicit descriptions of the Riemann surface by solving $P(w,z) = 0$ to obtain two ``functions'' $w(z)$ and $z(w)$. These are, of course, multivalued, and have branch point singularities whenever the corresponding derivatives, $dw/dz$ or $dz/dw$, do not exist. Since $P$ is just a polynomial,
and since $$P = 0 \implies \frac{\partial P}{\partial w} \frac{dw}{dz} + \frac{\partial P}{\partial z}= 0,$$
the derivative $dw/dz$ exists unless $\frac{\partial P}{\partial w}$ vanishes. One can easily show, furthermore that this is the condition for the polynomial $P$, considered as a polynomial in $w$, to have a repeated root at some value of $z$.
The branch points of these functions then define a Riemann surface in the usual way: one makes arbitrary branch cuts, lifts the complex plane to a multi-sheeted cover and obtains a single-valued function on $\mathcal{M}^g$.
It is important to stress, however, that these two descriptions of the same Riemann surface (one arising from branch points of $w(z)$ and one from $z(w)$) are quite different.
Indeed, one is an $n$-sheeted cover and the other is $m$-sheeted. Moreover, their branch points are not the same.
Now, we are interested in the problem of finding the solutions for the unknown mass $w$ that result as we vary the input mass parameter $z$. The description of $\mathcal{M}^g$ that is relevant for us is therefore the one provided by the function $w(z)$. (If we were interested in the inverse problem of solving for $z$ given $w$, the appropriate description would be in terms of $z(w)$; we repeat that these two descriptions differ in their branch structure.)
We may now ask what happens as we vary the input mass parameter $z$ along a trajectory in $\mathbb{C}$ that goes along the real axis from some initial value towards the origin, where the nature of the mass-shell constraint changes, such that the number of constraints increases. We already know that the behaviour of the solutions must be completely smooth, except for possible branch point singularities. We also expect that the number of solutions must decrease at the origin. We now ask what this implies for the Riemann surface. There are three possibilities, which we discuss in turn.
One possibility is that, due to the logical abhorrence mentioned above, some of the solutions of the polynomial simply cease to become solutions of the system of multiple equations. We shall see it explicitly in the examples.
The second possibility is that some of the roots go towards the point at infinity. Whilst perfectly acceptable from the point of view of the compact Riemann surface, we would no longer regard these as physical solutions. In our examples, this only happens for special kinematic configurations.
The third possibility, which is of most interest to us, is that the polynomial has a repeated root, or equivalently, that $w(z)$ has a branch point, at the origin in $z$. If so, in the neighbourhood of the branch point, multiple solutions will lie close together, leading to a correlation between correct and incorrect solutions, if one of those solutions is the correct solution.
We note that the trajectories followed by the roots as the input parameter moves towards the branch point at the origin may be highly non-trivial, as the reader may see by glancing ahead at Figures \ref{fig:2d}, \ref{fig:lqkW}, and \ref{fig:lqkT}, which illustrate the later examples. The left-hand column of each Figure shows the trajectories, projected from $\mathcal{M}^g$ into the complex plane, followed by the roots in an event. We shall discuss these in more detail later. For now, we note that the roots can indeed coalesce at branch points, that they can move away from the branch point before moving towards it, and also that they can reverse, or otherwise change, their direction, following a cusped trajectory. The cusps do not correspond to singular branch points of the description of the Riemann surface in terms of $w(z)$, which, as we discussed above, arise when $dz/dw$ vanishes (and are forced to lie on the real axis in the projected $w$-plane, given that the coefficients of $P(w,z)$ are real and that we follow a real trajectory in $z$). Rather, they arise at the branch points of the dual description of $\mathcal{M}^g$ in terms of $z(w)$, where $dw/dz$ vanishes. Indeed, at such points, then writing $w$, $z$ in terms of their real and imaginary parts, $w = u + iv$, $z = t$, we have that
$d u/d t= d v/d t = 0$, whence $d v/d u$ is undefined. A classic example is the cycloid curve, $u= t - \sin t, v =1 - \cos t$, which despite being a smooth map from $t$ to $(u,v)$ has cusps at the points where $d u / d t = 0$ and $dv/du$ is undefined.
\subsection{Combinatorics}
The limit as one of the intermediate masses goes to zero is also interesting from the point of view of the problem of combinatorial ambiguities. There are two types of combinatorial ambiguities. One arises when different visible particles along a decay chain are indistinguishable in particle detectors. The second arises when
new particles are pair produced, and the subsequent decays involve identical (or rather indistinguishable) final states. In particular, if the branching ratio for one decay dominates over all others, then the decay chains will be identical (modulo charge conjugation), leading to an ambiguity in assigning observed final state particles to one or other decay chain.
If such ambiguities are truly ambiguous, then the only robust manner in which to proceed is to consider all possible assignments in solving the system of constraints. For a $p$-fold ambiguity, one must solve the constraint system $p$ times, obtaining $p$ copies of all solutions, both right and wrong. Of course, only one of these solutions is the correct one.
Now, in the limit that an intermediate mass goes to zero, it is easy to see that ambiguities in the arrangement of visible particles further down the chain are irrelevant, in the sense that the solutions of the constrained system after permutation are the same as the original solutions. Why? In the limit that an intermediate mass goes to zero, all subsequent particles (which must also be massless) must be emitted collinearly. They may be fully characterized by the fraction of the energy of the parent particle that they carry, such that the order of emissions is irrelevant.
Since permutations down the chain are irrelevant in the limit that the mass vanishes, and since we expect smooth behaviour in the solutions as the mass varies (for a wrong permutation, we are simply solving a different polynomial, and we still have a Riemann surface, albeit a different one), then for small intermediate masses, we should find that solutions of the permuted equations are close to right or wrong solutions of the equations with the correct particle assignment, for which the wrong solutions may, in turn, be close to the right solution.
We now pause to remark that there is an important distinction between the reality properties of solutions (right and wrong) of the right equations and those of the wrong equations, {\em viz.} those obtained by a wrong permutation. In the former case (in the absence of measurement errors), we are guaranteed that one of the solutions (the right one) is real. (For an even polynomial, we are also guaranteed that there exists another real solution, which may or may not lie close to the right solution.) When we solve the polynomial equation corresponding to a wrong particle assignment, we are not guaranteed any real solutions. Nevertheless, according to the arguments above, we expect solutions lying close to the real solution, but possibly off the real axis, in the limit than an intermediate mass is small.
Na\"{\i}vely therefore, we can reduce the combinatorial ambiguity by only accepting solutions that are real. As we shall now discuss, this is not necessarily the optimal strategy in the presence of measurement errors.
\subsection{Mismeasurements}
In the presence of measurement errors, none of the solutions obtained is the right solution. Moreover, we are not even guaranteed to have any real solutions of our polynomial equation, even with the correct particle assignment. This then re-opens the question of whether one should insist on real solutions, as in \cite{Cheng:2008mg}, or whether one should accept all complex solutions, or only those whose imaginary part is small, according to some criterion.
Let us now consider this issue in more detail. At least in the presence of arbitrarily small measurement errors, it makes sense to retain only real solutions. Indeed, since we are solving real polynomials, the complex solutions may only occur in complex-conjugate pairs. Starting from the limit in which measurement errors vanish and one solution is the truly right solution, we see that this solution must remain real as we increase the measurement error, unless the measurement error is so large that the right solution can `pair up' with another solution and move off the axis. In order to do so, the error in the solution for the mass resulting from the measurement error must be comparable to the distance between the right solution and another wrong solution. If this distance were of order of the mass itself, then one could argue that one should reject complex solutions. Indeed, for such a solution to arise from the right, real solution would require a large measurement error, in which case the event should probably have been discarded in the first place.
Unfortunately, we have argued above that the distance between right and wrong mass solutions in the complex plane is not necessarily of the order of the mass itself.
On the contrary, we have argued that right and wrong solutions may coalesce in the limit that an intermediate mass becomes small. So it is easier than one might expect for the right, real solution to become complex in the presence of measurement errors. Of course, if one has a good understanding of the size of typical measurement errors, one could choose whether or not to accept complex solutions. In the absence of such an understanding, it would perhaps make more sense to accept all complex solutions.
\subsection{Classification of event reconstruction}
Before discussing specific examples, let us attempt to categorize the different kinds of event reconstruction that one may envisage and give examples of them.
We will show that several different kinds of reconstruction problem can be viewed as extensions of a basic momentum reconstruction problem.
The basic problem we consider is to reconstruct the energy and momentum of one or more invisible particles, in a single collision event, in which the masses of all particles are assumed to be known. One example relevant for colliders is the leptonic decay of a $W$-boson, where the neutrino has four unknown energy-momentum components, but there are four constraints, namely the two mass-shell constraints for the $W$ and the neutrino, and the two missing transverse momentum constraints. A second example is the di-leptonic decay of pair-produced top quarks. Here, there are two neutrinos and eight unknown energy-momentum components, but there are also eight constraints, if all the masses are known.
Now consider a momentum reconstruction problem of this type, but in which there are more constraints than unknowns. Of course, one can still solve for the momenta if all the masses are known, but one can go further, since the constraints then imply relations between the masses of particles involved in an event. This then gives the possibility of reconstructing not only the momenta, but also some or all of the masses. Indeed, even with just one event, then if one already knows some of the masses, one may be able to solve for the others. With the masses known, one can then go back and reconstruct the momenta in that one event or indeed in any other. As a trivial example, one can always turn a momentum reconstruction problem into a mass reconstruction problem by adding one more particle (of unknown mass) at the head of a decay chain. As an example, taking the decay of a $W$-boson above, one may add a top quark that decays to it (together with a $b$) and solve for the mass of the top. This is precisely what we did in the introduction.
One may go even further: given that a single event of this type implies relations between the masses, one can attempt to reconstruct all of the masses by simply combining events. A possible collider example (which we shall study further later on) is given by pairs of cascade decays, each with three visible particles on each chain. There are eight unknown energy-momentum components (in 3+1 dimensions), but ten mass-shell and two missing transverse momentum conditions. Thus each event can be reduced to two relations on the particle masses. If the chains are assumed to be identical, such that there are only four independent masses in the chains, one needs two events to reconstruct all masses. If the chains are not identical, one needs four events.
In conclusion, we see that various mass reconstruction problems can be viewed as extensions of the basic momentum reconstruction problem.
\section{Examples}
\label{sec:examples}
\subsection{Single chain decays \label{sec:single}}
We have already given a algebraic discussion of one example, namely that of leptonic decays of the top quark. We saw that, once the $W$-boson mass is known and the neutrino mass is assumed negligible, one may solve a quadratic equation for the mass of the top quark; this quadratic equation reduces to a linear equation in the limit that $m_W/m_t \rightarrow 0$ and this in turn leads to a correlation between the right and wrong solutions for small, but non-vanishing $m_W/m_t$.
We would now like to generalize this example further and show that its behaviour may be understood via simple, geometric arguments.
Consider a single decay chain in $D+1$ spacetime dimensions, $\dots \rightarrow C+ \dots \rightarrow B+2 + \dots \rightarrow A + 1+2+ \dots$, with visible particles $1,2,\dots$, terminating in an invisible particle $A$. We assume that the visible particles are all massless and that the masses of all states are known, and that we wish to solve for the unknown energy-momentum components of $A$. We assume that $0 \leq d \leq D$ of the spatial momentum components of $A$ can be inferred via some kind of missing energy measurement. By analogy with a collider physics experiment (and in a slight abuse of terminology), we will call these the `transverse' directions; the unmeasured momentum directions will be called `longitudinal'. We thus have $D+1-d$ unknowns and we may solve for these provided we have an equal (or greater) number of mass-shell constraints. We therefore need a chain containing (at least) $D-d$ visible particles.
As it stands, this is a momentum reconstruction problem. We may turn it into a mass reconstruction problem by adding one more parent particle of unknown mass at the top of the chain. This adds one more unknown (the parent mass), together with one more mass-shell constraint, so the system remains constrained. There are then $D-d+1$ visible particles. In the example of the top quark decay described above, we have $D=3$, $d=2$, such that we need one visible particle (the lepton), to solve for the momentum of the neutrino, and two visible particles (the lepton and the bottom quark) to solve for the mass of the top.
The general constrained single cascade just described can be easily understood in a geometrical way, given the following lemma: provided we only consider Lorentz boosts in the subspace that is orthogonal to the transverse directions, then, if we consider two frames $F$ and $F^\prime$ related by such a Lorentz boost $\Lambda$, then the boost of a solution of the equations written in frame $F$ is itself a solution of the equations written in the boosted frame $F^\prime$. This is obviously true for the right solution, but the Lorentz invariance of the mass-shell constraints guarantees that it holds equally true for wrong solutions as well. An immediate corollary is that, if the unknown being solved for is a mass, the solution will be the same in any two such Lorentz frames.
We stress that the lemma does not hold if one considers boosts in the transverse directions, since there is no sense in which the missing-momentum constraints are Lorentz covariant. One does not even know how to boost the measured missing transverse momentum to another frame, since the result depends on the unknown missing energy.
With the lemma in hand, it is easy to see what happens. In general,
we can boost to a frame in which all the longitudinal momenta of the $D-d$ visible particles required for the momentum reconstruction problem
are linearly dependent, spanning a $D-d-1$-dimensional subspace. For example, we can boost to the longitudinal centre of mass frame of the $D-d$ visibles, in which the longitudinal momenta sum to zero. Figure \ref{fig:reflection} shows the particles in this frame in a decay with three visible particles. In this frame, the mass-shell constraints are invariant under a change in sign of the longitudinal momentum component of $A$ that is orthogonal to the subspace spanned by the visible longitudinal momenta. The two solutions of the equations ({\em viz.} the energy-momenta of $A$) are thus degenerate in this frame, with the exception of that orthogonal component, for which the two solutions are equal in magnitude, but opposite in sign. In a different frame, the longitudinal boost will of course mix up the components, such that none will be degenerate in general.
\begin{figure}\begin{center}
\includegraphics[width=0.6\linewidth]{fig/reflection.eps}
\caption{The right and wrong solutions for the longitudinal momenta in the boosted frame, for a single chain decay with three visible particles.}
\label{fig:reflection}
\end{center}\end{figure}
We note that this argument does not work in $D=1 \implies d=0$, because there is then only one (massless) visible particle, and no finite boost will take us to its rest frame. Indeed, explicit solution in that case shows that there is only ever one solution. More generally, whenever $D-d=1$, implying only one visible particle, the argument applies only if we can do a boost to the longitudinal rest frame of the visible particle. Since the particle is massless, we may do so only if the transverse momentum is non-vanishing.
Now, what happens when one of the intermediate masses is sent to zero? Then the energy-momentum of $A$ is necessarily collinear with the energy-momentum of visible $1$ in the above frame (and indeed in any other frame). Thus, the momentum component of $A$ orthogonal to the subspace spanned by the visibles is zero, and the two solutions for the energy-momentum of $A$ are degenerate in all components, in this frame, as are boosts thereof.
Turning now to the related mass measurement problem (with one more particle added to the chain), we see that, for all intermediate masses non-vanishing,
there will be two solutions for the mass of the added parent particle, obtained by plugging the two values for the energy-momentum of $A$ into the mass shell constraint for the parent particle. By the lemma, these two values will be the same in all frames related by longitudinal boosts. When an intermediate mass vanishes, the two values for the reconstructed energy-momenta of $A$ are the same, and so are the two values for the reconstructed parent mass.
Finally, when any intermediate mass is small compared to the mass that preceeded it in the chain, the two reconstructed parent mass values should lie close together.
These properties are all confirmed by an explicit algebraic analysis. In the momentum reconstruction problem, one obtains a quadratic equation (unless $D-d=1$ and the transverse momentum of particle $1$ is taken to zero, in which case the coefficient of the quadratic term goes to zero) which reduces to a linear equation in the limit that an intermediate mass vanishes. For $D=3$ and $d=2$, one has a simple generalization of the top decay example considered in Section~\ref{sec:top}, viz.\ a single decay chain $C\to B+2$, $B\to A+1$ with an invisible particle $A$ carrying away the missing transverse momentum $\slashed{\vp}$. In the case of top decay, $A,B,C,1,2=\nu,W,t,l,b$. Taking the masses of $A$ and $B$ as known and neglecting those of the visible decay products $1$ and $2$, we have in analogy with (\ref{eq:top})
\begin{align}
\label{eq:single}
E_1\Delta E_A &= q_1\Delta q_A = \frac{E_1 q_1}{\mathbf{p}_1^2}\sqrt{(m_B^2 -m_A^2+ 2\mathbf{p}_1\cdot \slashed{\vp})^2 - 4 \mathbf{p}_1^2 (\slashed{\vp}^2+m_A^2) },\\
\Delta m_C^2 &= 2(E_2 \Delta E_A - q_2 \Delta q_A).
\end{align}
Then for $m_B\to 0$ all the solution differences vanish since we must have $m_A\leq m_B$. We also see that the two solutions for $m_C$, but not for $E_A$ and $q_A$, coincide when visible particle 2 is soft, corresponding to $m_C=m_B$.
Since the difference between right and wrong mass solutions vanishes when the intermediate mass takes its maximum and minimum values and is non-vanishing elsewhere, then the wrong solution (which is necessarily real) must change direction when we follow a trajectory that covers the full range of intermediate masses.
This illustrates in a rather extreme way one type of behaviour we described earlier in our discussion of Riemann surfaces: in this case, not only does the wrong root move away from the right root before returning, but the fact that it is also forced to be real means it traverses a cusp of angle $\pi$ as it does so. As discussed, this must correspond to a branch point of the dual description of the Riemann surface.
Figure~\ref{fig:single} shows the trajectories followed by the wrong solutions as the intermediate mass $m_B$ is decreased from $m_C$ to zero, in a sample of twenty events. As expected, each trajectory begins and ends at the right solution ($m_C =1$), but departs from it in the intervening region. Moreover, whilst
for the majority of events the wrong solution lies close to the right solution throughout the trajectory (including the red vertical line, which corresponds to the kinematics of top quark decays), the discrepancy can be large. Finally, we see that the trajectories can change direction more than once.
\begin{figure}\begin{center}
\includegraphics[angle=90,width=0.8\linewidth]{fig/mass_det_single0.eps}
\caption{Solutions for the mass-squared of the parent particle $C$ in the single decay chain $C\to B+2,\;B\to A+1$, as functions of $m_B$, for 20 ``typical'' events. The correct solution $m_C=1$ is shown in black; the incorrect ones are in blue. While $m_B$ is varied, the events have fixed decay angles in the parent ($B$ and $C$) rest frames, distributed isotropically. The vertical red line corresponds to the kinematics of top quark decay.
\label{fig:single}}
\end{center}\end{figure}
There is yet one more interesting property of this decay chain, which is not evident in our geometrical description. We find that the difference in right and wrong solutions is independent of the mass of $A$, as one varies the mass of the intermediate $B$, whilst keeping the decay angles of all particles constant, as measured in their rest frames. This behaviour is easily demonstrated from Eq.~(\ref{eq:single}). The momenta of particles 1 and $A$ in the rest frame of $B$ have magnitude $p^* = (m_B^2-m_A^2)/2m_B$. Writing $\slashed{\vp} = \mathbf{p}_B-\mathbf{p}_1$, the argument of the square root in (\ref{eq:single}) can then be expressed as
\begin{equation}
4\left[(m_B p^* +\mathbf{p}_B\cdot\mathbf{p}_1)^2 -\mathbf{p}_1^2(m_B^2+\mathbf{p}_B^2)\right]\,.
\end{equation}
\def\mathbf{n}{\mathbf{n}}
Now as we are assuming particle 1 to be massless, its 4-momentum in the collider frame is of the form $p_1^\mu=p^*n^\mu$ where $n^\mu=(n^0,\mathbf{n},n^3)$ is a function of the 4-momentum of $B$ and the direction of 1 in $B$'s rest frame. The important point is that $n^\mu$ is independent of $m_A$. Then (\ref{eq:single}) can be written as
\begin{equation}
\Delta m_C^2 = \frac 4{\mathbf{n}^2}(E_2 n^3 - q_2 n^0)\sqrt{(m_B +\mathbf{p}_B\cdot\mathbf{n})^2 -\mathbf{n}^2(m_B^2+\mathbf{p}_B^2)}\,,
\end{equation}
which is manifestly independent of $m_A$, as the $p^*$ dependence has cancelled. It follows that, for given production and decay distributions of $C$ and $B$, the curves shown are really only functions of the ratio $m_B/m_C$. Accordingly we have marked the point corresponding to top decay as $m_B = m_C m_W/m_t = 0.46$.
Turning this argument around, we can say that, if the mass ratio $m_B/m_C$ has been determined, then the distribution of the wrong solutions provides information on the decay angular distributions, independent of the mass of $A$.
\begin{figure}\begin{center}
\includegraphics[angle=270,width=0.9\linewidth]{fig/mE.eps}
\caption{
The distributions of $\Delta m^2_E /(m_E^2-m_D^2)$ for the decay chain
$E \rightarrow D + 4, D\rightarrow C +3, C \rightarrow B +2, B\rightarrow A + 1$,
showing the correlation between right and wrong solutions.
}
\label{fig:mE}
\end{center}\end{figure}
Another interesting single decay chain is that with four visible particles,
$E\to D+4, D\to C+3, C\to B+2, B\to A+1$.
In this case, given the masses of particles $A,B,C,D$, one obtains a quadratic equation for the mass of the parent particle $E$ without any missing momentum measurement. The difference between the solutions takes the form
\begin{equation}
\Delta m_E^2 = (m_E^2 - m_D^2)f(m_B/m_C,m_B/m_D;\Omega)\,,
\end{equation}
where $\Omega$ represents the dependence on the decay angles. Thus again the distribution of the wrong mass solutions is independent of the invisible particle mass $m_A$. The function $f$ is complicated but vanishes as $m_C$ and/or $m_B\to 0$. Therefore the solutions coalesce in these limits, and also as $m_D\to m_E$.
The fact that the wrong mass solution is forced to lie close to the right mass solution
in the limit of either large or small intermediate masses leads to a correlation in the distribution of wrong solutions and the right solution in a sample of multiple events. This is illustrated in Figure \ref{fig:mE}, which shows the distribution of the wrong mass solution obtained in a decay chain with four visible particles, for varying values of the intermediate masses.
Shorter single chain examples where no missing energy measurement is available may also be relevant for collider physics. Here one needs to combine information from multiple events (making the hypothesis that each corresponds to the same signal), in order to obtain a constrained system \cite{Nojiri:2003tu,Kawagoe:2004rz}.
\subsection{Double chain decays \label{sec:double}}
Let us now turn to pair produced particles, each of which undergoes a cascade decay to an invisible particle. We label one chain as before, and the other with primes: $\dots \rightarrow C^\prime + \dots \rightarrow B^\prime +2^\prime + \dots \rightarrow A^\prime + 1^\prime +2^\prime + \dots$. It is not necessary to assume that the two decay chains are identical, or even of the same length.
Now, in the absence of measured missing energy, the constraints on the two cascade decays are decoupled from each other; we can, thus, apply independent Lorentz boosts to the two chains and show that, as above, solutions exhibit a pairwise degeneracy in the limit that intermediate masses vanish.
Even in the presence of missing energy constraints that couple the two chains, we may be able to reconstruct the two cascade decays individually, in which case the arguments
of the previous section still go through. Let us consider double chains, with $n$ and $m$ visible particles, in $D+1$ spacetime dimensions with $d$ measured missing momenta.
Assuming all masses are known, to solve for the momenta of the two invisible particles $A$ and $A^\prime$, we must have that $2D = d + n +m$. For example if $D=3,d=2,n=3,m=1$, we can first solve for the $n=3$ chain, ignoring the missing energy and then reconstruct the $m=1$ chain using the missing energy. Again, this may be converted into a mass reconstruction problem by adding two parent particles, at the top of each chain, or indeed one parent particle at the top of both chains.
Novel cases arise when we cannot decouple the two chains. The simplest example is $D=d=2, n=m=1$. Whilst this example is not obviously relevant for hadron collider physics, it nevertheless provides a useful illustration of what may happen in situations that are relevant for colliders, such as $D=3,d=2,n=m=2$.
This $D=d=2, n=m=1$ example can, by elimination, be reduced to a quartic equation for the invisible particle momenta, with four complex roots, of which either two or four must be real, in the absence of combinatorics and measurement errors. We solve the quartic equation numerically for several events corresponding to the topology with a single particle $C$ at the head of two identical decay chains. In the limit that the masses of $B$ and $B^\prime$ vanish, the system of constraints collapses to a linear equation. Indeed, in each chain, the visible particle 1 or $1^\prime$ is forced to be collinear with the invisible particle $A$ or $A^\prime$, such that we have the equations $p_A = \alpha p_1$ and $p_{A^\prime} = \alpha^\prime p_{1^\prime}$, with $\alpha, \alpha^\prime$ unknown. Plugging these into the two transverse missing momentum constraints gives a unique solution for $\alpha, \alpha^\prime$ and hence for all the other unknowns.
Thus, provided solutions do not move off to infinity or cease to become solutions as
$m_{B,B^\prime} \rightarrow 0$, then all four solutions must coalesce at that point, with the three wrong solutions lying on top of the right solution.
A complication arises when we try to solve for the mass of the parent particle $C$ at the head of the two chains. This quantity involves the energies of the invisible particles $A,A'$ as well as their 3-momenta. For the real solutions we can legitimately demand that these energies be positive, but for the complex solutions we have to accept either sign, making four mass solutions for each solution of the quartic equation.
In Figure \ref{fig:2d}, we illustrate what happens for four typical events. In the left-hand column, we show all sixteen mass solutions in the complex plane, whereas in the right-hand column we show only the real solutions (the number of solutions is therefore not constant).
In all events, we find that the solutions do indeed coalesce in fours as $m_{B,B^\prime} \rightarrow 0$, one set corresponding in the limit to positive energies for $A$ and $A'$ and the correct mass $m_C$, and the others to unphysical energies for one or both of $A$ and $A'$.
\begin{figure}[ht!]\begin{center}
\includegraphics[angle=90,width=0.399\linewidth]{fig/mass_det_2d_traj1.ps}
\includegraphics[angle=90,width=0.399\linewidth]{fig/mass_det_2d_plot1.ps}
\includegraphics[angle=90,width=0.399\linewidth]{fig/mass_det_2d_traj2.ps}
\includegraphics[angle=90,width=0.399\linewidth]{fig/mass_det_2d_plot2.ps}
\includegraphics[angle=90,width=0.399\linewidth]{fig/mass_det_2d_traj3.ps}
\includegraphics[angle=90,width=0.399\linewidth]{fig/mass_det_2d_plot3.ps}
\includegraphics[angle=90,width=0.399\linewidth]{fig/mass_det_2d_traj4.ps}
\includegraphics[angle=90,width=0.399\linewidth]{fig/mass_det_2d_plot4.ps}
\caption{Solutions for the mass-squared of $C$ in the double decay chain $C\to B+B',\;B\to A+1,\;B'\to A'+1'$
in 2+1 dimensions, as functions of $m_B$, for four ``typical'' events. The correct solution is $m_C=1$. While $m_B$ is varied, each event has fixed decay angles in the parent rest frames. On the left: trajectories of all 16 solutions in the complex mass-squared plane. Each trajectory starts with a cross at $m_B=m_C/2$ (sometimes outside the region shown) and ends with a square at $m_B=0$. The intervening points correspond to uniform steps in $m_B$. On the right: corresponding plots of the real solutions versus $m_B$. \label{fig:2d}}
\end{center}\end{figure}
A more realistic example for collider physics was studied in \cite{Cheng:2008mg,Cheng:2009fw}, where pair decays with three visible particles in each chain were considered: $D\to C+3, C\to B+2, B\to A+1$ and similarly for $D'\ldots1'$.
In a single event, there are eight energy-momentum unknowns, together with eight mass-shell constraints and two measured missing transverse momenta, implying two relations between the eight masses along the two chains. If one makes the further hypothesis that the chains are identical, then from two events one obtains four relations between four masses, meaning that one can solve for all masses in the chain. In \cite{Cheng:2008mg}, it was shown that the system of constraints could be reduced to a single polynomial equation of degree eight in one of the masses. The strategy for dealing with wrong solutions and wrong combinatorics was simply to accept all real solutions and a correlation between right and wrong solutions of the type we describe was observed in numerical simulations.
Again, it is a simple matter to show that, in the limit that $m_{C,C^\prime} \rightarrow 0$, this eighth order equation reduces to a linear equation. (The same is true if $m_{B,B^\prime} \rightarrow 0$.) Indeed, in each event and in each chain the visible particle 1 is forced to be collinear with the invisible particle $A$, such that we have four equations of the form $p_A = \alpha p_1$, with $\alpha$ unknown. Plugging these into the four transverse missing momentum constraints (two components for each of two events) constitute four equations in the four unknowns $\alpha$, with a unique solution.
In \cite{Cheng:2008mg}, the relevant masses were taken from the SUSY benchmark point SPS1a, and were $m_{A,B,C} = 97,143,180$ GeV and $m_D=$ either 565 or 571 GeV (for up or down squarks, respectively). Since $m_C$ is substantially less that $m_D$, we therefore conjecture that all eight complex solutions lie close to the the right solution, leading to a correlation between right and wrong solutions over many events. This was indeed observed for the real solutions in \cite{Cheng:2008mg}; the complex solutions were not retained.
Our arguments also permit us to make a useful statement with regard to combinatorics. We have argued that permutations of visible particles 1 and 2 should be irrelevant in the limit that $m_{C,C^\prime} \rightarrow 0$. Now, in the decay chains considered in \cite{Cheng:2008mg}, particles 1 and 2 are either electrons or muons, leading to an eight- (for $2\mu 2e$) or sixteen-fold ambiguity (for $4\mu$ or $ 4e$) per event, or a 64-, 128-, or 256- fold ambiguity per pair of events. But in the limit that $m_{C,C^\prime} \rightarrow 0$, we argue that permutation of visibles 1 and 2 is irrelevant, in the sense that the solutions obtained after the permutation will be the same as those obtained beforehand. This translates to sixteen irrelevant permutations for a pair of chains and for a pair of events. If we made the na\"{\i}ve assumption that a relevant permutation will lead to a polynomial that has no real solutions, then we would conjecture that one should find precisely sixteen times as many real solutions when one includes combinatorics as compared to when combinatorial ambiguities are removed. In \cite{Cheng:2008mg}, a sample of one hundred events was considered, corresponding to 4,950 event pairs, with 11,662 real solutions in total, without combinatorics. This corresponds to 4,069 event pairs with the minimum number of two real solutions, and 881 with the maximum number of four real solutions. Now, with combinatorics, one must solve 120 times as many degree eight polynomials, but we predict that the number of real solutions will increase by a factor of only sixteen and furthermore that these will be correlated with the right solution. In fact, 185,867 real solutions are obtained in \cite{Cheng:2008mg}, a factor of 15.93 increase compared to the situation without combinatorial ambiguities! Moreover, the pattern of correlation between right and wrong solutions is not changed once one includes combinatorics, as we expect. The mere fact that an odd number of real solutions was obtained in \cite{Cheng:2008mg} once wrong combinations were included shows that one cannot expect perfect agreement: the algorithm used to solve the eighth-order polynomials will, presumably, sometimes fail to converge. Moreover, we only expect the permutations to be truly irrelevant in the limit that $m_{C,C^\prime} \rightarrow 0$; for non-vanishing $m_{C,C^\prime}$ the number of solutions ought to change. Finally, polynomials obtained from relevant permutations may still have real solutions. Indeed, though they are just random polynomials, they have real coefficients and their zeroes are more likely to lie on the real line than, say, on any other straight line drawn through the origin in the complex plane.
\subsubsection{Di-leptonic top decays}
Another relevant example of pair cascade decays occurs already in the Standard Model, namely decays of pair produced top quarks in the di-leptonic channel. There, each top quark decays to a bottom quark and a $W$-boson, which subsequently decays to a charged lepton and an invisible neutrino. Since the masses of all particle involved (including the top quark) are relatively well-known, one can attempt to reconstruct the neutrinos' momenta event-by-event. Indeed, there are eight unknowns (the two four-momenta of the neutrinos), together with eight constraints (the six mass-shell constraints and the two missing transverse momentum constraints). Such a reconstruction, if it can be achieved in practice, would be useful, for example, for a likelihood based test of spin correlations\cite{Melnikov:2011ai}. It has previously been shown that the system of constraints can be reduced to a single, quartic equation in one unknown \cite{Sonnenschein:2006ud}. Here we remark only that, in the limit that the $W$-boson mass can be neglected compared to the top quark mass, the system of constraints reduces to a linear equation in a single unknown. (The arguments are much the same as those given above; we do not repeat them here.) Thus, we again expect a correlation between the right and wrong solutions of the quartic, given the fairly small mass ratio between the $W$ and the top. This effect should enhance our ability to measure spin correlations between pairs of top quarks.
\subsection{Massless particle decays}\label{sec:massless}
In \cite{Gripaios:2010hv} search strategies were discussed for composite leptoquarks coupled to third-generation quarks and leptons, which were argued to give a generic and striking signature for models of strongly-coupled electroweak symmetry breaking that can be consistent with constraints from flavour physics \cite{Gripaios:2009dq}. One challenging final state discussed there was the decay of pair produced leptoquarks, each to a top quark and a $\tau$-lepton, with one top decaying hadronically and the other decaying leptonically.
Assuming the leptoquarks are rather massive (existing constraints suggest that their masses should exceed a couple of hundred GeV), then one can neglect the mass of the $\tau$-lepton,
such that the neutrino or neutrinos emitted in the $\tau$ decay may be assumed to be collinear with the visible products of the $\tau$ decay. With this assumption, one is able to solve for the unknown leptoquark mass, given the known masses of the final state particles.
To wit, on the one hand, there are seven unknowns, namely the leptoquark mass, the energy fractions carried off by the neutrinos in the two $\tau$ decays, and the four momentum of the neutrino from the leptonic $W$ decay. On the other hand there are seven constraints, namely the two missing transverse momentum constraints, the mass shell constraints for the two leptoquarks, and the mass shell constraints for the leptonically-decaying top, its daughter $W$, and its daughter neutrino.
It was shown in \cite{Gripaios:2010hv} that this system of seven constraints in seven unknowns can be reduced to a single quartic equation in the energy fraction of one of the neutrinos coming from a $\tau$ decay. It was also observed that there exists a correlation between the right and wrong solutions of the quartic.
We now ask whether this can be understood in the light of the arguments presented here. To do so, let us consider what happens to the four solutions of the quartic as one of the intermediate masses is taken to zero. To begin with, we consider the limit in which the mass of the $W$ may be neglected compared to the mass of the top quark. Then, one may show that the system of equations collapses to a single, linear equation. Indeed, as $m_W \rightarrow 0$, the four-momentum of the neutrino coming from the $W$-decay must be proportional to the four-momentum of the lepton coming from the $W$-decay; the only unknown is the constant of proportionality, or, equivalently, the energy fraction carried off by the lepton. Then, the mass shell constraint for the top quark, together with the two missing transverse momentum constraints, make up a set of three equations that are linear in three unknowns, namely the energy fractions carried off by the neutrinos in the two $\tau$ decays and the $W$ decay. Thus, there is a unique solution for the energy fractions and indeed the other unknowns.
However, it is not the case that the quartic equation collapses to a linear equation. Rather, what happens is that the quartic equation collapses to a cubic equation; one solution of this cubic is, of course, the right solution, whereas the other two solutions are simply not solutions of the the original constraint equations, in the limit.
This is immediately evident from Figure \ref{fig:lqkW}, where we exhibit numerical solutions of the quartic equation for four typical events. Again, in the left-hand column, we show all four solutions in the complex plane, whereas in the right-hand column we show only the real solutions. In all events, we see that one of the wrong solutions coalesces with the right solution in the limit, but the other two wrong solutions retain a non-vanishing imaginary part in the limit. These complex solutions cannot be solutions of the original system of constraints in the limit, since we saw that the original constraints reduce to three real, linear equations in the three unknown energy fractions, with a unique, real solution. When inserted into the mass shell condition for the leptoquark, these yield a single, real solution for the leptoquark mass.
Hence there is only a two-fold coalescence of solutions in the limit $m_W \rightarrow 0$, rather than the four-fold coalescence that we might have expected. Nevertheless, this will lead to a correlation between two of the real solutions of the quartic equation.
Figure \ref{fig:lqkW} also shows (in red) that there are solutions which may be discarded on the grounds of being unphysical. In the case at hand, we expect that the energy fractions carried off by invisible particles in decays should not exceed unity. Thus, at least in the limit that measurement errors and combinatorics were under control, one would have grounds for rejecting these solutions, even though they result in real leptoquark masses.
\begin{figure}[ht!]\begin{center}
\includegraphics[angle=0,width=0.8\linewidth]{fig/lq_Wplotraj.eps}
\caption{Leptoquark mass solutions for four ``typical'' events, as functions of W mass, for a true leptoquark mass of 1 TeV. On the left: trajectories of all the solutions in the complex mass-squared plane. Each trajectory starts with a cross at $m_W=0.17$ TeV and ends with a square at $m_W=0$ . The intervening points correspond to uniform steps in $m_W$. On the right: corresponding plots of the real solutions versus $m_W$. The red portions of the curves correspond to unphysical values of one or both $\tau$ jet energy fractions.\label{fig:lqkW}}
\end{center}\end{figure}
\begin{figure}[ht!]\begin{center}
\includegraphics[angle=0,width=0.8\linewidth]{fig/lq_plotraj.eps}
\caption{Leptoquark mass solutions for four ``typical'' events, as functions of top mass, for a true leptoquark mass of 1 TeV. On the left: trajectories of all the solutions in the complex mass-squared plane. Each trajectory starts with a cross at $m_{\rm top}=1$ TeV and ends with a square at $m_{\rm top}=0$ . The intervening points correspond to uniform steps in $m_{\rm top}$. On the right: corresponding plots of the real solutions versus $m_{\rm top}$. The red portions of the curves correspond to unphysical values of one or both $\tau$ jet energy fractions.\label{fig:lqkT}}
\end{center}\end{figure}
We may also consider what happens in the limit that the mass of the top quark is assumed to be negligible compared to that of the leptoquark. This obviously implies that the $W$-boson mass may also be neglected, as described above, but it qualitatively changes the behaviour of the solutions, as shown in Figure \ref{fig:lqkT}. Indeed, one can show that neglecting the top mass implies, on its own, that the system reduces from a quartic equation to a quadratic equation. As we have repeatedly described, this implies either that roots go to infinity (which we do not observe), or that roots cease to become solutions of the original system of constraints, or that roots coalesce. For a generic event, we begin with two real and two complex roots. Taking the top mass to zero forces us to have two real roots (since the system reduces to a quadratic equation), but these cannot be the two real roots that we started with, since we know that these coalesce in the limit that the $W$ mass vanishes, which is implied by the vanishing of the top mass.
Thus, the two complex roots must also both become real (and coalesce with each other) in the limit that the top mass vanishes, and indeed this is what we see in all four events in Figure \ref{fig:lqkT}.
Event 2 in Figure \ref{fig:lqkT} illustrates dramatically the kind of cusp behaviour that we described in Section~\ref{sec:riemann}, arising from branch points of the dual description of the Riemann surface.
\subsection[Higgs to $\tau\tau$ decay]{\boldmath Higgs to $\tau\tau$ decay}\label{sec:htautau}
In Section~\ref{sec:massless} we considered the decay of a very massive object into a top quark and a $\tau$-lepton, the latter being so highly boosted that it was a good approximation to neglect its mass and treat its decay products as collinear. If we make the same approximation for a Higgs boson in the favoured mass range $115<m_h<150$ GeV decaying into $\tau\tau$, the kinematics can be reconstructed unambiguously from the visible decay products and the missing transverse momentum. On the other hand the boost is not so large and, especially after taking into account detector resolution and acceptance, the reconstruction of the Higgs mass may not be optimal.
One can avoid the collinearity assumption by making use of information on the $\tau$ decay vertices.
The most useful and best measured attributes of these are their {\em impact parameters}. The impact parameter $\mathbf{b}$ is the displacement of a decay vertex in a direction perpendicular to that of the visible decay momentum, in this case the $\tau$ jet momentum $\mathbf{p}_j$. Then the invisible momentum $\mathbf{p}_\nu$ must lie in the $(\mathbf{b},\mathbf{p}_j)$ plane, so we can write $\mathbf{p}_\nu=x\mathbf{b}+y\mathbf{p}_j$. For hadronic $\tau$ decays, the invisible momenta are carried by single neutrinos and so their four-momenta are fixed by $x$ and $y$ for each decay. These four quantities are subject to two linear missing-$p_T$ constraints and two quadratic $\tau$ mass-shell constraints, giving four solutions and hence a fourfold ambiguity in the reconstructed Higgs mass. However, from our previous arguments the mass hierarchy $m_h\gg m_\tau$ implies that the solutions should tend to be clustered together.
We have investigated this reconstruction method using a sample of 50,000 simulated LHC ($pp$ at 14 TeV) events in which a Higgs boson of mass 130 GeV is produced by vector boson fusion and decays into $\tau\tau$. The event generator was Herwig++ version 2.5.0~\cite{Bahr:2008pv,Gieseke:2011na}, with parton showering, multiple parton interactions, hadronization and the built-in $\tau$-decay package~\cite{Grellscheid:2007tt}. The detector simulation was Delphes version 1.9~\cite{Ovyn:2009tx} with its $\tau$-identification algorithm and the ATLAS simulation card. Vertex information is not provided by Delphes, so we used the hadron-level positions from Herwig++ after gaussian smearing with the r.m.s. values expected for the ATLAS experiment~\cite{Aad:2009wy} (10.5 $\mu$m for the impact parameter). For the analysis, we demanded two $\tau$-tagged hadronic jets with $p_T>10$ GeV and $|\eta|<2$, resulting in 1467 events remaining after cuts.
\begin{figure}[htb!]\begin{center}
\includegraphics[width=0.7\linewidth,angle=270]{fig/all_effects_posi_E.eps}
\caption{Higgs mass reconstructed from simulated detector-level $h\to \tau\tau$ events using impact parameter information, compared with the collinear approximation.\label{fig:htautau}}
\end{center}\end{figure}
Figure~\ref{fig:htautau} shows the Higgs mass reconstructed from the detector-level data using the above method. All solutions with positive real parts of both reconstructed neutrino energies are included. We see that after resolution smearing most of the solutions are complex, but the mass resolution from taking their real parts is just as good as that of the real solutions, and substantially better than that of the collinear approximation. Furthermore, because each solution represents a full reconstruction of the kinematics, there may be scope for further improvement by weighting solutions according to the relevant decay matrix elements.
\section{Discussion and Conclusions}
Reconstruction of missing energy events may be important for many physics
analyses at colliders. Even in the Standard Model, missing energy is ubiquitous, in the form of neutrinos, which are invisible in the detectors.
Reconstruction of Standard Model events may be useful for, for example, improved measurements of the top quark mass, or for identifying the presence of spin correlations in pair production of top quarks.
Reconstruction may prove to be even more important for physics beyond the Standard Model, not only because we hope to see missing energy in events
involving dark matter particles, but also because it is not so easy in the case of new physics to specify the signal hypothesis, that is to say, the lagrangian.
A significant complication affecting reconstruction of energies, momenta, and masses in missing energy events at colliders is the presence of multiple solutions. As we have seen, the number of solutions can be large (sixteen in one of the examples we considered). This is compounded by the presence of combinatorial ambiguities and measurement errors, which further increase the number of solutions and make it less easy to decide which of the multiple solutions is the correct one.
In the worst case scenario, one would have to accept all solutions, correct or incorrect, real or complex, with the risk that the ``signal'' of correct solutions would be overwhelmed by the ``background'' of incorrect solutions.
Here, we have shown that this problem is mitigated by the existence of mechanisms by which the incorrect solutions are correlated with the correct ones. Specifically, we found that correct and incorrect solutions may coincide in the limit that intermediate masses in cascade decay chains either are negligible, or are degenerate with the masses of particles further up the chain, such that the emitted particles are either collinear, or soft, respectively. Furthermore, these same limits can also lead to combinatorial ambiguities becoming irrelevant, in the sense that the same solutions are obtained before and after a permutation of particles. The correlations between correct and incorrect solutions, which are perfect at either end of the interval of possible intermediate particle masses, persist throughout the intermediate mass interval.
We saw that these phenomena have a natural description in terms of the theory of Riemann surfaces, and studied several examples relevant to colliders, for processes both in and beyond the Standard Model.
We hope that our results provide some insight into the general problem of reconstructing events with missing energy and that they will be useful to those who seek to do so in today's colliders.
More specifically, we have shown that the closeness of correct and incorrect solutions means that complex solutions can occur even in the presence of small measurement errors. Whilst existing analyses discard complex solutions (on the grounds that they must correspond to events with large mismeasurements) we recommend that future analyses retain all solutions, with a consequent increase in the available statistics. As discussed in the introduction and Section~\ref{sec:htautau}, the ongoing searches at the LHC for Higgs bosons in decays to pairs of $\tau$ leptons would seem to be a good place to begin.
\section*{Acknowledgments}
BMG thanks N. Orantin for discussions. Part of the work of BMG and BW was performed at the Kavli Institute for Theoretical Physics, University of California, Santa Barbara, supported in part by the National Science Foundation under Grant No. NSF PHY05-51164. BW also acknowledges the support of a Leverhulme Trust Emeritus Fellowship, and thanks the CERN Theory Group for hospitality.
\providecommand{\href}[2]{#2}\begingroup\raggedright |
\section{Introduction}
Variational methods are endemic in physics and engineering. Commonly, physical symmetries are built into the Lagrangian; these correspond to conservation laws via Noether's Theorem. In classical mechanics, for instance, invariance under translation in time is associated with conservation of energy, invariance under rotations in the base space yields conservation of angular momentum, and so on. Noether's Theorem applies only to systems of Euler--Lagrange equations that are in Kovalevskaya form (see Olver (1993) for details). For other Euler--Lagrange systems, each nontrivial variational symmetry still yields a conservation law, but there is no guarantee that it will be nontrivial.
It is still not widely known that Noether's famous paper on variational problems (Noether, 1918) includes a second theorem, which applies to variational symmetries that depend on arbitrary locally smooth functions of all of the independent variables. (Henceforth, we restrict attention to functions that are locally smooth in each continuous variable.) Noether's Second Theorem states that such symmetries exist if and only if there exist differential relations between the Euler--Lagrange equations. In this case the Lie algebra of variational symmetries is infinite-dimensional and the Euler--Lagrange equations are not in Kovalevskaya form. Brading (2002) is an interesting and accessible history of the mathematics and physics of Noether's two theorems and Weyl's gauge symmetries. Brading observes that Noether's Second Theorem has received scant attention in the physics literature, despite its importance for equations that have gauge symmetries.
Noether's Theorem deals primarily with finite-dimensional Lie algebras of variational symmetry generators, whereas Noether's Second Theorem addresses the infinite-dimensional case when generators depend on arbitrary functions of \textit{all} independent variables. However, there is an intermediate case that occurs commonly: generators may depend upon functions that satisfy some constraints. For instance, the variational problem
\[
\delta_u\int \tfrac{1}{2}\big(u_{,x}^2-u_{,t}^2\big)\, \upd x\, \upd t =0
\]
for the one-dimensional wave equation is invariant under Lie point transformations whose characteristic is of the form $\gtt(x-t)$, where $\gtt$ is an arbitrary function. Therefore the Lie algebra of variational symmetry generators is infinite-dimensional. However, the differential constraint
\[
\gtt_{,x}+\gtt_{,t}=0
\]
applies; there are no generators that depend upon an unconstrained arbitrary function $\gtt(x,t)$. Rosenhaus (2002) examines this type of problem from the global viewpoint, determining which spatial boundary conditions are necessary for the existence of integral quantities that are conserved in time. However, the effect of constraints on local conservation laws has remained an open problem until now.
Some important variational problems have variational symmetry generators that depend on several arbitrary functions which are linked by constraints. Noether's Second Theorem has not yet been generalized to deal with this type of problem.
A discrete variational problem may be a model of a discrete process, or it may be a discretization of a continuous variational problem. Either way, discrete analogues
of Noether's two theorems are important. Noether's Theorem for finite difference variational problems has now been well-studied; see, for example, Dorodnitsyn (2001), Hickman \& Hereman (2003) and
Hydon \& Mansfield (2004). If a discretization preserves a variational symmetry, the discrete version of
Noether's Theorem guarantees that the corresponding conservation law is also preserved.
Christiansen \& Halvorsen (2011) have constructed a gauge symmetry-preserving discretization of the Maxwell--Klein--Gordon equations for a charged particle in an electromagnetic field. They used the discrete analogue of Noether's Theorem to obtain a linear relationship between the discrete Euler--Lagrange equations; this was viewed (correctly) as the satisfaction of a physically-important constraint. So there is evidently a need for a discrete analogue of Noether's Second Theorem.
The current paper has three purposes. First, we give a simple local proof of Noether's Second Theorem for a single arbitrary function and extend this to deal with variational symmetries that depend on several arbitrary functions. Then we show what happens locally when the arbitrary functions are constrained in some way. Finally, we transfer these ideas to discrete variational problems. Some well-known systems are used to illustrate the application of our results.
\section{Symmetries of Lagrangians: the continuous case}
We begin by summarizing a few key facts about variational symmetries; for details, see Olver (1993).
We work locally, using independent variables $\mathbf{x} = (x^1,\dots, x^p)$ and dependent variables $\mathbf{u}=(u^1,\dots,u^q)$. Derivatives of each $u^\alpha$ are written in the form $u^\alpha_{,\mathbf{J}}$, where $\mathbf{J}=(j^1,\dots,j^p)$ is a multi-index; here each $j^i$ is a non-negative integer that denotes the number of derivatives with respect to $x^i$, and so $u^\alpha_{,\mathbf{0}}\equiv u^\alpha$. The \textit{total derivative} with respect to $x^i$ is the operator
\[
D_i=\frac{\p}{\p x^i}+\sum_{\alpha,\mathbf{J}}\frac{\p u^\alpha_{,\mathbf{J}}}{\p x^i}\frac{\p}{\p u^\alpha_{,\mathbf{J}}}\,,
\]
and we shall use the following shorthand:
\[
\mathbf{D}_\mathbf{J}=D_1^{j^1}D_2^{j^2}\dots D_p^{j^p},\qquad (\mathbf{D}_\mathbf{J})^{\bdag}=(-\mathbf{D})_\mathbf{J}=(-1)^{j^1+\cdots+j^p}\mathbf{D}_\mathbf{J}.
\]
Here $\bdag$ denotes the formal adjoint.
Henceforth, we adopt the Einstein summation convention to avoid a plethora of summation symbols.
Consider the action
\[
\mathcal{L}[\mathbf{u}] = \int L(\mathbf{x},[\mathbf{u}])\,\upd \mathbf{x},
\]
where $[\mathbf{u}]$ in the integrand is shorthand for $\mathbf{u}$ and finitely many of its derivatives. The Euler--Lagrange equations are obtained by varying each $u^\alpha$ and integrating by parts:
\beq\label{oldfunc}
0=\delta_{\mathbf{u}}\mathcal {L}[\mathbf{u}]\equiv\int\frac{\p L}{\p u^\alpha_{,\mathbf{J}}}\delta u^\alpha_{,\mathbf{J}}\,\upd \mathbf{x}=\int\frac{\p L}{\p u^\alpha_{,\mathbf{J}}}\mathbf{D}_\mathbf{J}(\delta u^\alpha)\,\upd \mathbf{x}=\int \mathbf{E}_\alpha (L) \delta u^\alpha \,\upd \mathbf{x}.
\eeq
Here the Euler operator $\mathbf{E}_\alpha$, which corresponds to variations in $u^\alpha$, is defined by
\[
\mathbf{E}_\alpha (L)=(-\mathbf{D})_\mathbf{J}\left(\frac{\p L}{\p u^\alpha_{,\mathbf{J}}}\right).
\]
Following Olver, we have assumed that variations are restricted to ensure that there are no contributions from the boundary of the domain. With this formal approach, two Lagrangians $L$ and $L'$ lead to the same set of Euler--Lagrange equations,
\beq\label{ELbasic}
\mathbf{E}_\alpha(L) =0,\qquad \alpha = 1,\dots,q,
\eeq
if and only if $L-L'$ is a total divergence.
Typically, the set of all generalized (or Lie--B\"{a}cklund) symmetries of the Euler--Lagrange equations (\ref{ELbasic}) is not a group. However, just as for point symmetries, one can find infinitesimal generators (in evolutionary form, prolonged to all orders),
\[
X=Q^\alpha\frac{\p }{\p u^\alpha}+(D_iQ^\alpha)\frac{\p }{\p D_i u^\alpha}+\dots= \big(\mathbf{D}_{\mathbf{J}}Q^{\alpha}\big)\frac{\p}{\p u^\alpha_{,\mathbf{J}}},
\]
by solving the linearized symmetry condition,
\[
X\big(\mathbf{E}_\beta(L)\big)=0\quad\text{on solutions of (\ref{ELbasic})},\qquad \beta=1,\dots,q.
\]
The solution is a $q$-tuple $\mathbf{Q(\mathbf{x},[\mathbf{u}])}=(Q^1,\dots,Q^q)$, which is called the characteristic of the symmetries generated by $X$. The set of characteristics of all generalized symmetries of the Euler--Lagrange equations (\ref{ELbasic}) is a Lie algebra, whose Lie bracket is
\beq\label{Lieprod}
[\mathbf{Q}_1,\mathbf{Q}_2]=X_1(\mathbf{Q}_2)-X_2(\mathbf{Q}_1).
\eeq
Variational symmetries (in the broadest sense) are generalized symmetries whose infinitesimal generators satisfy the additional condition
\beq\label{xldiv}
X(L) = D_i\Big(P_0^i(\mathbf{x},[\mathbf{u}])\Big),
\eeq
for some functions $P_0^1,\dots, P_0^p$. This condition ensures that the variational problem is invariant under such symmetries. The set of characteristics of all variational symmetries is a Lie algebra with the Lie bracket (\ref{Lieprod}). Integrating (\ref{xldiv}) by parts yields an expression of the form
\beq
Q^{\alpha}\mathbf{E}_{\alpha}(L) = D_i\big(P^i(\mathbf{x},[\mathbf{u}])\big).
\label{first}
\eeq
Consequently, each variational symmetry characteristic $\mathbf{Q}$ yields a conservation law
\beq\label{N1CL}
D_i\big(P^i(\mathbf{x},[\mathbf{u}])\big)=0\quad\text{on solutions of (\ref{ELbasic})}.
\eeq
This is the local version of Noether's first (and best-known) theorem on variational symmetries.
\section{A local approach to Noether's Second Theorem}
Olver (1993) states and proves Noether's Second Theorem in the following form (using our notation).
\begin{theorem} The variational problem (\ref{oldfunc}) admits an infinite-dimensional group of variational symmetries whose characteristics $\mathbf{Q}(\mathbf{x},[\mathbf{u};\gtt])$ depend on an arbitrary function $\gtt(\mathbf{x})$ (and its derivatives) if and only if there exist differential operators $\mathcal{D}^1,\dots,\mathcal{D}^q$, not all zero, such that
\beq\label{N2Ol}
\mathcal{D}^1\mathbf{E}_1(L) + \cdots + \mathcal{D}^q\mathbf{E}_q(L)\equiv 0
\eeq
for all $\mathbf{x}, \mathbf{u}$.
\end{theorem}
The theorem as stated is restricted to variational problems whose variational symmetries \textit{do} form a group; such problems are uncommon. However, if `group of variational symmetries' is replaced by `Lie algebra of variational symmetry generators,' the theorem applies to all variational problems (\ref{oldfunc}). We shall use the more general wording henceforth.
The `if' part of Olver's proof is straightforward. Multiply (\ref{N2Ol}) by an arbitrary function $\gtt(\mathbf{x})$ and integrate by parts to get an expression of the form (\ref{first}), where
\[
Q^\alpha(\mathbf{x},[\mathbf{u};\gtt])=\big(\mathcal{D}^\alpha\big)^{\bdag} (\gtt).
\]
Then reverse the steps that lead from (\ref{xldiv}) to (\ref{first}) to show that $\mathbf{Q}$ is the characteristic of a variational symmetry for each $\gtt(\mathbf{x})$. Consequently, if there are no characteristics of variational symmetries that depend on a completely arbitrary function of $\mathbf{x}$, no relation of the form (\ref{N2Ol}) exists.
The `only if' part of the proof is quite long, so we do not reproduce it here. Instead, we use the following constructive proof, which is much simpler. The key is to regard $\gtt$ as an additional dependent variable and to apply the Euler--Lagrange operator $\mathbf{E}_\gtt$ to (\ref{first}), giving
\beq\label{EgQE}
\mathbf{E}_\gtt\big\{Q^\alpha(\mathbf{x},[\mathbf{u};\gtt])\mathbf{E}_{\alpha}(L)\big\}\equiv 0.
\eeq
This is the required differential relation between the Euler--Lagrange equations. Explicitly, it amounts to
\beq\label{diffrel1}
(-\mathbf{D})_\mathbf{J}\left(\frac{\p Q^\alpha(\mathbf{x},[\mathbf{u};\gtt])}{\p \gtt_{,\mathbf{J}}}\,\mathbf{E}_{\alpha}(L)\right)\equiv 0.
\eeq
If $\mathbf{Q}$ is linear in $\gtt$ and its derivatives, this relation is independent of $\gtt$. Otherwise, one can linearize $\mathbf{Q}$ by taking its Fr\'{e}chet derivative with respect to $\gtt$, as explained by Olver; we discuss an alternative approach later in this section.
Typically the characteristic is independent of high-order derivatives of $\gtt$ and so (\ref{diffrel1}) is easy to determine; the following example illustrates this.
[Note: in each example we use whatever notation is standard, so that readers can see the results in a familiar form, without having to translate from the notation that we use for the general theory.]
\bex\refstepcounter{exple}\label{gaugeex}
The interaction of a scalar particle of mass $m$ and charge $e$ with an electromagnetic field has a variational formulation with well-known gauge symmetries.
The independent variables are the standard flat space-time coordinates $\mathbf{x}=(x^0, x^1, x^2, x^3)$, where $x^0$ denotes time. The dependent variables are the complex-valued scalar $\psi$ (which is the wavefunction for the particle), its complex conjugate, $\psi^*$,
and the real-valued electromagnetic four-potential $A^{\mu}$, $\mu=0,1,2,3$; indices are raised and lowered using the metric $\eta=\text{diag}\{-1,1,1,1\}$. The Lagrangian is
\begin{equation}\label{emLag}
L= \frac14 F_{\mu\nu}F^{\mu\nu} + \left(\nabla_{\mu}\psi\right)\left(\nabla^{\mu}\psi\right)^*+m^2\psi\psi^*
\end{equation}
where
$$F_{\mu\nu}=A_{\nu,\mu}-A_{\mu,\nu},\qquad \nabla_{\mu}=D_{\mu}+\ri e A_{\mu}.$$
Therefore the Euler--Lagrange equations are
\begin{align*}
0&= \mathbf{E}_\psi (L) \equiv -\left(\nabla_\mu\nabla^\mu\psi\right)^{\! *}+m^2\psi^*,\\
0&= \mathbf{E}_{\psi^*}(L) \equiv -\nabla_\mu\nabla^\mu\psi+m^2\psi,\\
0&= \mathbf{E}_\sigma (L) \equiv \ \ri e\psi\left(\nabla_\sigma \psi\right)^{\! *} - \ri e\psi^*\nabla_\sigma\psi+\eta_{\sigma\alpha}F^{\alpha\beta}_{,\beta},
\end{align*}
where $\mathbf{E}_\sigma$ is obtained by varying $A^\sigma$.
The variational symmetries include the gauge symmetries
$$\psi\mapsto \exp(-\ri e\lambda)\psi,\qquad \psi^*\mapsto \exp(\ri e\lambda)\psi^*,\qquad A^{\sigma}\mapsto A^{\sigma}+\eta^{\sigma\alpha}\lambda_{,\alpha},$$
where $\lambda$ is an arbitrary real-valued function of $\mathbf{x}$; the Lagrangian $L$ is invariant under these transformations, so $XL=0$.
The characteristics of the gauge symmetries have components
$$Q^{\psi}=-\ri e \psi \gtt,\qquad Q^{\psi^*}=\ri e \psi^* \gtt,\qquad Q^{\sigma}=\eta^{\sigma\alpha}\gtt_{,\alpha},$$
where $\gtt$ is an arbitrary real-valued function of $\mathbf{x}$. Consequently
\[
\mathbf{E}_\gtt\big\{ Q^{\psi}\mathbf{E}_\psi (L)+ Q^{\psi^*}\mathbf{E}_{\psi^*}(L) +Q^{\sigma} \mathbf{E}_\sigma (L)\big\}=-\ri e\psi\mathbf{E}_\psi (L)+\ri e\psi^*\mathbf{E}_{\psi^*}(L)-D_\alpha\left(\eta^{\sigma\alpha} \mathbf{E}_\sigma (L)\right),
\]
and so Noether's Second Theorem yields the differential relation
\[
-\ri e\psi\mathbf{E}_\psi (L)+\ri e\psi^*\mathbf{E}_{\psi^*}(L)-D_\alpha\left(\eta^{\sigma\alpha} \mathbf{E}_\sigma (L)\right)\equiv 0.
\]
\eex
Our proof of the `only if' part of Noether's Second Theorem is essentially a local version of Noether's original proof (see Noether, 1918). Like Noether's proof, it extends immediately to variational symmetries whose characteristics depend on $R$ independent arbitrary functions $\gttb=(\gtt^1(\mathbf{x}),\dots, \gtt^R(\mathbf{x}))$ and their derivatives. This gives $R$ differential relations between the Euler--Lagrange equations:
\beq\label{EgrQE}
\mathbf{E}_{\gtt^r}\big\{Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])\mathbf{E}_{\alpha}(L)\big\}\equiv(-\mathbf{D})_\mathbf{J}\left(\frac{\p Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])}{\p \gtt^r_{,\mathbf{J}}}\,\mathbf{E}_{\alpha}(L)\right)= 0,\quad r=1,\dots,R.
\eeq
A useful observation is that this set of relations is invariant under any locally invertible change of the arbitrary functions. For suppose that each $\gttb^r$ is a locally invertible function of a new set of arbitrary functions $\htt^\rho,\ \rho=1,\dots,R$. Then the identity
\beq\label{EhrQE}
\mathbf{E}_{\htt^\rho}\big\{Q^\alpha(\mathbf{x},[\mathbf{u};\gttb(\httb)])\mathbf{E}_{\alpha}(L)\big\}\equiv
\frac{\p \gtt^r(\httb)}{\p \htt^\rho}\Big(\mathbf{E}_{\gtt^r}\big\{Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])\mathbf{E}_{\alpha}(L)\big\}\Big) \Big{|}_{\gttb=\gttb(\httb)}
\eeq
shows that the set of relations
\[
\mathbf{E}_{\htt^\rho}\big\{Q^\alpha(\mathbf{x},[\mathbf{u};\gttb(\httb)])\mathbf{E}_{\alpha}(L)\big\}=0, \qquad \rho=1,\dots,R,
\]
is obtained by taking an invertible linear combination of the original relations (\ref{EgrQE}). So if there exists a locally invertible change of variables such that $\mathbf{Q}$ is linear in $\httb$, (\ref{EgrQE}) is equivalent to a set of relations that is independent of any arbitrary functions.
It is convenient to regard (\ref{EgrQE}) as `new' Euler--Lagrange equations that arise when $\mathbf{g}$ is varied in the functional
\beq\label{newfunc}
\hat{\mathcal{L}}[\mathbf{u};\mathbf{g}] = \int \hat{L}(\mathbf{x},[\mathbf{u};\mathbf{g}])\,\upd \mathbf{x},
\eeq
where
\beq\label{newL}
\hat{L}(\mathbf{x},[\mathbf{u};\mathbf{g}])=Q^\alpha(\mathbf{x},[\mathbf{u};\mathbf{g}])\mathbf{E}_{\alpha}(L(\mathbf{x},[\mathbf{u}])).
\eeq
To prove that the relations (\ref{EgrQE}) are independent, suppose that the converse is true; then there exist differential operators $\hat{\mathcal{D}}^r$ such that
\[
\hat{\mathcal{D}}^1\mathbf{E}_{\gtt^1}(\hat{L}) + \cdots + \hat{\mathcal{D}}^R\mathbf{E}_{\gtt^R}(\hat{L})\equiv 0.
\]
By the `if' part of Noether's Second Theorem, there exist variational symmetries of (\ref{newfunc}) whose generator is of the form
\[
\hat{X}=D_{\mathbf{J}}\Big\{\big(\hat{\mathcal{D}}^r\big)^{\bdag} (\hat{\gtt})\Big\}\frac{\p}{\p \gtt^r_{,\mathbf{J}}}\,,
\]
where $\hat{\gtt}$ is an arbitrary function. In other words, there exist functions $\hat{P}^i(\mathbf{x},[\mathbf{u};\gttb;\hat{\gtt}])$ such that
\[
\hat{X}\Big\{ Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])\Big\}\mathbf{E}_{\alpha}(L(\mathbf{x},[\mathbf{u}]))=\hat{X}\big(\hat{L}\big)=D_i\big(\hat{P}^i\big).
\]
Consequently, for arbitrary $\gttb$ and $\hat{\gtt}$,
\[
\hat{Q}^\alpha(\mathbf{x},[\mathbf{u};\gttb;\hat{\gtt}])=\hat{X}\Big\{ Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])\Big\}
\]
is a characteristic of variational symmetries for the original variational problem (\ref{oldfunc}). This implies that the set of variational symmetries for (\ref{oldfunc}) depends on more than $R$ independent arbitrary functions, contradicting our original assumption. Hence the relations (\ref{EgrQE}) cannot be dependent.
Conversely, given $R$ independent relations between the Euler--Lagrange equations, Olver's `if' proof applied to each one shows that there are characteristics of variational symmetries that depend on $R$ independent arbitrary functions, namely
\beq\label{grchar}
Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])=\big(\mathcal{D}^\alpha_r\big)^{\bdag} (\gtt^r).
\eeq
Thus we arrive at the general form of Noether's Second Theorem.
\begin{theorem} The variational problem (\ref{oldfunc}) admits an infinite-dimensional Lie algebra of variational symmetry generators whose characteristics $\mathbf{Q}(\mathbf{x},[\mathbf{u};\gttb])$ depend on $R$ independent arbitrary functions $\gttb=(\gtt^1(\mathbf{x}),\dots, \gtt^R(\mathbf{x}))$ and their derivatives if and only if there exist differential operators $\mathcal{D}^\alpha_r$ that yield $R$ independent differential relations,
\beq\label{N2HM1}
\mathcal{D}^\alpha_r\mathbf{E}_\alpha(L)\equiv 0,\qquad r=1,\dots,R,
\eeq
between the Euler--Lagrange equations. Given the relations (\ref{N2HM1}), the corresponding characteristics are (\ref{grchar})
Conversely, given the characteristics, the corresponding differential relations are (\ref{EgrQE}).
\end{theorem}
\section{Constrained variational symmetries}
Constraints on arbitrary functions in $\mathbf{Q}$ arise from the linearized symmetry condition for the Euler--Lagrange equation, coupled with the (linear) requirement that the symmetries are variational. Therefore we now suppose that the characteristics of variational symmetries of (\ref{oldfunc}) depend on functions $\gttb$ that are subject to $S$ linear differential constraints,
\beq\label{constr}
\mathscr{D}_{sr}(\gtt^r)=0,\qquad s=1,\dots,S,
\eeq
where each $\mathscr{D}_{sr}$ is a differential operator. We assume that this set of constraints is complete, in the sense that it yields no additional integrability conditions.
The constraints can be incorporated into the Lagrangian (\ref{newL}), yielding
\beq\label{newLcon}
\hat{L}(\mathbf{x},[\mathbf{u};\gttb])=Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])\mathbf{E}_{\alpha}(L(\mathbf{x},[\mathbf{u}]))-
\nu^s\mathscr{D}_{sr}(\gtt^r),
\eeq
where $\nu^1,\dots,\nu^s$ are Lagrange multipliers. Taking variations of (\ref{newfunc}) with respect to $\gttb$, we obtain
\beq\label{EgQEcon}
(-\mathbf{D})_\mathbf{J}\left(\frac{\p Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])}{\p \gtt^r_{,\mathbf{J}}}\,\mathbf{E}_{\alpha}(L)\right)\equiv(\mathscr{D}_{sr})^{\bdag}(\nu^s),\qquad r=1,\dots,R.
\eeq
If $S<R$, one may be able to eliminate the undetermined functions $\nu^s$ from (\ref{EgQEcon}), as will be illustrated by the next example. The elimination of a particular set ${\cal S}$ of dependent variables from a differential system may be accomplished using algorithms in differential algebra. Hubert (2000) and Hubert (2003) provide a careful and extensive summary of the state of the art to date. Differential elimination algorithms have been implemented in the {\sc Maple} library packages {\tt diffalg} and
{\tt rifsimp}. These algorithms come with a `certificate'; provided that the correct elimination term ordering is
used, the elimination ideal is generated by those equations ${\cal E}$ in the output that are independent of the variables in ${\cal S}$. This means that any equation that is algebraically or differentially derived from the
input set, and is free of the elements of ${\cal S}$, can be derived from ${\cal E}$. If ${\cal E}$ is empty then the elimination ideal
itself is empty; in this case, none of the equations that can be derived from the input set are independent of all variables in ${\cal S}$.
Incidentally, the elimination algorithm can be used to test directly whether or not an Euler--Lagrange system
has a differential relation such as (\ref{N2HM1}): simply replace $\{E^{\alpha}(L)=0\}$
with $\{E^{\alpha}(L)-f^{\alpha}=0\}$ and eliminate ${\cal S}=\{u^{\alpha}\}$. The resulting equations for $\{f^{\alpha}\}$ codify the ideal of differential relations on the expressions $\{E^{\alpha}(L)\}$.
There are subtleties for nonlinear systems; indeed, the proof that the output of the algorithms has
the stated properties is one of the most subtle and complex problems of differential algebra.
\bex\refstepcounter{exple}\label{twoarbs}
Suppose that $\mathbf{Q}$ depends on two functions $\gtt^1(x^1,x^2)$ and $\gtt^2(x^1,x^2)$ that are subject to the single constraint
\beq
\gtt^1_{,1} + \gtt^2_{,2} =0
\label{tenth}
\eeq
(This type of constraint occurs when the variational symmetries include all area-preserving diffeomorphisms of the plane.)
Then (\ref{EgQEcon}) amounts to
\begin{eqnarray*}
(-\mathbf{D})_\mathbf{J}\left(\frac{\p Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])}{\p \gtt^1_{,\mathbf{J}}}\,\mathbf{E}_{\alpha}(L)\right)&\equiv& -\nu_{,1}, \\
(-\mathbf{D})_\mathbf{J}\left(\frac{\p Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])}{\p \gtt^2_{,\mathbf{J}}}\,\mathbf{E}_{\alpha}(L)\right)&\equiv& -\nu_{,2}.
\end{eqnarray*}
By eliminating the Lagrange multiplier $\nu$, we obtain a single differential relation between the Euler--Lagrange equations:
\beq
(-D_2)(-\mathbf{D})_\mathbf{J}\left(\frac{\p Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])}{\p \gtt^1_{,\mathbf{J}}}\,\mathbf{E}_{\alpha}(L)\right) + D_1 (-\mathbf{D})_\mathbf{J}\left(\frac{\p Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])}{\p \gtt^2_{,\mathbf{J}}}\,\mathbf{E}_{\alpha}(L)\right)\equiv 0.
\label{eleventh}
\eeq
This is precisely the same constraint as would be obtained by using the local solution of (\ref{tenth}),
\[
\gtt^1=\gtt_{,2},\quad \gtt^2 =-\gtt_{,1}
\]
(where $\gtt$ is entirely arbitrary) in Noether's Second Theorem. However the Lagrange multiplier method has the advantage of working even when (\ref{constr}) cannot be solved explicitly.\eex
When the system of constraints (\ref{constr}) does not allow the multipliers $\nu^s$ to be eliminated, the most that one can do is to construct conservation laws of the Euler--Lagrange equations, as follows. First determine a particular solution of (\ref{EgQEcon}) for $\nu^s$; this can be done by inspection (which is fast), or by the homotopy method (which is systematic). The resulting solution is then substituted into
\beq\label{gclaw}
\nu^s\mathscr{D}_{sr}(\gtt^r)-\gtt^r(\mathscr{D}_{sr})^{\bdag}(\nu^s)=0\qquad\text{when}\quad \mathbf{E}_{1}(L)=\cdots=\mathbf{E}_{q}(L)=0,
\eeq
yielding a set of conservation laws that depends on $\gttb$. This set is independent of which particular solution is used. It can be shown that if $\mathbf{Q}$ is homogeneous with respect to $\gttb$, the conservation law (\ref{gclaw}) is a multiple of the one that is obtained from Noether's (first) Theorem.
\bex\refstepcounter{exple}\label{weex}
As we discussed in the Introduction, the Lagrangian
\[
L=\tfrac{1}{2}\big((u_{,x})^2-(u_{,t})^2\big)
\]
for the one-dimensional scalar wave equation,
\beq\label{wave}
u_{,tt}-u_{,xx}=0,
\eeq
admits variational symmetries with partly-constrained characteristics. Indeed, the most general variational symmetry characteristic is $Q=\gtt^1(x,t)+\gtt^2(x,t)$, where
\[
\gtt^1_{,x}+\gtt^1_{,t}=0,\qquad \gtt^2_{,x}-\gtt^2_{,t}=0.
\]
So the modified Lagrangian (\ref{newLcon}) is
\[
\hat{L}=(u_{,tt}-u_{,xx})(\gtt^1+\gtt^2)-\nu^1(\gtt^1_{,x}+\gtt^1_{,t})-\nu^2(\gtt^2_{,x}-\gtt^2_{,t}).
\]
\eex
Taking variations with respect to $\gttb=(\gtt^1,\gtt^2)$ gives the identities
\[
u_{,tt}-u_{,xx}\equiv-(\nu^1_{,x}+\nu^1_{,t}),\qquad u_{,tt}-u_{,xx}\equiv-(\nu^2_{,x}-\nu^2_{,t}),
\]
which are satisfied by
\[
\nu^1=u_{,x}-u_{,t},\qquad \nu^2=u_{,x}+u_{,t}.
\]
Then the conservation laws (\ref{gclaw}) amount to
\[
\big\{(\gtt^1+\gtt^2)u_{,x}-(\gtt^1-\gtt^2)u_{,t}\big\}_{,x}+\big\{(\gtt^1-\gtt^2)u_{,x}-(\gtt^1+\gtt^2)u_{,t}\big\}_{,t}=0\ \, \text{when (\ref{wave}) holds}.
\]
There are no trivial conservation laws with $Q$ nonzero.
\bex\refstepcounter{exple}\label{swwex} For a more substantial example that produces interesting conservation laws, consider the two-dimensional shallow water equations from Lagrangian fluid mechanics.
Each fluid particle is labelled by its position $(a^1,a^2)$ at some reference time, $t=0$, and $(x,y)=(x(a^1,a^2,t),y(a^1,a^2,t))$ is the position of
the fluid particle at time $t$. We set $\dot{x}=x_{,t} $ and $\dot{y}=y_{,t}$; from here on only the derivatives with respect to the label variables $a^i$ are denoted by subscripts.
Salmon (1983) introduced a Lagrangian for the shallow water equations; with constant Coriolis parameter $f$, it amounts to
$$ L([x,y])= \frac12 \left (\dot{x}^2+\dot{y}^2\right) + fx\dot{y}-\frac12 g h,$$
where $$ h=\frac1{x_{,1}y_{,2}-x_{,2}y_{,1}},$$ is the fluid depth and
$g$ is the constant of gravity. To translate between the Eulerian and Lagrangian viewpoints, we need the identities
$$\frac{\partial}{\partial x}= h\left( y_{,2}\frac{\partial}{\partial a^1}-y_{,1}\frac{\partial}{\partial a^2}\right),\qquad
\frac{\partial}{\partial y}= h\left( -x_{,2}\frac{\partial}{\partial a^1}+x_{,1}\frac{\partial}{\partial a^2}\right).$$
Calculating the Euler--Lagrange system yields the shallow water equations
\begin{equation}\label{sww}
E_x(L)\equiv-\ddot{x} + f \dot{y} - g \frac{\partial h}{\partial x}=0,\qquad E_y(L)\equiv-\ddot{y} - f \dot{x} - g \frac{\partial h}{\partial y}=0.\end{equation}
These are supplemented by the continuity equation,
\[
\dot{h}+ h\left(\frac{\partial \dot{x}}{\partial x}+\frac{\partial \dot{y}}{\partial y}\right)=0,
\]
which is {\em not} an Euler--Lagrange equation in Salmon's formulation.
The variational symmetries include the well-known particle relabelling symmetries; these are arbitrary area-preserving diffeomorphisms of label space, whose characteristics are (Bila, Mansfield \& Clarkson, 2006)
$$ Q^x = x_{,i}\gtt^i,\qquad Q^y = y_{,i}\gtt^i,\qquad \text{where}\quad\dot{\gtt}^1=\dot{\gtt}^2=\gtt^1_{,1}+\gtt^2_{,2}=0.$$
Taking the three constraints on the functions $\gttb=(\gtt^1,\gtt^2)$ into account, the modified Lagrangian is
$$\hat{L} = x_{,i}\gtt^iE_x(L) + y_{,i}\gtt^iE_y(L) -\nu^1\dot{\gtt}^1 -\nu^2\dot{\gtt}^2 -\nu^3\left(\gtt^1_{,1}+\gtt^2_{,2}\right).$$
Taking variations with respect to $\gttb$, we obtain
$$x_{,i} E_x(L)+y_{,i} E_y(L) \equiv -\dot{\nu}^i-\nu^3_{,i}, \qquad i=1,2,$$
which is satisfied by
\begin{align*}
\nu^1&=\dot{x}x_{,1}+\dot{y}y_{,1}+fxy_{,1},\\
\nu^2&=\dot{x}x_{,2}+\dot{y}y_{,2}+fxy_{,2},\\
\nu^3&=-\tfrac{1}{2} (\dot{x}^2+\dot{y}^2) -fx\dot{y} +gh.
\end{align*}
With these solutions $\nu^i$, the resulting conservation laws (\ref{gclaw}) are
\[
D_t(\gtt^1\nu^1+\gtt^2\nu^2)+(\gtt^1\nu^3)_{,1}+(\gtt^2\nu^3)_{,2}=0\qquad \text{when (\ref{sww}) holds}.
\]
In particular, taking $(\gtt^1,\gtt^2)$ to be $(1,0)$ and $(0,1)$ in turn yields
\[
\dot{\nu}^1+\nu^3_{,1}=0,\qquad \dot{\nu}^2+\nu^3_{,2}=0\qquad \text{when (\ref{sww}) holds},
\]
which were discovered in a (somewhat complicated) multisymplectic setting by Hydon (2005). Differentiating the first of these with respect to $a^2$, the second with respect to $a^1$ and eliminating $\nu^3_{,12}$, we obtain the well-known conservation law for potential vorticity,
\[
D_t\left\{\frac{1}{h}\left(\frac{\partial \dot{y}}{\partial x}-\frac{\partial \dot{x}}{\partial y} +f\right)\right\}=0\qquad \text{when (\ref{sww}) holds}.
\]
So in this example, our extension of Noether's Second Theorem has led quickly and easily to some fundamental conservation laws. Conservation of potential vorticity also arises directly from Noether's Theorem, but considerable effort is needed to obtain it (see Bila, Mansfield \& Clarkson, 2006).
\eex
\section{From continuous to discrete: Noether's Second Theorem}\label{sec:dis}
For difference equations, the dependent variables are again $\mathbf{u} = (u^1,\dots,u^q)$, but now the independent variables are $\mathbf{n} = (n^1,\dots,n^p)$, where each $n^i$ is an integer. The forward shift operators, $S_i$, are defined by
\[
S_i:n^j\mapsto n^j+\delta_i^j,\qquad S_i f(\mathbf{n})=f(S_i\mathbf{n}),
\]
for all functions $f$; here $\delta_i^j$ is the Kronecker delta. The adjoint of $S_i$ is the backward shift $(S_i)^{\bdag}=S_i^{-1}:n^j\mapsto n^j-\delta_i^j$. We also use the multiple shifts, $\mathbf{S}_{\mathbf{J}} = S_1^{j_1}\cdots S_p^{j_p}$, and the forward difference operators,
\[
\tl{D}_i=S_i-\text{id},
\]
where $\text{id}$ denotes the identity map. (Here and in what follows, we use notation that is similar to the continuous case in order to highlight the similarities. To avoid confusion, we use a tilde to distinguish an operator from its continuous counterpart.)
For difference variational problems, the action is of the form
\[
\mathcal{L}[\mathbf{u}] = \sum_{\mathbf{n}}L\big(\mathbf{n},[\mathbf{u}]\big),
\]
where now $[\mathbf{u}]$ denotes $\mathbf{u}(\mathbf{n})$ and finitely many of its shifts. Then the Euler--Lagrange equations are
\beq
\tl{\mathbf{E}}_{\alpha}(L)\equiv \mathbf{S}_{-\mathbf{J}}\left\{\frac{\p L}{\p \mathbf{S}_{\mathbf{J}}u^{\alpha}}\right\}=0,\qquad \alpha =1,\dots,q.
\label{twelve}
\eeq
Generalized symmetries of the Euler--Lagrange equations have generators of the form
$$\tl{X}=\mathbf{S}_{\mathbf{J}}\big(Q^{\alpha}(\mathbf{n},[\mathbf{u}])\big) \frac{\partial}{\partial \mathbf{S}_{\mathbf{J}}u^{\alpha}}\,$$
that satisfy the linearized symmetry condition,
\[
\tl{X}\big(\tl{\mathbf{E}}_{\alpha}(L)\big) =0\qquad\text{when (\ref{twelve}) holds}.
\]
Here $Q^{\alpha}(\mathbf{n},[\mathbf{u}]),\ \alpha = 1,\dots,q$, are the components of the characteristic, $\mathbf{Q}(\mathbf{n},[\mathbf{u}])$.
The symmetries are \textit{variational} if they do not change the action, that is, if there exist functions $P_0^i\big(\mathbf{n},[\mathbf{u}]\big)$ such that
\beq
\tl{X}(L) \equiv \tl{D}_i P_0^i\big(\mathbf{n},[\mathbf{u}]\big),
\label{fourteen}
\eeq
Summing (\ref{fourteen}) by parts yields the following condition, which is analogous to (\ref{first}):
\beq
Q^{\alpha}\tl{\mathbf{E}}_{\alpha}(L) = \tl{D}_i P^i\big(\mathbf{n},[\mathbf{u}]\big),
\label{fifteen}
\eeq
for some functions $ P^i$. So Noether's Theorem applies just as in the continuous case: every variational symmetry characteristic yields a conservation law of the Euler--Lagrange equations, namely
\[
\tl{D}_i P^i\big(\mathbf{n},[\mathbf{u}]\big)=0,\qquad\text{when (\ref{twelve}) holds}.
\]
With the above notation, the analogue of Noether's Second Theorem for the difference calculus of
variations is as follows.
\begin{theorem} The difference variational problem $\delta_\mathbf{u}\tl{\mathcal{L}}[\mathbf{u}] =0$ admits an infinite-dimensional Lie algebra of variational symmetry generators whose characteristics $\tl{\mathbf{Q}}(\mathbf{n},[\mathbf{u};\gttb])$ depend on $R$ independent arbitrary functions $\gttb=(\gtt^1(\mathbf{n}),\dots, \gtt^R(\mathbf{n}))$ and their derivatives if and only if there exist difference operators $\tl{\mathcal{D}}^\alpha_r$ that yield $R$ independent difference relations,
\beq\label{N2HM2}
\tl{\mathcal{D}}^\alpha_r\tl{\mathbf{E}}_\alpha(L)\equiv 0,\qquad r=1,\dots,R,
\eeq
between the Euler--Lagrange equations. Given the relations (\ref{N2HM2}), the corresponding characteristics are
\beq
Q^\alpha(\mathbf{n},[\mathbf{u};\gttb])=\big(\tl{\mathcal{D}}^\alpha_r\big)^{\bdag} (\gtt^r).
\eeq
Conversely, given the characteristics, the corresponding difference relations are
\beq\label{EgrQEd}
\tl{\mathbf{E}}_{\gtt^r}\big\{Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])\tl{\mathbf{E}}_{\alpha}(L)\big\}\equiv \mathbf{S}_{-\mathbf{J}}\left(\frac{\p Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])}{\p \mathbf{S}_\mathbf{J}\gtt^r}\,\tl{\mathbf{E}}_{\alpha}(L)\right)= 0,
\eeq
and consequently
\[
\tl{\mathcal{D}}^\alpha_r=\left\{ \mathbf{S}_{-\mathbf{J}}\left(\frac{\p Q^\alpha(\mathbf{x},[\mathbf{u};\gttb])}{\p \mathbf{S}_\mathbf{J}\gtt^r}\right)\right\}\mathbf{S}_{-\mathbf{J}}\,.
\]
\end{theorem}
The proof completely replicates our proof of the continuous version, so we shall omit it.
The identity (\ref{EhrQE}) holds for difference equations, so the set of relations is invariant under locally invertible changes of the arbitrary functions.
\bex\refstepcounter{exple}\label{gaugeex2}
For the interaction of a scalar particle with an electromagnetic field (see Example 1), Christiansen \& Halvorsen (2011) discovered a finite difference approximation that preserves the gauge symmetries as variational symmetries. The mesh $\mathbf{x}(\mathbf{n})$ is uniformly-spaced in each direction, with step lengths
\[
h^\mu=\big(S_\mu-\text{id}\big) \big(x^\mu(\mathbf{n})\big),\qquad \mu=0,\dots,3.
\]
It is useful to introduce the scaled forward difference operators
\[
\ol{D}_\mu=\tl{D}_\mu/h^\mu,\qquad \mu=0,\dots,3,
\]
so that $\ol{D}_\mu$ tends to $D_\mu$ in the limit as $h^\mu$ tends to zero. Note that
\[
\ol{D}_\mu^{\,\bdag}=-\frac{\text{id}-S_\mu^{-1}}{h^\mu}\,;
\]
the adjoint of the (scaled) forward difference operator is the negative of the (scaled) backward difference operator.
Let $\tl{\psi}(\mathbf{n})$ denote the approximation to the wavefunction $\psi$ at $\mathbf{x}(\mathbf{n})$. Christiansen and Halvorsen
used a Yee discretization to approximate $A_\mu$ on the edge that connects the points $\mathbf{x}(\mathbf{n})$ and $S_\mu\mathbf{x}(\mathbf{n})$; we denote this approximation by $\tl{A}_\mu(\mathbf{n})$. As before, indices are raised and lowered using the metric $\eta=\text{diag}\{-1,1,1,1\}$. Up to a sign, the Lagrangian for the scheme is
\begin{equation}\label{emLagd}
L= \frac14 \tl{F}_{\mu\nu}\tl{F}^{\mu\nu} + \left(\tl{\nabla}_{\mu}\tl{\psi}\right)\left(\tl{\nabla}^{\mu}\tl{\psi}\right)^*+m^2\tl{\psi}\tl{\psi}^*
\end{equation}
where, for each $\mu,\nu=0,\dots,3$,
$$\tl{F}_{\mu\nu}=\ol{D}_\mu\tl{A}_{\nu}-\ol{D}_\nu\tl{A}_{\mu},$$
$$\tl{\nabla}_{\mu}=\frac{1}{h^\mu}\left\{S_\mu-\exp\!\big(\!-\ri e h^\mu \tl{A}_{\mu}\big)\text{id}\right\}=\ol{D}_{\mu}+\frac{1}{h^\mu}\left\{1-\exp\!\big(\!-\ri e h^\mu \tl{A}_{\mu}\big)\right\}\text{id}.$$
Therefore the Euler--Lagrange equations that are obtained by varying $\tl{\psi},\ \tl{\psi}^*$ and $\tl{A}^\sigma$ are
\begin{align*}
0&= \tl{\mathbf{E}}_{\tl{\psi}} (L) \equiv \tl{\nabla}_\mu^{\bdag}\big(\tl{\nabla}^\mu\tl{\psi}\big)^{\! *}+m^2\tl{\psi}^*,\\
0&= \mathbf{E}_{\tl{\psi}^*}(L) \equiv \big(\tl{\nabla}_\mu^{\bdag}\big)^{\! *}\tl{\nabla}^\mu\tl{\psi}+m^2\tl{\psi},\\
0&= \tl{\mathbf{E}}_\sigma (L) \equiv \ \ri e\exp\!\big(\!-\ri e h^\sigma \tl{A}_{\sigma}\big)\tl{\psi}\big(\tl{\nabla}_\sigma \tl{\psi}\big)^{\! *} - \ri e\exp\!\big(\ri e h^\sigma \tl{A}_{\sigma}\big)\tl{\psi}^*\tl{\nabla}_\sigma\tl{\psi}-\eta_{\sigma\alpha}\ol{D}_\beta^{\,\bdag}\tl{F}^{\alpha\beta}.
\end{align*}
Note that $$\tl{\nabla}_\mu^{\bdag}=\ol{D}_{\mu}^{\,\bdag}+\frac{1}{h^\mu}\left\{1-\exp\big(\!-\ri e h^\mu \tl{A}_{\mu}\big)\right\}\text{id}$$ tends to $-\nabla_\mu^*$ as $h^\mu$ approaches zero, so the scheme has the correct limiting behaviour.
The above discretization preserves the gauge symmetries as variational symmetries. Explicitly, these are
$$\tl{\psi}\mapsto \exp(-\ri e\lambda)\tl{\psi},\qquad \tl{\psi}^*\mapsto \exp(\ri e\lambda)\tl{\psi}^*,\qquad \tl{A}^{\sigma}\mapsto \tl{A}^{\sigma}+\eta^{\sigma\alpha}\ol{D}_\alpha\lambda,$$
where now $\lambda$ is an arbitrary real-valued function of $\mathbf{n}$. As in the continuous case, the Lagrangian is invariant, so $\tl{X}L=0$.
The characteristics of the gauge symmetries have components
$$Q^{\tl{\psi}}=-\ri e \tl{\psi}\,\gtt,\qquad Q^{\tl{\psi}^*}\!=\ri e \tl{\psi}^*\gtt,\qquad Q^{\sigma}=\eta^{\sigma\alpha}\ol{D}_\alpha \gtt,$$
where $\gtt$ is an arbitrary real-valued function of $\mathbf{n}$. Consequently the discrete version of Noether's Second Theorem immediately yields the difference relation
\[
-\ri e\tl{\psi}\tl{\mathbf{E}}_{\tl{\psi}} (L)+\ri e\tl{\psi}^*\tl{\mathbf{E}}_{\tl{\psi}^*}(L)+\ol{D}_{\alpha}^{\,\bdag}\left(\eta^{\sigma\alpha} \tl{\mathbf{E}}_\sigma (L)\right)\equiv 0,
\]
which Christensen and Halvorsen obtained with the aid of Noether's first theorem.
\eex
\section{Variational symmetries with difference constraints}
If the functions $\gttb$ are subject to a (complete) set of $S$ linear difference constraints,
\beq\label{constrd}
\tl{\mathscr{D}}_{sr}\gtt^r=0,\qquad s=1,\dots,S,
\eeq
we incorporate these using the same approach as for differential equations. First form the new Lagrangian,
\beq\label{newLcond}
\hat{L}(\mathbf{n},[\mathbf{u};\gttb])=Q^\alpha(\mathbf{n},[\mathbf{u};\gttb])\tl{\mathbf{E}}_{\alpha}(L(\mathbf{n},[\mathbf{u}]))-
\nu^s\tl{\mathscr{D}}_{sr}(\gtt^r),
\eeq
where $\nu^1,\dots,\nu^s$ are Lagrange multipliers. Then take variations with respect to $\gttb$ to obtain
\beq\label{EgQEcond}
\mathbf{S}_{-\mathbf{J}}\left(\frac{\p Q^\alpha(\mathbf{n},[\mathbf{u};\gttb])}{\p \mathbf{S}_{\mathbf{J}}\gtt^r}\,\tl{\mathbf{E}}_{\alpha}(L)\right)\equiv(\tl{\mathscr{D}}_{sr})^{\bdag}(\nu^s),\qquad r=1,\dots,R.
\eeq
Just as in the continuous case, these relations yield conservation laws of the Euler--Lagrange equations (\ref{twelve}), namely
\beq\label{CLd}
\nu^s\tl{\mathscr{D}}_{sr}\gtt^r-\gtt^r(\tl{\mathscr{D}}_{sr})^{\bdag}(\nu^s)=0\qquad \text{on solutions of (\ref{twelve})}.
\eeq
It may also be possible to eliminate the Lagrange multipliers from (\ref{EgQEcond}) to obtain difference relations between the Euler--Lagrange equations.
\bex
The lattice Korteweg--de Vries (KdV) equation is
\beq\label{lkdv}
U_{1,1}-U_{0,0}-c\left(\frac{1}{U_{1,0}}-\frac{1}{U_{0,1}}\right)=0,\qquad c\neq 0,
\eeq
where $U_{i,j}$ denotes the value of the dependent variable $U$ at the point $(n^1+i,n^2+j)$; see Grammaticos \textit{et al.} (1991) for details. This is not an Euler--Lagrange equation as it stands, but it can be turned into one by introducing a potential, $u$, such that
\[
U_{i,j}=u_{i,j-1}-u_{i-1,j}.
\]
Then (\ref{lkdv}) amounts to
\[
\tl{\mathbf{E}}_u(L)\equiv u_{1,0}-u_{0,1}+u_{-1,0}-u_{0,-1}-c\left(\frac{1}{u_{1,-1}-u_{0,0}}-\frac{1}{u_{0,0}-u_{-1,1}}\right)=0,
\]
where
\[
L[u]=u_{0,0}(u_{1,0}-u_{0,1})+c\ln(u_{1,0}-u_{0,1}).
\]
Clearly, the variable $U$ is unaffected by gauge transformations
\[
u_{i,j}\mapsto u_{i,j}+\lambda_{i,j},\qquad\text{where}\quad \lambda_{1,0}=\lambda_{0,1},
\]
so these are symmetries of the Euler--Lagrange equation. Their characteristic is $Q=\gtt_{0,0}$, where $\gtt_{1,0}-\gtt_{0,1}=0$; therefore
\[
\tl{X}(L)=\gtt_{0,0}(u_{1,0}-u_{0,1})=\tl{D}_1\big(\gtt_{\!\,-1,0}\,u_{0,0}\big)+\tl{D}_2\big(\!-\gtt_{0,\!\,-1}\,u_{0,0}\big),
\]
and so the gauge symmetries are variational. Now vary $\gtt$ for the new Lagrangian
\[
\hat{L}[u;\gtt]=Q\tl{\mathbf{E}}_u(L[u])-\nu_{0,0}(\gtt_{1,0}-\gtt_{0,1})
\]
to obtain the relation
\[
\tl{\mathbf{E}}_u(L[u])\equiv \nu_{\!\,-1,0}-\nu_{0,\!\,-1}.
\]
One obvious solution is
\beq\label{nu00}
\nu_{0,0}=u_{0,0}-u_{1,1}+\frac{c}{u_{1,0}-u_{0,1}}\,;
\eeq
the corresponding conservation laws (\ref{CLd}) are
\[
\tl{D}_1\big(\gtt_{0,0}\,\nu_{\!\,-1,0}\big)+\tl{D}_2\big(\!-\gtt_{0,0}\,\nu_{0,\!\,-1}\big)=0\qquad\text{when}\quad \tl{\mathbf{E}}_u(L)=0.
\]
Note that by setting $\nu_{0,0}=0$ in (\ref{nu00}), we obtain the well-known discrete potential KdV equation,
\beq\label{dpkdv}
(u_{1,1}-u_{0,0})(u_{1,0}-u_{0,1})=c.
\eeq
So the constraint on the gauge symmetry provides a link between the lattice KdV equation and
its related potential form. This observation applies similarly to many integrable
difference equations.
\eex
\section{Conclusion}
We have shown there is a complete correspondence between Noether's Second Theorem for differential and difference equations, and that the same is true where characteristics of variational symmetries depend on partly-constrained functions. Our approach yields, \textit{mutatis mutandis}, the corresponding results for differential-difference equations.
|
\section{Introduction}
Two basic cosmological questions are how do galaxies form and how do they achieve their present structure. The current approach to answer these questions involves the comparison of structure-formation models to galaxy observations at low- and high-redshift. The stellar masses of galaxies can increase through major mergers, accretion of smaller satellite systems, and star formation by converting gas into stars; some of this gas can be externally accreted. Guo \& White (2008) showed that the relative importance of these three modes is a strong function of the stellar mass in the specific galaxy. Galaxy growth through major mergers depends strongly on stellar mass, minor mergers contribute more to galaxy growth than major mergers at all redshifts and stellar masses, and in galaxies significantly less massive than the Milky Way, star formation dominates the growth.
Gavazzi \& Scodeggio (1996) found that the star formation in late-type galaxies (Sa and later) is most probably regulated by the total mass of a galaxy. Their assumptions in reaching this conclusion were that galaxies develop in isolation, that they all have solar metallicity, that the star formation rate (SFR) decreases exponentially with time with an e-folding time depending on the galaxy mass, and that the initial mass function is ``Salpeter'', with stars from 0.1 to 120 M$_{\odot}$.
Balogh et al. (2004) identified two galaxy families in the colour distribution of Sloan Digital Sky Survey (SDSS) objects; a blue star-forming family and a red, passively-evolving population. They concluded that transitions between the two families can take place on short timescales of a few Gyr. Such transitions can be caused by mass influx from a merger with a small, gas-rich galaxy (Martin et al. 2007).
In normal, inactive galaxies, the stellar velocity dispersion of the bulge $\sigma_*$ tracks the maximal rotation velocity of the disk roughly as $v_m \simeq 1.7 \times \sigma_*$ (Whitmore et al. 1979). Although the theoretical basis of this correlation is still not clear, and the $v_m - \sigma_*$ relation is not as tight as has been claimed (e.g., Pizzella et al. 2005, specifically for LSB galaxies), nevertheless an empirical relation between $v_m$ and $\sigma_*$ does exist (Courteau et al. 2007). This implies that $\sigma_*$ can, in principle, be estimated from $v_m$. Since $v_m$ can be measured from HI observations for galaxies that are sufficiently gas-rich, in the absence of a resolved rotation curve $v_m$ can be estimated from a single-dish HI measurement while the inclination angle of the disk can be derived from optical observations. Given the rotation velocity, it is possible to deduce the total galaxy luminosity through the Tully-Fisher relation (Tully \& Fisher 1977) and the dynamical mass of the system given an estimate of the disk size from optical imaging. It is not clear, however, whether these assumptions can be applied also to any random sample of galaxies, although this seems to be the case at least for the ``All Digital HI catalog'' objects (ADHIC: Courtois et al. 2009).
The Arecibo Legacy Fast ALFA (ALFALFA) survey is an on-going, second generation, blind extragalactic HI survey using the seven-feed L-band focal plane array (ALFA) at the Arecibo Observatory (AO). The seven beams have sizes of 3\arcmin.3 along the azimuth direction and 3\arcmin.8 along the zenith angle direction. ALFALFA is performed by drift-scanning ribbons of the sky and repeating these after a few months. The combination of multiple scans with the fast one-second sampling (14 samples per source beam transit time) allows the centroiding of HI sources to much less than a beam width (Giovanelli et al. 2005a, b). ALFALFA will eventually survey more than 7000 square degrees of the high galactic latitude sky and is expected to detect more than 25000 extragalactic HI sources up to a redshift of 18000 km s$^{-1}$ exploiting AO's superior sensitivity, angular resolution and digital technology. The survey aims primarily to probe the faint end of the HI mass function (HIMF) and will provide a complete HI census in the surveyed sky area. The ALFALFA survey strategy was described by Giovanelli et al. (2005a) and survey results for selected sky regions were presented by Giovanelli et al. (2005b, 2007), Saintonge et al. (2007) and Kent et al. (2008). A revision of the HI mass function and a determination of $\Omega_{HI}$ from the 40\% of the total survey area with complete source extraction was published by Martin et al. (2010). The source detection algorithm and method were described by Saintonge (2007). ALFALFA is the best deep and unbiased HI survey now in existence.
The ALFALFA survey yields a number of parameters for each detection. The ones most often used are the position and radial velocity (to characterize the location of each object in 3D space), and the total HI flux (to derive the HI mass). We use here an additional parameter, $w(50)$ that is the width of the HI profile at 50\% of the peak flux density, to select a sample of objects with HI profiles wider than 550 km s$^{-1}$ from the published ALFALFA catalogs and associated data products, and study it. The $w(50)$ parameter can be used to estimate the maximal rotation velocity of the gas in a galaxy (e.g., Courtois et al. 2009), since for disk galaxies $w(50)=2\times v_{max}sin(i)$.
There could be a number of reasons why a galaxy should show a very wide HI profile in a single-dish observation: it could have a high mass implying a fast asymptotic rotation velocity, it could originate in a galaxy showing not only regular rotation but also some chaotic dynamics, or it could result from the detection of a confused binary or multiple galaxy system where the entire HI profile in the relatively wide ALFALFA beam is contributed by two or more objects that may be strongly interacting (as found by e.g., Bothun et al. 1982). We expect to be able to distinguish between these two possibilities from inspecting the optical galaxy images and the galaxies' HI profiles, and from the characterization of the CHMD's galaxy neighbourhood, thus being able to select a clean, HI-based, sample of high-mass objects.
Trachternach et al. (2009) studied the baryonic Tully-Fisher (TF) relation in a sample of very low-mass dwarf galaxies. They concluded that the baryonic TF relation is followed by all rotationally-dominated galaxies. Since the relation was established using low-to-intermediate rotational velocities, it is interesting to see how well it reproduces the behaviour of high rotational velocities galaxies.
High-mass objects observed at present represent strong deviations of the initial density fluctuation field in the early Universe. It is possible that interacting galaxies also originate from such strong density fluctuations. In this case, the surroundings of galaxies with wide HI profiles should be characterized by a higher galaxy density than at random locations. This also is tested in our paper, where we adopt cosmology-corrected quantities: H$_0$ = 73 km s$^{-1}$ Mpc$^{-1}$, $\Omega_{matter}$=0.27 and $\Omega_{vacuum}$=0.73, as in NED.
The connection between HI gas and galaxy mass at the high-mass end was recently studied by Catinella et al. (2010). They found that the gas-to-stellar-mass ratio decreases with stellar mass and stellar surface mass density. Since we will be deriving both quantities here, it will be interesting to compare our results with those of Catinella et al.
\section{The sample}
\label{txt:sample}
The sample studied here consists of 28 objects selected from the ALFALFA observations. We stress that the selection was done on the ALFALFA data set available only to the ALFALFA collaboration, including sources not yet published. All the objects with recession velocity smaller than 12000 km s$^{-1}$, HI profile width $w(50)\geq$550 km s$^{-1}$ and ALFALFA detection code 1 (implying high signal-to-noise and high confidence detection) were selected from the ALFALFA region that has been fully covered by the survey, i.e., 7$^h$.5$\leq$RA$\leq$16$^h$.5 and +4$^{\circ}\leq$Dec$ \leq +16^{\circ}$, which is $\sim$20\% of the final ALFALFA survey coverage. An ALFALFA detection code of 1 implies a S/N$\geq$6.5, but this cannot be translated into a flux density limit since the S/N depends, among others, on the HI line width (see e.g. Giovanelli et al. 2007). As an example, a source with a flux density of 0.72 Jy km sec$^{-1}$ would be detected at 5$\sigma$ if its line width would be 200 km sec$^{-1}$ (Giovanelli et al. 2005b, Figure 8).
The sample is therefore complete according to the selection criteria listed here and is unbiased regarding optical brightness or surface brightness criteria. The volume surveyed to detect these 28 objects is 7.2$\times10^5$ Mpc$^3$; the corresponding volume density of these very rare objects is 3.9$\times10^{-5}$ Mpc$^{-3}$. By selecting objects closer than 12000 km sec$^{-1}$ we avoid the outer edge of the ALFALFA survey, which is the velocity bin 120000$\leq v \leq$18000 km s$^{-1}$.
The lowest recession velocity present in our sample is 4933 km s$^{-1}$ excluding also the low-velocity segment of each ALFALFA data cube that contains ``local'' objects, such as the Virgo Cluster
The region has full SDSS coverage; it is thus possible to extract broad-band total magnitudes and colors as well as images with reasonable resolution for the morphological classification of the selected objects. SDSS also yields radial velocities for the sample objects and for some of their optical neighbours, while ALFALFA yields radial velocities for the optically-faint but HI-rich neighbours. We aim to characterize the neighbourhood galaxy density of each object by the distance to its nearest catalogued neighbor and by the number of objects in a preset volume, say within 3h$^{-1}$ Mpc projected distance and 300 km s$^{-1}$ velocity difference.
The galaxies selected from the ALFALFA survey are listed in Table 1. We give there the name of the object as used in the ALFALFA survey (the AGC designator), the derived HI position and the optical center of the galaxy from SDSS (both in J2000 coordinates), its heliocentric recession velocity and error (in parentheses) in km s$^{-1}$, the full width at half-maximum of the HI line profile and its error (in parentheses) in km s$^{-1}$, and the HI flux integral FI and its error (in parentheses) in Jy km s$^{-1}$.
The heliocentric velocity is measured as the midpoint between the channels where the flux density drops to 50\% of each of the two peaks (or of one, if only one is present) at each side of the spectral feature and is measured in km s$^{-1}$, as explained e.g., in Giovanelli et al. (2007). The velocity width of the source line profile, w(50), is measured at the 50\% outer level of each of the two peaks and is corrected for instrumental broadening.
We checked carefully all 28 objects in the original selection to reject those that could have a wide HI profile by being a confused pair or multiple object, or by being affected by interactions to show a disturbed appearance. This was done by retaining only galaxies where a two-horned profile could clearly be seen, and by inspecting the SDSS image and radial velocity data to identify morphological disturbances or possible nearby companions. Information about the inspection of SDSS images and spectral data is given in Appendix A; the HI profiles are shown in Appendix B. The galaxies identified here as having an intrinsic wide HI profile, and retained as such, are marked in bold font in Table 1.
\begin{table*}
\caption{ALFALFA galaxies with wide HI profiles}
\label{t:sample}
{\small
\begin{tabular}{cccccc} \hline
Object & HI ellipse & Optical & v$_{50}$ & w(50) & FI
\\
AGC name & (J2000) $\alpha, \delta$ & (J2000) $\alpha, \delta$ & km s$^{-1}$ & km s$^{-1}$ & Jy km s$^{-1}$
\\ \hline
181756 & 082656.0+073011 & 082654.6+072953 & 9323(43) & 685(86) & 4.26(0.13) \\
192281 & 095900.0+130305 & 095859.9+130309 & 10470(3) & 885(5) & 3.62(0.17) \\
{\bf 205190} & 104801.9+142431 & 104802.2+142429 & 9683(4) & 698(7) & 2.6(0.12) \\
{\bf 200589} & 104837.6+125846 & 104834.1+125859 & 10768(3) & 553(7) & 1.73(0.11) \\
006042 & 105611.8+094544 & 105615.4+094515 & 9921(6) & 568(12) & 1.61(0.1) \\
201697 & 105744.9+151755 & 105744.8+151827 & 10955(3) & 650(7) & 1.48(0.1) \\
{\bf 006066} & 105859.7+063141 & 105859.3+063121 & 11807(2) & 667(4) & 6.25(0.14) \\
{\bf 222042} & 120305.6+051023 & 120305.0+051028 & 11569(5) & 555(9) & 2.37(0.14) \\
226077 & 122518.4+160740 & 122514.2+160713 & 9250(2) & 582(4) & 2.1(0.12) \\
226111 & 125934.9+150248 & 125936.0+150257 & 10678(1) & 566(3) & 2.38(0.1) \\
233609 & 130711.6+133918 & 130711.6+133933 & 8143(6) & 624(11) & 2.53(0.13) \\
{\bf 008375} & 131954.6+155105 & 131956.3+155101 & 7007(34) & 696(68) & 2.39(0.14) \\
{\bf 008379} & 132021.8+062019 & 132020.3+062010 & 11967(6) & 569(11) & 2.93(0.12) \\
{\bf 008475} & 132925.7+110015 & 132925.8+110028 & 6835(1) & 591(2) & 15.71(0.14) \\
{\bf 008488} & 133003.3+132511 & 133002.3+132458 & 7363(5) & 596(11) & 4.03(0.13) \\
008559 & 133453.9+135008 & 133455.9+134956 & 7025(12) & 572(23) & 4.31(0.12) \\
008766 & 135133.1+140515 & 135130.9+140530 & 6996(8) & 622(16) & 3.1(0.13) \\
008902 & 135903.1+153411 & 135902.8+153357 & 7559(3) & 691(6) & 20.14(0.13) \\
{\bf 230914} & 140004.6+120630 & 140004.4+120641 & 11848(5) & 580(11) & 3.89(0.13) \\
008943 & 140213.5+080235 & 140213.1+080212 & 4933(49) & 669(98) & 1.67(0.1) \\
{\bf 009031} & 140740.4+145204 & 140739.3+145151 & 11864(3) & 709(6) & 2.76(0.16) \\
{\bf 248977} & 144110.9+145327 & 144112.3+145324 & 9153(19) & 560(38) & 2.13(0.13) \\
{\bf 009624} & 145735.0+081726 & 145735.8+081706 & 11068(18) & 599(36) & 2.74(0.12) \\
009788 & 151536.1+081735 & 151538.7+081803 & 10178(9) & 564(17) & 3.77(0.14) \\
009794 & 1516 9.1+103033 & 151610.8+103034 & 6406(2) & 630(5) & 22.03(0.13) \\
009838 & 152512.3+070823 & 152512.2+070916 & 10267(9) & 575(19) & 2.59(0.15) \\
{\bf 260110} & 160448.5+140741 & 160448.1+140744 & 10160(8) & 610(15) & 5.99(0.16) \\
{\bf 010272} & 161257.3+095143 & 161256.9+095201 & 5124(31) & 580(62) & 1.72(0.1) \\
\hline
\end{tabular}
}
\end{table*}
We identified 14 galaxies with wide HI profiles that did not show signs of interaction nor had nearby ``significant'' neighbour galaxies; these might be truly high-mass objects. The objects chosen from the ALFALFA selection as candidate high-mass galaxies (CHMGs, see below) were observed in the R-band and in rest-frame H$\alpha$, using the Wise Observatory narrow-band filter set. These observations
allow the determination of the H$\alpha$ equivalent width. This, together with the absolutely calibrated $ugriz$ magnitudes from SDSS
and other public-domain data where available, allows the determination of the total H$\alpha$ emission line flux and of an approximate star formation rate (SFR) and star formation history (SFH) for each object.
\subsection{H$\alpha$ observations}
H$\alpha$ imaging of 14 CHMDs was carried out at the Wise Observatory on 8 nights, from March 2009 to April 2010, using the 40-inch telescope with the PI CCD camera (pixel scale of 0.6 arcsec$\cdot$pixel$^{-1}$) and narrow-band H$\alpha$ filters centered
approximately on the wavelength of the redshifted H$\alpha$ line of each galaxy. At least three 20-minute dithered exposures in the narrow-band H$\alpha$ filter which best fits its redshift and three 5-minute dithered exposures in R were obtained for each galaxy.
The images in the same band were debiased, flat-fielded, sky-subtracted, and combined into a final image for each filter.
In order to derive the net line emission contribution (that includes also that of the
[NII] lines), we subtracted the continuum contribution from a properly scaled R-band image. The scaling was done using stars, assuming that stars would not contribute
specific emission or absorption features in the narrow redshifted band, and that their contribution in R would only be continuum. This is justified, since even the lowest redshift H$\alpha$ filter we used is significantly distant from the zero-redshift H$\alpha$ line that could appear in the spectra of the scaling stars.
H$\alpha$ equivalent widths (EW) were obtained using uncalibrated flux measurements of the galaxies in the R and net-H$\alpha$ images and the transmission curves of the H$\alpha$ and R filters. Calibrated H$\alpha$ flux measurements were obtained using the measured EW values and the continuum flux per frequency at the H$\alpha$ redshifted wavelength, estimated by linearly interpolating between the SDSS measured $r$ and $i$ fluxes.
\subsection{SDSS and 2MASS data}
Catalog data were extracted for each sample galaxy from the SDSS
and are listed in Table~\ref{t:sample2}. The table lists a shortened name for each object obtained by truncating the position of the object from column 1 of Table~\ref{t:sample} to retain the first four right ascension digits, the declination sign and two declination digits, and the total $ugriz$ magnitudes from SDSS.
Table~\ref{t:sample3}
lists the H$\alpha$ emission line equivalent width in \AA\, and flux in erg cm$^{-2}$ sec$^{-1}$ determined as explained above, the GALEX FUV and NUV magnitudes with errors for the few objects where these data are available, and the 2MASS J, H, and K$_s$ magnitudes with errors. Given the lack of approximately uniform coverage by GALEX of the objects in our sample, we decided to only list the FUV and/or NUV magnitudes without using them for further investigations.
\begin{table*}
\caption{SDSS total photometry for the sample galaxies}
\label{t:sample2}
\begin{tabular}{cccccc} \hline
ACG & $u$ & $g$ & $r$ & $i$ & $z$ \\ \hline
181756 &17.42$\pm$0.02 &15.42$\pm$0.00 &14.48$\pm$0.00 &14.05$\pm$0.00&13.66$\pm$0.00 \\
192281 &17.95$\pm$0.02 &16.35$\pm$0.00 &15.51$\pm$0.00&15.05$\pm$0.00&14.71$\pm$0.01\\
205190 &17.09$\pm$0.02 &14.93$\pm$0.00&13.82$\pm$0.00&13.52$\pm$0.00&12.92$\pm$0.00\\
200589 &16.76$\pm$0.01 &14.77$\pm$0.00&13.87$\pm$0.00&13.42$\pm$0.00&13.10$\pm$0.00\\
006042 &16.35$\pm$0.01 &14.40$\pm$0.00&13.54$\pm$0.00&13.15$\pm$0.00&12.81$\pm$0.00\\
201697 &17.82$\pm$0.03 &16.07$\pm$0.00&15.22$\pm$0.00&14.79$\pm$0.00&14.44$\pm$0.00\\
006066 &16.98$\pm$0.02 &15.02$\pm$0.00&14.16$\pm$0.00&13.75$\pm$0.00&13.38$\pm$0.00\\
222042 &17.00$\pm$0.02 &15.00$\pm$0.00 &14.12$\pm$0.00&13.68$\pm$0.00&13.35$\pm$0.00\\
226077 &17.10$\pm$0.02 &15.59$\pm$0.00 &14.85$\pm$0.00&14.47$\pm$0.00&14.17$\pm$0.00\\
226111 &17.07$\pm$0.01 &15.27$\pm$0.00&14.40$\pm$0.00&13.97$\pm$0.00&13.59$\pm$0.00\\
233609 &17.71$\pm$0.02 &16.71$\pm$0.00&16.41$\pm$0.00&16.22$\pm$0.01&15.92$\pm$0.01\\
008375 &16.11$\pm$0.01 &14.16$\pm$0.00&13.32$\pm$0.00&12.87$\pm$0.00&12.54$\pm$0.00\\
008379 &16.68$\pm$0.02 &14.91$\pm$0.00&14.08$\pm$0.00&13.66$\pm$0.00&13.33$\pm$0.00\\
008475 &15.74$\pm$0.01 &13.70$\pm$0.00&12.74$\pm$0.00&12.33$\pm$0.00&11.93$\pm$0.00\\
008488 &16.20$\pm$0.01 &14.15$\pm$0.00&13.25$\pm$0.00&12.80$\pm$0.00&12.49$\pm$0.00\\
008559 &16.11$\pm$0.01 &14.05$\pm$0.00&13.07$\pm$0.00&12.58$\pm$0.00&12.23$\pm$0.00\\
008766 &15.66$\pm$0.01 &13.78$\pm$0.00&12.90$\pm$0.00&12.41$\pm$0.00&12.01$\pm$0.00\\
008902 &16.10$\pm$0.01 &14.09$\pm$0.00&13.21$\pm$0.00&12.81$\pm$0.00&12.41$\pm$0.00\\
230914 &16.57$\pm$0.01 &14.71$\pm$0.00&13.89$\pm$0.00&13.48$\pm$0.00&13.18$\pm$0.00\\
008943 &15.25$\pm$0.01 &13.30$\pm$0.00&12.45$\pm$0.00&12.03$\pm$0.00&11.74$\pm$0.00\\
009031 &16.83$\pm$0.01 &14.98$\pm$0.00&14.07$\pm$0.00&13.58$\pm$0.00&13.13$\pm$0.00\\
248977 &17.77$\pm$0.02 &15.94$\pm$0.00&15.14$\pm$0.00&14.74$\pm$0.00&14.40$\pm$0.00\\
009624 &16.22$\pm$0.01 &14.32$\pm$0.00&13.45$\pm$0.00&12.99$\pm$0.00&12.64$\pm$0.00\\
009788 &16.83$\pm$0.02 &15.13$\pm$0.00&14.22$\pm$0.00&13.70$\pm$0.00&13.28$\pm$0.00\\
009794 &19.14$\pm$0.03 &17.13$\pm$0.00&16.18$\pm$0.00&15.62$\pm$0.00&14.96$\pm$0.00\\
009838 &16.73$\pm$0.02 &14.82$\pm$0.00&13.89$\pm$0.00&13.42$\pm$0.00&13.04$\pm$0.00\\
260110 &17.13$\pm$0.01 &15.10$\pm$0.00&14.18$\pm$0.00&13.72$\pm$0.00&13.33$\pm$0.00\\
010272 &15.79$\pm$0.01 &13.78$\pm$0.00&12.86$\pm$0.00&12.34$\pm$0.00&11.98$\pm$0.00\\ \hline
\end{tabular}
Note: Magnitude errors smaller than 0.01 are reported here as 0.00.
\end{table*}
\begin{table*}
\caption{H$\alpha$, GALEX, and 2MASS total photometry for the galaxies in the sample}
\label{t:sample3}
\begin{tabular}{cccccccc} \hline
ACG & EW H$\alpha$ & F$_{H\alpha}$ & FUV & NUV & J & H & K$_s$ \\
& \AA\ & $\times10^{-13}$ & AB mag & AB mag & mag & mag & mag \\ \hline
181756& - & - & - & - & 12.39$\pm$0.03 & 11.66$\pm$0.04 & 11.51$\pm$0.06 \\
192281 & - & - & - & 17.79$\pm$0.02 & 13.57$\pm$0.05 & 12.72$\pm$0.06 & 12.31$\pm$0.06 \\
205190 & 14$\pm$02 & 1.1$\pm$0.2 & - & - & 11.70$\pm$0.02 & 10.99$\pm$0.03 & 10.66$\pm$0.04 \\
200589 & 19$\pm$03 & 1.6$\pm$0.3 & - & - & 11.84$\pm$0.02 & 11.15$\pm$0.03 & 10.85$\pm$0.04 \\
006042 & - & - & - & - & 11.29$\pm$0.03 & 10.60$\pm$0.03 & 10.46$\pm$0.06 \\
201697 & - & - & - & - & 11.29$\pm$0.03 & 11.00$\pm$0.03 & 10.46$\pm$0.06 \\
006066 & 19$\pm$06 & 1.2$\pm$0.4 & - & - & 11.82$\pm$0.04 & 11.18$\pm$0.05 & 10.96$\pm$0.08 \\
222042 & 21$\pm$04 & 1.4$\pm$0.2 & - & - & 12.04$\pm$0.04 & 11.50$\pm$0.05 & 11.09$\pm$0.07 \\
226077 & - & - & - & - & 13.44$\pm$0.08 & 12.67$\pm$0.10 & 12.23$\pm$0.09 \\
226111 & - & - & - & - & 12.43$\pm$0.03 & 11.63$\pm$0.04 & 11.38$\pm$0.04 \\
233609 & - & - & - & - & 14.12$\pm$0.11 & 13.59$\pm$0.17 & 13.10$\pm$0.16 \\
008375 & 10$\pm$04 & 1.3$\pm$0.6 & - & 19.30$\pm$0.06 & 11.26$\pm$0.02 & 10.60$\pm$0.02 & 10.30$\pm$0.03 \\
008379 & 25$\pm$08 & 1.7$\pm$0.5 & - & 17.56$\pm$0.05 & 12.30$\pm$0.05 & 11.53$\pm$0.06 & 11.23$\pm$0.08 \\
008475 & 34$\pm$06 & 7.8$\pm$1.5 & 19.17$\pm$0.14 & - & 10.29$\pm$0.03 & 9.63$\pm$0.03 & 9.33$\pm$0.04 \\
008488 & 16$\pm$18 & 2.3$\pm$2.7 & 20.42$\pm$0.30 & - & 11.12$\pm$0.02 & 10.56$\pm$0.03 & 10.24$\pm$0.04 \\
008559 & - & - & - & - & 10.88$\pm$0.02 & 10.26$\pm$0.03 & 10.05$\pm$0.03 \\
008766 & - & - & - & - & 10.75$\pm$0.02 & 10.00$\pm$0.02 & 9.67$\pm$0.03 \\
008902 & - & - & - & - & 11.09$\pm$0.02 & 10.23$\pm$0.02 & 9.95$\pm$0.03 \\
230914 & 29$\pm$06 & 2.3$\pm$0.5 & 20.99$\pm$0.32 & - & 12.01$\pm$0.03 & 11.26$\pm$0.04 & 10.97$\pm$0.04 \\
008943 & - & - & - & - & 10.30$\pm$0.02 & 9.63$\pm$0.03 & 9.37$\pm$0.03 \\
009031 & 32$\pm$08 & 2.3$\pm$0.6 & 22.34$\pm$0.58 & - & 11.96$\pm$0.03 & 11.13$\pm$0.04 & 10.79$\pm$0.04 \\
248977 & 09$\pm$02 & 0.2$\pm$0.05 & - & 20.69$\pm$0.32 & 13.02$\pm$0.05 & 12.65$\pm$0.10 & 12.18$\pm$0.09 \\
009624 & 26$\pm$04 & 3.2$\pm$0.5 & - & 18.28$\pm$0.06 & 11.46$\pm$0.03 & 10.73$\pm$0.03 & 10.56$\pm$0.05 \\
009788 & - & - & - & - & - & - & - \\
009794 & - & - & - & - & 10.68$\pm$0.02 & 9.80$\pm$0.03 & 9.48$\pm$0.03 \\
009838 & - & - & - & 19.21$\pm$0.12 & 11.82$\pm$0.03 & 11.23$\pm$0.04 & 10.90$\pm$0.05 \\
260110 & 27$\pm$05 & 1.7$\pm$0.3 & - & - & 12.11$\pm$0.02 & 11.42$\pm$0.03 & 11.10$\pm$0.03 \\
010272 & 23$\pm$03 & 5.1$\pm$0.6 & 21.28$\pm$0.31 & 19.05$\pm$0.09 & 10.72$\pm$0.02 & 9.98$\pm$0.02 & 9.74$\pm$0.03 \\ \hline
\end{tabular}
Notes:\\
(1) Magnitude errors smaller than 0.01 are reported here as 0.00. \\
(2) 201697 and 009788 have no 2MASS data. \\
(3) 2MASS data for ACG 248977 are those listed for 2MASX J14411233+1453242. \\
(4) Only the candidate high-mass galaxies have H$\alpha$ measurements. The fluxes are given in units of erg cm$^{-2}$ s$^{-1}$.
\end{table*}
\section{Results}
\label{txt:results}
The basic question stated above is whether the wide HI profiles detected by the ALFALFA survey are produced by single, non-interacting galaxies or by two or more galaxies that are interacting and are too close together to be resolved by the rather wide ALFALFA beam. This has been shown to be the case for the galaxy with the widest HI profile reported by Bothun et al. (1982).
The relevant information is contained in the ALFALFA HI profile. We expect this profile to have a clean two-horned shape for non-interacting disky galaxies. Courtois et al. (2009) explained this in their description of some ``good, bad, and ugly'' profiles in the ADHIC: PGC 71392, in their Figure 6, shows a clean profile with $w(50)$=989 km sec$^{-1}$ and is a high inclination S0/Sa (see also Giovanelli et al. 1986); PGC 10314 has $w(50)$=33 km sec$^{-1}$ and appears to be face-on, and the two nearby galaxies PGC 68870 and PGC 68878 show contamination of the HI profile of one galaxy by that of the other.
We checked the ALFALFA profiles of all galaxies in Table~\ref{t:sample} and
show in Figure~\ref{Fig:Good-bad-ugly} examples of the three kinds of profiles from among the objects in our sample.
\begin{figure*}
{\centering
\includegraphics[clip=,angle=0,width=11.3cm]{HI160448.0+140752_small.EPS}
\includegraphics[clip=,angle=0,width=11.3cm]{HI133453.6+135015_small.EPS}
\includegraphics[clip=,angle=0,width=11.3cm]{HI122517.8+160754_small.EPS}
}
\caption{%
{\it Top panel:} A ``good'' HI profile; ACG 260110, a fairly isolated galaxy.
{\it Middle panel:} A ``bad'' profile; ACG 008559, an interacting galaxy with a strong warp or tilt of the outer disk.
{\it Bottom panel:} An ``ugly'' profile; ACG 226077, a galaxy whose HI profile is affected not only by the nearby neighbour that is included in the ALFALFA beam, but also by radio interference for velocities below $\sim$9000 km sec$^{-1}$.
}
\label{Fig:Good-bad-ugly}
\end{figure*}
\subsection{High-mass candidate galaxies and their neighbourhoods}
We found here that, in general, the 28 sample galaxies have large physical sizes and are very luminous. The largest object is 82 kpc wide while the smallest is only $\sim$12 kpc. In many cases the target galaxy had nearby objects that would have been included in the ALFALFA beam. In cases where redshifts were available, some objects were shown to be physical companions of the target galaxy. We also found that some companions showed emission-line spectra and blue continua, indicating fairly recent star formation possibly triggered by a past interaction. Other companions had spectra of early-type galaxies but showed also Balmer absorption lines; such cases could possibly be similar to E+A galaxies (Dressler \& Gunn 1983) indicating the presence of a relatively large population of A and B stars along with an old population dominated by G, K, and M spectral types, representing possibly a star formation episode about 1 Gyr ago, with the stellar bulk being much older.
Only 14 galaxies among the 28 selected from the ALFALFA data set appear to be fairly isolated, since they do not seem to have SDSS companions closer than $\sim$10 arcmin that could be confused in the derivation of the HI line width; these are candidates for being high-mass objects and are listed in Table 4. Here and in Appendix A we give additional details about each object and about their neighbourhood properties. In particular, we add in column 10 of Table 4 the number of identified neighbours in a 3h$^{-1}$ Mpc and 300 km sec$^{-1}$ volume around the galaxy; these could well be physical companions that may have interacted with the galaxy in the not-too-distant past. The data for the neighbourhood search are primarily from NED, supplemented by information from ALFALFA.
Table 4 lists the luminosity of the candidate galaxies in the $r$-band using M$_{\odot}$(r)=+4.52 mag
(from www.ucolick.org/$\sim$cnaw/sun.html),
the size of
their major axes in kpc using the angular size and the luminosity-distance from NED, the inclination $i$ calculated from the SDSS isophotal axes a and b
and cos(i)=b/a, the maximal rotation velocity corrected for inclination from v$_{max} = \frac {w(50)}{2 \times sin(i)}$, where $w(50)$ is the HI profile width listed in Table 1, the dynamical mass from
M$_{dyn} = \frac {v^2_{max} R}{G}$,
where R is the maximal extent of the visible galaxy with the power-of-ten of the value
shown in parentheses, M$_*$ an indicative stellar mass derived from the $r$-band luminosity by assuming all stars are K5V (this estimate will be refined below), and M$_{HI}$, the mass in atomic hydrogen from the ALFALFA flux integral and the luminosity distance listed in NED, using M$_{\rm HI}$=2.36$\times 10^5$ D$^2_{\rm Mpc} \times$FI,
where D$_{\rm Mpc}$ is the distance to the object in Mpc. Note that all estimates were made disregarding Milky Way and internal extinction, and also not correcting M$_{\rm HI}$ for the helium content to arrive at an estimate of the diffuse baryon masses. Also, Bernardi et al. (2010) list the $r$-band absolute magnitude of the Sun as +4.67; this would imply a 15\% brighter luminosity for all the objects. Column 9 of the table shows
F=$\frac{M_{dyn}}{M_{\rm gas}+M_{\rm stars}}$, the ratio of the dynamical mass to the total baryonic mass (stars and gas, now correcting for the helium content with M$_{gas}$=1.3$\times$M$_{\rm HI }$). Figures~\ref{f:candidates0},~\ref{f:candidates1},~\ref{f:candidates2},
and~\ref{f:candidates3} in Appendix B show the HI profiles of these CHMD.
The table also lists the error in the derived parameters; in some cases e.g. ACG 009624, the error in the derived dynamical mass is similar to the value itself. In most cases, the rather large errors are dominated by the error in the inclination.
We stress that the calculated value for M$_{dyn}$ is a lower limit, since it assumes that the HI distribution extends only as far as the 25 mag arcsec$^{-1}$ isophote of the optical image; in most disk galaxies this is not the case and the HI extends further out of the luminous disk by a factor of a few. On the other hand, this calculation also assumes that the gas dynamics are dominated by circular rotation. M$_*$ is also a lower limit, since we did not correct the photometry for Milky Way and internal extinction.
\section{Discussion}
\label{txt:discussion}
\subsection{Simplified star formation history for CHMDs}
The availability of the integrated SDSS colours allows a comparison of the 14 candidate high-mass galaxies with the relations found by Wu et al (2007) between the SDSS colour indices and the absolute magnitudes of galaxies. This indicates that, in most cases, the integrated colours can be produced by evolved stellar populations (SPs). In disk galaxies such colours correspond to high-luminosity systems, as their Figure 3 shows. The CHMDs are thus definitely on the ``red sequence'' of galaxies.
The present star formation rate (SFR) can be derived from the H$\alpha$ luminosity, following e.g., Pflamm-Altenburg, Weidner, \& Kroupa (2007). They give a formula to convert the H$\alpha$ luminosity to an SFR (in M$_{\odot}$ yr$^{-1}$) valid for L(H$\alpha)\geq2.8\times10^{36}$ erg s$^{-1}$:
\begin{equation}
SFR=L(H\alpha)/X
\end{equation}
where $X$=1.89$\times10^{41}$ erg s$^{-1}$ for a canonical (broken) IMF, or 3.3$\times10^{41}$ erg s$^{-1}$ for a Salpeter IMF.
The resultant SFRs for the objects with H$\alpha$ line emission, listed in the last column of Table 4 for a canonical IMF, are rather modest with the median SFR=2 M$_{\odot}$ yr$^{-1}$ and the highest value at 4.3 M$_{\odot}$ yr$^{-1}$ (for 009624), or $\sim$40\% lower for a Salpeter IMF. We are, therefore, not witnessing very strong SF events but rather mild ones, implying a possibility to maintain the SF process at its present level, given the detected total HI content, over many Gyr by having many and frequent such SF events. In fact, among all galaxies we find the shortest HI consumption time (for 009624) to be $\sim$4 Gyr.
Given this result, we decided to test first a possibility that the CHMDs have been forming stars at a constant rate since their formation. Our data set, in particular the availability of SDSS and 2MASS colours, allows a rough estimation of the global star formation history (SFH) in the candidate high-mass disk galaxies. This procedure aims to best-fit the measured global colours of a CHMD with colours predicted by accepted population synthesis programs and has been explained in Zitrin \& Brosch (2008) and in Zitrin, Brosch \& Bilenko (2009). The baseline synthesized colours originate from the Bruzual \& Charlot (2003) GALAXEV library. To allow for slight variations in the simplest scenario, we modeled a composite of two constant SFR models where we allowed each SF process to start at a different time and to produce different fractions of the light detected at present from each galaxy. Note that a single stellar population (SP) formed with a constant SFR would result automatically from such a procedure as one of the two SPs, where the second population would not contribute to the present-day light. The criterion by which the proper population mix was chosen was the lowest reduced-$\chi^2$ of the fit. We can reject this constant SFR scenario, based on the resultant $\chi^2$ values, when comparing with the $\chi^2$ values of the other tested scenarios.
Since we could not reproduce the colours with a constant SFR, even when involving two SPs formed at different times with constant SFRs, we tested two other possible SF scenarios; scenario A combining a continuous star formation process with a single $\delta$-function SF burst while allowing any metallicity value for each of the two SPs, and scenario B assuming two ``instantaneous'' $\delta$-function bursts occuring at different times, but fixing the metallicity value to either solar or 2.5 times higher. The parameters that change are the start times of the two processes and their relative proportions in the present-day light collected from the CHMD.
We found that scenario A produced exclusively ``old'' galaxies, where only $\sim$1/3 of the objects showed a relatively young population that was nevertheless $\sim$3 Gyr old and contributed up to 34\% of the present light. Scenario B, on the other hand, allowed the presence of a reasonable fraction of young to very young stars in most objects, less than 10 Myr old, producing $\sim$10-30\% of the observed light and yielded $\sim3\times$ lower reduced $\chi^2$ values than scenario A.
Based on the resultant minimal $\chi^2$ we can, therefore, reject the scenario of a SP formed with a constant SFR combined with an SP formed by a recent SF burst, We thus adopt a scenario where most of the stars in a CHMD were formed a long time ago, perhaps close to a Hubble time ago, during a short period but a small fraction of their SP formed in a recent SF event. This fraction of young stars now produces 10-30\% of the light collected from the galaxy.
It is tempting at this point to speculate what could be the possible very recent SF trigger, a few tens of Myr at most, in a galaxy that seems to have been simply aging with time. One possibility is that the SF process is fed by stellar ejecta and not by externally accreted material. This recycling of material through stars can be done by gentle mass loss via stellar winds and planetary nebulae formation, with $\sim$40-50\% of the stellar mass returned to the ISM and only a small fraction of ISM ($\sim$10\% of the total) coming from supernovae (e.g., Pozzetti et al. 2007, Matig \& Bournaud 2010).
Another option could be a minor merger with a gas-rich companion that rejuvenated the old disk, since our optical survey and the shapes of the HI profiles do not seem to support major mergers. We point out that signs of a past but recent interaction could possibly be detected with 2D kinematic mapping, either optically or by mapping the HI with a synthesis telescope. It is also possible that some of the other 14 galaxies not identified as CHMDs could be transition objects, following the recent accretion of a gas-rich object, indicating that with this sample we are seeing ``real-time'' evolution of massive disks from the red to the blue sequence, or at least into the green valley.
These two options can be checked against the location of the star-forming regions. If the SF trigger is accumulation of stellar ejecta, then the SF should be more concentrated that the general galaxy light. If the trigger is external gas accumulation by IGM accretion (e.g., Salim \& Rich 2010) then the young stars are expected form outside (most of) the disk and not show a particular central condensation.
We tested this by blinking the net-H$\alpha$ images against the R-band images and indicate the results in the last column of Table~\ref{t:massive}. Code C implies that the line emission is centrally concentrated while code D indicates that it is diffuse. The results show that in most cases the H$\alpha$ emission is centrally concentrated, supporting the possibility that the source of material for the formation of the young stellar population is recycling of internal ISM produced by stellar ejecta. We cannot rule out influx of fresh material, at least for the objects with a mixed classification (C/D). However, it is significant that in no case have we observed a galaxy where the HII regions were exclusively relegated to the periphery of the object.
\subsection{Stellar mass estimate}
We estimated above the stellar mass by assuming that all stars were K5V; this is certainly not the case, since the stellar population models showed the best fits to be a composite of a population almost a Hubble time ago with a small fraction formed at most a few tens of Myr ago. The young population contributes 10-30\% of the visible light. We can use this result to refine the estimate of the stellar mass of each galaxy. We shall first explain the method, then apply it to each CHMG.
The mass estimate uses a value for the M/L to convert from measured luminosity to mass; above we used the value for a K5V star. However, the M/L values for different populations e.g., by using the top panel from Figure 5 of Anders et al. (2009), indicate that M/L$_V\simeq2\times10^{-2}$ M$_{\odot}$L$^{-1}_{\odot}$ for a stellar population 10 Myr old, while it is $\sim$5 M$_{\odot}$L$^{-1}_{\odot}$ for a population $\sim10^{10}$ yr old.
If the stellar composition of a galaxy of total luminosity $L$ is produced by only two stellar populations, say population 1 (young) and population 2 (old), which have (M/L)$_1$ and (M/L)$_2$ respectively, and if each population produces a fraction X or (1-X) of the total luminosity recorded for the galaxy, then the mass of first population is:
\begin{equation}
M_1=X \times L \times (M/L)_1
\end{equation}
whereas for the second population
\begin{equation}
M_2=(1-X) \times L \times(M/L)_2
\end{equation}
The mass ratio of the two stellar populations, for the CHMGs studied here, is
\begin{equation}
M_1/M_2=\frac{X}{1-X} \times \frac{(M/L)_1}{(M/L)_2} \simeq\frac{0.2}{0.8} \times \frac{0.02}{5}
\end{equation}
The young population contributes $\sim$0.1\% to the total mass, thus assuming that all the mass is in the old stars is certainly not a bad approximation. Since the total stellar mass in the CHMGs is of order 9$\times10^{10}$ M$_{\odot}$, the mass of the newly formed stellar population should be $\sim10^8$ M$_{\odot}$ and the time to form it, at a typical rate of 2 M$_{\odot}$ yr$^{-1}$ would be some 50 Myr. This rough estimate justifies our conclusion that in these CHMGs we see a recent SF event in an old galaxy.
\subsection{CHMDs and the Tully-Fisher relation}
Catinella et al. (2010) studied a sub-sample of an unbiased collection of galaxies with stellar masses M$_*\geq$10$^{\rm 10}$ M$_{\odot}$ at nearby redshifts (0.025$\leq$z$\leq$0.05). The objects were primarily selected from the GALEX and SDSS data sets; those analyzed in the paper, some 200 galaxies, have HI measurements from Arecibo. The preliminary results of Catinella et al. show a strong decrease in the M$_{HI}$/M$_*$ ratio with stellar mass and with the stellar surface density in M$_{\odot}$ kpc$^{-2}$, $\mu_*$=M$_*$/(2$\pi R^2_{50,z})$, with $R_{50,z}$ being the radius
containing 50\% of the Petrosian flux in the z-band. Specifically, M$_{HI}$/M$_*$ decreases by about one order of magnitude for 10$\leq$log M$_*\leq$11.3 and by two orders of magnitude for 8.0$\leq$log $\mu_* \leq$9.5 (see their Figure 10).
The CHMD behaviour shows a similar trend of the HI-to-stellar mass ratio to decrease with the stellar mass increase, except for the aberrant object 248977 discussed below.
We found that in all cases M$_{dyn}$ was 2.5--7.5$\times$ larger than the sum of M$_*$ and
M$_{\rm HI}$ corrected for the helium content albeit with rather large errors (see Table 4)
This seems to indicate that all galaxies studied here have significant but not excessive amounts of dark matter. The stellar mass to gas mass ratios found here are similar to those of field Sa and Sab galaxies (Read \& Trentham 2005).
Another interesting result concerns the immediate neighbourhood of the high-mass candidate galaxies. One would expect massive galaxies to reside in high galaxy density regions yet, as column 10 of Table 4 shows, these neighbourhoods are relatively sparse. In fact, some of the more massive candidates (e.g., 006066 or 009624) have very few possible neighbours within 3h$^{-1}$ Mpc and 300 km sec$^{-1}$.
The shape of HI profiles themselves is an accepted criterion to indicate possible signs of interaction (e.g., Haynes et al. 1998). Although the ALFALFA profiles shown in Figures~\ref{f:candidates0},~\ref{f:candidates1},~\ref{f:candidates2},
and~\ref{f:candidates3} of Appendix B are not as clean and ``high-quality'' as those shown by Haynes et al., we can state with reasonable confidence that the profiles appear to be symmetric in the areal symmetry criterion. We thus identified a small population of high-mass disk galaxies that are fairly isolated and do not appear to have had
interactions in the last few Gyr
The sub-sample of candidate high-mass objects that do not appear to show obvious signs of interaction offers a possibility to check the mass vs. rotational velocity at the high-v$_{max}$ end.
The mass vs. rotational velocity (a.k.a., the Tully-Fisher=TF) relation, was shown e.g., in Figure 4 of Giovanelli et al. (1986) and is shown here for the CHMDs in Figure~\ref{f:HiMass_plot}.
ACG 248977, a small galaxy extending less than 10 kpc as shown in the SDSS images, yet with
a very wide HI profile (Figure~\ref{f:candidates2}, 3$^{\rm rd}$ row-right), is one of the objects that deviates significantly from the Giovanelli et al. (1986) relation and from the relation suggested here for the high v$_{max}$ galaxies, rotating $\sim3\times$ faster than what its $r$-band luminosity would allow, or by being fainter by one order of magnitude than its v$_{max}$ would predict. It is possible that we are not observing pure rotation in this object and that HI synthesis observations, or a single-dish observation with significantly higher S/N will indicate chaotic motion of HI clouds. Note also that the apparent size of HI14110.4+145339, as inspected by us, seems to be about three times larger overall than the listed SDSS major axis (Appendix A). If this is true, then its M$_{dyn}$ would be larger by the same factor alleviating the discrepancy. On the other hand, we also note that the HI profile is affected by RFI, as the plotted baseline shows. In adopting the listed w(50) we have, therefore, accepted essentially a lower limit to the profile width implying a lower limit to M$_{dyn}$
Eliminating ACG 248977, and fitting a linear regression to the log(M$_{dyn}$) vs. log(v$_{max}$) distribution, yields a correlated set (correlation coeffcient 0.73) with
\begin{equation}
log(M_{dyn})=(8.4\pm0.9)+(1.25\pm0.35)\times log(v_{max})
\end{equation}
We plotted this relation with a dashed line in Figure~\ref{f:HiMass_plot}.
Many of the 14 galaxies discussed here appear to fit well the upper-right part of the
Giovanelli et al. (1986) plot but the three objects with the highest v$_{max}$ values lie below the extrapolation of the Giovanelli et al. relation. This was also the case for UGC 12591 (see Giovanelli et al. 1986). We note in this context that the study of the high-mass end of the TF relation (Noordermeer \& Verheijen 2007) found a similar result,
that galaxies with v$_{max} \geq$200 km sec$^{-1}$ were rotating faster than expected (or were less luminous). However, Meyer et al. (2008) found slopes of $\sim$4 for lower mass galaxies (v$_{max}\leq$300 km sec$^{-1}$) measured by HIPASS, as also found by Trachternach et al. (2009) for 11 very low-mass dwarf galaxies for which they have HI synthesis maps. Our result, a slope of approximately unity, together with the results of Giovanelli et al. and of Noordermeer \& Verheijen, indicate a fundamental difference between the high-mass and lower mass galaxies in the definition of the TF relation at the high-mass end of the disks.
\begin{table*}
\caption{Candidate high-mass galaxies}
\label{t:massive}
\begin{tabular}{ccccclcccccc} \hline
ACG & L$_r$ & $a$ & $i$ & v$_{max}$ & M$_{dyn}$ & M$_*$ & M$_{HI}$ & F & N & SFR & HII\\
& L$_{\odot}$& kpc & $^{\circ}$ & km s$^{-1}$ & M$_{\odot}$ & M$_{\odot}$ & M$_{\odot}$ & & 3h$^{-1}$Mpc & M$_{\odot}$yr$^{-1}$ & \\ \hline
205190 & 3.8(10) & 25 & 46 & 485 & 7.2(11) & 8.1(10) & 1.2(10) & 7.5 & 24 & 1.1 & C \\
& & & 14 & 115 & 2.4 & & 0.06 & 2.5 & & 0.2 \\ \hline
200589 & 4.5(10) & 26 & 61 & 316 & 2.9(11) & 9.6(10) & 1.0(10) & 2.7 & 14 & 2.0 & C \\
& & & 10 & 29 & 0.4 & & 0.06 & 0.4 & & 0.4 \\ \hline
006066 & 4.3(10) & 41 & 77 & 342 & 5.4(11) & 9.1(10) & 4.4(10) & 3.6 & 3 & 1.8 & C \\
& & & 07 & 11 & 0.4 & & 0.09 & 0.3 & & 0.6 \\ \hline
222042 & 4.1(10) & 28 & 66 & 304 & 2.9(11) & 8.7(10) & 1.6(10) & 2.7 & 18 & 2.1 & C \\
& & & 09 & 22 & 0.3 & & 0.10 & 0.3 & & 0.3 \\ \hline
008375 & 3.1(10) & 25 & 73 & 364 & 3.7(11) & 6.6(10) & 5.9(09) & 7.3 & 6 & 0.7 & C \\
& & & 07 & 37 & 0.6 & & 0.35 & 1.3 & & 0.3 \\ \hline
008379 & 4.4(10) & 33 & 74 & 296 & 3.5(11) & 9.4(10) & 2.1(10) & 2.9 & 7 & 2.7 & C \\
& & & 07 & 46 & 0.7 & & 0.08 & 0.6 & & 0.8 \\ \hline
008475 & 5.1(10) & 28 & 57 & 352 & 3.9(11) & 1.1(11) & 3.7(10) & 2.5 & 60 & 4.0 & D \\
& & & 12 & 46 & 0.7 & & 0.04 & 0.4 & & 0.8 \\ \hline
008488 & 3.7(10) & 28 & 69 & 319 & 3.2(11) & 7.9(10) & 1.1(10) & 4.0 & 48 & 1.4 & C \\
& & & 09 & 25 & 0.4 & & 0.03 & 0.6 & & 1.6 \\ \hline
230914 & 5.2(10) & 33 & 67 & 315 & 3.7(11) & 1.1(11) & 2.7(10) & 2.5 & 14 & 3.6 & C \\
& & & 09 & 11 & 0.4 & & 0.08 & 0.3 & & 0.8 \\ \hline
009031 & 4.4(10) & 27 & 66 & 388 & 4.6(11) & 9.4(10) & 1.9(10) & 3.9 & 8 & 3.6 & C/D \\
& & & 09 & 30 & 0.6 & & 0.11 & 0.6 & & 0.9 \\ \hline
248977 & 1.0(10) & 6 & 48 & 377 & 9.6(10) & 2.1(10) & 9.0(09) & 2.9 & 5 & 0.2 & C \\
& & & 14 & 85 & 3.1 & & 0.54 & 0.9 & & 0.05 \\ \hline
009624 & 7.0(10) & 21 & 34 & 536 & 6.8(11) & 1.5(11) & 1.7(10) & 4.0 & 5 & 4.3 & C \\
& & & 20 & 270 & 4.9 & & 0.07 & 2.9 & & 0.7 \\ \hline
260110 & 2.9(10) & 14 & 34 & 545 & 4.7(11) & 6.2(11) & 2.9(10) & 4.7 & 19 & 1.9 & C/D \\
& & & 20 & 273 & 3.3 & & 0.09 & 3.3 & & 0.3 \\ \hline
010272 & 2.5(10) & 16 & 63 & 325 & 1.9(11) & 5.3(10) & 2.2(09) & 3.4 & 14 & 1.5 & C \\
& & & 10 & 43 & 0.4 & & 0.13 & 0.7 & & 0.2 \\
\hline
\end{tabular}
\\ Notes: \\
(a) Powers-of-ten are given in parentheses. \\
(b) The last column, HII, gives the classification of the HII region distribution: C=concentrated and D=diffuse. \\
(c) The second line for each entry gives the error in the parameter listed immediately above it, in the same units and with the same power-of-ten as in the first line.
\end{table*}
\begin{figure}
\label{f:log(vmax)_vs_Log(DynMss)}
\includegraphics[clip=,angle=-90,width=14cm]{Fig2_newest.eps}
\caption{Logarithmic plot of the maximal rotational velocity vs. the total dynamical mass for the CHMDs. The general trend is for a relatively flat relation, i.e. the maximal rotational velocity being $\sim$independent of galaxy dynamical mass, with one object (ACG 248977) deviating considerably. The error bars are not plotted, for clarity, but the errors are listed in Table 4. The relation derived in equation 5 is plotted as a dashed line.}
\label{f:HiMass_plot}
\end{figure}
We stress that in Figure 2 we have not plotted the error bars; these are relatively large and their inclusion (in a log-log plot) would result in a figure that would not be understandable. However, as mentioned above, the formal errors are listed in Table 4.
\section{Summary}
\label{txt:summ}
\begin{enumerate}
\item We described a complete sample of 28 galaxies with wide HI profiles selected from the ALFALFA HI survey that could be candidate high-mass objects.
\item Using on-line data bases, we showed that some of the objects could have been confused by nearby neighbour galaxies that could have been included in the radio telescope beam producing the wide HI profiles.
\item We eliminated those, as well as objects showing obvious signs of interaction in SDSS images, and reduced the sample to 14 objects lacking visible signs of interaction and/or very close neighbours. We propose these as candidate high-mass disk galaxies (CHMDs).
\item We checked the degree of isolation of the CHMDs and found that some may indeed be located in low galaxy density neighbourhoods, contrary to expectations that high-mass galaxies would be found in high galaxy density regions.
\item We investigated simple models for the star formation histories of the CHMDs and found that, in most cases, these objects show signs of young and massive star formation. Tests with different star formation scenarios indicated that most of the stars formed about a Hubble time ago, with 10-30\% of the light produced by very young stars formed at most a few 10s of Myrs ago.
\item We found that the mass of the young stellar component is a small fraction of the total stellar mass; the CHMGs are thus showing signs of rejuvenation of their old stellar populations. With the current star formation rate, the young stellar population can be produced in a few tens of Myr; the HI reservoir detected by ALFALFA suffices to maintain the star formation for about a Hubble time.
\item We calculated dynamical masses from the HI profile width, the optical inclination, and the optical radius, and compared those with the HI and stellar masses, finding that the luminous matter makes up only $\sim$30\% of the dynamical mass.
\item We investigated the location of these CHMDs in the Tully-Fisher relation and found that they deviate significantly from the relation defined by lower-mass galaxies.
\end{enumerate}
\section*{Acknowledgments}
The authors would like to acknowledge the work of the entire ALFALFA collaboration team in observing, flagging, and extracting the catalog of galaxies used in this work. We are grateful for the selection of galaxies from the ALFALFA data set provided by Martha Haynes and Riccardo Giovanelli. We acknowledge some constructive remarks from an anonymous referee.
Funding for the Sloan Digital Sky Survey (SDSS) has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Aeronautics and Space Administration, the National Science Foundation, the U.S. Department of Energy, the Japanese Monbukagakusho, and the Max Planck Society. 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 University of Chicago, Fermilab, the Institute for Advanced Study, the Japan Participation Group, the Johns Hopkins University, Los Alamos National Laboratory, the Max-Planck-Institute for Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, University of Pittsburgh, Princeton University, the United States Naval Observatory, and the University of Washington.
This publication makes use of data products from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation.
This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
|
\section{Introduction}
$~~~$ The state-of-the-art understanding on the subject of higher spin supersymmetric
multiplets was established in a work by Kuzenko, Postnikov, and Sibiryakov \cite{Kuzenko:1993jp}.
In fact, they established two such formulations for each and every possible value of the superspin
$Y$.
These formulations are based on the introduction of constrained compensating superfields. The
goal of this work is to re-examine these schemes in order to be able to reproduce their results and,
if possible, to discover new formulations in the case of half odd superspins. This is exactly what
will happen in the following. Their results will emerge naturally from our algorithm as a possible
way a theory of higher, half odd massless superspins can be formulated.
In an accompany paper \cite{Gates:2011qb} devoted to the study of massless 4D, $\cal N$ $=$ 1 higher integer
superspins, we developed an algorithm that was able to do two things:
\begin{itemize}
\item generate all known results for massless 4D, $\cal N$ $=$ 1 higher integer superspins
up to that point, and
\item introduce a new formulation of the theory.
\end{itemize}
After the success of this algorithm in the investigation of higher, integer superspins we would
like to apply a similar way of thinking in the case of half odd superspins.
The conceptual backbone of the method followed, can be summarized as following:\\
$~~~~~$ Step 1) find the main physical superfield\footnote{This fixes the
index structure, mass dimensions and says something about reality of this\\
$~~~~~~$ main physical superfield. Within the context of
supergravity, this main physical superfield \\
$~~~~~~$
is known as the `superconformal submultiplet.'},
that will be used to \newline
$~~~~~~~~~\,~~~~~~$ construct the theory,\\
$~~~~~$ Step 2) find the most general free action which is quadratic to this \newline
$~~~~~~~~~\,~~~~~~$
superfield,\\
$~~~~~$ Step 3) find the gauge transformation of the main superfield,\\
$~~~~~$ Step 4) find the type of superfield(s) we have to introduce as \newline
$~~~~~~~~~\,~~~~~~$
compensators,\\
$~~~~~$ Step 5) find the possible gauge transformation{s} of the com-
\newline
$~~~~~~~~~\,~~~~~~$
pensators which on-shell give just the degrees of \newline
$~~~~~~~~~\,~~~~~~$
freedom needed, and \\
$~~~~~$ Step 6) check invariances of the action with respect to all trans-
\newline
$~~~~~~~~~\,~~~~~~$ formations.
\section{General Action and Gauge Transformations}
$~~~$ The goal is to develop a theory for massless half odd superspin $Y$ $=$ $s+\frac{1}
{2}$, for integers $s$. This means the highest superspin projection operator acting on the `main
superfield' used to develop the theory must generate an object with an odd number of indices
($2s+1$). As suggested by supergravity theory, the fundamental superfield for this theory should
be a bosonic superfield with an even number of indices, $s$ undotted and $s$ dotted ($H_{{\alpha}(s)
{\dot{\alpha}}(s)}$). Furthermore, its highest spin component (which is the completely symmetric piece of
the $\theta\, \bar{\theta}$ term, $h_{{\alpha}(s+1){\dot{\alpha}}(s+1)}$) must propagate on-shell. But for that to
happen, it must have mass dimensions one ($[h]=1$) and according to the Fronsdal action of
massless integer spins (which must be the bosonic piece of our theory) it also needs to be real.
Therefore our theory must be constructed in terms of a real bosonic superfield $H_{{\alpha}(s){\dot{\alpha}}(s)}$
with zero mass dimensions ($[H]=0$). The most general action that can be written for such an
object has the form:
\begin{equation}
\eqalign{
S=\int d^8z\Big\{
&~c_1 H^{{\alpha}(s){\dot{\alpha}}(s)}{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}(s){\dot{\alpha}}(s)}~+~
c_2 H^{{\alpha}(s){\dot{\alpha}}(s)}\Box H_{{\alpha}(s){\dot{\alpha}}(s)}\cr
&+c_3 H^{{\alpha}(s){\dot{\alpha}}(s)}\pa_{{\alpha}_s{\dot{\alpha}}_s}\pa^{{\gamma}{\dot{\gamma}}}H_{{\gamma}{\alpha}(s-1){\dot{\gamma}}{\dot{\alpha}}(s-1)}\cr
&+c_4 H^{{\alpha}(s){\dot{\alpha}}(s)}[{\rm D}_{{\alpha}_s},{\bar{\rm D}}_{{\dot{\alpha}}_s}][{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}]H_{{\gamma}{\alpha}(s-1)
{\dot{\gamma}}{\dot{\alpha}}(s-1)}\Big\} ~~~. {~~~~~~~~~~~}
}
\label{equ01}
\end{equation}
In writing this action, we have also made an assumption that parity violating terms should
be excluded. If this assumption is not used then an additional term of the form
\begin{equation}
\eqalign{
S_{P-violation}=\int d^8z\Big\{
H^{{\alpha}(s){\dot{\alpha}}(s)}\pa_{{\alpha}_s{\dot{\alpha}}_s} [{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}]H_{{\gamma}{\alpha}(s-1)
{\dot{\gamma}}{\dot{\alpha}}(s-1)}\Big\} ~~~.
}
\label{equ02}
\end{equation}
may be considered\footnote{In principle we could repeat the whole analysis including this term and
show that it's coefficient \newline $~\,~~\,~~$ will vanish. But just knowing that the final results are
Frondal's actions for bosons and fermions
\newline $~\,~\,~~~$and they preserves parity, allows us to set this term to zero from the very beginning.}.
The massless property of the theory suggest there must be an underlying gauge
symmetry. This symmetry, of course, must respect the highest superspin projection
operator. Taking this into account there is only one option.
The gauge transformation of $H_{{\alpha}(s){\dot{\alpha}}(s)}$ must be of the form:
\begin{equation}
\delta H_{{\alpha}(s){\dot{\alpha}}(s)}=\frac{1}{s!}{\bar{\rm D}}_{({\dot{\alpha}}_s}L_{{\alpha}(s){\dot{\alpha}}(s-1))}-\frac{1}{s!}{\rm D}_{({\alpha}_s}
\bar{L}_{{\alpha}(s-1)){\dot{\alpha}}(s)} ~~~,
\label{equ03}
\end{equation}
written in terms of some complex gauge parameter superfield $L_{{\alpha}(s){\dot{\alpha}}(s-1))}$.
The change of the above action under this transformation is :
$$
\eqalign{
\delta S=\int d^8z \Bigg\{
&~\[-2c_1+2c_2+\frac{2}{s}c_3+2\frac{2s+1}{s}c_4\]H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}_{{\dot{\alpha}}_s}
{\rm D}^2{\bar{\rm D}}^2L_{{\alpha}(s){\dot{\alpha}}(s-1)} \cr
}
$$
\begin{equation}
\eqalign{ {~~\,~~~~} {~~~~~~~} {~~~~~~~}
&+2c_2 H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}^2{\rm D}^2{\bar{\rm D}}_{{\dot{\alpha}}_s}L_{{\alpha}(s){\dot{\alpha}}(s-1)}\cr
&+\[-\frac{2}{s}c_3+2\frac{2s+1}{s}c_4\]H^{{\alpha}(s){\dot{\alpha}}(s)}{\rm D}_{{\alpha}_s}{\bar{\rm D}}^2{\rm D}^{{\gamma}}
{\bar{\rm D}}_{{\dot{\alpha}}_s}L_{{\gamma}{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
&+\[2c_3-2c_4\]H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}_{{\alpha}_s}{\bar{\rm D}}^2{\rm D}^{{\gamma}}L_{{\gamma}{\alpha}(s-1)
{\dot{\alpha}}(s-1)}\cr
&-\[\frac{s-1}{s}\]\[2c_3-2c_4\]H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}_{{\alpha}_s}{\bar{\rm D}}_{{\dot{\alpha}}_{
s-1}}{\rm D}^{{\gamma}}{\bar{\rm D}}^{{\dot{\gamma}}}L_{{\gamma}{\alpha}(s-1){\dot{\gamma}}{\dot{\alpha}}(s-2)}\cr
&+c.c.\Bigg\} ~~~.
} \label{equ04}
\end{equation}
It is obvious that the above action is not invariant under the proposed gauge transformation. There
are two ways around this. One way is to impose differential constraints
(using either D or $\Bar {\rm D}$) on the gauge parameter superfield $L_{{\alpha}(s){\dot{\alpha}}(s-1))}$.
In general this procedure leads to the `ghost-for-ghost' phenomenon in a quantum
theory \cite{Siegel:1980}. We wish to avoid this.
The other way, is to introduce a set of compensators.
In order to keep the propagationing degrees of freedom down to the minimal number and
in order to have on-shell an irreducible representation of the Super-Poincare group, we
need exactly one `propagating' compensator\footnote{This is a superfield of mass
dimensions 0 or 1/2.} and some arbitrary number of auxiliary compensators\footnote{
These are superfields of mass dimensions 1}.
~The propagating compensator must satisfy several constraints. It must provide the
extra degrees of freedom in order to complete the irreducible representation and the
rest of its components must vanish on shell. In principle there are two options, it can
be either bosonic or fermionic. In the first case, the gauge transformation of a bosonic
compensator has to be\footnote{This is fixed just by considering the index structure and mass
dimensions.} of the form ${\rm D}^{{\alpha}_s}L_{{\alpha}(s){\dot{\alpha}}(s-1)}+{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{L}_{{\alpha}(s-1){\dot{\alpha}}(s)}$.
But with a transformation like that we can not gauge away all the degrees of freedom,
besides the ones needed for the irreducible representation. So this option can not
lead to the desired result. Therefore the propagating compensator must be a fermionic
superfield
$\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}$. This is in accord with the feature of all previous studied
theories, where the statistics of the main superfield and the compensator opposite to
one another.
Also the gauge transformation of the fermionic compensator $\Upsilon$ must be such that,
it satisfies the following:
\begin{itemize}
\item the component $\Upsilon^{(1,0)S}_{{\alpha}(s+1){\dot{\alpha}}(s-1)}$ must be gauged away. This
can be done if
\begin{equation}
{\rm D}_{({\alpha}_{s+1}}\Upsilon_{{\alpha}(s)){\dot{\alpha}}(s-1)}|\sim \text{~some component of the gauge
parameter (algebraicly)}\nonumber
\end{equation}
\item the component $\Upsilon^{(0,1)S}_{{\alpha}(s){\dot{\alpha}}(s)}$ must be gauged away. This can
be done if
\begin{equation}
{\bar{\rm D}}_{({\dot{\alpha}}_s}\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1))}|\sim \text{~some component of the gauge parameter
(algebraicly)}\nonumber
\end{equation}
\item the component $\Upsilon^{(0,1)A}_{{\alpha}(s){\dot{\alpha}}(s-2)}$ must be gauged away. This can
be done if
\begin{equation}
{\bar{\rm D}}^{{\dot{\gamma}}}\Upsilon_{{\alpha}(s){\dot{\gamma}}{\dot{\alpha}}(s-2)}|\sim \text{~some component of the gauge parameter
(algebraicly)}\nonumber
\end{equation}
\item the component $\Upsilon^{(1,1)(A,A)}_{{\alpha}(s-1){\dot{\alpha}}(s-2)}$ must propagate on shell
and have a specific gauge transformation. This can be done if
\begin{equation}
[{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}]\Upsilon_{{\gamma}{\alpha}(s-1){\dot{\gamma}}{\dot{\alpha}}(s-2)}|\sim \pa^{{\gamma}{\dot{\gamma}}}\text{~some component
of the gauge parameter} \nonumber
\end{equation}
\end{itemize}
The above constraints and equation (\ref{equ04}) will be our guideline. Based on equation
(\ref{equ04}), we must find all possible ways that we can introduce a fermionic compensator
with mass dimensions 1/2 and with the specific index structure $\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}$, which
has a gauge transformation that satisfies all the above constraints. This can be done only
by two ways.
~Since $L_{{\alpha}(s){\dot{\alpha}}(s-1)}$ and $\bar{L}_{{\alpha}(s-1){\dot{\alpha}}(s)}$ are the only gauge parameters
available, the gauge transformation of $\Upsilon$ must include at least one of them. Their mass
dimensions are -1/2 ($[L]=[\bar{L}]=-1/2$), so we need 2 ${\rm D}({\bar{\rm D}})$'s to build something
with mass dimensions 1/2. Therefore, the transformation of $\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}$ must include at
least one of the following terms.
\begin{equation}
\eqalign{
&A) {~~~} {\bar{\rm D}}^2L_{{\alpha}(s){\dot{\alpha}}(s-1)} ~~~, \cr\nonumber
&B) {~~~} \frac{1}{s!}{\bar{\rm D}}^{{\dot{\alpha}}_s}{\rm D}_{({\alpha}_s}\bar{L}_{{\alpha}(s-1)){\dot{\alpha}}(s)} ~~~, \cr
&C) {~~~}{\rm D}^2L_{{\alpha}(s){\dot{\alpha}}(s-1)} ~~~, \cr
&D) {~~~} \frac{1}{s!}{\rm D}_{({\alpha}_s}{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{L}_{{\alpha}(s-1){\dot{\alpha}}(s)} ~~~.
}
\end{equation}
In order, these types of possible transformations to have a hope to give
something desirable, they need to be completed appropriately so the above constraints
are satisfied. The minimal way to do that is the following:
\begin{equation}
\eqalign{
&A) {~~~} {\bar{\rm D}}^2L_{{\alpha}(s){\dot{\alpha}}(s-1)}+{\rm D}^{{\alpha}_{s+1}}\Lambda_{{\alpha}(s+1){\dot{\alpha}}(s)} ~~~, \cr\nonumber
&B) {~~~} \frac{1}{s!}{\bar{\rm D}}^{{\dot{\alpha}}_s}{\rm D}_{({\alpha}_s}\bar{L}_{{\alpha}(s-1)){\dot{\alpha}}(s)}+{\rm D}^{{\alpha}_{s+1}}
\Lambda_{{\alpha}(s+1){\dot{\alpha}}(s)} ~~~, \cr
&C) {~~~}{\rm D}^2L_{{\alpha}(s){\dot{\alpha}}(s-1)}+{\bar{\rm D}}^{{\dot{\alpha}}_s}V_{{\alpha}(s){\dot{\alpha}}(s)} ~~~.
}
\end{equation}
the last case was eliminated through this last requirement.
~Now it is really straightforward to check if any of the above transformations can be
used together with (\ref{equ04}) in order to introduce the fermionic compensator. Just
by observing (\ref{equ04}) we see that case C can not happen and we are left with two
possibilities. Case (A) can arise from the first term of (\ref{equ04}) and Case (B) can arise
from the third term. Next we will study this two cases.
\section{The Higher Superspin KPS-Series}
For Case (A) consider, we impose
\begin{equation}
c_2=c_3=c_4=0 ~~~,
\label{equ05}
\end{equation}
so then equation (\ref{equ04} )becomes:
\begin{equation}
\eqalign{
\delta S&=\int d^8z\Big\{-2c_1H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}^2{\bar{\rm D}}^2L_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\Big\}
\cr
&=\int d^8z\Big\{-2c_1H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}^2\[{\bar{\rm D}}^2L_{{\alpha}(s){\dot{\alpha}}(s-1)}+{\rm D}^{{\alpha}_{s+1}}
\Lambda_{{\alpha}(s+1){\dot{\alpha}}(s-1)}\]+c.c.\Big\} ~~~.
}
\label{equ06}
\end{equation}
At this point we can introduce a fermionic compensator $\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}$ with the following
gauge transformation:
\begin{equation}
\delta\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}={\bar{\rm D}}^2L_{{\alpha}(s){\dot{\alpha}}(s-1)}+{\rm D}^{{\alpha}_{s+1}}\Lambda_{{\alpha}(s+1){\dot{\alpha}}(s)}
~~~,
\label{equ07}
\end{equation}
and in order to construct a fully invariant action and give to the compensator some dynamics
we have to add to the initial action some more terms
\begin{itemize}
\item Add a counter term, which cancels the chang of the initial action:
\begin{equation}
S_c=\int d^8z~2c_1H^{{\alpha}(s){\dot{\alpha}}(s)}\left({\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}-{\rm D}_{{\alpha}_s}{\bar{\rm D}}^2
\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\right)
\label{equ08}
\end{equation}
\item Add a kinetic energy term for the compensator (the most general free action quadratic to
$\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}$)
\begin{equation}
\eqalign{
S_{k.e}=\int d^8z\Big\{
&~h_1\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&+h_2\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\bar{\rm D}}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&+h_3\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\bar{\rm D}}^{{\dot{\alpha}}_s}{\rm D}_{{\alpha}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\cr
&+h_4\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}_{a_s}{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\Big\}
}
\label{equ09}
\end{equation}
\end{itemize}
The full action is thus given by
\begin{equation}
\eqalign{ {~~~~~~~~~}
S_{Full}=\int d^8z\Big\{
&~c_1 H^{{\alpha}(s){\dot{\alpha}}(s)}{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}(s){\dot{\alpha}}(s)}\cr
&+2c_1H^{{\alpha}(s){\dot{\alpha}}(s)}\left({\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}-{\rm D}_{{\alpha}_s}{\bar{\rm D}}^2\bar{
\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\right)\cr
&+h_1\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&+h_2\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\bar{\rm D}}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&+h_3\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\bar{\rm D}}^{{\dot{\alpha}}_s}{\rm D}_{{\alpha}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\cr
&+h_4\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}_{a_s}{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\Big\}
~~~. }
\label{equ10}
\end{equation}
Now we can define the superfields ${\bm {{\cal G}}}_{{\alpha}(s){\dot{\alpha}}(s)}$ and ${\bm{{\cal T}}}_{{\alpha}(s)
{\dot{\alpha}}(s-1)}$ as the variations of the full action with respect to the superfields $H_{{\alpha}(s)
{\dot{\alpha}}(s)}$ and $\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}$. These variations yield respectively,
\begin{equation}
\eqalign{
{\bm{{\cal G}}}_{{\alpha}(s){\dot{\alpha}}(s)}=&~2c_1{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}(s){\dot{\alpha}}(s)}
\,+\, \frac{2c_1}{s!}\left({\bar{\rm D}}_{({\dot{\alpha}}_s}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1))}-{\rm D}_{({\alpha}_s}{\bar{\rm D}}^2\bar{
\Upsilon}_{{\alpha}(s-1)){\dot{\alpha}}(s)}\right) ~~~,
}
\label{equ11}
\end{equation}
and
\begin{equation}
\eqalign{ {~~~}
{\bm{{\cal T}}}_{{\alpha}(s){\dot{\alpha}}(s-1)}=&~2c_1{\rm D}^2{\bar{\rm D}}^{{\dot{\alpha}}_s}H_{{\alpha}(s){\dot{\alpha}}(s)}
\,+\, 2h_1{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)} \,+\, 2h_2{\bar{\rm D}}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}\cr
&+\, \frac{h_3}{s!}{\bar{\rm D}}^{{\dot{\alpha}}_s}{\rm D}_{({\alpha}_s}\bar{\Upsilon}_{{\alpha}(s-1)){\dot{\alpha}}(s)}
\,+\, \frac{h_4}{s!} {\rm D}_{(a_s}{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{{\alpha}(s-1)){\dot{\alpha}}(s)} ~~~.
}
\label{equ12}
\end{equation}
The invariance of the full action under the above gauge transformations, forces a set of
constraints that must be satisfied. These are the Bianchi Identities which are going
to determine all the free parameters.
\begin{equation}
\eqalign{
&{\bar{\rm D}}^{{\dot{\alpha}}_s}{\bm{{\cal G}}}_{{\alpha}(s){\dot{\alpha}}(s)}+{\bar{\rm D}}^2{\bm{{\cal T}}}_{{\alpha}(s){\dot{\alpha}}(s-1)}=0 ~~~, \cr
& {\rm D}_{({\alpha}_{s+1}}{\bm{{\cal T}}}_{{\alpha}(s)){\dot{\alpha}}(s-1)}=0 ~~~.
}
\label{equ13}
\end{equation}
The solution of the first one is:
\begin{equation}
\eqalign{
&h_1=-\frac{s+1}{s}c_1 ~~~, ~~~
h_4=2c_1 ~~~,
}
\label{equ14}
\end{equation}
and the solution of the second one is:
\begin{equation}
\eqalign{
&h_2=0 ~~~,~~~
h_3=0 ~~~.
}
\label{equ15}
\end{equation}
Therefore the final action is:
\begin{equation}
\eqalign{
S=\int d^8z\Big\{
&~c_1H^{{\alpha}(s){\dot{\alpha}}(s)}{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\dot{\gamma}}}H_{{\alpha}(s){\dot{\alpha}}(s)}\cr
&+2c_1H^{{\alpha}(s){\dot{\alpha}}(s)}\left({\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}
-{\rm D}_{{\alpha}_s}{\bar{\rm D}}^2\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\right)\cr
&-\[\frac{s+1}{s}\]c_1\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&+2c_1\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}_{a_s}{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\Big\} ~~~,
}
\label{equ16}
\end{equation}
and it is invariant under the gauge transformations
\begin{equation}
\eqalign{
&\delta H_{{\alpha}(s){\dot{\alpha}}(s)}=\frac{1}{s!}{\bar{\rm D}}_{({\dot{\alpha}}_s}L_{{\alpha}(s){\dot{\alpha}}(s-1))}-\frac{1}{s!}{\rm D}_{({\alpha}_s}
\bar{L}_{{\alpha}(s-1)){\dot{\alpha}}(s)} ~~~~,\cr
&\delta\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}={\bar{\rm D}}^2L_{{\alpha}(s){\dot{\alpha}}(s-1)}+{\rm D}^{{\alpha}_{s+1}}\Lambda_{{\alpha}(s+1){\dot{\alpha}}(s)}
~~~.
}
\label{equ17}
\end{equation}
This theory is equivalent to that developed by Kuzenko, Postnikov, and Sibiryakov (KPS) \cite{
Kuzenko:1993jp}, once one solves the constraints that appear in their description (as done in \cite{Gates:2010td}).
Therefore without any further examination we can conclude that this action, with that set of transformations
describes a massless half odd superspin ($Y=s+1/2$).
\section{The Higher Superspin B-Series}
For Case (B) we impose
\begin{equation}
\eqalign{
c_2&=0 ~~~, \cr
-2c_1+\frac{2}{s}c_3+2\frac{2s+1}{s}c_4&=0~~\Rightarrow c_1=\frac{1}{s}\left[c_3+(2s+1)c_4\right]
~~~.
}
\label{equ18}
\end{equation}
so that equation (\ref{equ04}) becomes:
$$
\eqalign{
\delta S=\int d^8z \, {\Big \{ }
&\[\frac{2}{s}\](c_3-(2s+1)c_4)H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}^2\Big[\frac{1}{s!}{\bar{\rm D}}^{{\dot{\gamma}}}{\rm D}_{
({\alpha}_s}\bar{L}_{{\alpha}(s-1)){\dot{\gamma}}{\dot{\alpha}}(s-1)}\Big]\cr
+&\[\frac{2}{s}\](c_3-(2s+1)c_4)H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}^2\Big[{\rm D}^{{\alpha}_{s+1}}\Lambda_{
{\alpha}(s+1){\dot{\alpha}}(s-1)}\Big]\cr
+& c.c.\cr
-&2(c_3-c_4)H^{{\alpha}(s){\dot{\alpha}}(s)}[{\rm D}_{{\alpha}_s},{\bar{\rm D}}_{{\dot{\alpha}}_s}]\Big[{\bar{\rm D}}^2{\rm D}^{{\gamma}}L_{{\gamma}{\alpha}(s-1){\dot{\alpha}}(s-1)}
\Big]\cr
}
$$
\begin{equation}
\eqalign{
{~~~~~~~~~~~~~~~~~~~~} -&2(c_3-c_4)H^{{\alpha}(s){\dot{\alpha}}(s)}[{\rm D}_{{\alpha}_s},{\bar{\rm D}}_{{\dot{\alpha}}_s}]\Big[{\rm D}^2{\bar{\rm D}}^{{\dot{\gamma}}}\bar{L}_{{\alpha}(s-1)
{\dot{\gamma}}{\dot{\alpha}}(s-1)}\Big]\cr
-&2(c_3-c_4)H^{{\alpha}(s){\dot{\alpha}}(s)}[{\rm D}_{{\alpha}_s},{\bar{\rm D}}_{{\dot{\alpha}}_s}]\Big[-\frac{s-1}{s!}{\bar{\rm D}}_{({\dot{\alpha}}_{s-1}}
{\rm D}^{{\gamma}}{\bar{\rm D}}^{{\dot{\gamma}}}L_{{\gamma}{\alpha}(s-1){\dot{\gamma}}{\dot{\alpha}}(s-2))}\Big]\cr
-&2(c_3-c_4)H^{{\alpha}(s){\dot{\alpha}}(s)}[{\rm D}_{{\alpha}_s},{\bar{\rm D}}_{{\dot{\alpha}}_s}]\Big[-\frac{s-1}{s!}{\rm D}_{({\alpha}_{s-1}}
{\bar{\rm D}}^{{\dot{\gamma}}}{\rm D}^{{\gamma}}\bar{L}_{{\gamma}{\alpha}(s-2)){\dot{\gamma}}{\dot{\alpha}}(s-1)}\Big] {\Big \} }~~~.
}
\label{equ19}
\end{equation}
We introduce two compensators: \\
1) A fermionic propagating compensator $\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}$ with mass dimensions 1/2 \newline
$~~~~$
and the following gauge transformation
\begin{equation}
\delta\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}=\frac{1}{s!}{\bar{\rm D}}^{{\dot{\gamma}}}{\rm D}_{({\alpha}_s}\bar{L}_{{\alpha}(s-1)){\dot{\gamma}}{\dot{\alpha}}(s-1)}
+{\rm D}^{{\alpha}_{s+1}}\Lambda_{{\alpha}(s+1){\dot{\alpha}}(s-1)} ~~~.
\label{equ20}
\end{equation}
2) A real auxiliary bosonic compensator $B_{{\alpha}(s-1){\dot{\alpha}}(s-1)}$, with mass dimensions 1
\newline $~~~~$
which
transforms as
\begin{equation}
\eqalign{
\delta B_{{\alpha}(s-1){\dot{\alpha}}(s-1)}=&~{\bar{\rm D}}^2{\rm D}^{{\alpha}_s}L_{{\alpha}(s){\dot{\alpha}}(s-1)}+{\rm D}^2{\bar{\rm D}}^{{\dot{\alpha}}_s}
\bar{L}_{{\alpha}(s-1){\dot{\alpha}}(s)}\cr
&-\[\frac{s-1}{s!}\]{\bar{\rm D}}_{({\dot{\alpha}}_{s-1}}{\rm D}^{{\gamma}}{\bar{\rm D}}^{{\dot{\gamma}}}L_{{\gamma}{\alpha}(s-1){\dot{\gamma}}{\dot{\alpha}}(s-2))}\cr
&-\[\frac{s-1}{s!}\]{\rm D}_{({\alpha}_{s-1}}{\bar{\rm D}}^{{\dot{\gamma}}}{\rm D}^{{\gamma}}\bar{L}_{{\gamma}{\alpha}(s-2)){\dot{\gamma}}{\dot{\alpha}}(s-1)}
~~~.}
\label{equ21}
\end{equation}
To create an invariant action and give dynamics to the compensators we have to add the
following terms:
\begin{itemize}
\item A counter term which will cancel the change of the initial action
\begin{equation}
\eqalign{
S_c=\int d^8z\Big\{&-\[\frac{2}{s}\](c_3-(2s+1)c_4)H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}^2\Upsilon_{
{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&+2(c_3-c_4)H^{{\alpha}(s){\dot{\alpha}}(s)}[{\rm D}_{{\alpha}_s},{\bar{\rm D}}_{{\dot{\alpha}}_s}]B_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\Big\}
~~~, }
\label{equ22}
\end{equation}
\item A kinetic energy term for both the compensators (the most general action for $\Upsilon$ and $B$)
\begin{equation}
\eqalign{
S_{k.e}=\int d^8z\Big\{
&~eB^{{\alpha}(s-1){\dot{\alpha}}(s-1)}B_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
&+h_1\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&+h_2\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\bar{\rm D}}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&+h_3\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\bar{\rm D}}^{{\dot{\alpha}}_s}{\rm D}_{{\alpha}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\cr
&+h_4\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}_{a_s}{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\Big\}
~~~,}
\label{equ23}
\end{equation}
\item An interaction term among compensators (as in principle, such a term can exist)
\begin{equation}
S_{int}=\int d^8z \Big\{bB^{{\alpha}(s-1){\dot{\alpha}}(s-1)}\[D^{{\alpha}_s}\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{
{\alpha}(s-1){\dot{\alpha}}(s)}\]\Big\} ~~~.
\label{equ24}
\end{equation}
\end{itemize}
Therefore the full action is
\begin{equation}
\eqalign{
S=\int d^8z\Big\{
&~\[\frac{1}{s}\](c_3+(2s+1)c_4) H^{{\alpha}(s){\dot{\alpha}}(s)}{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}(s){\dot{\alpha}}(s)}\cr
&+c_3 H^{{\alpha}(s){\dot{\alpha}}(s)}\pa_{{\alpha}_s{\dot{\alpha}}_s}\pa^{{\gamma}{\dot{\gamma}}}H_{{\gamma}{\alpha}(s-1){\dot{\gamma}}{\dot{\alpha}}(s-1)}\cr
&+c_4 H^{{\alpha}(s){\dot{\alpha}}(s)}[{\rm D}_{{\alpha}_s},{\bar{\rm D}}_{{\dot{\alpha}}_s}][{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}]H_{{\gamma}{\alpha}(s-1){\dot{\gamma}}{\dot{\alpha}}(s-1)}\cr
&-\[\frac{2}{s}\](c_3-(2s+1)c_4)H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}\cr
&+\[\frac{2}{s}\](c_3-(2s+1)c_4)H^{{\alpha}(s){\dot{\alpha}}(s)}{\rm D}_{{\alpha}_s}{\bar{\rm D}}^2\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\cr
&+2(c_3-c_4)H^{{\alpha}(s){\dot{\alpha}}(s)}[{\rm D}_{{\alpha}_s},{\bar{\rm D}}_{{\dot{\alpha}}_s}]B_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
&+eB^{{\alpha}(s-1){\dot{\alpha}}(s-1)}B_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
&+h_1\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&+h_2\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\bar{\rm D}}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&+h_3\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\bar{\rm D}}^{{\dot{\alpha}}_s}{\rm D}_{{\alpha}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\cr
&+h_4\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}_{a_s}{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\cr
&+bB^{{\alpha}(s-1){\dot{\alpha}}(s-1)}\left(D^{{\alpha}_s}\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}
\right)\Big\}
~~~.
} \label{equ25}
\end{equation}
The invariance of this action under the corresponding gauge transformations is guaranteed
by the satisfaction of the following two Bianchi identities
\begin{equation}
\eqalign{ {~~~~}
0=&~{\rm D}^{{\alpha}_s}{\bm{{\cal G}}}_{{\alpha}(s){\dot{\alpha}}(s)}-\frac{1}{s!}{\rm D}^{{\alpha}_s}{\bar{\rm D}}_{({\dot{\alpha}}_s}{\bm{{\cal T}}}_{
{\alpha}(s){\dot{\alpha}}(s-1))}
~~~, \cr
&+\frac{1}{s!}{\bar{\rm D}}_{({\dot{\alpha}}_s}{\rm D}^2{\bm{{\cal Y}}}_{{\alpha}(s-1){\dot{\alpha}}(s-1))}-\frac{s-1}{s!s!}{\rm D}_{({\alpha}_{s-1}}
{\bar{\rm D}}_{({\dot{\alpha}}_s}{\rm D}^{{\gamma}}{\bm{{\cal Y}}}_{{\gamma}{\alpha}(s-2)){\dot{\alpha}}(s-1))}=0 ~~~, \cr
0=&~ {\rm D}_{({\alpha}_{s+1}}{\bm{{\cal T}}}_{{\alpha}(s)){\dot{\alpha}}(s-1)} ~~~,
} \label{equ26}
\end{equation}
where ${\bm{{\cal G}}}_{{\alpha}(s){\dot{\alpha}}(s)},~{\bm{{\cal T}}}_{{\alpha}(s){\dot{\alpha}}(s-1)},~{\bm{{\cal Y}}}_{{\alpha}(s-1)
{\dot{\alpha}}(s-1)}$ are the variations of the action with respect the corresponding superfields $H_{
{\alpha}(s){\dot{\alpha}}(s)},~\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}$, and $B_{{\alpha}(s-1){\dot{\alpha}}(s-1)}$.
The solution of the first bianchi identity gives:
$$
\eqalign{
h_1 &=-\[\frac{1}{s}\]\left(c_3-(2s+1)c_4\right) ~~~, ~~~
h_3 =0 ~~~, \cr
h_4 &=\[\frac{2(s+1)}{s^2}\]\left(c_3-(2s+1)c_4\right) ~~~, \cr
}
$$
\begin{equation}
\eqalign{
e &=\frac{1}{2}b-(c_3-c_4) ~~~, \cr
b&=-h_4=-\[\frac{2(s+1)}{s^2}\]\left(c_3-(2s+1)c_4\right) ~~~, \cr
b&=\[\frac{2(2s+1)}{s}\]\left(c_3-c_4\right) ~~~.
}
\label{equ27}
\end{equation}
the last two equations will give a relationship among $c_3$ and $c_4$
\begin{equation}
c_4=\[\frac{2s^2+2s+1}{(2s+1)^2}\]c_3 ~~~.
\label{equ28}
\end{equation}
The second Binachi identity has as a solution:
\begin{equation}
\eqalign{
&h_2=0 ~~,~~~
h_3=0 ~~~.
}
\label{equ29}
\end{equation}
So the full action takes the form
\begin{equation}
\eqalign{
S=\int d^8z\Big\{
&~\[\frac{2(s+1)^2}{s(2s+1)}\]c_3H^{{\alpha}(s){\dot{\alpha}}(s)}{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}(s){\dot{\alpha}}(s)}\cr
&+c_3 H^{{\alpha}(s){\dot{\alpha}}(s)}\pa_{{\alpha}_s{\dot{\alpha}}_s}\pa^{{\gamma}{\dot{\gamma}}}H_{{\gamma}{\alpha}(s-1){\dot{\gamma}}{\dot{\alpha}}(s-1)}\cr
&+ \[\frac{2s^2+2s+1}{(2s+1)^2}\]c_3H^{{\alpha}(s){\dot{\alpha}}(s)}[{\rm D}_{{\alpha}_s},{\bar{\rm D}}_{{\dot{\alpha}}_s}][{\rm D}^{{\gamma}},
{\bar{\rm D}}^{{\dot{\gamma}}}]H_{{\gamma}{\alpha}(s-1){\dot{\gamma}}{\dot{\alpha}}(s-1)}\cr
&+\[\frac{4s}{2s+1}\]c_3H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}\cr
&-\[\frac{4s}{2s+1}\]c_3H^{{\alpha}(s){\dot{\alpha}}(s)}{\rm D}_{{\alpha}_s}{\bar{\rm D}}^2\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\cr
&+\[\frac{4s(s+1)}{(2s+1)^2}\]c_3H^{{\alpha}(s){\dot{\alpha}}(s)}[{\rm D}_{{\alpha}_s},{\bar{\rm D}}_{{\dot{\alpha}}_s}]B_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
&+\[\frac{2(s+1)^2}{(2s+1)^2}\]c_3B^{{\alpha}(s-1){\dot{\alpha}}(s-1)}B_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
&+\[\frac{2s}{2s+1}\]c_3\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&-\[\frac{4(s+1)}{2s+1}\]c_3\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}_{a_s}{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{{\alpha}(s-1)
{\dot{\alpha}}(s)}\cr
&+\[\frac{4(s+1)}{2s+1}\]c_3B^{{\alpha}(s-1){\dot{\alpha}}(s-1)}\left(D^{{\alpha}_s}\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+
{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\right)\Big\}
~~~. }
\label{equ30}
\end{equation}
At this point we can use the equation of motion of the auxiliary superfield $B_{{\alpha}(s-1){\dot{\alpha}}(s-1)}$
\begin{equation}
\eqalign{
&{\bm{{\cal Y}}}_{{\alpha}(s-1){\dot{\alpha}}(s-1)}=0\Rightarrow\cr
&B_{{\alpha}(s-1){\dot{\alpha}}(s-1)}=-\[\frac{s}{s+1}\][{\rm D}^{{\alpha}_s},{\bar{\rm D}}^{{\dot{\alpha}}_s}]H_{{\alpha}(s){\dot{\alpha}}(s)}\cr
&~~~~~~~~~~~~~~~~~~~-\[\frac{2s+1}{s+1}\]\left(D^{{\alpha}_s}\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+{\bar{\rm D}}^{{\dot{\alpha}}_s}
\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\right) ~~~,
}
\label{equ31}
\end{equation}
in order to integrate it out and simplify the action. So our final action is:
\begin{equation}
\eqalign{
S=\int d^8z\Big\{
&~\[\frac{2(s+1)^2}{s(2s+1)}\]c_3H^{{\alpha}(s){\dot{\alpha}}(s)}{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}(s){\dot{\alpha}}(s)}\cr
&+c_3 H^{{\alpha}(s){\dot{\alpha}}(s)}\pa_{{\alpha}_s{\dot{\alpha}}_s}\pa^{{\gamma}{\dot{\gamma}}}H_{{\gamma}{\alpha}(s-1){\dot{\gamma}}{\dot{\alpha}}(s-1)}\cr
&+ \[\frac{1}{(2s+1)}\]c_3H^{{\alpha}(s){\dot{\alpha}}(s)}[{\rm D}_{{\alpha}_s},{\bar{\rm D}}_{{\dot{\alpha}}_s}][{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}]H_{{\gamma}{\alpha}(s-1)
{\dot{\gamma}}{\dot{\alpha}}(s-1)}\cr
&-\[\frac{4s}{2s+1}\]c_3H^{{\alpha}(s){\dot{\alpha}}(s)}{\rm D}_{{\alpha}_s}{\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}^{{\gamma}}\Upsilon_{{\gamma}{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
&+\[\frac{4s}{2s+1}\]c_3H^{{\alpha}(s){\dot{\alpha}}(s)}{\bar{\rm D}}_{{\dot{\alpha}}_s}{\rm D}_{{\alpha}_s}{\bar{\rm D}}^{{\dot{\gamma}}}\bar{\Upsilon}_{{\alpha}(s-1)
{\dot{\gamma}}{\dot{\alpha}}(s-1)}\cr
&-\[\frac{2(s+1)}{2s+1}\]c_3\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
&+\[\frac{4s}{2s+1}\]c_3\Upsilon^{{\alpha}(s){\dot{\alpha}}(s-1)}{\rm D}_{a_s}{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{{\alpha}(s-1){\dot{\alpha}}(s)}\Big\}
~~~, }
\label{equ32}
\end{equation}
and it is invariant under the following gauge transformations
\begin{equation}
\eqalign{
&\delta H_{{\alpha}(s){\dot{\alpha}}(s)}=\frac{1}{s!}{\bar{\rm D}}_{({\dot{\alpha}}_s}L_{{\alpha}(s){\dot{\alpha}}(s-1))}-\frac{1}{s!}{\rm D}_{({\alpha}_s}\bar{L
}_{{\alpha}(s-1)){\dot{\alpha}}(s)}\cr
&\delta\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}=\frac{1}{s!}{\bar{\rm D}}^{{\dot{\gamma}}}{\rm D}_{({\alpha}_s}\bar{L}_{{\alpha}(s-1)){\dot{\gamma}}{\dot{\alpha}}(s-1)}+{\rm D}^{
{\alpha}_{s+1}}\Lambda_{{\alpha}(s+1){\dot{\alpha}}(s-1)}
~~~.}
\label{equ33}
\end{equation}
From this action we can calculate the following superfields
$$
\eqalign{
{\bm{{\cal G}}}_{{\alpha}(s){\dot{\alpha}}(s)}=&~\[\frac{4(s+1)^2}{s(2s+1)}\]c_3{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}(s){\dot{\alpha}}(s)}\cr
&+\frac{2c_3}{s!s!}\pa_{({\alpha}_s({\dot{\alpha}}_s}\pa^{{\gamma}{\dot{\gamma}}}H_{{\gamma}{\alpha}(s-1)){\dot{\gamma}}{\dot{\alpha}}(s-1))}\cr
&+ \[\frac{2}{(2s+1)}\]\frac{c_3}{s!s!}[{\rm D}_{({\alpha}_s},{\bar{\rm D}}_{({\dot{\alpha}}_s}][{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}]H_{{\gamma}{\alpha}(s-1)){\dot{\gamma}}
{\dot{\alpha}}(s-1))}\cr
&-\[\frac{4s}{2s+1}\]\frac{c_3}{s!s!}{\rm D}_{({\alpha}_s}{\bar{\rm D}}_{({\dot{\alpha}}_s}{\rm D}^{{\gamma}}\Upsilon_{{\gamma}{\alpha}(s-1)){\dot{\alpha}}(s-1))}\cr
&+\[\frac{4s}{2s+1}\]\frac{c_3}{s!s!}{\bar{\rm D}}_{({\dot{\alpha}}_s}{\rm D}_{({\alpha}_s}{\bar{\rm D}}^{{\dot{\gamma}}}\bar{\Upsilon}_{{\alpha}(s-1)){\dot{\gamma}}{\dot{\alpha}}(
s-1))}\cr
~~~,}
$$
\begin{equation}
\eqalign{
{\bm{{\cal T}}}_{{\alpha}(s){\dot{\alpha}}(s-1)}=&-\[\frac{4s}{2s+1}\]\frac{c_3}{s!}{\rm D}_{({\alpha}_s}{\bar{\rm D}}^{{\dot{\gamma}}}{\rm D}^{{\gamma}}H_{
{\gamma}{\alpha}(s-1)){\dot{\alpha}}(s-1)}\cr
&-\[\frac{4(s+1)}{2s+1}\]c_3{\rm D}^2\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}\cr
&+\[\frac{4s}{2s+1}\]\frac{c_3}{s!}{\rm D}_{(a_s}{\bar{\rm D}}^{{\dot{\alpha}}_s}\bar{\Upsilon}_{{\alpha}(s-1)){\dot{\alpha}}(s)}
~~~,}
\label{equ34}
\end{equation}
which satisfhy the bianchi identities for this action
\begin{equation}
\eqalign{
0=&~{\rm D}^{{\alpha}_s}{\bm{{\cal G}}}_{{\alpha}(s){\dot{\alpha}}(s)}-\frac{1}{s!}{\rm D}^{{\alpha}_s}{\bar{\rm D}}_{({\dot{\alpha}}_s}{\bm{{\cal T}}}_{{\alpha}(s)
{\dot{\alpha}}(s-1))}\cr
0=&~ {\rm D}_{({\alpha}_{s+1}}{\bm{{\cal T}}}_{{\alpha}(s)){\dot{\alpha}}(s-1)}
~~~.}
\label{equ35}
\end{equation}
It is also straightforward to prove that they satisfy another identity
\begin{equation}
\eqalign{
{\rm D}^{{\alpha}_{2s+1}}{\bm{\cal W}}_{{\alpha}(2s+1)}=&~\[\frac{s(2s+1)}{4(s+1)^2}\]\frac{1}{c_3}\pa_{({\alpha}_{2s}}
{}^{{\dot{\alpha}}_s}\dots\pa_{({\alpha}_{s+1}}{}^{{\dot{\alpha}}_1}{\bm{{\cal G}}}_{{\alpha}(s)){\dot{\alpha}}(s)}\cr
&+i\[\frac{s^2}{4(s+1)^2}\]\frac{1}{c_3}{\rm D}_{({\alpha}_{2s}}{\bar{\rm D}}^2\pa_{({\alpha}_{2s-1}}{}^{{\dot{\alpha}}_{s-1}}\dots\pa_{
({\alpha}_{s+1}}{}^{{\dot{\alpha}}_1}{\bm{{\cal T}}}_{{\alpha}(s)){\dot{\alpha}}(s-1)}\cr
&+\[\frac{s^2}{4(s+1)^2}\]\frac{1}{c_3}{\rm D}_{({\alpha}_{2s}}\pa_{({\alpha}_{2s-1}}{}^{{\dot{\alpha}}_{s}}\dots\pa_{({\alpha}_{s}}
{}^{{\dot{\alpha}}_1}\bar{{\bm {{}\cal T}}}_{{\alpha}(s-1)){\dot{\alpha}}(s)}
~~~, }
\label{equ36}
\end{equation}
with
\begin{equation}
{\bm{\cal W}}_{{\alpha}(2s+1)}={\bar{\rm D}}^2{\rm D}_{({\alpha}_{2s+1}}\pa_{({\alpha}_{2s}}{}^{{\dot{\alpha}}_{s}}\dots\pa_{({\alpha}_{s+1}}
{}^{{\dot{\alpha}}_1}H_{{\alpha}(s)){\dot{\alpha}}(s)} ~~~.
\label{equ37}
\end{equation}
That means that on-shell (${\bm{{\cal T}}}={\bm{{\cal G}}}=0$) the object ${\bm{\cal
W}}_{{\alpha}(2s+1)}$ satisfies the
equations
\begin{equation}
\eqalign{
&{\rm D}^{{\alpha}_{2s+1}}{\bm{\cal W}}_{{\alpha}(2s+1)}=0 ~~~, ~~~
{\bar{\rm D}}_{{\dot{\alpha}}}{\bm{\cal W}}_{{\alpha}(2s+1)}=0 ~~~,
}
\label{equ38}
\end{equation}
therefore it describes a massless half odd superspin. Now we know that this theory,
on-shell has an irreducible representation propagating. The last thing we need to check
is whether these are the only degrees of freedom propagating or if there are more. The
easiest way to do that is to go to components notation and calculate the action in the
Wess-Zumino gauge. If the only thing propagating is this half odd supermultiplet,
the components action must be the Fronsdal action for bosons and fermions respectively.
~Because of the gauge transformation, we have the freedom to gauge away some
of the components. Specifically:\footnote{The definition of symmetric and antisymmetric
pieces of a field is the following $$\Phi_{{\gamma}{\alpha}(s-1)}=\Phi^{(S)}_{{\gamma}{\alpha}(s-1)}+\frac{s-1}{s!}
C_{{\gamma}({\alpha}_{s-1}}\Phi^{(A)}_{{\alpha}(s-2))}~~,~~\Phi^{(S)}_{{\gamma}{\alpha}(s-1)}=\frac{1}{s!}\Phi_{({\gamma}{\alpha}(s-1))}
~~,~~\Phi^{(A)}_{{\alpha}(s-2)}=C^{{\gamma}{\alpha}_{s-1}}\Phi_{{\gamma}{\alpha}(s-1)}$$
Furthermore the notation $\Phi^{(m,n)}$ represents the $\theta^{m}\bar{\theta}^n$
component in the taylor series of the superfield $\Phi$}\\
\begin{tabular}{c c}
Bosons: & Fermions: \\
\begin{tabular}{| l | l |}
\hline
Component & Gauged away by\\
\hline
$H^{(0,0)}_{{\alpha}(s){\dot{\alpha}}(s)}$ & $Re\left[L^{(0,1)(S)}_{{\alpha}(s){\dot{\alpha}}(s)}\right]$\\
\hline
$H^{(2,0)}_{{\alpha}(s){\dot{\alpha}}(s)}$ & $L^{(2,1)(S)}_{{\alpha}(s){\dot{\alpha}}(s)}$\\
\hline
$H^{(1,1)(A,S)}_{{\alpha}(s-1){\dot{\alpha}}(s+1)}$ & $\bar{L}^{(2,1)(S)}_{{\alpha}(s-1){\dot{\alpha}}(s+1)}$\\
\hline
$\Upsilon^{(1,0)(S)}_{{\alpha}(s+1){\dot{\alpha}}(s-1)}$ & $\Lambda^{(2,0)}_{{\alpha}(s+1){\dot{\alpha}}(s-1)}$\\
\hline
$\Upsilon^{(1,0)(A)}_{{\alpha}(s-1){\dot{\alpha}}(s-1)}$ & $\bar{L}^{(2,1)(A)}_{{\alpha}(s-1){\dot{\alpha}}(s-1)}$\\
\hline
$\Upsilon^{(0,1)(S)}_{{\alpha}(s){\dot{\alpha}}(s)}$ & $\Lambda^{(1,1)(A,S)}_{{\alpha}(s){\dot{\alpha}}(s)}$\\
\hline
$\Upsilon^{(0,1)(A)}_{{\alpha}(s){\dot{\alpha}}(s-2)}$ & $\Lambda^{(1,1)(A,A)}_{{\alpha}(s){\dot{\alpha}}(s-2)}$\\
\hline
$\Upsilon^{(1,2)(S)}_{{\alpha}(s+1){\dot{\alpha}}(s-1)}$ & $\Lambda^{(2,2)}_{{\alpha}(s+1){\dot{\alpha}}(s-1)}$\\
\hline
\end{tabular}
&
\begin{tabular}{| l | l |}
\hline
Component & Gauged away by\\
\hline
$H^{(1,0)(S)}_{{\alpha}(s+1){\dot{\alpha}}(s)}$ & $L^{(1,1)(S,S)}_{{\alpha}(s+1){\dot{\alpha}}(s)}$\\
\hline
$H^{(1,0)(A)}_{{\alpha}(s-1){\dot{\alpha}}(s)}$ & $\bar{L}^{(2,0)}_{{\alpha}(s-1){\dot{\alpha}}(s)}$\\
\hline
$\Upsilon^{(0,0)}_{{\alpha}(s){\dot{\alpha}}(s-1)}$ & $\Lambda^{(1,0)(A)}_{{\alpha}(s){\dot{\alpha}}(s-1)}$\\
\hline
$\Upsilon^{(0,2)}_{{\alpha}(s){\dot{\alpha}}(s-1)}$ & $\Lambda^{(1,2)(A)}_{{\alpha}(s){\dot{\alpha}}(s-1)}$\\
\hline
$\Upsilon^{(1,1)(S,S)}_{{\alpha}(s+1){\dot{\alpha}}(s)}$ & $\Lambda^{(2,1)(S)}_{{\alpha}(s+1){\dot{\alpha}}(s)}$\\
\hline
$\Upsilon^{(1,1)(S,A)}_{{\alpha}(s+1){\dot{\alpha}}(s-2)}$ & $\Lambda^{(2,1)(A)}_{{\alpha}(s+1){\dot{\alpha}}(s-2)}$\\
\hline
$\Upsilon^{(1,1)(A,S)}_{{\alpha}(s-1){\dot{\alpha}}(s)}$ & $\bar{L}^{(2,2)}_{{\alpha}(s-1){\dot{\alpha}}(s)}$\\
\hline
\end{tabular}
\\
\end{tabular}
So in the Wess-Zumino gauge for the two superfields are:
\begin{equation}
\eqalign{
H_{{\alpha}(s){\dot{\alpha}}(s)}=&~\theta^{{\alpha}_{s+1}}\bar{\theta}^{{\dot{\alpha}}_{s+1}}h_{{\alpha}(s+1){\dot{\alpha}}(s+1)}-
\frac{s}{s!s!}\theta_{({\alpha}_{s}}\bar{\theta}_{({\dot{\alpha}}_{s}}h_{{\alpha}(s-1)){\dot{\alpha}}(s-1))}\cr
&+\frac{1}{\sqrt{2}}\bar{\theta}^2\theta^{{\alpha}_{s+1}}\psi_{{\alpha}(s+1){\dot{\alpha}}(s)}+\frac{1}{\sqrt{2}}
\theta^2\bar{\theta}^{{\dot{\alpha}}_{s+1}}
\bar{\psi}_{{\alpha}(s){\dot{\alpha}}(s+1)}\cr
&+\frac{1}{\sqrt{2}s!}\theta^2\bar{\theta}_{({\dot{\alpha}}_s}\psi_{{\alpha}(s){\dot{\alpha}}(s-1)}-\frac{1}{\sqrt{2}s!}
\bar{\theta}^2
\theta_{({\alpha}_s}\bar{\psi}_{{\alpha}(s-1)){\dot{\alpha}}(s)}\cr
&+\theta^2\bar{\theta}^2{A}_{{\alpha}(s){\dot{\alpha}}(s)} ~~~,
} \label{equ39}
\end{equation}
and
\begin{equation}
\eqalign{
\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}=&~\theta^2\[\rho_{{\alpha}(s){\dot{\alpha}}(s-1)}+\frac{1}{\sqrt{2}}\psi_{{\alpha}(s){\dot{\alpha}}(s-1)}\]\cr
&+\frac{1}{\sqrt{2}s!(s-1)!}\theta_{({\alpha}_s}\bar{\theta}_{({\dot{\alpha}}_{s-1}}\psi_{{\alpha}(s-1)){\dot{\alpha}}(s-2))}\cr
&+\theta^2\bar{\theta}^{{\dot{\alpha}}_s}\[v_{{\alpha}(s){\dot{\alpha}}(s)}+iw_{{\alpha}(s){\dot{\alpha}}(s)}-\frac{s}{2s+1}A_{{\alpha}(s)
{\dot{\alpha}}(s)}\right.\cr
&~~~~~~~~~~~~-i\frac{s^2+2s+2}{2(s!)^2}\pa_{({\alpha}_s({\dot{\alpha}}_s}h_{{\alpha}(s-1)){\dot{\alpha}}(s-1))}\cr
&~~~~~~~~~~~~+i\left.\frac{s}{2}\pa^{{\alpha}_{s+1}{\dot{\alpha}}_{s+1}}h_{{\alpha}(s+1){\dot{\alpha}}(s+1)}\]\cr
&+\frac{s-1}{s!}\theta^2\bar{\theta}_{({\dot{\alpha}}_{s-1}}\[U_{a(s){\dot{\alpha}}(s-2))}+i\frac{s+1}{s!}
\pa_{({\alpha}_s}{}^{{\dot{\gamma}}}h_{{\alpha}(s-1)){\dot{\gamma}}{\dot{\alpha}}(s-2))}\]\cr
&+\frac{s}{(s+1)!}\bar{\theta}^2\theta_{({\alpha}_s}\[S_{{\alpha}(s-1)){\dot{\alpha}}(s-1)}+i P_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\]\cr
&+\theta^2\bar{\theta}^2\[\beta_{{\alpha}(s){\dot{\alpha}}(s-1)}+\frac{i}{2\sqrt{2}s!(s-1)!}\pa_{({\alpha}_s({\dot{\alpha}}_{
s-1}}\psi_{{\alpha}(s-1)){\dot{\alpha}}(s-2))}\right.\cr
&~~~~~~-\frac{i}{\sqrt{2}(s+1)!}\pa_{{\alpha}_s}{}^{{\dot{\alpha}}_s}\bar{\psi}_{{\alpha}(s-1)){\dot{\alpha}}(s)}
+\left.\frac{i}{2}\frac{s}{(s+1)!}\pa_{({\alpha}_s}{}^{{\dot{\alpha}}_s}\bar{\rho}_{{\alpha}(s-1)){\dot{\alpha}}(s)}\]}\label{equ40}
\end{equation}
From mass dimensions arguments we can tell immediately that the components $A,~U,~S,~P,~\rho,~\beta$ are
auxiliary fields, so they cannot appear with derivatives in the component action. The rest of the degrees of freedom
left are exactly those that compose the half odd superspin supermultiplet and therefore, the action in components
has to be the Fronsdal action.
To see in details how all this takes place, we substitute the component field expansions from the above expression
for the superfields to the action(\ref{equ32}). The bosonic piece is:
\begin{equation}
\eqalign{
S_{Bosons}=\int d^4x&~\[\frac{2(s+1)^2}{s(2s+1)}\]c_3h^{{\alpha}(s+1){\dot{\alpha}}(s+1)}\Box h_{{\alpha}(s+1)
{\dot{\alpha}}(s+1)}\cr
-&\[\frac{(s+1)^3}{s(2s+1)}\]c_3h^{{\alpha}(s+1){\dot{\alpha}}(s+1)}\pa_{{\alpha}_{s+1}{\dot{\alpha}}_{s+1}}\pa^{{\gamma}{\dot{\gamma}}}h_{
{\gamma}{\alpha}(s){\dot{\gamma}}{\dot{\alpha}}(s)}\cr
+&\[\frac{2(s+1)^3}{(2s+1)}\]c_3h^{{\alpha}(s+1){\dot{\alpha}}(s+1)}\pa_{{\alpha}_{s+1}{\dot{\alpha}}_{s+1}}\pa_{{\alpha}_s
{\dot{\alpha}}_s}h_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
-&\[\frac{2(s+1)^3}{s}\]c_3h^{{\alpha}(s-1){\dot{\alpha}}(s-1)}\Box h_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
-&\[\frac{(s+1)^3(s-1)^2}{s(2s+1)}\]c_3h^{{\alpha}(s-1){\dot{\alpha}}(s-1)}\pa_{{\alpha}_{s-1}{\dot{\alpha}}_{s-1}}\pa^{
{\gamma}{\dot{\gamma}}}h_{{\gamma}{\alpha}(s-2){\dot{\gamma}}{\dot{\alpha}}(s-2)}\cr
&+\[\frac{4s}{2s+1}\]c_3 S^{{\alpha}(s-1){\dot{\alpha}}(s-1)}S_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
&+\[\frac{4s}{2s+1}\]c_3 P^{{\alpha}(s-1){\dot{\alpha}}(s-1)} P_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
&+\[\frac{4(s+1)^3-16s^4}{s(2s+1)}\]c_3A^{{\alpha}(s){\dot{\alpha}}(s)}{{{A}}}_{{\alpha}(s){\dot{\alpha}}(s)}\cr
&+4c_3v^{{\alpha}(s){\dot{\alpha}}(s)}v_{{\alpha}(s){\dot{\alpha}}(s)}\cr
&-\[\frac{4}{2s+1}\]c_3w^{{\alpha}(s){\dot{\alpha}}(s)}w_{{\alpha}(s){\dot{\alpha}}(s)}\cr
&+\[\frac{2(s+1)(s-1)}{s(2s+1)}\]c_3U^{{\alpha}(s){\dot{\alpha}}(s-2)}U_{{\alpha}(s)
{\dot{\alpha}}(s-2)}+c.c.
}
\end{equation}
The component fields above all correspond to the zero-$\theta$ limit of a corresponding
superfield. The equations of motions for the auxiliary superfields are:
\begin{equation}
\eqalign{
&{{{A}}}_{{\alpha}(s){\dot{\alpha}}(s)}=0 ~~~, \cr
&S_{{\alpha}(s-1){\dot{\alpha}}(s-1)}=0 ~~~, \cr
&P_{{\alpha}(s-1){\dot{\alpha}}(s-1)}=0 ~~~, \cr
&v_{{\alpha}(s){\dot{\alpha}}(s)}=0 ~~~, \cr
&U_{{\alpha}(s){\dot{\alpha}}(s-2)}=0 ~~~, \cr
&w_{{\alpha}(s){\dot{\alpha}}(s)}=0 ~~~. }
\label{equ43}
\end{equation}
and the final action for the propagating bosonic components is
\begin{equation}
\eqalign{
S_{Bosons}=\int d^4x&~\[\frac{2(s+1)^2}{s(2s+1)}\]c_3h^{{\alpha}(s+1){\dot{\alpha}}(s+1)}\Box h_{{\alpha}(s+1)
{\dot{\alpha}}(s+1)}\cr
-&\[\frac{(s+1)^3}{s(2s+1)}\]c_3h^{{\alpha}(s+1){\dot{\alpha}}(s+1)}\pa_{{\alpha}_{s+1}{\dot{\alpha}}_{s+1}}\pa^{{\gamma}{\dot{\gamma}}}h_{
{\gamma}{\alpha}(s){\dot{\gamma}}{\dot{\alpha}}(s)}\cr
+&\[\frac{2(s+1)^3}{(2s+1)}\]c_3h^{{\alpha}(s+1){\dot{\alpha}}(s+1)}\pa_{{\alpha}_{s+1}{\dot{\alpha}}_{s+1}}\pa_{{\alpha}_s
{\dot{\alpha}}_s}h_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
-&\[\frac{2(s+1)^3}{s}\]c_3h^{{\alpha}(s-1){\dot{\alpha}}(s-1)}\Box h_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
-&\[\frac{(s+1)^3(s-1)^2}{s(2s+1)}\]c_3h^{{\alpha}(s-1){\dot{\alpha}}(s-1)}\pa_{{\alpha}_{s-1}{\dot{\alpha}}_{s-1}}\pa^{
{\gamma}{\dot{\gamma}}}h_{{\gamma}{\alpha}(s-2){\dot{\gamma}}{\dot{\alpha}}(s-2)}
~~~.}
\label{equ44}
\end{equation}
By setting $c_3=\frac{s(2s+1)}{2(s+1)^2}$ we obtain:
\begin{equation}
\eqalign{
S_{Bosons}=\int d^4x&~h^{{\alpha}(s+1){\dot{\alpha}}(s+1)}\Box h_{{\alpha}(s+1)
{\dot{\alpha}}(s+1)}\cr
-&\frac{s+1}{2}h^{{\alpha}(s+1){\dot{\alpha}}(s+1)}\pa_{{\alpha}_{s+1}{\dot{\alpha}}_{s+1}}\pa^{{\gamma}{\dot{\gamma}}}h_{
{\gamma}{\alpha}(s){\dot{\gamma}}{\dot{\alpha}}(s)}\cr
+&(s+1)sh^{{\alpha}(s+1){\dot{\alpha}}(s+1)}\pa_{{\alpha}_{s+1}{\dot{\alpha}}_{s+1}}\pa_{{\alpha}_s
{\dot{\alpha}}_s}h_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
-&(s+1)(2s+1)h^{{\alpha}(s-1){\dot{\alpha}}(s-1)}\Box h_{{\alpha}(s-1){\dot{\alpha}}(s-1)}\cr
-&\frac{(s+1)(s-1)^2}{2}h^{{\alpha}(s-1){\dot{\alpha}}(s-1)}\pa_{{\alpha}_{s-1}{\dot{\alpha}}_{s-1}}\pa^{
{\gamma}{\dot{\gamma}}}h_{{\gamma}{\alpha}(s-2){\dot{\gamma}}{\dot{\alpha}}(s-2)}
~~~,}
\end{equation}
which is the Fronsdal action for a propagating spin-($s+1$) bosonic field. For the
limiting value of $s$ $=$ 1, this is the linearized Einstein-Hilbert action.
The fermionic piece of the action is:
$$
\eqalign{
S_{Fermions}=\int d^4x &\[\frac{4(s+1)^2}{s(2s+1)}\]ic_3\bar{\psi}^{{\alpha}(s){\dot{\alpha}}(s+1)}\pa^{
{\alpha}_{s+1}}{}_{{\dot{\alpha}}_{s+1}}\psi_{{\alpha}(s+1){\dot{\alpha}}(s)} {~~~~~~~~~~~~} {~~~~~~~~~~~~~~} \cr
}
$$
\begin{equation}
\eqalign{
~~~~~~&-\[\frac{2}{s}\]ic_3\bar{\psi}^{{\alpha}(s-1){\dot{\alpha}}(s)}\pa^{{\alpha}_s}{}_{{\dot{\alpha}}_s}\psi_{{\alpha}(s){\dot{\alpha}}(s-1)}\cr
&+\[\frac{2(s+1)}{2s+1}\]ic_3\psi^{{\alpha}(s+1){\dot{\alpha}}(s)}\pa_{{\alpha}_{s+1}{\dot{\alpha}}_{s}}\psi_{{\alpha}(s){\dot{\alpha}}(s-1)
}+c.c.\cr
&-\[\frac{2(s+1)^2}{s(2s+1)}\]ic_3\psi^{{\alpha}(s){\dot{\alpha}}(s-1)}\pa_{{\alpha}_s{\dot{\alpha}}_{s-1}}\psi_{{\alpha}(s-1){\dot{\alpha}}(
s-2)}+c.c.\cr
&-\[\frac{2(s+1)^2}{s(2s+1)}\]ic_3\bar{\psi}^{{\alpha}(s-2){\dot{\alpha}}(s-1)}\pa^{{\alpha}_{s-1}}{}_{{\dot{\alpha}}_{s-1}}
\psi_{{\alpha}(s-1){\dot{\alpha}}(s-2)}\cr
&+\[\frac{4(s+1)}{2s+1}\]c_3\beta^{{\alpha}(s){\dot{\alpha}}(s-1)}\rho_{{\alpha}(s){\dot{\alpha}}(s-1)}+c.c.\cr
~~~. }
\end{equation}
The equation of motion for the fermionic auxiliary fields are
\begin{equation}
\rho_{{\alpha}(s){\dot{\alpha}}(s-1)}=0 ~~~, ~~~\beta_{{\alpha}(s){\dot{\alpha}}(s-1)}=0
\label{equ47}
\end{equation}
and the action for propagating fermions takes it's final form:
\begin{equation}
\eqalign{
S_{Fermions}=\int d^4x &\[\frac{2(s+1)^2}{s(2s+1)}\]ic_3\bar{\psi}^{{\alpha}(s){\dot{\alpha}}(s+1)}\pa^{
{\alpha}_{s+1}}{}_{{\dot{\alpha}}_{s+1}}\psi_{{\alpha}(s+1){\dot{\alpha}}(s)}\cr
&-\[\frac{2}{s}\]ic_3\bar{\psi}^{{\alpha}(s-1){\dot{\alpha}}(s)}\pa^{{\alpha}_s}{}_{{\dot{\alpha}}_s}\psi_{{\alpha}(s){\dot{\alpha}}(s-1)}\cr
&+\[\frac{2(s+1)}{2s+1}\]ic_3\psi^{{\alpha}(s+1){\dot{\alpha}}(s)}\pa_{{\alpha}_{s+1}{\dot{\alpha}}_{s}}\psi_{{\alpha}(s){\dot{\alpha}}(s-1)
}+c.c.\cr
&-\[\frac{2(s+1)^2}{s(2s+1)}\]ic_3\psi^{{\alpha}(s){\dot{\alpha}}(s-1)}\pa_{{\alpha}_s{\dot{\alpha}}_{s-1}}\psi_{{\alpha}(s-1){\dot{\alpha}}(
s-2)}+c.c.\cr
&-\[\frac{2(s+1)^2}{s(2s+1)}\]ic_3\bar{\psi}^{{\alpha}(s-2){\dot{\alpha}}(s-1)}\pa^{{\alpha}_{s-1}}{}_{{\dot{\alpha}}_{s-1}}
\psi_{{\alpha}(s-1){\dot{\alpha}}(s-2)}
~~~. }
\label{equ48}
\end{equation}
Let's set the value for $c_3=\frac{s(2s+1)}{2(s+1)^2}$ as in the bosonic case,
the action becomes
\begin{equation}
\eqalign{
S_{Fermions}=\int d^4x~~ &i\bar{\psi}^{{\alpha}(s){\dot{\alpha}}(s+1)}\pa^{
{\alpha}_{s+1}}{}_{{\dot{\alpha}}_{s+1}}\psi_{{\alpha}(s+1){\dot{\alpha}}(s)}\cr
&-\[\frac{2s+1}{(s+1)^2}\]i\bar{\psi}^{{\alpha}(s-1){\dot{\alpha}}(s)}\pa^{{\alpha}_s}{}_{{\dot{\alpha}}_s}\psi_{{\alpha}(s){\dot{\alpha}}(s-1)}\cr
&+\[\frac{s}{s+1}\]i\psi^{{\alpha}(s+1){\dot{\alpha}}(s)}\pa_{{\alpha}_{s+1}{\dot{\alpha}}_{s}}\psi_{{\alpha}(s){\dot{\alpha}}(s-1)
}+c.c.\cr
&-i\psi^{{\alpha}(s){\dot{\alpha}}(s-1)}\pa_{{\alpha}_s{\dot{\alpha}}_{s-1}}\psi_{{\alpha}(s-1){\dot{\alpha}}(
s-2)}+c.c.\cr
&-i\bar{\psi}^{{\alpha}(s-2){\dot{\alpha}}(s-1)}\pa^{{\alpha}_{s-1}}{}_{{\dot{\alpha}}_{s-1}}
\psi_{{\alpha}(s-1){\dot{\alpha}}(s-2)}
~~~. }
\end{equation}
which is the Fronsdal action for spin-($s+1/2$).
Therefore we conclude only an irreducible supermultiplet propagates on-shell
and therefore the action(\ref{equ32}) describes a massless half odd superspin
$Y$ $=$ $s$+1/2.
The counting of the off-shell degrees of freedom for this action including all the auxiliary
fields is:
\begin{center}
\begin{tabular} {| c | c | c |}
\hline
Component Field(s) & Bosonic & Fermionic\\
\hline
$h_{{\alpha}(s+1){\dot{\alpha}}(s+1)}$~/~$h_{{\alpha}(s-1){\dot{\alpha}}(s-1)}$ & $s^2 + 2 s + 3$ & {}\\
\hline
$\psi_{{\alpha}(s+1){\dot{\alpha}}(s)}$~/~$\psi_{{\alpha}(s){\dot{\alpha}}(s-1)}$~/~$\psi_{{\alpha}(s-1){\dot{\alpha}}(s-2)}$ & {} & $4 (
s^2 + s + 1 ) $\\
\hline
$A_{{\alpha}(s){\dot{\alpha}}(s)}$ & $(s+1)^2$ & {}\\
\hline
$\rho_{{\alpha}(s){\dot{\alpha}}(s-1)}$ & {} & $2 s (s+1) $\\
\hline
$U_{{\alpha}(s){\dot{\alpha}}(s-2)} $ & $ 2(s+1)(s-1)$ & {}\\
\hline
$v_{{\alpha}(s){\dot{\alpha}}(s)} $ & $(s+1)^2$ & {}\\
\hline
$w_{{\alpha}(s){\dot{\alpha}}(s)}$ & $(s+1)^2$ & {}\\
\hline
$S_{{\alpha}(s-1){\dot{\alpha}}(s-1)}$ & $s^2$ & {}\\
\hline
$P_{{\alpha}(s-1){\dot{\alpha}}(s-1)}$ & $s^2$ & {}\\
\hline
$\beta_{{\alpha}(s){\dot{\alpha}}(s-1)}$ & {} & $2 s (s+1) $\\
\hline\hline
$~$ & $8s^2+8s+4$ & $8s^2+8s+4$\\
\hline
\end{tabular}
\end{center}
~~The results in (\ref{equ35}), (\ref{equ36}), and (\ref{equ37}), taken together
with the component expansions in (\ref{equ39}) and (\ref{equ40}), and the
component results discussed thereafter are very revealing...when one considers them
for the special case of the $s$ $=$ 1 theory\footnote{When we take the limit to $s=1$
we have to keep in mind, that fields with a negative \newline $~~~~~~$ number of
undotted or dotted indices must vanish. That is true because these compo- \newline
$~~~~~~$ nents do not exist in the $\theta$ expansion of the superfields. Specifically
the component \newline $~~~~~~$ $\psi_{{\alpha}(s-1){\dot{\alpha}}(s-2)}$ which is the
antisymmetric component of the $\theta\bar{\theta}$ term of the $\Upsilon$ superfield
\newline $~~~~~~$
doesn't exist in the $s=1$ limit, so it must disappear from the action}.
~~The first component level off-shell description of supergravity was provided in 1977
in a work by Breitenlohner \cite{nonmin1}. Two years later and in a subsequent
series of papers \cite{nonmin2}, these results were put into the context of the general
superspace formalism for 4D, $\cal N$ $=$ 1 superfield supergravity. These old results
and the special case of the higher spin $s$ $=$ 1 results for B-series discussed above
match perfectly. This is especially clear from an examination of the auxiliary fields in
the table immediately above. In $s$ $=$ 1 limit, only the $U$ auxiliary boson must
be set to zero and the remaining fields are the well known ones of the non-minimal
off-shell 4D, $\cal N$ $=$ 1 SG multiplet.
Another way to see this, one can initially compare the results of the current paper
for the component expansions given in (\ref{equ39}) and (\ref{equ40}) to the similar
expansions given in equation (4.9) of the first work in Ref. \cite{nonmin2}. The
(\ref{equ35}), (\ref{equ36}), and (\ref{equ37}), can be compared with the linearized
versions of the results found in the remaining papers of Ref. \cite{nonmin2}. In other
words, the implication of our present effort reveals that the non-minimal formulation
of Breitenlohner is the lowest member of a class of arbitrary higher superspin, off-shell
formulation of massless supermultiplets!
$$
\vCent
{\setlength{\unitlength}{1mm}
\begin{picture}(-20,-140)
\put(-56,-61){\includegraphics[width=4.5in]{Tower}}
\end{picture}}
$$
\vskip2.5in
\section{Perspectives On Future Investigations}
$~~~$ In the current work, we have been able to advance the state-of-the-art with
regard to the understanding of 4D, $\cal N $ $=$ 1 superfields and the issue of higher
spin supermultiplets. The discovery of the B-series of superfield theories, suggests
that many features of off-shell 4D, $\cal N $ $=$ 1 supergravity may well persists
in the cases of higher spin ($s$ $>$ 1). The gauge transformation law in (\ref{equ03})
for $s$ $=$ 1 is known to define the superspace superconformal group. It thus seems
reasonable for values of $s$ $\ne$ 1 to use this as a definition of the 4D, $\cal N $ $=$ 1
superspace superconformal group acting on the entire B-series of theories. Furthermore,
there is no obvious reason not to use this to define a 4D, $\cal N $ $=$ 1 superspace
superconformal group for the KPS-series also.
If it is accepted that the gauge transformation law of $H_{{\alpha}(s){\dot{\alpha}}(s)}$ defines a 4D, $\cal
N $ $=$ 1 superspace superconformal group, the second equation in (\ref{equ33}) has an
obvious interpretation. In the case of $s$ $=$ 1 limit, the superfield $\Upsilon_{{\alpha}(s){\dot{\alpha}}(s-1)}$ is
known to constitute a conformal compensator whose functions is to break the 4D, $\cal N $
$=$ 1 superspace superconformal group down to the 4D, $\cal N $ $=$ 1 superspace super
Poincar\' e group. Once more it is suggestive that this interpretation can be carried over
to the entirety of the B-series and as well to the KPS-series (though the governing equations
for the KPS-series are given by (\ref{equ17})). For both cases we have verified the existence
of field strength superfields ${\bm{\cal W}}_{{\alpha}(2s+1)}$, $ {\bm{{\cal G}}}_{{\alpha}(s){\dot{\alpha}}(
s)}$ and ${\bm{{\cal T}}}_{{\alpha}(s){\dot{\alpha}}(s-1)}$ which occur for both the B-series and the
KPS-series.
In a future work, we will revisit all of these results in the context of a Fock-space formulation.
We conjecture that all the structures we have met in this investigation will likely generalize
to such a formulation. Should this be the case, then we may have a new avenue to
ask questions of covariant superstring field theory. Can there exist a limit of covariant
superstring field theory which recovers all the structure found in a Fock space extension
of our current work?
${~~~}$ \newline
${~~~~~}$``{\it {Never express yourself more clearly than you are able to think.
}}"${~~~}$ \newline
\newline $~~~~~~~$ -- Niels Bohr
\newline ${~~~}$ \newline
\noindent
{\Large\bf Acknowledgments}
This research was supported in part by the endowment of the John S.~Toll Professorship, the University of Maryland
Center for String \& Particle Theory, National Science Foundation Grant PHY-0354401. This work is also supported
by U.S. Department of Energy (D.O.E.) under cooperative agreement DEFG0205ER41360. SJG offers additional
gratitude to the M.\ L.\ K. Visiting Professorship and to the M.\ I.\ T.\ Center for Theoretical Physics for support and
hospitality extended during the undertaking of this work.
\newpage
{\bf \Large {Appendix: Recovering the missing $s=1$ pieces}}
~~~For the purpose of completeness we study separately the case of $s=1$. This limit is special
because in a practical level the index structure of the entire theory gets simplified and \eqref{equ04},
which was a guideline, becomes simpler as well. Furthermore it is known that there exist two dual,
well studied, theories of supergravity, the minimal and the new minimal. We would like to find if and
how they emerge from our construction. For this purpose we will not bother, with the $s=1$ limits of
the two theories described above. Instead we will search for different routes that could lead to a
consistent theory.
There are a couple of interesting observations that one can make. Using the table above that counts
the off-shell degrees of freedom, in the $s=1$ limit gives the answer twenty. Minimal and new minimal
formulations of supergravity are known to have twelve off-shell degrees of freedom. So it can not be a
limit of the above theories. This means that in our framework there must be a different mechanism that is
capable of generating these theories. We would like to find this.
The second and more important observation is that the superfield $H$ in the s=1 limit includes all the
propagating bosonic and fermionic degrees of freedom need to construct the Fronsdal action. It is very
easy to verify that, just be looking equation \eqref{equ39} and \eqref{equ40}. One can check that in the
$s=1$ limit all the $h$'s and $\psi$'s components are in the taylor expansion of the $H_{{\alpha}{\dot{\alpha}}}$. This
means that the compensator looses one of it's roles, to provide the extra degrees of freedom needed
for the irreducible representation. It's sole purpose purpose now is to guarantee the gauge invariance
of the action. This infers that the set of all compensators must be auxiliary superfields (so their mass
dimensions must be one) and as a result their gauge transformation must be made out of 3 ${\rm D}$'s
(${\bar{\rm D}}$'s) acting on the only gauge parameter available $L_{{\alpha}}$. So we must look for compensators
that transform like
\begin{center}
\begin{tabular}{c c}
\begin{minipage}{1.9in
\begin{itemize}
\item ~${\bar{\rm D}}^2{\rm D}^{{\alpha}}L_{{\alpha}}$
\end{itemize}
\end{minipage}
&
\begin{minipage}{1.9in
\begin{itemize}
\item ~${\rm D}^{{\alpha}}{\bar{\rm D}}^2L_{{\alpha}}$
\end{itemize}
\end{minipage}
\\
\begin{minipage}{1.9in
\begin{itemize}
\item ~${\rm D}^2{\bar{\rm D}}_{{\dot{\alpha}}}L_{{\alpha}}$
\end{itemize}
\end{minipage}
&
\begin{minipage}{1.9in
\begin{itemize}
\item ~${\bar{\rm D}}_{{\dot{\alpha}}}{\rm D}^2L_{{\alpha}}$
\end{itemize}
\end{minipage}
\end{tabular}
\end{center}
The last two possibilities will introduce compensators with exactly the same index structure (and therefore fields content) as the main superfield, that is why we will not allow them. So there are two cases left that correspond at the minimal and new minimal formulations of supergravity.
The starting action is the $s=1$ version of \eqref{equ01} and the change of this action under the gauge transformation
\eqref{equ03} is
\begin{equation}
\eqalign{
\delta S=\int d^8z\Bigg\{
&~\[-2c_1+2c_2+2c_3+6c_4\]H^{{\alpha}{\dot{\alpha}}}{\bar{\rm D}}_{{\dot{\alpha}}}
{\rm D}^2{\bar{\rm D}}^2L_{{\alpha}} \cr
&+2c_2 H^{{\alpha}){\dot{\alpha}}}{\bar{\rm D}}^2{\rm D}^2{\bar{\rm D}}_{{\dot{\alpha}}}L_{{\alpha}}\cr
&+\[-2c_3+6c_4\]H^{{\alpha}{\dot{\alpha}}}{\rm D}_{{\alpha}}{\bar{\rm D}}^2{\rm D}^{{\gamma}}
{\bar{\rm D}}_{{\dot{\alpha}}}L_{{\gamma}}\cr
&+\[2c_3-2c_4\]H^{{\alpha}{\dot{\alpha}}}{\bar{\rm D}}_{{\dot{\alpha}}}{\rm D}_{{\alpha}}{\bar{\rm D}}^2{\rm D}^{{\gamma}}L_{{\gamma}}\cr
&+c.c.\Bigg\} ~~~.
}
\end{equation}
Setting $c_2=0$ gives
\begin{equation}
\eqalign{
\delta S=\int d^8z\Bigg\{
&~H^{{\alpha}{\dot{\alpha}}}\Big\{(-2c_3+6c_4){\rm D}_{{\alpha}}{\bar{\rm D}}_{{\dot{\alpha}}}-(-2c_1+2c_3+6c_4){\bar{\rm D}}_{{\dot{\alpha}}}{\rm D}_{{\alpha}}\Big\}{\rm D}^{{\gamma}}{\bar{\rm D}}^2L_{{\gamma}}\cr
&+2(c_3-c_4)H^{{\alpha}{\dot{\alpha}}}{\bar{\rm D}}_{{\dot{\alpha}}}{\rm D}_{{\alpha}}{\bar{\rm D}}^2{\rm D}^{{\gamma}}L_{{\gamma}}\cr
&+c.c.\Bigg\} ~~~.
}
\end{equation}
If $-2c_3+6c_4=-2c_1+2c_3+6c_4 \Rightarrow c_1=2c_3$
then
\begin{equation}
\eqalign{
\delta S=\int d^8z\Bigg\{
&~(-2c_3+6c_4)H^{{\alpha}{\dot{\alpha}}}\[{\rm D}_{{\alpha}},{\bar{\rm D}}_{{\dot{\alpha}}}\]\Big\{{\rm D}^{{\gamma}}{\bar{\rm D}}^2L_{{\gamma}}+{\bar{\rm D}}^{{\dot{\gamma}}}{\rm D}^2\bar{L}_{{\dot{\gamma}}}\Big\}\cr
&+2(c_3-c_4)H^{{\alpha}{\dot{\alpha}}}{\bar{\rm D}}_{{\dot{\alpha}}}{\rm D}_{{\alpha}}\Big\{{\bar{\rm D}}^2{\rm D}^{{\gamma}}L_{{\gamma}}\Big\}+c.c.\Bigg\} ~~~.
}
\end{equation}
At this point, we introduce two compensators, a real scalar $U$ with mass dimensions $[U]=1$ and a complex
scalar $\sigma$ with mass dimensions $[\sigma]=1$. Their transformations are defined to be
\begin{equation}
\eqalign{
&\delta U={\rm D}^{{\gamma}}{\bar{\rm D}}^2L_{{\gamma}}+{\bar{\rm D}}^{{\dot{\gamma}}}{\rm D}^2\bar{L}_{{\dot{\gamma}}}\cr
&\delta {\sigma}={\bar{\rm D}}^2{\rm D}^{{\gamma}}L_{{\gamma}}
}
\end{equation}
We also add to the action the following terms so the compensators have dynamics
\begin{equation}
\eqalign{
S_{c}=\int d^8z\Bigg\{
&~~-(-2c_3+6c_4)H^{{\alpha}{\dot{\alpha}}}\[{\rm D}_{{\alpha}},{\bar{\rm D}}_{{\dot{\alpha}}}\]U\cr
&~~-2(c_3-c_4)H^{{\alpha}{\dot{\alpha}}}{\bar{\rm D}}_{{\dot{\alpha}}}{\rm D}_{{\alpha}}\sigma +c.c.\Bigg\}\cr
{}\cr
S_{k.e}=\int d^8z\Bigg\{
&~~bU^2+e{\sigma}\bar{{\sigma}}+f{\sigma}\s+f^*\bar{{\sigma}}\bar{{\sigma}}\Bigg\}\cr
S_{int.}=\int d^8z\Bigg\{
&~~gU({\sigma}+\bar{{\sigma}})\Bigg\}
}
\end{equation}
The full action is
\begin{equation}
\eqalign{
S=\int d^8z\Bigg\{
~&2c_3 H^{{\alpha}{\dot{\alpha}}}{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}{\dot{\alpha}}}\cr
+&c_3 H^{{\alpha}{\dot{\alpha}}}\pa_{{\alpha}{\dot{\alpha}}}\pa^{{\gamma}{\dot{\gamma}}}H_{{\gamma}{\dot{\gamma}}}\cr
+&c_4 H^{{\alpha}{\dot{\alpha}}}[{\rm D}_{{\alpha}},{\bar{\rm D}}_{{\dot{\alpha}}}][{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}]H_{{\gamma}{\dot{\gamma}}}\cr
-&(-2c_3+6c_4)H^{{\alpha}{\dot{\alpha}}}\[{\rm D}_{{\alpha}},{\bar{\rm D}}_{{\dot{\alpha}}}\]U\cr
-&2(c_3-c_4)H^{{\alpha}{\dot{\alpha}}}{\bar{\rm D}}_{{\dot{\alpha}}}{\rm D}_{{\alpha}}\sigma +c.c.\cr
+&bU^2+e{\sigma}\bar{{\sigma}}+f{\sigma}\s+f^*\bar{{\sigma}}\bar{{\sigma}}\cr
+&gU({\sigma}+\bar{{\sigma}})\Bigg\}
}
\end{equation}
The invariance of the above action under the gauge transformations gives the following
Bianchi identity
\begin{equation}
{\bar{\rm D}}^{{\dot{\alpha}}}\mathcal{G}_{{\alpha}{\dot{\alpha}}}-{\bar{\rm D}}^2{\rm D}_{{\alpha}}\mathcal{E}^1-{\rm D}_{{\alpha}}{\bar{\rm D}}^2\mathcal{E}^2=0
\end{equation}
where $\mathcal{G}_{{\alpha}{\dot{\alpha}}},~\mathcal{E}^1,~\mathcal{E}^2$ are the variations of the full action
with respect the superfields $H_{{\alpha}{\dot{\alpha}}},~U,~{\sigma}$
\begin{equation}
\eqalign{
\mathcal{G}_{{\alpha}{\dot{\alpha}}}=~&4c_3{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}{\dot{\alpha}}}+2c_3\pa_{{\alpha}{\dot{\alpha}}}\pa^{{\gamma}{\dot{\gamma}}}H_{{\gamma}{\dot{\gamma}}}\cr
+&2c_4[{\rm D}_{{\alpha}},{\bar{\rm D}}_{{\dot{\alpha}}}][{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}]H_{{\gamma}{\dot{\gamma}}}-(-2c_3+6c_4)\[{\rm D}_{{\alpha}},{\bar{\rm D}}_{{\dot{\alpha}}}\]U\cr
-&2(c_3-c_4){\bar{\rm D}}_{{\dot{\alpha}}}{\rm D}_{{\alpha}}\sigma+(c_3-c_4){\rm D}_{{\alpha}}{\bar{\rm D}}_{{\dot{\alpha}}}\bar{{\sigma}}\cr
{}\cr
\mathcal{E}^1=~&(2c_3-6c_4)\[{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}\]H_{{\gamma}{\dot{\gamma}}}+2bU\cr
+&g{\sigma}+g\bar{{\sigma}}\cr
{}\cr
\mathcal{E}^2=~&2(c_3-c_4){\rm D}^{{\gamma}}{\bar{\rm D}}^{{\dot{\gamma}}}H_{{\gamma}{\dot{\gamma}}}+e\bar{{\sigma}}+2f{\sigma}\s+gU
}
\end{equation}
The solution of the Bianchi identity is
\begin{center}
\begin{tabular}{c c}
\begin{minipage}{1.9in}
\begin{itemize}
\item ${\sigma}$ is chiral
\end{itemize}
\end{minipage}
&
\begin{minipage}{1.9in
\begin{itemize}
\item $c_3=c_4$
\end{itemize}
\end{minipage}
\\
\begin{minipage}{1.9in
\begin{itemize}
\item $e=0$
\end{itemize}
\end{minipage}
&
\begin{minipage}{1.9in
\begin{itemize}
\item $b=6c_4$
\end{itemize}
\end{minipage}
\\
\begin{minipage}{1.9in
\begin{itemize}
\item $g=4c_4$
\end{itemize}
\end{minipage}
\end{tabular}
\end{center}
Hence the action takes the form
\begin{equation}
\eqalign{
S=\int d^8z\Bigg\{
~~&2c_4 H^{{\alpha}{\dot{\alpha}}}{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}{\dot{\alpha}}}\cr
&+c_4 H^{{\alpha}{\dot{\alpha}}}\pa_{{\alpha}{\dot{\alpha}}}\pa^{{\gamma}{\dot{\gamma}}}H_{{\gamma}{\dot{\gamma}}}\cr
&+c_4 H^{{\alpha}{\dot{\alpha}}}[{\rm D}_{{\alpha}},{\bar{\rm D}}_{{\dot{\alpha}}}][{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}]H_{{\gamma}{\dot{\gamma}}}\cr
&-4c_4H^{{\alpha}{\dot{\alpha}}}\[{\rm D}_{{\alpha}},{\bar{\rm D}}_{{\dot{\alpha}}}\]U\cr
&+6c_4U^2\cr
&+4c_4U({\sigma}+\bar{{\sigma}})\Bigg\}
}
\end{equation}
which is invariant under the gauge transformations:
\begin{equation}
\eqalign{
& \delta H_{{\alpha}{\dot{\alpha}}}={\bar{\rm D}}_{{\dot{\alpha}}}L_{{\alpha}}-{\rm D}_{{\alpha}}\bar{L}_{{\dot{\alpha}}}\cr
&\delta U={\rm D}^{{\gamma}}{\bar{\rm D}}^2L_{{\gamma}}+{\bar{\rm D}}^{{\dot{\gamma}}}{\rm D}^2\bar{L}_{{\dot{\gamma}}}\cr
&\delta {\sigma}={\bar{\rm D}}^2{\rm D}^{{\gamma}}L_{{\gamma}}
}
\end{equation}
The equation of motion for the superfield $U$ is
\begin{equation}
\eqalign{
\mathcal{E}^1=0 \Rightarrow U=\frac{1}{3}\[{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}\]H_{{\gamma}{\dot{\gamma}}}-\frac{1}{3}\left({\sigma}+\bar{{\sigma}}\right)
}
\end{equation}
and the action becomes:
\begin{equation}
\eqalign{
S_1=\int d^8z\Bigg\{
~~&2c_4 H^{{\alpha}{\dot{\alpha}}}{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}{\dot{\alpha}}}\cr
&+c_4 H^{{\alpha}{\dot{\alpha}}}\pa_{{\alpha}{\dot{\alpha}}}\pa^{{\gamma}{\dot{\gamma}}}H_{{\gamma}{\dot{\gamma}}}\cr
&+\frac{1}{3}c_4 H^{{\alpha}{\dot{\alpha}}}[{\rm D}_{{\alpha}},{\bar{\rm D}}_{{\dot{\alpha}}}][{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}]H_{{\gamma}{\dot{\gamma}}}\cr
&+i\frac{4}{3}c_4H^{{\alpha}{\dot{\alpha}}}\pa_{{\alpha}{\dot{\alpha}}}\left(\bar{{\sigma}}-{\sigma}\right)\cr
&-\frac{4}{3}c_4{\sigma}\bar{{\sigma}}\Bigg\}
}
\end{equation}
which is invariant under the transformations:
\begin{equation}
\eqalign{
& \delta H_{{\alpha}{\dot{\alpha}}}={\bar{\rm D}}_{{\dot{\alpha}}}L_{{\alpha}}-{\rm D}_{{\alpha}}\bar{L}_{{\dot{\alpha}}}\cr
&\delta {\sigma}={\bar{\rm D}}^2{\rm D}^{{\gamma}}L_{{\gamma}}
}
\end{equation}
The action $S_1$, up to redefinitions, is the minimal supergravity formulation.
Instead of using the equation of motion for the superfield $U$, we can use the equation of
motion for the superfield ${\sigma}$ then we get:
\begin{equation}
\mathcal{E}^2=0 \Rightarrow {\bar{\rm D}}^2U=0
\end{equation}
therefore $U$ is now a linear compensator and the action is
\begin{equation}
\eqalign{
S_2=\int d^8z\Bigg\{
~~&2c_4 H^{{\alpha}{\dot{\alpha}}}{\rm D}^{{\gamma}}{\bar{\rm D}}^2{\rm D}_{{\gamma}}H_{{\alpha}{\dot{\alpha}}}\cr
&+c_4 H^{{\alpha}{\dot{\alpha}}}\pa_{{\alpha}{\dot{\alpha}}}\pa^{{\gamma}{\dot{\gamma}}}H_{{\gamma}{\dot{\gamma}}}\cr
&+c_4 H^{{\alpha}{\dot{\alpha}}}[{\rm D}_{{\alpha}},{\bar{\rm D}}_{{\dot{\alpha}}}][{\rm D}^{{\gamma}},{\bar{\rm D}}^{{\dot{\gamma}}}]H_{{\gamma}{\dot{\gamma}}}\cr
&-4c_4H^{{\alpha}{\dot{\alpha}}}\[{\rm D}_{{\alpha}},{\bar{\rm D}}_{{\dot{\alpha}}}\]U\cr
&+6c_4U^2\Bigg\}
}
\end{equation}
which is invariant under the transformations
\begin{equation}
\eqalign{
& \delta H_{{\alpha}{\dot{\alpha}}}={\bar{\rm D}}_{{\dot{\alpha}}}L_{{\alpha}}-{\rm D}_{{\alpha}}\bar{L}_{{\dot{\alpha}}}\cr
&\delta U={\rm D}^{{\gamma}}{\bar{\rm D}}^2L_{{\gamma}}+{\bar{\rm D}}^{{\dot{\gamma}}}{\rm D}^2\bar{L}_{{\dot{\gamma}}}
}
\end{equation}
This action is the new-minimal supergravity formulation
|
\section{Introduction}
\label{sec:intro}
The on-going work by the United Nations, following the major report of
the \citet{IPCC2007} in 2007 and the Climate Change Conferences in
Copenhagen and Cancun (Mexico), centres on the prediction of future
climate change, a key feature of which would be any suggestion of a
sudden and (possibly) irreversible abrupt change called a tipping
point (\citet{Lenton2008,Scheffer2009a}). Many tipping points, such as
the switching on and off of ice-ages, are well documented in
paleo-climate studies over millions of years of the Earth's history.
Predicting such tippings in advance using time-series data derived
from preceding behaviour is now seen as a major challenge, impinging,
for example on the possible use of geo-engineering
(\citet{Launder2010}). Techniques introduced by \citet{Held2004} and
\citet{Livina2007} draw on the assumption that the tipping events are
governed by a bifurcation in an underlying nonlinear dissipative
dynamical system. Specifically, researchers search for a slowing down
of intrinsic transient responses within the data, which is predicted
to occur before most bifurcational instabilities. This is done by
determining a so-called propagator, estimated from the correlation
between successive elements of a window sliding along the time series
\jsmod{(this estimate is called $\kappa_\mathrm{ACF}$ in this
paper)}. This propagator, such as the AR(1) mapping coefficient, is
a measure of the linear decay rate and should increase to unity at a
tipping instability (corresponding to a decrease of the linear decay
rate to zero).
Prediction techniques can be tested on climatic computer models, but
more challenging is to try to predict real ancient climate tippings,
using their preceding geological data provided by ice cores,
sediments, etc. Using this preceding data alone, the aim would be to
see to what extent the actual tipping could have been predicted in
advance.
One such study by \citet{Livina2007} looks at the end of the most
recent ice age and the associated Younger Dryas event, about $11,500$
years ago, when the Arctic warmed by $7\celsius$ in $50$ years. This
pioneering study used a time series derived from Greenland ice-core
paleo-temperature data. A second such study (one of eight made by
\citet{Dakos2008}), using data from tropical Pacific sediments, gives
a good prediction for the end of `greenhouse' Earth about $34$ million
years ago when the climate tipped from a tropical state into an
icehouse state. [Lenton \emph{et al.} this volume]\ gives a complete overview of
current techniques for extracting early-warning signals from time
series and compares them on realistic models and a range of
paleo-climate time series.
A \jsmod{remark} that is made during the analysis of [Lenton \emph{et al.} this volume]\ is
that the absolute values of the extracted quantities have no direct
bearing on the tipping probability over time unless one knows also the
size of the \emph{basin of attraction}. This basin is determined in
the simplest cases by the dominant nonlinear term in the underlying
equations of motion, which is the subject of this paper.
\section{Concepts from Nonlinear Dynamics}
\label{sec:concepts}
The core techniques for analysis of time series of climate data aim to
extract the linear decay rate toward (a possibly drifting)
noise-perturbed equilibrium \citep{Held2004,Livina2007}. These
techniques can be directed at the Earth's climate as a whole, or at
the relevant climate sub-systems described by \citet{Lenton2008} as
\emph{tipping elements}. Our aim here is to augment this linear information
with information about nonlinear features of the underlying dynamics,
by examining to what extent we can identify nonlinear softening and
include it into the list of early-warning signs for climate tipping.
\subsection{Shrinking basins around dangerous bifurcations}
\label{sec:basins}
As we approach a dangerous bifurcation, from which a nonlinear
dissipative dynamical system will experience a finite jump to a remote
alternative state, the current attractor is located within a shrinking
basin of attraction. Correspondingly, the maintenance of the state
will become increasingly precarious in the presence of noisy
disturbances.
Thinking in terms of an underlying potential energy function (the
existence of which is theoretically well-defined for a saddle-node
fold and some other bifurcations) the parabolic shape of its graph
will become increasingly perturbed by softening nonlinear features.
The relevant bifurcations are the so-called \emph{codimension-one}
events (\citet{Thompson1994,Thompson2002}), namely those that can be
typically encountered under a gradual change of a single control
parameter. The complete list of the dangerous codimension-one
bifurcations is given in Table~\ref{tab:bifs}, where we give a brief
description of the nature of the shrinking basin. More details can be
found in \citet{Thompson2010}.
\begin{table}
\centering
\begin{tabbing}
\textbf{(a) L}\=\textbf{ocal Saddle-Node Bifurcations}\hspace*{10em}\=\\
\> Static Fold (saddle-node of fixed point) \> one-sided basin shrinkage\\
\>Cyclic Fold (saddle-node of cycle)\>one-sided basin shrinkage\\
\textbf{(b) Local Subcritical Bifurcations}\\
\>Subcritical Hopf\>complete basin shrinkage\\
\> Subcritical Neimark-Sacker (secondary Hopf)\>complete basin shrinkage\\
\>Subcritical Flip (period-doubling)\>complete basin shrinkage \\
\textbf{(c) Global Bifurcations}\\
\>Saddle Connection (homoclinic connection)\>
outside shrinkage around cycle\\
\> Regular-Saddle Catastrophe (boundary crisis)\>
fractal basin shrinkage\\
\>Chaotic-Saddle Catastrophe (boundary crisis)\>
fractal basin shrinkage
\end{tabbing}
\caption{Dangerous Bifurcations and the behaviour of the basin of attraction}
\label{tab:bifs}
\end{table}
Focusing first on the local bifurcations (headings (a) and (b) of
Table~\ref{tab:bifs}) the shrinking boundaries are illustrated
schematically in Figure~\ref{fig:c360}, with the control parameter plotted
horizontally and the response plotted vertically.
\begin{figure}[th]
\centering
\includegraphics[scale=0.3]{c360_Dangerous_local_bifs_FFCL_BW}
\caption{Basin boundary transformations at dangerous
bifurcations. (a) the fold bifurcations and (b) the subcritical
bifurcations. Solid curves denote stable paths while broken curves
denote unstable paths. The basin is shown in grey.}
\label{fig:c360}
\end{figure}
In the first picture, Figure~\ref{fig:c360}(a), we illustrate both
types of the saddle-node fold bifurcation: the static fold involving
a path of equilibrium fixed-points, and the cyclic fold involving a
trace of periodic orbits. For these the basin shrinkage is
one-sided. The basin is bounded by the unstable side of the
parabolically folding path.
Next, in Figure~\ref{fig:c360}(b), we illustrate the local subcritical
bifurcations of which the Hopf, flip and Neimark are codimension-one
(we remember that the more familiar static pitchfork bifurcations are
codimension-two, being structurally unstable against a
symmetry-breaking perturbation). Here we see an unstable solution,
which controls the basin boundary, shrinking parabolically around the
stable solution, from which the (noise-free) system jumps out of our
field of view as the control parameter, $\mu$, reaches its critical value
of $\mu_C$.
The static fold and the Hopf bifurcation have a theoretically
well-defined potential energy surface that neatly summarizes the
shrinking basin as illustrated in Figure~\ref{fig:c361}. Note that
Figure~\ref{fig:c361}(a) will be discussed more fully in
section~\ref{sec:concepts}(\ref{sec:fold}).
\begin{figure}[th]
\centering
\includegraphics[scale=0.15]{c361_Fold_Hopf_Ux_FFCL_BW}
\caption{Total potential energy transformations in (a) the
saddle-node fold and (b) the Hopf bifurcation (via
averaging). Black balls denote stable equilibrium states while
white balls denote unstable states.}
\label{fig:c361}
\end{figure}
\jsrem{For other dangerous bifurcations, where the existence of an underlying
potential surface lacks any theoretical backing (and indeed may be
technically impossible because of fractal features in the dynamics)
there are nevertheless occasions when it is useful and practical to
consider such a surface for escape predictions in the presence of
significant noise as we shall demonstrate in sectio
}
\subsection{A closer look at the fold}
\label{sec:fold}
The fold is the simplest and most common way in which an equilibrium
of a nonlinear dissipative dynamical system can lose its
stability. Having only a single active degree of freedom ($x$) and
being observable under the variation of a single control parameter
($\mu$) it can be illustrated on a graph of $x$ against $\mu$ as a
smooth curve that simply reaches an extreme value (a maximum, say) of
the control $\mu$, as shown in Figure~\ref{fig:c360}(a). So although
it is traditionally called a saddle-node bifurcation, it does not
exhibit any obvious `bifurcation' of a path, but rather just a smooth
folding of an existing path. In the case of a static fold (to which we
shall largely restrict our attention) the path is a trace of
equilibrium fixed points, while for a cyclic fold the path would be a
trace of periodic solutions.
Assuming that we are close to a fold, the intrinsic damping of the
system will have become super-critical so the system will have the
non-oscillatory response of an over-damped particle sliding (in thick
oil, as we might imagine) on a notional potential energy surface
$U(x,\mu)$. This is illustrated for the fold in
Figure~\ref{fig:c361}(a), where we show the potential energy surface
erected over the $(x,\mu)$ base plane. Here the solid part of the
equilibrium path, corresponding to energy minima, is a stable node,
while the dashed part, corresponding to energy maxima (more generally
a geometrical saddle in higher dimensions) is an dynamically unstable
saddle.
We see that as the control $\mu$ is slowly increased towards its
critical value, $\mu_C$, the stable equilibrium solution gets closer
and closer to the hill-top potential barrier, and so gets
progressively more precarious in the presence of noisy disturbances.
With $\mu$ increasing at an infinitesimal rate, noise will ensure
escape over the hill-top before $\mu$ reaches $\mu_C$. If, however,
$\mu$ increases rapidly, this early escape may be delayed or supressed
as we have examined quantitatively elsewhere (\citet{Thompson2011}).
\begin{figure}[th]
\centering
\includegraphics[scale=0.3]{c357_Climate_cusp_FFCL_BW}
\caption{The cusp catastrophe and its associated pattern of
folds. The notional axes show how such a cusp might arise in a
climate tipping scenario.}
\label{fig:c357}
\end{figure}
To give a greater perspective to our view of the fold it is useful to
introduce another control parameter $\mu_2$ such that our parameter
space is now represented jointly by the $(\mu_1,\mu_2)$-plane. We can
now erect an equilibrium surface, of height $x$, over the
two-dimensional control plane as illustrated in
Figure~\ref{fig:c357}. Here we see two fold lines whose projections
divide the control plane into regimes exhibiting either one or three
co-existing solutions. These fold lines coalesce and vanish at the
cusp. This cusp is a codimension-two phenomenon, which is typically
only observed when we have independent control of both parameters
($\mu_1$ and $\mu_2$). This is made transparent by the fact that a
typical path in control space cannot be expected to pass precisely
under the cusp point; by contrast, if a parameter path passes through
a fold line all near-by paths also cross this fold line, which is what
`the fold has codimension-one' means. For an appropriate choice of
path in the ($\mu_1,\mu_2)$-plane through the cusp one sees a
pitchfork bifurcation (\citet{Thompson2002}).
We will consider the case in which the main equilibrium sheet (in
grey) is stable while the narrow inverted sheet (white) is
unstable.
A climate system will have a large number of possible control
parameters, but in trying to predict a tipping point we will be
studying a particular evolution which corresponds to a particular
route in control space; this is precisely why we are only likely to
encounter a fold, and not a cusp in Figure~\ref{fig:c357}. To
explore possible scenarios that might underlie a tipping study, it is
now useful to examine some possible routes in our two-parameter model
of Figure~\ref{fig:c357}.
Consider, first, the simple scenario in which the system parameters
take path $\overline{BB}$ leading transversally through a fold
line. This is the classical encounter with a fold that is implicitly
envisaged in earlier work. In the presence of noise a premature tip
from N occurs with a certain probability depending on the noise level
and the speed with which the path is traversed, as analysed by
\citet{Thompson2010}. If the noise level were particularly high, a
jump from N could easily be perceived as a purely noise-induced jump
with no bifurcation, even though it is actually caused by the
proximity to the fold: this could hold even if there were no movement
in the control space at all, the system having rested at N throughout
([Ashwin \emph{et al.} this volume]\ call this N-tipping). Another scenario, not
specifically illustrated in the figure, could arise if the route in
control space approached the fold line, but then turned back away from
it. Analysis would then show a temporary decrease of the strength of
the attraction (with the AR(1) coefficient moving towards unity) which
might be discounted as a false alarm, even though the system did
approach a fold and a noise-induced escape was temporarily probable.
Finally, in the scenario of path $\overline{AA}$, one would observe a
temporary increase of the AR(1) coefficient, even though no
bifurcation is apparent but rather a gradual shift of the steady
state. This shows that a rapid transition can be related to a nearby
fold that was just barely missed. Similarly, the scenario called
R-tipping by [Ashwin \emph{et al.} this volume]\ occurs if one changes the other
control parameter in Figure~\ref{fig:c357} sufficiently rapidly from
the path $\overline{AA}$ to N (avoiding the folds). \jsmod{In this
scenario the system jumps to the other stable equilibrium despite
having never crossed a bifurcation curve.}
\section{Time series analysis for leading order terms --- qualitative
overview}
\label{sec:timeseries}
The extraction of the underlying dynamics from time series is divided
into subtasks of increasing difficulty. Assume that we have a time
series coming from a process that is either stationary or has a slowly
drifting underlying parameter $\mu$ as suggested by the paths in
Figure~\ref{fig:c357}. In order to predict (or identify) the dynamics
one attempts to extract the leading orders of the dominant components
of the underlying equations of motion.
\jsmod{
\begin{figure}[tbh]
\centering
\includegraphics[width=0.9\textwidth]{snfdata_rev-10}
\caption{Comparison of time series generated by linear (left
column, (a-1) to (a-4)) and nonlinear process (right column,
(b-1) to (b-4)). Both time series are stationary and have
identical linear decay rates. As measures for nonlinearity we
used skewness and the coefficient $c_\mathrm{emp}$ (as defined
below in equation~\eqref{eq:fp1}). Parameters: $\mu(0)=2$,
$N=1\,223$, $\Delta t=0.1$, $\sigma=1$, window length for linear
and nonlinear analysis, $w=N/2$, $\epsilon=0.01$.}
\label{fig:snfdata}
\end{figure}
} We illustrate the steps using \jsmod{a time series} obtained from a
process which has \jsmod{the saddle-node normal form with a slowly
drifting normal form parameter $\mu$} as its deterministic part
(corresponding to Figure~\ref{fig:c361}(a)) and is influenced by
additive Gaussian noise. We permit the parameter $\mu$ to drift slowly
with speed $\epsilon$:
\begin{equation}
\label{eq:snfnoise}
\begin{split}
\d x&=\left[\mu-x^2\right]\d t +\sigma \d W_t \mbox{, where}\\
\d\mu& =-\epsilon \d t\mbox{.}
\end{split}
\end{equation}
Here $x$ is the response, $\mu$ is the control parameter, and $t$ is
the time. The Gaussian perturbation $\d W_t$ has zero mean and unit
variance such that $\sigma$ controls the noise induced variance added
to the dynamics of the deterministic saddle-node form $\dot
x=\mu-x^2$, which has the stable equilibrium
$x_\mathrm{eq}=\sqrt{\smash[b]{\mu}}$. The second equation prescribes
the linear sweep of $\mu$ from its starting value $\mu_0>0$ towards
zero at rate $\epsilon\ll1$. For fixed $\mu>0$ the unperturbed system
($\sigma=0$) has a stable equilibrium at
$x=x_\mathrm{eq}=\sqrt{\smash[b]{\mu}}$ and an unstable equilibrium at
$x=-\sqrt{\smash[b]{\mu}}$.
In Section~\ref{sec:quant} we will use the saddle-node normal form
\eqref{eq:snfnoise} to \jsmod{test the ability of a range of estimators to
distinguish a time series generated by a process with a quadratic
nonlinearity from a linear time series. Two example time series
and their analysis are shown in Figure~\ref{fig:snfdata}. Both time
series have identical linear properties but differ in the dominant
nonlinear term of the deterministic part of the dynamics.The
softening of the full (estimated) nonlinear potential well
(corresponding to the illustration in Figure~\ref{fig:c361}(a)) is
displayed in Figure~\ref{fig:c365}. }
\jsrem{The left column (a) of Figure~\ref{fig:snfdata} shows a linear
time series generated by \eqref{eq:snfnoise} with densely taken
measurements (length of time series $N=28,803$, time step $\Delta
t=0.1$, drift speed $\epsilon=0.01$ for system parameter $\mu$). The
right column shows the results for a short time series (also
generated by \eqref{eq:snfnoise}) with sparser measurements.}
\subsection{Equilibrium --- zero order}
\label{sec:0order}
\noindent The location of the slowly drifting equilibrium and its
trend is estimated first. [Lenton \emph{et al.} this volume]\ call this step detrending
and propose either piecewise linear fitting or filtering with a
Gaussian kernel. The result of filtering with a Gaussian kernel is
shown as a thick grey curve in the row 1 of
Figure~\ref{fig:snfdata}. Every method for detrending requires a
bandwidth parameter (for example the width of the Gaussian
kernel). For Figure~\ref{fig:snfdata} we chose the bandwidth which the
kernel density estimation routine (Matlab's \texttt{kde} by
\citet{Botev2010}) offered. The thin black curve in
\jsmod{Figure~\ref{fig:snfdata}(b-1)} shows the true value of the
quasi-equilibrium (for $\epsilon=0$). It gives a visual estimate of how the
reliability of the zero-order estimates depends on the quality of the
data. The oscillatory deviations of the extracted equilibrium (thick
light grey curve) from its true value (thin black curve) shows that
the automatically chosen bandwidth was slightly smaller than the
theoretical optimum.
\subsection{Linear decay rate --- first order}
\label{sec:1order}
\noindent Assuming that the deterministic part of the underlying
process has a stable equilibrium $x_\mathrm{eq}$ (which is possibly
slowly drifting), and that the zero-order estimate is an approximation
for $x_\mathrm{eq}$, the next step is to estimate the linear decay
rate $\kappa$ toward $x_\mathrm{eq}$ and the trend of $\kappa$ over
time. Generally, if the underlying deterministic process is
high-dimensional then these estimates capture the dominant (that is,
smallest) decay rate. They are typically applied to the detrended time
series, that is, $\tilde x=x-x_\mathrm{eq}$, which is shown in the
second row of Figure~\ref{fig:snfdata}.
\begin{itemize}
\item \textbf{ACF} The $k$-step (usually one-step) autocorrelation
function (ACF) $\alpha$ is fitted directly to the detrended time
series: $\tilde x_{k+1}=\alpha \tilde x_k$. This was introduced by
\citet{Held2004} as degenerate finger-printing for the analysis of
climate time series. The ACF $\alpha$ is related to $\kappa$ via
\begin{displaymath}
\alpha=\exp(-\kappa_\mathrm{ACF}\,\Delta t)\mbox{.}
\end{displaymath}
The estimate for $\kappa_\mathrm{ACF}$ is shown in
Figure~\ref{fig:snfdata}(a-3) and (b-3).
\item \textbf{DFA} Detrended fluctuation analysis has been introduced
by \citet{Livina2007} for climate time series, see also
[Lenton \emph{et al.} this volume]\ for a short description of how one extracts $\kappa$
from this analysis.
\item \textbf{Quasi-stationary density} If one assumes that the
deterministic dynamics of the detrended time series $\tilde x$ is
essentially that of an overdamped particle moving in a potential
well $U(\tilde x)$, that is,
\begin{equation}
\label{eq:potwell}
\d \tilde x=-\partial_xU(\tilde x)\d t+\sigma \d W_t\mbox{,}
\end{equation}
then the linear decay rate equals the second derivative
$\partial_{xx}U(0)$ of $U$ at the bottom of the well, $\tilde
x=0$. Moreover, the stationary density $p(x)$ is related to the
shape of the potential well $U$ via the stationary Fokker-Planck equation
\begin{align}
\frac{1}{2}\partial_{xx} p(\tilde x)&=
\partial_x\left[-\sigma^{-2}\partial_xU(\tilde x)\,p(\tilde
x)\right]\mbox{, which gives, integrated once,}\nonumber\\
\label{eq:fp}
\frac{1}{2}\partial_x p(\tilde x)&=
-\sigma^{-2}\partial_xU(\tilde x)\,p(\tilde x)+\sigma^{-2}c\mbox{,}
\end{align}
where $c$ is the constant of integration. This constant satisfies
$c=\lim_{\tilde x\to\pm\infty}\partial_xU(\tilde x)p(\tilde x)$, and
can be interpreted as the flow rate ($c<0$ indicates flow toward
$-\infty$). Remembering that the drift speed $\epsilon$ of the
system parameter $\mu$ is small, the true time-dependent probability
density $p$ still satisfies \eqref{eq:fp} approximately at each
$\mu$:
\begin{equation}
\label{eq:fpdrift}
\frac{1}{2}\partial_x p(\tilde x,\mu)=
-\sigma^{-2}\partial_xU(\tilde x,\mu)\,p(\tilde x,\mu)+
\sigma^{-2}c(\mu)+O(\epsilon)\mbox{.}
\end{equation}
In \eqref{eq:fpdrift} we included the dependence of all quantities
on $\mu$ expressly and remember that $\dot \mu(t)$ is of order
$\epsilon$. Assuming quasi-stationarity, we neglect the
order-$\epsilon$ terms and determine the coefficients,
$-\sigma^{-2}\partial_xU$ and $\sigma^{-2}c$, by fitting the
empirical density $p_\mathrm{emp}(x)$ from a time window
$[t-w/2,t+w/2]$ to the Equation~\eqref{eq:fpdrift}. Specifically,
$c_\mathrm{emp}=\sigma^{-2}c$ is a scalar and $\partial_xU(0)$ is
equal to $0$ after detrending such that one has to fit two
coefficients, $\kappa_U$ and $c_\mathrm{emp}$ using
\eqref{eq:fpdrift} if one truncates $\partial_xU$ after first order:
\begin{equation}
\label{eq:fp1}
\frac{1}{2}\partial_x p_\mathrm{emp}(\tilde x)=
-\kappa_U \tilde x\,p_\mathrm{emp}(\tilde x)+c_\mathrm{emp}\mbox{.}
\end{equation}
In problems where no escape is possible one can drop the term $c$ in
\eqref{eq:fp} (and, thus, in \eqref{eq:fpdrift} and \eqref{eq:fp1}),
and simplify \eqref{eq:fp} to
\begin{equation}\label{eq:logu}
U=-\frac{\sigma^2}{2}\log p\mbox{.}
\end{equation}
\citet{Livina2011} used relation \eqref{eq:logu} (fitting of $U$
with higher-order polynomials) to detect potential wells in time
series that included rapid transitions. Fitting $U$ with a parabola
(where $U(0)$ is irrelevant and $\partial_xU(0)=0$) is equivalent to
using the variance $\sigma^2_\mathrm{emp}$ of $p_\mathrm{emp}(\tilde
x)$ if the density $p_\mathrm{emp}$ is
Gaussian. \citet{Ditlevsen2010} state that monitoring the empirical
variance $\sigma^2_\mathrm{emp}$ helps to avoid erroneous detection
of false trends.
\end{itemize}
[Lenton \emph{et al.} this volume]\ compare these three estimates,
\jsmod{$\kappa_\mathrm{ACF}$, the DFA-based estimate for $\kappa$, and
the variance $\sigma^2_\mathrm{emp}$ (which is inversely
proportional to $\kappa_U$),} in detail using time series arising in
climate models and in palaeoclimate records. Methods based on
properties of the spectrum of the time series were proposed and
investigated by \citet{Kleinen2003,Biggs2009} but are not discussed
here. \jsmod{Figure~\ref{fig:snfdata} shows in row~3 how the
estimates of $\kappa_U$ from the quasi-stationary density compare to
the ACF estimates $\kappa_\mathrm{ACF}$. For both time series the
estimates are quantitatively close to the theoretical value of
$\kappa$ (which is indicated by the thin black line in
Figure~\ref{fig:snfdata}(a-3,b-3)).}
\jsrem{Figure~\ref{fig:snfdata}(b-3) shows that even for sparse and
short time series the estimates for the linear decay rate are
qualitatively correct but that discerning trends such as in
Figure~\ref{fig:snfdata}(a-3) is possible only for longer time
series.} \jsmod{We also observe that in
Figure~\ref{fig:snfdata}(a-3,b-3) the local trends of
$\kappa_\mathrm{ACF}$ from the autocorrelation estimate and of
$\kappa_U$ based on the quasi-stationary density are strongly
correlated, so it is unclear if monitoring both quantities really
prevents false alarms.} \jsrem{As the underlying decay rate has no
downward trend one might call the local downward trend of the
estimates a `false alarm'.} \jsrem{However, the parameters of the
system make a noise-induced escape (or N-tipping in the notation of
\mbox{[Ashwin \emph{et al.} this volume]\ }) likely. Escape occurs indeed at time
$t=40$ in Figure~\ref{fig:snfdata}(b-1).}
There is one notable difference between the AR(1) and the DFA estimate
on one hand and the distribution-based estimate $\kappa_U$ on the
other hand: the estimate for $\kappa_U$ based on the quasi-stationary
density only looks at the distribution of the values $\tilde x_k$ in
the window of interest but does not care about the order in which they
appear. This is in contrast to, for example, the AR(1) estimate which
determines the correlation between each value $\tilde x_k$ and its
predecessor.
\subsubsection*{Estimate of the input noise amplitude}
\noindent If the AR analysis shows the presence of a single positive
dominant coefficient (which gives evidence that the underlying
dynamics has really one distinct direction in which the decay is
slowest) then one can combine both estimates of the decay rate to
obtain an estimate of the amplitude $\sigma$ of the additive
noise. \jsmod{Equations \eqref{eq:fp} and \eqref{eq:fp1} imply that
$\kappa_U$ is an estimate for $\sigma^{-2}\kappa$ where $\kappa$ and
$\sigma^2$ are the true linear decay rate and input noise
amplitude. Thus, if we replace the unknown $\kappa$ by the estimate
$\kappa_\mathrm{ACF}$ we can use the relation between $\kappa_U$ and
the true $\kappa$ to estimate $\sigma^2$:}
\begin{equation}
\sigma_\mathrm{emp}^2=\kappa_\mathrm{ACF}/\kappa_U\label{eq:sigmaemp}
\end{equation}
where $\kappa_\mathrm{ACF}$ is the estimate of the linear decay rate
obtained from the autocorrelation and $\kappa_U$is the estimate
obtained from the quasi-stationary density $p_\mathrm{emp}$. \jsmod{This is
an alternative to using the residual of the least-squares fit that one
obtains when estimating $\kappa_\mathrm{ACF}$. We found that the
residual systematically underestimates the noise level in the normal
form examples shown in Fig.~\ref{fig:snfdata}.}
\subsection{Nonlinearity --- basin boundary}
\label{sec:nonlin}
\noindent One problem left open by the linear analysis is that the
quantitative value of the estimated linear decay rate $\kappa$ and the
estimated noise-level do not give any indication for the probability
of tipping. What the linear methods estimate is the $\alpha$ and the
$\sigma$ of a discrete process
\begin{equation}\label{eq:lindisc}
\tilde x_{k+1}=\alpha \tilde x_k+\sigma\eta_k
\end{equation}
(where the $\eta_k$ are assumed to be independent and normally
distributed random numbers). The discrete process \eqref{eq:lindisc}
is the linearisation of the time-$\Delta t$ map of the continuous
process \eqref{eq:potwell}. An estimate for $\alpha$ less than but
close to unity does not necessarily indicate that we are close to a
stability boundary for the nonlinear problem \eqref{eq:potwell}.
Rather, it implies that the time step $\Delta t$ between successive
measurements has been small compared to the mean decay time to half,
$\log 2/\kappa$, of the process:
\begin{displaymath}
\alpha(t_k)=\exp\left(-\kappa(t_k)\Delta t\right)=
1-\kappa(t_k)\Delta t+O\left((\Delta t)^2\right)\mbox{.}
\end{displaymath}
The \emph{trend} of the estimated $\alpha$ is an indicator for
incipient tipping because (if extrapolated in any way) it gives an
estimate for the time to tipping if one ignores the possibility of
early escape. However, the trend of $\alpha$ is less certain even in
the artificial time series of Figure~\ref{fig:snfdata}. Similarly, the
estimated noise amplitude $\sigma$ has to be compared to the
coefficient in front of the leading nonlinear term of the right-hand
side (which equals $-1$ in \eqref{eq:snfnoise}). The ratio between
$\sigma$ and nonlinear term enters the estimates for the probability
of tipping if there is no discernible upward trend in $\alpha$
(N-tipping), or for the probability of early escape if the trend of
$\alpha$ points upwards (see \citet{Thompson2011}). So without at
least an order-of-magnitude estimate of the leading nonlinear term the
linear methods only provide an estimate for the tipping time if there
is a clearly discernible trend in $\alpha$, and even then this
estimate discounts early escape.
\jsmod{Figure~\ref{fig:snfdata} illustrates this problem. Apart from a
slow trend both time series in Figure~\ref{fig:snfdata} have
identical properties at the linear level. However, while for the
time series in Figure~\ref{fig:snfdata}(b-1) the median time to
tipping is $t=100$ according to \citet{Thompson2011} (indeed, it
escapes shortly after $t=120$), for the time series in
Figure~\ref{fig:snfdata}(a-1) the extrapolation (correctly) predicts
tipping (or rather, linear instability) at $t\approx320$. Thus, we
observe that the linear properties of the time series alone are not
sufficient to provide estimates about tipping, for example, the
median time to tipping. The difference between the time series in
Figure~\ref{fig:snfdata}(a-1) and Figure~\ref{fig:snfdata}(b-1) is
in the dominant nonlinear term of the deterministic part. That this
term plays a role is not surprising because tipping is a feature of
nonlinear systems. Extracting its presence from a time series is a
necessary step that has to follow the linear analysis.} \jsmod{
\begin{figure}[th]
\centering
\includegraphics[width=0.6\textwidth]{c365_Fold_FFCL_PUB_rev}
\caption{Nonlinear potential energy surfaces extracted from noisy
time series using sliding windows ($w=N/2$) for the saddle node
fold \protect\jsmod{($U_\mathrm{emp}=
-\sigma_\mathrm{emp}^2\log(p_\mathrm{emp})/2$ where
$\sigma_\mathrm{emp}$ was estimated using \eqref{eq:sigmaemp})}.
At either end of the surface we have plotted the known function
$U(x,t)=x^3/3-\mu(t)x$ ($\sigma=1$, $\epsilon=0.01$ and
$\mu(0)=2$). Red denotes a positive deviation, blue a negative
deviation from the parabola given by linear theory.}
\label{fig:c365}
\end{figure}}
The approaches presented in the following all generalise the estimate
for $\kappa_U$, based on the quasi-stationary density, by looking at
the full nonlinear potential well $U_\mathrm{emp}(\tilde x)$
associated to the quasi-stationary density $p_\mathrm{emp}(\tilde x)$
via the Fokker-Planck
equation~\eqref{eq:fpdrift}. Figure~\ref{fig:c365} illustrates how
$U_\mathrm{emp}=-\sigma^2\log(p_\mathrm{emp})/2$ looks like for the
time series shown in Figure~\ref{fig:snfdata} column (b). The
realisation of the noisy, evolving time series is shown in the base
plane together with the equilibrium path $\mu=x^2$, and the estimated
(empirical) potential, $U_\mathrm{emp}$ is shown as the coloured
surface. The estimate for the potential well is taken in a sliding
window, which explains why the surface can only be determined in the
central region of the time series with half the length of the window
unrepresented at either end. The numerically estimated equilibrium
trend $x_\mathrm{eq}$ is displayed as a black curve at the bottom of
the valley of $U_\mathrm{emp}$. For comparison with this
$U_\mathrm{emp}$ we display at either end the real (nonlinear)
potential $U(x,t)$ corresponding to \eqref{eq:snfnoise}, which at the
later time is already showing the hill-top to the left.
The colour in Figure~\ref{fig:c365} shows the deviation of the
empirical potential from the linear-theory parabola, which is not
itself displayed. The softening, highlighted by the blue colouration,
is seen on the left hand side which is approaching the
hill-top. Transparency of the colour is used to show the number of
data points that support a particular part of the surface
$U_\mathrm{emp}(\tilde x)$ (as given by the empirical quasi-stationary density
$p_\mathrm{emp}(\tilde x)$).
Several methods have been proposed to detect nonlinearity in the
potential well in the form of this type of softening (or tilting).
\begin{itemize}
\item \textbf{Skewness} \citet{Guttal2008,Guttal2008a} studied how the
closeness of a bifurcation influences the skewness $\gamma$ of the
empirical quasi-stationary density from time series. As this
approach is also based on the empirical quasi-stationary density
$p_\mathrm{emp}$, it generalises the estimate $\kappa_U$ to
non-parabolic features of the well $U$ and non-Gaussian features of
the stationary density. \citet{Guttal2008,Guttal2008a} proposed to
look at the \emph{trend} of the skewness $\gamma$ over time to
detect incipient bifurcations. \jsmod{Figure~\ref{fig:snfdata}(b-4)
shows how the skewness, taken over a single window, changes. It is
noticeably negative (compare to the empirical skewness from the
linear time series in Fig~\ref{fig:snfdata}(a-4)) but no trend is
discernable.}
\item \textbf{Quasi-stationary density} \citet{Livina2011} generalised
the potential well analysis for the quasi-stationary density beyond
the estimate $\kappa_U$ for the linear decay rate, fitting the
potential well to higher-order polynomials of even degree using
relation~\eqref{eq:logu}. This analysis was not used to attempt
prediction of transition probabilities from time series but it was
applied to time series that already included all transitions to
count the number of wells of $U$ depending on time $t$ (which
corresponds to the number of modes (peaks) of the empirical
quasi-stationary density $p_\mathrm{emp}$). See also
[Beaulieu \emph{et al.} this volume]\ for methods that detect change points in time
series.
\item \textbf{Drift ratio} If one assumes that the underlying system
parameter is approaching a fold more or less linearly
($\dot\mu(t)=\epsilon>0$) and one is sufficiently close to the fold
already then the ratio between the drift $\Delta x_\mathrm{eq}$ of
the equilibrium estimate and the increase of the estimated linear
decay rate $\Delta \kappa$ estimates the quadratic term in the
normal form \citep{Thompson2011}.
\end{itemize}
\section{Quantitative estimates of the nonlinear coefficients}
\label{sec:quant}
The colour shading of Figure~\ref{fig:c365} suggests that the
nonlinear term in the underlying deterministic system shows up in the
quasi-stationary density $p_\mathrm{emp}$. This raises two questions.
First, can the level of deviation from linear theory be distinguished
with confidence from random chance? That is, can one state, when
looking at a time series, that a significant quadratic term is
present? Second, is it possible to quantify the size of the nonlinear
term with reasonable level of certainty?
Finding a significant nonlinearity is far easier than estimating its
precise value because biases introduced during the analysis of the
equilibrium and decay rate (zero-order and first-order analysis) may
distort the value of the nonlinear estimate but may still keep it
significantly different from zero. We will demonstrate this in
Section~\ref{sec:geolog} for two paleo-climate time series.
\jsmod{\subsection{Properties of the stationary density of the
saddle-node normal form}}
\label{sec:fokkerplanck}
\begin{figure}[th]
\centering
\includegraphics[width=0.8\textwidth]{skewnessdeco}
\caption{Dependence of mean, variance $\operatorname{var} p$ and skewness $\gamma$
of the empirical stationary density for the saddle-node normal
form with additive Gaussian noise \eqref{eq:snfnoise} \protect\jsmod{as
computed directly from the Fokker-Planck equation~\eqref{eq:fp}
with $-\partial_xU(x)=\mu-x^2$ and $\sigma=1$. The escape rate
$c$ is defined by the constant of integration in
Equation~\eqref{eq:fp}.}}
\label{fig:skewness}
\end{figure}
Figure~\ref{fig:skewness} shows the dependence of the mean, the
variance $\operatorname{var} p$, and the skewness $\gamma$ of the empirical
stationary density $p_\mathrm{emp}$ on the system parameter $\mu$ for
\eqref{eq:snfnoise} as computed using the stationary Fokker-Planck
equation \eqref{eq:fp}. Note that this
\jsmod{corresponds to the stationary process where the
random realisation is re-injected at $+\infty$ whenever it has
escaped to $-\infty$.}
We observe that the dependence of
$\gamma$ is strongly nonlinear for the saddle-node normal form with
noise ($\dot\mu(t)=\epsilon=0$, $\sigma=1$ in
Equation~\eqref{eq:snfnoise}). This means that trends of the skewness
are likely to be difficult to ascertain when approaching a
fold. However, the presence of skewness is an indicator for the
nonlinearity uniformly for all parameters $\mu$, and its sign
indicates the direction of escape.
Note also how the nonlinearity affects the empirical mean and the
variance $\operatorname{var} p$, which is no longer inversely proportional to the
linear decay rate $\kappa$. This shift of the mean and the additional
variance is in part due to the fraction of trajectories that escape,
in part due to the non-parabolic shape of the well at some distance
from its local minimum.
\jsmod{\subsection{Practical estimates of the nonlinear term from time
series}}
\label{sec:estnonlin}
We can estimate the nonlinearity directly as the deviation of the
empirical potential well $U$ from a parabola as suggested by
figure~\ref{fig:c365} (we are only interested in $\partial_xU$). The
simplest approach (apart from looking at skewness $\gamma$) is to fit
$\kappa_U$ and $c_\mathrm{emp}$ to the empirical density
$p_\mathrm{emp}$ using Equation~\eqref{eq:fp1}. If $\partial_xU(\tilde
x)$ is indeed linear then $c_\mathrm{emp}$ would be zero such that
$c_\mathrm{emp}$ is a signed scalar measure for the deviation of
$\partial_xU(\tilde x)$ from linearity, serving as a proxy for the
present nonlinearity. A second alternative is to expand
$-\partial_xU(\tilde x)$ to second order, incorporating a quadratic
nonlinearity directly into $\partial_xU$ with an unknown coefficient
$N_2$. This leads to
\begin{align}
\nonumber -\sigma^{-2}\partial_xU(\tilde x)&= -\kappa_U \tilde
x+N_2\tilde x^2\mbox{,\quad such that we fit}\\
\label{eq:fp2}
\frac{1}{2}\partial_x p_\mathrm{emp}(\tilde x)&=
\left[-\kappa_U \tilde x+N_2\tilde x^2\right]\,
p_\mathrm{emp}(\tilde x)+c_{\mathrm{emp},2}\mbox{.}
\end{align}
Both quantities, $c_\mathrm{emp}$ and $N_2$, measure the deviation of
$-\sigma^{-2}\partial_xU(\tilde x)$ from its linear approximation
$\kappa_U\tilde x$. The main difference between them is the weighting
($\tilde x^2$ for $N_2$, uniform for $c_\mathrm{emp}$). \jsmod{In
summary, the step-by-step procedure to extract an estimate of the
quadratic term (or a proxy for it) from a time series $x_k$ is the
following.
\begin{enumerate}
\item Detrend the time series $x_k$ (see
Section~\ref{sec:timeseries}(\ref{sec:0order})). We call the
detrended time series $\tilde x_k$. The following steps are
performed for each suitable time $t$ to obtain $\kappa_U(t)$,
$c_\mathrm{emp}(t)$, $N_2(t)$ or skewness $\gamma(t)$:
\item Choose a length $w$ of sliding windows and obtain the
empirical stationary density $p_\mathrm{emp}(\tilde x)$ for the
sliding window centered at $t$. (We used \texttt{kde1d} to obtain
$p_\mathrm{emp}(\tilde x)$. For non-stationary time series we used
the sliding window length $w=N/2$.)
\item Use relation \eqref{eq:fp1} to find $\kappa_U(t)$ and
$c_\mathrm{emp}(t)$ from $p_\mathrm{emp}$. Use relation
\eqref{eq:fp2} to find $N_2(t)$ (and another estimate for
$\kappa_U$) from $p_\mathrm{emp}$. Compute the empirical skewness
$\gamma(t)$ from $p_\mathrm{emp}$.
\end{enumerate}
}
\jsmod{
\subsection{Comparison of estimates using saddle-node normal form}
\label{sec:snftest}
}
\begin{figure}[th]
\centering
\includegraphics[width=0.9\textwidth]{snftestdeco}
\caption{Quartiles of lengths (d), and estimates $c_\mathrm{emp}$
(a), $N_2$ (b), skewness $\gamma$ (c), $\kappa_\mathrm{ACF}$, and
$\kappa_U$ (both (e)) for time series generated by the saddle-node
normal form \eqref{eq:snfnoise} with noise ($\epsilon=0$, $\mu=1$)
and for linear time series generated by
\eqref{eq:ou}. \protect\jsmod{Without nonlinearity transition would occur
at $\mu=0$.}}
\label{fig:snftest}
\end{figure}
Figure~\ref{fig:snftest} shows how the estimates
for nonlinearity behave for the saddle-node normal form with noise,
Equation~\eqref{eq:snfnoise} with $\epsilon=0$ and
$\sigma=1$.
Each panel shows the quartiles of the distribution of the estimate for
$100$ realisatons of time series generated by
\eqref{eq:snfnoise}. Each realisation was run until it reached $2000$
data points or $x=-1$ (indicating escape). Panel (a) shows the
estimate $c_\mathrm{emp}$ obtained by linear-least-squares fitting of
the approximate stationary Fokker-Planck equation~\eqref{eq:fp1} to
the empirical stationary density $p_\mathrm{emp}$, panel (b) shows the
quadratic coefficient $N_2$ obtained from \eqref{eq:fp2}, and panel
(c) shows the skewness $\gamma$ of $p_\mathrm{emp}$. For comparison,
we generate also $100$ linear time series using
\begin{equation}
\label{eq:ou}
\d x=-2\sqrt{\mu}x\d t+\d W_t
\end{equation}
and then fit $c_\mathrm{emp}$, $N_2$ and $\gamma$ for these time
series, too. Figure~\ref{fig:snftest} compares the quartiles of the
distributions for time series generated by the saddle-node normal form
\eqref{eq:snfnoise} and the linear equation \eqref{eq:ou}. If there is
small overlap in the distribution then the quantity is a good
indicator for the presence of a quadratic nonlinear term. We observe
that this is the case for all quanitites as long as the time series
has a length of $2000$ (see Figure~\ref{fig:snftest}(d) for the
distribution of time series lengths). Naturally, the uncertainty, and,
hence, the overlap, increases when the time series is shorter due to
escape from the potential well to $-\infty$ . This is the case for
smaller $\mu$ in Figure~\ref{fig:snftest} (see \citet{Thompson2011}
for a quantitative study of escape probabilities for small $\mu$). We
also notice that all quantities have a systematic deviation from the
theoretical value of the quantity they supposedly estimate: $N_2$
should be equal to $-1$ according to \eqref{eq:snfnoise}, the real
escape rate $c$ has a much smaller modulus than $c_\mathrm{emp}$
(compare Figure~\ref{fig:skewness}(a)), and the skewness $\gamma$ has
theoretically a much larger modulus (compare
Figure~\ref{fig:skewness}(b)). The deviation for $N_2$ is relatively
small, and is likely caused by the kernel density estimate for
$p_\mathrm{emp}$. The large modulus of $c_\mathrm{emp}$ shows that the
linear term $-\kappa_U\tilde x$ is a poor fit for $-\partial_xU(\tilde
x)$, which makes $c_\mathrm{emp}$ a measure of how much the odd
symmetry is broken by $-\partial_xU(\tilde x)$, rather than an
estimate for the escape rate. We note that $c_{\mathrm{emp},2}$ as
fitted in \eqref{eq:fp2} has a realistic modulus (close to zero, not
shown in Figure~\ref{fig:snftest}). The bias of the skewness $\gamma$
is mainly due to our restriction to time series that stay inside the
potential well. However, the characteristic dip of $\gamma$ seen in
Figure~\ref{fig:skewness}(b) is still visible in
Figure~\ref{fig:snftest}(c). \jsmod{\ref{sec:skewness}
addresses the dependence of the empirical skewness on all method
parameters.
We also note that the estimates for the linear decay rate,
$\kappa_U$ and $\kappa_\mathrm{ACF}$ are more accurate than the the
estimates for the nonlinearity but by their nature they cannot
distinguish the time series generated by the saddle-node normal form
from a linear surrogate.}
The summative conclusion from Figure~\ref{fig:snftest} is that one can
detect the underlying nonlinearity of the deterministic part of
Equation~\eqref{eq:snfnoise} by observing $c_\mathrm{emp}$, $N_2$ or
the skewness $\gamma$ if the time series is moderately long. One can
expect this to be true in the more general case of a time series
generated by a system with deterministic dynamics close to a fold
and additive noise. While one can in principle recover the underlying
parameter $\mu$ from the estimates for $\kappa_U$, $N_2$ and $\sigma$
(see \citet{Thompson2011}), the uncertainty in $N_2$
propagates dramatically into uncertainty for $\mu$ (one has to divide
by $N_2$).
\jsmod{The results of Figure~\ref{fig:snftest} show the stationary
case ($\epsilon=0$). For slowly drifting time series the same
analysis will then have to be applied to parts of the time series in
sliding windows. We demonstrate this for two geological times series
in Section~\ref{sec:geolog}. For time series with rapidly drifting
system parameter $\epsilon$ the estimates for the softening,
$c_\mathrm{emp}$, $N_2$ and the skewness $\gamma$, will not give a
detectable difference from the corresponding linear time series
before tipping. The conclusion one would draw from this absence of
detectable softening would be correct: tipping happens when the
linear decay rate $\kappa$ reaches the critical value $0$ (or
slightly later, see \citet{Thompson2011}). The same holds if the
noise level is low, which is equivalent to rapid drift (see
\citet{Thompson2011} for the transformation). In practice one uses
the argument the other way round: if the nonlinear estimates do not
give a value significantly different from zero one will conclude that
the ratio between noise level and drifting speed is small.}
\section{Studies of Geological Time Series}
\label{sec:geolog}
We now apply our nonlinear investigations to two ancient climate
tippings. The first is the major transition marking the end of the
latest ice age which occurred about $17,000$ years before the present
(yrs BP). The data used for this is a temperature reconstruction from
the Vostok ice core deuterium record (\citet{Petit1999}). The second
more recent tipping is the ending of the Younger Dryas using the
grayscale from the Cariaco basin sediments in Venezuela
(\citet{Hughen2000}). This Younger Dryas event was a curious cooling,
about $11,500$ years ago, just as the Earth was warming up after the
last ice age. It ended in a dramatic tipping point, when the Arctic
warmed by $7\celsius$ in $50$ years. This sudden ending has been
related (\citet{Houghton2004}) to a \emph{switching-on} of the global
oceanic thermohaline circulation (THC). This switch-on is known from
extensive theoretical studies (\citet{Dijkstra2005}) to be at a
saddle-node fold arising as a perturbation of a sub-critical pitchfork
(see, for example, \citet{Rahmstorf2000} and \citet{Thompson2011}).
\begin{figure}[ht]
\centering
\includegraphics[width=\textwidth]{c366_End_last_ice_age_Younger_D_FFCL_PUB}
\caption{Two predicted potential energy surfaces for (a) the end of
the last glaciation, using the Vostok ice-core record, and (b) the
end of the Younger Dryas event when the Arctic warmed by
$7\celsius$ in $50$ years, using the grey-scale of basin sediment
in Cariaco, Venezuela. The colour shows the deviation from a
(time-dependent) parabola fitted to $U_\mathrm{emp}(x,t)$, three
of which are illustrated above. Red signifies a positive
deviation, blue a negative deviation. Time is given in years
before the present (BP). \protect\jsmod{The plotted potential was
obtained via
$U_\mathrm{emp}=-\sigma_\mathrm{emp}^2\log(p_\mathrm{emp})/2$
and \eqref{eq:sigmaemp} and sliding windows of half the length
of the record.}}
\label{fig:c366}
\end{figure}
Under linear time-series analysis of the directly preceding data,
neither of these two events shows a strong trend in the stability
propagator (see Figure~\ref{fig:records}). Meanwhile sample estimates
of the underlying potential functions are shown in
Figure~\ref{fig:c366}.
In Figure~\ref{fig:c366}(a), for the end of the last glaciation, we
see clearly that the well is softening (falling beneath the parabolic
fit) on the high-temperature end while hardening (rising above the
parabola) on the low-temperature end. So there is a strong nonlinear
signal that a jump to higher temperatures may be pending (as indeed it
was). We show that this signal is statistically significant in our
full analysis in Figure~\ref{fig:records}. By comparison, the study of
the Younger Dryas, illustrated in Figure~\ref{fig:c366}(b), provided
no significant nonlinear conclusions.
\begin{figure}[th]
\centering
\includegraphics[width=0.8\textwidth]{recordsdeco}
\caption{Linear and nonlinear coefficients for two paleo-climate
time series. (a,c,e,g): End of last glaciation (snapshot of data
from \citet{Petit1999}); (b,d,f,h): End of Younger Dryas (snapshot
of data from \citet{Hughen2000}). Only the black part of the data
in panels (a) and (b) was used in the analysis. Panel (c) and (d)
show the time series after detrending. Panel (e) and (f) show the
linear indicators $\kappa_\mathrm{ACF}$ and $\kappa_U$. Panel (g)
and (h) show the means of the estimates $c_\mathrm{emp}$, $N_2$
and the skewness $\gamma$, compared to zero (thin vertical black
line) and a histogram of estimates sampled from $500$ random
linear time series generated by the linear process \eqref{eq:ou}
with
$-2\sqrt{\smash[b]{\mu}}=\kappa_\mathrm{ACF}$. \protect\jsmod{Sliding
window length $w=N/2$.}}
\label{fig:records}
\end{figure}
For Figure~\ref{fig:records} we detrended both time series using
Gaussian filtering (see Figures~\ref{fig:records}(c) and (d)) and
estimated the linear indicators $\kappa_\mathrm{ACF}$ and
$\kappa_U$. We observe weak to non-existent trends, leaving the
evidence inconclusive at the linear level (grey and black curves
Figure~\ref{fig:records}(e) and (f)). The ratio between
$\kappa_\mathrm{ACF}$ and $\kappa_U$, both shown in
Figure~\ref{fig:records}(e) and (f), gives an estimate of the variance
$\sigma^2$ of the noise input. \jsmod{The unit for
$\kappa_\mathrm{ACF}$ is \emph{units of $x$ per time step} (so, for
example for Fig.~\ref{fig:records} (e) $\celsius/\Delta t$), and,
correspondingly, the unit for $\kappa_U$ is $\Delta t/\celsius$
(unit of $\kappa_\mathrm{ACF}$ divided by unit for $\sigma^2$). Both
units are arbitrary such that we do not indicate the scale of the
$y$-axis in Figure~\ref{fig:records}(e,f).}
\jsmod{[Lenton \emph{et al.} this volume]\ discuss the linear analysis of both time
series in greater detail. Their time series \emph{Vostok}
corresponds to Figure~\ref{fig:records}(a), their time series
\emph{Cariaco} corresponds to Figure~\ref{fig:records}(b). (Note
that [Lenton \emph{et al.} this volume]\ use the deuterium proxy directly whereas
Figure~\ref{fig:records}(a) shows the temperature reconstruction.)
Specifically, Figures 2 and 4, and Figures 9 and A1 in
[Lenton \emph{et al.} this volume]\ discuss how the results depend on method
parameters such as detrending bandwidth and sliding window. Their
analysis finds that the presence of a significant trend depends
strongly on both parameters. See [Lenton \emph{et al.} this volume]\ , Table~1 (rows
\emph{Vostok} and \emph{Cariaco}), for a summary of the evidence at
the linear level.}
At the nonlinear level, the time series for the end of the last
glaciation has a strong quadratic nonlinearity in $\partial_xU(\tilde
x)$ where the potential well $U$ is presumed to govern the
deterministic part of the dynamics (Figure~\ref{fig:records}(g)). The
indicators $c_\mathrm{emp}$, $N_2$ and the skewness $\gamma$ in
Figure~\ref{fig:records}(g) are all far removed from what can be
expected by chance in a linear time series. The histogram in the
background of Figure~\ref{fig:records}(g,h) has been sampled from
$500$ random linear time series generated by the linear process given
in Equation \eqref{eq:ou} with
$-2\sqrt{\smash[b]{\mu}}=\kappa_\mathrm{ACF}$ where
$\kappa_\mathrm{ACF}$ is taken from the estimate shown in
Figure~\ref{fig:records}(e). The percentages at the top of
Figure~\ref{fig:records}(g,h) express how far in the tail of the
histogram the \jsmod{mean of the} quantity extracted from the
geological time series is. \jsrem{ $50$\% corresponds to the median of
the histogram, a percentage smaller than $50$\% gives the percentage
of linear realisations that were further away from the median than
the geologocal data.} \jsmod{We warn that the three quantities
$\gamma$, $N_2$ and $c_\mathrm{emp}$ are not really three
independent indicators as they all depend on the same estimate of
the empirical density $p_\mathrm{emp}$.}
Visual inspection of the well shape in Figure~\ref{fig:c366} confirms
that the well is softening (bending downward) on the high-temperature
end and hardening (bending upward) on the low-temperature end. So, at
the nonlinear level the time series data close to the bottom of the
well gives already evidence for a propensity to escape toward larger
temperatures. The time series for the end of the Younger Dryas does
not show evidence for strong nonlinearity of the underlying dynamics
at the second-order level (Figure~\ref{fig:records}(h)). The values
for $c_\mathrm{emp}$, $N_2$ and the skewness $\gamma$ can all be
explained by randomness as the histograms of estimates for the $500$
linear time series (as the histograms in Figure~\ref{fig:records}(h)
show). Note that we did not include the previous transition (visible
at the left end in Figure~\ref{fig:records}(b)) into our analysis as
our aim was to infer the nonlinearity exclusively from the data near
the tipping equilibrium.
\section{Conclusion}
\label{sec:conc}
\jsmod{The main message of the paper is that the linear analysis
investigated by [Lenton \emph{et al.} this volume]\ on its own, even if all estimates
are accurate, does not contain all information necessary to estimate
the probability of tipping over time.} The result of the linear
analysis is an estimate of the decay rate $\kappa$\jsmod{, which is taken}
relative to the time spacing of the measurements. Even if this decay
rate has an identifiable trend to zero the probability of tipping over
time is determined by the dominant nonlinear coefficient $N_2$ in
conjunction with $\kappa$ and the noise level.
We found that the dominant nonlinear coefficient $N_2$ is much harder
to estimate accurately, so we also looked at proxies that indicate
nonlinearity such as the skewness $\gamma$ (as proposed by
\citet{Guttal2008}) or the coefficient $c_\mathrm{emp}$ from
Equation~\eqref{eq:fp1} (which is also a scalar signed measure for
deviation from linear theory). Our study of the saddle-node normal
form suggests that both proxies, $\gamma$ and $c_\mathrm{emp}$, give
values that are easier to distinguish from random chance than the
estimate for $N_2$ itself. A significantly non-zero value of either of
these proxies indicates that a quadratic nonlinear term is present and
gives its sign.
However, we found the uncertainty for moderately long time series
($N=2000$) too large to translate the proxy back into a quantitatively
reliable estimate for the normal form parameter (this would be
necessary to read off the probability for tipping from the tables in
\citet{Thompson2011}). We also found that the nonlinear proxies do not
have discernible trends when one approaches tipping, so it makes sense
only to extract their overall mean from the time series but not the
trend.
\bibliographystyle{unsrtnat}
|
\section{Introduction}
\object{Q 01323-4037} and \object{Q 0132-4037} are listed as two separate quasi-stellar objects (QSOs) in the last editions of the \emph{catalogue of quasars and active nuclei} \citep{VeronCetty:2003p565,VeronCetty:2006p231,VeronCetty:2010p564}. According to this catalog, the two objects (A and B for simplicity) have similar redshift, $z_\mathrm A=2.100$, $z_\mathrm B=2.150$. Their angular separation in the sky, as computed from the J2000 coordinates ($\alpha_\mathrm A=01$ $34$ $32.5$, $\delta_\mathrm A=-40$ $22$ $08$ and $\alpha_\mathrm B=01$ $34$ $32.2$, $\delta_\mathrm B=-40$ $21$ $33$, respectively), is approximately $35$ arcsec, making the seeming QSO pair a suitable candidate for a tomographic study of the inter-galactic medium along close lines of sight.
We observed the two objects in November, 2010 with the single target, medium resolution spectrograph X-shooter \citep{DOdorico:2006p815} at the \emph{Very Large Telescope} (VLT), in the context of a GTO program. Fig.~\ref{fig:chart} reproduces the finding chart used for our observation, with the position of A and B according to \citet{VeronCetty:2010p564}. We extracted the 1D spectra of the objects for the three X-shooter arms (UVB, VIS, and NIR) using the release 1.2.0 of the X-shooter reduction pipeline \citep{Goldoni:2006p199} and the ESO-MIDAS package. Fig.~2 shows the spectra obtained after flat-field correction, bias and sky subtraction, and flux calibration. Much to our surprise, the observation revealed that B is not a QSO, unlike A, which displays a typical QSO spectrum with a strong Lyman $\alpha$ emission at $\lambda\simeq 3820$ \AA. Apparently, a spurious object has been included in the V\'eron-Cetty \& V\'eron catalog by mistake.
\begin{figure}[htbp]
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{cupani-et-al_2011fg1.eps}}
\caption{Finding chart with the position of object A (a QSO) and object B (a star mistakenly identified as a separate QSO). The image is centered on object A (J2000 coordinates $\alpha_\mathrm A=01$ $34$ $32.5$, $\delta_\mathrm A=-40$ $22$ $08$).}
\label{fig:chart}
\end{center}
\end{figure}
\begin{figure*}[htbp]
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{cupani-et-al_2011fg2.eps}}
\caption{Spectra of object A (upper panels) and object B (lower panels) obtained with X-shooter. Flat-field correction, bias and sky subtraction, and flux calibration were performed on both spectra. Vertical dotted lines highlight the main emission features which allow to identify object A as a QSO. Similar features are totally absent in the spectrum of object B. [Color legend -- Blue: UVB arm; black: VIS arm; red: NIR arm.]}
\label{fig:spectra}
\end{center}
\end{figure*}
\section{Origin of the misidentification}
We argue that the misidentification of object B as a QSO is a consequence of poorly-constrained measurements on the object A, which is in fact the only QSO in a radius of $10$ arcmin. Different values of position and redshift attributed to A during the years were interpreted as the presence of \emph{two} QSOs in this region, one of which was incorrectly identified with B.
QSO A was first observed by \citet{Hoag:1977p647} at B1950 coordinates $\alpha=01$ $32$ $22.2$, $\delta=-40$ $37$ $18$. \citet{Osmer:1980p581} published the redshift of this objects, $z=2.15\pm 0.01$.\footnote{Quite interestingly, \citet{Hoag:1977p647} and \citet{Osmer:1980p581} are quoted in the NASA-IPAC Extragalactic Database as references for A, and in the SIMBAD Astronomical Database as references for B. This is a hint that some mismatch occurred in the identification of the two objects.} The QSO is included in the \emph{Automatic Plate Measurement} (APM) galaxy survey \citep{Maddox:1990p587} and in the catalog by \citet{Hewitt:1993p390} (labeled as B 0132-406). It is also listed in the catalog by \citet{Iovino:1996p605}, with B1950 coordinates $\alpha=01$ $32$ $21.2$, $\delta=-40$ $37$ $29$ and redshift $z=2.10$. Despite the slight differences, there is no doubt that the object is the same.
Unfortunately, the finding chart published by \citet{Hoag:1977p647} points to object B, not to object A, as the target QSO. This error (together with the poor accuracy of the published QSO position) is the most probable cause of the misidentification, which can be traced back to \citet{VeronCetty:1996p645}. Here the B1950 coordinates of the QSO appear to be $\alpha=01$ $32$ $21.12$, $\delta=-40$ $36$ $53.7$ (corresponding to J2000 coordinates $\alpha=01$ $34$ $32.27$, $\delta=-40$ $21$ $33.6$), quite different than the values previously measured, and coincident with the position of object B. The new coordinates appear in all editions of the V\'eron-Cetty \& V\'eron catalog until 2001 \citep{VeronCetty:1996p573,VeronCetty:1998p654,VeronCetty:2000p651,VeronCetty:2001p676}. During this period, all references to Q 0132-4037 \citep{Barkhouse:2001p656,Cutri:2003p658} point to object B instead of object A.
The real QSO reappears in \citet{VeronCetty:2003p565} as an additional entry Q 01323-4037, distinct from Q 0132-4037. The authors cite \citet{Iovino:1996p605} as a reference for the first one, and \citet{Osmer:1980p581} as a reference for the second one. The slight difference in position and redshift may have led them to believe there were two different QSO, even though both \citet{Osmer:1980p581} and \citet{Iovino:1996p605} had observed the same object. The error has been reproduced in the subsequent of the catalog until today \citep{VeronCetty:2006p231,VeronCetty:2010p564}.
To summarize: the two entries Q 01323-4037 and Q 0132-4037 in the last editions of V\'eron-Cetty \& V\'eron catalog correspond in fact to a single object, QSO A, with J2000 coordinates $\alpha=01$ $34$ $32.5$, $\delta=-40$ $22$ $08$ and redshift $z$ between $2.10$ and $2.15$. Object B, whose coordinates are associated with Q 0132-4037, is not a QSO and should not be considered as such. The spectrum of object B is that of a star with spectral type early M, as seen from weak (but clear) TiO absorption bands around 620 nm and 710 nm. The spectrum shows a good resemblance to the spectrum of the M1 III star \citep{Bagnulo:2003p887}
Hereafter, we suggest to address to object A as Q 0132-4037, and to stop using the ambiguous identifier Q 01323-4037 introduced by \citet{VeronCetty:2003p565}.
\section{Refined measurement of Q 0132-4037 redshift}
According to several studies \citep{Gaskell:1982p852,Vrtilek:1985p867,Hutchings:1987p872}, redshift estimated from forbidden lines are in agreement to $\sim$$100$ km s$^{-1}$ with the systemic redshifts of QSOs as determined by stellar absorption and H \textsc{i} 21 cm emission in the host galaxies. We estimated the redshift of Q 0132-4037 by a gaussian fitting of the narrow line [O \textsc{iii}] $\lambda 5007$, which appears unblended and quite prominent in the NIR part of the spectrum (fig.~\ref{fig:spectra}). We obtained $z_{\scriptsize{\textrm{[O \tiny{\textsc{iii}}]}}}=2.1568\pm 0.0081$. The same procedure was performed for the low-ionization line Mg \textsc{ii} $\lambda 2798$ and the H$\beta$ line, giving $z_{\scriptsize{\textrm{Mg \tiny{\textsc{ii}}}}}=2.1614\pm 0.0155$ and $z_{\mathrm{H}\beta}=2.1629\pm 0.0200$, respectively. The three estimates are in agreement within the uncertainty. The combined fiducial value of the systemic redshift is estimated as $z_\mathrm{sys}=2.1583\pm 0.0067$, confirming the value by \citet{Osmer:1980p581} and ruling out the value $z=2.10$ published by \citet{Iovino:1996p605}.
\begin{acknowledgements}
Based on observations collected at the European Southern Observatory, Chile, as part of program 086.A-0076. The Dark Cosmology Centre is funded by the Danish National Research Foundation. G.~C.~would like to thank Carlo Morossi for his help in recognizing the spectral type of object B.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
Extending quantum channels by allowing for Hamiltonian control
turns them into interesting and important examples of geometric control of
open systems. While for \emph{closed systems}, the theory of Lie
groups provides a rich structure to address questions of reachability,
accessibility, and controllability \cite{Jurdjevic97}, already
simple \emph{open quantum systems} come with the intricate
geometry of Lie semigroups \cite{HHL89,Lawson99}.
For instance, in most closed systems
the reachable set to an initial state $\rho_0$ simply is
the orbit $\mathcal O_{\bG}(\rho_0):=\{G\rho_0 G^{-1}\,|\, G\in\bG\}$
of a unitary subgroup $\bG$ whose Lie algebra can be identified
easily via Lie closure, while in {\em open systems} reachable sets are much
more difficult to determine explicitly.
Thus in view of controlling open quantum dynamics, in \cite{DHKS08}
we systematically related the framework of completely positive semigroups
\cite{Kraus71,Koss72,Koss72b,Choi75,GKS76,Lind76,Kraus83}, which is
well established in quantum physics, with the more recent
mathematical theory of Lie semigroups. An early example confined to
single-qubit systems can be found in \cite{Alt03}.
More precisely, for exploiting the power of systems and control theory
in open quantum dynamics, the system parameters have to be characterised
first, e.g., by input-output relations in the sense of quantum
process tomography. The decision problem whether the dynamics
of the quantum system thus specified is Markovian to good
approximation has recently been analysed \cite{Wolf08a,Wolf08b}.
Moreover (time-dependent) Markovian quantum channels
were elucidated from the viewpoint of divisibility \cite{Wolf08a}
thus paving the way to Lie semigroups \cite{DHKS08}. Following up,
this work sets out to determine the geometry of quantum channel
semigroups in terms of their tangent cones (Lie wedges)
for a number of coherently controlled standard unital channels
in a unified frame in line with \cite{KuDiHeTAC11a}.
For the first time, here we
explicitly parameterize the {\em set of all possible directions}
an open quantum system under coherent controls
may take --- its \emph{Lie wedge}. Thereby, we heavily exploit
the fact that the set of all reachable quantum maps governed
by a controlled Markovian master equation constitutes a Lie semigroup
\cite{DHKS08}.
Previous characterizations of reachable sets
for unital open quantum systems by majorization techniques,
e.g., \cite{Yuan10}, become increasingly inaccurate
once full controllability of the Hamiltonian part
(\mbox{condition (H)} {\em vide infra})
is violated, which for growing number of qubits happens
{\em in all experimentally realistic settings}.
In contrast, the Lie-semigroup tools presented here
do not require \mbox{condition (H)}
and carry over to multi-qubit systems
without the draw-back of increasing inaccuracy.
\section{Theory and Background}
We start out by recalling some basic notions and notations
of Lie subsemigroups \cite{HHL89} and their
application for characterising reachable sets of quantum control
systems modelled by Lindblad-Kossakowski master equations
\cite{DHKS08}.
\subsection{Lie Semigroups}\label{sec:LieSemiBasics}
To begin with, let $\bG$ be a matrix Lie group, i.e.~a
group which is (isomorphic to) a path-connected subgroup of
$GL(n,\R{})$ or $GL(n,\C{})$ for some $n \in \N{}$, and let
$\fg$ be its corresponding matrix Lie algebra. Thus $\fg$ is
(isomorphic to) a Lie subalgebra of $\gl(n,\R{})$ or $\gl(n,\C{})$.
Then a subset $\bS\subset\bG$ which is closed under the
group operation in the sense $\bS\cdot\bS\subseteq\bS$ and which
contains the identity $\unity$ is said to be a {\em subsemigroup}
of $\textbf{G}$.
The largest subgroup within $\bS$ is written $E(\bS):=\bS\cap\bS^{-1}$.
Furthermore, a closed convex cone $\fw \subset \fg$
is called a wedge. The largest linear subspace of $\fw$
is denoted $E(\mathfrak{w}):=\mathfrak{w}\cap(\mathfrak{-w})$
and it is termed the {\em edge of the wedge} $\fw$.
Now, $\fw\subseteq\fg$
is a {\em Lie wedge} of $\fg$ if it is invariant under
the adjoint action
of the subgroup generated by the edge $E(\mathfrak{w})$,
i.e.~if it satisfies
\begin{equation}
e^A \,\fw\, e^{-A} = \fw
\end{equation}
(or equivalently $e^{\adr_A}(\fw) = \fw$) for all $A\in E(\fw)$.
Note that the edge of a Lie wedge always
forms a Lie subalgebra of $\fg$.
Moreover, for any closed subsemigroup $\bS$ of $\bG$ we define
its tangent cone $L(\bS)$ at the identity $\unity$ by
\begin{equation}
L(\bS) := \{ A\in\fg\,|\, \exp(tA)\in \bS\text{\; for all\;} t\geq 0\}\;.
\end{equation}
Then one can show that $L(\bS)$ is a Lie wedge of
$\fg$ satisfying the identity
$
E\big(L(\bS)\big) = L\big(E(\bS)\big) .
$
Yet, the \/`local-to-global\/' correspondence
between Lie wedges and closed connected subsemigroups is
much more subtle than the correspondence between Lie (sub)algebras
and Lie (sub)groups: for instance, several connected subsemigroups may
share the same Lie wedge $\fw$ in the sense that $L(\bS) = L(\bS')$
for $\bS \neq \bS'$, or conversely there may be Lie wedges $\fw$
which do not correspond to any subsemigroup, i.e.~$\fw=L(\bS)$
fails for all subsemigroups $\bS\subset\bG$.
Therefore, one introduces the important notion of
a {\em Lie subsemigroup} $\bS$ which is characterised by the equality
\begin{equation}
\bS = \overline{\expt{\exp L(\bS)}}_S\;,
\end{equation}
where the closure is taken in $\bG$ and $\expt{\exp L(\bS)}_S$
denotes the subsemigroup
generated by $\exp L(\bS)$,
i.e.~$\expt{\exp L(\bS)}_S := \{e^{A_1}
\cdots e^{A_n}\,|\, n\in\N, \, A_1, \dots, A_n\in L(\bS)\}$.
Moreover, a Lie wedge $\fw$ is said to be {\em global} in $\bG$,
if there is a Lie subsemigroup $\bS\subset\bG$ such that
\begin{equation}
L(\bS)=\fw
\;.
\end{equation}
Thus, one has the identity $\bS=\overline{\expt{\exp \fw}}_S$.
Whenever a Lie wedge $\fw\subset\fg$ specialises to be
compatible with the Baker-Campbell-Hausdorff (BCH)
multiplication
\begin{equation}
A\star B := A+B+\tfrac{1}{2}[A,B]+ \dots
= \log(e^A e^B)\quad\forall A,B\in\fw\end{equation}
defined via the BCH series, it
is termed {\em Lie semialgebra}. For this to be the case, there
has to be an open BCH neighbourhood $\mathcal B\subset\fg$
of the origin in $\fg$ such that
$(\fw\cap \mathcal B)\star(\fw\cap \mathcal B) \subseteq\fw$.
An equivalent definition for
being a Lie semialgebra is given by the tangential condition
\begin{equation}\label{eqn:semialg-incl}
[A,T_A\fw] \subset T_A\fw \quad\text{for all $A \in \fw$}\;,
\end{equation}
where $T_A\fw$ denotes the tangent space of $\fw$ at $A$
defined by
\begin{equation}
T_A\fw := \big(A^\perp \cap \fw^*\big)^\perp\;.
\end{equation}
Here $A^\perp$ denotes the {\em orthogonal complement} of $A$ and
$\fw^* :=
\{A \in \fg \,|\, \langle A,B \rangle \geq 0 \text{ for all } B \in \fw\}$
the {\em dual wedge}---both taken with respect to the standard trace
inner product.
The conceptual importance of Lie semialgebras roots in
the fact that---in Lie semialgebras---the exponential map of
a zero-neighbourhood in $L(\bS)$ yields a $\unity$"~neighbourhood
in $\bS$.
In contrast, as soon as $\fw$ is merely a Lie wedge that fails
to carry the stronger structure of a Lie semialgebra, there will be
elements in $\bS$ that are arbitrary close to the identity
without belonging to any one-parameter semigroup completely
contained in $\bS$. For more details and a variety of
illustrative examples, we recommend \cite{HHL89} and
\cite{HofRupp97div}, where the respective introduction
does provide a lucid overview of the entire subject.
The connection between Lie semialgebras and time-independent Markovian
quantum channels has been worked out in detail in \cite{DHKS08}.
With these stipulations, the frame is set to describe the time
evolution of Markovian (i.e., memory-less) open quantum
systems in the differential geometric picture of Lie wedges.
\subsection{Markovian Quantum Dynamics and Quantum Channels}
Markovian quantum dynamics is conveniently described by
a linear autonomous differential equation
\begin{equation}\label{eqn:LE}
\dot X(t) = - \mathcal L \, X(t)\;,
\end{equation}
where $X(t)$ usually denotes the state of a quantum system
represented by its density operator $\rho(t)$,
i.e.~$\rho(t)=\rho(t)^\dagger$, $\rho(t) \geq 0$, and
$\tr\rho(t)=1$. Here and henceforth, $(\cdot)^\dagger$
denotes the adjoint (complex-conjugate transpose).
For ensuring complete positivity,
$\mathcal L$ has to be of Lindblad form
\cite{Lind76}, i.e.
\begin{equation}\label{eqn:master-rho-uncontrolled}
\mathcal L(\rho) = i\adr_{H}(\rho)+\Gamma_L(\rho)\;,
\end{equation}
with $\adr_{H_j}(\rho):= [H_j, \rho]$ and
\begin{equation}\label{eqn:GKS}
\Gamma_L(\rho) := \tfrac{1}{2}\sum_kV_k^\dagger V_k\rho
+ \rho V_k^\dagger V_k-2V_k\rho V_k^\dagger\;,
\end{equation}
Here, the {\em Hamiltonian} $H$ is assumed to be a Hermitian
\mbox{$N \times N$} matrix while the {\em Lindblad generators}
$\{V_k\}$ may be arbitrary \mbox{$N\times N$} matrices.
The resulting equation of motion \eqref{eqn:LE}
acts on the vector space of all Hermitian operators,
$\herm(N)$, and more precisely, leaves the set of all
density operators
$\pos_1(N) := \left\{\rho\in\herm(N)\,|\,\rho=\rho^\dagger,\rho\geq0, \tr\rho=1\right\}$
invariant.
\medskip
In \cite{DHKS08} it was shown that the set of all Lindblad
generators $\{-\mathcal{L}\}$ has an interpretation as a particular
Lie wedge. To see this, consider the group lift of \eqref{eqn:LE},
i.e.~now $X(t)$ denotes an element in the general linear group
$GL(\herm(N))$. Moreover, define the set of all completely
positive (cp), trace-preserving invertible linear operators
acting on $\herm(N)$ as $\bP^{cp}$, i.e.
\begin{equation*}
\bP^{cp} := \{T\in GL(\herm(N))\,|\, T \;\text{is cp and trace-preserving}\}
\end{equation*}
and let $\bP^{cp}_0$ denote its connected component of the identity.
Then, $\bP^{cp}$ is exactly the set of so-called invertible
\emph{quantum channels}. A quantum channel $T$ is said to
be \emph{time independent Markovian} or briefly \emph{Markovian},
if it is a solution of \eqref{eqn:LE}. Thus $T = {\rm e}^{-t \mathcal{L}}$
for some fixed Lindblad generators $\mathcal{L}$ and
some $t \geq 0$. Furthermore, $T$ is \emph{time dependent Markovian}
if it is a solution of \eqref{eqn:LE}, where now
$\mathcal{L} = \mathcal{L}(t)$ may vary in time
(for terminology see also \cite{Wolf08a,Wolf08b}).
Finally, we will denote
the set of all {\em time independent Markovian} and
{\em time dependent Markovian} quantum channels
by $\mathbf{MQC}$ and $\mathbf{TMQC}$, respectively.
Then, with regard to the work by Lindblad \cite{Lind76} and
Kossakowski \cite{GKS76}, one obtains the following result
\cite{DHKS08}:
\begin{enumerate}
\item[(a)]
The global {\em Lie wedge} of $\bP^{cp}_0$ is given by the
{\em set of all Lindblad generators} of the form
\begin{equation}
- \mathcal L:= -\big(i\adr_{H} + \Gamma_L\big)
\end{equation}
with $H \in \herm(N)$ and $\Gamma_L$ as in \eqref{eqn:GKS}.
\item[(b)]
The Lie semigroup
\begin{equation}
\overline{\expt{\exp L(\bP^{cp}_0)}}_S
\end{equation}
clearly contains $\mathbf{MQC}$ and moreover it exactly coincides with the closure of
$\mathbf{TMQC}$
thus excluding the {\em non}\/-Markovian ones in $\bP^{cp}_0$,
which is most remarkable.
\end{enumerate}
\noindent
While assertion (a) reformulates previous results by Lindblad
and Kossakowski \cite{Koss72,GKS76,Lind76}, part (b) is noteworthy
as it also says that $\bP^{cp}_0$ is {\em not} a Lie subsemigroup of
$GL(\herm(N))$.
\subsection{Coherently controlled Master Equations}
\label{subsec:notions}
\emph{Controlled} Markovian quantum dynamics is appropriately
addressed as right-invariant {\em bilinear control system}
\cite{DHKS08,dAll08,DiHeGAMM08,Elliott09}
\begin{equation}\label{eqn:sigma}
\dot \rho(t) = -\mathcal L_{u(t)}\big(\rho(t)\big)\quad,
\quad \rho(0) \in \pos_1(N)\;,
\end{equation}
where $\mathcal{L}_u$ now depends on some control variable
$u \in \R{}^m$.
Here, we focus on \emph{coherently controlled}
open systems. This means that $\mathcal{L}_u$ has the
following special from
\begin{equation}\label{eqn:master-rho}
\mathcal{L}_u(\rho) = -i\adr_{H_u}(\rho)-\Gamma_L(\rho)\quad\text{with}
\end{equation}
\begin{equation}\label{eqn:controlHam}
\adr_{H_u} := \adr_{H_d}+\sum_{j=1}^mu_j\adr_{H_j}\;.
\end{equation}
Note that the control terms $i\adr_{H_j}$ with {\em control Hamiltonians}
$H_j \in \herm(N)$ are usually switched by piecewise constant
{\em control amplitudes} $u_j(t)\in\R{}$.
The drift term of \eqref{eqn:master-rho} is then composed of two parts, (i)
the term $i\adr_{H_d}$ (in abuse of language sometimes called
\/`Hamiltonian\/' drift) accounting for the coherent time evolution
and (ii) a dissipative Lindblad part $\Gamma_L$.
So $\mathcal{L}_u$ denotes the {\em coherently controlled Lindbladian}.
As in the uncontrolled case,
system \eqref{eqn:sigma} acts on
the vector space of all Hermitian operators leaving the set of all
density operators invariant. Equivalently, one can regard
\eqref{eqn:sigma} as an affine system on
$\mathfrak{her}_0(N) := \{ H \in \herm(N)\;|\; \tr H = 0\}$.
In the following, we further impose {\em unitality}, i.e.
we assume $\Gamma_L(\unity)=0$. This ensures that
\eqref{eqn:sigma} actually yields a bilinear control system on
$\mathfrak{her}_0(N)$ instead of an affine one.
Therefore, it allows a group lift to
$GL\big(\mathfrak{her}_0(N)\big)$ which henceforth is
referred to as ($\Sigma$), i.e.
\begin{equation}\label{eqn:SIGMA}
(\Sigma)\quad
\dot X(t) = -\mathcal L_{u(t)} \, X(t) \,,\;
X(0) \in GL\big(\mathfrak{her}_0(N)\big)\;.
\end{equation}
The corresponding group lift in the affine case is more involved
\cite{DiHeGAMM08,DHKS08}.
Now, the {\em system semigroup} $\bP_\Sigma$ associated to ($\Sigma$)
reads
\begin{eqnarray}\label{eqn:Psemi}
\bP_\Sigma=\langle T_u(t)=
\exp(-t \mathcal L_u)\,|\,t\geq 0, u \in \R{}^m \rangle_S
\end{eqnarray}
and lends itself to exemplify the notion of a Lie wedge.
To distinguish between different notions of controllability in open
systems, we define three algebras: the \emph{control algebra} $\fk_c $,
the \emph{extended algebra} $\fk_d$, and the \emph{system algebra}
$\fs$ as follows
\begin{equation}\label{eqn:HamL-closure}
\begin{split}
\fk_c
& := \expt{i\adr_{H_j}\,|\, j=1,\dots, m}_{\sf Lie}\;,\\[2mm]
\fk_d
& := \expt{i\adr_{H_d}, i\adr_{H_j}\,|\, j=1,\dots, m}_{\sf Lie}\;,\\[2mm]
\fs
&:= \expt{\mathcal L_u|\, u_j\in\mathbb R }_{\sf Lie}\\[2mm]
&= \expt{i\adr_{H_d}+\Gamma_L, i\adr_{H_j}|\, j=1,\dots, m}_{\sf Lie}\;.
\end{split}
\end{equation}
Note that $\fs$ is different from $\fk_d$, because it
contains the entire
drift term $(i\adr_{H_d}+\Gamma_L)$ for the Lie closure,
while $\fk_d$ only takes its Hamiltonian component $i\adr_{H_d}$.
Then $(\Sigma)$ is said to fulfill condition (H), (WH), and (A),
respectively, if
\begin{eqnarray}
&(H) \qquad
\fk_c &=\; \adr_{\su(N)} \\
&(W\negthinspace H)\quad
\fk_d &=\; \adr_{\su(N)} \;\text{while}\;\;\fk_c \neq \adr_{\su(N)}\\
&(A)\qquad
\fs &=\; \gl(\herm_0(N))
\end{eqnarray}
While condition (A) respects a standard construction of
non-linear control theory \cite{Jurdjevic97,JS72} to express
accessibility, conditions (H) and (WH)
serve to characterize different types of {\em controllability}
of the Hamiltonian part of ($\Sigma$) in the absence
of relaxation: Condition (H) says that the Hamiltonian part is
fully controllable even {\em without} resorting to the drift
Hamiltonian, whereas condition (WH) yields full controllabilty
of the Hamiltonian part with the drift Hamiltonian being {\em necessary}.
We refer to the first scenario as \emph{(fully) $H$-controllable}
and to the second as satisfying the (WH)-condition.
Generically, open systems ($\Sigma$) given by \eqref{eqn:SIGMA}
meet the accessibility condition (A) \cite{Alt04,Diss-Indra}.
Finally, note that via
$\e^{i\adr_{H}}(\rho) = \e^{i H} \rho\, \e^{-iH}$
the Lie algebra $\adr_{\su(N)}$ generates the Lie group
$\Adr_{SU(N)} \iso PSU(N)$ here acting on $\herm_0(N)$
by conjugation.
\subsection[Lie Wedges for Coherently Controlled GKS"~Master
Equations]{Computing Lie Wedges for Controlled
Master Equations}
\label{subsec:LW-comp}
Here, the goal is to determine the (global) Lie wedge
of a coherently controlled unital open system ($\Sigma$) given in terms
of its Markovian master equation \eqref{eqn:SIGMA} of GKS"~Lindblad
form.
In view of the examples worked out in detail in Sec.~\ref{sec:geoR3},
here we sketch how to approximate a Lie wedge of a controlled
Markovian systems in two ways, (i) by an {\em inner approximation}
and (ii) by an {\em outer approximation} thus following
\cite{Lawson99,DHKS08}.
Moreover for unital systems, we present two results
which guarantee that the {\em inner approximation} is
global and thus coincides with the Lie wedge
$L(\overline{\bP}_\Sigma)$ sought for.
Let ($\Sigma$) be a unital open control system as in
\eqref{eqn:SIGMA} where, for simplicity, the system algebra
$\fs$ fulfills the accessibility condition (A).
Moreover, let
\begin{equation}
\Omega_\Sigma := \{\cLu | u \in \R{}^m\} \subset
\gl(\mathfrak{her}_0(N))
\end{equation}
be the set of all directions specified by \eqref{eqn:SIGMA}.
The {\em reachable set} $\Reach(\Omega_\Sigma,\unity)$
of ($\Sigma$) is defined
as the set of all states $X(T)$, $T\geq 0$ that can be reached
from the unity $X(0)=\unity$ under the dynamics of ($\Sigma$),
while the controls $u(t) \in \R{}^m$ are assumed to be
piecewise constant functions. In general, one could
allow for larger classes of {\em admissible} controls,
such as locally bounded or locally integrable ones. Yet, the closure
of the corresponding reachable sets will not differ
\cite{JS72,S83,Elliott09}.
Clearly, $\Reach(\Omega_\Sigma,\unity)$ takes the form of a
subsemigroup within the embedding Lie group $GL(\mathfrak{her}_0(N))$
in the sense of Sec.~\ref{sec:LieSemiBasics}. For instance,
restricting the control amplitudes $\{u_j\}$ to be piecewise constant yields the equality
$\Reach(\Omega_\Sigma,\unity) = \bP_\Sigma$.
More generally,
the following result holds.
\medskip
\begin{theorem}[\cite{Lawson99}]\label{thm:L-saturate}
Let $\bP_\Sigma$ be defined as in \eqref{eqn:Psemi}.
Then
\begin{equation}\label{reach}
\overline{\bP}_\Sigma \;=\; \overline{\Reach(\Omega_\Sigma,\unity)}
\;=\; \overline{\Reach(L(\bP_\Sigma),\unity)}\;.
\end{equation}
In particular, $\overline{\bP}_\Sigma$ is a Lie subsemigroup.
Furthermore, $L(\overline{\bP}_\Sigma)$ is the smallest global
Lie wedge containing $\Omega_\Sigma$ as well as
the largest subset $\Omega'$ of $\gl(\mathfrak{her}_0(N))$
which satisfies the equality
\begin{equation}
\overline{\Reach(\Omega',\unity)} \;=\;
\overline{\Reach(\Omega_\Sigma,\unity)}\;.
\end{equation}
Due to the last property,
the Lie wedge $L(\overline{\bP}_\Sigma)$
is also called the {\em Lie saturate} of $\Omega_\Sigma$,
cf.~\cite{JurdKupka81a,JurdKupka81b,Lawson99}.
\end{theorem}
\medskip
Unfortunately, for an arbitrary system ($\Sigma$), currently no
procedure is known to explicitly determine its global Lie wedge.
Yet there is a straightforward strategy to compute
an {\em inner approximation} \cite{Lawson99,DHKS08}.
It consists of the following steps:
\begin{enumerate}
\item[(1)]
form the smallest closed convex cone $\fw$ containing $\Omega_\Sigma$;
\item[(2)]
compute the edge $E(\fw)$ of the wedge and the
smallest Lie algebra $\fe$ containing $E(\fw)$,
i.e.~$\fe :=\langle E(\fw) \rangle_{\rm Lie}$;
\item[(3)] make the wedge invariant under the $\Adr$"~action
of $\mathfrak{e}$ by forming the set
$\bigcup_{A \in \mathfrak{e}} \Adr_{\exp A}(\fw)$;
\item[(4)]
update by taking the convex hull $\conv\{S\}$ of the set $S$ obtained in step (3);
\item[(5)]
repeat steps (2) through (4) until nothing new is added:
the resulting final wedge $\fw_0$ is henceforth referred
to as {\em inner approximation} to the global Lie wedge
$L(\overline{\bP}_\Sigma)$.
\end{enumerate}
Now, the crucial question arises whether the inner approximation
$\fw_0$ is global or not. If it is global, Theorem~\ref{thm:L-saturate}
guarantees that $\fw_0$ is equal to $L(\overline{\bP}_\Sigma)$.
Next we present two results which proved quite helpful to decide
the globality problem: The first one
yields a global {\em outer approximation} $\fw^0$ of
$L(\overline{\bP}_\Sigma)$.
Combining inner and outer approximation, the Lie wedge
$L(\overline{\bP}_\Sigma)$ sought for can be determined via the
inclusions
\begin{equation}
\fw_0 \subseteq L(\overline{\bP}_\Sigma) \subseteq \fw^0.
\end{equation}
Clearly, if the outer and inner approximations coincide, one is done.
The second one
based on the so-called \emph{Principal Theorem of Globality}
from \cite{HHL89} (see also Appendix~A)
provides
a \/`direct\/' method for proving globality. It will be the key
tool to show that the inner approximations given in the worked
examples of Secs.~\ref{sec:geoR3} and \ref{sec:q-channels} are
in fact {\em global} Lie wedges.
\medskip
\begin{theorem}[\cite{DHKS08}]\label{thm:LW-outer}
Let ($\Sigma$) be a unital controlled open system
as in \eqref{eqn:SIGMA}. If there exists a \emph{pointed}
cone $\mathfrak{c}$ in the set of all positive semidefinite
operators that act on $\mathfrak{her}_0(N)$ so that
\begin{enumerate}
\item[(1)] $\Gamma_L\in\fc$\\[-3mm]
\item[(2)] $[\fc,\fc]\subset\adr_{\su(N)}$\\[-3mm]
\item[(3)] $[\fc,\adr_{\mathfrak{su}(N)}]\subset (\fc-\fc)$\\[-3mm]
\item[(4)] $\Adr_U\fc\Adr_{U^\dagger}\subset \fc$ for all $U\in SU(N)$,
\end{enumerate}
then the subsemigroup associated to ($\Sigma$) follows the inclusion
$\overline{\bP}_\Sigma\subseteq\Adr_{SU(N)}\cdot\exp(-\mathfrak{c})$
and hence its Lie wedge obeys the relation
$L(\overline{\bP}_\Sigma)\subseteq\adr_{\mathfrak{su}(N)}\oplus(-\mathfrak{c})$,
i.e.~$\adr_{\su(N)}\oplus(-\fc)$ is a
global outer approximation to $L(\overline{\bP}_\Sigma)$.
\end{theorem}
\medskip
\begin{corollary}(\cite{DHKS08,Alt03})\label{HcontrolGlobalWedge}
Let ($\Sigma$) be a unital
single-qubit system satisfying condition (H)
with a generic\footnote{In \cite{DHKS08} Cor.~\ref{HcontrolGlobalWedge}
is stated under the above genericity
assumption; yet one can drop this additional condition.}
Lindblad term $\Gamma_L$.
Then $\overline \bP_{\Sigma}=\Adr_{SU(2)}\cdot\exp(-\fc)$,
where the cone
\begin{eqnarray*}
\fc:=\R{}_0^+\conv\left\{\Adr_U\,\Gamma_L\,\Adr_{U^\dagger} \;|\;
U\in SU(2)\right\}
\end{eqnarray*}
is contained in the set of all positive semidefinite elements in
$\mathfrak{gl}(\mathfrak{her}_0)$. Furthermore
$L(\overline{\textbf{P}}_{\Sigma})=
\adr_{\mathfrak{su}(2)}\oplus(-\mathfrak{c})$.
\end{corollary}
\medskip
\begin{theorem}\label{thm:globality-3}
Let ($\Sigma$) be a unital controlled open system
given by \eqref{eqn:SIGMA}. In addition assume that
($\Sigma$) meets the accessibility condition (A)
and that the Lie subgroup $\bK$ which corresponds
to the control algebra $\fk_c$ is closed within $SU(N)$. Then,
$\fw := \fk_c \oplus (-\fc)$ is a {\em global} Lie wedge in
$\gl(\mathfrak{her}_0(N))$, where
$
\fc :=
\R{}_0^+\conv \big\{
\Adr_{U}\,\big(\adr_{{\rm i}H_d}+\Gamma_L\big)\,\Adr^\dagger_{U}
\;|\; U \in \bK \big\}\;.
$
Moreover, $\fw$ is the global Lie wedge of ($\Sigma$), i.e.
$\fw = L(\overline{\bP}_\Sigma)$.
\end{theorem}
\begin{proof}(Sketch)
The full proof will be given elsewhere in a more general context.
For applying the \/`Principal Theorem of Globality\/' \cite{HHL89}
(see Appendix~A),
the following steps have to be established:
\begin{enumerate}
\item[(1)]
The edge of $\fw$ coincides with $\fk_c$.
\item[(2)]
$\fw$ is a Lie wedge in $\fg := \gl(\mathfrak{her}_0(N))$.
\item[(3)]
There exists a function $\varphi:GL(\mathfrak{her}_0(N)) \to \R{}$
such that its differential satisfies
${\rm d}\varphi(X) \, AX \geq 0$ for all
$X \in GL(\mathfrak{her}_0(N))$ and all $A \in \fw$.
\item[(4)]
The differential of $\varphi$ fulfills
${\rm d}\varphi(\unity) \, A > 0$ for all
$A \in \fw\setminus E(\fw)$.
\end{enumerate}
Note that step (3) is the essential one, and an appropriate candidate
for $\varphi$ is given by
$
X \mapsto - \langle X,X \rangle
:=
- \sum_{k=1}^{N^2-1}\tr\big(X(B_k)X(B_k)\big)
\;,
$
where $B_1, \dots, B_{N^2-1}$ is any orthonormal basis of
$\mathfrak{her}_0(N)$.
\end{proof}
\medskip
\noindent
As a useful tool,
we add the following Corollary,
which is put into a broader context in Appendix~A:
\medskip
\begin{corollary}[\cite{HHL89}]\label{HHLglobalcor}
Let $\bG$ be a Lie group with Lie algebra $\fg$ and let
$\fw_0\subseteq\fw$ be two Lie wedges in $\fg$.
Provided one has
$
\fw_0\setminus-\fw_0\subseteq\fw \setminus -\fw
$ [or equivalently $E(\fw_0)=E(\fw)\cap\fw_0$],
then $\fw_0$ is global in $\bG$ if the following conditions are
satisfied:
(i)
$\fw$ is global in $\bG$;
(ii)
the edge of $\fw_0$
is the Lie algebra of a closed Lie subgroup of $\bG$.
\end{corollary}
\subsection*{Guideline through Applications}
For illustrating the power of the Lie-semigroup formalism by
applications, we follow a two-fold route: Sec.~\ref{sec:geoR3}
addresses three paradigmatic types of bilinear control systems on
$\mathbb R^3$, where the control parts of the dynamics generate
easy-to-visualise rotations in $SO(3)$.
{\em Thus Sec.~\ref{sec:geoR3} is meant to be readable without any background in
quantum mechanics}, yet it directly corresponds to single-qubit systems
undergoing relaxation as the presented examples coincide with the
so-called {\em coherence-vector representation} of such
systems \cite{AlickiLendi87}.
Therefore, the results obtained in Sec.~\ref{sec:geoR3} can readily be transferred
to Sec.~\ref{sec:q-channels}, where we address quantum channels
in the customary explicit $\su(2)$"~representation of qubits. By the isomorphism
$\so(3)\iso\su(2)$, the geometry in Sec.~\ref{sec:geoR3} thus illustrates
key results in Sec.~\ref{sec:q-channels} for qubit channels.
\section{Geometry of Open Systems in $\mathbb R^3$}\label{sec:geoR3}
In this section, we discuss three simple introductory examples of \/`open\/'
systems, the geometry of which can be envisaged as rotations in $\mathbb R^3$
concomitant to relaxation. To fix notations, define the following
\begin{equation}\label{eqn:R3rot_gen}
H_x:=\left[\begin{smallmatrix} 0 &0 &0 \\ 0 &0 &-1 \\ 0 &1 &0 \end{smallmatrix}\right]\;,\quad
H_y:=\left[\begin{smallmatrix} 0 &0 &1 \\ 0 &0 &0 \\ -1 &0 &0 \end{smallmatrix}\right]\;,\quad
H_z:=\left[\begin{smallmatrix} 0 &-1 &0 \\ 1 &0 &0 \\ 0 &0 &0 \end{smallmatrix}\right]\;\quad
\end{equation}
as generators of the rotations $R_\nu(\theta):={\rm e}^{\theta H_\nu}$ reading
\begin{equation}\label{eqn:R3rots}
\begin{split}
R_x(\theta) &:=\left[\begin{smallmatrix} 1 &0 &0 \\ 0& \ct &-\st \\ 0 &\st &\ct \end{smallmatrix}\right],\;
R_y(\theta) :=\left[\begin{smallmatrix} \ct &0 &\st \\ 0 &1 &0 \\ -\st &0 &\ct \end{smallmatrix}\right],\\[2mm]
R_z(\theta) &:=\left[\begin{smallmatrix} \ct &-\st &0 \\ \st &\ct &0 \\ 0 &0 &1 \end{smallmatrix}\right]\;.
\end{split}
\end{equation}
So we have $\expt{H_x,H_y,H_z}_{\sf Lie}=\so(3)$ and thereby a basis for the skew-symmetric matrices
forming the $\fk$-part in the Cartan decomposition
$\gl(3,\mathbb R)=\so(3)\oplus\mathfrak{sym}(3)$, where the $\fp$-part is spanned
by the symmetric matrices
\begin{equation}\label{eqn:R3sym_gen}
p_x:=\left[\begin{smallmatrix} 0 &0 &0 \\ 0 &0 &1 \\ 0 &1 &0 \end{smallmatrix}\right]\;,\quad
p_y:=\left[\begin{smallmatrix} 0 &0 &1 \\ 0 &0 &0 \\ 1 &0 &0 \end{smallmatrix}\right]\;,\quad
p_z:=\left[\begin{smallmatrix} 0 &1 &0 \\ 1 &0 &0 \\ 0 &0 &0 \end{smallmatrix}\right]\;\quad
\end{equation}
and the diagonal $3 \times 3$-matrices $E_{ii} := e_ie_i^\top$ for $i=1,2,3$.
Recall that the skew-symmetric $\fk$-part and a symmetric $\fp$-part
obeying the usual commutator relations
$\comm \fk \fk \subseteq \fk$,
$\comm \fk \fp \subseteq \fp$ and $\comm \fp \fp \subseteq \fk$.
For later convenience, we note commutation relations for the above
basis in Tab.~\ref{tab:comm-tab}.
\subsection[{\em Example~1:} Fully H-Controllable System with General Relaxation Operator]
{{\bf Example~1}: Corresponds to a Qubit System with
Condition (H) Satisfied and General Relaxation Operator}
Using definitions from above, consider the control system
in $GL(3,\mathbb R)$ given by the equation
\begin{eqnarray}\label{eq:ex1}
\dot X=-(A+B_u)X \;,
\end{eqnarray}
where the control term $B_u:=u_xH_x+u_yH_y$ shall have independent
controls $u_x, u_y \in \R{}$, and the drift term
$A:=H_z+\Gamma_0$ is composed of a \/`Hamiltonian\/' component, $H_z$, and
a relaxation component given by the matrix
\begin{equation}\label{eqn:gamma-abc}
\Gamma_0:= \diag(a,b,c)
\end{equation}
with relaxation-rate constants $a,b,c\geq 0$.
Since $\expt{H_x,H_y}_{\sf Lie}=\so(3)$, system \eqref{eq:ex1}
satisfies in fact condition (H)
in the sense of Sec.~\ref{subsec:notions}
(i.e.~without resorting to the drift component $H_z$).
For explicitly computing the Lie wedge of \eqref{eq:ex1} we proceed
as in Sec.~\ref{subsec:LW-comp}. For the following calculations observe that
$H_x,H_y,H_z$ belong to the $\fk$-part, while $\Gamma_0$ is contained in
the $\fp$-part of \/`the\/' Cartan decomposition of $\gl(3,\R{})$.
\begin{table}[Ht!]
\caption{\label{tab:comm-tab}Commutation Table}
\begin{tabular}{c cccccc}
\hline\hline\\[-1.5mm]
$[H_\nu,(\cdot)]$ & $E_{11}$ & $E_{22}$ & $E_{33}$ & $p_x$ & $p_y$ & $p_z$\\[1mm]
\hline\\[-1mm]
$H_x$ &0 &$p_x$ &$-p_x$ &$-2\Delta_{23}$ &$-p_z$ &$p_y$ \\[0.5mm]
$H_y$ &$-p_y$ &$0$ &$p_y$ &$p_z$ &$-2\Delta_{31}$ &$-p_x$\\[0.5mm]
$H_z$ &$p_z$ &$-p_z$ &$0$ &$-p_y$ &$p_x$ &$-2\Delta_{12}$ \\[0.5mm]
\hline\hline\\[-1.5mm]
\multicolumn{3}{l} \mbox{define $\Delta_{ij}:=E_{ii}-E_{jj}$}
\end{tabular}
\end{table}
\medskip
\noindent
Step (1) of the algorithm gives the initial wedge approximation
\begin{eqnarray}
\fw_1 := \R{}H_x \oplus \R{}H_y \oplus (-\fc_1)\;,
\end{eqnarray}
where $\fc_1 := \R{}^+_0 \big(H_z+\Gamma_0\big)$.
In step (2) one then readily finds
\begin{equation}
E(\fw_1) = \R{}H_x \oplus \R{}H_y
\end{equation}
so $\fe = \langle H_x, H_y\rangle_{\rm Lie} = \so(3)$.
Hence, step (3) and (4) give
\begin{equation}
\fw_0 = \so(3)\;\oplus\; \mathbb R^-_0\,
\conv\mathcal O_{SO(3)}(\Gamma_0) \;,
\end{equation}
where
$\mathcal O_{SO(3)}(\Gamma_0)
:=\{\Theta \Gamma_0 \Theta^\top\,|\, \Theta\in SO(3)\}$
denotes the {\em orthogonal orbit} of $\Gamma_0$. Here we used
the trivial fact that $\so(3)$ is $\Adr_{SO(3)}$-invariant.
By a well-known result of Uhlmann\footnote{The result mentioned
is originally stated for density matrices and $SU(N)$.
However, the proof in \cite{Ando89} immediately carries over
to symmetric matrices and $SO(N)$ \cite{Uhlm71,Ando89,MarshallOlkin}.}
the convex hull of the isospectral set $\mathcal O_{SO(3)}(\Gamma_0)$
simplifies to
\begin{equation}
\conv \mathcal O_{SO(3)} (\Gamma_0) =
\{S \in \mathfrak{sym}(3) \,|\, S \prec \Gamma_0\} =: \mathcal M(\Gamma_0)
\;.
\end{equation}
Defining the pointed convex cone
$\fc_0 := \mathbb R^+_0\,\mathcal M(\Gamma_0)$, we obtain
\begin{equation}
\fw_0 = \so(3)\;\oplus\; (-\fc_0)
\;,
\end{equation}
as final inner approximation to the global Lie wedge
of \eqref{eq:ex1}.
\medskip
\begin{lemma}\label{ex1edgeprop}
The set $\fw_0$ is a Lie wedge of $\mathfrak{gl}(3,\mathbb{R})$.
Its edge $E(\fw_0)$ is given by $\mathfrak{so}(3)$.
\end{lemma}
\medskip
\noindent
\begin{proof} It suffices to show
that the edge of $\fw_0$ is given by $\mathfrak{so}(3)$.
Then the invariance of $\fw_0$ under the $\Adr$-action
of $E(\fw_0)$ is obviously guaranteed by construction.
Clearly, one has the inclusion $\mathfrak{so}(3) \subset E(\fw_0)$.
Conversely, let $W \in E(\fw_0)$.
Then, $W=A+B$ with $A\in\mathfrak{so}(3)$ and $B\in\fc_0$.
Since $-W\in E(\fw_0)$, there exits $A'\in\mathfrak{so}(3)$
and $B'\in\fc_0$ such that $-W= A'+B'$. Hence
$A+B=-(A'+B')$ and thus $A+A'=-(B+B')$. Since $\fc_0\in\mathfrak{sym}(3)$, it
is trivial that $\mathfrak{so}(3)\cap\fc_0=\{0\}$. Therefore,
$A+A' = -(B+B')= 0$ and hence $A'=-A$ and $B'=-B$.
Now, $B,-B\in\fc_0=\mathbb R^+_0\,\mathcal M(\Gamma_0)$.
But $\mathcal{M}(\Gamma_0)$ is contained in the set of all positive
semidefinite matrices and therefore we conclude $B=0$. Thus we
obtain $E(\fw_0)=\mathfrak{so}(3)$.
\end{proof}
\medskip
\begin{proposition}\label{example1globality}
The set $\fw_0 = \so(3)\oplus (-\fc_0)$ is
the {\em global} Lie wedge to control system \eqref{eq:ex1}.
\end{proposition}
\medskip
\begin{proof}
This is an immediate consequence of Theorem~\ref{thm:globality-3}
and the fact that \eqref{eq:ex1} comes from a unital GKS"~Lindblad
master equation in the coherence-vector representation.
\end{proof}
\medskip
\begin{remark}
Alternatively to the above proof, one could apply Corollary
\ref{HcontrolGlobalWedge}, because $\fw_0$ is a matrix representation
of the global Lie wedge described therein.
\end{remark}
\medskip
For general $\Gamma_0=\diag(a,b,c)$, the Lie wedge $\fw_0$ in
Example~1 does {\em not} carry the special structure of a Lie
{\em semialgebra}, cf.~Sec.~\ref{sec:LieSemiBasics}. This can be shown
by choosing a suitable $A' \in \fw_0$ which violates the
inclusion $[A', T_{A'}\fw_0] \subset T_{A'}\fw_0$.
--- In contrast, for the case\footnote{Note that this case
exactly corresponds to the (isotropic) depolarising quantum channel
discussed in Sec.~\ref{sec:q-channels}.}
$\Gamma_0=\lambda\cdot\unity$,
indeed we obtain a
Lie semialgebra, because the BCH-product $A \star B$ obviously stays
inside $\fw_0$, whenever $A,B \in \fw_0$. Further details and
proofs are given in Appendix~B.
\subsection[{\em Example~2:} System Satisfying (WH)-Condition with Invariant Relaxation Operator]
{{\bf Example~2}: Corresponds to a Qubit System with Condition (WH) Satisfied
and Control Invariant Relaxation Operator}
Consider the control system in $GL(3,\R{})$ given by
\begin{eqnarray}
\label{eq:ex2}
\dot X=-(A+B_u)\,X\;,
\end{eqnarray}
where the control term $B_u$ is of the form $B_u:=u H_y$
with $u\in \R{}$ and the drift term
$A:=\Gamma_0+H_z$ is composed of a \/`Hamiltonian\/' part, $H_z$, and
a relaxation part
\begin{eqnarray}\label{eq:drift2}
\Gamma_0:=\gamma\,\diag(1,0,1)
\end{eqnarray}
with $\gamma\geq0$. Since
$\langle H_y,H_z\rangle_{\sf Lie}=\mathfrak{so}(3)$
the system \eqref{eq:ex2} fulfills
condition (WH) in the sense of Sec.~\ref{subsec:notions}
but obviously not condition (H).
\medskip
Now in step (1) of the inner approximation procedure to the global
Lie wedge of \eqref{eq:ex2} one finds
\begin{equation}
\fw_1 = \mathbb R H_y \oplus (-\fc_1)
\end{equation}
with $\fc_1 := \mathbb R_0^+ (H_z + \Gamma_0)$ whose edge is given by
\begin{equation}
E(\fw_1)=\mathbb{R}H_y \;.
\end{equation}
In step (3) we include elements obtained by
conjugations generated by edge elements identified
in step (2), i.e.\ elements of the form
\begin{equation}\label{dissipativeExample2}
{\rm e}^{\theta H_y} \,\fc_1\, {\rm e}^{-\theta H_y}
=\lambda\;(\cos(\theta)H_z+\sin(\theta)H_x+\Gamma_0)\qquad
\end{equation}
for $\theta\in\R{}$ and $\lambda\geq 0$.
By orthogonality \mbox{$\langle\Gamma_0|H_\nu\rangle=0$}
for $\nu=x,y,z$ one readily gets a Hilbert space
$\mathcal{H}:=\mbox{span}\left\{H_x,H_z,\Gamma_0\right\}$,
in which the edge-invariant cone elements of \eqref{dissipativeExample2}
can be expanded using the following short-hand
\begin{eqnarray}\label{eqn:short-hand-scalprod}
\lambda\big(\sin(\theta)H_x+\cos(\theta)H_z+\Gamma_0\big) =:
\lambda\left[\begin{smallmatrix}\sin(\theta)\\ \cos(\theta)\\1\end{smallmatrix}\right]
\cdot\left[\begin{smallmatrix}H_x\\H_z\\\Gamma_0\end{smallmatrix}\right]\;.
\end{eqnarray}
Then its convex hull gives the final cone
\begin{eqnarray}\label{eqn:Ex2-cone}
\fc_0:= \R{}^+_0\conv \left\{
\left[\begin{smallmatrix}\sin(\theta)\\ \cos(\theta)\\1\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix}H_x\\H_z\\\Gamma_0\end{smallmatrix}\right]
\;|\; \theta \in \R{}\right\}
\end{eqnarray}
--- a classical $3$-dimensional \/`ice cone\/', cf.~Fig.~\ref{fig:Ex2}(a).
By construction, $\fc_0$ remains $\Adr_{\exp E(\fw)}$-invariant.
Finally, since $H_y\perp\fc_0$,
the Lie wedge itself admits the orthogonal decomposition
\begin{eqnarray}\label{eqn:Ex2-cone-a}
\fw_0:=\mathbb{R}H_y\oplus (-\fc_0)\;.
\end{eqnarray}
\begin{figure}
\begin{center}
\raisebox{35mm}{{\sf (a)}}
\includegraphics[scale=.9]{Ex2a}\\[2mm]
\raisebox{35mm}{{\sf (b)}}
\includegraphics[scale=.9]{Ex2b}
\end{center}
\caption{\label{fig:Ex2} The Lie wedge for the system of {\bf Example~2}
satisfying the (WH)-condition is four-dimensional: it contains (a) the convex cone $\fc_0$ (see
\eqref{eqn:Ex2-cone} with $\lambda=1$)
generated by $e^{\theta H_y}(\Gamma_0+H_z)e^{-\theta H_y}$ and shown in the projection into
the subspace spanned by $\{H_x,H_z,\Gamma_0\}$; and it comprises (b) the
prism-shaped wedge projected into the subspace spanned by $\{H_y,H_z,\Gamma_0\}$.
Note that the {\em edge of the wedge} is spanned by the
control Hamiltonian $H_y$. Both parts of the figure scale with $\lambda\to\infty$.
}
\end{figure}
\medskip
\begin{proposition}\label{example2weakglobality}
The set $\fw_0 = \mathbb{R}H_y\oplus(-\fc_0)$ is
the {\em global} Lie wedge to control system \eqref{eq:ex2}.
\end{proposition}
\medskip
\begin{proof}
The Lie wedge property of $\fw_0$ can be derived as in Example 1.
Then the globality of $\fw_0$ follows again from Theorem
\ref{thm:globality-3}.
\end{proof}
\begin{remark}
For clarity, let us denote the Lie wedge of Example~1
for $\Gamma_0$ as in \eqref{eq:drift2} by $\fw'_0$
while $\fw_0$ still refers to the Lie wedge of Example~2.
Clearly, one has $\fw_0\subset\fw'_0$ and
$E(\fw_0) = E(\fw'_0)\cap \fw_0$.
Hence globality of $\fw_0$ also follows
by Corollary \ref{HHLglobalcor} and
the globality of $\fw'_0$.
\end{remark}
\medskip
Note that the Lie wedge $\fw_0$ in Example~2 does {\em not} specialise
to the form of a Lie {\em semialgebra} as can readily be verified
by a counter example:
According to \eqref{eqn:Ex2-cone},
choose $B\in\fw_0$ as $B=\Gamma_0+H_x$
(recalling $\Gamma_0:=E_{11}+E_{33}$ and $A=\Gamma_0+H_z$).
Then by the commutator relations of Tab.~\ref{tab:comm-tab},
the BCH product
\begin{equation}
A\star B = 2\Gamma_0 + H_x + H_z + \tfrac{1}{2}(H_y + p_x + p_z)\, +\, \dots
\end{equation}
immediately leads outside the Lie wedge $\fw_0$, e.g.,
by the non-vanishing component $p_x + p_z$.
(NB: This argument can be made rigorous by introducing a scaling factor $t$
to give $tA\star tB$).
Finally, the edge of $\fw_0$ in Example~2
is $E(\fw_0)=\mathbb{R}H_y$ (for $\gamma>0$, see Fig.~\ref{fig:Ex2})
while in the limit of a closed system, i.e.\ for $\gamma=0$, it turns into
the entire Lie algebra $\mathfrak{so}(3)$.
\subsection[{\em Example~3:} System Satisfying (WH)-Condition with General Diagonal Relaxation Operator]
{{\bf Example~3}: Corresponds to a Qubit System with Condition
(WH) Satisfied and General Diagonal Relaxation Operator}
Consider the contol system in $GL(3,\R{})$ given by
\begin{eqnarray}\label{eq:ex3}
\dot X=-(A+B_u)X\;,
\end{eqnarray}
where $A:=\Gamma_0+H_z$, $B:=uH_y$, and
\begin{equation}
\Gamma_0 := \gamma\,\diag(1,1,2)
\end{equation}
with $u\in\mathbb R$ and $\gamma\geq 0$.
So for approximating the corresponding Lie wedge, we take the first
step to be
\begin{equation}
\fw_1 := \mathbb R H_y \oplus (-\fc_1)\;,
\end{equation}
with $\fc_1:=\mathbb R_0^+ (H_z + \Gamma_0)$ and edge given by the
span of $H_y$ --- the control \/`Hamiltonian\/'.
Again, in step (2) we identify the conjugation to be brought about by
$\Adr_{\exp E(\mathfrak{w})}$ acting on the drift terms such as
to give in step (3) the set
\begin{eqnarray}
\fc_3 = \mathbb R_0^+ \{R_y(\theta)(\Gamma_0+H_z)R_y(\theta)^\top\;|\; \theta\in \mathbb R\}
\end{eqnarray}
with the $\fk$"~component brought about by the conjugated drift
\begin{equation}
R_y(\theta) H_z R_y(\theta)^\top = \cos(\theta)H_z+\sin(\theta)H_x
\end{equation}
and the $\fp$-component reading
\begin{equation}\label{eqn:p-Ex3}
\begin{split}
R_y &(\theta) \Gamma_0 R_y(\theta)^\top \;=\;
\gamma \left[\begin{smallmatrix} 1+\sin^2(\theta) && 0 && \sin(\theta)\cos(\theta)\\
0 && 1 && 0\\
\sin(\theta)\cos(\theta) && 0 &&1+\cos^2(\theta)\end{smallmatrix}\right] \\[2mm]
& = \Gamma_0 + \tfrac{\gamma}{2} \Delta_{13}
-\tfrac{\gamma}{2}\big(\cos(2\theta)\Delta_{13} -\sin(2\theta) p_y\big) \\[2mm]
&=\tfrac{11+\cos(2\theta)}{12}\Gamma_0 + \tfrac{\gamma}{2}\big((1-\cos(2\theta))\Delta+\sin(2\theta)p_y\big)\;,\\
\end{split}
\end{equation}
where the matrices $\Delta_{ij}$ and $p_y$ are defined as in
Tab.~\ref{tab:comm-tab}, and, for the sake of orthogonality,
$\Delta:=\tfrac{2}{9}\unity+\tfrac{1}{18}\Delta_{12}+\tfrac{8}{9}\Delta_{13}$.
\medskip
Therefore, the Lie wedge can be expanded within the five-dimensional
Hilbert space
$\mathcal{H}:= \mbox{span}\{H_x, H_z, p_y, \Delta, \Gamma_0\}$
and the final inner approximation to the Lie wedge takes the form
\begin{equation}\label{eqn:Ex3-cone-a}
\fw_0:=\mathbb R H_y \oplus - \fc_0 \;,
\end{equation}
where $\fc_0$ is parameterised (again in the short-hand of
\eqref{eqn:short-hand-scalprod}) as
\begin{equation}\label{eqn:Ex3-cone}
\fc_0:=\R{}_0^+\conv\Big\{
\left[\begin{smallmatrix}
2\sin(\theta)\\2\cos(\theta)\\ \gamma\sin(2\theta) \\ \gamma(1-\cos(2\theta))\\[.5mm] (11+\cos(2\theta))/6
\end{smallmatrix}\right]
\cdot
\left[\begin{smallmatrix}
H_x\\H_z\\p_y\\ \Delta\\[.5mm] \Gamma_0
\end{smallmatrix}\right] \;\Big|\; \theta\in \R{} \Big\}.
\end{equation}
It is shown in Fig.~\ref{fig:Ex3}. --- As in Example~2,
letting $\Adr_{\exp E(\mathfrak{w})}$ act on the drift terms
adds no further elements to the edge of the wedge, so one gets:
\medskip
\begin{proposition}\label{example3weakglobality}
The set $\fw_0 = \mathbb{R}H_y\oplus(-\fc_0)$ is
the {\em global} Lie wedge to control system \eqref{eq:ex3}.
\end{proposition}
\medskip
\begin{proof}
The Lie wedge property of $\fw_0$ can be derived as in Example 1.
Then the globality of $\fw_0$ follows again from Theorem
\ref{thm:globality-3}.
\end{proof}
\begin{figure}
\begin{center}
\includegraphics[scale=0.39]{Ex23}
\end{center}
\caption{\label{fig:Ex3} (Colour online) The Lie wedge of the more general
system in {\bf Example~3} is five-dimensional. Here the surface
of the convex cone $\fc_0$ generated by $e^{\theta H_y}(\Gamma_0+H_z)e^{-\theta H_y}$
(see \eqref{eqn:Ex3-cone} with $\lambda=1$)
is projected into the subspace $\mbox{span}\{H_x,H_z,\Gamma_0\}$.
Since $[\Gamma_0,H_y]\neq 0$ it deviates from the rotational symmetry
of Example~2 given in the inner part for comparison, cp Fig.~\ref{fig:Ex2}(a).
The figure scales with $\lambda\to\infty$.
The prism-shaped projection into the subspace $\mbox{span}\{H_y,H_z,\Gamma_0\}$ (not shown) is similar
to that of Example~2 already given in Fig.~\ref{fig:Ex2}(b).
}
\end{figure}
\medskip
\noindent
Generalising the relaxation operator in Example~3 to
$\Gamma_0 := \gamma\,\diag(a,b,c)$
with $a,b,c,\gamma\geq 0$ results in a generalised $\fp$-component
replacing \eqref{eqn:p-Ex3} by
\begin{equation*}
\begin{split}
&R_y (\theta) \Gamma_0 R_y(\theta)^\top \;=\;
\gamma\left[\begin{smallmatrix} a+(c-a)\sin^2(\theta) && 0 && (c-a)\sin(\theta)\cos(\theta)\\
0 && b && 0\\
(c-a)\sin(\theta)\cos(\theta) && 0 &&c-(c-a)\sin^2(\theta)\end{smallmatrix}\right] \\[2mm]
& = \Gamma_0 + \gamma(c-a) \sin^2(\theta) \Delta_{13}
+\gamma(c-a) \sin(\theta)\cos(\theta) p_y \\[2mm]
& = \Gamma_0 + \tfrac{\gamma(c-a)}{2} \Delta_{13}
-\tfrac{\gamma(c-a)}{2}\big(\cos(2\theta)\Delta_{13} -\sin(2\theta) p_y\big)
\end{split}
\end{equation*}
This leads to a cone $\fc_0$ for \eqref{eqn:Ex3-cone-a} that keeps the structure of
\eqref{eqn:Ex3-cone} in a slightly more general form,
where Example~2 is readily reproduced by $c=a$, while Example~3
follows for $a=b=1$ and $c=2$.
\medskip
The Lie wedge in Example~3 and its generalised form treated above
do {\em not} take the form of a Lie {\em semialgebra} either.
Choose $B:= H_y$ from the wedge $\fw_0$ of \eqref{eqn:Ex3-cone-a}
and recall $A:=\Gamma_0 + H_z$.
Then the BCH product
\begin{equation}
A\star B = \Gamma_0 + H_y + H_z - \tfrac{1}{2}(H_x + p_y)\, +\, \dots
\end{equation}
leads outside the Lie wedge of \eqref{eqn:Ex3-cone-a}
since the component
$H_x + p_y$ is {\em not} within\footnote{This would require the equality $2\,\sin(\theta)=\sin(2\theta)$
to hold for non-trivial~$\theta$ beyond its actual solutions $\theta=0\negthickspace\mod\pi$.}
the cone \eqref{eqn:Ex3-cone}.
\medskip
As pointed out already, in this section, we have chosen a
representation in $\mathbb R^3$ in order to visualise the
Hamiltonian parts of the respective quantum dynamics by
$SO(3)$-rotations. In quantum mechanics,
this picture can be recovered in the so-called
{\em coherence-vector representation} \cite{AlickiLendi87}.
Therefore, when taking an explicit spin"~$\tfrac{1}{2}$
representation of $\adr_{\su(2)}$
in the following chapter,
the key results obtained here in
Examples~1 through~3
will show up again.
\section{Open Single-Qubit Quantum Systems}\label{sec:q-channels}
In this section, we analyze the standard single-qubit unital quantum systems
beyond their purely dissipative evolution by allowing for Hamiltonian
drifts and controls.
In view of steering open quantum systems, this is an important generalisation.
\subsection{Markovian Master Equation in Qubit Representation}
Based on the Pauli matrices
\begin{equation}\label{eqn:Paulis}
\sigma_x:=
\left[\begin{matrix} 0 & 1\\ 1 &0 \end{matrix}\right],\quad
\sigma_y:=
\left[\begin{matrix} 0 & -i\\ i &0 \end{matrix}\right],\quad
\sigma_z:=
\left[\begin{matrix} 1 & 0\\ 0 & -1 \end{matrix}\right]
\end{equation}
in this section we deliberately depart from the previous notation by using
the explicit spin-$\tfrac{1}{2}$ adjoint representation carrying the
spin-quantum number $j=\tfrac{1}{2}$ as prefactor in
given by
\begin{equation}\label{eqn:def-ad-sigma}
\hat{\sigma}_\nu:=
\tfrac{1}{2} (\unity_2\otimes\sigma_\nu - \sigma_\nu^\top\otimes\unity_2)\;,
\end{equation}
for $\nu\in\{x,y,z\}$, where $\otimes$ denotes
the Kronecker product of matrices.
One easily recovers the $\su(2)$ commutation relations
\begin{equation}
[i\,\hat{\sigma}_p,i\,\hat{\sigma}_q]= - \varepsilon_{pqr} \;i\,\hat{\sigma}_r
\end{equation}
to convince oneself of
$\widehat{\adr}_{\su(2)} := \expt{i\hat{\sigma}_x, i\hat{\sigma}_y,i\hat{\sigma}_z}_{\sf Lie}
\iso\adr_{\su(2)}$.
Here and henceforth we use $\varepsilon_{pqr}$ to discriminate even and odd permutations
of $(x,y,z)$ by their signs, i.e.\
$\varepsilon_{pqr}=+1$ if $(p,q,r)$ is an {\em even} permutation
of $(x,y,z)$, while $\varepsilon_{pqr}=-1$ for an {\em odd} permutation.
Thus for
a single open qubit system in the above representation
the controlled master equation \eqref{eqn:sigma} or rather
its group lift \eqref{eqn:SIGMA} takes the explicit
form
\begin{equation}
\dot X(t) =
-\Big(i\big(\hat{H}_d + \sum_j u_j\hat{H}_j\big)
+ \hat{\Gamma}_L\Big)\, X(t)\,.
\end{equation}
Here, $X(t)$ may be a density operator regarded (via the so-called
$\vec$-representation\footnote{Note that the $\vec$-representation
as well as the vector of coherence notation just provide different
\/`coordinates\/' for the abstract master equation \eqref{eqn:sigma}.
\cite{HJ2}) as an element in $\C{}^4$} or a qubit quantum channel represented in
$GL(4,\mathbb C)$. Moreover, $\hat{H}_d$ and $\hat{H}_j$ are
in general of the from
$\hat{H}_d :=
(\unity_2\otimes H_d - H_d^\top\otimes\unity_2)$
and similar $\hat{H}_j$.
To ensure complete positivity, the relaxation term $\hat{\Gamma}_L$
shall be again of Lindblad-Kossakowski form
which for the standard unital single-qubit systems (with $V_k$ Hermitian)
simply reads
$\hat{\Gamma}_L= 2\sum_k \gamma_k\;\hat{\sigma}_k^2$
to give the nicely structured generator
\begin{equation}\label{eqn:Lindblad-Lu}
\mathcal L_u =
i\big(\hat{H}_d + \sum_j u_j\hat{H}_j\big)
+\; 2 \negthickspace\negthickspace\sum_{k\in \{x,y,z\}}\negthickspace\negthickspace \gamma_k\;\hat{\sigma}_k^2\;.
\end{equation}
The generator is of this form because the $i\hat{H}$ terms are in the $\fk$-part
of the Cartan decomposition of $\gl(4,\mathbb C)$ into
skew-Hermitian ($\fk$) and Hermitian ($\fp$) matrices,
whereas the $\hat{\sigma}_k^2$ terms are in the $\fp$-part.
\subsection[Fully H-Controllable Channels]
{Single-Qubit Systems Satisfying Condition (H)}
Here we consider the class of fully Hamiltonian controllable
unital single-qubit systems whose dissipation is governed by
a \emph{single} Lindblad operator $\hat{\sigma}_k^2$
for some $k\in\{x,y,z\}$
i.e.~two of the three prefactors $\gamma_x,\gamma_y,\gamma_z$
have to vanish.
Similar to Example~1 of Sec.~\ref{sec:geoR3}, choose the controls
$\hat{\sigma}_x$ and $\hat{\sigma}_y$ to
see that such a system fulfills condition (H), since
$\expt{i\hat{\sigma}_x,i\hat{\sigma}_y}_{\sf Lie}=\widehat{\adr}_{\su(2)}$.
Then it is actually immaterial which single Pauli matrix is chosen
as the Lindblad operator $\hat{\sigma}^2_k$, because all of the
Pauli matrices are unitarily equivalent.
So without loss of generality, one may choose $k=z$,
i.e. $\gamma_x=0$, $\gamma_y=0$, and $\gamma_z =: \gamma$.
Therefore the fully Hamiltonian controllable version of the
bit-flip, phase-flip, and bit-phase-flip channels are dynamically
equivalent in as much as they have (up to unitary
equivalence) a common global Lie wedge
\begin{eqnarray}
\fw_0 :=\widehat{\adr}_{\su(2)}\oplus-\fc_0\;,
\end{eqnarray}
where the cone $\fc_0$ is defined by
\begin{equation}\label{Eqn;fullHcontrolcone1qubit}
\fc_0:=
\R{}_0^+\mbox{conv}\left\{\hat{U}\,\hat{\sigma}_z^2\,\hat{U}^\dagger
\;\big|\;U\in SU(2) \right\} \\[2mm] \quad
\end{equation}
with
\begin{equation}\label{eqn:U-superop}
\hat{U} := \bar U \otimes U\;.
\end{equation}
Clearly, the
wedge $\fw_0$ is \emph{global} by Corollary~\ref{HcontrolGlobalWedge}
or, alternatively, by Theorem \ref{thm:globality-3} and its edge $E(\fw_0)$ is
given by the Lie subalgebra $\widehat{\adr}_{\su(2)}$.
The above Lie wedge is isomorphic to the one in
Example~1 of Sec.~\ref{sec:geoR3} for the particular
choice that $\Gamma_0 = \diag(1,1,0)$.
\subsection[Controllable Channels Satisfying Condition (WH)~I: One Lindblad Operator]
{Single-Qubit Systems Satisfying Condition (WH): One Lindblad Operator}
Here we discuss an important class of standard single-qubit systems
which are particularly simple in three regards
\begin{enumerate}
\item[(i)] their dissipative term is governed by a single
Lindblad operator,
$\Gamma_0 :=
2\gamma\hat{\sigma}^2_k$
for some $k\in\{x,y,z\}$;
\item[(ii)] their switchable Hamiltonian control is brought
about by a single Hamiltonian $\hat{\sigma}_c$
for some $c\in\{x,y,z\}$;
\item[(iii)] their non-switchable Hamiltonian drift is
$\hat{\sigma}_d$ for some $d\in\{x,y,z\}$.
\end{enumerate}
Applying the algorithm for the inner approximation of the Lie wedge,
we get in step (1)
\begin{equation}
\fw^c_{dk}(1) :=
i\; \mathbb R \hat{\sigma}_c \oplus\,
-\mathbb R_0^+ \big(i\hat{\sigma}_d+2\gamma\hat{\sigma}_k^2\big)\;,
\end{equation}
where again we note the separation by $\fk$-$\fp$ components.
In step (2) we identify the span generated by the control $i\hat{\sigma}_c$
as the edge $E(\fw)$ of the wedge.
So the conjugation has to be by the control subgroup, i.e.\ by
$
e^{-i2\theta\hat{\sigma}_c} =
e^{+i\theta\sigma_c^\top}\otimes e^{-i\theta\sigma_c}$.
Thus in step (3)
one obtains as $\fk$"~component of the conjugated drift
\begin{equation}\label{Eqn:Kcomponenttrig}
\begin{split}
K_d^c(\theta) := &\quad e^{-i\theta\hat\sigma_c}(i\hat\sigma_d)e^{i\theta\hat\sigma_c}\\
= &\begin{cases} i\,\hat\sigma_d &\text{for $c=d$}\quad\\
i\,\cos(\theta)\hat\sigma_d+ i\,\varepsilon_{cdq}\sin(\theta)\hat\sigma_q &\text{else}
\end{cases}
\end{split}
\end{equation}
and as $\fp$-component
\begin{equation}\label{Eqn:Pcomponenttrig}
\begin{split}
P_k^c(\theta) := &\quad e^{-i\theta\hat\sigma_c}(2\gamma\,\hat\sigma_k^2)e^{i\theta\hat\sigma_c}\\[1mm]
= &\begin{cases} 2\gamma\,\hat\sigma_k^2 &\text{for $c=k$}\quad\\[1mm]
2\gamma\,\big(\cos(\theta)\hat\sigma_k+ \,\varepsilon_{ckr}\sin(\theta)\hat\sigma_r\big)^2 &\text{else}\;.
\end{cases}
\end{split}
\end{equation}
The last expression (for $c\neq k$) can be further resolved
using
the anticommutator $\left\{A,B\right\}_+:=AB+BA$
\begin{equation}\label{eqn:qubitchancor2}
\begin{split}
P^c_k(\theta)
&= 2\gamma\left[\begin{smallmatrix}\cos^2(\theta)\\
\sin^2(\theta)\\
\cos(\theta)\sin(\theta)\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix}\hat\sigma_k^2\\ \hat\sigma_r^2\\ \varepsilon_{ckr}\{\hat\sigma_k,\hat\sigma_r\}_+ \end{smallmatrix}\right]\,.\\[2mm]
&= \tfrac{\gamma}{2}\left[\begin{smallmatrix}2\\
1+\cos(2\theta)\\
1-\cos(2\theta)\\
\sin(2\theta)\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix}\unity\\-(\sigma^\top_k\otimes\sigma_k)\\
-(\sigma^\top_r\otimes\sigma_r)\\
-\varepsilon_{ckr} ((\sigma^\top_k\otimes\sigma_r)+(\sigma^\top_r\otimes\sigma_k))\end{smallmatrix}\right]\;,\\
\end{split}
\end{equation}
where the latter identity gives a decomposition into mutually orthogonal Pauli-basis elements.
To summarize, if the control Hamiltonian neither commutes with the Hamiltonian
part nor with the dissipative part of the drift, one obtains
in terms of the above $K_d^c(\theta)$ and $P_k^c(\theta)$
\begin{eqnarray}\label{eqn:QubitCone}
\fc_{dk}^c:= \mathbb{R}_0^+ \conv\{K_d^c(\theta)+P_k^c(\theta)\,|\,\theta\in\R{}\,\}
\end{eqnarray}
However, if $[\hat\sigma_c,\Gamma_0]=0$, then the convex cone in equation \eqref{eqn:QubitCone}
simplifies by $P_k^c(\theta)=\Gamma_0$ to
\begin{equation}\label{eqn:CommQubitCone}
\fc_{dk}^c
=\R{}_0^+\conv\Big\{\left[\begin{smallmatrix}\cos(\theta)\\\sin(\theta)\\1\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix}i\hat\sigma_d \\ i\varepsilon_{cdq}\hat\sigma_q\\\Gamma_0\end{smallmatrix}\right]
\;\Big|\; \theta \in \R{} \Big\}\quad
\end{equation}
in entire analogy to Example~2 of Sec.~\ref{sec:geoR3}.
The final Lie wedge admits the orthogonal decomposition
\begin{equation}
\fw^c_{dk} := i \mathbb R \,\hat\sigma_c \oplus - \fc^c_{dk}
\end{equation}
and moreover by Theorem \ref{thm:globality-3} (or alternatively by
Corollary \ref{HHLglobalcor}) it is {\em global}.
For $\gamma>0$, the edge $E(\fw)=\expt{i\,\hat\sigma_c}$ is again
the span generated by the control, yet it flips into the full algebra
$E(\fw)=\widehat{\adr}_{\su(2)}$ in the limit $\gamma = 0$.
The relation to Examples~2 and 3 of Sec.~\ref{sec:geoR3} is obvious:
Let a unital qubit system satisfy the (WH)-condition
and have a dissipative Lindbladian
$\Gamma_0:=2\gamma\hat\sigma_k^2$ induced by a single Lindblad
operator $V_k = \sigma_k$. If $[\hat\sigma_c,\Gamma_0]\neq 0$,
one arrives at a situation resembling Example~3, whereas if
$[\hat\sigma_c,\Gamma_0] = 0$, one obtains a result analogous to Example~2.
\subsection*{Application: Bit-Flip and Phase-Flip Channels}
Also the relation to standard {\em unital} qubit channels is immediate:
Note that in the bit-flip channel the noise is generated by
$\hat\sigma_x^2$, while it is
$\hat\sigma_y^2$ in the bit-phase-flip channel and $\hat\sigma_z^2$ in
the phase-flip channel, see Tab.~\ref{tab:q-channels}.
In the absence of any {\em coherent} drift or control brought about
by the respective Hamiltonians
$H_d=\hat \sigma_d$ or $H_j=\hat \sigma_j$, the Kraus representations
are standard. By allowing for drifts and controls, the Kraus rank $K$
of the channel usually increases to $K$=$4$
with exception of a single $\hat \sigma_d$ or
$\hat \sigma_j$ commuting with the single
Lindblad operator {$\hat \sigma_k^2$} keeping $K$=$2$.
Also the time dependences become more involved. Hence explicit results will be given elsewhere.
Under full H"~controllability, the Lie wedges of all the three
channels become equivalent as the Pauli
matrices and thus the corresponding noise generators are unitarily similar.
In contrast, for the case satisfying the (WH)"~condition, assume
a control system with a Hamiltonian drift term governed by $\hat\sigma_z$.
Upon including relaxation, now there are two different scenarios:
if the control Hamiltonian (indexed by $c\in\{x,y,z\}$)
commutes with the noise generator (indexed by $k\in\{x,y,z\}$), one
finds a situation as in Example~2 and \eqref{eqn:CommQubitCone}, otherwise the
scenario is more general as in \eqref{eqn:QubitCone}.
\begin{table*}[Ht!]
\caption{\label{tab:q-channels}
Controlled Single-Qubit Channels and Their Lie Wedges}
\vspace{-2mm}
\begin{center}
\begin{tabular}{l|ll ll}
\hline\hline\\
{\small Channel} & {\small Primary$^*$ Lindblad Operators} & {\small Primary$^*$ Kraus Operators}
& \multicolumn{2}{c}{\small ------------------------ Lie Wedges ------------------------- \phantom{XXXX}}\\
& & & case satisfying (WH)-condition & H-controllable case\\[1mm]
\hline\\[-0.8mm]
Bit Flip &
$V_1=\sqrt{a_{11}}\;\sigma_x$
&
$E_1=\sqrt{r_{11}}\;\sigma_x$
&
$\mathfrak{w}_{dx}^c=\langle
i\hat\sigma_c\rangle\oplus-\mathfrak{c}_{dx}^c$&
$\mathfrak{w}=\widehat{\adr}_{\mathfrak{su}(2)}\oplus-\mathfrak{c}_0$\\[1mm]
&&
$E_0=\sqrt{q_{11}}\;\unity$ & [see Eqns.~(\ref{eqn:QubitCone},\ref{eqn:CommQubitCone})]
& [see Eqn.~\eqref{Eqn;fullHcontrolcone1qubit}] \\[3mm]
Phase Flip &
$V_1=\sqrt{a_{22}}\;\sigma_z$
&
$E_1=\sqrt{r_{22}}\;\sigma_z$
&
$\mathfrak{w}_{dz}^c=\langle
i\hat\sigma_c\rangle\oplus-\mathfrak{c}_{dz}^c$&
---same as above---\\[1mm]
&&
$E_0=\sqrt{q_{22}}\;\unity$ & [see Eqns.~(\ref{eqn:QubitCone},\ref{eqn:CommQubitCone})]\\[3mm]
Bit-Phase Flip &
$V_1=\sqrt{a_{33}}\;\sigma_y$
&
$E_1=\sqrt{r_{33}}\;\sigma_y$
&
$\mathfrak{w}_{dy}^c=\langle
i\hat\sigma_c\rangle\oplus-\mathfrak{c}_{dy}^c$&
---same as above---\\[1mm]
&&
$E_0=\sqrt{q_{33}}\;\unity$ & [see Eqns.~(\ref{eqn:QubitCone},\ref{eqn:CommQubitCone})]\\[3mm]
\hline\\
Depolarizing &
$V_1=\sqrt{a_{11}}\;\sigma_x$
&
$E_1=\sqrt{r_1}\;\sigma_x$
&
$\mathfrak{w}_{d,xyz}^c=\langle
i\hat\sigma_c\rangle\oplus-\mathfrak{c}_{d,xyz}^c$&
$\fw=\widehat{\adr}_{\su(2)}\oplus -\fc_{xyz}$ \\[1mm]
&$V_2=\sqrt{a_{22}}\;\sigma_y$
&
$E_2=\sqrt{r_2}\;\sigma_y$
& \multicolumn{1}{l}{[see Eqn.~\eqref{Eqn:Kcomponenttrig} and Eqns.~(\ref{eqn:Pkkk},\ref{eqn:Pkkk-simple})]}
& [see Eqns.~(\ref{eqn:H-cont-depol-1},\ref{eqn:H-cont-depol-2})]\\[1mm]
&$V_3=\sqrt{a_{33}}\;\sigma_z$
&
$E_3=\sqrt{r_3}\;\sigma_z$
&&\\[1mm]
&&
$E_0=\sqrt{r_0}\;\unity$&&\\[2mm]
\multicolumn{5}{l}{\phantom{XX}}\\[-3mm]
\hline\hline\\
\multicolumn{5}{l}{$^*$) Primary operators are for purely dissipative time evolutions
(no Hamiltonian drift no control).
Then the time dependence of the}\\
\multicolumn{5}{l}{Kraus operators roots in the GKS matrix
$\{a_{ii}\}_{i=1}^3$. Define:
$ \lambda_1:=a_{22}+a_{33}, \lambda_2:=a_{22}-a_{33}, \lambda_3:=a_{11}+a_{22}$, and thereby
$q_{ii}:=\tfrac{1}{2}(1+e^{-a_{ii}t}), $}\\[1mm]
\multicolumn{5}{l}{$r_{ii}:=\tfrac{1}{2}(1-e^{-a_{ii}t})$ and $r_0:=\tfrac{1}{4}(1+e^{-\lambda_1 t}+e^{-\lambda_2 t}+e^{-\lambda_3 t}),\;
r_1:=\tfrac{1}{4}(1-e^{-\lambda_1 t}+e^{-\lambda_2 t}-e^{-\lambda_3 t}),\;
r_2:=\tfrac{1}{4}(1+e^{-\lambda_1 t}-e^{-\lambda_2 t}-e^{-\lambda_3 t})$,}\\[1mm]
\multicolumn{5}{l}{$r_3:=\tfrac{1}{4}(1-e^{-\lambda_1 t}-e^{-\lambda_2 t}+e^{-\lambda_3 t})$. ---
Under Hamiltonian drift and control the Kraus-rank gets $K=4$ except for one single control $H_c$ or drift $H_d$}\\
\multicolumn{5}{l}{that commutes with the only Lindblad operator $V_k$: in this case the Kraus rank is $K=2$. Time-dependences are involved
and will be given elsewhere.}\\
\multicolumn{5}{l}{The Pauli matrices $\sigma_\nu$ are defined in Eqn.~\eqref{eqn:Paulis}.
}\\
\end{tabular}
\end{center}
\end{table*}
\subsection[Controllable Channels Satisfying Condition (WH)~II: Several Lindblad Operators]
{Single-Qubit Systems Satisfying Condition (WH): Several Lindblad Operators}
Consider a unital qubit system satisfying the (WH)-condition and whose
Lindbladian $\Gamma_0$ is generated by $\ell = 2$ or $\ell = 3$ different
Lindblad operators $\hat\sigma_k^2$.
Then one obtains the following generalisations of the
symmetric component $P_k^c(\theta)\in\fc_{dk}^c$.\\[3mm]
For $\ell=2$ and $\sigma_c\perp\sigma_k$, $\sigma_c=\sigma_{k'}$
\begin{equation}
\begin{split}
\label{eqn:2Lind1comm}
P^c_{kk'}(\theta)
&= 2\left[\begin{smallmatrix}\gamma'\\\gamma\cos^2(\theta)\\
\gamma\sin^2(\theta)\\
\gamma\cos(\theta)\sin(\theta)\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix}\hat\sigma^2_{k'}\\ \hat\sigma^2_k\\ \hat\sigma_r^2\\
\varepsilon_{ckr}\{\hat\sigma_k,\hat\sigma_r\}_+ \end{smallmatrix}\right]\\[2mm]
= \tfrac{1}{2}&\left[\begin{smallmatrix} \gamma' \\ 2 (\gamma+\gamma')\\
\gamma(1+\cos(2\theta))\\
\gamma(1-\cos(2\theta))\\
\gamma\sin(2\theta)\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix} -(\sigma^\top_{k'}\otimes\sigma_{k'})\\\unity\\
-(\sigma^\top_k\otimes\sigma_k)\\
-(\sigma^\top_r\otimes\sigma_r)\\
-\varepsilon_{ckr} ((\sigma^\top_r\otimes\sigma_k)+(\sigma^\top_k\otimes\sigma_r))\end{smallmatrix}\right]. \\[2mm]
\end{split}
\end{equation}
while for $\ell=3$ and $\sigma_c\perp\sigma_k$, $\sigma_c\perp\sigma_{k'}$, $\sigma_c=\sigma_{k''}$
\begin{equation}\label{eqn:Pkkk}
\begin{split}
&P^c_{kk'k''}(\theta)
= 2\left[\begin{smallmatrix}\gamma''\\
\gamma\cos^2(\theta)+\gamma'\sin^2(\theta)\\
\gamma'\cos^2(\theta)+\gamma\sin^2(\theta)\\
(\gamma-\gamma')\cos(\theta)\sin(\theta)\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix}\hat\sigma^2_{k''} \\ \hat\sigma^2_k\\ \hat\sigma^2_{k'}\\
\varepsilon_{ckk'}\{\hat\sigma_k,\hat\sigma_{k'}\}_+ \end{smallmatrix}\right]\\[2mm]
= \tfrac{1}{2}&\left[\begin{smallmatrix}\gamma''\\
2(\gamma+\gamma'+\gamma'')\\
\gamma+\gamma'+(\gamma-\gamma')\cos(2\theta)\\
\gamma+\gamma'-(\gamma-\gamma')\cos(2\theta)\\
(\gamma-\gamma')\sin(2\theta)\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix}-(\sigma^\top_{k''}\otimes\sigma_{k''})\\
\unity\\-(\sigma^\top_k\otimes\sigma_k)\\
-(\sigma^\top_{k'}\otimes\sigma_{k'})\\
-\varepsilon_{ckk'} ((\sigma^\top_{k'}\otimes\sigma_k)+(\sigma^\top_k\otimes\sigma_{k'}))\end{smallmatrix}\right].\\[2mm]
\end{split}
\end{equation}
which for $\gamma=\gamma'=\gamma''$ simplifies to
\begin{eqnarray}\label{eqn:Pkkk-simple}
P^c_{kk'k''}(\theta)=\Gamma_0=2\gamma(\hat\sigma^2_k+\hat\sigma^2_{k'}+\hat\sigma^2_{k''})\;.
\end{eqnarray}
Note that \eqref{eqn:2Lind1comm} with $\gamma=\gamma'$ precisely
corresponds to Example~3 in Sec.~\ref{sec:geoR3}.
\subsection*{Application: Depolarising Channel}
Treating the depolarising channel also becomes immediate,
since one has three noise generators governed by all of $\hat\sigma_x,\hat\sigma_y$,
and $\hat\sigma_z$.
Thus the fully Hamiltonian controllable version of the
depolarising channel follows the bit-flip and phase-flip channels
in the structure of its global Lie wedge
\begin{eqnarray}\label{eqn:H-cont-depol-1}
\fw_0 :=\widehat{\adr}_{\su(2)}\oplus-\fc_{xyz}\;,
\end{eqnarray}
where the cone $\fc_{xyz}$ now reads
\begin{equation}\label{eqn:H-cont-depol-2}
\fc_{xyz}:=\R{}_0^+\mbox{conv}\big\{\hat{U}(\gamma_x\hat\sigma_x^2 + \gamma_y\hat\sigma_y^2
+\gamma_z\hat\sigma_z^2)\hat{U}^\dagger\;|\;U\in SU(2)\big\}
\end{equation}
with $\hat U$ of the form \eqref{eqn:U-superop}.
Again, the edge of the wedge is given by the entire algebra $E(\fw)=\widehat{\adr}_{\su(2)}$
and globality of the wedge follows by Theorem~\ref{thm:globality-3} or
Corollary~\ref{HcontrolGlobalWedge}. ---
Moreover, note that the Lie wedge in the fully Hamiltonian controllable depolarising channel with
{\em isotropic} noise takes the structure of a Lie semialgebra as (in the coherence-vector
representation) it corresponds to the special case of Example~1 in
Sec.~\ref{sec:geoR3}, where the relaxation
operator is a scalar multiple of the unity, $\Gamma_0=\lambda\cdot\unity$.
For {\em anisotropic} relaxation, however, this feature does not arise.
\medskip
If only condition (WH) is satisfied, there are two distinctions:
if the noise contributions are isotropic (i.e.\ with equal contribution by all the
Paulis through $\gamma_x=\gamma_y=\gamma_z$),
one finds a cone expressed by \eqref{Eqn:Kcomponenttrig} and \eqref{eqn:Pkkk-simple}.
However, in the generic anisotropic case, the cone can be expressed by
\eqref{Eqn:Kcomponenttrig} and \eqref{eqn:Pkkk}, see also Tab.~\ref{tab:q-channels}.
\section{Open Two-Qubit Quantum Systems}
In this section we extend the notions introduced in the previous
chapter to three types of two-qubit quantum systems with uncorrelated noise.
The two qubits will be denoted $\sA$ and $\sB$, respectively.
Moreover, we use the short-hands $\sigma_{\mu\nu}:=\sigma_\mu\otimes\sigma_\nu$
with $\mu,\nu\in\{x,y,z,1\}$, where $\sigma_1:=\unity$ as well as
the corresponding \/`commutator superoperators\/'
$\hat\sigma_{\mu\nu}:=\tfrac{1}{2}(\unity\otimes\sigma_{\mu\nu}-\sigma_{\mu\nu}^t\otimes\unity)$.
\subsection[Fully H"~Controllable Channels]{Fully H"~Controllable Two-Qubit Channels}
A fully Hamiltonian controllable two-qubit toy-model
system with switchable Ising-coupling is given by the master equation
\begin{eqnarray}\label{hcontrol2qubiteqn}
\dot\rho=
-\big(i \sum_j u_j\hat\sigma_j + \Gamma_0\big)\rho\;
\end{eqnarray}
where $\hat\sigma_j\in\{\hat\sigma_{x1},\hat\sigma_{y1};\hat\sigma_{1x},\hat\sigma_{1y};\hat\sigma_{zz}\}$
are the Hamiltonian control terms with amplitudes $\left\{u_j\right\}_{j=1}^5\in\mathbb{R}$.
Since $\expt{i\hat\sigma_j\,|\,j=1,2,\dots,5}_{\sf Lie} =\widehat{\adr}_{\su(4)}$,
the edge of the wedge is $E(\fw)=\widehat{\adr}_{\su(4)}$.
Following the algorithm for an inner approximation of the
Lie wedge, step (1) thus gives
\begin{eqnarray}
\fw_1:=\widehat{\adr}_{\mathfrak{su}(4)}\oplus-\mathbb{R}^+_0\Gamma_0\;.
\end{eqnarray}
Conjugating the dissipative component by the exponential map of the edge
and then taking the convex hull yields the convex cone
\begin{eqnarray}
\fc_0:=\mbox{conv}\big\{\lambda\hat U\Gamma_0 \hat U^{\dagger}| \hat U:=\bar U \otimes U, U\in SU(4), \lambda\geq 0\big\}\;,
\end{eqnarray}
which is the two-qubit analogue of the cone in Eqn.~\eqref{Eqn;fullHcontrolcone1qubit}.
The resulting Lie wedge
\begin{eqnarray}
\fw_0:=\widehat{\adr}_{\su(4)}\oplus-\fc_0
\end{eqnarray}
is global by Theorem~\ref{thm:globality-3}.
\subsection[Controllable Channels Satisfying (H)"~Condition Locally and (WH)"~Condition Globally]
{Two-Qubit Channels Satisfying the (H)"~Condition Locally and the (WH)"~Condition Globally}
By shifting the Ising coupling term from the set of switchable control Hamiltonians
into the (non-switchable) drift term, $\hat\sigma_d= \hat\sigma_{zz}$,
one obtains the realistic and actually widely occuring type of system
\begin{eqnarray}\label{eqn:HWHcontrol2qubit}
\dot\rho=
-\big(i \hat\sigma_d + i\sum_j u_j\hat\sigma_j + \Gamma_0\big)\rho\;
\end{eqnarray}
where now one just has the local control terms
$\hat\sigma_j\in\{\hat\sigma_{x1},\hat\sigma_{y1};\hat\sigma_{1x}, \hat\sigma_{1y}\}$.
Since $\expt{i\hat\sigma_{x1}, i\hat\sigma_{y1}}_{\sf Lie} =\widehat{\adr}_{\su_{\sA}(2)}\otimes\unity_{\sB}$,
whereas on the other hand $\expt{i\hat\sigma_{1x},i\hat\sigma_{1y}}_{\sf Lie} =\unity_{\sA}\otimes\widehat{\adr}_{\su_{\sB}(2)}$,
the edge of the wedge
\begin{equation}
E(\fw)=\widehat{\adr}_{\su_{\sA}(2)\widehat{\oplus}\su_{\sB}(2)}
\end{equation}
is in fact brought about by the Kronecker sum of local algebras
\begin{equation}
\su_{\sA}(2)\otimes\unity_{\sB} + \unity_{\sA}\otimes\su_{\sB}(2) =: \su_{\sA}(2)\widehat{\oplus}\su_{\sB}(2)
\end{equation}
forming the generator of the group of local unitary actions
\begin{equation}
\exp \big(\su_{\sA}(2)\widehat{\oplus}\su_{\sB}(2)\big) = SU_{\sA}(2)\otimes SU_{\sB}(2)\;.
\end{equation}
Remarkably, in this important class of open quantum-dynamical systems,
qubits \sA and \sB are {\em locally} (H)"~controllable, respectively,
while {\em globally} the system satisfies but the (WH)"~condition.
\medskip
The final Lie wedge in these systems reads as
\begin{eqnarray}
\fw_{dk}^{2\oplus 2}=\widehat{\adr}_{\su_{\sA}(2)\widehat{\oplus}\su_{\sB}(2)}\oplus-\fc_{dk}^{2\oplus 2}
\end{eqnarray}
with the convex cone
\begin{eqnarray}
\fc_{dk}^{2\oplus 2}:=\mathbb{R}^+_0 \conv\big\{K_{d}^{2\oplus 2}+P_{k}^{2\oplus 2}\big\}
\end{eqnarray}
being given in terms of the respective $\fk$ and $\fp$-components.
Here we use the short-hand of \eqref{eqn:U-superop} in the sense of
$\hat U_{2\otimes 2}:=\bar U_{2\otimes 2}\otimes U_{2\otimes 2}$
to arrive at
\begin{eqnarray}
K_d^{2\oplus 2}&:=&\{\hat U_{2\otimes 2}(i\hat \sigma_d)\hat U^\dagger_{2\otimes 2}\,|\,U_{2\otimes 2}\in SU(2)\otimes SU(2)\}\qquad\\
P_k^{2\oplus 2}&:=&\{\hat U_{2\otimes 2}(\Gamma_0)\hat U^\dagger_{2\otimes 2}\,|\,U_{2\otimes 2}\in SU(2)\otimes SU(2)\}\,.
\end{eqnarray}
As before, this immediately results from the initial wedge approximation by step (1)
\begin{eqnarray}
\fw_1^{2\oplus 2}:=\widehat{\adr}_{\mathfrak{su}_{\sA}(2)\widehat{\oplus}\mathfrak{su}_{\sB}(2)}\oplus
-\mathbb{R}^+_0(i\hat \sigma_d+\Gamma_0)
\end{eqnarray}
followed by conjugation with $\Adr_{\exp E(\fw)}=\Adr_{2\otimes 2}$ to give
\begin{equation}\label{eqn:2qubiKP}
K_d^{2\oplus 2}+P_k^{2\oplus 2}
:= \mathcal O_{SU(2)\otimes SU(2)}\big(i\hat \sigma_d+\Gamma_0\big)\;.\quad
\end{equation}
Step (3) then takes the convex hull. ---
To show globality, let $\fw'$ denote the global
Lie wedge corresponding to the fully H"~controllable
system given in Eqn.~ \eqref{hcontrol2qubiteqn}. Then $\fw_{dk}^{2\oplus 2}\subset\fw'$, and it can be
shown that $\fw_{dk}^{2\oplus 2}$ satisfies the conditions of Corollary \ref{HHLglobalcor}
and therefore is {\em global}.
\subsection[Controllable Channels Satisfying Only (WH)"~Condition]
{Two-Qubit Channels Satisfying Only the (WH)"~Condition}
In the final example of a two-qubit system, the independent local controls shall even
be limited to either $x$ or $y$-controls on the two qubits according to
\begin{eqnarray}\label{eqn:WHWHcontrol2qubit}
\dot\rho
=-\big(i(\hat\sigma_d + u_{\sA}\hat\sigma_{c1} + u_{\sB}\hat\sigma_{1c'})+\Gamma_0\big)\rho\;,
\end{eqnarray}
where now $\hat\sigma_d:=i\big(\hat\sigma_{z1} + \hat\sigma_{1z} + \hat\sigma_{zz}\big)$ and
$\hat\sigma_{c1}$ with a single $c\in\{x,y\}$ and likewise $\hat\sigma_{1c'}$ with
a single $c'\in\{x,y\}$ and $u_{\sA},u_{\sB}\in\mathbb{R}$.
Furthermore, assume the system undergoes {\em local uncorrelated noise}
in each of the two subsystems in the sense that the Lindblad operators are of local form
\begin{eqnarray}
V_k &\in &\{\sigma_{k1}\;|\; k\in\{x,y,z\}\}\\
V_{k'} &\in &\{\hat\sigma_{1k'}\;|\; k'\in\{x,y,z\}\}\;,
\end{eqnarray}
where $k$ and $k'$ are chosen independently $k,k'\in\{x,y,z\}$ so that
in the convention of \eqref{eqn:def-ad-sigma} one finds
\begin{equation}
\Gamma_0\;:=\;2\gamma\hat \sigma_{k1}^2\; +\; 2\gamma'\hat \sigma^2_{1k'}\;.
\end{equation}
This system satisfies but the (WH)"~condition both locally and globally,
the latter following from
\begin{equation}
\expt{i\hat\sigma_{c1}, i\hat\sigma_{1c'}, i\hat\sigma_d}_{\sf Lie} = \widehat{\adr}_{\su(4)}\;.
\end{equation}
\bigskip
The Lie wedge is given by
\begin{eqnarray}
\fw^{cc'}_{kk'}:=\expt{i\hat \sigma_c} + \expt{i\hat \sigma_{c'}} \oplus - \fc^{cc'}_{kk'}\;,
\end{eqnarray}
where the two-dimensional edge of the wedge is generated by the rays
$\expt{i\hat \sigma_{c1}}$, $\expt{i\hat\sigma_{1c'}}$ and the cone
\begin{equation}
\begin{split}
\fc^{cc'}_{kk'}:=\mathbb{R}_0^+ \conv\big\{& K^c(\theta)+K^{c'}(\theta') +K^{cc'}(\theta,\theta')\\
&+ P_k^c(\theta)+P^{c'}_{k'}(\theta')\,|\,\theta,\theta'\in\mathbb R\,\big\}
\end{split}
\end{equation}
is given in terms of the $\fk$- and $\fp$"~components
(setting $\theta:=u_{\sA}$ and $\theta':=u_{\sB}$
and using the relations in \eqref{Eqn:Kcomponenttrig}) as
\begin{equation}\label{eqn:two-qubit-K123}
K^c(\theta) + K^{c'}(\theta') + K^{cc'}(\theta,\theta')
= \left[\begin{smallmatrix}\cos(\theta)\\\sin(\theta)\\\cos(\theta')\\\sin(\theta')\\ \cos(\theta)\cos(\theta') \\
\cos(\theta)\sin(\theta')\\ \cos(\theta')\sin(\theta)\\ \sin(\theta)\sin(\theta') \end{smallmatrix}\right]
\cdot i
\left[\begin{smallmatrix}\hat \sigma_{z1}\\ \varepsilon_{czq}\hat \sigma_{q1}\\
\hat \sigma_{1z}\\ \varepsilon_{c'zq'} \hat \sigma_{1q'}\\
\hat \sigma_{zz}\\ \varepsilon_{c'zq'} \hat \sigma_{zq'}\\
\varepsilon_{czq} \hat \sigma_{qz}\\
\hat \sigma_{qq'} \end{smallmatrix}\right]\qquad
\end{equation}
and (as in \eqref{eqn:qubitchancor2})
\begin{equation}\label{eqn:two-qubit-P1}
\begin{split}
P^c_k&(\theta)
= 2\gamma\left[\begin{smallmatrix}\cos^2(\theta)\\
\sin^2(\theta)\\
\cos(\theta)\sin(\theta)\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix}\hat \sigma^2_{k1}\\
\hat \sigma^2_{r1}\\
\varepsilon_{ckr}\{\hat \sigma_{k1},\hat \sigma_{r1}\}_+ \end{smallmatrix}\right]\\[2mm]
&= \tfrac{\gamma}{2}\left[\begin{smallmatrix}2\\
1+\cos(2\theta)\\
1-\cos(2\theta)\\
\sin(2\theta)\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix}\unity\\-(\sigma^\top_k\otimes\sigma_k)\otimes\unity_{\sB}\\
-(\sigma^\top_r\otimes\sigma_r)\otimes\unity_{\sB}\\
-\varepsilon_{ckr} ((\sigma^\top_k\otimes\sigma_r)+(\sigma^\top_r\otimes\sigma_k))\otimes\unity_{\sB}\end{smallmatrix}\right]\\[2mm]
\end{split}
\end{equation}
as well as
\begin{equation}\label{eqn:two-qubit-P2}
\begin{split}
P^{c'}_{k'}&(\theta')
= 2\gamma'\left[\begin{smallmatrix}\cos^2(\theta')\\
\sin^2(\theta')\\
\cos(\theta')\sin(\theta')\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix}\hat \sigma_{1k'}^2\\
\hat \sigma_{1r'}^2\\
\varepsilon_{c'k'r'}\{\hat \sigma_{1k'}^2,
\hat \sigma_{1r'}^2\}_+ \end{smallmatrix}\right]\\[2mm]
&= \tfrac{\gamma}{2}\left[\begin{smallmatrix}2\\
1+\cos(2\theta')\\
1-\cos(2\theta')\\
\sin(2\theta')\end{smallmatrix}\right]\cdot
\left[\begin{smallmatrix}\unity\\-\unity_{\sA}\otimes(\sigma^\top_k\otimes\sigma_k)\\
-\unity_{\sA}\otimes(\sigma^\top_r\otimes\sigma_r)\\
-\varepsilon_{c'k'r'} \unity_{\sA}\otimes((\sigma^\top_{k'}\otimes\sigma_{r'})+(\sigma^\top_{r'}\otimes\sigma_{k'}))
\end{smallmatrix}\right].\\[2mm]
\end{split}
\end{equation}
To see this, observe that by step (1), the initial wedge approximation is given by
\begin{equation}
\fw_1:= \expt{i \hat \sigma_c}+\expt{i\hat \sigma_{c'}}
\oplus-\mathbb R^+_0(i\hat \sigma_d+2\gamma\hat \sigma_k^2+2\gamma'\hat \sigma^2_{k'})\,,\\[2mm]
\end{equation}
which has to be conjugated by $\Adr_{\exp(E(\fw))}$. As usual, the edge of
the wedge is invariant under such a conjugation, so we need
only determine the effects on the drift components of the system
as is done in Eqns.~\eqref{eqn:two-qubit-K123} through \eqref{eqn:two-qubit-P2}.
Moreover, the wedge is global by application of Corollary \ref{HHLglobalcor}.
\medskip
Now, the generalisation to systems with more than two qubits satisfying the (H)"~ or (WH)"~condition is obvious:
assuming uncorrelated noise, the $\fp$"~parts of the Lie wedges can be immediately
extended on the grounds of the previous description, since all processes are
local on each qubit. Though straightforward, calculating the $\fk$"~components becomes
a bit more tedious: but the many-body coherences have to be considered just as in \eqref{eqn:two-qubit-K123}.
\section{Outlook: Approximating Reachable Sets}
Knowing the {\em global} Lie wedge of a {\em coherently controlled}
Markovian system provides a convenient means
to efficiently approximate its reachable sets.
As in the case of a Lie algebra, the image of the wedge $\fw$ under
the exponential map yields a first approximation of the corresponding
Lie semigroup $\bS$. Unfortunately, this image is in general only
a proper subset of $\bS$---this, however, may happen also
for Lie algebras when the corresponding Lie group is non-compact.
Therefore, one has to allow for finite products of the form
${\rm e}^{A_1}{\rm e}^{A_2}\cdots{\rm e}^{A_\ell}$ with
$A_1,A_2,\dots, A_\ell \in \fw$ to obtain the entire semigroup $\bS$.
Although the minimal number $\ell_*$ of factors to generate $\bS$
(called \emph{number of intrinsic control-switches})
is in general unknown, this approach provides a much more effective
parametrization of the reachable sets than the standard method
which works with the original control directions and
piecewise constant controls as parameter space. Thereby one can optimize
target functions almost directly over the reachable sets thus
complementing standard optimal control methods of open systems
\cite{Teo91,JPB_decoh,Rabitz07}. Particularly simple are systems whose Lie wedges
do carry a Lie-semialgebra structure (like in isotropoic
depolarising channels). Here
one knows {\em a priori} that only a few (or sometimes even
zero) intrinsic control-switches are necessary, so some control
problems may actually be solved by {\em constant controls}.
\section{Conclusions}
We have generalised standard unital quantum channels (bit-flip,
phase-flip, bit-phase-flip, and depolarising) by allowing for different
degree of coherent Hamiltonian control.
For the first time, here we have characterized their
respective global Lie wedges governing {\em all directions}
the controlled open system can possibly take.
The results have been further generalised
to various types of two-qubit systems with uncorrelated noise.
Since controlled multi-qubit channels can be treated likewise,
the geometrical Lie-semigroup approach taken is anticipated
to find wide applications in quantum systems theory and engineering:
this is because knowing the global Lie wedge of a controlled
Markovian system paves the way to efficiently approximate its reachable sets.
Thus this
knowledge will be very useful for improving
known bounds (cf.~\cite{Yuan10}) on the corresponding system semigroup
$\bP_\Sigma$ in follow-up work.
Finally, our results demonstrate that the Lie wedges
associated to most of the controlled quantum systems
do \emph{not} take the special form of Lie semialgebras,
an important exception being the fully controlled isotropic
depolarising channel.
\section{Appendix}
\subsection{The Principal Theorem of Globality}\label{app:A}
For the reader's convenience, we state the `\/Principal Globality
Theorem\/' with minor simplifications. For the full version and
its (quite involved) proof we refer to \cite{HHL89}, which we sketch
in the sequel.
Let $\bG$ be a matrix Lie group with Lie algebra $\fg$, so
\begin{equation}
\fg \, X := \{AX \;|\; A \in \fg\}
\end{equation}
can be envisaged as {\em tangent space} $T_X \bG $ at $X\in\bG$,
while $T\bG$ and $T^*\bG$ shall denote the {\em tangent bundle}
and, respectively, {\em cotangent bundle} of $\bG$. Thus, one
has the isomorphisms
\begin{equation}
T\bG \cong \fg \times \bG
\quad\text{and}\quad
T^*\bG \cong \fg^* \times \bG.
\end{equation}
Now, let $\fw$ be any wedge of $\fg$. A {\em 1-form} on $\bG$
is a smooth cross section of the cotangent bundle,
i.e.~$\omega: \bG\to T^*\bG$ with $\omega(X)\in T^*_X\bG$.
Moreover, $\omega$ is called
\begin{enumerate}
\item[(1)]
{\em exact} if there exists a smooth function
$\varphi: \bG \to \R{}$ such that ${\rm d}\varphi = \omega$;
\item[(2)] {\em $\fw$"~positive} at $X\in\bG$ if
$\langle\omega(X),AX\rangle \geq 0$ for all $A\in \fw$;
\item[(3)] {\em strictly $\fw$"~positive} at $X\in\bG$ if
$\fw$"~positivity holds at $X\in\bG$ and one has
$\expt{\omega(X),AX}>0$ for all $A\in\fw\setminus -\fw$;
\end{enumerate}
\noindent
The existence of a strictly $\fw$"~positive 1"~form is ensured
in the following scenario \cite{HHL89}:
If $\bG$ is a Lie group with Lie algebra $\fg$ and
$\bH$ a closed subgroup with Lie algebra $\fh$,
then for any Lie wedge $\fw\subset\fg$ whose edge
$E(\fw)$ coincides with $\fh$ one can construct strictly
$\fw$"~positive 1"~forms on $\bG$.
Note, however, that these 1"~forms on $\bG$ are in general
\emph{not} exact. Yet, whenever exactness can be guaranteed
in addition, one has the following equivalences.
\medskip
\begin{theorem}[\cite{HHL89}]\label{thm:PGT}
Let $\bG$ denote a finite-dimensional real matrix Lie group with
Lie algebra $\fg$ and let $\fw$ be a Lie wedge of $\fg$.
Moreover, let $\mathfrak{g}_0:=\langle \fw \rangle_{\rm Lie}$ be the Lie
subalgebra generated by $\fw$ and let $\mathbf{G}_0$ be the corresponding
Lie subgroup of $\bG$. Further, assume that $\mathbf{G}_0$ is closed within
$\bG$. Then the following statements are equivalent:
\begin{enumerate}
\item[(a)]
$\fw$ is global in $\bG$.
\item[(b)]
$\fw$ is global in $\mathbf{G}_0$.
\item[(c)]
There is a closed connected subgroup $\bH$ of $\mathbf{G}_0$
with $L(\bH) = E(\fw)$ and a $1$-from $\omega$ on $\mathbf{G}_0$
which satisfies the following conditions:
\begin{enumerate}
\item[(i)]
$\omega$ is exact.
\item[(ii)]
$\omega$ is $\fw$-positive for all $X \in \mathbf{G}_0$.
\item[(iii)]
$\omega$ is strictly $\fw$-positive at the identity $\unity$.
\end{enumerate}
\end{enumerate}
\end{theorem}
\medskip
Now, the following consequence of the `\/Principal Globality Theorem\/'
already mentioned in the main text is a useful tool
whenever a {\em global} Lie wedge $\fw$ embracing the Lie wedge $\fw_0$
of interest is already known.
\medskip
{\em Corollary~\ref{HHLglobalcor}~(\cite{HHL89})}
Let $\bG$ be a Lie group with Lie algebra $\fg$ and let
$\fw_0\subseteq\fw$ be two Lie wedges in $\fg$.
Provided
\begin{eqnarray}
\fw_0\setminus-\fw_0\subseteq\fw \setminus -\fw \;,
\end{eqnarray}
then $\fw_0$ is global in $\bG$ if the following conditions are
satisfied:
\begin{enumerate}
\item[(i)]
$\fw$ is global in $\bG$.
\item[(ii)]
The edge of $\fw_0$
is the Lie algebra of a closed Lie subgroup of $\bG$.
\end{enumerate}
\medskip
In other words, if the edge of the wedge follows the intersection
$E(\fw_0)=E(\fw)\cap\fw_0$
and $\fw$ is global, then $\fw_0$ is also a global Lie wedge,
whenever $\exp E(\fw_0)$ generates a closed subgroup.
\subsection{Lie Semialgebra Structure in Example 1}\label{app:B}
In Sec.~\ref{sec:geoR3} we stated that the Lie wedge of
Example~1
\begin{equation}
\fw_0 := \so(3)\;\oplus\; (-\fc_0),
\end{equation}
where $\fc_0 := \R{}^+_0\mathcal M(\Gamma_0)$ and
$M(\Gamma_0):=
\{S \in \sym(3) \,|\, S \prec \Gamma_0\}$,
is in fact a {\em Lie semialgebra} for $\Gamma_0 = \lambda \cdot \unity$
(corresponding to the isotropic depolarising channel),
whereas it fails to be a Lie semialgebra for any other $\Gamma_0$.
Recall, here $\sym(3)$ is the set of all symmetric
$3 \!\times\! 3$-matrices. For proving the above statement,
we distinguish the following
cases\footnote{Although case (iii) seems to be quite similar to
case (ii), its proof is more involved and a helpful preparation
of the general case (iv).}:
\begin{enumerate}
\item[(i)]
$\Gamma_0$ is a multiple of the identity, thus
we can assume without loss of generality $\Gamma_0 := \unity$;
\item[(ii)]
$\Gamma_0$ has zero as eigenvalue with multiplicity $2$, thus we
can assume $\Gamma_0 := \diag(1,0,0)$.
\item[(iii)]
$\Gamma_0$ has an eigenvalue different to zero with multiplicity $2$,
thus without loss of generality $\Gamma_0 :=\mbox{diag}(1,1,0)$.
\item[(iv)]
$\Gamma_0$ has three distinct eigenvalues, i.e.
$\Gamma_0 :=\mbox{diag}(a,b,c)$ with $a > b > c \geq 0$.
\end{enumerate}
In all cases, the identification of the dual wedge
of $\fw_0$ is crucial for the compution of the
tangent space $T_A\fw_0$ at $A \in \fw_0$ via \eqref{eqn:semialg-incl}.
Therefore, we first provide an auxiliary result characterizing
the dual cone of $\fc_0$ within $\sym(3)$.
\begin{lemma}
\label{dual}
Let $\Gamma_0 := \diag(a,b,c)$ with $a \geq b \geq c \geq0$ and let
$\fc_0 := \R{}^+_0\mathcal M(\Gamma_0)$ with
$ \mathcal{M}(\Gamma_0) :=
\{S \in \sym(3) \,|\, S \prec \Gamma_0\}$.
Then the dual cone of $\fc_0$ within $\sym(3)$ is given by
\begin{eqnarray}
\mathfrak{c}_{\mathfrak{p}}^*
:=\left\{S \in \sym(3) \,|\,
c \lambda_1(S) + b \lambda_2(S) + a \lambda_3(S) \geq 0\right\},
\end{eqnarray}
provided $\lambda_1(S) \geq \lambda_2(S) \geq \lambda_3(S)$
are the eigenvalues of $S$.
\end{lemma}
\begin{proof}
By definition, one has the equivalence $S \in \fc_{\fp}^*$ if
and only if $\langle S, S'\rangle\geq 0$ for all $S'\in\fc_0$.
Since $\fc_0 = \R{}_0^+ \conv \mathcal{O}_{SO(3)}(\Gamma_0)$
this condition reduces to
$\langle S, \Theta \Gamma_0\Theta^\top \rangle \geq 0$
for all $\Theta \in SO(3)$. Then von Neumann's inequality
\cite{NEUM-37} provides the equivalence:
$\langle S, \Theta \Gamma_0\Theta^\top \rangle \geq 0$
for all $\Theta \in SO(3)$ if and only if
$c \lambda_1(S) + b \lambda_2(S) + a \lambda_3(S) \geq 0 $,
where $\lambda_1(S) \geq \lambda_2(S) \geq \lambda_3(S)$
are the eigenvalues of $S$. Hence the result follows.
\end{proof}
Now, we are prepared to prove the above claim about the
Lie semialgebra property of $\fw_0$
\begin{proof} (i) In case $\Gamma_0=\unity$, the pointed cone
$\fc_0 := \R{}^+_0\mathcal M(\Gamma_0)$ equals the ray
$\R{}^+_0 \unity$. By Lemma \ref{dual}, we obtain
$\mathfrak{c}_{\mathfrak{p}}^* = \{S \in \sym(3) \,|\, \tr S \geq 0\}$
and thus
$\fc_0^* = \mathfrak{sl}(3,\mathbb R) \oplus \mathbb R^+_0 \Gamma_0$,
where $\mathfrak{sl}(3,\R{})$ denotes the set of all
$3 \!\times\! 3$-matrices with trace zero.
Hence
\begin{equation}
\fw^*_0 = \so(3)^\perp \cap \fc^*_0 = \{S \in \sym(3)\;|\; \tr S \geq 0\}\;.
\end{equation}
For $A := \lambda \unity + \Omega \in \fw_0$ with $\lambda \geq 0$
and $\Omega \in \so(3)$ it follows
\begin{equation}
A^\perp \cap \fw^*_0 =
\begin{cases}
\{S \in \sym(3)\;|\; \tr S = 0\} & \text{for } \lambda > 0,\\[2mm]
\{S \in \sym(3)\;|\; \tr S \geq 0\} & \text{for } \lambda = 0,
\end{cases}
\end{equation}
and thus
\begin{equation}
T_A \fw_0 = (A^\perp \cap \fw^*_0)^\perp =
\begin{cases}
\so(3) \oplus \R{} \unity & \text{for } \lambda > 0,\\[2mm]
\so(3) & \text{for } \lambda = 0.
\end{cases}
\end{equation}
Thereby the inclusion
$[A, T_A \fw_0] \subset T_A \fw_0$ is obviously always satisfied
and hence $\fw_0$ {\em is a Lie semialgebra for
$\Gamma_0=\unity$}.
\medskip
\noindent
(ii) In case $\Gamma_0=\diag(1,0,0)$, it is easy to see that the
pointed cone $\fc_0 := \R{}^+_0\mathcal M(\Gamma_0)$ actually
consists of all positive semidefinite
$3 \!\times\! 3$-matrices. Here, Lemma \ref{dual} yields
$\mathfrak{c}_{\mathfrak{p}}^* = \fc_0$. This reflects the
well-known fact that the cone of all positive semidefinite
matrices is self-dual \emph{within the space of all symmetric
matrices}. Hence
\begin{equation}
\fw^*_0 = \so(3)^\perp \cap \fc^*_0 = \sym(3) \cap \fc^*_0 = \fc_0\;.
\end{equation}
Now, for
$
A := \Gamma_0 + H_z =
\left[\begin{smallmatrix}
1 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{smallmatrix}\right]
+
\left[\begin{smallmatrix}
0 & -1 & 0\\
1 & 0 & 0\\
0 & 0 & 0
\end{smallmatrix}\right]
\in \fw_0
$
we obtain
\begin{equation}
A^\perp \cap \fw^*_0 =
\left\{
\begin{bmatrix}
0 & 0\\
0 & S
\end{bmatrix}
\;\Big|\;
S\; \in \sym(2)\,,\, S\geq 0 \right\}
\end{equation}
and therefore
\begin{equation}
T_A \fw_0 = (A^\perp \cap \fw^*_0)^\per
=
\so(3) \oplus
\text{span}
\left\{
\Gamma_0,
p_y,
p_z
\right\}.
\end{equation}
Finally, for disproving the inclusion
$[A, T_A \fw_0] \subset T_A \fw_0$ consider the
commutator of $A\in\fw_0$ and
$
B := p_z
\in\;T_A \fw_0\;.
$
It follows
$
[A,B] =
-H_z
+
\diag(-2,2,0)
$
which clearly violates the inclusion $[A, T_A \fw_0] \subset T_A \fw_0$.
Thus $\fw_0$ is {\em not a Lie semialgebra for $\Gamma_0=\diag(1,0,0)$}.
\medskip
\noindent
(iii) In case $\Gamma_0=\diag(1,1,0)$, we obtain by Lemma
\ref{dual} the following description
\begin{equation*}
\fw^*_0 = \so(3)^\perp \cap \fc^*_0 = \fc^*_{\fp}=
\{S \in \sym(3) \,|\, \lambda_2(S) +\lambda_3(S) \geq 0\}\;.
\end{equation*}
Now, let $A:=\Gamma_0+H_y$. Then, it is easy to see that
\begin{equation*}
\begin{split}
A^\perp &\cap \fw^*_0 \supseteq \\
& \R{}^+_0\text{conv}
\left\{
\diag(1,-1,1),
\diag(-1,1,1),
(p_z+E_{33})
\right\}\,.
\end{split}
\end{equation*}
Moreover,
for $S \in A^\perp \cap \fw^*_0$ one has the conditions
\begin{equation*}
\langle\Gamma_0,S\rangle = 0
\quad\text{and}\quad
\langle\Theta\Gamma_0\Theta^\top,S\rangle \geq 0
\end{equation*}
for all $\Theta \in SO(3)$.
Now, differentiating the second condition with respect to
$\Theta \in SO(3)$ shows $\langle[\Omega,\Gamma_0],S\rangle = 0$
for all $\Omega \in \so(3)$, i.e.~$[\Omega,\Gamma_0]$
belongs to $(A^\perp \cap \fw^*_0)^\perp$. Thus one has
\begin{equation*}
T_A \fw_0 = (A^\perp \cap \fw^*_0)^\per
\supseteq \so(3) \oplus \R{}A \oplus
\text{span}
\left\{
p_x,
p_y
\right\}.
\end{equation*}
Hence, counting dimensions finally yields
\begin{equation*}
T_A \fw_0 = \so(3) \oplus
\text{span}
\left\{
\Gamma_0,
p_x,
p_y
\right\}\,.
\end{equation*}
To disprove the set inclusion $[A, T_A \fw_0] \subset T_A \fw_0$
consider the commutator of $A\in\fw_0$ and
$
B :=
p_y
\in\;T_A \fw_0\;.
$
The computation is left to the reader (see Tab.~\ref{tab:comm-tab}). The
result clearly violates the inclusion $[A, T_A \fw_0] \subset T_A \fw_0$
and thus $\fw_0$ is {\em not a Lie semialgebra} for
$\Gamma_0=\diag(1,1,0)$ either.
\medskip
\noindent
(iv) For $\Gamma_0=\diag(a,b,c)$ with $a > b > c \geq 0$ and
$A := \Gamma_0 + H_\nu$ with $\nu \in \{x,y,z\}$, the same
arguments as above show that $T_A \fw_0$ is given by
\begin{equation*}
T_A \fw_0 = \so(3) \oplus
\text{span}
\left\{
\Gamma_0,
p_x,
p_y,
p_z
\right\}\,.
\end{equation*}
Therefore, an appropriate choice of $B=p_\nu$ with
$\nu \in \{x,y,z\}$ demostrates again that $\fw_0$ is
{\em not a Lie semialgebra} in the general case
$\Gamma_0=\diag(a,b,c)$ with $a > b > c \geq 0$ either.
\end{proof}
Note that in all the above cases the tangent space
of $\fw_0$ has the following form
\begin{equation*}
T_A \fw_0 = \so(3) \oplus \R{}\Gamma_0 \oplus
\rT_{\Gamma_0} \mathcal{O}_{SO(3)}(\Gamma_0)\,,
\end{equation*}
where the tangent space
of the orbit $\mathcal{O}_{SO(3)}(\Gamma_0)$
at $\Gamma_0$ is given by
$\rT_{\Gamma_0} \mathcal{O}_{SO(3)}(\Gamma_0)
= \{[\Omega,\Gamma_0] \,|\, \Omega \in \so(3)\}$.
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{intro}
\subsection{}
The guiding theme of this note is the relation between the combinatorics
of a projective hyperplane arrangement and algebro-geometric and
intersection-theoretic invariants of the union of the hyperplanes.
Our results are as follows. We work over a field $k$. We consider a hyperplane
arrangement ${\mathscr A}$ in ${\mathbb{P}}^n$, and the corresponding central arrangement
${\widehat{\cA}}$ in $k^{n+1}$. We denote by $A\subseteq {\mathbb{P}}^n$ the union of the
hyperplanes in ${\mathscr A}$. We let $\chi_{{\widehat{\cA}}}(t)$ be the characteristic
polynomial of ${\widehat{\cA}}$, and denote by $\underline{\chi_{{\widehat{\cA}}}}(t)$ the quotient
$\chi_{\widehat{\cA}}(t)/(t-1)$ (also a polynomial).
We let $M({\mathscr A})$ be the complement ${\mathbb{P}}^n \smallsetminus A$.
\begin{theorem}\label{Grothintro}
Let ${\mathbb{L}}$ be the class of the affine line in the Grothendieck group of $k$-varieties
$K(\Var_k)$. Then $\underline{\chi_{\widehat{\cA}}}({\mathbb{L}})$ evaluates the class of the
complement $M({\mathscr A})$ in~$K(\Var_k)$.
\end{theorem}
If the arrangement is defined over a finite field, Theorem~\ref{Grothintro}
shows how to recover the Poincar\'e polynomial of an arrangement by counting
points, an observation made by several authors (see e.g., Theorem~2.69 in
\cite{MR1217488}). Placing this elementary result in the Grothendieck ring
$K(Var_k)$ for arbitrary $k$ has not-so-elementary applications: for example,
we note that the Orlik--Solomon theorem relating the characteristic polynomial
to the Poincar\'e polynomial of the complement of an arrangement
(\cite{MR558866}) is an immediate consequence of this result
(Corollary~\ref{OSredux}).
Another application is the determination of the stable rational equivalence
class of $M({\mathscr A})$
(Corollary~\ref{stableb}).
\subsection{}
As observed in \cite{MR2504753}, Proposition~2.2, the information carried by the
{Gro\-then\-dieck} class of a union of linear subspaces is equivalent to the information
in the image of its `Chern--Schwartz--MacPherson' (CSM) class in $A_*{\mathbb{P}}^n$.
Therefore, Theorem~\ref{Grothintro} may be recast in terms of CSM classes:
\begin{theorem}\label{CSMint}
The Chern--Schwartz--MacPherson class of
the complement $M({\mathscr A})$ may be obtained by
replacing $t^k$ with $[{\mathbb{P}}^k]$ in $\underline{\chi_{\widehat{\cA}}}(t+1)$.
\end{theorem}
For a reminder on Chern--Schwartz--MacPherson classes, see~\S\ref{CSMintro}.
J.~Huh has also observed that the additivity properties of these classes lead to
formulas for the characteristic polynomial: see Remark~26 in~\cite{Huh}.
In the particular case of {\em free\/} arrangements, we prove the following:
\begin{theorem}\label{freeintro}
Let ${\mathscr A}$ be a projective arrangement such that the corresponding affine arrangement
${\widehat{\cA}}$ is free. Then the Chern--Schwartz--MacPherson class of the complement
$M({\mathscr A})$ equals $c(\Omega^1_{{\mathbb{P}}^n}(\log A)^\vee)\cap [{\mathbb{P}}^n]$.
\end{theorem}
We view Theorem~\ref{CSMint} as the primary result,
as it holds for arbitrary arrangements, and Theorem~\ref{freeintro} as a computation
in the particular case of free arrangements. We obtain Theorem~\ref{freeintro} as
a consequence of a result of {Musta\c{t}\v{a}}{} and Schenck (\cite{MR1843320}).
It is natural to question whether a similar result may hold for a substantially more
general situation. The following conjecture appears to be consistent with all
known cases:
\begin{conj}
If $X$ is a locally quasi-homogeneous free divisor in a nonsingular variety~$V$,
then ${c_{\text{SM}}}(1\hskip-3.5pt1_{V\smallsetminus X})$ equals $c(\Omega^1_V(\log X)^\vee) \cap [V]$.
\end{conj}
This statement holds if $X$ is a divisor with simple normal crossings in $V$,
and (as verified in Theorem~\ref{freeintro}) if $V={\mathbb{P}}^n$ and $X$ is a free hyperplane
arrangement. The restriction to locally quasi-homogeneous divisors is suggested by
the case $V=$ surface, studied by Xia Liao \cite{Liao}.
(We recall that a hypersurface is `locally quasi-homogeneous' if at each point
it admits a weighted homogeneous equation, with positive weights, with respect
to some set of analytic parameters.)
\subsection{}
The subtlest part of the information carried by the Chern--Schwartz--MacPherson
class of a hypersurface in a nonsingular variety amounts to the {\em Segre class\/}
of its singularity (`Jacobian') subscheme. Extracting this information and applying
Theorem~\ref{CSMint} gives the following
(in characteristic~$0$):
\begin{corol}\label{Segreintro}
The characteristic polynomial $\chi_{\widehat{\cA}}(t)$ for a hyperplane arrangement ${\mathscr A}$
in ${\mathbb{P}}^n$ may be recovered from the degree of $A$ and
the image in $A_*{\mathbb{P}}^n$ of the Segre class of the singularity subscheme of $A$.
\end{corol}
The precise relation between these invariants is given in~Theorem~\ref{segrethm}.
We note that Wakefield and Yoshinaga have proven (\cite{MR2424913}) that (essential)
hyperplane arrangements may be recovered from their singularity subschemes.
Corollary~\ref{Segreintro} shows that the Segre class of the same scheme
suffices in order to recover the most basic combinatorial information of the
arrangement. Putting together Theorem~\ref{segrethm} and Corollary~\ref{OSredux}
yields the following surprisingly simple relation: for any hyperplane arrangement
in ${\mathbb{P}}^n_{\mathbb{C}}$ and $k\le n$
\begin{equation*}
\tag{*}
\rk H^k(M({\mathscr A}),{\mathbb{Q}}) = \sum_{i=0}^k \binom k i (d-1)^{k-i} \sigma_i\quad,
\end{equation*}
where $d$ is the degree of the arrangement and the integers $\sigma_i$ are
determined by the push-forward of the Segre class of the singularity subscheme~$S$:
\[
\sum_{i=0}^n \sigma_i \cap [{\mathbb{P}}^{n-i}] = [{\mathbb{P}}^n]-\iota_* s(S,{\mathbb{P}}^n) \quad.
\]
This equality should be compared with Huh's formula expressing the Betti numbers
of the complement as mixed multiplicities; see the proof of Corollary~25 in \cite{Huh}.
A referee points out that hyperplane arrangements are the only hypersurfaces
for which (*) holds. Indeed, the rank of $H^1$ of the complement of a
hypersurface is one less than the number of distinct irreducible components
(\cite{MR1194180}, Chapter~4, Proposition~1.3). If (*) holds for a degree~$d$
hypersurface, then the rank of $H^1$ equals $d-1$, and it follows that the
hypersurface consists of $d$ hyperplanes.
\subsection{}
Finally, we discuss positivity of Chern classes of hyperplane arrangements.
Chern--Schwartz--MacPherson classes arising in
combinatorial situations have a mysterious tendency to be effective. For example,
the Chern--Schwartz--MacPherson class of a toric variety is effective
(\cite{MR2209219}); CSM classes of Schubert varieties are conjecturally effective
(\cite{MR2448279}, \S4); and all CSM classes of graph hypersurfaces computed to
date are effective (cf.~\cite{MR2504753}, Conjecture~1.5). It is natural to inquire
about the positivity of CSM classes of hyperplane arrangements.
\begin{theorem}\label{positivintro}
\begin{itemize}
\item The CSM class of a generic hyperplane arrangement $X\subseteq {\mathbb{P}}^n$ of
degree $d\le n+3$ (for $n$ even), resp.~$d\le n+4$ (for $n$ odd) is effective.
\item The CSM class of every free hyperplane arrangement of degree $\le n$
in ${\mathbb{P}}^n$, $n\le 8$, is effective.
\end{itemize}
\end{theorem}
However, in \S\ref{posifree} we exhibit a free arrangement of degree~$9$ in
${\mathbb{P}}^9$ that has non-effective Chern--Schwartz--MacPherson class. The effectivity
of the CSM class of an arrangement has a straightforward combinatorial
interpretation, which we give in Proposition~\ref{effe}.
\medskip
I thank an anonymous referee for valuable comments.
\section{Grothendieck classes}\label{grothsec}
\subsection{}\label{genera}
We work over a field $k$. For considerations involving Chern--Schwartz--MacPherson
classes, $k$ will be assumed to be algebraically closed, of characteristic~$0$
(see the comments at the beginning of \S\ref{CSMintro}).
We consider an arrangement ${\mathscr A}$ of distinct hyperplanes $H_i$ in ${\mathbb{P}}^n$,
$i=1,\dots,d$, $d\ge 2$. We also consider the corresponding arrangement
${\widehat{\cA}}$ of hyperplanes ${\widehat{H}}_i\subset V=k^{n+1}$; this has the advantage of being
{\em central.\/} While we are interested in the projective geometry of ${\mathscr A}$, basic
definitions and results in the literature are more often given for the associated affine
arrangement ${\widehat{\cA}}$.
We will denote by $A$ and ${\widehat{A}}$, respectively the union of the hyperplanes in ${\mathscr A}$
and ${\widehat{\cA}}$, respectively. We will view~$A$ as a reduced singular hypersurface
of~${\mathbb{P}}^n$ of degree~$d\ge 2$.
We will be especially interested in the complements $M({\widehat{\cA}})=V\smallsetminus {\widehat{A}}$,
$M({\mathscr A})={\mathbb{P}}^n\smallsetminus A$. Note that $M({\widehat{\cA}})$ is a trivial $k^*$-fibration over
$M({\mathscr A})$.
We will denote by $L({\widehat{\cA}})$ the poset of intersections $\cap_{i\in J} {\widehat{H}}_i$,
partially ordered by reverse inclusion. (For $x,y\in L({\widehat{\cA}})$ we will write
interchangeably $y\supseteq x$ or $y\le x$ to denote that the subspace
$y$ contains the subspace $x$, i.e., that $y$ precedes $x$ in the poset $L({\widehat{\cA}})$.)
The space $V$ itself is viewed as
the intersection over $J=\emptyset$, and is denoted $0\in L({\widehat{\cA}})$. As ${\widehat{\cA}}$
is central, $L({\widehat{\cA}})$ also has a maximum $1$, corresponding to $\cap_i {\widehat{H}}_i$.
The arrangement is {\em essential\/} if $\cap_i {\widehat{H}}_i$ is the origin; this assumption
will not be needed in this paper.
The {\em M\"obius function\/} of $L({\widehat{\cA}})$ is defined on pairs $x\le y$ by the
following prescription:
\begin{align*}
\mu(x,x) &= 1 \quad\text{for all $x\in L({\widehat{\cA}})$} \\
\sum_{x\le z\le y} \mu(x,z) &= 0 \quad\text{for all $x<y$ in $L({\widehat{\cA}})$.}
\end{align*}
Write $\mu(x)$ for $\mu(0,x)$.
The {\em characteristic polynomial\/} of the arrangement ${\widehat{\cA}}$ is
\[
\chi_{\widehat{\cA}}(t):=\sum_{x\in L({\widehat{\cA}})} \mu(x) t^{\dim x}=t^{n+1}+\cdots\quad.
\]
Note that $\chi_{\widehat{\cA}}(1)=\sum_{0\le z\le 1} \mu(0,z)=0$.
The {\em Poincar\'e polynomial\/} of the arrangement~is
\[
\pi_{\widehat{\cA}}(t):=\sum_{x\in L({\widehat{\cA}})} \mu(x) (-t)^{\codim x} = (-t)^{n+1}\cdot \chi_{\widehat{\cA}}(-t^{-1})
\quad.
\]
In any extension in which $\chi_{\widehat{\cA}}(t)$ factors completely,
$\chi_{\widehat{\cA}}(t)=(t-\alpha_1)\cdots (t-\alpha_{n+1})$, we have
\begin{align*}
\pi_{\widehat{\cA}}(t) = (-t)^{n+1}\cdot \chi_{\widehat{\cA}}(-t^{-1})
&=(-t)^{n+1} \left(-\frac 1t-\alpha_1\right) \cdots \left(-\frac 1t-\alpha_{n+1}\right) \\
& = (1+\alpha_1 t)\cdots (1+\alpha_{n+1} t)\quad.
\end{align*}
Note that $\pi_{{\widehat{\cA}}}(-1)=\chi_{{\widehat{\cA}}}(1)=0$:
we may assume that $\alpha_{n+1}=1$, and we let
\[
\underline{\chi_{\widehat{\cA}}}(t)=(t-\alpha_1)\cdots (t-\alpha_n)=\frac{\chi_{\widehat{\cA}}(t)}{t-1}
\quad,\quad
\underline{\pi_{\widehat{\cA}}}(t)=(1+\alpha_1 t)\cdots (1+\alpha_n t)= \frac{\pi_{\widehat{\cA}}(t)}
{1+t}\quad.
\]
Thus, $\underline{\chi_{\widehat{\cA}}}(t)$ and $\underline{\pi_{\widehat{\cA}}}(t)$ are polynomials of degree
at most $n$. The results in this paper are most naturally expressed in terms of these
polynomials. As we will see in a moment,
$\underline{\pi_{\widehat{\cA}}}(t)$ has nonnegative coefficients.
\subsection{}
Our first task is the computation of the Grothendieck class of the complement $M({\mathscr A})$
in terms of the characteristic polynomial. We denote by $[X]$ the class of
a $k$-variety $X$ in the Grothendieck ring $K(\Var_k)$ of varieties.
The class of ${\mathbb{A}}^1_k$ is denoted~${\mathbb{L}}$.
\begin{theorem}\label{Gclassc}
With notation as above:
\[
[M({\mathscr A})]=\underline{\chi_{\widehat{\cA}}}({\mathbb{L}})\quad.
\]
\end{theorem}
\begin{proof}
For $x\in L({\widehat{\cA}})$, let
\[
x^\circ = x \smallsetminus \cup_{y> x} y
\]
be the complement of the union of smaller intersections.
Write $[x]={\mathbb{L}}^{\dim x}$, $[x^\circ]$ for the classes in $K(\Var_k)$ of the corresponding
subsets of $V$. Then
\[
[y] = \sum_{x\ge y} [x^\circ]\quad.
\]
Applying M\"obius inversion (\cite{MR1217488}, Proposition~2.39), this gives
\[
[y^\circ] = \sum_{x\subseteq y} \mu(y,x) [x]
\]
and in particular
\[
[M({\widehat{\cA}})] = [0^\circ] = \sum_{x\in L({\widehat{\cA}})} \mu(0,x)[x]
=\sum_{x\in L({\widehat{\cA}})} \mu(x){\mathbb{L}}^{\dim x} = \chi_{\widehat{\cA}}({\mathbb{L}})\quad.
\]
Since $M({\widehat{\cA}})$ fibers over $M({\mathscr A})$, with $k^*$ fibers, we have
$[M({\widehat{\cA}})]=({\mathbb{L}}-1) \cdot [M({\mathscr A})]$, and the stated formula follows.
\end{proof}
\subsection{}
If ${\mathscr A}$ is defined over a finite field ${\mathbb{F}}_q$, the content of Theorem~\ref{Gclassc}
is that the information carried by the characteristic polynomial of ${\widehat{\cA}}$ may be
recovered from counting points of $A$ over ${\mathbb{F}}_{q^r}$. This observation is of
course not new, see \cite{MR1409420} (e.g.~Theorem~2.2) for a thorough study of
characteristic polynomials of arrangements defined over finite fields.
For complex arrangements, Theorem~\ref{Gclassc} has the following immediate
consequence:
\begin{corol}
With notation as above, the Deligne-Hodge polynomial of the complement
$M({\mathscr A})$ of a complex hyperplane arrangement in ${\mathbb{P}}^n$ equals
$\underline{\chi_{\widehat{\cA}}}(uv)$.
\end{corol}
\noindent The Deligne-Hodge polynomial is the polynomial
$\sum_{p,q} e^{p,q}(X) u^p v^q$, where
$e^{p,q}(X)=\sum_k (-1)^k h^{p,q} (H_c^k(X,{\mathbb{Q}}))$. It is determined by the
Grothendieck class of $X$, by its well-known additivity/multiplicativity properties
(\cite{MR873655}). As the polynomial for ${\mathbb{L}}$ is $uv$, the statement follows
immediately from Theorem~\ref{Gclassc}.
Taking into account the fact the mixed Hodge structure on the cohomology of
the complement of a hyperplane arrangement is pure (\cite{MR1131042}), we
obtain the following:
\begin{corol}\label{OSredux}
$\underline{\chi_{\widehat{\cA}}}(t)=\sum_{k=0}^n (-1)^{n+k} \rk H_c^{n+k}(M({\mathscr A}),{\mathbb{Q}})\,t^k$.
\end{corol}
\noindent (To keep track of weights: the Hodge structures of $H_c^{n+k}$ and $H^{n-k}$
are compatible through the Poincar\'e pairing, see~\cite{MR873655}, \S1.4 (f);
$H^{n-k}$ is of type $(n-k,n-k)$ by~\cite{MR1131042}; thus $H_c^{n+k}$
is of type $(k,k)$.)
As $\rk H_c^{n+k}(M({\mathscr A}),{\mathbb{Q}})=\rk H^{n-k}(M({\mathscr A}),{\mathbb{Q}})$, this statement is
equivalent to
\[
\underline{\pi_{\widehat{\cA}}}(t)=\sum_{k=0}^n \rk H^k(M({\mathscr A}),{\mathbb{Q}})\,t^k\quad.
\]
This shows that the coefficients of $\underline{\pi_{\widehat{\cA}}}(t)$ are nonnegative.
Since $M({\widehat{\cA}})\cong M({\mathscr A})\times k^*$, this also proves
\[
\pi_{\widehat{\cA}}(t)=\sum_{k=0}^{n+1} \rk H^k(M({\widehat{\cA}}),{\mathbb{Q}})\,t^k\quad,
\]
a classic result of Orlik and Solomon (\cite{MR1217488}, Theorem 5.93).
The approach presented here appears particularly straightforward, since it
shows that this result follows directly from Theorem~\ref{Gclassc}, which is
a trivial consequence of M\"obius inversion.
\subsection{}
Another piece of information that may be derived from the Grothendieck class is the
stable birational equivalence class.
Two projective nonsingular irreducible
varieties $X$, $Y$ are {\em stably birational\/} if $X\times {\mathbb{P}}^m$
is birational to $Y\times {\mathbb{P}}^n$ for some $m$, $n$. Stable birational equivalence
classes of complete nonsingular varieties generate a ring ${\mathbb{Z}}[SB]$, with addition
defined by disjoint union and multiplication by product.
Every variety (possibly noncomplete or singular) has a well-defined class in
${\mathbb{Z}}[SB]$. The ring ${\mathbb{Z}}[SB]$ is defined and studied in \cite{MR1996804}.
\begin{corol}\label{stableb}
Let ${\mathscr A}$ be a complex hyperplane arrangement in ${\mathbb{P}}^n$, and let $A\subset
{\mathbb{P}}^n$ be the union of its components.
Then the class of $A$ in ${\mathbb{Z}}[SB]$ equals
$1-\underline{\chi_{{\widehat{\cA}}}}(0)=1-(-1)^n \rk H^n(M({\mathscr A}),{\mathbb{Q}})$.
\end{corol}
\begin{proof}
The ring ${\mathbb{Z}}[SB]$ is isomorphic to $K(\Var)/({\mathbb{L}})$ (\cite{MR1996804},
Theorem~2.3 and Proposition~2.7). Setting ${\mathbb{L}}=0$ in Theorem~\ref{Gclassc}
shows that $[M({\mathscr A})]$ is congruent to the constant term of $\underline{\chi_{\widehat{\cA}}}(0)$ in
${\mathbb{Z}}[SB]$. Since $[{\mathbb{P}}^n]=1$ in ${\mathbb{Z}}[SB]$, it follows that the stable
birational equivalence class of $A$ equals $1-\underline{\chi_{\widehat{\cA}}}(0)\in {\mathbb{Z}}
\subseteq {\mathbb{Z}}[SB]$.
The equality $\underline{\chi_{\widehat{\cA}}}(0)=(-1)^n \rk H^n(M({\mathscr A}),{\mathbb{Q}})$ follows from
Corollary~\ref{OSredux}.
\end{proof}
\section{Chern--Schwartz--MacPherson classes}\label{CSMsec}
\subsection{}\label{CSMintro}
We now assume the ground field $k$ to be algebraically closed, of characteristic~$0$.
This assumption on the characteristic could likely be relaxed: in any environment
in which resolution of singularities is available, one may define a class satisfying
inclusion-exclusion and the basic normalization property of
Chern--Schwartz--MacPherson classes (see~\cite{MR2282409}, Definition~4.4
and~\S3.3); these are the only tools needed in this section. However,
characteristic~$0$ is necessary for the main covariance property of these classes
recalled below (see~\cite{MR2282409}, \S5.2), and some results on these classes
are currently only known in characteristic zero. Hence, we prefer to conservatively
adopt this assumption, at the price of limiting the scope of the results. (Many
interesting arrangements can only be realized in positive characteristic.)
Recall that there is a homomorphism from the group of constructible functions on a variety
$X$ to the Chow group of $X$, $\varphi \mapsto c_*(\varphi)$, covariant with respect
to proper maps and such that $c_*(1\hskip-3.5pt1_X)$ equals the total Chern class of the tangent
bundle of~$X$ if $X$ is nonsingular (\cite{MR0361141}; Example~19.1.7 in \cite{85k:14004};
\cite{MR2282409}). The key covariance property of $c_*$ amounts to the fact that if
$\alpha: X\to Y$ is proper, and $\varphi$ is a constructible function, then
$c_*(\alpha_*(\varphi))=\alpha_*(c_*(\varphi))$. Here the push-forward of constructible
functions is defined by taking Euler characteristics of fibers.
We consider this theory over $X={\mathbb{P}}^n$. Every subvariety (possibly singular, noncomplete)
$Y\subseteq {\mathbb{P}}^n$ determines a {\em Chern--Schwartz--MacPherson\/} (CSM) class
${c_{\text{SM}}}(Y):=c_*(1\hskip-3.5pt1_Y)\in A_*{\mathbb{P}}^n$.
\begin{theorem}\label{CSM}
Let ${\mathscr A}$ be a hyperplane arrangement in ${\mathbb{P}}^n$; denote by $A$ the union of the
hyperplanes in ${\mathscr A}$, by $M({\mathscr A})$ the complement of $A$ in ${\mathbb{P}}^n$, and let
$\underline{\chi_{\widehat{\cA}}}(t)$ be the polynomial introduced in \S\ref{grothsec}. Then
\[
{c_{\text{SM}}}(M({\mathscr A}))=\underline{\chi_{\widehat{\cA}}}(t+1)\quad,
\]
where the right-hand side is interpreted as a class in $A_*{\mathbb{P}}^n$ by replacing
$t^k$ by $[{\mathbb{P}}^k]$, $k=0,\dots,n$.
\end{theorem}
This result may be obtained by applying the formula relating CSM classes and
Grothendieck classes for varieties obtained by elementary set-theoretic operations
on linear subspaces (Proposition~2.2 in \cite{MR2504753}): the CSM class is
obtained by viewing the Grothendieck class as a polynomial in ${\mathbb{T}}=[k^*]$
and replacing ${\mathbb{T}}^r$ with~$[{\mathbb{P}}^r]$. Since ${\mathbb{L}}={\mathbb{T}}+1$,
Theorem~\ref{CSM} follows immediately from Theorem~\ref{Gclassc}.
In more conventional notation, Theorem~\ref{CSM} states the following:
\[
{c_{\text{SM}}}(M({\mathscr A}))=\left(h^{n+1} \chi_{{\widehat{\cA}}}\left(1+\frac 1h\right)\right)\cap [{\mathbb{P}}^n]
=\left(h^n \underline{\chi_{\widehat{\cA}}}\left(1+\frac 1h\right)\right)\cap [{\mathbb{P}}^n]
\quad,
\]
where $h$ denotes the hyperplane class in ${\mathbb{P}}^n$, and the expressions on the
right should be interpreted as the polynomials obtained by expanding them.
For the sake of completeness and ease of reference in later sections, we
give here a direct proof of these formulas.
\begin{proof}
The second formula is just a restatement of the first one. To obtain the first one,
consider the functions from $L({\widehat{\cA}})$ to the abelian group of constructible functions
on $k^{n+1}$, defined by $x \mapsto 1\hskip-3.5pt1_x$, $x\mapsto 1\hskip-3.5pt1_{x^\circ}$.
(Here $x^\circ$ is as in the proof of Theorem~\ref{Gclassc}.)
We have
\[
1\hskip-3.5pt1_y = \sum_{x\ge y} 1\hskip-3.5pt1_{x^\circ}\quad,
\]
and hence
\[
1\hskip-3.5pt1_{y^\circ} = \sum_{x\subseteq y} \mu(y,x) 1\hskip-3.5pt1_x
\]
by M\"obius inversion. In particular,
\[
1\hskip-3.5pt1_{M({\widehat{\cA}})} = \sum_{x\in L({\widehat{\cA}})} \mu(x) 1\hskip-3.5pt1_x\quad.
\]
For positive dimensional $x$, the characteristic functions $1\hskip-3.5pt1_x$ on $k^{n+1}$ are
pull-backs of the corresponding characteristic functions from ${\mathbb{P}}^n$. It follows that
\[
1\hskip-3.5pt1_{M({\mathscr A})} = \sum_{x\in L({\widehat{\cA}})} \mu(x) 1\hskip-3.5pt1_{\underline x}
\]
on ${\mathbb{P}}^n$, where $\underline x$ denotes the projective subspace corresponding to
$x\subset V$ (and $\underline x=\emptyset$ if $x$ is the origin in $V$).
Applying MacPherson's natural transformation, this shows that
\[
{c_{\text{SM}}}(M({\mathscr A})) = \sum_{x\in L({\widehat{\cA}})} \mu(x) {c_{\text{SM}}}(\underline x)\in A_*{\mathbb{P}}^n\quad.
\]
Now $\underline x\cong {\mathbb{P}}^{\dim x-1}$, and hence
\[
{c_{\text{SM}}}(\underline x) = (1+h)^{\dim x} h^{n+1-\dim x}\cap [{\mathbb{P}}^n]
\]
as a class in ${\mathbb{P}}^n$.
(In particular ${c_{\text{SM}}}(\underline x)=0$ if $x$ is the origin in $V$, as it should, since
$h^{n+1}\cap [{\mathbb{P}}^n]=0$.) It follows that
\[
{c_{\text{SM}}}(M({\mathscr A})) = \left(h^{n+1} \sum_{x\in L({\widehat{\cA}})} \mu(x) \frac{(1+h)^{\dim x}}{h^{\dim x}}
\right) \cap [{\mathbb{P}}^n]
=\left(h^{n+1} \chi_{{\widehat{\cA}}}\left(1+\frac 1h\right)\right) \cap [{\mathbb{P}}^n]\quad,
\]
as stated.
\end{proof}
\begin{corol}\label{poinc}
With the same notation:
\[
{c_{\text{SM}}}(M({\mathscr A}))=(1+h)^n \underline{\pi_{\widehat{\cA}}}\left(\frac {-h}{1+h}\right)\cap [{\mathbb{P}}^n]
=\pi_{{\widehat{\cA}}}\left(\frac {-h}{1+h}\right)\cap \left(c(T{\mathbb{P}}^n)\cap [{\mathbb{P}}^n]\right)\quad.
\]
\end{corol}
\begin{example}\label{norcro}
Suppose ${\mathscr A}$ is {\em generic,\/} i.e., it consists of $d$ hyperplanes meeting with
normal crossings. Then
\[
{c_{\text{SM}}}(M({\mathscr A}))=c(\Omega(\log A)^\vee)\cap [A]=\frac 1{(1+h)^d}\cap
(c(T{\mathbb{P}}^n)\cap [{\mathbb{P}}^n])\quad,
\]
see e.g., Theorem~1 in \cite{MR2001d:14008}.
By Corollary~\ref{poinc}, then:
\[
\pi_{{\widehat{\cA}}}\left(\frac {-h}{1+h}\right)=\frac 1{(1+h)^d}
\]
modulo $h^{n+1}$. Setting $t=\frac {-h}{1+h}$, i.e., $h=\frac {-t}{1+t}$, gives
\[
\pi_{{\widehat{\cA}}}(t)\equiv (1+t)^d\mod t^{n+1}\quad.
\]
The coefficient of $t^{n+1}$ in $\pi_{{\widehat{\cA}}}(t)$ is then determined by the fact
that $\pi_{{\widehat{\cA}}}(-1)=0$.
For instance, $\pi_{{\widehat{\cA}}}(t)=(1+t)^d$ for $d\le n+1$.
The {\em Boolean arrangement,\/} where ${\widehat{\cA}}$ consists of the $n+1$ coordinate
hyperplanes in $V\cong k^{n+1}$, is of this type; cf.~\S2.3 in \cite{MR1217488}.
\hfill$\lrcorner$\end{example}
\section{Free arrangements}\label{free}
\subsection{}
The key feature of Example~\ref{norcro} is the (well known) fact that if $D$ is a divisor
with normal crossings
in a nonsingular variety, then the bundle $\Omega^1(\log D)$ is locally free and its
total Chern class computes the Chern--Schwartz--MacPherson class of the complement
of $D$. Thus, for normal crossings arrangements the left-hand side of the formulas
in Theorem~\ref{CSM} and Corollary~\ref{poinc} may be viewed as the Chern class
of a bundle, and the formulas may be interpreted as an alternative computation of
this class.
In this section we show that this interpretation extends to {\em free arrangements,\/}
by which we mean projective arrangements ${\mathscr A}$ such that the corresponding
affine arrangements ${\widehat{\cA}}$ are free in the sense of \cite{MR1217488}.
For these arrangements, the sheaf $\Omega^1_{{\mathbb{P}}^n}(\log A)$ of differential
$1$-forms with logarithmic poles along the union $A$ of the hyperplanes in ${\mathscr A}$
is locally free. We will prove:
\begin{theorem}\label{freethm}
For free arrangements ${\mathscr A}$ in ${\mathbb{P}}^n$,
\begin{equation*}
\tag{$\dagger$}
{c_{\text{SM}}}(M({\mathscr A}))=c(\Omega^1_{{\mathbb{P}}^n}(\log A)^\vee)\cap [{\mathbb{P}}^n]\quad.
\end{equation*}
\end{theorem}
We recall that sections of the logarithmic sheaf $\Omega^1_{{\mathbb{P}}^n}(\log A)$
are differential forms~$\omega$ such that both $f\omega$ and $f d\omega$ are
regular, where $f=0$ is the equation for $A$. This definition, due to Saito,
generalizes Deligne's definition for normal crossing divisors. In general, a divisor
$D$ of a variety $X$ is said to be `free' if this sheaf is locally free.
\subsection{}
Taken together, Theorem~\ref{CSM} and Theorem~\ref{freethm} give a relation
between the characteristic polynomial of an arrangement ${\mathscr A}$ and the
total Chern class
of $\Omega^1_{{\mathbb{P}}^n}(\log A)$ in the case of free arrangements.
Theorem~\ref{CSM} may be viewed as a generalization of this relation,
as it holds for arbitrary projective arrangements---notwithstanding the fact that its
proof is completely trivial, while the proof of the particular case of free arrangements
in the form of Theorem~\ref{freethm} requires some actual work. By our good fortune,
the main ingredient in this work may be found in a paper of M.~{Musta\c{t}\v{a}}{} and
H.~Schenck; Theorem~\ref{freethm} will be obtained as a consequence of Theorem~4.1
in \cite{MR1843320}.
This is one instance in which MacPherson's functorial theory of Chern
classes clearly provides the `right' generalization of Chern classes to noncomplete
(and/or singular) varieties, and it is tempting to guess that the equality in
Theorem~\ref{freethm} may hold for more general free divisors. Work of Xia Liao
(\cite{Liao}) shows that for a reduced curve $X$ in a nonsingular surface $V$, the
Chern--Schwartz--MacPherson class of the complement equals the Chern
class of the corresponding bundle of logarithmic derivations only if the Milnor
and Tjurina numbers of the singularities of $X$ agree. This indicates that a hypothesis
of local quasi-homogeneity is likely necessary for a generalization of
Theorem~\ref{freethm}; we conjecture as much in the introduction.
In any case, the question of comparing the CSM
class of the complement of a divisor $D$ and the Chern classes of the corresponding
sheaf of differential forms with logarithmic poles along $D$ appears to be interesting
and approachable. Theorem~5.13 in \cite{denhamschulze} indicates that a correction
term will be necessary if this sheaf is not locally free, as it provides such a term for
the corresponding generalization of Theorem~4.1 in \cite{MR1843320}
to the `locally tame' case.
\subsection{}\label{prelim}
The formula in Theorem~\ref{freethm} holds if $A$ is any divisor with simple
normal crossings in a nonsingular variety. For a proof of this elementary fact,
see e.g., Theorem~1 in \cite{MR2001d:14008} or Proposition 15.3 in \cite{MR1893006};
this observation may in fact be used to give an alternative treatment of CSM classes
(\cite{MR2282409}). The obvious strategy to prove Theorem~\ref{freethm} would
therefore be to apply resolution of singularities and reduce to the case of normal crossing
divisors. The behavior of the left-hand side of $(\dagger)$ through a resolution is
controlled by the covariance of CSM classes:
\begin{lemma}\label{trilem}
Let $X$ be a variety, and let $Y\subseteq X$ be a subscheme. Let $\rho: \Til X \to X$
be a proper map, and $Y'\subseteq \Til X$ any subscheme such that $\rho$ restricts
to an isomorphism of the complements $M(Y')$ of $Y'$ in $\Til X$ and $M(Y)$ of
$Y$ in $X$. Then
\[
{c_{\text{SM}}}(M(Y))=\rho_* {c_{\text{SM}}}(M(Y'))\quad.
\]
\end{lemma}
\begin{proof}
Under the hypotheses specified in the statement, $\rho_*(1\hskip-3.5pt1_{M(Y')})=1\hskip-3.5pt1_{M(Y)}$,
and the equality follows then immediately from the covariance property of
Chern--Schwartz--MacPherson classes (recalled in \S\ref{CSMintro}).
\end{proof}
Lemma~\ref{trilem} reduces the proof of Theorem~\ref{freethm} to verifying that
the Chern class of the bundle of differential forms with logarithmic poles is preserved
under push-forward for suitable blow-ups. The difficulty lies in the fact that the
bundle itself is {\em not\/} preserved under blow-ups.
However, Silvotti (\cite{MR1484696}) has analyzed the behavior of the logarithmic
bundle under blow-ups in the case of arrangements, and we feel that his analysis
should suffice in order to obtain a proof of Theorem~\ref{freethm}.
In fact, Silvotti proves (\cite{MR1484696}, Proposition~4.5) that if
$\Omega^1_X(\log D)$ splits as a direct sum of line bundles ${\mathscr L}_1,\dots,{\mathscr L}_n$,
then the corresponding bundle $\Omega^1_{\decor X}(\log \decor D)$ in the blow-up
along a subvariety~$Y$ also splits, and in fact
\[
\Omega^1_{\decor X}(\log \decor D) \cong
(\sigma^*{\mathscr L}_1 \otimes (-\mu_1 E)) \oplus \cdots \oplus
(\sigma^*{\mathscr L}_n \otimes (-\mu_n E))
\]
for non-negative integers $\mu_1,\dots,\mu_n$, where $E$ denotes the exceptional
divisor. (Terao proved that the splitting does occur in the case of free hyperplane
arrangements, cf.~Proposition~5.1 in \cite{MR1484696}.)
A proof of Theorem~\ref{freethm} follows easily if one could show that
{\em at most $\codim Y-1$ of the numbers $\mu_i$ are nonzero.\/}
In fact, given that Theorem~\ref{freethm} does hold, it seems that this must indeed
be the case. It would be nice to have a direct proof of this fact.
\subsection{}\label{compa}
Theorem~4.1 in~\cite{MR1843320} provides us with an alternative approach to
Theorem~\ref{freethm}. Denote by $\Omega^1$ the {\em module\/} of differential
forms with logarithmic poles along the central arrangement ${\widehat{\cA}}$ in $k^{n+1}$.
This is a graded module, hence it defines a coherent sheaf $\Til{\Omega^1}$ on
${\mathbb{P}}^n$. Under the assumption that the arrangement is free, $\Til{\Omega^1}$
is a rank-$(n+1)$ locally free sheaf on ${\mathbb{P}}^n$.
\begin{theorem}[{Musta\c{t}\v{a}}-Schenck]\label{MusSch}
If ${\widehat{\cA}}$ is an essential free arrangement, then
\[
c(\Til{\Omega^1})=\pi_{\widehat{\cA}} (h)\quad,
\]
where $h=c_1({\mathscr O}_{{\mathbb{P}}^n}(1))$.
\end{theorem}
\begin{remark}
In fact, {Musta\c{t}\v{a}}{} and Schenck prove the equality in Theorem~\ref{MusSch}
under the weaker hypothesis that the arrangement is {\em locally\/} free.
Also, we note that while the statement of our Theorem~\ref{freethm} assumes
the ground field to be algebraically closed of characteristic~$0$,
this assumption is not needed in the result of {Musta\c{t}\v{a}}{} and Schenck.
In Theorem~\ref{MusSch}, the right-hand side should be parsed as the truncation
of the Poincar\'e polynomial modulo $h^{n+1}$, as $h^{n+1}=0$ in ${\mathbb{P}}^n$.
Also, we recall that `essential' means that the intersection of all hyperplanes in
the projective arrangement ${\mathscr A}$ is empty.
\hfill$\lrcorner$
\end{remark}
We now prove Theorem~\ref{freethm} as a corollary of Theorems~\ref{CSM}
and~\ref{MusSch}.
\begin{lemma}\label{exseq}
Let ${\mathscr A}$ be a free arrangement. Then there is an exact sequence
\[
\xymatrix{
0 \ar[r] & \Omega^1_{{\mathbb{P}}^n}(\log A) \ar[r] &
\Til{\Omega^1}\otimes {\mathscr O}_{{\mathbb{P}}^n}(-1) \ar[r] & {\mathscr O}_{{\mathbb{P}}^n} \ar[r] & 0\quad.
}
\]
\end{lemma}
\begin{proof}
The Euler derivation
$x_0 \frac{\partial}{\partial x_0} + \cdots + x_n \frac{\partial}{\partial x_n}$ defines
an epimorphism $\Til{\Omega^1} \to {\mathscr O}_{{\mathbb{P}}^n} (1)$ (cf.~Proposition~4.27 in
\cite{MR1217488} for the dual statement). The shifted epimorphism
$\Til{\Omega^1}(-1) \to {\mathscr O}_{{\mathbb{P}}^n}$ is the natural extension of the standard
epimorphism ${\mathscr O}_{{\mathbb{P}}^n}(-1)^{\oplus(n+1)} \to {\mathscr O}_{{\mathbb{P}}^n}$
whose kernel defines the sheaf of differential forms over ${\mathbb{P}}^n$ (as in
\cite{MR0463157}, II.8.13). The kernel of this epimorphism is then the sheaf
of meromorphic differential forms satisfying the same conditions as the forms
in $\Omega^1$, and this is the definition of $\Omega^1_{{\mathbb{P}}^n}(\log A)$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{freethm}]
First, we note that we may assume the arrangement to be essential. Indeed,
every arrangement ${\mathscr A}$ in ${\mathbb{P}}^n$ is a cone over an essential arrangement
${\mathscr A}'$ in ${\mathbb{P}}^{n-k}$, for some $k\ge 0$. Assume the formula ($\dagger$)
in Theorem~\ref{freethm} is known for ${\mathscr A}'$:
\[
{c_{\text{SM}}}(M({\mathscr A}'))=c(\Omega^1_{{\mathbb{P}}^{n-k}}(\log A')^\vee)\cap [{\mathbb{P}}^{n-k}]\quad.
\]
Inductively, it suffices to show that the formula for ${\mathscr A}$ follows from this for $k=1$.
Write ${c_{\text{SM}}}(M({\mathscr A}'))=g(h)\cap [{\mathbb{P}}^{n-1}]$, for a polynomial $g$ of degree $\le n-1$;
we are assuming that $c(\Omega^1_{{\mathbb{P}}^{n-1}}(\log A')^\vee)=g(h)$.
We have ${c_{\text{SM}}}(A')=f(h)\cap [{\mathbb{P}}^{n-1}]$ for $f(h)=(1+h)^n-h^n-g(h)$, hence
by Proposition~5.2 in \cite{MR2504753}
\[
{c_{\text{SM}}}(A) = (1+h) f(h)\cap [{\mathbb{P}}^n] + [{\mathbb{P}}^0]\quad,
\]
and therefore
\begin{align*}
{c_{\text{SM}}}(M({\mathscr A}))&={c_{\text{SM}}}({\mathbb{P}}^n)-{c_{\text{SM}}}(A) \\
&=\big(((1+h)^{n+1}-h^{n+1})-((1+h) ((1+h)^n-h^n-g(h)) +h^n)\big) \cap [{\mathbb{P}}^n] \\
&=(1+h) g(h) \cap [{\mathbb{P}}^n]
\end{align*}
\noindent (The more combinatorially minded reader may reach the same conclusion
as a consequence of Theorem~\ref{CSM}.)
On the other hand, by Lemma~\ref{exseq} we have
\[
c(\Omega^1_{{\mathbb{P}}^n}(\log A')) = c(\Til {{\Omega'}^1}(-1))
\]
where ${\Omega'}^1$ is the module of differentials with logarithmic poles along ${\mathscr A}'$.
Under the assumption that ${\mathscr A}$ is free so is ${\mathscr A}'$, and $\Omega^1={\Omega'}^1
\oplus k$. Therefore $c(\Til{\Omega^1}(-1))=(1-h)c(\Til{{\Omega'}^1}(-1))$, and again
by Lemma~\ref{exseq} we get
$c(\Omega^1_{{\mathbb{P}}^n}(\log A))= (1-h) c(\Til {{\Omega'}^1}(-1))$, and hence
\[
c(\Omega^1_{{\mathbb{P}}^n}(\log A))^\vee= (1+h)\, c(\Til{{\Omega'}^1}(-1)^\vee)
=(1+h)\, g(h)\quad.
\]
It follows that
\[
{c_{\text{SM}}}(M({\mathscr A}))=(1+h) \,g(h)\cap [{\mathbb{P}}^n]=c(\Omega^1_{{\mathbb{P}}^n}(\log A))^\vee
\cap [{\mathbb{P}}^n]\quad,
\]
which is ($\dagger$) for ${\mathscr A}$, as claimed.
Therefore, we may assume that the arrangement is essential. By Theorem~\ref{MusSch},
\[
c(\Til{\Omega^1})=\pi_{\widehat{\cA}} (h)=(1+h)\, \underline{\pi_{\widehat{\cA}}}(h)
\]
in $A^*{\mathbb{P}}^n$ (that is, modulo $h^{n+1}$). Using Lemma~\ref{exseq}, it follows that
\[
\underline{\pi_{\widehat{\cA}}}(h) \equiv (1+h)^{-1} c(\Til{\Omega^1})\mod h^{n+1}
\equiv c(\Omega^1_{{\mathbb{P}}^n}(\log A)\otimes {\mathscr O}_{{\mathbb{P}}^n}(1)) \mod h^{n+1}\quad,
\]
and hence
\[
\underline{\pi_{\widehat{\cA}}}(h) = c(\Omega^1_{{\mathbb{P}}^n}(\log A)\otimes {\mathscr O}_{{\mathbb{P}}^n}(1))
\]
as polynomials in $h$, since both sides have degree $\le n$.
Now (as in \S\ref{genera}) we factor $\underline{\chi_{\widehat{\cA}}}(t)=(t-\alpha_1)\cdots (t-\alpha_n)$
over an extension\footnote{According to a theorem of Terao, the polynomial of a free
arrangement actually factors over ${\mathbb{Z}}$; cf.~\S\ref{posifree}. This is not needed here.}
of ${\mathbb{Q}}$,
and note that $\underline{\pi_{\widehat{\cA}}}(t) = (1+\alpha_1 t)\cdots (1+\alpha_n t)$.
With this notation, we have shown
\[
c(\Omega^1_{{\mathbb{P}}^n}(\log A)\otimes {\mathscr O}_{{\mathbb{P}}^n}(1))
=(1+\alpha_1 h)\cdots (1+\alpha_n h)\quad,
\]
and it follows that
\begin{align*}
c(\Omega^1_{{\mathbb{P}}^n}(\log A)^\vee) &=(1-\alpha_1 h+h)\cdots (1-\alpha_n h+h) \\
&=h^n\left(1+\frac 1h-\alpha_1\right) \cdots \left(1+\frac 1h-\alpha_n\right) \\
&=h^n \underline{\chi_{\widehat{\cA}}}\left(1+\frac 1h\right)
\end{align*}
with the usual caveat that the right-hand side must be interpreted as the polynomial
obtained by expanding it.
By Theorem~\ref{CSM} this proves
\[
c(\Omega^1_{{\mathbb{P}}^n}(\log A)^\vee) \cap [{\mathbb{P}}^n]
={c_{\text{SM}}}(M({\mathscr A}))\quad,
\]
and we are done.
\end{proof}
\begin{remark}
The projective version of Theorem~4.1 from~\cite{MR1843320} used above is also given
in~\cite{denhamschulze}, \S5, and generalized to locally tame arrangements.
\hfill$\lrcorner$
\end{remark}
\section{Segre classes of singularity subschemes}\label{segresec}
\subsection{}
In~\cite{MR2424913}, M.~Wakefield and M.~Yoshinaga prove that any (essential)
projective arrangement ${\mathscr A}$ may be reconstructed from the {\em singularity
subscheme\/} $S$ of the hypersurface $A\subseteq {\mathbb{P}}^n$, that is, the subscheme
defined by the partial derivatives of an equation for~$A$. In this section we prove
that the polynomial $\underline{\pi_{\widehat{\cA}}}(t)$ determines and is determined
by the degree of the arrangement and the Segre class (cf.~\cite{85k:14004},
Chapter~4) of the singularity subscheme in ${\mathbb{P}}^n$.
As we will use Chern--Schwartz--MacPherson classes for this result, we still work
over algebraically closed fields of characteristic~$0$. The following statement
makes sense over any field, but we do not know whether it holds in such
generality.
\begin{theorem}\label{segrethm}
Let $\iota:S\hookrightarrow {\mathbb{P}}^n$ be the singularity subscheme of an arrangement
${\mathscr A}$ in ${\mathbb{P}}^n$. (That is, $S$ is defined by the partial derivatives of an equation
for the hypersurface~$A$.) Let $\sigma_i\in {\mathbb{Z}}$ be such that
\[
[{\mathbb{P}}^n]-\iota_* s(S,{\mathbb{P}}^n) = \sum_{i=0}^n \sigma_i h^i \cap [{\mathbb{P}}^n] \in A_*{\mathbb{P}}^n\quad.
\]
Then
\[
\underline{\pi_{\widehat{\cA}}}(t) = \sum_{k=0}^n
\left(\sum_{i=0}^k \binom k i (d-1)^{k-i} \sigma_i\right) t^k\quad.
\]
\end{theorem}
\noindent Matching this formula with the expression for $\underline{\pi_{\widehat{\cA}}}(t)$ obtained
in the wake of Corollary~\ref{OSredux} yields the formula given in the introduction
for the ranks of the cohomology of the complement.
\begin{proof}
By Corollary~\ref{poinc},
\begin{equation*}
\tag{$\dagger$}
{c_{\text{SM}}}(A)=c(T{\mathbb{P}}^n)\cap [{\mathbb{P}}^n] - (1+h)^n
\underline{\pi_{\widehat{\cA}}}\left(\frac{-h}{1+h}\right) \cap [{\mathbb{P}}^n]\quad.
\end{equation*}
On the other hand, by Theorem~I.4 in \cite{MR2001i:14009},
\[
{c_{\text{SM}}}(A)=c(T{\mathbb{P}}^n)\cap \left(\frac{d h}{1+dh}\cap [{\mathbb{P}}^{n-1}]
+\frac 1{1+dh} \cap \big((\iota_* s(S,{\mathbb{P}}^n))^\vee \otimes_{{\mathbb{P}}^n} {\mathscr O}(dh)\big)\right)
\]
where $d$ is the degree of the arrangement, and this expression uses notation
given in~\cite{MR2001i:14009}, \S1.4. Writing $\iota_* s(S,{\mathbb{P}}^n)=\sum_{i=0}^n
s_i [{\mathbb{P}}^i]$, this means
\begin{equation*}
\tag{$\ddagger$}
{c_{\text{SM}}}(A)=(1+h)^{n+1}\left(\frac{d h}{1+dh}
+\frac 1{1+dh} \sum_{i=0}^n \frac{s_i\cdot (-h)^{n-i}}{(1+dh)^{n-i}}
\right)\cap [{\mathbb{P}}^n]
\end{equation*}
Comparing $(\dagger)$ and $(\ddagger)$ gives the following equality of series
modulo $h^{n+1}$:
\[
\underline{\pi_{\widehat{\cA}}}\left(\frac{-h}{1+h}\right)
=\frac {1+h}{1+dh} - \frac {1+h}{1+dh}
\sum_{i=0}^n \frac{s_i\cdot (-h)^{n-i}}{(1+dh)^{n-i}}\quad,
\]
and setting $t=-h/(1+h)$ yields
\[
\underline {\pi_{\widehat{\cA}}}(t) \equiv \frac 1{1-(d-1)\, t}\left(
1- \sum_{i=0}^n s_i\cdot \left(\frac t{1-(d-1)\, t}\right)^{n-i}\right)
\mod t^{n+1}\quad.
\]
Now with notation as in the statement we have $\sigma_0=1$ and $\sigma_i=
-s_{n-i}$ for $i>0$, and therefore
\[
\underline {\pi_{\widehat{\cA}}}(t) \equiv \frac 1{1-(d-1)\, t} \,
\sum_{i=0}^n \sigma_i\cdot \left(\frac t{1-(d-1)\, t}\right)^i
\mod t^{n+1}\quad.
\]
The statement follows immediately from this equality.
\end{proof}
Using notation as in \S1.4 of~\cite{MR2001i:14009}, the formula given in
Theorem~\ref{segrethm} may be rewritten as
\begin{equation*}
\tag{*}
\underline {\pi_{\widehat{\cA}}}(h)\cap [{\mathbb{P}}^n]
=\frac 1{1-(d-1) h}\cap \big(([{\mathbb{P}}^n]-\iota_*s(S,{\mathbb{P}}^n))\otimes_{{\mathbb{P}}^n} {\mathscr O}(-(d-1)h)\big)
\quad.
\end{equation*}
This is occasionally convenient in concrete computations, see~e.g., Example~\ref{book}.
\subsection{}
We illustrate Theorem~\ref{segrethm} with a few examples, in which the computation
of the Poincar\'e polynomial of the arrangement can also be performed easily with
standard techniques. Algorithms computing Segre classes may be implemented in
software systems such as Macaulay2 (\cite{M2}); one such implementation is
described in~\cite{MR1956868}. See Example~\ref{counter} for an illustration of
the use of such a routine.
\begin{example}
The three transversal intersections of the configuration of Example~\ref{fourlinesex}
count for one point each in the Segre class of the singularity subscheme. To
evaluate the contribution of the triple intersection, write it in local coordinates as
the singularity subscheme of $xy(x+y)=0$; the Jacobian ideal is then
\[
(2xy+y^2,x^2+2xy)
\]
and it follows that the contribution to the Segre class is $4$ points.
Thus $(\sigma_0,\sigma_1,\sigma_2)=(1,0,-7)$, and Theorem~\ref{segrethm}
gives
\[
\underline{\pi_{\widehat{\cA}}}(t)=1+3t+t^2\quad.
\]
Therefore $\pi_{\widehat{\cA}}(t)=(1+t)\underline{\pi_{\widehat{\cA}}}(t)=1+4t+5t^2+2t^3$, i.e.,
$\chi_{\widehat{\cA}}(t)=t^3-4t^2+5t-2$, as it should.
\hfill$\lrcorner$
\end{example}
\begin{example}
Let ${\mathscr A}$ consist of three planes in ${\mathbb{P}}^3$, with equation $xyz=0$.
\end{example}
\begin{wrapfigure}{l}{0.45\textwidth}
\begin{center}
\includegraphics[width=0.3\textwidth]{threeplanes}
\end{center}
\end{wrapfigure}
The singularity subscheme $S$ is supported on three lines; it is defined by the
ideal $(yz,xz,xy)$, so it is the intersection of three quadrics $Q_1,Q_2,Q_3$.
The Segre class $s(S,{\mathbb{P}}^3)$ is $3[{\mathbb{P}}^1]+m[{\mathbb{P}}^0]$ for some integer~$m$.
One way to evaluate $m$ is
the following: the intersection product of the three quadrics must be $8$
by B\'ezout's theorem, and can be evaluated by applying the `basic
construction' (Proposition~6.1 (a) in \cite{85k:14004}) to the fiber diagram
\[
\xymatrix{
S=Q_1\cap Q_2\cap Q_3 \ar[r] \ar[d] & {\mathbb{P}}^3 \ar[d] \\
Q_1\times Q_2\times Q_3 \ar[r] & {\mathbb{P}}^3\times {\mathbb{P}}^3\times {\mathbb{P}}^3
}
\]
Writing $h_i$ for the hyperplane class in the $i$-th copy of ${\mathbb{P}}^3$,
this gives
\[
8=\int (1+2h_1)(1+2h_2)(1+2h_3)\cap (3[{\mathbb{P}}^1]+m[{\mathbb{P}}^0]) = 18+m\quad,
\]
from which $m=-10$.
With notation as in Theorem~\ref{segrethm} we have $(\sigma_0,\dots,\sigma_3)
=(1,0,-3,10)$, from which
\[
\underline{\pi_{\widehat{\cA}}}(t)=1+2t+(4-3)t^2+(8-3\cdot 2 \cdot 3+10)t^3 = 1+2t+t^2=(1+t)^2\quad.
\]
Therefore $\pi_{\widehat{\cA}}(t)=(1+t)^3$. Of course this agrees with Example~\ref{norcro}, since
${\mathscr A}$ is a generic arrangement.
\hfill$\lrcorner$
\medskip
\begin{example}\label{book}
Let ${\mathscr A}$ consist of $d$ hyperplanes in the pencil of hyperplanes containing a fixed
codimension-$2$ subspace in ${\mathbb{P}}^n$.
\end{example}
\begin{wrapfigure}{l}{0.45\textwidth}
\begin{center}
\includegraphics[width=0.28\textwidth]{book}
\end{center}
\end{wrapfigure}
The singularity subscheme $S$ is supported on ${\mathbb{P}}^{n-2}$. To evaluate its
Segre class, blow-up along this subspace; if $E$ is the exceptional divisor,
the Segre class of the latter pushes forward to the Segre class of ${\mathbb{P}}^{n-2}$,
by the birational invariance of Segre classes:
$
\frac{E}{1+E} \mapsto s({\mathbb{P}}^{n-2},{\mathbb{P}}^n) = \frac 1{(1+h)^2}\cap [{\mathbb{P}}^{n-2}]
$.
A straightforward computation shows that the singularity subscheme pulls
back to $(d-1)$ times the exceptional divisor. Therefore (again by birational
invariance) $s(S,{\mathbb{P}}^n)$ is the push-forward of $(d-1)E/(1+(d-1)E)$, and
matching terms gives
\[
\iota_* s(S,{\mathbb{P}}^n) = \frac 1{(1+(d-1) h)^2}\cap (d-1)^2 [{\mathbb{P}}^{n-2}]\quad.
\]
Using (*), we get that
\begin{align*}
\underline{\pi_{\widehat{\cA}}}(h)\cap [{\mathbb{P}}^n] &=
\frac 1{1-(d-1)h}\cap \left([{\mathbb{P}}^n]-\frac{(d-1)[{\mathbb{P}}^{n-2}]}{(1+(d-1)h)^2}
\otimes {\mathscr O}(-(d-1)h)\right) \\
&=\frac 1{1-(d-1)h}\cap \left([{\mathbb{P}}^n]-(d-1)^2 [{\mathbb{P}}^{n-2}]\right) \\
&=(1+(d-1)h)\cap [{\mathbb{P}}^n]\quad.
\end{align*}
Therefore $\underline{\pi_{\widehat{\cA}}}(t)=1+(d-1)t$. It follows that
$\pi_{{\widehat{\cA}}}(t)=(1+(d-1)t)(1+t)=1+dt+(d-1)t^2$, and $\chi_{\widehat{\cA}}(t)=t^{n+1}-d t^n+(d-1)t^{n-1}$.
(This is of course also evident from the poset associated with this arrangement.)
\hfill$\lrcorner$
\medskip
\section{Positivity}\label{posit}
\subsection{}
One problem that prompted us to take a more careful look at hyperplane arrangements
is the issue of {\em positivity\/} of Chern--Schwartz--MacPherson classes. In the nonsingular
case, positivity of Chern classes is well understood; for example, $c(TX)\cap [X]$ is
effective if $TX$ is generated by global sections (cf.~\cite{85k:14004}, Example~12.1.7).
We know of no such statement for Chern classes of singular varieties, and preciously
few examples are known: the CSM class of a toric variety is represented by an effective
cycle (this follows from ``Ehlers' formula'', see e.g.,~\cite{MR1197235}), and CSM
classes of Schubert varieties are conjecturally effective. It is natural to ask the following
\begin{ques}
Denote by $A\subseteq {\mathbb{P}}^n$ the union of the hyperplanes of an arrangement ${\mathscr A}$.
For which arrangements ${\mathscr A}$ is ${c_{\text{SM}}}(A)$ effective?
\end{ques}
Here, by `effective' we mean that ${c_{\text{SM}}}(A)\in A_*{\mathbb{P}}^n$ should be represented by
an effective cycle. By Theorem~\ref{CSM}, ${c_{\text{SM}}}(A)$ is determined by the
characteristic polynomial of ${\widehat{\cA}}$, so this is a combinatorial question.
\subsection{}
Here is the explicit translation of effectivity in combinatorial terms:
\begin{prop}\label{effe}
Let ${\mathscr A}$ be a hyperplane arrangement in~${\mathbb{P}}^n$, and let $\mu$ be the
corresponding M\"obius function, as in \S\ref{genera}. Then ${c_{\text{SM}}}(A)$ is
effective if and only if all coefficients of the polynomial
\[
-\sum_{x\ne 0} \mu(x) (t+1)^{\dim x}
\]
are nonnegative.
\end{prop}
In fact, the coefficient of $t^k$ in this expression equals the coefficient of
$[{\mathbb{P}}^{k-1}]$ in ${c_{\text{SM}}}(A)$, for $k\ge 1$; the constant term equals~$1$.
\begin{proof}
In the (direct) proof of Theorem~\ref{CSM} we obtained the equality
\[
{c_{\text{SM}}}(M({\mathscr A})) = \sum_{x\in L({\widehat{\cA}})} \mu(x) {c_{\text{SM}}}(\underline x)\quad,
\]
where $\underline x$ denotes the projective subspace of ${\mathbb{P}}^n$
determined by $x$. The summand corresponding to $x=0$ is
$\mu(0) {c_{\text{SM}}}(\underline 0)=c(T{\mathbb{P}}^n)\cap [{\mathbb{P}}^n]$. Thus
\[
{c_{\text{SM}}}(A)=c(T{\mathbb{P}}^n)\cap [{\mathbb{P}}^n] - {c_{\text{SM}}}(M({\mathscr A}))
= -\sum_{x\ne 0} \mu(x) {c_{\text{SM}}}(\underline x)\quad.
\]
Now $\underline x\cong {\mathbb{P}}^{\dim x-1}$, so
\[
{c_{\text{SM}}}(\underline x) = \sum_{k=1}^{\dim x} \binom{\dim x}{k} [{\mathbb{P}}^{k-1}]\quad.
\]
Hence the coefficient of $[{\mathbb{P}}^{k-1}]$ in ${c_{\text{SM}}}(A)$ equals
\[
-\sum_{x\ne 0\,:\, \dim x\ge k} \mu(x)\binom{\dim x}k\quad,
\]
that is, the coefficient of $t^k$ in
\[
-\sum_{x\ne 0} \mu(x) (t+1)^{\dim x}\quad.
\]
This holds for $k\ge 1$. On the other hand, since $\sum_x \mu(x)=0$ and $\mu(0)=1$,
the constant term in this expression is $1>0$. It follows that ${c_{\text{SM}}}(A)$ is effective
if and only if all coefficients of this polynomial are nonnegative, which is the
statement.
\end{proof}
\begin{example}\label{fourlinesex}
We illustrate Proposition~\ref{effe} with a simple example.
\begin{center}
\includegraphics[scale=.5]{fourlines}
\end{center}
The poset and M\"obius function for the arrangement on the left (in ${\mathbb{P}}^2$)
are given on the right; the $0$ element of the lattice, i.e., $k^3$, is at the bottom.
The polynomial appearing in Proposition~\ref{effe} is
\[
-(-4 (t+1)^2 + 5(t+1) -2) = 4 t^2 +3t+1\quad.
\]
As the coefficients are all positive, the CSM class of this arrangement is effective;
this class equals $4[{\mathbb{P}}^1]+3 [{\mathbb{P}}^0]$.
\hfill$\lrcorner$
\end{example}
\subsection{}
Heuristically, arrangements of low degree should be more likely to have effective
CSM class. This is the case for generic arrangements:
\begin{prop}
Let ${\mathscr A}$ be a {\em generic\/} arrangement of $d\ge 1$ distinct hyperplanes in~${\mathbb{P}}^n$.
Then ${c_{\text{SM}}}(A)$ is effective for $n=1$ and all $d$, and for for $n>1$ and
\begin{itemize}
\item $n$ even, $d\le n+3$,
\item $n$ odd, $d\le n+4$,
\end{itemize}
and it is not effective otherwise.
\end{prop}
\begin{proof}
The arrangement ${\mathscr A}$ is generic precisely when $A$ is a divisor with simple normal
crossings. As seen in Example~\ref{norcro}, the CSM class of the complement is
${c_{\text{SM}}}(M({\mathscr A}))=\frac 1{(1+h)^d}\cap (c(T{\mathbb{P}}^n)\cap [{\mathbb{P}}^n])$, and hence
${c_{\text{SM}}}(A)$ equals{\small
\[
\left (1-\frac 1{(1+h)^d}\right) (1+h)^{n+1} \cap [{\mathbb{P}}^n]
=\sum_{k=0}^n \left(\binom{n+1}k - (-1)^k \binom{k+d-n-2}k\right) h^k\cap[{\mathbb{P}}^n]
\quad.
\]}
The statement is easy to verify from this expression. If $n$ is even
and $d\ge n+4$, the coefficient of $[{\mathbb{P}}^0]$ (i.e., the Euler characteristic of $A$) is
bound by
\[
(n+1)-\binom{n+2}2 < 0\quad;
\]
if $n>1$ is odd and $d\ge n+5$, the coefficient of $[{\mathbb{P}}^1]$ is bound by
\[
\binom{n+1}2 - \binom{n+2}3<0\quad,
\]
so the class is not effective outside of the specified range.\end{proof}
\begin{example}
The smallest generic arrangement with non-effective CSM class consists of
six lines in ${\mathbb{P}}^2$:
\begin{center}
\includegraphics[scale=.5]{sixlines}
\end{center}
The polynomial in the statement of Proposition~\ref{effe} is
\[
-(6\cdot (-1)\cdot (t+1)^2 + 15\cdot 1\cdot (t+1) -10)
=6 t^2 -3 t+1\quad,
\]
and not all its coefficients are nonnegative. The CSM class of this arrangement
is $6[{\mathbb{P}}^1]-3[{\mathbb{P}}^0]$; its Euler characteristic is $-3$.
\hfill$\lrcorner$
\end{example}
\subsection{}\label{posifree}
Another source of interesting examples comes from free arrangements.
\begin{prop}\label{freeposi}
For $n\le 8$, every free arrangement ${\mathscr A}$ of $d\le n$ hyperplanes in ${\mathbb{P}}^n$ has
effective CSM class.
\end{prop}
\begin{proof}
According to a theorem of Terao (\cite{MR608532}; see also \cite{MR1217488},
Chapter~4), the characteristic polynomial of a free central arrangement ${\widehat{\cA}}$
in $k^{n+1}$ factors over ${\mathbb{Z}}$:
\[
\chi_{{\widehat{\cA}}}(t)=(t-d_1)\cdots (t-d_n)\cdot (t-d_{n+1})\quad,
\]
where the $d_i$'s are the `exponents' of the arrangement, i.e., the
degrees of the generators of the (free) module of ${\widehat{\cA}}$-derivations
(Definitions~4.5 and~4.25 in~\cite{MR1217488}). One of the exponents
necessarily equals~$1$ (cf.~\S\ref{genera}); the sum of the exponents
equals the number of hyperplanes in the arrangement (\cite{MR1217488},
Proposition~4.26).
Thus, we may assume that the characteristic polynomial of the arrangement is
\[
(t-d_1)\cdots (t-d_n) (t-1)
\]
with $n\le 8$, $d_i\in {\mathbb{N}}$, $d_1+\cdots + d_8 \le n-1$. With the aid of a
computer, applying Proposition~\ref{effe} to all these cases is straightforward.
\end{proof}
Based on Proposition~\ref{freeposi} and the case of generic arrangements, one
may be tempted to guess that arrangements of $d\le n$ hyperplanes in ${\mathbb{P}}^n$
have effective CSM class. The following is the smallest counterexample
to this statement, for free arrangements.
\begin{example}\label{counter}
The polynomial
\[
\left|
\begin{matrix}
x_0 & x_0^3 & x_0^5 \\
x_1 & x_1^3 & x_1^5 \\
x_2 & x_2^3 & x_2^5
\end{matrix}
\right|
= x_0 x_1 x_2 (x_0-x_1)(x_0-x_2)(x_1-x_2)(x_0+x_1)(x_0+x_2)(x_1+x_2)
\]
defines a free arrangement of $9$ lines in ${\mathbb{P}}^2$, with exponents $1$, $3$, $5$.
The corresponding $9$ planes in $k^3$ meet along $13$ distinct lines; $6$ of these
lines lie on $2$ planes, $4$ on $3$, and $3$ on $4$ planes. It follows that the
M\"obius function takes values $1$ at $6$ lines, $2$ at $4$, and $3$ at $3$.
It also follows that the value of the M\"obius function for the affine arrangement
at the origin is $-15$.
\begin{center}
\includegraphics[scale=.5]{ninefree}
\end{center}
The cone over this projective arrangement in ${\mathbb{P}}^9$ is a free arrangement ${\mathscr A}$
of $9$ hyperplanes, with characteristic polynomial
\[
\chi_{{\widehat{\cA}}}(t)=t^{10}-9t^9+(6\cdot 1+4\cdot 2+3\cdot 3) t^8-15t^7= (t-5)(t-3)(t-1)t^7\quad.
\]
Using the criterion in Proposition~\ref{effe}, we compute
\begin{multline*}
-(-9 (t+1)^9 + 23(t+1)^8 -15 (t+1)^7) \\
= 9 t^9+58 t^8 + 155 t^7+217 t^6+161 t^5 +49 t^4 \underline{-7 t^3-5 t^2}+2 t+1
\quad,
\end{multline*}
verifying that the Chern--Schwartz--MacPherson class of this arrangement is not effective.
We end by remarking that the CSM routine described in \cite{MR1956868} and
implemented in Macaulay2 offers a quick verification of this computation: the
calculation of this CSM class from the equation of the arrangement takes
about $.1$ seconds on a laptop computer. The same routine may be used to
compute the Segre class of the singularity subscheme, giving with notation as in
Theorem~\ref{segrethm}
\[
(\sigma_0,\dots,\sigma_9)=(1,0,-49,664,-6528,54272,-389120,2260992,-7340032,
-58720256).
\]
Applying Theorem~\ref{segrethm} gives then
\[
\underline{\pi_{\widehat{\cA}}}(t)=1+8t+15t^2=(1+3t)(1+5t)\quad,
\]
in agreement with the combinatorial computation shown above.
\hfill$\lrcorner$
\end{example}
\newpage
|
\section{Introduction}
G and K giant stars are expected to have low lithium abundances because Li is destroyed in main-sequence stars except in the outermost layers (1$\%$ - 2$\%$ by mass). As the star ascends the red giant branch of the H-R diagram, the convective envelope deepens, diluting the existing Li and further reducing the observable Li \citep{1989ApJS...71..293B}. This effect is seen in most G and K giant stars, though a small minority ($\sim 1 \%$) exist that are unexpectedly rich in Li \citep[e.g.,][]{1982ApJ...255..577W,1990AJ.....99.1225P,1993ApJ...403..708F}.
Lithium abundance calculations are sensitive to temperature variations, so knowing the effective temperature ($T_{\rm eff}$) of a star is vital in estimating its Li content. HD~148293 was discovered to be Li rich by \citet{1989ApJS...71..293B}, who estimated its temperature using published photometry and color-temperature relations. Their value was also used by \citet[][hereafter CB00]{2000AandA...359..563C}, who placed HD 148293 on an H-R diagram and found it was at a phase known as the ``bump in the luminosity function''. This phase is characterized by an outwardly-moving hydrogen shell, leading to a short-lived phase of Li production before it is rapidly destroyed as evolution continues. Only low-mass stars that contain a highly degenerate helium core and later experience the helium flash pass through this stage and spend a mere 3$\%$ of their ascent on the red giant branch at the bump ($\sim$80,000 years, CB00).
By directly measuring the angular diameter of HD 148293, we are able to calculate its $T_{\rm eff}$ when combined with other observed quantities, such as interstellar absorption and bolometric corrections. We then modified the H-R diagram presented in CB00 to support their claim of proximity to the red-giant bump. Section 2 describes our observing procedure, Section 3 discusses how HD 148293's angular diameter and $T_{\rm eff}$ were determined, and Section 4 explores the physical implications of the new measurements.
\section{Interferometric observations}
Interferometric observations were obtained using the CHARA Array, a six element Y-shaped optical-infrared interferometer located on Mount Wilson, California \citep{2005ApJ...628..453T}. All observations used the pupil-plane ``CHARA Classic'' beam combiner in the $K'$-band at 2.14~$\mu$m while visible wavelengths (470-800 nm) were used for tracking and tip/tilt corrections. The observing procedure and data reduction process employed here are described in \citet{2005ApJ...628..439M}. We observed HD~148293 over two nights using two telescope pairs with different baseline lengths: 30 July 2010 using the E2-W2 pair with a baseline of approximately 156 m and 31 July 2010 using the W2-S2 pair at approximately 177 m.\footnote{The three arms of the CHARA Array are denoted by their cardinal directions: ``S'', ``E'', and ``W'' are south, east, and west, respectively. Each arm bears two telescopes, numbered ``1'' for the telescope farthest from the beam combining laboratory and ``2'' for the telescope closer to the lab. The ``baseline'' is the distance between the telescopes.}
Two calibrators (HD 145454 and HD 147321) were selected to be single single stars with expected visibility amplitudes $>$95$\%$ so they were nearly unresolved on the baselines used, which meant uncertainties in the calibrator's diameter did not affect the target's diameter calculation as much as if the calibrator star had a significant angular size on the sky. We interleaved calibrator and target star observations so that every target was flanked by calibrator observations made as close in time as possible, which allowed us to convert instrumental target and calibrator visibilities to calibrated visibilities for the target.
To check for possible unseen close companions that would contaminate our observations, we created spectral energy distribution (SED) fits based on published $UBVRIJHK$ photometric values obtained from the literature for each calibrator to establish diameter estimates. We combined the photometry with Kurucz model atmospheres\footnote{Available to download at http://kurucz.cfa.harvard.edu.} based on $T_{\rm eff}$ and log~$g$ values to calculate limb-darkened angular diameters for the calibrators. The stellar models were fit to observed photometry after converting magnitudes to fluxes using \citet[][$UBVRI$]{1996AJ....112..307C} and \citet[][$JHK$]{2003AJ....126.1090C}. The photometry, $T_{\rm eff}$ and log~$g$ values, and resulting limb-darkened angular diameters for the calibrators are listed in Table \ref{calibrators}. There were no hints of excess emission associated with a low-mass stellar companion or circumstellar disk in the calibrators' SED fits (see Figure \ref{seds}).
\begin{deluxetable}{lccl}
\tablewidth{0pc}
\tablecaption{Calibrator Information.\label{calibrators}}
\tablehead{ \colhead{Parameter} & \colhead{HD 145454} & \colhead{HD 147321} & \colhead{Source} }
\startdata
$U$ magnitude & 5.35 & 6.22 & \citet{Mermilliod} \\
$B$ magnitude & 5.42 & 6.07 & \citet{Mermilliod} \\
$V$ magnitude & 5.44 & 5.99 & \citet{Mermilliod} \\
$R$ magnitude & 5.46 & 5.99 & \citet{2003AJ....125..984M} \\
$I$ magnitude & 5.50 & 5.98 & \citet{2003AJ....125..984M} \\
$J$ magnitude & 5.37 & 5.79 & \citet{2003tmc..book.....C} \\
$H$ magnitude & 5.43 & 5.82 & \citet{2003tmc..book.....C} \\
$K$ magnitude & 5.43 & 5.77 & \citet{2003tmc..book.....C} \\
$T_{\rm eff}$ (K) & 9772 & & \citet{1999AandA...352..555A} \\
log $g$ (cm s$^{-2}$) & 4.13 & & \citet{1999AandA...352..555A} \\
$T_{\rm eff}$ (K) & & 8600 & \citet{2010yCat.2300....0L} \\
log $g$ (cm s$^{-2}$) & & 4.2 & \citet{2010yCat.2300....0L} \\
$\theta_{\rm LD}$ (mas) & 0.268$\; \pm \;$0.015 & 0.240$\; \pm \;$0.010 & \\
\enddata
\end{deluxetable}
\begin{figure}[h]
\includegraphics[width=0.5\textwidth]{f1.ps}
\caption{SED fits for the calibrator stars HD 145454 and HD 147321. The diamonds are fluxes derived from $UBVRI JHK$ photometry (left to right) and the solid lines are the Kurucz stellar models of the stars. See Table \ref{calibrators} for the values used to create the fits.}
\label{seds}
\end{figure}
\section{Determination of angular diameter and $T_{\rm eff}$}
The observed quantity of an interferometer is defined as the visibility ($V$), which is fit to a model of a uniformly-illuminated disk (UD) that represents the observed face of the star. Diameter fits to $V$ were based upon the UD approximation given by $V = 2 J_1(x) / x$, where $J_1$ is the first-order Bessel function and $x = \pi B \theta_{\rm UD} \lambda^{-1}$, where $B$ is the projected baseline at the star's position, $\theta_{\rm UD}$ is the apparent UD angular diameter of the star, and $\lambda$ is the effective wavelength of the observation \citep{1992ARAandA..30..457S}. A more realistic model of a star's disk involves limb-darkening (LD), and relationship incorporating the linear LD coefficient $\mu_{\lambda}$ \citep{1974MNRAS.167..475H} is:
\begin{equation}\begin{split}
V = \left( {1-\mu_\lambda \over 2} + {\mu_\lambda \over 3} \right)^{-2}
\times \\
\left[(1-\mu_\lambda) {J_1(\rm x) \over \rm x} + \mu_\lambda {\left( \frac{\pi}{2} \right)^{1/2} \frac{J_{3/2}(\rm x)}{\rm x^{3/2}}} \right] .
\end{split}\end{equation}
Table \ref{calib_visy} lists the Modified Julian Date (MJD), projected baseline ($B$) at the time of observation, projected baseline position angle ($\Theta$), calibrated visibility ($V$), and error in $V$ ($\sigma V$) for HD 148293.
\begin{deluxetable}{cccccc}
\tablewidth{0pc}
\tablecaption{HD 148293's Calibrated Visibilities.\label{calib_visy}}
\tablehead{\colhead{Calib} & \colhead{ } & \colhead{$B$} & \colhead{$\Theta$} & \colhead{ } & \colhead{ }
\\
\colhead{HD} & \colhead{MJD} & \colhead{(m)} & \colhead{(deg)} & \colhead{$V$} & \colhead{$\sigma V$} \\ }
\startdata
147321 & 55407.165 & 210.93 & 251.2 & 0.518 & 0.077 \\
& 55407.174 & 211.75 & 253.7 & 0.441 & 0.068 \\
& 55407.182 & 212.40 & 255.9 & 0.465 & 0.063 \\
& 55407.191 & 212.95 & 258.2 & 0.439 & 0.061 \\
& 55407.200 & 213.41 & 260.4 & 0.542 & 0.062 \\
& 55407.208 & 213.77 & 262.7 & 0.504 & 0.057 \\
& 55407.217 & 214.03 & 264.9 & 0.426 & 0.055 \\
145454 & 55408.275 & 156.09 & 243.2 & 0.759 & 0.072 \\
& 55408.283 & 156.15 & 246.0 & 0.768 & 0.061 \\
& 55408.291 & 156.18 & 248.7 & 0.752 & 0.035 \\
& 55408.304 & 156.22 & 252.8 & 0.750 & 0.057 \\
& 55408.312 & 156.24 & 255.7 & 0.765 & 0.080 \\
147321 & 55408.275 & 156.09 & 243.2 & 0.685 & 0.045 \\
& 55408.283 & 156.15 & 246.0 & 0.728 & 0.052 \\
& 55408.291 & 156.18 & 248.7 & 0.730 & 0.033 \\
& 55408.304 & 156.22 & 252.8 & 0.763 & 0.052 \\
& 55408.312 & 156.24 & 255.7 & 0.747 & 0.072 \\
\enddata
\tablecomments{The projected baseline position angle ($\Theta$) is calculated to be east of north.}
\end{deluxetable}
The LD coefficient was obtained from \citet{1995AandAS..114..247C} after adopting the $T_{\rm eff}$ and log~$g$ values for HD 148293 and the resulting LD angular diameter is listed in Table \ref{parameters}. The errors on the UD and LD diameters are 4$\%$ and the difference between the UD and LD diameters is on the order of a few percent, and the final angular diameter is little affected by the choice of $\mu_{\lambda}$\footnote{A 20$\%$ change in the $\mu_{\lambda}$ leads to a change in the measured LD diameter is less than $1\%$}. Additionally, the combination of the interferometric measurement of the star's angular diameter plus the \emph{Hipparcos} parallax \citep{2007hnrr.book.....V} allowed us to determine the star's physical radius. This result is also listed in Table \ref{parameters}.
\begin{deluxetable}{lcl}
\tablewidth{0pc}
\tablecaption{HD 148293 Stellar Parameters.\label{parameters}}
\tablehead{ \colhead{Parameter} & \colhead{Value} & \colhead{Reference} }
\startdata
[Fe/H] & +0.08 & \citet{1997AandAS..124..299C} \\
$V$ magnitude & 5.25 & \citet{Mermilliod} \\
$K$ magnitude & 2.83$\; \pm \;$0.11 & \citet{1969tmss.book.....N} \\
$A_{\rm V}$ & 0.04 & \citet{2005AandA...430..165F} \\
BC & 0.36$\; \pm \;$0.10 & \citet{1999AandAS..140..261A} \\
$F_{\rm BOL}$ (10$^{-8}$ erg s$^{-1}$ cm$^{-2}$) & 28.9$\; \pm \;$2.8 & Calculated here \\
$\theta_{\rm UD}$ (mas) & 1.439$\; \pm \;$0.059 (4$\%$) & Measured here \\
$\theta_{\rm LD}$ (mas) & 1.480$\; \pm \;$0.060 (4$\%$) & Measured here \\
$R_{\rm linear}$ ($R_\odot$) & 14.3$\; \pm \;$0.6 (4$\%$) & Measured here \\
$T_{\rm eff}$ (K) & 4640$\; \pm \;$100 & \citet{1989ApJS...71..293B} \\
$T_{\rm eff}$ (K) & 4460$\; \pm \;$141 (3$\%$) & Measured here \\
\enddata
\end{deluxetable}
For the $\theta_{\rm LD}$ fit, the errors were derived via the reduced $\chi^2$ minimization method \citep{2003psa..book.....W,1992nrca.book.....P}: the diameter fit with the lowest $\chi^2$ was found and the corresponding diameter was the final $\theta_{\rm LD}$ for the star. The errors were calculated by finding the diameter at $\chi^2 + 1$ on either side of the minimum $\chi^2$ and determining the difference between the $\chi^2$ diameter and $\chi^2 +1$ diameter. In calculating the diameter errors in Table \ref{parameters}, we adjusted the estimated visibility errors to force the reduced $\chi^2$ to unity because when this is omitted, the reduced $\chi^2$ is well under 1.0, indicating we are overestimating the errors in our calibrated visibilities. Figure \ref{hd148293} shows the LD diameter fit for HD 148293. Though several points lie outside the 1-$\sigma$ line of the diameter fit, the errors in the individual visibility points overlap the diameter fit itself.
\begin{figure}[h]
\includegraphics[width=0.35\textwidth, angle=90]{f2.ps}
\caption{HD 148293 LD disk diameter fit. The solid line represents the theoretical visibility curve for a star with the best fit $\theta_{\rm LD}$, the dotted lines are the 1$\sigma$ error limits of the diameter fit, the filled circles are the calibrated visibilities, and the vertical lines are the measured errors.}
\label{hd148293}
\end{figure}
Once $\theta_{\rm LD}$ was determined interferometrically, the $T_{\rm eff}$ was calculated using the relation
\begin{equation}
F_{\rm BOL} = {1 \over 4} \theta_{\rm LD}^2 \sigma T_{\rm eff}^4,
\end{equation}
where $F_{\rm BOL}$ is the bolometric flux and $\sigma$ is the Stefan-Bolzmann constant. $F_{\rm BOL}$ was determined in the following way: HD 148293's $V$ and $K$ magnitudes were dereddened using the extinction curve described in \citet{1989ApJ...345..245C} and its interstellar absorption ($A_{\rm V}$) value was from \citet{2005AandA...430..165F}. The intrinsic broad-band color ($V-K$) was calculated and the bolometric correction (BC) was determined by interpolating between the [Fe/H] = 0.0 and +0.2 tables from \citet{1999AandAS..140..261A}. They point out that in the range of 6000 K $\geq T_{\rm eff} \geq$ 4000 K, their BC calibration is symmetrically distributed around a $\pm$0.10 magnitude band when compared to other calibrations, so we assigned the BC an error of 0.10. The bolometric flux was then determined by applying the BC and the $T_{\rm eff}$ was calculated. See Table \ref{parameters} for a summary of these parameters.
\section{Results and discussion}
As a check to our measured diameter, limb-darkened angular diameters were estimated using two additional methods: (1) by producing an SED fit (see Figure \ref{sed}) as described in Section 2, where $UBV$ photometry is from \citet{Mermilliod}, $RI$ photometry is from \citet{2003AJ....125..984M}; and $JHK$ photometry is from \citet{2003tmc..book.....C}; and (2) using the relationship described in \citet{2004AandA...426..297K} between the ($V-K$) color and log $\theta_{\rm LD}$. Our measured $\theta_{\rm LD}$ is 1.480$\; \pm \;$0.060 mas, the SED fit estimates 1.418$\; \pm \;$0.085 mas, and the color-diameter relationship produces 1.465$\; \pm \;$0.692 mas. Because 2MASS measurements saturate at magnitudes brighter than $\sim$3.5 in the $K$-band even when using the shortest exposure time\footnote{Explanatory Supplement to the 2MASS All Sky Data Release and Extended Mission Products, http://www.ipac.caltech.edu/2mass/releases/allsky/doc/.}, we used the $K$ magnitude from the Two-Micron Sky Survey \citep{1969tmss.book.....N} for the color-diameter determination.
\begin{figure}[h]
\includegraphics[width=0.35\textwidth, angle=90]{f3.ps}
\caption{HD 148293 SED fit. The diamonds are fluxes derived from $UBVRI JHK$ photometry (left to right) and the solid line is the Kurucz stellar model of a star with $T_{\rm eff}$ = 4571 K and log $g$ = 2.41 from \citet{1999AandA...352..555A}. There were no quoted errors for the $UBVRI$ measurements, $\sigma_J \sim 7\%$, $\sigma_H \sim 7$, and $\sigma_K \sim 12\%$, which are not indicated on the plot.}
\label{sed}
\end{figure}
The main sources of errors for the three methods are uncertainties in visibility measurements for the interferometric measurement, uncertainties in the comparison between observed fluxes and the model fluxes for a given $T_{\rm eff}$ and log~$g$ for the SED estimate, and uncertainties in the $K$ magnitude for the color-diameter determination. All three diameters agree within their errors but our interferometric measurements provide an error approximately 40$\%$ and 12$\times$ smaller than the latter two methods, respectively.
We also wanted to compare our $T_{\rm eff}$ with those obtained from the literature (see Table \ref{temps}). Values range from 4390 K \citep{2010yCat.2300....0L} to 4693$\pm$24 K \citep{2011AandA...525A..71W}. The methods are varied and are based on spectroscopic observations, photometric measurements, spectral typing, color-$T_{\rm eff}$ relationships, the correlation between absorption line features and stellar parameters, and the star's position on the color-magnitude diagram. Most temperatures do not have errors indicated in the source papers. Our measurement of 4460$\pm$141 K is on the cooler side of the values from the literature, though the entire range spans only $\sim$300 K.
\begin{deluxetable}{lll}
\tablewidth{0pc}
\tablecaption{HD 148293 $T_{\rm eff}$ from the Literature.\label{temps}}
\tablehead{ \colhead{$T_{\rm eff}$ (K)} & \colhead{Reference} & \colhead{Method Used} }
\startdata
4390 & \citet{2010yCat.2300....0L} & Based on spectral type \\
4420 & \citet{2008AstL...34..785G} & Photometric measurements \\
4420 & \citet{2003AJ....125..359W} & Based on spectral type \\
4560 & \citet{1999AandAS..140..261A} & ($B-V$) - $T_{\rm eff}$ relationship \\
4571$\pm$12 & \citet{1999AandA...352..555A} & Position on color-magnitude diagram \\
4585$^{+287}_{-189}$ & \citet{2006ApJ...638.1004A} & Photometric measurements \\
4640$\pm$100 & \citet{1989ApJS...71..293B} & Photometry and color-$T_{\rm eff}$ relations \\
4640 & \citet{2007ApJS..171..146S} & Relation between absorption line features \\
& & and stellar parameters \\
4650 & \citet{1990ApJS...74.1075M} & Color - $T_{\rm eff}$ relationship \\
4650 & \citet{1997AandAS..124..299C, 2001AandA...373..159C} & Spectroscopic measurements \\
4693$\pm$24 & \citet{2011AandA...525A..71W} & Spectroscopic measurements \\
4460$\pm$141 & Measured here & Interferometric measurement \\
\enddata
\end{deluxetable}
With our newly calculated $T_{\rm eff}$, we are able to place HD 148293 on the H-R diagram from CB00. Figure \ref{hr} shows a reproduction of their Figure 1 with Li-rich giant stars plotted along with evolutionary tracks for a range of stellar masses. We used the YZVAR stellar model \citep{2008AandA...484..815B, 2009AandA...508..355B} for HD 148293's metallicity of +0.08 \citep{1997AandAS..124..299C}.
\begin{figure}[h]
\includegraphics[width=0.35\textwidth, angle=90]{f4.ps}
\caption{H-R diagram for the Li-rich giants. The small circles are the stars from CB00, the large circle is their original placement of HD 148293, and the large triangle is our new placement. The evolutionary tracks are marked by the stars' masses, and the shaded region indicates the red giant bump region. The luminosity and temperature error bars for the Charbonnel \& Balachandran points are from uncertainties in the Hipparcos parallaxes and from the literature, respectively. Our error bars are based on interferometric observations, uncertainties in Hipparcos parallaxes, and bolometric corrections.}
\label{hr}
\end{figure}
CB00 hypothesized that HD 148293 is at the red-giant bump, even though the $T_{\rm eff}$ they used from \citet{1989ApJS...71..293B} places the star slightly to the left of the bump. They claim the temperature shift required to place HD 148293 in the red-giant bump is not unreasonable ($\sim$200 K), given the uncertainties in the $T_{\rm eff}$ measurement. Our $T_{\rm eff}$ is 180 K cooler than the value they used and places HD 148293 closer to and within the error bar of the red-giant bump, supporting their hypothesis. CB00 do not seem to believe that the slight shift in $T_{\rm eff}$ significantly affects the Li abundance calculated by Brown et al., so it would appear likely that we are seeing HD 148293 during a very brief stage in its evolutionary process where Li is being produced or was produced very recently.
\acknowledgements
The CHARA Array is funded by the National Science Foundation through NSF grant AST-0606958 and by Georgia State University through the College of Arts and Sciences, and by the W.M. Keck Foundation. STR acknowledges partial support by NASA grant NNH09AK731. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France. This publication makes use of data products from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation.
|
\section{Introduction}\label{intro}
Fractional derivative constitutive equations (CEs) have been found to accurately predict, among others, the stress relaxation of viscoelastic fluids in the glass transition and glassy (high frequency) states. The experimental behavior of storage $G'$ and loss $G''$ moduli (obtained upon using the time - temperature superposition principle - see \cite{pal2}, \cite{pal3}) of a linear, narrow molecular weight series of polybutadienes is exceptionally well predicted by linearized fractional derivative models as can be reckoned from \cite{pal1}. Polybutadienes are of utter importance for the tire industry, for manufacturing certain solid propergols, etc. Similar excellent agreements between frequency sweep experimental data obtained on other polymers (e.g. polystyrenes) and theoretical predictions of linear fractional derivative models are reported in \cite{fr,hey1, main}.
The object of study is the below given objective, fractional derivative viscoelastic (incompressible) fluid constitutive equation (CE) (see \cite{pal4})
\begin{eqnarray}\label{ce}
\lefteqn{ \mathbf{S}(t)+\lambda^{\alpha}\mathbf{F}(t)\bigg\{\int_{-\infty}^{t}\mu_{1}(t-\tau)
\mathbf{F}^{-1}(\tau)\stackrel{\bigtriangledown}{\mathbf{S}}(\tau)\left[\mathbf{F}^{-1}(\tau)\right]^{T}
\mathrm{d} \tau \bigg\}
\mathbf{F}(t)^{T} } \nonumber\\
& & =G\lambda^{\beta}\mathbf{F}(t)\bigg\{\int_{-\infty}^{t}\mu_{2}(t-\tau)
\mathbf{F}^{-1}(\tau)\mathbf{A}_{1}(\tau)\left[\mathbf{F}^{-1}(\tau)\right]^{T}\mathrm{d} \tau\bigg\}
\mathbf{F}(t)^{T}
\end{eqnarray}
Function $\mathbf{S}$ is the (objective) stress tensor and $\stackrel{\bigtriangledown}{\mathbf{S}}$ its objective upper convected derivative defined by (with $D/Dt$ denoting the material derivative and $\mathbf{L}$ the velocity gradient; see for example \cite{hp2},\cite{mor}, \cite{tig}):
\begin{equation} \label{ody}
\stackrel{\bigtriangledown}{\mathbf{S}}=\frac{D {\bf S}}{D
t}-\mathbf{L}\mathbf{S}-\mathbf{S}\mathbf{L}^{T},
\end{equation}
Function $\mathbf{F}$ is the strain gradient and $\mathbf{A}_{1}=\nabla\mathbf{u}+(\nabla\mathbf{u})^T={\bf L}+{\bf L}^T$ is the first Rivlin-Ericksen tensor. The model parameters are such that $0<\lambda$, $0<\alpha<\beta<1$. $\mu_{1,2}(t)$ are two memory kernels given by:
\begin{equation}\label{mu}
\mu_{1}(t-\tau)=\frac{(t-\tau)^{-\alpha}}{\Gamma(1-\alpha)},\,\,
\mu_{2}(t-\tau)=\frac{(t-\tau)^{-\beta}}{\Gamma(1-\beta)}
\end{equation}
The stability of the rest state is now investigated using the linearized theory. As shown in \cite{pal4} and \cite{pal7}, it is first assumed that the stress tensor $\mathbf{S}=\mathit{O}(\epsilon)$ and the deformation
gradient $\mathbf{F}(t)=\mathbf{1}+\epsilon\mathbf{J}(t)+\mathit{O}(\epsilon^{2})$. Since $\mathbf{L}=\dot{\mathbf{F}}\mathbf{F}^{-1}$ (see for ex. \cite{hp2}, \cite{ddj}, \cite{mor}), $\mathbf{L}=\mathit{O}(\epsilon)$. Hence the velocity $\mathbf{u}=\mathit{O}(\epsilon)$, and the first Rivlin-Ericksen tensor $\mathbf{A}_{1}=\mathit{O}(\epsilon)$ as well. Therefore, keeping only terms
of $\mathit{O}(\epsilon)$, within the linear response theory eq.\eqref{ce} reduces to :
\begin{equation}\label{law2}
\mathbf{S}(t)+\lambda^{\alpha}\int_{-\infty}^{t}\mu_{1}(t-\tau)\frac{\partial \mathbf{S}(\tau)}{\partial
\tau}\mathrm{d}\tau=G\lambda^{\beta}\int_{-\infty}^{t}\mu_{2}(t-\tau)\mathbf{A}_{1}(\tau)\mathrm{d} \tau
\end{equation}
Next, assume the fluid is contained in a bounded volume $\Omega\subset\mathbb{R}^3$ whose boundary $\partial\Omega$ is sufficiently smooth, and set in motion at $t=0$. The CE in eq.\eqref{law2} then takes the form:
\begin{equation}\label{law3}
\mathbf{S}(t)+\lambda^{\alpha}\int_{0}^{t}\mu_{1}(t-\tau)\frac{\partial \mathbf{S}(\tau)}{\partial
\tau}\mathrm{d}\tau=G\lambda^{\beta}\int_{0}^{t}\mu_{2}(t-\tau)\mathbf{A}_{1}(\tau)\mathrm{d} \tau
\end{equation}
The above may be re-written in condensed form using the Caputo operators $D^{\alpha}_{t}$ and $I^{1-\beta}_{t}$ as:
\begin{equation}\label{law4}
\mathbf{S}(t)+\lambda^{\alpha}\, D^{\alpha}_{t}\mathbf{S}=G\lambda^{\beta}\, I^{1-\beta}_{t}\mathbf{A}_{1}
\end{equation}
where for an absolutely continuous function $f:\mathbb{R}_+\to \mathbb{C}$:
\begin{equation}\label{cfd}
D^{\alpha}_{t}f(t)=\frac{1}{\Gamma(1-\alpha)}
\int^{t}_{0}\frac{f'(\tau)}{(t-\tau)^\alpha} \mathrm{d}\tau
\end{equation}
and for $f\in L^1_{\text{loc}}(\mathbb{R}_+)$,
\begin{equation}\label{cfi}
I^{1-\beta}_{t}f(t)=\frac{1}{\Gamma(1-\beta)}\int^{t}_{0}
\frac{f(\tau)}{(t-\tau)^\beta} \mathrm{d}\tau
\end{equation}
As shown in \cite{pal4,pal7}, investigating the stability of the rest state is tantamount to studying the existence and the uniqueness of solutions to the following initial boundary value problem (IBVP):
\begin{subeqnarray}\label{bvp1}
& & \frac{\partial\mathbf{u}}{\partial t}=-\nabla p+\nabla\cdot\mathbf{S} \label{bvp1 1}\\
& & \mathbf{S}+\lambda^\alpha D^{\alpha}_{t}\mathbf{S}=G\lambda^\beta I^{1-\beta}_{t}\mathbf{A}_{1},\quad \mathbf{A}_{1}=\nabla\mathbf{u}+(\nabla\mathbf{u})^T \label{bvp1 2}\\
& & \mathbf{\nabla}\cdot\mathbf{u}=0,\quad\text{in}\quad \lbrack 0,+\infty \lbrack \times
\Omega,\, \Omega\subset\mathbb{R}^3 \label{bvp1 3}\\
& & \mathbf{u}=\mathbf{0},\quad\text{in} \quad \lbrack 0,
+\infty \lbrack \times \partial\Omega \label{bvp1 4}\\
& & \mathbf{u}(t=0)=\mathbf{u}_0,\quad \mathbf{S}(t=0)=\mathbf{S}_0 \label{bvp1 5} \\
\end{subeqnarray}
In the above system of equations we assume ${\bf u}: [0,+\infty[ \times \Omega \to\mathbb{R}^3$, $\nabla\cdot\mathbf{u}_0=0$,
$p:[0,+\infty[ \times \Omega \to\mathbb{R}$, $\mathbf{S}: [0,+\infty[ \times \Omega \to \mathscr{M}_{3,3}(\mathbb{R})$, $0<\alpha<\beta<1$. Denote $\delta=\beta-\alpha>0$.
A change of variables on $({\bf x},t)$ can be performed to eliminate the CE parameters $\lambda$ and $G$ (see \cite{pal7}). This is carried out only for convenience; in no way the generality of this paper results is shrinked down. Therefore, from now on assume $\lambda=G=1$.
At this stage recall that an existence result for the initial boundary value problem given in equations \eqref{bvp1}was presented in \cite{pal7}. The present paper, which is a continuation of \cite{pal7}, is organized as follows:
\begin{itemize}
\item Section \ref{wfbvp} presents the weak formulation of the boundary value problem.
\item Section \ref{wf} is devoted to proving the existence and uniqueness of the solutions. We further on use the existence theorem obtained in \cite{pal7} to state a general existence and uniqueness result.
\item Section \ref{so} deals with the functional framework within which the solution continuity at $t=0$ is proved.
\item Section \ref{n} presents the proof of the solution continuity at $t=0$.
\item Section \ref{ie} contains results on the solution smoothness.
\end{itemize}
\section{Weak formulation of the IBVP}\label{wfbvp}
All time-depending functions involved in the current stability analysis, save for when stated otherwise, are causal functions (i.e. set equal to zero on $\mathbb{R}_-$). Hence the convolution in time is simply $(f\ast g)(t):=\displaystyle \int_0^t f(s)g(t-s)\mathrm{d} s$.
We first present the weak formulation of the boundary value problem eqs.\eqref{bvp1}: find ${\bf u}\in\mathscr{C}^0([0,+\infty[,\,L^2(\Omega)^3) $ $\cap L^1_{\text{loc}}(\mathbb{R}_+,H^1_0(\Omega)^3)$, $\nabla\cdot{\bf u}=0$, $[{\bf S}]_{ij}\in\mathscr{C}^0 \left( ]0,+\infty[,\,L^2(\Omega)\cap L^1_{\text{loc}}\left(\mathbb{R}_+ ,L^2(\Omega) \right) \right)$, $i,j=1,2,3$, such that for any test-functions $\forall\boldsymbol{ \theta}\in\left(H^1_0(\Omega)\right)^3,\, \nabla\cdot\boldsymbol{ \theta}=0,\,\forall {\bf a}\in\left(\mathscr{D}(\Omega)\right)^3,\,\forall
\psi\in\mathscr{C}^{\infty}_{00}([0,+\infty[)$, where $\mathscr{C}^{\infty}_{00}([0,+\infty[) $ denotes the space of $\mathscr{C}^{\infty} $ class functions that vanish in a neighborhood of $+\infty$, the following equations hold true:
\begin{equation}\label{fv1}
\psi(0)\int_{\Omega}{\bf u}_0({\bf x})\cdot\boldsymbol{\theta}({\bf x}) \mathrm{d} {\bf x}+\int^{+\infty}_{0}\int_{\Omega}{\bf u}(t,{\bf x})\cdot\boldsymbol{\theta}({\bf x})\psi'(t) \mathrm{d} {\bf x}\mathrm{d} t =\int^{+\infty}_{0}\int_{\Omega}({\bf S}(t,{\bf x})\colon\nabla\boldsymbol{\theta}({\bf x}))\psi(t) \mathrm{d} {\bf x} \mathrm{d} t
\end{equation}
\begin{eqnarray}\label{fv2}
& & -\int^{+\infty}_{0}\int_{\Omega}\frac{\psi(\tau)}{\Gamma(1-\alpha)\tau^\alpha}[{\bf S}_0]_{ij}({\bf x})[{\bf a}]_j({\bf x})\mathrm{d} {\bf x} \mathrm{d} \tau \nonumber\\[0.5cm]
& & -\int^{+\infty}_{0}\int^{+\infty}_{\tau}\int_{\Omega}\frac{\psi'(t)}{\Gamma(1-\alpha)\tau^\alpha}[{\bf S}]_{ij}(t-\tau,{\bf x})[{\bf a}]_j({\bf x})\mathrm{d} {\bf x}\mathrm{d} t\mathrm{d}\tau \nonumber\\[0.5cm]
& & +\int^{+\infty}_{0}\int_{\Omega}[{\bf S}]_{ij}(t,{\bf x})[{\bf a}]_j({\bf x})\psi(t)\mathrm{d} {\bf x} \mathrm{d} t= \nonumber\\[0.5cm]
& & -\frac{1}{\Gamma(1-\beta)}\int^{+\infty}_{0}\int^{t}_{0}\int_{\Omega}\frac{\psi(t)}{(t-\tau)^\beta}\left\{\left(\nabla\cdot{\bf a}\right)[{\bf u}]_i+[{\bf u}\cdot\nabla{\bf a}]_i\right\}(\tau,{\bf x})\mathrm{d} {\bf x}\mathrm{d}\tau \mathrm{d} t
\end{eqnarray}
Summation over repeated indices is understood in equations \eqref{fv1} and \eqref{fv2} above.
We now detail the functional framework. Let $V=\{{\bf h}\in \displaystyle\left(H^1_0(\Omega)\right)^3\, \text{s.t.}\, \nabla\cdot {\bf h}=0\}$ be the Hilbert space endowed with the inner product:
\begin{equation}\label{hils}
\left\langle {\bf f}|{\bf g} \right\rangle_V=\displaystyle\mathop{\sum}_{i,j=1}^{3}\int_{\Omega}\frac{\partial f_i}{\partial x_j}\frac{\overline{\partial g_i}}{\partial x_j}({\bf x})\mathrm{d} {\bf x}
\end{equation}
and denote $\|\,\|_{V}$ the corresponding norm.
The closure of $V$ in $(L^2(\Omega))^3 $ is denoted by $H$, the later space being endowed with the inner product:
\begin{equation}\label{hsd}
\left\langle {\bf f}|{\bf g} \right\rangle_H=\displaystyle\mathop{\sum}_{i=1}^{3}\int_{\Omega}f_i\overline{g}_i({\bf x})\mathrm{d} {\bf x}
\end{equation}
with $\|\,\|_{H}$ being the corresponding norm. Let $0<\lambda_1\leq\lambda_2\leq\dots\lambda_n\leq\dots \displaystyle \mathop{\longrightarrow}_{n\rightarrow +\infty} +\infty $ and ${\bf w}_i\in V,\,i\in\mathbb{N}^\ast$, be the eigenvalues and the corresponding eigenfunctions of the Stokes operator in $H$, i.e.:
\begin{equation}\label{eig1}
\forall \boldsymbol{\phi}\in V,\, \langle {\bf w}_k\,|\, \boldsymbol{\phi}\rangle_V=\lambda_k\langle {\bf w}_k\,|\, \boldsymbol{\phi}\rangle_H,\quad\text{where}\, \|{\bf w}_k\|_H=1
\end{equation}
\section{The solution existence and uniqueness}\label{wf}
To prove the solution uniqueness, we first eliminate ${\bf S}$ from equations \eqref{fv1} and \eqref{fv2}.
Denote $\mathscr{L}:=\{f\in L^1_{\text{loc}}(\mathbb{R}_+)\,\text{s.t.}\,\exists M>0,\,\text{so that}\, fe^{-Mt}\in L^1(\mathbb{R}_+) \} \subset L^1_{\text{loc}}(\mathbb{R}_+) $. Next, let $f\in L^1_{\text{loc},\mathbb{R}_+}(\mathbb{R})$. For any $a\in \mathbb{R}$ and $\alpha\in ]0,1[$, define $D^\alpha _{t,a}f$ by:
\begin{equation}\label{u1}
\left\langle D^\alpha _{t,a}f,\varphi \right\rangle =\dfrac{1}{\Gamma(1-\alpha)}\left[ -a\int_0^{+\infty}\dfrac{\varphi(\tau)}{\tau^\alpha}\mathrm{d} \tau-\int_0^{+\infty}f(t)\left(\int_0^{+\infty}\dfrac{\varphi'(t+\tau)}{\tau^\alpha}\mathrm{d} \tau \right)\mathrm{d} t \right]
\end{equation}
for any test function $ \varphi\in \mathscr{D}(\mathbb{R})$.
Observe that $D^\alpha _{t,a}f\in\mathscr{D}'(\mathbb{R})$. Moreover, for any $ f\in \mathscr{L}$, one easily sees that $ x\geq M$, $e^{-xt}D^\alpha _{t,a}f\in\mathscr{S}'(\mathbb{R})$.
The hat $\widehat{(\,)}$ notation to be used below stands for the usual Laplace transform.
\begin{proposition}\label{up1}
Let $f\in \mathscr{L}$. Then, for $\forall s\in \mathbb{C}$, with $\mathrm{Re}(s)$ large enough, one has $\widehat{D^\alpha _{t,0}f}(s)=s^\alpha\hat{f}(s)$.
\end{proposition}
\begin{proof}
Let $M>0$ such that $e^{-Mt} f\in L^1(\mathbb{R}_+)$. For any $s\in\mathbb{C},\,\mathrm{Re}(s)\geq M$ and $\varphi(t)=e^{-st}$, Eq.\eqref{u1} - still valid for this particular choice of $\varphi(t)=e^{-st}\notin \mathscr{D}(\mathbb{R})$ - gives:
\begin{equation}\label{up1 1}
\left\langle D^\alpha _{t,0}f,e^{-st} \right\rangle =-\dfrac{1}{\Gamma(1-\alpha)} \int_0^{+\infty}f(t)\left(\int_0^{+\infty}\dfrac{-se^{s+\tau}}{\tau^\alpha}\mathrm{d} \tau \right)\mathrm{d} t
\end{equation}
As $\displaystyle\int_0^{+\infty}\dfrac{e^{-s\tau}}{\tau^\alpha}\mathrm{d} \tau=\dfrac{\Gamma(1-\alpha)}{s^{1-\alpha}}$, for $\mathrm{Re}(s)\geq M>0$ one gets:
\begin{equation}\label{up1 2}
\widehat{D^\alpha _{t,0}f}(s)=\left\langle D^\alpha _{t,0}f,e^{-st} \right\rangle =-\dfrac{s}{\Gamma(1-\alpha)} \int_0^{+\infty}e^{-st}f(t)\Gamma(1-\alpha)s^{\alpha-1}\mathrm{d} t=s^\alpha\hat{f}(s)
\end{equation}
\end{proof}
The following classical result (see \cite{cm}) is stated here within our functional framework. Recall first that (see also equations 14 and 15 in \cite{pal7}):
\begin{equation}\label{wf1}
W_0(t)=\dfrac{\sin(\alpha\pi)}{\pi} \displaystyle\int_0^{+\infty}\dfrac{e^{-rt}r^{\alpha-1}}{r^{2\alpha}+2r^{\alpha}\cos(\alpha\pi)+1}\mathrm{d} r,\, t\geq 0
\end{equation}
\begin{equation}\label{wf2}
E_\alpha(t)=\dfrac{\sin(\alpha\pi)}{\pi}\displaystyle\int_0^{\infty}\dfrac{r^\alpha e^{-rt}}{r^{2\alpha}+2r^\alpha \cos(\alpha\pi)+1}\mathrm{d} r, \, t> 0
\end{equation}
\begin{proposition}\label{up2}
Let $F\in \mathscr{L}$. Then, for any $ a\in \mathbb{R}$ and any $ \alpha\in ]0,1[$, the equation
\begin{equation}\label{up2 1}
D^\alpha_{t,a}f+f=F
\end{equation}
has a unique solution $f\in \mathscr{L}$, given by $f=E_\alpha\displaystyle \ast F+aW_0$.
\end{proposition}
\begin{proof}
\textit{Existence}:
Assume $F\in \mathscr{C}^1(\mathbb{R}_+)$. Then, $f=E_\alpha\displaystyle \ast F+aW_0$ is a solution of Eq.\eqref{up2 1} (cf \cite{cm}). Now, if one assumes that $F\in \mathscr{L}$, then there exists $ (F_n)_{n\in \mathbb{N}^\ast}$, $F_n\in \mathscr{C}^1(\mathbb{R}_+)$ such that $F_n \displaystyle\mathop{\longrightarrow}_{n\to+\infty} F $ in $L^1_{\text{loc}}(\mathbb{R}_+)$. Since $E_\alpha\in L^1_{\text{loc}}(\mathbb{R}_+)$, then $f_n=E_\alpha\displaystyle \ast F_n+aW_0 \displaystyle\mathop{\longrightarrow}_{n\to+\infty} E_\alpha\displaystyle\mathop{\ast}_{(t)}F+aW_0 $ in $L^1_{\text{loc}}(\mathbb{R}_+)$. Hence the equation $D^\alpha_{t,a}f_n+f_n=F_n$, for $n\to+\infty$, becomes $D^\alpha_{t,a}f+f=F$.
We must next show that $f\in \mathscr{L}$. Notice first that $\|aW_0\|_{\infty}\leq |a|W_0(t)$. Therefore $aW_0\in\mathscr{L}$. Denote $g=e^{-Mt}F$; choose $M>0$ so that $e^{-Mt}F\in L^1(\mathbb{R})$. Then $\left|E_\alpha\displaystyle\mathop{\ast}_{(t)}F \right|=\left|\displaystyle \int _0^t E_\alpha(t-s)e^{Ms}g(s)\mathrm{d} s \right|\leq e^{Mt} \left[ E_\alpha\displaystyle\mathop{\ast}_{(t)}g \right]$. Since $g\in L^1(\mathbb{R}_+)$, and $E_\alpha\in L^1(\mathbb{R}_+)$, then $E_\alpha\displaystyle \ast g\in L^1(\mathbb{R}_+)$. Finally $E_\alpha\displaystyle \ast F\in \mathscr{L}$ and $f=E_\alpha\displaystyle \ast F+aW_0\in \mathscr{L}$.
\textit{Uniqueness}:
Let $f,g\in \mathscr{L}$ be two solutions of Eq.\eqref{up2 1}. Then $D_{t,0}^\alpha (f-g)+(f-g)=0$, from which it follows that $(s^\alpha+1)\widehat{f-g}(s)=0$, for $Re(s)$ large enough. Therefore $\widehat{f-g}(s)=0$, thus $f=g$.
\end{proof}
We shall use the following result to prove the uniqueness property:
\begin{lemma}\label{ul1}
For any $g\in\mathscr{L}$, $I_t^{1-\beta}g\in\mathscr{L}$.
\end{lemma}
\begin{proof}
Since $g\in\mathscr{L}$, there exists $G\in L^1(\mathbb{R}_+)$ and $M>0$ such that $g=Ge^{-Mt}$. Therefore, for $t\geq0$ a.e., $\left|I_t^{1-\beta}g(t) \right|\leq K \left|\displaystyle\int_0^t \dfrac{g(t-u)}{u^\beta}\mathrm{d} u \right| \leq K \displaystyle\int_0^t |G(t-u)|\dfrac{e^{-M(t-u)}}{u^\beta}\mathrm{d} u \leq K e^{-Mt}\left(|G|\displaystyle \ast u_\beta e^{-Mt} \right) $. Now, $G \in L^1(\mathbb{R}_+)$, $u_\beta e^{-Mt}\in L^1(\mathbb{R}_+)$ leads to $\left(|G|\displaystyle \ast u_\beta e^{-Mt} \right)\in L^1(\mathbb{R}_+)$. Therefore, $\left|I_t^{1-\beta}g(t) \right|\leq e^{Mt}H(t)$, with $H\in L^1(\mathbb{R}_+)$, which gives $I_t^{1-\beta}g\in\mathscr{L}$.
\end{proof}
Making use of Proposition \ref{up2} and of Lemma \ref{ul1}, we get:
\begin{corollary}[\textbf{Solution uniqueness}]\label{uc1}
Let ${\bf u}_0\in H$ and ${\bf S}_0\in L^2(\Omega)^9$. The system of equations Eqs.\eqref{fv1}-\eqref{fv2} has at most one solution that belongs to the functional space $ \mathscr{F}:=\{ ({\bf u},{\bf S}) \in \left[ \mathscr{C}^0 (\mathbb{R}_+,H)\cap L^1_{\text{loc}}(\mathbb{R}_+,V) \right] \times \mathscr{C}^0\left(]0,+\infty[,L^2(\Omega)^9 \right)$, such that $\|\nabla{\bf u}\|_{L^2(\Omega)^9}\in \mathscr{L},\,\| {\bf S}\|_{L^2(\Omega)^9}\in \mathscr{L} \} $.
\end{corollary}
\begin{proof}
Let $({\bf u},{\bf S})\in \mathscr{F}$ be a solution to Eqs.\eqref{fv1}-\eqref{fv2}. For any test function $\boldsymbol{\varphi}\in \mathscr{D}(\Omega)^9$, as a consequence of Eq.\eqref{fv2} and of the fact that ${\bf u}\in L^1_{\text{loc}}(\mathbb{R}_+,V)$, one has
\begin{equation}\label{uc1 1}
D^\alpha_{t,\left\langle{\bf S}_0|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9}}\left\langle{\bf S}|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9}+ \left\langle{\bf S}|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9}=I^{1-\beta}_t\left( \left\langle{\bf A}_1|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9} \right)
\end{equation}
However, $\left| \left\langle{\bf S}|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9} \right|\leq \|{\bf S}\|_{L^2(\Omega)^9} \|\boldsymbol{\varphi}\|_{L^2(\Omega)^9} $. Since $\|{\bf S}\|_{L^2(\Omega)^9}\in \mathscr{L}$ and $\|\nabla{\bf u}\|_{L^2(\Omega)^9}\in \mathscr{L}$, we infer that $\left\langle{\bf S}|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9}\in \mathscr{L}$, and $\left\langle{\bf A}_1|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9}\in \mathscr{L}$. Now Lemma \ref{ul1} implies $I_t^{1-\beta}\left\langle {\bf A}_1|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9}\in\mathscr{L}$, and Proposition \ref{up2} leads to
\begin{eqnarray}\label{uc1 2}
\left\langle {\bf S}|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9} & = & E_\alpha \ast I^{1-\beta}_t\left( \left\langle{\bf A}_1|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9} \right)+\left\langle {\bf S}_0|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9}W_0 \nonumber\\
& = & \rho \ast \left\langle{\bf A}_1|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9} + \left\langle {\bf S}_0|\boldsymbol{\varphi} \right\rangle_{L^2(\Omega)^9}W_0
\end{eqnarray}
Notice that Eq.\eqref{uc1 2} still holds true for $\boldsymbol{\varphi} \in L^2(\Omega)^9$.
Let $ {\boldsymbol \theta}\in V$, $\psi\in \mathcal{C}_{00}^{+\infty}([0,+\infty[) $.
We deduce from \eqref{uc1 2} and \eqref{fv1} that:
\begin{eqnarray}
& & \psi(0)\int_{\Omega} {\bf u}_0({\bf x})\cdot {\boldsymbol \theta}({\bf x})\mathrm{d}{\bf x}+ \int_0^{+\infty}\int_{\Omega} {\bf u}(t,{\bf x}) \cdot {\boldsymbol \theta}({\bf x})\psi'(t)\mathrm{d}{\bf x}\mathrm{d} t \nonumber \\
& & =-\int_0^{+\infty} \left( \rho \ast \left\langle{\bf A}_1|\nabla{\boldsymbol \theta} \right\rangle_{L^2(\Omega)^9} + \left\langle {\bf S}_0|\nabla{\boldsymbol \theta} \right\rangle_{L^2(\Omega)^9}W_0 \right)(t)\psi(t) \mathrm{d} t \label{uc1 3}
\end{eqnarray}
We search for ${\bf u}\in L^1_{\text{loc}}(\mathbb{R}_+,V)$. In this case, for almost every $t>0$, ${\bf u}$ can be expressed as
\begin{equation}\label{wf3}
{\bf u}(t)=\displaystyle \sum_{q=1}^{+\infty}\alpha_q(t){\bf w}_q
\end{equation}
the series being convergent in $V$. It follows, by taking ${\boldsymbol \theta}={\bf w}_k$ and $\psi\in\mathscr{D}(]0,+\infty[)$ in equation \eqref{uc1 3}, that $\alpha'_k= -\lambda_k \left(\rho \ast \alpha_k \right)-b_k\sqrt{\lambda_k}W_0$, with $b_k:=\displaystyle \int_{\Omega}\left( {\bf S}_0:\nabla{\bf w}_k\right)({\bf x})\mathrm{d} {\bf x}$, the equality holding true in $ \mathscr{D}'(]0,+\infty[)$. Recall that $\alpha_k=\left\langle {\bf u}|{\bf w}_k \right\rangle_H \in \mathscr{C}^0(\mathbb{R}_+)$. As $W_0\in \mathscr{C}^0(\mathbb{R}_+)$, then necessarily $\alpha_k\in \mathscr{C}^1(\mathbb{R}_+)$. However, (cf \cite{pal7}) the Cauchy's initial value problem
\begin{subeqnarray}\slabel{al:a}
\alpha'_k(t) & = & -\lambda_k \left( \rho \ast \alpha_k \right)(t) - b_k \sqrt{\lambda_k} W_0(t) \slabel{al:a1}\\
\alpha_k(0) & = & \alpha_k^0 \slabel{al:a2}
\end{subeqnarray}
has a unique solution in $\mathscr{C}^1(\mathbb{R}_+)$. The uniqueness of the solution ${\bf u}$ is thus proved, and that of ${\bf S}$ follows.
\end{proof}
We now state an existence and uniqueness result. Denote $\mathscr{C}^0_b(\mathbb{R}_+^\ast,V):=\{ {\bf u} \in \mathscr{C}^0 (\mathbb{R}_+^\ast,V) \, \text{s.t.} \, \displaystyle\mathop{\sup}_{t\geq1}\|{\bf u}(t)\|_V < +\infty \}$. The functional space $\mathscr{C}^0_b(\mathbb{R}_+^\ast,L^2(\Omega)^9)$ is defined in a similar way.
\begin{theorem}[\textbf{First Existence and Uniqueness Theorem}]\label{ut1}
Let ${\bf u}_0 \in H$, ${\bf S}_0 \in L^2(\Omega)^9$. Then the boundary value problem given by the system of equations \eqref{fv1}-\eqref{fv2} has a unique solution
\begin{equation}\label{ut11}
({\bf u},{\bf S}) \in \left[ \mathscr{C}^0 (\mathbb{R}_+,H)\cap\mathscr{C}^0_b (\mathbb{R}_+^\ast,V)\cap L^1_{\text{loc}}(\mathbb{R}_+ ,V) \right] \times \left[ \mathscr{C}^0_b (\mathbb{R}_+^\ast,L^2(\Omega)^9)\cap L^1_{\text{loc}}(\mathbb{R}_+ ,L^2(\Omega)^9) \right]
\end{equation}
Moreover, ${\bf u}(0)={\bf u}_0$.
\end{theorem}
\begin{proof}
The existence of at least one solution
$$({\bf u},{\bf S}) \in \left[ \mathscr{C}^0 (\mathbb{R}_+,H)\cap\mathscr{C}^0 (\mathbb{R}_+^\ast,V)\cap L^1_{\text{loc}}(\mathbb{R}_+ ,V) \right] \times \left[ \mathscr{C}^0 (\mathbb{R}_+^\ast ,L^2(\Omega)^9)
\cap L^1_{\text{loc}}(\mathbb{R}_+ ,L^2(\Omega)^9) \right] $$
follows from Theorem 8.4 in \cite{pal7}.
It remains to be proved that the solution $({\bf u},{\bf S})$ obtained in \cite{pal7} satisfies
\begin{equation}\label{uta4}
\displaystyle\mathop{\sup}_{t\geq1} \|{\bf u}(t)\|_V+\displaystyle\mathop{\sup}_{t\geq1} \|{\bf S}(t)\|_{L^2(\Omega)^9}<+\infty
\end{equation}
This is essentially contained in the arguments presented in \cite{pal7}. Indeed, since ${\bf u}\in \mathscr{C}^0 (\mathbb{R}_+,V)$, we write ${\bf u}=\displaystyle\sum_{k=1}^{+\infty}\alpha_k(t){\bf w}_k$, $t>0$. From equation 125 in \cite{pal7} one gets
\begin{eqnarray}\label{wf4}
\lambda_k|\alpha_k(t)|^2 & \leq & M \bigg\{ \left[ \left(\dfrac{1}{t^{ \delta/2}}+\dfrac{1}{t^{1- \delta/2}} \right)^2+
\dfrac{\left(t\lambda_k^{1/(2-\delta)}\right)^{2(1-\delta/2)}}{t^{2(1-\delta/2)}}e^{-2\gamma t\lambda_k^{1/(2-\delta)}} \right]|\alpha_k^0|^2\nonumber\\
& + & \left( \dfrac{1}{t^{2(1-\delta)}}+\dfrac{(t\lambda_k^{1/(2-\delta)})^{2(1-\delta)}}{t^{2(1-\delta)}}e^{-2at\lambda_k^{1/(2-\delta)}}\right)|b_k|^2 \bigg\}
\end{eqnarray}
In the above, $\delta=(\alpha-\beta)\in]0,1[$, $a>0$, $\gamma>0$, and $\alpha_k^0 =\langle {\bf u}_0 | {\bf w}_k \rangle_{H}$. Clearly ${\bf u}_0\in L^2(\Omega)^3$, ${\bf S}_0\in L^2(\Omega)^9$ implies that $\displaystyle\sum_{k=1}^{+\infty}|\alpha_k^0|^2<+\infty$ and $\displaystyle\sum_{k=1}^{+\infty}|b_k |^2<+\infty$. Hence
\begin{equation}\label{uta5}
\|{\bf u}\|^2_V=\displaystyle\sum_{k=1}^{+\infty}\lambda_k|\alpha_k(t)|^2 \leq \dfrac{M}{t^{\inf (\delta,2-2\delta)}}\mathop{\longrightarrow}_{t\to+\infty}0
\end{equation}
and $ {\bf u}_0\in\mathscr{C}_b^0(\mathbb{R}_+^\ast,V)$.
We use the equation that defines ${\bf S}$ given right below equation 137 in \cite{pal7}. Then:
\begin{equation}\label{uta1}
\|{\bf S}\|_{L^2(\Omega)^9}\leq M \left[ \displaystyle \sum_{k=1}^{+\infty}\lambda_k |\rho\ast\alpha_k(t)|^2+|W_0(t)| \|{\bf S}_0\|_{L^2(\Omega)^9} \right]
\end{equation}
From \eqref{wf1} we see that
\begin{equation}\label{uta2}
|W_0(t)|\displaystyle \mathop{\longrightarrow}_{t\to+\infty}0
\end{equation}
Moreover, (see equation 130 in \cite{pal7})
$$\lambda_k |\rho\ast\alpha_k(t)|^2 \leq M \left[ \left( \dfrac{1}{t^\delta}+\dfrac{\left( t\lambda_k^{1/(2-\delta)}\right)^\delta}{t^\delta}\exp\left[-at\lambda_k^{1/(2-\delta)}\right] \right)|\alpha_k^0|^2 + \left( \dfrac{1}{t^{2\epsilon}}+\exp(-2at) \right)|b_k|^2 \right],\epsilon>0, a>0$$
Hence
\begin{equation}\label{uta3}
\displaystyle\sum_{k=1}^{+\infty}\lambda_k|\alpha_k(t)|^2 \leq \dfrac{M}{t^{\inf (\delta,2\epsilon)}}, \quad \text{for}\quad t\geq 1
\end{equation}
Finally (see \eqref{uta1}-\eqref{uta3}): $\|{\bf S}\|_{L^2(\Omega)^9} \displaystyle \mathop{\longrightarrow}_{t\to+\infty}0$.
It follows that ${\bf S}\in\mathscr{C}^0_b(\mathbb{R}_+^\ast,L^2(\Omega)^9)$. The existence of at least one solution belonging to the functional space of \eqref{ut11} is thus proved.
The uniqueness of such a solution results from Corollary \ref{uc1} and from the fact that
$$\left[ \mathscr{C}^0 (\mathbb{R}_+,H)\cap\mathscr{C}^0_b (\mathbb{R}_+^\ast,V)\cap L^1_{\text{loc}}(\mathbb{R}_+,V) \right] \times \left[ \mathscr{C}^0_b (\mathbb{R}_+^\ast,L^2(\Omega)^9) \cap L^1_{\text{loc}}(\mathbb{R}_+ ,L^2(\Omega)^9) \right] \subset \mathscr{F}$$.
\end{proof}
\section{Functional spaces}\label{so}
In order to prove the continuity of the $({\bf u},{\bf S})$ at $t=0$ we recall several classical functional spaces (see also \cite{lim} and \cite{rt}).
Denote ${\boldsymbol \epsilon}_k=\displaystyle \dfrac{\nabla {\bf w}_k}{\sqrt{\lambda_k}}$, $k\in\mathbb{N}^\ast$. Let $\Pi:L^2(\Omega)^9\to L^2(\Omega)^9$ be the orthogonal projection operator of $L^2(\Omega)^9 $ onto $\left[\text{Vect}({\boldsymbol \epsilon}_k)_{k\in\mathbb{N}^\ast} \right]^{\bot}$, $\theta\geq0$.
For any ${\bf f}\in L^2(\Omega)^9$, denote $\|{\bf f}\|^2_{D_\theta}:=\displaystyle \sum_{q=1}^{+\infty}\lambda_q^\theta |\left\langle {\bf f}|{\boldsymbol \epsilon}_q \right\rangle_{L^2(\Omega)^9} |^2+\|\Pi({\bf f})\|^2_{L^2(\Omega)^9}$. Let $D_\theta :=\{{\bf f} \in L^2(\Omega)^9 \,\text{s.t.}\,\|f\|_{D_\theta}<+\infty \}$.
For any ${\bf f},{\bf g} \in D_\theta $, denote $\left\langle {\bf f}|{\bf g}\right\rangle_{D_\theta}:=\displaystyle \sum_{q=1}^{+\infty}\lambda_q^\theta \left\langle {\bf f}|{\boldsymbol \epsilon}_q\right\rangle_{L^2(\Omega)^9} \overline{\left\langle {\bf g}|{\boldsymbol \epsilon}_q \right\rangle} _{L^2(\Omega)^9} +\left\langle \Pi({\bf f})|\Pi({\bf g})\right\rangle_{_{L^2(\Omega)^9}} $. The functional space $(D_\theta, \left\langle\,|\,\right\rangle_{D_\theta})$ is a Hilbert space.
For any ${\bf f}\in H$, let $\| {\bf f} \|^2_{H_\theta}:=\displaystyle \sum_{q=1}^{+\infty}\lambda_q^\theta |\left\langle {\bf f}|{\bf w}_q \right\rangle_{H}|^2$. Next, let $H_\theta :=\{{\bf f}\in H,\, \|{\bf f}\|_{H_\theta}<+\infty \}$. For any $f,g\in H_\theta$, $\left\langle f|g \right\rangle_{H_\theta}:=\displaystyle \sum_{k=1}^{+\infty}\lambda_q^\theta \left\langle f|{\bf w}_k \right\rangle_{H} \overline{\left\langle g|{\bf w}_k \right\rangle}_{H}$. As the sequence $({\bf w}_k)_{k\in\mathbb{N}^\ast}$ is complete in $L^2(\Omega)^9 $, the functional space $(H_\theta,\left\langle \quad|\quad\right\rangle_{H_\theta})$ is a Hilbert space.
Remark that, for any $0\leq \theta \leq \theta' \leq1 \leq \theta'' $, one has:
\begin{equation}\label{so1}
H=H_0\hookleftarrow H_{\theta}\hookleftarrow H_{\theta'}\hookleftarrow H_{1}=V \hookleftarrow H_{\theta''}
\end{equation}
\begin{equation}\label{so2}
L^2(\Omega)^9=D_0\hookleftarrow D_{\theta}\hookleftarrow D_{\theta'}\hookleftarrow D_{1} \hookleftarrow D_{\theta''}
\end{equation}
The above injections are dense; use of them will be often made from now on.
The following $\Delta_\theta$ spaces are closely related to the $D_\theta$ ones. Let $P:L^2(\Omega)^9 \to L^2(\Omega)^9$ be the orthogonal projection operator from $L^2(\Omega)^9$ onto $\left[\displaystyle \bigcup_{k=1}^{+\infty} \left\{ {\boldsymbol \epsilon}_k,{\boldsymbol \epsilon}_k^T \right\} \right]^{\perp}$.
Let $\theta\in\mathbb{R}$. For any element ${\bf f} \in L^2(\Omega)^9$, denote
\begin{equation}\label{sow1}
\|{\bf f}\|^2_{\Delta_\theta}=\displaystyle\sum_{q=1}^{+\infty}\lambda_q^\theta \left| \langle {\bf f} | {\boldsymbol \epsilon}_q \rangle_{L^2(\Omega)^9} \right|^2 + \displaystyle\sum_{q=1}^{+\infty}\lambda_q^\theta \left| \langle {\bf f} | {\boldsymbol \epsilon}_q^T \rangle_{L^2(\Omega)^9} \right|^2 + \left\| P({\bf f}) \right\|^2_{L^2(\Omega)^9}
\end{equation}
For any $\theta\geq0$, $\Delta_\theta :=\{ {\bf f} \in L^2(\Omega)^9 \,\text{s.t.}\, \left\| {\bf f} \right\|_{\Delta_\theta}<+\infty \}$. The functional space $\Delta_\theta$ endowed with the inner product defined as: $\forall {\bf f} \in \Delta_\theta$, $\forall {\bf g} \in \Delta_\theta$,
\begin{equation}\label{sow2}
\langle {\bf f} | {\bf g} \rangle_{\Delta_\theta} = \displaystyle\sum_{q=1}^{+\infty}\lambda_q^\theta \langle {\bf f} | {\boldsymbol \epsilon}_q \rangle_{L^2(\Omega)^9} \overline{ \langle {\bf g} | {\boldsymbol \epsilon}_q \rangle}_{L^2(\Omega)^9} + \displaystyle\sum_{q=1}^{+\infty}\lambda_q^\theta \langle {\bf f} | {\boldsymbol \epsilon}_q^T \rangle_{L^2(\Omega)^9} \overline{ \langle {\bf g} | {\boldsymbol \epsilon}_q^T \rangle}_{L^2(\Omega)^9} + \langle P({\bf f})|P({\bf g}) \rangle_{L^2(\Omega)^9}
\end{equation}
is a Hilbert space.
For any $\theta<0$, let $\Delta_\theta$ denote the topological dual space of $\Delta_{-\theta}$, i.e. $\Delta_\theta=\left(\Delta_{-\theta}\right)'$. The space $\Delta_\theta$ is the completion of $(L^2(\Omega)^9,\|\,\|_{\Delta_\theta})$. Next, note that whenever $0\leq \gamma\leq \tilde{\gamma}$, the following injections $L^2(\Omega)^9=\Delta_0 \hookleftarrow \Delta_\gamma \hookleftarrow \Delta_{\tilde{\gamma}}$ are dense. It results that $\Delta_{-\tilde{\gamma}} \hookleftarrow \Delta_{-\gamma} = \left(\Delta_\gamma \right)' \hookleftarrow \left( L^2(\Omega)^9 \right)' \simeq L^2(\Omega)^9= \Delta_0 \hookleftarrow \Delta_\gamma \hookleftarrow \Delta_{\tilde{\gamma}}$, invoking the fact that $L^2(\Omega)^9 $ and $\left( L^2(\Omega)^9 \right)'$ are isomorphic to each other.
Next, for any ${\bf f}\in H_0^1(\Omega)^9$, one has:
\begin{eqnarray}\label{so3 3}
\|{\bf f}\|_{\Delta_1}^2 & = & \sum_{q=1}^{+\infty}\lambda_q \left|\langle {\bf f}|{\boldsymbol \epsilon}_q \rangle_{L^2(\Omega)^9} \right|^2+\sum_{q=1}^{+\infty}\lambda_q \left|\langle {\bf f}|{\boldsymbol \epsilon}_q^T \rangle_{L^2(\Omega)^9} \right|^2+\|P({\bf f})\|^2_{L^2(\Omega)^9} \nonumber\\
& = & \sum_{q=1}^{+\infty} \left|\langle \nabla\cdot {\bf f}|{\bf w}_q \rangle_{L^2(\Omega)^3} \right|^2+\left|\langle \nabla\cdot {\bf f}^T|{\bf w}_q \rangle_{L^2(\Omega)^3} \right|^2+\|P({\bf f})\|^2_{L^2(\Omega)^3} \label{so3 2} \nonumber\\
& \leq & K\|{\bf f}\|^2_{H_0^1(\Omega)^9}
\end{eqnarray}
due to the Poincar\'e's inequality. Consequently $H_0^1(\Omega)^9\hookrightarrow \Delta_1$ and the restriction $r:\Delta'_1\to H^{-1}(\Omega)^9 $, such that
$T \stackrel{r}{\mapsto} T|_{H^1_0(\Omega)^9} $ is continuous.
\begin{lemma}\label{sxl4}
Let $\theta\in\mathbb{R}_+$.
\begin{enumerate}[(a)]
\item\label{a}The sequence $({\boldsymbol \epsilon}_k)_{k\in \mathbb{N}^\ast}$ is orthogonal in $D_\theta$. Moreover, $\|{\boldsymbol \epsilon}_k\|_{D_\theta}=\lambda_k^{\theta/2}$.
\item\label{b} The sequence $({\boldsymbol \epsilon}^T_k)_{k\in \mathbb{N}^\ast}$ is orthogonal in $D_\theta$.
\item\label{c} The sequence $({\boldsymbol \epsilon}_k+{\boldsymbol \epsilon}^T_k)_{k\in \mathbb{N}^\ast}$ is orthogonal in $D_\theta$, and $\|{\boldsymbol \epsilon}_k+{\boldsymbol \epsilon}^T_k\|_{D_\theta}=(1+\lambda_k^{\theta})^{1/2}$.
\item\label{d} The sequence $({\bf w}_k)_{k\in \mathbb{N}^\ast}$ is orthogonal in $H_\theta$, and $\|{\bf w}_k\|_{H_\theta}=\lambda_k^{\theta/2}$.
\item\label{e} Let ${\bf f}\in L^2(\Omega)^9$. Denote ${\bf f}=\displaystyle \sum_{k=1}^{+\infty}a_k {\boldsymbol \epsilon}_k + \displaystyle \sum_{k=1}^{+\infty}b_k {\boldsymbol \epsilon}^T_k + P({\bf f})$. Then:
$\|{\bf f}\|^2_{\Delta_{-\theta}}=\displaystyle \sum_{k=1}^{+\infty} \lambda_k^{-\theta} \left( |a_k|^2+|b_k|^2\right) + \|P({\bf f})\|^2_{L^2(\Omega)^9}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Observe that (cf. eq.\eqref{eig1}) for any $(k,q)\in \mathbb{N}^{\ast 2} $:
\begin{equation}\label{sol1}
\left\langle{\boldsymbol \epsilon}_k|{\boldsymbol \epsilon}_q \right\rangle _{L^2}=\left\langle{\bf w}_k|{\bf w}_q \right\rangle _{L^2}=\delta_{kq}
\end{equation}
On the other hand, since $\nabla\cdot {\bf w}_k=0$,
\begin{eqnarray}\label{sol3}
\left\langle{\boldsymbol \epsilon}_k|{\boldsymbol \epsilon}^{T}_q \right\rangle _{L^2} & = & \sum_{i,j}\int_\Omega \dfrac{\partial ({\bf w}_k)_i }{\partial x_j}({\bf x})\overline{\dfrac{\partial ({\bf w}_q)_j }{\partial x_i}}({\bf x}) \mathrm{d} {\bf x} \label{sol2} \nonumber\\
& = & \int_\Omega \left(\sum_{i}\dfrac{\partial ({\bf w}_k)_i }{\partial x_j}({\bf x}) \right)\left(\overline{ \sum_{j}\dfrac{\partial ({\bf w}_q)_j }{\partial x_i}}({\bf x})\right) \mathrm{d} {\bf x}=0
\end{eqnarray}
Hence:
\begin{equation}\label{sol4}
\left\langle{\boldsymbol \epsilon}_k|{\boldsymbol \epsilon}^{T}_q \right\rangle _{L^2} =0
\end{equation}
The statements (a) to (e) result from Eqs.\eqref{sol1}-\eqref{sol4}.
\end{proof}
Except for the injection $H_{2k} \hookrightarrow H^{2k}(\Omega)^3$ (see below), the following description of the spaces will not be used in this paper.
Let first $\theta\in[0,1]$.
Let $\Lambda:V \to H$, such that $\Lambda\displaystyle\left(\sum_{k\in\mathbb{N}^\ast}a_k{\bf w}_k \right) = \displaystyle\sum_{k\in\mathbb{N}^\ast}a_k\sqrt{\lambda_k}{\bf w}_k $. Then, for any $ ({\bf u},{\bf v})\in V^2$, $\left\langle\Lambda{\bf u}|\Lambda{\bf v} \right\rangle_H =\left\langle{\bf u}|{\bf v} \right\rangle_V$, and $H_\theta=D(\Lambda^\theta)=[V,H]_{1-\theta}$, where $[V,H]_{1-\theta}$ stands for the holomorphic interpolation of spaces $V$ and $H$, and $D(\Lambda^\theta)$ for the domain of $\Lambda^\theta$. Denote $H^0_0(\Omega)\equiv L^2(\Omega)$. Let the canonical injection $H\stackrel{i}{\hookrightarrow} H^0_0(\Omega)^3$ and $V\stackrel{i|_V}{\hookrightarrow} H^1_0(\Omega)^3$ be its restriction.
Then $\displaystyle H_\theta =\displaystyle [V,H]_{1-\theta}\stackrel{i|_{[V,H]_{1-\theta}}}{\hookrightarrow}[H^1_0(\Omega)^3,H^0_0(\Omega)^3]_{1-\theta}\hookrightarrow H^\theta_0(\Omega)^3$ (the last continuous injection $\hookrightarrow$ boils down to an equality $=$ whenever $\theta\neq 1/2$). Let now $\gamma_n$ denote the normal-trace application. It is well known that $H=\{ {\bf u}\in H^0(\Omega)^3,\,\text{s.t.}\,\nabla\cdot{\bf u} =0, \,\gamma_n({\bf u})=0 \}$. From the preceding arguments it results that we have the continuous injection $H_\theta \hookrightarrow H\cap H^\theta_0(\Omega)^3=\{ {\bf u}\in H^\theta_0(\Omega)^3,\,\text{s.t.}\,\nabla\cdot{\bf u} =0, \,\gamma_n({\bf u})=0 \}$, with the space $H\cap H^\theta_0(\Omega)^3$ being endowed with the $H^\theta_0(\Omega)^3$ topology.
Let now $\theta\in\mathbb{N}^\ast$.
As quoted on page 106 in \cite{rt}, $H_2=D(\Lambda^2)=H^2(\Omega)^3\cap V$. Also, invoking Agmon - Douglis - Nirenberg's Theorem as stated on page 832 in \cite{dl}, leads to $H_{2k}=D(\Lambda^{2k})\hookrightarrow H^{2k}(\Omega)^3\cap V$, $k\in\mathbb{N}^\ast$. Here $H^{2k}(\Omega)$ are classical Sobolev spaces.
\section{The continuity of the solution $({\bf u},{\bf S})$ at $t=0$}\label{n}
From now on $({\bf u},{\bf S})$ denotes the solution to equations \eqref{fv1}-\eqref{fv2}, with initial data $({\bf u}_0,{\bf S}_0)\in H\times L^2(\Omega)^9$ (see Theorem \ref{ut1}). In order to prove continuity results we recall several representation formulas for ${\bf u}$ and ${\bf S}$. First, functions $\alpha_k$, $k\in \mathbb{N}^\ast$, are defined by equations \eqref{al:a1}-\eqref{al:a2}. Equivalently, for $x\in\mathbb{R}_+$ (see \cite{pal7}),
\begin{equation} \label{per3 2}
\alpha_k(t) = \dfrac{1}{2\pi}\left[\lim_{A\rightarrow+\infty} \int_{-A}^{+A} T_{\lambda_k} (x+iy)e^{(x+iy)t}\mathrm{d} y \alpha_k^0
- \lim_{A\rightarrow+\infty} \int_{-A}^{+A}
(T_{\lambda_k}w) (x+iy)e^{(x+iy)t}\mathrm{d} y \sqrt{\lambda_k} b_k \right]
\end{equation}
with $\displaystyle T_\mu (s)=\dfrac{s^{1-\beta}(s^\alpha+1)}{s^{2-\beta}(s^\alpha+1)+\mu}$, $\displaystyle w (s)=\dfrac{1}{s^{1-\alpha}(s^\alpha+1) }$, $\mu\in\mathbb{R}_+$, $s\in \mathbb{C}-\mathbb{R}_-$, and $\alpha_k^0=\langle {\bf u}_0|{\bf w}_k \rangle_H$, $b_k=\displaystyle\int_\Omega ({\bf S}_0:\nabla {\bf w}_k)({\bf x})\mathrm{d}{\bf x} $.
Notice that eq.\eqref{per3 2} is given in \cite{pal7} only for $x\geq M$. The general result ($x\in \mathbb{R}_+$) follows from a simple use of the Cauchy formula; details are omitted. Regarding function ${\bf S}$, recall the following formula from \cite{pal7}:
\begin{equation}\label{t14 3}
{\bf S}=\sum_{k=1}^{+\infty}(\rho\ast\alpha_k)\otimes(\nabla{\bf w}_k+\nabla^T{\bf w}_k)+W_0\otimes {\bf S}_0
\end{equation}
Notation $h=f\otimes g$ means $h(x,y)=f(x)g(y)$. As quoted in \cite{pal7}, the series in \eqref{t14 3} converges in $\mathscr{C}^0(\mathbb{R}_+^\ast,L^2(\Omega)^9)$ and in $L^2_{\text{loc}}(\mathbb{R}_+ ,L^2(\Omega)^9)$. Here $\rho=E_\alpha\ast \dfrac{t^{-\beta}}{\Gamma(1-\beta)}$, and $0<\rho(t)\leq k t^{-\delta}$ (see \cite{pal7}), $\delta=\beta-\alpha$.
The following estimate will give the continuity at $t=0$ of $({\bf u},{\bf S})$.
\begin{lemma}\label{sol5}
For any $\mu_0>0$, $\exists M>0$, such that $\forall ({\bf x},t)\in(\mathbb{R}_+)^3 \times \mathbb{R}_+ $, and $\forall \mu\geq\mu_0$, we have:
\begin{equation}\label{sol6}
\displaystyle\mathop{\lim}_{A\to\infty}\displaystyle\left|\int_{-A}^{+A}T_\mu w(x+iy)e^{(x+iy)t} \mathrm{d} y \right| \leq \displaystyle\int_{-\infty}^{+\infty}\left|(T_\mu w)(iy) \right|\mathrm{d} y \leq \dfrac{M}{\mu^{1/(2-\delta)}}
\end{equation}
\end{lemma}
\begin{proof}
Whenever $y>0$, $\displaystyle \left|(T_\mu w)(iy) \right|=\dfrac{1}{|y|^\delta}\dfrac{1}{\left|\mu+\left(y e^{i\pi/2} \right)^{2-\delta}+ \left(y e^{i\pi/2} \right)^{2-\beta} \right|} $. Therefore,
\begin{eqnarray}\label{sol7 5}
\left|\mu+\left(y e^{i\pi/2} \right)^{2-\delta}+ \left(y e^{i\pi/2} \right)^{2-\beta} \right| & \geq & \left|\mathrm{Im} \left[ \mu e^{i\pi(\beta-2)/2} + y^{2-\delta} e^{i\pi\alpha/2}+y^{2-\beta} \right] \right| \nonumber\\
& \geq & \left|-\mu \sin(\pi\beta/2)+y^{2-\delta}\sin(\pi\alpha/2) \right| \nonumber\\
& \geq & \mu \sin(\pi\beta/2)-y^{2-\delta}\sin(\pi\alpha/2) \nonumber\\
& \geq & K\mu, \,\text{for}\,y\leq \mu^{1/(2-\delta)}
\end{eqnarray}
The constant $K=\sin(\pi\beta/2)-\sin(\pi\alpha/2)>0$ is independent of $\mu$. Moreover,
\begin{eqnarray}\label{sol8 5}
\left|\mu+\left(y e^{i\pi/2} \right)^{2-\delta}+ \left(y e^{i\pi/2} \right)^{2-\beta} \right| \label{sol8 1} & \geq & \left|\mathrm{Im} \left[ \mu+ \left(y e^{i\pi/2} \right)^{2-\delta} + \left(y e^{i\pi/2} \right)^{2-\beta} \right] \right| \label{sol8 2} \nonumber\\
& = & \left|y^{2-\delta} \sin(\pi-\pi\delta/2)+y^{2-\beta}\sin(\pi-\pi\beta/2) \right| \nonumber\\
& = & \left|y^{2-\delta} \sin(\pi\delta/2)+y^{2-\beta}\sin(\pi\beta/2) \right| \nonumber\\
& \geq & Ky^{2-\delta}, \,\text{for}\,y\geq \mu_0^{1/(2-\delta)}
\end{eqnarray}
From the above estimates we infer that:
\begin{eqnarray}\label{sol9 3}
\int_{0}^{+\infty}|(T_\mu w)(iy)|\mathrm{d} y & = & \int_0^{\mu^{1/(2-\delta)}}\dfrac{\mathrm{d} y}{y^\delta\left|\mu+\left(y e^{i\pi/2} \right)^{2-\delta}+ \left(y e^{i\pi/2} \right)^{2-\beta} \right|} \nonumber\\
& + & \int_{\mu^{1/(2-\delta)}}^{+\infty}\dfrac{\mathrm{d} y}{y^\delta\left|\mu+\left(y e^{i\pi/2} \right)^{2-\delta}+ \left(y e^{i\pi/2} \right)^{2-\beta} \right|} \nonumber\\
& \leq & \int_0^{\mu^{1/(2-\delta)}} \dfrac{K}{\mu y^\delta}\mathrm{d} y+\int_{ \mu^{1/(2-\delta)}}^{+\infty}\dfrac{K}{y^\delta y^{2-\delta}}\mathrm{d} y \nonumber\\
& \leq & \dfrac{M}{\mu^{1/(2-\delta)}}
\end{eqnarray}
A similar estimate can be obtained for $\displaystyle\int_{-\infty}^0|(T_\mu w)(iy)|\mathrm{d} y $. Combining these results achieves the proof.
\end{proof}
Denote, $\delta=\beta-\alpha$, $\omega=\delta/(2-\delta)$ and notice that $0<\omega<\delta<1$. From now on we shall sometimes write $\alpha_k ({\bf u}_0,{\bf S}_0 )$ instead of $\alpha_k $; of course $\alpha_k $ is linear w.r.t. initial data $ ({\bf u}_0,{\bf S}_0 )$. Most of the following estimates are already proved in \cite{pal7}, save for those derived from Lemma \ref{sol5}.
\begin{proposition}\label{sot1}
Let ${\bf u}_0\in H$, ${\bf S}_0\in L^2(\Omega)^9$. Then exists $\exists M>0$, such that, $\forall t \in \mathbb{R}_+$ and $\forall k \in \mathbb{N}^\ast$,
\begin{enumerate}[(i)]
\item \label{i} $|\alpha_k(t)|^2\leq M \left(|\alpha_k^0|^2+\lambda_k^{-\omega}|b_k|^2\right)$.
\item \label{ii} $\lambda_k|\alpha_k(t)|^2\leq M \left(\dfrac{|\alpha_k^0|^2}{t^{2-\delta}}+\dfrac{|b_k|^2}{t^{2-2\delta}} \right)$.
\item \label{iii} for any $\mu \in [0,1]$ and any $\tau \in [0,1]$,
$$ |\alpha_k(t)|^2\leq M \left(\dfrac{|\alpha_k^0|^2}{\lambda_k^{\mu} t^{\mu(2-\delta)}}+\dfrac{|b_k|^2}{\lambda_k^{\tau+(1-\tau)\omega} t^{2\tau(1-\delta)}} \right)$$
\item \label{iv} for any $\mu \in [0,1]$ and any $\tau \in [0,1]$,
$$\lambda_k |\rho \displaystyle \ast \alpha_k |^2(t) \leq M \left( \dfrac{|\alpha_k^0|^2}{\lambda_k^{\mu-1} t^{\mu(2-\delta)+2(\delta-1)}}+\dfrac{|b_k|^2}{\lambda_k^{-(1-\tau)(1-\omega)} t^{-2(1-\tau)(1-\delta) }} \right)$$
\end{enumerate}
\end{proposition}
\begin{proof}
(i) \\
Since $\alpha_k$ is linear with respect to $({\bf u}_0,{\bf S}_0)$, we have, by eq.(121) in Theorem 8.1 in \cite{pal7} and eq.\eqref{per3 2}:
\begin{equation}\label{sot1 1}
|\alpha_k ({\bf u}_0,{\bf S}_0)|^2\leq 2|\alpha_k({\bf u}_0,0)|^2+2|\alpha_k(0,{\bf S}_0)|^2 \leq M \left( |\alpha_k^0|^2+\left| \int_{-\infty}^{+\infty}(T_{\lambda_k} w)(iy)e^{iyt}\mathrm{d} y\sqrt{\lambda_k}b_k \right|^2\right)
\end{equation}
Invoking Lemma \ref{sol5} we get
\begin{equation}\label{sot1 4}
|\alpha_k({\bf u}_0,{\bf S}_0)|^2\leq M \left( |\alpha_k^0|^2 +\dfrac{|b_k|^2 \lambda_k}{\lambda_k^{2 (2-\delta)}}\right)
\end{equation}
which gives (i).\\
(ii)\\
Estimate (ii) is obtainable right away from eq.(122) in Theorem 8.1 in \cite{pal7}, with $M_T$ instead of $M$. The proof that $M$ can be chosen independently of $T$ is deferred until Corollary \ref{ic1} in Section \ref{ie}. Hence we take here $M$ independent of $T$ and proceed further on.\\
(iii)\\
Notice first that (ii) above gives
\begin{equation}\label{sot1 5}
\left| \alpha_k ({\bf u}_0,0) \right|^2(t) \leq \dfrac{M|\alpha_k^0|^2}{\lambda_k t^{2-\delta}}
\end{equation}
and
\begin{equation}\label{sot1 6}
\left| \alpha_k (0,{\bf S}_0) \right|^2(t) \leq \dfrac{M| b_k |^2}{\lambda_k t^{2-2\delta}}
\end{equation}
Next, combining eq.\eqref{sot1 5} and eq.\eqref{sot1 4} with ${\bf S}_0={\bf 0}$ on one hand, and eq.\eqref{sot1 6} and eq.\eqref{sot1 4} with ${\bf u}_0={\bf 0}$ on the other, making further use of eq.\eqref{sot1 1} leads to estimate (iii).\\
(iv)\\
Since $0\leq \rho(t) \leq K/t^\delta$, estimate (iii) gives
\begin{eqnarray*}
\sqrt{\lambda_k} \left| \rho \displaystyle \ast \alpha_k \right|(t) & \leq & \sqrt{\lambda_k} \left( \rho \displaystyle \ast |\alpha_k| \right)(t) \\
& \leq & M \left(\dfrac{\sqrt{\lambda_k} |\alpha_k^0| }{\lambda_k^{\mu/2} t^{\mu(2-\delta)/2+\delta-1} } + \dfrac{\sqrt{\lambda_k} |b_k|}{ \lambda_k^{[\tau+\omega(1-\tau)]/2} t^{\tau(1-\delta) +\delta-1} } \right),
\end{eqnarray*}
from which (iv) is obtained.
\end{proof}
In order to work on spaces $H_\theta$ and $D_\gamma$, we need to reformulate Proposition \ref{sot1}. Let $[\quad]_+$ denote the positive part of a real number.
\begin{lemma}\label{sxl3}
Let ${\bf u}_0\in H$, ${\bf S}_0\in L^2(\Omega)^9$, and $0\leq\gamma\leq\theta\leq\gamma+1$. Then $\exists M>0$, so that $\forall t\geq0$,
\begin{enumerate}[(i)]
\item \label{i}
$$\lambda_k^\theta |\alpha_k(t)|^2\leq M\left( \lambda_k^\theta |\alpha_k^0|^2+\dfrac{|\lambda_k|^\gamma |b_k|^2}{t^{2(1-\delta)[(\theta-\gamma-\omega)/(1-\omega)]_+}} \right) $$
\item \label{ii}
$$\lambda_k^\gamma \left\langle \lambda_k|\rho \displaystyle \ast \alpha_k(t)|^2 (t)\right\rangle\leq M \left(\lambda_k^\theta|\alpha_k^0|^2 t^{\delta(\theta-\gamma-\omega)/\omega}+\lambda_k^\gamma|b_k|^2 \right) $$
\end{enumerate}
\end{lemma}
\begin{proof}
(i) \\
Assume $0\leq\gamma\leq\theta\leq\gamma+\omega$. Multiplying (i) of Theorem \ref{sot1} by $\lambda_k^\theta$ leads to $\lambda_k^\theta|\alpha_k(t)|^2\leq A\left[ \lambda_k^\theta|\alpha_k^0|^2+\lambda_k^{\theta-\omega}|b_k|^2 \right] $. Since $\displaystyle\mathop{\min}_{k\geq1}\lambda_k>0$ and $\theta-\omega\leq\gamma $, one gets $\lambda_k^\theta|\alpha_k(t)|^2\leq M\left[ \lambda_k^\theta|\alpha_k^0|^2+\lambda_k^{\gamma}|b_k|^2 \right] $.
Assume now that $0\leq\gamma+\omega\leq\theta\leq \gamma+1$. Use part (iii) of Theorem \ref{sot1} with $\mu=0$ and $\tau=\dfrac{\theta-\gamma-\omega}{1-\omega}\in[0,1] $. A simple calculation leads to $\tau+(1-\tau)\omega=\tau(1-\omega)+\omega=(\theta-\gamma-\omega)+\omega=\theta-\gamma$. Henceforth:
\begin{equation*}
|\alpha_k(t)|^2 \leq M \left( |\alpha_k^0|^2+ \dfrac{|b_k|^2}{\lambda_k^{\theta-\gamma} t^{2(1-\delta)(\theta-\gamma-\omega)/(1-\omega)}} \right) \\
\end{equation*}
which leads to (i).\\
(ii) \\
Assume $0\leq\gamma\leq\theta\leq \gamma+1$. Letting $\mu=\gamma-\theta+1 \in [0,1]$ and $\tau=1$ in part (iv) of Lemma \ref{sot1}, gives $\mu(2-\delta)+2(\delta-1)=(\gamma-\theta)(2-\delta)+\delta=-\delta(\theta-\gamma-\omega)/\omega$. Hence
$$ \lambda_k |\rho\displaystyle \ast \alpha_k|^2(t) \leq
M \left( \dfrac{|\alpha_k^0|^2}{\lambda_k^{ \gamma-\theta }t^{-\delta(\theta-\gamma-\omega)/ \omega }}
+|b_k|^2 \right)$$
which gives (ii).
\end{proof}
As a consequence, we have:
\begin{corollary}\label{soc}
Let $0\leq \gamma\leq\theta\leq\gamma+1$, ${\bf u}_0\in H_\theta$, ${\bf S}_0\in D_\gamma$. Then ${\bf u}\in \mathscr{C}^{0}(\mathbb{R}^\ast_+,H_{\gamma+1})$. In addition,
\begin{enumerate}[(a)]
\item \label{a} whenever $0\leq \gamma\leq\theta\leq\gamma+\omega $, ${\bf u}\in \mathscr{C}^{0}(\mathbb{R}_+,H_\theta)$, and ${\bf S}\in \mathscr{C}^{0}(\mathbb{R}^\ast_+,D_\gamma)$; moreover, ${\bf u}(0)={\bf u}_0$.
\item \label{b} whenever $0\leq \gamma+\omega\leq\theta $, ${\bf u}\in \mathscr{C}^{0}(\mathbb{R}_+,H_{\gamma+\omega})$, and ${\bf S}\in \mathscr{C}^{0}(\mathbb{R}_+,D_\gamma)$; moreover, ${\bf u}(0)={\bf u}_0$ and ${\bf S}(0)={\bf S}_0$.
\end{enumerate}
In both cases, for any $t\geq0$,
\begin{equation}\label{soc1}
\|{\bf u}(t)\|_{H_\theta}\leq M \left( \|{\bf u}_0\|_{H_\theta}+\dfrac{\|{\bf S}_0\|_{D_\gamma}}{t^{(1-\delta)[(\theta-\gamma-\omega)/(1-\omega)]_+}} \right)
\end{equation}
\begin{equation}\label{soc2}
\|{\bf S}(t)\|_{D_\gamma}\leq M \left( t^{\delta(\theta-\gamma-\omega)/(2\omega)}\|{\bf u}_0\|_{H_\theta}+\|{\bf S}_0\|_{D_\gamma}\right)
\end{equation}
\end{corollary}
\begin{proof}
We first prove that ${\bf u}\in \mathscr{C}^{0}\left(\mathbb{R}^\ast_+,H_{\gamma+1}\right)$.
From (d) of Lemma \ref{sxl4} and (ii) in Proposition \ref{sot1}, we reckon that, for any $N\leq M$ and $t\in[T_1,T_2]$, where $0<T_1<T_2$,
\begin{equation}\label{soc3:a2}
\left\| \displaystyle\sum_{k=N}^{M}\alpha_k(t) {\bf w}_q \right\|^2_{H_{\gamma+1}} = \displaystyle\sum_{k=N}^{M} |\alpha_k(t)|^2\lambda_k^{\gamma+1}
\leq M_T \displaystyle\sum_{k=N}^{+\infty} \left[\dfrac{\lambda_k^\gamma|\alpha_k^0|^2}{T_1^{2-\delta}} + \dfrac{\lambda_k^\gamma|b_k |^2}{T_1^{2(1-\delta)}} \right]
\end{equation}
Since $\gamma\leq\theta$, we have that ${\bf u}_0\in H_\theta\subset H_\gamma$. Also ${\bf S}_0\in H_\gamma$. Hence Eq.\eqref{soc3:a2} implies that
\begin{equation}\label{socx3}
\displaystyle\mathop{\sup}_{t\in[T_1,T_2]} \left\| \displaystyle\sum_{k=N}^{M}\alpha_k(t) {\bf w}_q \right\|_{H_{\gamma+1}} \displaystyle\mathop{\longrightarrow}_{N\to+\infty}0
\end{equation}
Finally, as $\alpha_k\in \mathscr{C}^{0}\left(\mathbb{R}_+\right)$, from \eqref{socx3} above we deduce that ${\bf u}=\left( \displaystyle\sum_{k=N}^{+\infty} \alpha_k \otimes {\bf w}_q \right)\in \mathscr{C}^{0}\left(\mathbb{R}^\ast_+,H_{\gamma+1}\right)$.
Next we proceed with the rest of the proof.\\
(a) \\
Let us prove that ${\bf S}\in \mathscr{C}^{0}(\mathbb{R}^\ast_+,D_\gamma)$. For any $N \leq M$, by (c) of Lemma \ref{sxl4} and by (ii) of Lemma \ref{sxl3},
\begin{eqnarray}\label{sx:x}
\left\|\sum_{k=N}^{M}\left(\rho \ast \alpha_k \right)(t)\left(\nabla{\bf w}_k+\nabla^T{\bf w}_k \right) \right\|^2_{D_\gamma} & = & \sum_{k=N}^{M}|\rho\mathop{\ast}_{(t)}\alpha_k|^2(1+\lambda_k^\gamma)\lambda_k \nonumber\\
& \leq & K\sum_{k=N}^{+\infty} \left[ \lambda^\theta_k |\alpha_k^0|^2 t^{\delta(\theta-\gamma-\omega)/\omega}+\lambda^\gamma_k |b_k|^2 \right]
\end{eqnarray}
Since ${\bf u}_0\in H_\theta$, and ${\bf S}_0\in D_\gamma$, then $\displaystyle \sum_{k=1}^{+\infty}\left(\rho \displaystyle \ast \alpha_k \right)(t) \left(\nabla{\bf w}_k+\nabla^T{\bf w}_k \right) $ is uniformly convergent w.r.t. $t$ on any compact subset $[T_0,T_1]\subset\mathbb{R}_+^\ast$, in ${D_\gamma}$. Given that $\alpha_k\in\mathscr{C}^0(\mathbb{R}_+)$, and that $\rho\in L^1_{\text{loc}}(\mathbb{R}_+)$ - and hence $(\rho\displaystyle \ast \alpha_k )\in \mathscr{C}^0(\mathbb{R}_+)$ -, we conclude that $\displaystyle \sum_{k=1}^{+\infty}\left(\rho \ast \alpha_k \right) \otimes \left(\nabla{\bf w}_k+\nabla^T{\bf w}_k \right) \in \mathscr{C}^0([T_0,T_1], D_\gamma)$. As $W_0\in \mathscr{C}^0(\mathbb{R}_+)$, one gets that (see \eqref{t14 3})
\begin{equation}\label{soc4}
{\bf S}=\displaystyle \sum_{k=1}^{+\infty}\left(\rho \ast \alpha_k \right) \otimes \left(\nabla{\bf w}_k+\nabla^T{\bf w}_k \right)+W_0\otimes{\bf S}_0 \in \mathscr{C}^0(\mathbb{R}_+, D_\gamma)
\end{equation}
The inequality Eq.\eqref{soc2} results from eq.\eqref{sx:x} by letting $N=1$, $M=+\infty$, and from the fact that $\| W_0(t){\bf S}_0 \|_{D_\gamma} \leq \|W_0\|_{\infty}\|{\bf S}_0\|_{D_\gamma}$ for any $t\in\mathbb{R}_+$, since $W_0\in L^{\infty}(\mathbb{R}_+)\cap \mathscr{C}^0(\mathbb{R}_+)$. In a similar way we prove that ${\bf u} \in \mathscr{C}^0(\mathbb{R}_+, H_\theta)$.
The inequality eq.\eqref{soc1} is a consequence of (i) in Lemma \ref{sxl3} and of the fact that ${\bf u}\in\mathscr{C}^0(\mathbb{R}_+, H_\theta)$. Finally, as $H_{\theta}\hookrightarrow H$ and $\displaystyle\mathop{\lim}_{t\to0}{\bf u}(t)={\bf u}_0$ in $H$ (see Theorem 8.4 in \cite{pal7}), $\displaystyle\mathop{\lim}_{t\to0}{\bf u}(t)={\bf u}_0$ in $H_{\theta}$. \\
(b)\\
Whenever $\theta\geq \omega+\gamma$, ${\bf u}_0\in H_{\gamma+\omega}$ and ${\bf S}_0\in D_{\gamma}$. The inequalities (i) and (ii) of Lemma \ref{sxl3}, for $\theta'=\omega+\gamma$ and $\gamma'=\gamma$, read
\begin{eqnarray*}
& & \lambda_k^{\theta'}|\alpha_k(t)|^2\leq M\left(\lambda_k^{\theta'}|\alpha_k^0|^2 +\lambda_k^{\gamma'}|b_k|^2\right) \\
& & \lambda_k^{\gamma'} \left( \lambda_k| \rho \ast \alpha_k|^2(t) \right) \leq M \left( \lambda_k^{\theta'}|\alpha_k^0|^2+\lambda_k^{\gamma'}|b_k|^2 \right)
\end{eqnarray*}
The proof of (i) of Lemma \ref{sxl3} entails the uniform convergence (with respect to $t$ on $[0,+\infty[$) of
$\displaystyle \sum_{k=1}^{+\infty}\left(\rho \displaystyle \ast \alpha_k \right)(t)\left(\nabla{\bf w}_k+\nabla^T{\bf w}_k \right)$ in $D_\gamma$.
Therefore ${\bf S}\in \mathscr{C}^0(\mathbb{R}_+, D_\gamma)$. Moreover, as $W_0(0)=1$, we get (see \eqref{t14 3}) ${\bf S}(0)=\displaystyle \sum_{k=1}^{+\infty}\left(\rho \ast \alpha_k \right)(0)\left(\nabla{\bf w}_k+\nabla^T{\bf w}_k \right)+W_0(0) {\bf S}_0= {\bf S}_0$. Arguing as in (a) above one proves that: ${\bf u}\in \mathscr{C}^0(\mathbb{R}_+, H_{\gamma+\omega})$ and ${\bf u}(0)={\bf u}_0$.
\end{proof}
In the case $\gamma\geq0$ and $\theta=\gamma+\omega$ we get the continuity of $({\bf u},{\bf S})$ at $t=0$:
\begin{theorem}[\textbf{Existence Theorem}] \label{sot2}
Let $\gamma\geq0$. Assume that ${\bf u}_0\in H_{\gamma+\omega}$ and ${\bf S}_0\in D_{\gamma}$. Then the system of Eqs.\eqref{fv1}-\eqref{fv2} has at least one solution $({\bf u},{\bf S})\in \left[\mathscr{C}^0\left(\mathbb{R}_+,H_{\gamma+\omega} \right)\cap \mathscr{C}^0(\mathbb{R}^\ast_+,V)\cap L^p_{\text{loc}}(\mathbb{R}_+,V) \right] \times \mathscr{C}^0(\mathbb{R}_+,D_\gamma) $, $p\in [1,2/(2-\delta)[$. Moreover, $\exists A>0$, such that $\forall t\in\mathbb{R}_+$,
\begin{equation*}
\|{\bf u}(t)\|_{H_{\gamma+\omega}}+\|{\bf S}(t)\|_{D_{\gamma}}\leq A\left(\|{\bf u}_0\|_{H_{\gamma+\omega}}+\|{\bf S}_0\|_{D_{\gamma}} \right)
\end{equation*}
\end{theorem}
The statement (a) of Corollary \ref{soc} says that, for any $0\leq \gamma \leq \theta \leq \gamma+\omega $, ${\bf u}_0\in H_\theta$, ${\bf S}_0\in D_\gamma$, and ${\bf S}\in \mathscr{C}^0(\mathbb{R}^\ast_+,H_\gamma)$. This does not ensure continuity at $t=0$. Nevertheless, for $0\leq \gamma \leq \theta \leq \gamma+\omega $, $\theta\geq\omega$ and still holding on the assumptions ${\bf u}_0\in H_\theta$ and ${\bf S}_0\in D_\gamma$, we have ${\bf S}_0\in D_{\theta-\omega}$. Therefore (see Theorem \ref{sot2}) ${\bf S}\in \mathscr{C}^0(\mathbb{R}_+,D_{\theta-\omega})$, and ${\bf u}\in \mathscr{C}^0(\mathbb{R}_+,D_{\theta})$.
From now on we shall focus on the case $0 \leq \gamma \leq \theta \leq \omega < 1$. Proceeding as previously we get:
\begin{lemma}\label{sxl5}
Assume that $0\leq \gamma \leq \theta \leq \omega \leq 1$, ${\bf u}_0\in H_\theta$ and ${\bf S}_0\in D_\gamma$. Then $\exists M>0$, such that for any $t\geq0$,
\begin{equation}\label{sxl51}
\lambda_k^\theta \left| \alpha_k(t) \right|^2 \leq M \left(\lambda_k^\theta \left| \alpha_k^0 \right|^2+\lambda_k^{\theta-\omega }\left| b_k \right|^2 \right)
\end{equation}
and
\begin{equation}\label{sxl52}
\lambda_k^{\theta-\omega}(\lambda_k|\rho\displaystyle \ast \alpha_k|^2(t))\leq M \left[ \lambda_k^{\theta}|\alpha^0_k|^2+\lambda_k^{\theta-\omega}|b_k|^2 \right] )
\end{equation}
\end{lemma}
\begin{proof}
Eq.\eqref{sxl51} follows from part (i) in Proposition \ref{sot1}. Next, we use part (iv) in Lemma \ref{sot1} with $\mu=1-\omega$ and $\tau=1$. It gives $\mu-1=-\omega$ and $\mu(2-\delta)+2(\delta-1)=[1-\delta/(2-\delta)](2-\delta)+2\delta-2=0$. Henceforth,
\begin{equation*}
\lambda_k (|\rho\displaystyle \ast \alpha_k|^2(t))\leq M \left(\lambda_k^\omega|\alpha^0_k|^2+ |b_k|^2 \right)
\end{equation*}
which ends the proof.
\end{proof}
Hence:
\begin{corollary}\label{soc5}
Assume $0\leq \gamma \leq \theta $, ${\bf u}_0\in H_\theta$ and ${\bf S}_0\in D_\gamma$.
\begin{enumerate}[(a)]
\item \label{a} whenever $\theta\geq\omega$, ${\bf u}\in\mathscr{C}^0(\mathbb{R}_+,H_\theta)$, and ${\bf S}\in\mathscr{C}^0(\mathbb{R}_+,D_{\theta-\omega})$. Moreover, $\exists A>0$ s.t. $\forall t\geq0$:
\begin{equation}\label{syc51}
\left\|{\bf u}(t)\right\|_{H_\theta}+\left\|{\bf S}(t)\right\|_{D_{\theta-\omega}} \leq A\left(\left\|{\bf u}_0\right\|_{H_\theta}+\left\|{\bf S}_0\right\|_{D_{\theta-\omega}} \right)
\end{equation}
\item \label{b} whenever $\theta \leq \omega $, ${\bf u}\in\mathscr{C}^0(\mathbb{R}_+,H_\theta)$, and ${\bf S}\in\mathscr{C}^0(\mathbb{R}_+,\Delta_{\theta-\omega})$. Moreover, $\exists A>0$ s.t. $\forall t\geq0$,
\begin{equation}\label{syc52}
\left\|{\bf u}(t)\right\|_{H_\theta}+\left\|{\bf S}(t)\right\|_{\Delta_{\theta-\omega}} \leq A\left(\left\|{\bf u}_0\right\|_{H_\theta}+\left\|{\bf S}_0\right\|_{\Delta_{\theta-\omega}} \right)
\end{equation}
\end{enumerate}
\end{corollary}
\begin{proof}
(a)\\
The proof is a direct consequence of the discussion preceding Lemma \ref{sxl5}.\\
(b)\\
That ${\bf u}\in\mathscr{C}^0(\mathbb{R}_+,H_\theta)$ is a consequence of part (a) of Corollary \ref{soc}. Next, Lemma \ref{sxl5} and part (e) of Lemma \ref{sxl4} imply that, for any $M\leq N$,
\begin{eqnarray}\label{syc53}
\left\| \displaystyle \sum_{k=M}^{N} \left( \rho\displaystyle \ast \alpha_k \right)(t)\left(\nabla{\bf w}_k+\nabla^T{\bf w}_k \right) \right\|^2_{\Delta_{\theta-\omega}} & = & 2 \displaystyle \sum_{k=M}^{N} \lambda_k^{\theta-\omega} \left( \lambda_k |\rho\displaystyle \mathop{\ast}_{(t)}\alpha_k|^2 \right) \nonumber\\
& \leq & M \displaystyle \sum_{k=M}^{N} \left( \lambda_k^\theta|\alpha^0_k|^2+\lambda_k^{\theta-\omega} |b_k|^2 \right)
\end{eqnarray}
Since ${\bf u}_0\in H_\theta$, ${\bf S}_0\in D_\gamma \hookrightarrow L^2(\Omega)^9 \hookrightarrow \Delta_{\theta-\omega}$, we get by Eq.\eqref{syc53} and Eq.\eqref{t14 3} that
\begin{equation}\label{syc54}
\left( {\bf S}-W_0 \otimes {\bf S}_0 \right) \in \mathscr{C}^0(\mathbb{R}_+,\Delta_{\theta-\omega})
\end{equation}
Moreover, $W_0\in \mathscr{C}^0(\mathbb{R}_+)$ and ${\bf S}_0 \in D_\gamma \hookrightarrow \Delta_{\theta-\omega}$. Therefore, by Eq.\eqref{syc54}, ${\bf S}\in\mathscr{C}^0(\mathbb{R}_+,\Delta_{\theta-\omega})$. Inequality Eq.\eqref{syc52} follows right away after invoking Eq.\ref{sxl51}, Eq.\eqref{t14 3}, Eq.\eqref{syc53} with $M=1$ and $N=+\infty$, and that $W_0\in L^\infty(\mathbb{R}_+)$ and ${\bf S}_0\in \Delta_{\theta-\omega}$.
\end{proof}
\begin{remark}\label{rksc}
Using part (b) in Corollary \ref{soc5} and by a density argument one may prove that, for ${\bf u}_0\in H$ and ${\bf S}_0\in \Delta_{-\omega}$, the system of equations \eqref{bvp1} has a weak solution $({\bf u},{\bf S})\in \mathscr{C}^0(\mathbb{R}_+,H)\times\mathscr{C}^0(\mathbb{R}_+,\Delta_{-\omega})$. Of course the integrals have to be replaced by inner product functionals.
\end{remark}
\begin{corollary}\label{soc6}
Assume $0\leq \gamma \leq \theta \leq \omega \leq 1$, ${\bf u}_0\in H_\theta$ and ${\bf S}_0\in D_\gamma$. Then ${\bf u}\in\mathscr{C}^0(\mathbb{R}_+,H_\theta)$, and ${\bf S}\in\mathscr{C}^0(\mathbb{R}_+,H^{-1}(\Omega)^9)$. Moreover ${\bf u}(0)={\bf u}_0$, ${\bf S}(0)={\bf S}_0$.
\end{corollary}
\begin{proof}
Corollary \ref{soc5} states that ${\bf S}\in\mathscr{C}^0(\mathbb{R}_+,\Delta_{\theta-\omega})$. Therefore (see Section \ref{so}), the mapping of $\mathbb{R}_+$ into $H^{-1}(\Omega)^9$ defined by $\mathbb{R}_+ \displaystyle\stackrel{{\bf S}}{\rightarrow}\Delta_{\theta-\omega} \displaystyle\stackrel{i}{\hookrightarrow}\Delta_{-1} \displaystyle\stackrel{r}{\rightarrow}H^{-1}(\Omega)^9$, is continuous; the Corollary statement follows right away.
\end{proof}
Hence, to the first existence and uniqueness theorem, we can add the following conclusion: ${\bf S}\in\mathscr{C}^0(\mathbb{R}_+,\Delta_{-\omega})\hookrightarrow \mathscr{C}^0(\mathbb{R}_+,H^{-1}(\Omega)^9)$ and ${\bf S}(0)={\bf S}_0$.
We now give a second existence and uniqueness Theorem in $H_{\gamma+\omega}\times D_\gamma$ spaces.
\begin{theorem}[\textbf{Second Existence and Uniqueness Theorem}]\label{ut2}
Let $\gamma\geq0$. Assume that ${\bf u}_0 \in H_{\gamma+\omega}$, ${\bf S}_0 \in D_{\gamma }$. Then the boundary value problem given by the system of equations \eqref{fv1}-\eqref{fv2} has a unique solution
\begin{equation}\label{ut21}
({\bf u},{\bf S}) \in \left[ \mathscr{C}^0 (\mathbb{R}_+,H_{\gamma+\omega}) \cap \mathscr{C}^0_b (\mathbb{R}_+^\ast,V)\cap L^1_{\text{loc}}(\mathbb{R}_+ ,V) \right] \times \left[ \mathscr{C}^0 (\mathbb{R}_+,D_\gamma) \cap L^\infty(\mathbb{R}_+,D_\gamma) \right]
\end{equation}
Moreover, there exists $A>0$, independent of ${\bf u}$, such that, for any $t\geq0$,
\begin{equation}\label{ut22}
\|{\bf u}(t)\|_{H_{\gamma+\omega}}+\|{\bf S}(t)\|_{D_{\gamma}} \leq A \left( \|{\bf u}_0\|_{H_{\gamma+\omega}}+\|{\bf S}_0\|_{D_{\gamma}} \right)
\end{equation}
Lastly, ${\bf u}(0)={\bf u}_0$, ${\bf S}(0)={\bf S}_0$.
\end{theorem}
\begin{proof}
The solution uniqueness is a consequence of the following inclusions: $H_{\gamma+\omega} \hookrightarrow H$, $D_{\gamma} \hookrightarrow L^2(\Omega)^9$, $\left[ \mathscr{C}^0 (\mathbb{R}_+,H_{\gamma+\omega}) \cap \mathscr{C}^0_b (\mathbb{R}_+^\ast,V)\cap L^1_{\text{loc}}(\mathbb{R}_+ ,V) \right] \times \left[ \mathscr{C}^0 (\mathbb{R}_+,D_\gamma) \cap L^\infty(\mathbb{R}_+,D_\gamma) \right] \subset \mathscr{F}$, and of Corollary \ref{uc1}.
The existence of a solution $({\bf u},{\bf S}) \in \left[ \mathscr{C}^0 (\mathbb{R}_+,H_{\gamma+\omega}) \cap \mathscr{C}^0 (\mathbb{R}_+^\ast,V)\cap L^1_{\text{loc}}(\mathbb{R}_+ ,V) \right] \times \left[ \mathscr{C}^0 (\mathbb{R}_+,D_\gamma) \right]$ follows from Theorem \ref{sot2}. In addition, the last estimate in Theorem \ref{sot2} grants that ${\bf S}\in L^\infty(\mathbb{R}_+,D_\gamma)$. Next, that $\|{\bf u}(t)\|_V\displaystyle \mathop{\longrightarrow}_{t\to+\infty} 0$ was proved in the first existence and uniqueness Theorem presented above. Based on this fact, we infer that ${\bf u} \in \mathscr{C}^0_b (\mathbb{R}_+^\ast,V)$, which ends the proof of solution existence.
The estimate Eq.\eqref{ut22} is a consequence of Theorem \ref{sot2}.
\end{proof}
\section{The smoothness of solutions.}\label{ie}
The following estimates will be used in proving the smoothness of solutions. They generalize those previously obtained in \cite{pal7}.
\begin{proposition}\label{ie1}
$\exists M>0$, s.t. $\forall (\tau,\chi)\in[0,1]^2$, $\forall(x,t)\in\mathbb{R}_+$, $\forall\mu\geq\lambda_1$, one has:
\begin{enumerate}[(a)]
\item \label{a} $\left| \displaystyle\lim_{A\to+\infty} \int_{-A}^{+A}T_\mu (x+iy)e^{(x+iy)t}\mathrm{d} y \right|\leq \dfrac{M}{\mu^{\chi/(2-\delta)}t^\chi}$
\item \label{b} $\sqrt{\mu}\left| \displaystyle\lim_{A\to+\infty} \int_{-A}^{+A}T_\mu w(x+iy)e^{(x+iy)t}\mathrm{d} y \right|\leq \dfrac{M}{\mu^{[\omega(1-\tau)+\tau]/2}t^{\tau(1-\delta)}}$
\end{enumerate}
\end{proposition}
\begin{proof}
We only have to prove these estimates for $\chi=0$ and $\chi=1$, $\tau=0$ and $\tau=1$. \\
(a)\\
The case $\chi=0$ has already been addressed in Lemma 7.4 and Lemma 7.2 in \cite{pal7}.
We now prove the case $\chi=1$. Lemma 7.4 and inequality 70 in \cite{pal7} give, for suitable $\kappa>0$ and $B>0$,
\begin{eqnarray}\label{ie01}
\left| \displaystyle\lim_{A\to+\infty} \int_{-A}^{+A}T_\mu (x+iy)e^{(x+iy)t}\mathrm{d} y \right| & \leq & K \left[\displaystyle\int_{0}^{+\infty}\left|T_\mu(ze^{i\pi})-T_\mu(ze^{-i\pi})\right|e^{-zt}\mathrm{d} z+e^{-\kappa t \mu^{1/(2-\delta)}} \right] \nonumber \\
& \leq & K \left[ \int_{0}^{+\infty} \dfrac{u^{1-\delta}+\lambda_1^{-(1-\delta)/(2-\delta)}}{(u^{2-\delta}+B)^2}e^{-u t \mu^{1/(2-\delta)}}\mathrm{d} u + e^{-\kappa t \mu^{1/(2-\delta)}} \right]
\end{eqnarray}
It implies that:
\begin{equation}\label{ie12}
\left| \displaystyle\lim_{A\to+\infty} \int_{-A}^{+A}T_\mu (x+iy)e^{(x+iy)t}\mathrm{d} y \right| \leq K \left[ \int_{0}^{+\infty} e^{-u t \mu^{1/(2-\delta)}}\mathrm{d} u + \dfrac{1}{t \mu^{1/(2-\delta)}}\right] \leq \dfrac{A}{t \mu^{1/(2-\delta)}}
\end{equation}
which gives the statement in (a) for $\chi=1$. \\
(b)\\
The case $\tau=0$ is addressed in Lemma \ref{sol5}. The case $\tau=1$: from (iii) in Lemma 7.5 in \cite{pal7} we get
\begin{equation}\label{ie13}
\sqrt{\mu}\left| \displaystyle\lim_{A\to+\infty} \displaystyle\int_{-A}^{+A}T_\mu w (x+iy) e^{(x+iy)t}\mathrm{d} y \right|\leq \dfrac{K}{\mu t^{1-\delta}}+\dfrac{K e^{-at \mu^{1/(2-\delta)}} }{\mu^{1/2-\delta}}\leq \dfrac{M}{\mu t^{1-\delta}}
\end{equation}
as $\displaystyle\mathop{\sup}_{t \mu^{1/(2-\delta)}\geq0}\left|\left(t \mu^{1/(2-\delta)}\right)^{1-\delta}e^{-a t \mu^{1/(2-\delta)}} \right|\leq M<+\infty$.
\end{proof}
As a consequence we have the following extensions of estimates (iii) and (iv) of Theorem \ref{sot1}.
\begin{corollary}\label{ic1}
Let ${\bf u}_0\in H$, ${\bf S}_0\in L^2(\Omega)^9$. Then:
\begin{enumerate}[(a)]
\item \label{a} $\exists M >0$, $\forall T\geq0$, , s.t. $\forall \gamma \in [0,\dfrac{1}{2-\delta}]$, $\forall \tau \in [0,1]$, $\forall t \in [0,T]$ and $\forall k\in \mathbb{N}^\ast$,
$$|\alpha_k(t)|^2\leq M \left( \dfrac{|\alpha_k^0|^2}{\lambda_k^{2\gamma}t^{2\gamma(2-\delta)}} + \dfrac{|b_k|^2}{\lambda_k^{\tau+\omega(1-\tau)}t^{2\tau(1-\delta)}} \right)$$
\item \label{b} $\exists M >0$, $\forall T\geq0$, s.t. $\forall \gamma \in [0,1]$, $\forall \tau \in [0,\dfrac{1}{2-\delta}]$, $\forall t \in [0,T]$ and $\forall k\in \mathbb{N}^\ast$,
$$\lambda_k \left| \rho\displaystyle \ast \alpha_k \right|^2(t) \leq M
\left( \dfrac{|\alpha_k^0|^2}{\lambda_k^{2\gamma-1}t^{2[\gamma(2-\delta)+\delta-1]}} + \dfrac{|b_k|^2}{\lambda_k^{-(1-\tau)(1-\omega)}t^{-2(1-\tau)(1-\delta)}} \right)$$
\end{enumerate}
\end{corollary}
\begin{proof}
(a)\\
The statement in (a) follows from \eqref{per3 2} and Proposition \ref{ie1} (with $\gamma=\dfrac{\chi}{2-\delta}$). \\
(b) \\
The statement follows from (a) above by convolution.\\
\end{proof}
We deduce from Corollary \ref{ic1}:
\begin{proposition}\label{o1}
Let $\gamma\geq0$, $\eta>0$. Then:
\begin{enumerate}[(a)]
\item \label{a} Assume ${\bf u}_0\in H_{1+\gamma+\omega}$ and ${\bf S}_0\in D_{1+\gamma}$. Then ${\bf u}\in \mathscr{C}^1(\mathbb{R}_+,H_{\gamma})$.
\item \label{b} Assume ${\bf u}_0\in H_{3+\gamma-\omega+\eta}$ and ${\bf S}_0\in D_{3+\gamma }$. Then ${\bf u}\in \mathscr{C}^2(\mathbb{R}_+^\ast,H_{\gamma})\cap W_{\text{loc}}^{2,p}(\mathbb{R}_+,H_{\gamma})$, $p\in[1,1/(1-\alpha)[$.
\end{enumerate}
\end{proposition}
\begin{proof}
(a)\\
Based on eq.\eqref{al:a1} and $W_0\in L^{\infty}(\mathbb{R}_+)$, we infer that:
\begin{equation}\label{o1 1}
\lambda_k^{\gamma}|\alpha_k'(t)|^2 \leq \lambda_k^{1+\gamma} \left( \lambda_k \left| \rho\displaystyle \ast \alpha_k\right|^2(t) \right)+\lambda_k^{1+\gamma}|b_k|^2 \|W_0\|_\infty^2
\end{equation}
Now, Lemma \ref{sxl3} with $\theta=\gamma+\omega$, leads to
\begin{equation}\label{o1 2}
\lambda_k^{1+\gamma}\left( \lambda_k \left| \rho\displaystyle \ast \alpha_k\right|^2(t) \right) \leq A\lambda_k \left( \lambda_k^{\gamma+\omega}|\alpha_k^0|^2+\lambda_k^{\gamma}|b_k|^2 \right)
\end{equation}
From Eqs.\eqref{o1 1}-\eqref{o1 2},
\begin{equation}\label{o1 3}
\lambda_k^{\gamma}|\alpha_k'(t)|^2 \leq A \left( \lambda_k^{1+\gamma+\omega}|\alpha_k^0|^2+\lambda_k^{1+\gamma}|b_k|^2 \right)
\end{equation}
Since ${\bf u}_0\in H_{1+\gamma+\omega}$ and ${\bf S}_0\in D_{1+\gamma}$, $\lambda_k^\gamma |\alpha_k'(t)|^2<+\infty$. Hence $\displaystyle\sum_{k=1}^{+\infty} \alpha_k'\otimes {\bf w}_k$ converges in $\mathscr{C}^0(\mathbb{R}_+,H_{ \gamma})$. Since by Lemma \ref{sxl3} $\,\displaystyle\sum_{k=1}^{+\infty} \alpha_k\otimes {\bf w}_k$ converges in $\mathscr{C}^0(\mathbb{R}_+,H_{1+ \gamma})$, it also converges in $ \mathscr{C}^0(\mathbb{R}_+,H_{\gamma})$. Finally ${\bf u}\in \mathscr{C}^1(\mathbb{R}_+,H_{\gamma})$. \\
(b)\\
Observe that $1+\gamma+\omega \leq 3+\gamma-\omega+\eta$ and $1+\gamma \leq 3+\gamma $. Consequently ${\bf u}_0\in H_{1+\gamma+\omega}$, ${\bf S}_0\in H_{1+\gamma}$. Next, (a) above ensures that ${\bf u}\in \mathscr{C}^1(\mathbb{R}_+,H_{\gamma})$. We now deduce several estimates for the second order derivatives. From eq.\eqref{al:a1} it follows that, for $t>0$, $\alpha_k''(t)=-\lambda_k \left( \rho \displaystyle \ast \alpha_k' \right)(t) -\sqrt{\lambda_k}b_k W_0'(t)=\lambda_k^2 \left(\rho \displaystyle \ast \rho\displaystyle \ast \alpha_k \right)(t)+\lambda_k^{3/2}b_k \left(\rho \displaystyle \ast W_0 \right)(t)-\sqrt{\lambda_k}b_k W_0'(t) \in \mathscr{C}^0(\mathbb{R}_+^\ast)$ since $\alpha_k \in \mathscr{C}^0(\mathbb{R}_+)$, $ W_0 \in \mathscr{C}^0(\mathbb{R}_+) \cap \mathscr{C}^1(\mathbb{R}_+^\ast)$ and $0\leq \rho(t)\leq kt^{-\delta}$.
Let now $\epsilon>0$ be small enough. Part (a) in Corollary \ref{ic1} with $\gamma=(1-\epsilon)/(2-\delta)\in\left[0,\dfrac{1}{2-\delta}\right]$, $\tau=1 $ and $|W'_0(t)|\leq \dfrac{k}{t^{1-\alpha}}$ (see \cite{pal7}), gives
\begin{eqnarray}\label{xot1}
\lambda_k^{\gamma/2} |\alpha_k''(t)| & \leq & M_T \lambda_k^{2+\gamma/2} \left[ t^{1-2\delta} \displaystyle \ast \left( \dfrac{|\alpha_k^0|}{\lambda_k^{(1-\epsilon)/(2-\delta)}t^{1-\epsilon}}+\dfrac{|b_k|}{\lambda_k^{1/2 }t^{1-\delta}} \right) \right] \nonumber\\
& + & M_T \lambda_k^{(1+\gamma)/2}\left( \dfrac{\lambda_k }{t^{\delta-1}}+\dfrac{ 1 }{t^{1-\alpha}} \right)|b_k| \nonumber\\
& \leq & M_T \left( \dfrac{\lambda_k^{(3+\gamma)/2}}{t^{\delta-1}}+\dfrac{\lambda_k^{(1+\gamma)/2}}{t^{1-\alpha}} \right) |b_k| \nonumber\\
& + & M_T\dfrac{\lambda_k^{2-(1-\epsilon)/(2-\delta)+\gamma/2}}{t^{2\delta-\epsilon-1}}|\alpha_k^0|
\end{eqnarray}
Observe that, for $\epsilon>0$ small enough, $2(3+\gamma)/2 \leq 3+\gamma $ and $2(\gamma+1)/2 \leq 3+\gamma $. Also, for $\epsilon>0$ small enough, $2\left[2+\gamma/2-(1-\epsilon)/(2-\delta) \right] \leq 4+\gamma-2/(2-\delta)+\eta$.
Hence, by \eqref{xot1}, and since ${\bf u}_0 \in H_{3+\gamma-\omega+\eta}=H_{4+\gamma-2/( 2-\delta )+\eta}$ and ${\bf S}_0 \in D_{3+\gamma}$, we get ${\bf u} \in \mathscr{C}^2(\mathbb{R}^\ast_+,H_\gamma)$.
\end{proof}
Proceeding as before (see Proposition \ref{o1}) one obtains the following smoothness properties:
\begin{proposition}\label{ot1}
Let $\gamma\geq0$, $\eta>0$.
\begin{enumerate}[(a)]
\item \label{a} Assume ${\bf u}_0\in H_{2+\gamma-\omega+\eta}$, ${\bf S}_0\in D_{2+\gamma}$. Then ${\bf S}\in \mathscr{C}^1(\mathbb{R}_+^\ast,D_\gamma) \cap W_{\text{loc}}^{1,p}(\mathbb{R}_+,D_\gamma)$, for any $p\in [1,1/(1-\alpha)[$.
\item \label{b} Assume ${\bf u}_0\in H_{4+\gamma-\omega+\eta}$, ${\bf S}_0\in D_{4+\gamma}$. Then ${\bf S}\in \mathscr{C}^2(\mathbb{R}_+^\ast,D_\gamma) $.
\end{enumerate}
\end{proposition}
\begin{proof}
(a)\\
We limit the proof to the case $0<\eta<\omega$. Denote $\tilde{\omega}=\omega-\eta$; hence $0<\tilde{\omega}<\omega<1$.
Since ${\bf u}_0\in H_{2+\gamma-\tilde{\omega}}$, ${\bf S}_0\in D_{2+\gamma-\tilde{\omega}-\omega}$, based on Theorem \ref{sot2}, we have that ${\bf S} \in \mathscr{C}^0(\mathbb{R}_+,D_{2+\gamma-\tilde{\omega}-\omega})$. Next, as $0<\tilde{\omega}<\omega<1$ entails $\gamma<2+\gamma-\tilde{\omega}-\omega$, one gets ${\bf S} \in \mathscr{C}^0(\mathbb{R}_+,D_\gamma)$.
Next we obtain an estimate for $\displaystyle {\bf S}'(t)$. Observe first that $W_0\in L^\infty(\mathbb{R}_+)$ and that $\alpha_k'(t)=-\lambda_k \left( \rho\displaystyle \ast \alpha_k \right)(t)-\sqrt{\lambda_k}b_k W_0(t)$. Therefore:
$$\lambda_k^{\gamma/2}\sqrt{\lambda_k}\left| \rho\displaystyle \ast \alpha_k' \right|\leq A \lambda_k^{(\gamma+1)/2}\left[ \lambda_k \left( \rho\displaystyle \ast \rho\displaystyle \ast |\alpha_k| \right)+\sqrt{\lambda_k} \left(\rho\displaystyle \ast |b_k| \right) \right](t) $$
Now, part (a) of Corollary \ref{ic1}, with $\tilde{\delta}=2\tilde{\omega}/(1+\tilde{\omega})<2\omega/(1+\omega)=\delta$, $\gamma=1/(2-\tilde{\delta})$, $\tilde{\delta}\in ]0,\delta[$, and $\tau=1$, gives
$$|\alpha_k(t)| \leq M_T \left( \dfrac{|\alpha_k^0|}{\lambda_k^{1/(2-\tilde{\delta})}t^{(2-\delta)/(2-\tilde{\delta})}}+\dfrac{|b_k|}{\sqrt{\lambda_k}t^{1-\delta}} \right)$$
However, $\left| \rho\displaystyle \ast \rho\right|(t) \leq K t^{1-2\delta}$. Also, since $0< \tilde{\delta}<\delta$, then $(2-\delta)/(2-\tilde{\delta})\in[0,1[$. One infers that
\begin{eqnarray}\label{ot11}
\lambda_k^{\gamma/2}\sqrt{\lambda_k}\left| \rho\displaystyle \ast \alpha_k' \right|(t) & \leq & M_T \lambda_k^{1+\gamma/2} \bigg[ \sqrt{\lambda_k} \left( \dfrac{1}{t^{2\delta-1}}\displaystyle \ast \dfrac{|\alpha_k^0|}{\lambda_k^{1/(2-\tilde{\delta})}t^{(2-\delta)/(2-\tilde{\delta})}} \right) \nonumber\\
& + & \sqrt{\lambda_k} \left( \dfrac{1}{t^{2\delta-1}}\displaystyle \ast \dfrac{|b_k|}{\sqrt{\lambda_k}t^{1-\delta}} \right) + \dfrac{|b_k|}{t^{\delta-1}} \nonumber\\
& \leq & M_T \lambda_k^{1+\gamma/2} \left( \lambda_k^{-\tilde{\omega}/2}t^{2-2\delta-(2-\delta)/(2-\tilde{\delta})}|\alpha_k^0|+t^{1-\delta}|b_k| \right)
\end{eqnarray}
Let $a=2-2\delta-(2-\delta)/(2-\tilde{\delta})=1-2\delta+(\delta-\tilde{\delta})/(2-\tilde{\delta}) $, and $b=1-\delta>0$. Then:
\begin{equation}\label{69}
\left| \lambda_k^{\gamma/2} \sqrt{\lambda_k}\left( \rho\displaystyle \ast \alpha_k' \right) \right|^2 (t) \leq M_T \left[ \lambda_k^{2+\gamma-\tilde{\omega}}|\alpha_k^0|^2 t^{2a}+\lambda_k^{2+\gamma}t^{2b}|b_k|^2 \right]
\end{equation}
Recall that - as stated in (c) of Lemma \ref{sxl4} - that $\left( \nabla {\bf w}_k+ \nabla^T {\bf w}_k\right)_{k\in\mathbb{N}^\ast }$ is an orthogonal sequence of functions that belongs to $D_\gamma$, and $\left\| \nabla {\bf w}_k+ \nabla^T {\bf w}_k \right\|^2_{D_\gamma} = (1+\lambda_k^{\gamma})\lambda_k$. Consequently, using the estimate given above and that ${\bf u}_0\in H_{2+\gamma-\tilde{\omega}}$, ${\bf S}_0\in D_{2+\gamma}$, leads to the fact that $\displaystyle \sum_{k=1}^{+\infty}\left( \rho \displaystyle \ast \alpha_k' \right) \otimes \left( \nabla {\bf w}_k+ \nabla^T {\bf w}_k\right)$ converges in $\mathscr{C}^0(\mathbb{R}_+^\ast,D_\gamma)$. Next, from Eq in \cite{pal7}, $W_0'\otimes {\bf S}_0\in \mathscr{C}^0(\mathbb{R}_+^\ast,D_\gamma)$. Therefore $\displaystyle \sum_{k=1}^{+\infty}\left( \rho \displaystyle \ast \alpha_k' \right) \otimes \left( \nabla {\bf w}_k+ \nabla^T {\bf w}_k\right)+W_0'\otimes {\bf S}_0\in \mathscr{C}^0(\mathbb{R}_+^\ast,D_\gamma)$. Hence ${\bf S} \in \mathscr{C}^1(\mathbb{R}_+^\ast,D_\gamma)$.
Whenever $a\geq0$, by Eq.\eqref{69}, $\displaystyle \sum_{k=1}^{+\infty}\left( \rho \displaystyle \ast \alpha_k' \right) \otimes \left( \nabla {\bf w}_k+ \nabla^T {\bf w}_k\right)+W_0'\otimes {\bf S}_0$ belongs to $\mathscr{C}^0(\mathbb{R}_+ ,D_\gamma)$, thus belongs to $L^p_{\text{loc}}(\mathbb{R}_+ ,D_\gamma)$ for any $1 \leq p < +\infty$.
Now, whenever $a<0$, $-a-(1-\alpha)=2\delta-2-(\delta-\tilde{\delta})/(2-\delta)-1+\alpha=[-3-(\delta-\tilde{\delta})/(2-\delta)+(2\delta+\alpha)]<0$. We conclude that $\displaystyle \sum_{k=1}^{+\infty}\left( \rho \displaystyle \ast \alpha_k' \right) \otimes \left( \nabla {\bf w}_k+ \nabla^T {\bf w}_k\right)$ converges in $L^q_{\text{loc}}(\mathbb{R}_+ ,D_\gamma)$, for any $q\in [1,1/(1-\alpha)]$. Moreover, from \cite{pal7} we observe that $\| W_0'(t){\bf S}_0 \|_{D_\gamma}\leq K/t^{1-\alpha}$. It implies that $W_0'\otimes{\bf S}_0\in L^q_{\text{loc}}(\mathbb{R}_+,D_\gamma)$ for any $q\in [1,1/(1-\alpha)[$. Therefore
$$\displaystyle \sum_{k=1}^{+\infty}\left( \rho \displaystyle \ast \alpha_k' \right) \otimes \left( \nabla {\bf w}_k+ \nabla^T {\bf w}_k\right)+W_0'\otimes{\bf S}_0 \in L^q_{\text{loc}}(\mathbb{R}_+ ,D_\gamma)$$,
for any $q\in [1,1/(1-\alpha)[$ and irrespective of whether $a$ is positive or negative.
Eventually ${\bf S}\in W_{\text{loc}}^{1,q}(\mathbb{R}_+,D_\gamma)$ for any $q\in [1,1/(1-\alpha)[$.\\
(b)\\
The proof is omitted.\\
\end{proof}
From Proposition \ref{o1} we can infer the existence of smooth solutions to eqs.\eqref{bvp1}. Assume that ${\bf u}_0 \in H_{5+\omega}$ and ${\bf S}_0 \in D_{5}\cap \mathscr{C}^1(\overline{\Omega})^9$. Then, the solution $({\bf u},{\bf S})$ the existence of which is granted by Theorem \ref{ut1} of Section \ref{wf}, complies with the statement (a) of Proposition \ref{o1}, that is ${\bf u} \in \mathscr{C}^1(\mathbb{R}_+,H_{4 })$. Since $H_{4 } \hookrightarrow H^4(\Omega) \hookrightarrow \mathscr{C}^2(\overline{\Omega})$ (see Section \ref{so} and by Sobolev's injection), one has ${\bf u} \in \mathscr{C}^1(\mathbb{R}_+,\mathscr{C}^2(\overline{\Omega})^3 ) $ and $\left( \nabla {\bf u}+ \nabla^T {\bf u} \right) \in \mathscr{C}^1(\mathbb{R}_+,\mathscr{C}^1(\overline{\Omega})^9 ) $. One also has ${\bf S}_0\in \mathscr{C}^1(\overline{\Omega})^9 $, $W_0 \in \mathscr{C}^0(\mathbb{R}_+) \cap \mathscr{C}^1(\mathbb{R}_+^\ast)$ and $\rho \in L^1_{\text{loc}}(\mathbb{R}_+)$. Consequently ${\bf S}=\rho\ast\left( \nabla {\bf u}+ \nabla^T {\bf u} \right)+W_0 \otimes {\bf S}_0 \in \mathscr{C}^0(\mathbb{R}_+ ,\mathscr{C}^1(\overline{\Omega})^9 ) \cap \mathscr{C}^1(\mathbb{R}_+^\ast,\mathscr{C}^1(\overline{\Omega})^9 )$. All the precedent arguments eventually lead to the conclusion that $({\bf u},{\bf S}) \in \mathscr{C}^1( \mathbb{R}_+ ,\mathscr{C}^2(\overline{\Omega})^3 ) \times \left[ \mathscr{C}^0(\mathbb{R}_+ ,\mathscr{C}^1(\overline{\Omega})^9 ) \cap \mathscr{C}^1(\mathbb{R}_+^\ast ,\mathscr{C}^1(\overline{\Omega})^9 ) \right] $, whenever ${\bf u}_0 \in H_{5+\omega}$ and ${\bf S}_0 \in D_{5 } \cap \mathscr{C}^1(\overline{\Omega})^9$.
\section{Final comments}
Fractional calculus has a long history that parallels the classical analysis \cite{mr,py,skm}. It has long been used in modeling natural phenomena: for a quick glimpse see for example \cite{ade1,ade2,ade4,ams,agr,bal1,cm,jd1,jd2,dz, gds,han4,k1,ko2,lion,ne1,mak,ok,py,manero2,sered2,tpx,xly,yz2,yz}, and references cited therein. In particular, fractional derivative CEs have been found to accurately predict stress relaxation of viscoelastic fluids in the glass transition and glassy (high frequency) states.
The results presented here enrich and complement the linear stability analysis within the framework of variational/weak solutions initiated in \cite{pal7}. We have proved results regarding existence, uniqueness, smoothness and continuity at $t=0$ of the solution to the initial boundary value problem stated in Section \ref{intro}. Moreover, this work is related to that of Shaw, Whiteman and co-workers on the well posedness, existence and uniqueness of weak solutions for similar in nature hereditary - type integral models (see for example \cite{sw1},\cite{sw2},\cite{sw3},\cite{sw4}), as well as to that reported in \cite{pal5}, \cite{pal6},\cite{vtb}.
The matter of the stability of the original nonlinear CE is an open question on which future work shall focus.
\section{Aknowledgements}
The authors are grateful to Emeritus Associate Professor Michel Charnay, P\^ole de Math\'ematiques, INSA-Lyon, for kind support and encouragements.
\section{Bibliography}
|
\section{Introduction}
Since the pioneering works of Fisher \cite{F} and KPP \cite{KPP},
wave propagations for reaction-diffusion models in fluid mechanics,
physical, chemical and biological systems have been investigated
intensively. For example, such wave behaviors with front propagation
into unstable states can be frequently seen in fluid dynamical
experiments on Taylor-Couette\cite {AC,LMW} and Rayleigh-Benard
system \cite{FS}, and in chemical wave experiment \cite{SS,VRCY} as
well as in population dynamics, combustion and biological invasions
in a \textit{homogeneous} environment \cite{AW1,AW2}. The theory of
spreading speeds and monostable traveling waves has been developed
to monotone semiflows in such a way that it can be applied to
various evolution systems admitting the comparison principle (see
\cite {Wein1982,Lui,LiangZhao}). For the stability of monostable
waves including the minimal wave fronts in some regular and
time-delayed reaction-diffusion equations in spatially homogeneous
habitat, we refer to \cite{VVV,gourley,M1,M2,M3,MOZ} and the
references therein. The stability for bistable waves to the regular
or time-delayed reaction-diffusion equations has also been
intensively studied by many people, for example, see
\cite{Mac-F,Matano,Kapitula,Smith-Zhao,X,VVV} and the references
therein.
Recently, the study of traveling waves and spreading speeds in \textit
heterogeneous} media or periodically fragmented environments has been
drawing more attention, see, e.g., the survey paper \cite{X}. In a series of
works, Berestyski et al. \cite{B2005,B2008,B2010} and Hamel \cite{H2008}
established the existence of traveling fronts of the following equation
\begin{equation}
u_t-\nabla \cdot (A(x)\nabla u)=f(x,u),\;\;\;x\in \mathbb{R}^n,
\label{1}
\end{equation}
where the diffusion matrix $A(x)$ and the reaction term $f$ both are
periodic in $x=(x_1,x_2,...,x_n)$ in the sense that there exist constants
L_i>0,\;1\leq i\leq n$, such that
\[
A(x+L_i{\bf e}_i)=A(x), \quad f(x+L_i{\bf e}_i,u)=f(x,u), \quad
\forall 1\leq i\leq n,
\]
with ${\bf e}_i$ being the $i$-th vector in the standard base of
$\mathbb{R}^n$ (i.e., its $i$-th component is $1$ and all other
components are $0$). The matrix $A(x)$ is also assumed to be
symmetric
and lies in the space $C^{1,\alpha }(\mathbb{R}^n).$ The reaction term
f(x,u)$ is a continuous function of class $C^{0,\alpha }$ with respect to $x$
locally uniformly in $u\in \mathbb{R}$. The derivative $\frac{\partial f}
\partial u}$ is assumed to exist and be continuous in $u \in \mathbb{R}$ uniformly for $x$ in $\mathbb{R
^n.$ For instance, $f$ can be taken as $f(x,u)=u(\mu (x)-u), $ with
the positive periodic function $\mu (x)$ usually being the
carrying capacity of the periodic environment. Furthermore, for the matrix
A,$ it is usually assumed to be bounded, and uniformly elliptic in the sense
that
\begin{equation}
\sum_{1\leq i,j\leq n}a_{i,j}(x)\xi _i\xi _j\geq \alpha _0|\xi
|^2,\;\;\alpha _0>0 \label{1a}
\end{equation}
for any vector $\xi \in \mathbb{R}^n.$
In the special case where the period $L_i=0,$ namely, the matrix $A$
and the reaction term $f$ are independent of $x,$ model (\ref{1}) is
a generalization of the simple Fisher-KPP model
\begin{equation}
\left\{
\begin{array}{l}
u_t-\Delta u=f(u)\vspace{3.01pt} \\
f(0)=f(1)=0,\;\;f(u)>0,\,\forall u\in (0,1).
\end{array}
\right. \label{2}
\end{equation}
It is well known that for any $c\geq 2$, system (\ref{2}) with
$f(u)=u(1-u)$ has a planar traveling wave solution $u(t,x)=U(x\cdot
\mathbf{e}+ct)$ with $|\mathbf{e}|=1$, where $-\mathbf{e}$ is the
direction of propagation, while it has no such traveling wave for
any $c<2$.
The existence of traveling fronts to (\ref{1}) has been studied
recently in \cite{B2005,H2008}. For spreading speeds and traveling waves of
other types of biological evolution systems in a periodic habitat, we
further refer to \cite{Wein2002,JinZhao, WengZhao} and the references
therein. However, as mentioned in the survey paper \cite[page 185]{X}, the stability analysis
of traveling wavefronts in periodic media is a very challenging research
topic. In particular, the author of \cite{X} emphasized that the dynamics of
the slow waves moving at the minimal speeds requires ``a more delicate
argument and awaits further investigation''.
In this paper, we will concentrate on the stability and uniqueness
of the wavefront of (\ref{1}), an extension of the classical
KPP-Fisher model in heterogeneous media. Motivated by the work of
\cite{Moet} on the classical Fisher-KPP equation and our recent
study \cite{MOZ} on nonlocal time-delayed reaction-diffusion
equations, we further investigate the stability of wavefronts for
the periodic spatial-temporal reaction-diffusion equations by the
Green function method. A \emph{new and easy to follow} approach in
terms of a wave transformation is developed. The wavefront is proved
to be exponentially stable in the case where the wave-speed is
greater than the minimal one (also called the critical speed), and
algebraically stable in the case of the minimal speed. Furthermore,
we extend our results to the case of time-periodic media. In the
final concluding remark, we illustrate how the stability result can
be directly used to obtain the uniqueness of the wavefront with a
given speed.
After we finished our first draft of this paper, we happened to
obtain an important preprint \cite{H2} from Drs. Roques and Hamel's
personal websites, where
they investigated the stability and uniqueness of pulsating wavefronts for
\ref{1}) as well. It is seen that both papers are focusing on the
open problem of the stability of wavefronts. However, there exist
some major differences.
Firstly, our developed method is quite different from
that in \cite{H2}. Secondly, we prove that all non-critical
pulsating wavefronts are exponentially stable in the form of
$O(e^{-\mu t})$ for some $\mu
>0$, and the critical pulsating traveling wave is algebraically
stable in the form of $O(t^{-n/2})$, while there are no convergence
rates in the global stability results of \cite{H2}. Thirdly, we
assume that the product of the initial perturbation and the weight
function is in $L^1(R^n)$, while it is assumed in \cite{H2} that
initial functions behave like a front.
\section{Exponential and algebraical stability}
In this section, we are going to investigate the global stability of planar
traveling waves to the spatial periodic equation \eqref{1}. Let
w(x,x\cdot \mathbf{e}+ct)$ be a given planar traveling wavefront
with wave speed $c$ for some direction $\mathbf{e}$
($|\mathbf{e}|=1,$ $-\mathbf{e}$ is the direction of propagation)
such that $w(x+L_i{\bf e}_i,\xi)=w(x,\xi), \, \, \forall x\in
\mathbb{R}^n,\, \xi=x\cdot \mathbf{e}+ct \in \mathbb{R},\, 1\leq
i\leq n$. Without loss of generality, we can always assume that
$\mathbf{e}=\mathbf{e}_1=(1,0,...,0)$ after a rotation of
coordinates. Note that the existence and the asymptotic behavior of
$w(x,\xi ),$ $\xi =x_1+ct,$ can be obtained from \cite
{B2005,H2008}.
For equation (\ref{1}) with $f(x,0)=0$ to have wavefronts connecting
two steady states, the following sub-linearity of the function $f$
is also assumed in \cite{B2005,H2008}:
\begin{equation}
s\rightarrow \frac{f(x,s)}s\text{ is decreasing in }s>0\text{ for
any }x\in \mathbb{R}^n\mathbf{,} \label{2.1}
\end{equation}
and
\begin{equation}
f(x,s)\leq 0,\;s\geq M\text{ for some positive constant }M. \label{2.2}
\end{equation}
We have the following simple observation.
\begin{lem}
If condition (\ref{2.1}) holds, then we have
\begin{equation}\label{2.1a}
f_u(x,u)\leq f_u(x,0),\;\;\;\forall x\in \mathbb{R}^n,\, u\geq 0,
\end{equation}
where $f_u(x,u)$ is the partial derivative of $f$ with respect to
the variable $u$.
\end{lem}
\begin{proof}
Define
\[
\quad g(x,s):=\frac{f(x,s)}s,\, \, \forall x\in \mathbb{R}^n, \,
s>0.
\]
Then \eqref{2.1} implies that the function $g$ is decreasing in
$s>0$, and hence, $\frac{\partial g}{\partial s}\leq 0$. Since
$f(x,s)=sg(x,s)$, we have
\[
\frac{\partial f}{\partial s}=g(x,s)+s\frac{\partial g}{\partial
s}\leq g(x,s)\leq g(x,0):=\frac{\partial f(x,0)}{\partial s}.
\]
This completes the proof.
\end{proof}
The non-negative steady-states of (\ref{1}), involving in the
boundary condition of wavefront at $\pm \infty ,$ are the
$C^{2,\alpha }(\mathbb{R}^n)$-solutions $p(x)$ defined by
\[
\left\{
\begin{array}{l}
\ \nabla \cdot (A(x)\nabla p(x))+f(x,p(x))=0,\vspace{3pt} \\
p(x)\geq 0,\;\;p(x+L_i{\bf e}_i)=p(x), \, \forall 1\leq i\leq n.
\end{array}
\right.
\]
Obviously, the constant $0$ is a trivial steady-state and its local
stability is associated with an eigenvalue problem. Let $\mu _1$ be the
principal eigenvalue of the operator $L_0$ defined by
\[
L_0\phi =\ \nabla \cdot (A(x)\nabla \phi )+f_u(x,0)\phi ,
\]
that is,
\begin{equation}
\left\{
\begin{array}{l}
\ \nabla \cdot (A(x)\nabla \phi )+f_u(x,0)\phi (x)=\mu _1\phi (x),\vspace{3p
} \\
\phi (x)\geq 0,\;\;\phi(x+L_i{\bf e}_i)=\phi(x), \, \forall 1\leq
i\leq n.
\end{array}
\right. \label{2.3}
\end{equation}
If $\mu _1>0,$ then we say the zero solution is linearly unstable. In this
case, there exists a unique positive and $L$-periodic steady state solution
p(x)$ such that every positive solution $u(t,x)$ of (\ref{1}) with $L
-periodic initial function converges to $p(x)$ uniformly in $x\in
\mathbb{R}^n$ as $t\rightarrow \infty $, see \cite{B2005a}.
Now we can see that when $\mu _1>0,$ there are two non-negative
steady-states $p_1(x)=0$ and $p_2(x)=p(x),$ which has the same
mono-stability as in the Fisher-KPP equation. The heteroclinic connection
between $p_1(x)$ and $p_2(x)$ gives rise to a traveling wavefronts we are
seeking. In other words, by a traveling wave solution, we mean a particular
solution $u(t,x)=w(x,\xi ),\xi =x_1+ct$, which satisfies equation (\ref{1})
subject to the boundary condition
\begin{equation}
\lim_{s\rightarrow -\infty }w(x,s)=0,\;\lim_{s\rightarrow \infty
}w(x,\infty )=p(x),\text{ uniformly in }x\in \mathbb{R}^n.
\label{2.4}
\end{equation}
Substituting the relation $u(t,x)=w(x,\xi )$ with $\xi=x_1+ct$ into (\ref{1}), we obtain the
wave profile equation
\begin{equation}
\left\{
\begin{array}{l}
c\frac{\partial w}{\partial \xi }=\ \nabla \cdot (A(x)\nabla
w(x,x_1+ct))+f(x,w),\vspace{3pt} \\
w(x,-\infty )=0,\;w(x,\infty )=p(x),\;\;\;\text{uniformly in }x\in
\mathbb{R} ^n.
\end{array}
\right. \label{2.5}
\end{equation}
Now we consider the case where $\mu _1>0,$ that is, the steady state
$0$ is unstable. The following result on the existence of spatially
periodic traveling wavefronts is from \cite {B2005,H2008}.
\noindent \textbf{Theorem A} (\cite{B2005,H2008}). Assume that $\mu
_1>0$ and $f$ satisfies (\ref{2.1}) and (\ref{2.2}). Then for any
$c\geq c^{*}$, there
exists a positive traveling wave $u(t,x)=w(x,x_1+ct)=w(x,\xi )$ satisfying
\ref{1}) and (\ref{2.4}), namely \eqref{2.5}, where the minimal speed $c^{*}$
is determined by
\begin{equation}
c^{*}=\min \{c:\;\exists \text{ }\lambda >0,\text{ such that }\mu _c(\lambda
)=0\}. \label{2.5a}
\end{equation}
Here $\mu _c(\lambda )$ is the principal eigenvalue of the operator
\begin{eqnarray}
-L_{c,\lambda }\psi &=&-\nabla \cdot (A(x)\nabla \psi )-2\lambda \mathbf
e\cdot }\left( A(x)\nabla \psi \right) \label{2.6} \\
&&-\left[ \lambda \nabla \cdot (A(x)\mathbf{e}^T)+\lambda ^2\mathbf{e\cdot
\left( A(x)\mathbf{e}^T\right) -\lambda c+f_u(x,0)\right] \psi , \nonumber
\end{eqnarray}
that is,
\[
-L_{c,\lambda }\psi =\mu _c(\lambda )\psi \ \mbox{ with } \,
\psi(x+L_i{\bf e}_i)=\psi(x), \, \forall x\in \mathbb{R}^n,\, 1\leq
i\leq n.
\]
The derivation of $c^{*}$ in \cite{B2005,H2008} is based on the
technique developed by Aronson and Weinberger \cite{AW1,AW2} in
terms of a monotone-semiflow defined from equation (\ref{1}). Here
we can give an alternative, also easy to follow, explanation
directly from the wave profile equation (\ref{2.5}). This idea
originates from study of the KPP-Fisher equation and now is called
the technique of stability analysis of positive wavefronts near the
unstable steady state in their phase plane.
Due to biological or physical reason, we always expect a non-negative
wavefront, and as such, near the far-field $\xi =-\infty $ we assume $w\sim
v(x)e^{\lambda \xi }$ for some positive constant $\lambda $ and positive
periodic function $v(x),$ both of which are dependent on the wave speed $c.$
Substituting this relation into (\ref{2.5}) and ignoring the higher terms,
we then get
\begin{equation}
-\nabla \cdot \left( A(x)\nabla v\right) -2\lambda \mathbf{e\cdot }\left(
A(x)\nabla v\right) -\left[ \lambda \nabla \cdot (A(x)\mathbf{e}^T)+\lambda
^2\mathbf{e\cdot }\left( A(x)\mathbf{e}^T\right) -\lambda c+f_u(x,0)\right]
v=0, \label{v}
\end{equation}
which implies that the operator $-L_{c,\lambda }$ defined in
(\ref{2.6}) should have zero as its principal eigenvalue. The
minimal speed $c$ is the smallest of such $c$ that (\ref{v}) holds
for $\lambda >0$. This agrees with the definition in (\ref{2.5a}).
Furthermore, in terms of $\lambda ,$ the operator $-L_{c,\lambda }$
is a polynomial of degree 2 and the result in \cite{H2008} shows
that the minimal speed $c^{*}$ is well-defined and has the following
properties.
\noindent \textbf{Theorem B} (\cite{H2008}). \textit{When $c>c^{*},$
there exists two constants $\lambda _1(c),\lambda
_2(c),\;\;0<\lambda _1(c)<\lambda _2(c)$ such that $\mu _c(\lambda
_1(c))=\mu _c\left( \lambda _2(c)\right) =0,\;\mu _c(\lambda )>0$
for any $\lambda \in (\lambda _1(c),\lambda _2(c))$; while in the
critical case $c=c^{*}$, we have $\mu _{c^*}(\lambda )\leq 0$, and
$\mu _{c^*}(\lambda ^{*})=0$ when $\lambda ^{*}=\lambda
_1(c^{*})=\lambda _2(c^{*})$.}
To present the exponential stability of wavefronts, we also impose a natural
condition on the positive steady-state $p_2=p(x).$ We assume that it is
linearly stable in the sense that the principal eigenvalue $\bar{\mu}_1$ of
the operator $L_1,$ defined by
\[
L_1\phi =\nabla \cdot (A(x)\nabla \phi )+f_u(x,p(x))\phi ,
\]
is negative, that is,
\begin{equation}
\left\{
\begin{array}{l}
\nabla \cdot (A(x)\nabla \phi )+f_u(x,p(x))\phi (x)=\bar{\mu}_1\phi (x)
\vspace{3pt} \\
\phi (x)\geq 0,\;\;\phi(x+L_i{\bf e}_i)=\phi(x), \, \forall 1\leq
i\leq n
\end{array}
\right. \label{2.7}
\end{equation}
hold for some negative constant $\bar{\mu}_1$ and non-zero function
$\phi$.
The subsequent result shows that the condition $\bar{\mu}_1<0$ can
be easily realized.
\begin{lem}
Assume that
\begin{enumerate}
\item[(C)] $f_u(x,u)\leq \frac{f(x,u)}{u},\, \forall u>0, \, x\in \mathbb{R}^n$, and
there exists $x_0\in \mathbb{R}^n$ such that $f_u(x_0,u)<
\frac{f(x_0,u)}{u}, \, \forall u>0$.
\end{enumerate}
Then any possible positive $L$-periodic steady state $p(x)$ of
(\ref{1}) is linearly stable.
\end{lem}
\begin{proof}
Let $g(x,u):=f(x,u)/u,\, \forall u>0$, and let $\mu (m(x))$ be the
principal eigenvalue of the operator $L_1$ with $f_u(x,p(x))$
replaced by a continuous and $L$-periodic weight function $m(x)$.
Since $p(x)=p(x+L)$ satisfies the elliptic equation and only the
principal eigenvalue has positive eigenfunction, it is easy to see
that $\mu (g(x,p(x))=0$. Since $f_u(x,p(x))\leq g(x,p(x))$ and
$f_u(x_0,p(x_0))<g(x_0,p(x_0))$, the monotonicity of the principal
eigenvalue with respect to the weight function implies that $\mu
(f_u(x,p(x)))<\mu (g(x,p(x)))=0$. Thus, $p(x)$ is linearly stable.
\end{proof}
For any given wavefront $w(x,\xi)$, $\xi=x_1+ct$ with speed $c\geq
c^{*}$, we define a weight function
\begin{equation}
W(\xi )=\begin{cases} e^{-\lambda (\xi -\xi _0)}, & \xi \leq \xi _0, \ \ \
\lambda \in \left[ \lambda _1(c),\lambda _2(c)\right] ,\vspace{3pt}\; \\
1,&\xi > \xi _0, \end{cases} \label{W}
\end{equation}
where $\xi _0$ is chosen in such a way that the wavefront $w(x,\xi ),$ $\xi
\geq \xi _0,$ is very close to the positive steady-state $p_2=p(x)$, that
is, $\xi _0$ is chosen so that
\begin{equation}
|w(x,\xi )-p(x)|<\bar{\varepsilon} ,\;\;\forall \xi \geq \xi _0,
\label{2.11a}
\end{equation}
uniformly in $x\in \mathbb{R}^n$ for any given sufficient small
constant $\bar{\varepsilon}
>0.$
Now we are in a position to prove the following stability result.
\begin{thm}
\label{thm2.1} Assume that \eqref{2.1} and \eqref{2.2} hold, $\mu
_1>0$, and $\bar{\mu}_1<0$.
Let $u(t,x)$ be the solution (\ref{1}) with the initial condition
u_0(x)$ satisfying $0\leq u_0(x)\leq p(x)$, and $w(x,x_1+ct)$ be a
spatially periodic traveling wavefront with $c\geq c^{*}.$ If the
initial data $u_0(x),$ having at most finite points of
discontinuity, satisfies
\begin{equation}
W(x_1)(u_0(x)-w(x,x_1))\in L^1(\mathbb{R}^n\mathbf{),} \label{2.9}
\end{equation}
then the following statements are valid:
(1) For $\xi =x_1+ct\leq \xi _0,$ it holds
\begin{equation}
|u(t,x)-w(x,x_1+ct)|<C(1+t)^{-\frac n2}e^{-\mu _c(\lambda )t}; \label{2.10}
\end{equation}
(2) For $\xi =x_1+ct\geq \xi _0,$ it holds
\begin{equation}
|u(t,x)-w(x,x_1+ct)|<C(1+t)^{-\frac n2}e^{-\min \{\mu _c(\lambda ),\frac{
\bar{\mu}_1}2\}t}. \label{2.11}
\end{equation}
Here $\lambda $, as mentioned in the weight function, is any number in the
interval $\left[ \lambda _1(c),\lambda _2(c)\right] .$
\end{thm}
\begin{proof}
For the given initial function $u_0(x)$, we define
\begin{equation}
\left\{
\begin{array}{l}
U_0^{+}(x)=\max \{u_0(x),\;w(x,x_1)\},\;x\in \mathbb{R}^n\mathbf{,\vspace{3p
}} \\
U_0^{-}(x)=\min \{u_0(x),\;w(x,x_1)\},\;x\in \mathbb{R}^n,
\end{array}
\right. \label{2.14}
\end{equation}
where $w(x,x_1)$ is the initial function of the traveling wave
solution $w(x,x_1+ct)$.
Let $U^{\pm }(t,x)$ be the solutions of (\ref{1}) with the initial data
U_0^{\pm }(x),$ respectively, that is,
\begin{equation}
\left\{
\begin{array}{l}
\dfrac{\partial U^{\pm }}{\partial t}=\nabla \cdot \left( A(x)\nabla U^{\pm
}\right) +f(x,U^{\pm }),\vspace{4pt} \\
U^{\pm }(0,x)=U_0^{\pm }(x),\;x\in \mathbb{R}^n\mathbf{.}
\end{array}
\right. \label{2.14a}
\end{equation}
Since $0\leq U_0^{-}\leq w(x,x_1)\leq U_0^{+}\leq p(x),$ the comparison
principle implies that
\[
0\leq U^{-}(t,x)\leq w(x,x_1+ct)\leq U^{+}(t,x)\leq p(x),
\]
and
\[
0\leq U^{-}(t,x)\leq u(t,x)\leq U^{+}(t,x)\leq p(x).
\]
Obviously, we have
\[
U^{-}(t,x)-w(x,x_1+ct)\leq u(t,x)-w(x,x_1+ct)\leq U^{+}(t,x)-w(x,x_1+ct),
\]
or
\[
|u(t,x)-w(x,x_1+ct)|\leq \max \{|U^{+}(t,x)-w(x,x_1+ct)|,\
|U^{-}(t,x)-w(x,x_1+ct)|\}.
\]
As such, to prove (\ref{2.10}) and (\ref{2.11}), we proceed to show
\begin{equation}
|U^{\pm }(t,x)-w(x,x_1+ct)|\leq C(1+t)^{-\frac n2}e^{-\mu _c(\lambda )t},\
\text{ for }\ \xi \leq \xi _0 \label{2.15}
\end{equation}
and
\begin{equation}
|U^{\pm }(t,x)-w(x,x_1+ct)|\leq C(1+t)^{-\frac n2}e^{-\min \{\mu _c(\lambda
),\frac{-\bar{\mu}_1}2\}t},\;\;\text{for }\;\xi >\xi _0. \label{2.16}
\end{equation}
In what follows, we give a detailed proof of (\ref{2.15}) and (\ref{2.16})
for $U^{+}(t,x)$. The similar arguments work for $U^{-}(t,x)$, too, and hence, we
omit them.
Recall that both $U^{+}(t,x)$ and $w(x,x_1+ct)$ satisfy the first equation
of (\ref{2.14a}). Note that $\xi =x_1+ct.$ Thus we define
\[
V(t,x)=U^{+}(t,x)-w(x,x_1+ct).
\]
It is easy to see that $V(t,x)$ is a nonnegative solution of the
following equation
\[
\begin{cases}
V_t=\nabla \cdot \Big( A \nabla V \Big) +f(x,w+V)-f(x,w)
\vspace{4pt} \\
V(0,x )=U_0^{+}(x )-w(x ,x_1 )\geq 0,
\end{cases}
\]
with the initial data $V(0,x)$ satisfying
\begin{equation}
W(x_1)V(0,x)=W(x_1)\left( U_0^{+}(x)-w(x,x_1)\right) \in
L^1(\mathbb{R}^n), \label{2.17}
\end{equation}
where $W(x_1)$ is the weight function defined in (\ref{W}).
Since the non-linear function $f$ is sub-linear and satisfies (\ref{2.1a}),
it follows that
\begin{equation}
f(x,w+V)-f(x,w)\leq f_u (x,0)V. \label{Q?}
\end{equation}
We then have an inequality
\begin{equation}
\begin{cases} V_t\leq \nabla \cdot \Big( A \nabla V \Big) +f_u (x,0)V,\vspace{4pt} \\ V(0,x )=U_0^{+}(x )-w(x ,x_1 )\geq 0. \end{cases}
\label{2.18}
\end{equation}
Let $v(x),$ dependent on $c$ and $\lambda ,$ be the eigenfunction of
operator $-L_{c,\lambda }$ with the principal eigenvalue $\mu _c(\lambda ).$
From Theorem B, when $c\geq c^{*},$ $\mu _c(\lambda )$ is real and $v(x)>0$
is well-defined. By a crucial wave transformation
\begin{equation}
V(t,x)=v(x)e^{\lambda (x_1+ct -\xi _0)}\bar{V}(t,x)e^{-\mu
_c(\lambda )t},\;\;\lambda \in [\lambda _1(c),\lambda _2(c)],
\label{2.19}
\end{equation}
it then follows from (\ref{2.18}) that $\bar{V}(t,x)$ satisfies
\begin{equation}
\begin{cases} \bar{V}_t\leq \nabla \cdot \Big( A \nabla \bar{V} \Big)
+g(t,x) \cdot \nabla \bar{V},\vspace{4pt} \\
\bar{V}(0,x)=v^{-1}(x)e^{-\lambda (x_1 -\xi _0)}V(0,x)\geq 0,
\end{cases} \label{2.20}
\end{equation}
where
\begin{equation}
g=2A\left( \frac{\nabla v}v+\lambda \mathbf{e}\right) . \label{2.21}
\end{equation}
To get (\ref{2.20}), we have made use of the following formulas
\[
\nabla (uv)=u\nabla v+v\nabla u,\;\;\;\;\nabla \cdot (vA\nabla u)=v\nabla
\cdot (A\nabla u)+\nabla v\cdot (A\nabla u).
\]
It is easy to see that $\bar{V}$ in (\ref{2.20}) can be estimated by
\begin{equation}
\bar{V}(t,x)\leq \int_{\mathbb{R}^n}G(t,x-y)\bar{V}(0,y)dy,\quad
\forall x\in \mathbb{R}^n,\, t\geq 0, \label{2.22}
\end{equation}
where $G(t,x-y)$ is the fundamental solution (i.e., the Green function) so
that $\int G(t,x-y)u_0(y)dy$ satisfies the following partial differential
equation
\[
u_t=\nabla \cdot (A\nabla u)-g\cdot \nabla u,\;\;\;\;u(0,x)=u_0(x),
\]
with
\[
\int_{\mathbf{R}}G(t,y)dy=1.
\]
Furthermore, since $A$ is spatially periodic and uniformly elliptic,
it follows from \cite[(6.12) in page 24]{FR} that
\begin{equation}
|G(t,x-y)|\leq \frac C{t^{\frac n2}}e^{-\frac{|x-y|^2}{4\alpha _1t}}
\label{2.23}
\end{equation}
for some $\alpha _1>0$, where $n$ is the spatial dimension. By
(\ref{2.17}), we understand that $\bar{V}(0,x)\in
L^1(\mathbb{R}^n).$ From (\ref{2.22}), together with (\ref{2.23}),
we can obtain
\[
\bar{V}(t,x)\leq Ct^{-\frac
n2}\int_{\mathbb{R}^n}|\bar{V}(0,y)|dy,\quad \forall x\in
\mathbb{R}^n, \, t\geq 0.
\]
Since the solution $\bar{V}$ for $t$ near zero has no singularity, the term
t^{-\frac n2\text{ }}$ can be replaced by $(1+t)^{-n/2}$. This technique is
frequently used in the heat equation. Returning to (\ref{2.19}), we then
have
\begin{eqnarray}
V(t,x) &=&|U^{+}(t,x)-w(x,x_1+ct)| \nonumber \\
&=&v(x)e^{\lambda (x_1+ct -\xi _0)}\bar{V}(t,x)e^{-\mu _c(\lambda
)t} \nonumber
\\
&\leq &Ce^{\lambda (x_1+ct -\xi _0)}(1+t)^{-\frac{n}{2}}e^{-\mu
_c(\lambda )t} \label{vv}
\end{eqnarray}
for all $x\in \mathbb{R}^n, \, t\geq 0$. When $\xi=x_1+ct \leq \xi _0$, we have
e^{\lambda (\xi -\xi _0)}\leq 1$. This proves (\ref{2.10}).
Next for $\xi=x_1+ct >\xi _0,$ we know that
\[
w(x,\xi )\leq w(x,\xi)+V\leq p(x).
\]
Since we have chosen $\xi _0$ so that $w(x,\xi )$ is very close to
$p(x)$ uniformly in $x\in \mathbb{R}^n$ when $\xi >\xi _0$ (for
precision, see (\ref{2.11a})), as such for $\xi =x_1+ct >\xi _0 $,
by the regularity of $f$, we have
\[
f(x,w+V)-f(x,w)\leq (f_u(x,p(x))+\varepsilon )V
\]
and
\[
\begin{cases}
V_t\leq \nabla \cdot (A\nabla V)
+(f_u (x,p(x))+\varepsilon
)V,
\vspace{4pt} \\
V(0,x)=U_0^{+}(x )-w(x,x_1)\geq 0
\end{cases}
\]
for some small constant $\varepsilon $ satisfying $0<\varepsilon <-\bar{\mu
_1$, where $\bar{\mu}_1$ is defined in (\ref{2.7}) with
eigenfunction $\phi (x)>0$. Under the transformation
\begin{equation} \label{moz1}
V(t,x)=\bar{V}(t,x)\phi (x)e^{(\bar{\mu}_1+\varepsilon )t},
\end{equation}
it then follows that for all $\xi=x_1+ct >\xi _0$, $\bar{V}(t,x)$
satisfies
\begin{equation}
\bar{V}_t\leq \nabla \cdot (A\nabla \bar{V})+2\left( A\frac{\nabla \phi
\phi \right) \cdot \nabla \bar{V},\ \ \ \ \bar{V}(0,x)=\phi ^{-1}(x)V(0,x),
\label{2.24}
\end{equation}
with
\[
\bar{V}(t,x)\leq V(t,x)\phi ^{-1}(x)e^{-(\bar{\mu}_1+\varepsilon )t}
\]
on the line $x_1+ct=\xi_0$. Since on the boundary line $\xi
=x_1+ct=\xi _0,$ (\ref{vv}) yields
\begin{equation}\label{moz2}
V(t,x)\leq C(1+t)^{-\frac n2}e^{-\mu _c(\lambda )t},
\end{equation}
which implies that on the line $x_1+ct=\xi _0,$
\[
\bar{V}(t,x)\leq C(1+t)^{-\frac n2}e^{-\mu _c(\lambda )t}e^{\mu
_2t},
\]
where $\mu _2=-\bar{\mu}_1-\varepsilon >0$. Let
\[
\mu _{\min }=\min \{\mu _c,\frac{-\bar{\mu}_1}2\}.
\]
Thus, applying \eqref{moz2} to \eqref{moz1}, we obtain that
on the line $x_1+ct=\xi _0,$
\[
\bar{V}(t,x)\leq C(1+t)^{-\frac n2}e^{-\mu _{\min }t}e^{\mu _2t}.
\]
Obviously, $C(1+t)^{-\frac n2}e^{-\mu _{\min }t}e^{\mu _2t}$ is an upper
solution of the equation
\[
u_t=\nabla \cdot (A\nabla u)+(2A\frac{\nabla \phi }\phi )\cdot
\nabla u,\quad t\geq 0
\]
for some constant $C$. By the parabolic comparison principle, it
then follows that
\[
V(t,x) =\bar{V}(t,x)\phi (x)e^{(\bar{\mu}_1+\varepsilon )t} \leq
C(1+t)^{-\frac n2}e^{-\mu _{\min }t},\, \, \forall x\in
\mathbb{R}^n,\, t\geq 0.
\]
This completes the proof.
\end{proof}
The following result is a straightforward consequence of Theorem
\ref{thm2.1}.
\begin{thm}
\label{thm2.2} Assume that \eqref{2.1} and \eqref{2.2} hold,
$\mu_1>0$, and $\bar{\mu}_1<0$.
Let $u(t,x)$ be the solution (\ref{1}) with the initial condition
u_0(x),\;0\leq u_0(x)\leq p(x),$ and $w(x,x_1+ct)$ be a traveling
wavefront with $c\geq c^{*}.$ If the initial data $u_0(x),$ having
at most finite points of discontinuity, satisfies
\[
W(x_1)(u_0(x)-w(x,x_1))\in L^1(\mathbb{R}^n\mathbf{)},
\]
then the following statements are valid:
(1) For any $c>c^{*}$ and $\lambda $ $\in (\lambda _1(c),\lambda _2(c)),$
the wavefront $w(x,x+ct)$ is exponentially stable in the sense that
\begin{equation}
\sup_{x\in \mathbb{R}^n}|u(t,x)-w(x,x_1+ct)|\leq Ce^{-\mu
_3t},\;\;\quad \text{for some }\,\,\mu _3>0; \label{2.12}
\end{equation}
(2) For $c=c^{*}$, or $c>c^{*}$ with $\lambda =\lambda _1(c),$ the wavefront
is algebraically stable in the sense that
\begin{equation}
\sup_{x\in \mathbb{R}^n}|u(t,x)-w(x,x_1+ct)|\leq C(1+t)^{-\frac n2}.
\label{2.13}
\end{equation}
\end{thm}
\begin{proof}
For any given $c>c^{*}$ and $\lambda \in (\lambda _1(c),\lambda
_2(c)),$
it follows from Theorem B that $\mu _c(\lambda )>0.$ Let $\mu _3=\min \{$
\mu _c(\lambda ),\mu _2\}.$ Then we have (\ref{2.12}) from (\ref{2.10}) and
\ref{2.11}). When $c=c^{*},$ or $c>c^{*}$ with $\lambda =\lambda
_1(c),$ we then get $\mu _c(\lambda )=0,$ and a combination of
(\ref{2.10}) and (\ref {2.11}) yields (\ref{2.13}).
\end{proof}
\section{An extension to the time-periodic case}
We have already studied the system which is periodic in space. In this
section, we extend our results to the case when the system is also periodic
in time.
We consider the following parabolic equation
\begin{equation}
u_t-\nabla \cdot (A(t,x)\nabla u)=f(t,x,u),\;\;x\in \mathbb{R}^n,
\label{5.1}
\end{equation}
where the matrix $A$ and the nonlinear term $f$ are periodic both in $t$ and
$x,$ that is,
\[
A(t+T,x)=A(t,x),\;\;f(t+T,x,u)=f(t,x,u),\;
\]
and
\[
A(t,x+L_i{\bf e}_i)=A(t,x),\;f(t,x+L_i{\bf e}_i,u)=f(t,x,u),\, \,
\forall 1\leq i\leq n
\]
for some positive real number $T>0$ and $n$-dimensional vector
L=(L_1,L_2,...,L_n)$.
As in \cite[page 364-366]{H2008}, we assume that the uniformly
elliptic symmetric matrix $A(t,x)$ is of class $C_{t,x}^{1,\alpha
/2;1,\alpha }$ and the function $f(t,x,u)$ is continuous, of class
$C^{0,\alpha /2;0,\alpha }$ with respect to $(t,x)$ locally
uniformly in $u\in \mathbb{R}$. We also assume that $\frac{\partial
f}{\partial u}$ exists and is continuous in $u$ uniformly for
$x\in\mathbb{R}^n$ and $t \in \mathbb{R}$ . The existence of
traveling waves $w(t,x,x_1+ct)$ was studied in \cite[page
364-366]{H2008} and \cite{nadin} with the sub-linearity assumption
on the reaction
term $f.$ Assume that there are two periodic states $p_1(t,x)=0$ and
p_2(t,x)>0$ of (\ref{5.1}). The system is called mono-stable if the state
p_1(t,x)$ is unstable and the other state $p_2(t,x)$ is stable. In what
follows, we assume that $\mu _1>0$ and $\bar{\mu}_1<0,$ where $\mu _1$ is
the principal eigenvalue of the linearized operator around $0$:
\[
\varphi (t,x)\rightarrow \varphi _t-\nabla \cdot (A(t,x)\nabla \varphi
)-f_u(t,x,0)\varphi ,\;\;\varphi (t,x)\text{ is periodic in }t\text{ and }x,
\]
and $\bar{\mu}_1$ is the the principal eigenvalue of the linearized operator
around $p_2(t,x)$:
\[
\varphi (t,x)\rightarrow \varphi _t-\nabla \cdot (A(t,x)\nabla \varphi
)-f_u(t,x,p_2)\varphi ,\;\;\varphi (t,x)\text{ is periodic in }t\text{ and
x.
\]
Traveling wavefronts are solutions of (\ref{5.1}) in the form $u(t,x)=w(t,x
\mathbf{e}\cdot x+ct)=w(t,x,x_1+ct),$ where $\mathbf{e}=(1,0,...,0),$ which
is periodic in the first two variables $t$ and $x$ and satisfies
\[
\lim_{\xi \rightarrow -\infty }w(t,x,\xi )=p_1(t,x)=0,\;\;\lim_{\xi
\rightarrow \infty }w(t,x,\xi )=p_2(t,x)>0
\]
uniformly for all $t$ and $x.$ The minimal speed $c^{*}$ is defined by
\[
c^{*}=\min \{c:\;\exists \text{ }\lambda \text{ so that }\mu _c(\lambda
)=0\},
\]
where $\mu (\lambda )$ is the principal eigenvalue of the operator
\[
v(t,x)\rightarrow v_t+c\lambda v-\nabla \cdot (A(t,x)\nabla v)-2\lambda
\mathbf{e\cdot }\left( A\nabla v\right) -\lambda \nabla \cdot (A\mathbf{e
)-\lambda ^2\mathbf{e\cdot }\left( A\mathbf{e}\right) -f_u(t,x,0)v,\;
\]
that is, there exists a periodic function $v(t,x)>0$ such that
\[
v_t+c\lambda v-\nabla \cdot (A(t,x)\nabla v)-2\lambda \mathbf{e\cdot }\left(
A\nabla v\right) -\lambda \nabla \cdot (A\mathbf{e})-\lambda ^2\mathbf
e\cdot }\left( A\mathbf{e}\right) \ -f_u(t,x,0)v=\mu _c(\lambda )v.
\]
It then follows that when $c>c^{*},\;$there exist $0<\lambda _1(c)<\;\lambda
_2(c)$ such that $\mu _c(\lambda _1)=\mu _c(\lambda _2)=0$ and $\mu
_c(\lambda )>0$ for $\lambda \in (\lambda _1,\lambda _2).$ In particular,
when $c=c^{*},$ we have $\lambda _1(c^{*})=\;\lambda _2(c^{*})$ and $\mu
_c(\lambda )\leq 0$ for all $\lambda .$
The weight function is defined as in (\ref{W}). The inequality (\ref{2.18})
now becomes
\[
\left\{
\begin{array}{l}
V_t\leq \nabla \cdot \left( A\nabla V\right) +f_u^{\prime }(t,x,0)V,\vspace
4pt} \\
V(0,x)=U_0^{+}(x)-w(0,x,x_1)\geq 0.
\end{array}
\right.
\]
Correspondingly, the transformation (\ref{2.19}) is replaced by
\[
V(t,\xi )=v(t,x)e^{\lambda (x_1+ct -\xi _0)}\bar{V}(t,x)e^{-\mu
_c(\lambda )t},\;\;\lambda \in [\lambda _1(c),\lambda _2(c)].
\]
All the remaining arguments in section 2 still work. Therefore, we have the
following results.
\begin{thm}
Let $u(t,x)$ be a solution (\ref{5.1}) with initial condition
u_0(x),\;0\leq u_0(x)\leq p_2(t,x),$ and $w(t,x,x_1+ct)$ be a traveling
wavefront with wave speed $c\geq c^{*}.$ If the initial data $u_0(x),$
having at most finite points of discontinuity, satisfies
\begin{equation*}
W(x_1)(u_0(x)-w(0,x,x_1))\in L^1(\mathbb{R}^n\mathbf{),}
\label{5.3}
\end{equation*}
then the following statements are valid:
(1) For $\xi =x_1+ct\leq \xi _0$, it holds
\begin{equation*}
|u(t,x)-w(t,x,\xi )|<C(1+t)^{-\frac n2}e^{-\mu _c(\lambda )t}; \label{5.4}
\end{equation*}
(2) For $\xi =x_1+ct\geq \xi _0$, it holds
\begin{equation*}
|u(t,x)-w(t,x,\xi )|<C(1+t)^{-\frac n2}e^{-\min \{\mu _c(\lambda ),\frac{
\bar{\mu}_1}2\}t}. \label{5.5}
\end{equation*}
Here $\lambda $ is any number in the interval $\left[ \lambda _1(c),\lambda
_2(c)\right] .$
\end{thm}
\begin{thm}
Let $u(t,x)$ be a solution (\ref{5.1}) with initial condition
u_0(x),\;0\leq u_0(x)\leq p_2(t,x),$ and $w(t,x,x_1+ct)$ be a
traveling wavefront with wave speed $c\geq c^{*}.$ If the initial
data $u_0(x),$ having at most finite points of discontinuity,
satisfies
\[
W(x_1)(u_0(x)-w(0,x,x_1))\in L^1(\mathbb{R}^n\mathbf{),}
\]
then the following statements are valid:
(1) For any $c>c^{*}$ and $\lambda \in (\lambda _1(c),\lambda _2(c)),$ the
wavefront $w(t,x,x_1+ct)$ is exponentially stable in the sense that
\begin{equation*}
\sup_{x\in \mathbb{R}^n}|u(t,x)-w(t,x,x_1+ct)|\leq Ce^{-\mu
_3t},\;\;\quad \text{for some}\,\,\,\mu _3>0; \label{5.6}
\end{equation*}
(2) For $c=c^{*}$, or $c>c^{*}$ with $\lambda =\lambda _1(c),$ the wavefront
is algebraically stable in the sense that
\begin{equation*}
\sup_{x\in \mathbb{R}^n}|u(t,x)-w(t,x,x_1+ct)|\leq C(1+t)^{-\frac
n2}. \label{5.7}
\end{equation*}
\end{thm}
\section{A concluding remark}
The stability analysis above, together with asymptotic behavior of
spatially periodic wavefronts $w(x,x_1+ct)$ near the far field, can
be directly used to obtain the uniqueness of the wave with a given
speed. This then may provide a new and insightful approach that is
different from the technique in \cite{H2}.
To illustrate this idea, we first recall a result from \cite[Theorem 1.3]
{H2008}.
\noindent \textbf{Theorem C} (\cite{H2008}). \textit{If }$c>c^{*},$\textit{\
then there exists a constant }$B>0$\textit{\ such that }
\begin{equation}
w(x,\xi )\sim Be^{\lambda _1(c)\xi }v(x),\;\;\xi \rightarrow -\infty
\label{3.1}
\end{equation}
\textit{where }$v(x)$\textit{\ is the eigenfunction defined in (\ref{v}). If
}$c=c^{*},$\textit{\ then there exists a constant }$B>0$\textit{\ such that }
\begin{equation}
w(x,\xi )\sim B|\xi |e^{\lambda _1(c^{*})\xi }v(x),\;\;\xi \rightarrow
-\infty \label{3.2}
\end{equation}
\textit{where }$\lambda _1(c^{*})=\lambda _2(c^{*})=\lambda ^{*}$\textit{\
and }$v(x)$\textit{\ is the eigenfunction in (\ref{v}) with }$c=c^{*}.$
Indeed, we can further claim that if $c>c^{*},$ then the wave profile
w(x,\xi ) $ has the following asymptotical behavior:
\begin{equation}
w(x,\xi )=Be^{\lambda _1(c)\xi }v(x)+O(e^{\alpha \xi }),\;\;\;B>0,\;\alpha
>\lambda _1(c),\;\;\;\xi \rightarrow -\infty . \label{3.3}
\end{equation}
To see this, we may assume from (\ref{3.2}) that $w(x,\xi )=Be^{\lambda
_1(c)\xi }v(x)(1+$ $\bar{w}(x,\xi )).$ Substituting this into (\ref{2.5}),
we can get an equation for $\bar{w}(x,\xi ).$ For this new equation, after a
simple asymptotic analysis near $\xi =-\infty ,$ we can find that $\bar{w
(x,\xi )$ decays to zero exponentially as $\xi \rightarrow -\infty
.$ This confirms our claim. Now for any given $c>c^{*},$ suppose
that there are two spatially periodic wave profiles $w_1(x,\xi )$
and $w_2(x,\xi )$ of (\ref{1}). A shift of distance $s$ in $w_2$
then enables $w_2(x,\xi +s)$ to share the same leading term in
(\ref{3.3}) as that of $w_1.$ Let $\lambda $ in the weight function
be $\lambda _1(c).$ Then when $t=0,$ we have $e^{-\lambda
_1(c)(x_1-\xi _0)}(w_1(x,x_1)-w_2(x,x_1+s))\in L^1(\mathbb{R}^n)$.
By Theorem \ref{thm2.2}, it follows that
\[
\lim_{t\to \infty}\sup_{x\in
\mathbb{R}^n}\left|w_1(x,x_1+ct)-w_2(x,x_1+ct +s)\right|=0.
\]
Using a change of variable $\xi=x_1+ct$, we then see that
\begin{equation}\label{unique}
\lim_{t\to \infty}\sup_{(\xi,x_2,\cdots,x_n)\in
\mathbb{R}^n}\left|w_1((\xi-ct,x_2,\cdots,
x_n),\xi)-w_2((\xi-ct,x_2,\cdots, x_n),\xi+s)\right|=0.
\end{equation}
By the $L_1$-periodicity of $w_1((x_1,x_2,\cdots,x_n),\xi)$ and
$w_2((x_1,x_2,\cdots, x_n),\xi+s)$ with respect to $x_1\in
\mathbb{R}$, it easily follows from (\ref{unique}) that
\[
w_1(x,\xi)=w_2(x,\xi +s),\,
\, \forall x=(x_1,x_2,\cdots, x_n)\in \mathbb{R}^n,\, \xi\in
\mathbb{R}.
\]
This shows that for any given $c>c^{*}$, the wave profile $w(x,\xi
)$ is unique up to translation in $\xi $. In the case where
$c=c^{*},$ the same idea applies, but more delicate details are
needed.
\
\noindent {\bf Acknowledgements.} We are very grateful to two
anonymous referees for careful reading and helpful suggestions which
led to an improvement of our original manuscript.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.